Index: /tags/pap_branch_080910/psModules/src/imcombine/.cvsignore
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/.cvsignore	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/.cvsignore	(revision 19481)
@@ -0,0 +1,6 @@
+.deps
+.libs
+Makefile
+Makefile.in
+*.la
+*.lo
Index: /tags/pap_branch_080910/psModules/src/imcombine/Makefile.am
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/Makefile.am	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/Makefile.am	(revision 19481)
@@ -0,0 +1,37 @@
+noinst_LTLIBRARIES = libpsmodulesimcombine.la
+
+libpsmodulesimcombine_la_CPPFLAGS = $(SRCINC) $(PSMODULES_CFLAGS)
+libpsmodulesimcombine_la_LDFLAGS  = -release $(PACKAGE_VERSION)
+
+libpsmodulesimcombine_la_SOURCES = \
+	pmReadoutCombine.c	\
+	pmStack.c		\
+	pmStackReject.c		\
+	pmSubtraction.c		\
+	pmSubtractionEquation.c	\
+	pmSubtractionIO.c	\
+	pmSubtractionKernels.c	\
+	pmSubtractionMask.c	\
+	pmSubtractionMatch.c	\
+	pmSubtractionParams.c	\
+	pmSubtractionStamps.c	\
+	pmSubtractionThreads.c	\
+	pmPSFEnvelope.c
+
+pkginclude_HEADERS = \
+	pmImageCombine.h \
+	pmReadoutCombine.h	\
+	pmStack.h		\
+	pmStackReject.h		\
+	pmSubtraction.h		\
+	pmSubtractionEquation.h	\
+	pmSubtractionIO.h	\
+	pmSubtractionKernels.h	\
+	pmSubtractionMask.h	\
+	pmSubtractionMatch.h	\
+	pmSubtractionParams.h	\
+	pmSubtractionStamps.h	\
+	pmSubtractionThreads.h	\
+	pmPSFEnvelope.h
+
+CLEANFILES = *~
Index: /tags/pap_branch_080910/psModules/src/imcombine/demo_psftool.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/demo_psftool.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/demo_psftool.c	(revision 19481)
@@ -0,0 +1,194 @@
+// This is a very quick and dirty test program for pmPSFEnvelope
+//
+// gcc -g demo_psftool.c -o demo_psftool `psmodules-config --cflags --libs` --std=gnu99 -Wall
+//
+// ./demo_psftool test MOPS.skycell.0198767.wrp*.psf
+//
+
+#include <stdio.h>
+#include <pslib.h>
+#include <psmodules.h>
+
+// Add a single filename to the arguments as an array, so that it can be used with pmFPAfileBindFromArgs, etc
+static void fileList(const char *file, // The symbolic name for the file
+                     const char *name, // The name of the file
+                     const char *comment, // Description of the file
+                     pmConfig *config // Configuration
+    )
+{
+    psArray *files = psArrayAlloc(1); // Array with file names
+    files->data[0] = psStringCopy(name);
+    psMetadataAddArray(config->arguments, PS_LIST_TAIL, file, 0, comment, files);
+    psFree(files);
+    return;
+}
+
+
+
+int main(int argc, char *argv[])
+{
+    psLibInit(NULL);
+    pmConfig *config = pmConfigRead(&argc, argv, "PSF");
+    if (!config) {
+        psErrorStackPrint(stderr, "Error reading configuration.");
+        exit(PS_EXIT_CONFIG_ERROR);
+    }
+
+    psTraceSetLevel("psModules.imcombine", 5);
+    psTraceSetLevel("psModules.objects", 0);
+    psTraceSetLevel("psLib.math", 0);
+
+    pmModelClassInit();
+
+    psMetadataAddStr(config->arguments, PS_LIST_TAIL, "OUTPUT", 0, "Name of the output", argv[1]);
+
+    psArray *files = psArrayAlloc(argc - 2);
+    for (int i = 2; i < argc; i++) {
+        psString name = NULL;           // Name of file list
+        psStringAppend(&name, "INPUT_%d", i);
+
+        fileList(name, argv[i], "Input PSF", config);
+
+        pmFPAfile *file = pmFPAfileDefineFromArgs(NULL, config, "PSPHOT.PSF.LOAD", name);
+        psFree(name);
+        if (!file) {
+            psErrorStackPrint(stderr, "Can't define PSF file from %s --> %s", name, argv[i]);
+            psFree(files);
+            psFree(config);
+            exit(PS_EXIT_SYS_ERROR);
+        }
+
+        files->data[i - 2] = psMemIncrRefCounter(file);
+    }
+
+    pmFPAfile *outFile = pmFPAfileDefineOutput(config, NULL, "PSPHOT.PSF.SAVE");
+    if (!outFile) {
+        psErrorStackPrint(stderr, "Can't define output PSF file");
+        psFree(files);
+        psFree(config);
+        exit(PS_EXIT_SYS_ERROR);
+    }
+    outFile->save = true;
+
+    // XXX This is a bit dodgy; should be more rigorous for a real system
+    {
+        pmFPAview *phuView = pmFPAviewAlloc(0);
+        phuView->chip = 0;
+        if (!pmFPAAddSourceFromView(outFile->fpa, "Envelope PSF", phuView, outFile->format)) {
+            psError(PS_ERR_UNKNOWN, false, "Unable to add PHU to output.");
+            psFree(phuView);
+            return false;
+        }
+        psFree(phuView);
+    }
+
+    pmFPAview *view = pmFPAviewAlloc(0);
+
+    if (!pmFPAfileIOChecks(config, view, PM_FPA_BEFORE)) {
+        psErrorStackPrint(stderr, "Problem in I/O");
+        psFree(view);
+        psFree(files);
+        psFree(config);
+        exit(PS_EXIT_SYS_ERROR);
+    }
+
+    pmChip *chip;
+    while ((chip = pmFPAviewNextChip(view, outFile->fpa, 1)) != NULL) {
+        if (!pmFPAfileIOChecks(config, view, PM_FPA_BEFORE)) {
+            psErrorStackPrint(stderr, "Problem in I/O");
+            psFree(view);
+            psFree(files);
+            psFree(config);
+            exit(PS_EXIT_SYS_ERROR);
+        }
+
+#if 0
+        pmFPAfileActivate(config->files, "PSPHOT.PSF.SAVE", false);
+        pmCell *cell;
+        while ((cell = pmFPAviewNextCell(view, outFile->fpa, 1)) != NULL) {
+            if (!pmFPAfileIOChecks(config, view, PM_FPA_BEFORE)) {
+                psErrorStackPrint(stderr, "Problem in I/O");
+                psFree(view);
+                psFree(files);
+                psFree(config);
+                exit(PS_EXIT_SYS_ERROR);
+            }
+        }
+#endif
+
+        psArray *inputs = psArrayAlloc(files->n);
+        int numCols = 4501, numRows = 4751;
+        for (int i = 0; i < files->n; i++) {
+            pmFPAfile *file = files->data[i];
+            pmChip *chip = pmFPAviewThisChip(view, file->fpa);
+
+            pmPSF *psf = psMetadataLookupPtr(NULL, chip->analysis, "PSPHOT.PSF");
+            if (!psf) {
+                psErrorStackPrint(stderr, "Can't find PSF in file %d", i);
+                psFree(inputs);
+                psFree(files);
+                psFree(config);
+                exit(PS_EXIT_PROG_ERROR);
+            }
+
+#if 0
+            pmHDU *hdu = pmHDUGetLowest(file->fpa, chip, NULL);
+            int imaxis1 = psMetadataLookupS32(NULL, hdu->header, "IMAXIS1");
+            int imaxis2 = psMetadataLookupS32(NULL, hdu->header, "IMAXIS2");
+            if (imaxis1 == 0 || imaxis2 == 0) {
+                psErrorStackPrint(stderr, "Size of image %d can't be determined.", i);
+                psFree(inputs);
+                psFree(files);
+                psFree(config);
+                exit(PS_EXIT_SYS_ERROR);
+            }
+            if (numCols == 0 && numRows == 0) {
+                numCols = imaxis1;
+                numRows = imaxis2;
+            } else if (imaxis1 != numCols || imaxis2 != numRows) {
+                psErrorStackPrint(stderr, "Image %d differs in size: %dx%d vs %dx%d",
+                                  i, imaxis1, imaxis2, numCols, numRows);
+                psFree(inputs);
+                psFree(files);
+                psFree(config);
+                exit(PS_EXIT_SYS_ERROR);
+            }
+#endif
+
+            inputs->data[i] = psMemIncrRefCounter(psf);
+        }
+
+        pmPSF *psf = pmPSFEnvelope(numCols, numRows, inputs, 5, 20, "PS_MODEL_RGAUSS", 3, 3);
+        psFree(inputs);
+        if (!psf) {
+            psErrorStackPrint(stderr, "Can't generate envelope PSF.");
+            psFree(config);
+            exit(PS_EXIT_SYS_ERROR);
+        }
+
+        psMetadataAddPtr(chip->analysis, PS_LIST_TAIL, "PSPHOT.PSF", PS_DATA_UNKNOWN, "Envelope PSF", psf);
+        psFree(psf);
+        chip->data_exists = true;
+
+        if (!pmFPAfileIOChecks(config, view, PM_FPA_AFTER)) {
+            psErrorStackPrint(stderr, "Problem in I/O");
+            psFree(view);
+            psFree(files);
+            psFree(config);
+            exit(PS_EXIT_SYS_ERROR);
+            return false;
+        }
+    }
+
+    if (!pmFPAfileIOChecks(config, view, PM_FPA_AFTER)) {
+        psErrorStackPrint(stderr, "Problem in I/O");
+        psFree(view);
+        psFree(files);
+        psFree(config);
+        exit(PS_EXIT_SYS_ERROR);
+    }
+
+    psFree(config);
+
+    exit(PS_EXIT_SUCCESS);
+}
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmImageCombine.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmImageCombine.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmImageCombine.c	(revision 19481)
@@ -0,0 +1,683 @@
+/** @file  pmImageCombine.c
+ *
+ *  This file will perform image combination of several images of the
+ *  same field, produce a list of questionable pixels, then tag some
+ *  of those pixels as cosmic rays.
+ *
+ *  @author Paul Price, IfA (original prototype)
+ *  @author GLG, MHPCC
+ *
+ *  @version $Revision: 1.12 $ $Name: not supported by cvs2svn $
+ *  @date $Date: 2007-04-04 22:42:48 $
+ *
+ *  XXX: pmRejectPixels() has a known bug with the pmImageTransform() call.
+ *
+ *  Copyright 2004 Maui High Performance Computing Center, University of Hawaii
+ *
+ */
+
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <config.h>
+#include <stdio.h>
+#include <math.h>
+#include "pslib.h"
+
+#define PIXEL_LIST_BUFFER 100           // Size of the pixel list buffer
+
+// Data structure for use as a buffer in combining pixels
+typedef struct
+{
+    psVector *pixels;                   // Pixel values
+    psVector *masks;                    // Pixel masks
+    psVector *errors;                   // Pixel errors
+    psStats *stats;                     // Statistics to use with combination
+}
+combineBuffer;
+
+void combineBufferFree(combineBuffer *buffer)
+{
+    psFree(buffer->pixels);
+    psFree(buffer->masks);
+    psFree(buffer->errors);
+    psFree(buffer->stats);
+}
+
+combineBuffer *combineBufferAlloc(long numImages // Number of images that will be combined
+                                 )
+{
+    combineBuffer *buffer = psAlloc(sizeof(combineBuffer));
+    psMemSetDeallocator(buffer, (psFreeFunc)combineBufferFree);
+
+    buffer->pixels = psVectorAlloc(numImages, PS_TYPE_F32);
+    buffer->masks = psVectorAlloc(numImages, PS_TYPE_MASK);
+    buffer->errors = psVectorAlloc(numImages, PS_TYPE_F32);
+    buffer->stats = psStatsAlloc(PS_STAT_SAMPLE_MEAN | PS_STAT_SAMPLE_STDEV);
+
+    return buffer;
+}
+
+
+static bool combinePixels(psImage *combine, // Combined image, for output
+                          psArray *questionablePixels, // Array of rejection masks
+                          int x, int y, // Position in the images
+                          const psArray *images, // Array of input images
+                          const psArray *errors, // Array of input error images
+                          const psArray *masks, // Array of input masks
+                          psU32 maskVal, // Mask value
+                          psS32 numIter, // Number of rejection iterations
+                          psF32 sigmaClip, // Number of standard deviations at which to reject
+                          combineBuffer *buffer // Buffer for combination; to avoid multiple allocations
+                         )
+{
+    assert(combine);
+    assert(x >= 0 && x < combine->numCols);
+    assert(y >= 0 && y < combine->numRows);
+    assert(images);
+    int numImages = images->n;          // Number of images to combine
+    if (masks) {
+        assert(masks->n == numImages);
+    }
+    if (errors) {
+        assert(errors->n == numImages);
+    }
+    assert(numIter >= 0);
+    assert(sigmaClip > 0);
+    assert(!questionablePixels || (questionablePixels && questionablePixels->n == numImages));
+
+    if (buffer) {
+        psMemIncrRefCounter(buffer);
+    } else {
+        buffer = combineBufferAlloc(numImages);
+    }
+
+    psVector *pixelData = buffer->pixels; // Values for the pixel of interest
+    psVector *pixelErrors = buffer->errors; // Errors for the pixel of interest
+    psVector *pixelMasks = buffer->masks; // Masks for the pixel of interest
+    psStats *stats = buffer->stats;     // Statistics for combination
+
+    //
+    // Loop through each image, extract the pixel/mask/error data into psVectors.
+    //
+    if (!masks) {
+        psVectorInit(pixelMasks, 0);
+    }
+    if (!errors) {
+        pixelErrors = NULL;
+    }
+    for (int i = 0; i < numImages; i++) {
+        // Set the pixel data
+        psImage *image = images->data[i]; // Image of interest
+        pixelData->data.F32[i] = image->data.F32[y][x];
+        // Set the pixel mask data, if necessary
+        if (masks) {
+            psImage *mask = masks->data[i]; // Mask of interest
+            pixelMasks->data.U8[i] = mask->data.U8[y][x];
+        }        // Set the pixel error data, if necessary
+        if (errors) {
+            psImage *error = errors->data[i]; // Error image of interest
+            pixelErrors->data.F32[i] = error->data.F32[y][x];
+        }
+    }
+
+    //
+    // Iterate on the pixels, rejecting outliers
+    //
+    for (int iter = 0; iter < numIter; iter++) {
+        // Combine all the pixels, using the specified stat.
+        if (!psVectorStats(stats, pixelData, pixelErrors, pixelMasks, maskVal)) {
+            combine->data.F32[y][x] = NAN;
+            psFree(buffer);
+            return false;
+        }
+        float combinedPixel = stats->sampleMean; // Value of the combination
+
+        if (iter == 0) {
+            // Save the value produced with no rejection, since it may be useful later
+            // (if the rejection turns out to be unnecessary)
+            combine->data.F32[y][x] = combinedPixel;
+        }
+
+        //
+        // Reject all pixels that lie more that sigmaClip standard deviations from
+        // the combined pixel value.
+        //
+        int numRejects = 0;     // Number of rejections
+        float stdev = stats->sampleStdev;
+        for (int i = 0; i < numImages; i++) {
+            if (!(pixelMasks->data.U8[i] & maskVal) &&
+                    fabs(pixelData->data.F32[i] - combinedPixel) > sigmaClip * stdev) {
+                // Reject pixel as questionable
+                numRejects++;
+                pixelMasks->data.U8[i] = maskVal;
+                if (questionablePixels) {
+                    // Mark the pixel as questionable
+                    psPixels *qp = questionablePixels->data[i]; // Questionable pixels for this image
+                    int qpNum = qp->n; // Number of QPs in the image of interest
+                    if (qpNum >= qp->nalloc) {
+                        // Grow dynamically, if required
+                        qp = psPixelsRealloc(qp, qp->nalloc + PIXEL_LIST_BUFFER);
+                        questionablePixels->data[i] = qp;
+                    }
+                    qp->data[qpNum].x = x;
+                    qp->data[qpNum].y = y;
+                    qp->n++;
+                }
+            }
+        }
+
+        //
+        // If the number of rejected pixels is zero, then there's no point continuing the loop.
+        //
+        if (numRejects == 0) {
+            break;
+        }
+
+        //
+        // XXX: Is it possible to have all pixels rejected?  If so, we should exit the loop.
+        //
+    }
+
+    psFree(buffer);
+    return true;
+}
+
+psImage *pmCombineImages(
+    psImage *combine,                   ///< Combined image (output)
+    psArray **questionablePixels,       ///< Array of rejection masks
+    const psArray *images,              ///< Array of input images
+    const psArray *errors,              ///< Array of input error images
+    const psArray *masks,               ///< Array of input masks
+    psU32 maskVal,                      ///< Mask value
+    const psPixels *pixels,             ///< Pixels to combine
+    psS32 numIter,                      ///< Number of rejection iterations
+    psF32 sigmaClip                     ///< Number of standard deviations at which to reject
+)
+{
+    psTrace("psModules.imcombine", 3, "Calling pmCombineImages(%ld)\n", images->n);
+
+    PS_ASSERT_ARRAY_NON_NULL(images, NULL);
+    PS_ASSERT_INT_POSITIVE(images->n, NULL);
+    long numImages = images->n;          // Number of images
+    int numCols = ((psImage*)images->data[0])->numCols; // Size in x
+    int numRows = ((psImage*)images->data[0])->numRows; // Size in y
+
+    if (combine) {
+        PS_ASSERT_IMAGE_NON_NULL(combine, NULL);
+        PS_ASSERT_IMAGE_SIZE(combine, numCols, numRows, NULL);
+    }
+    if (questionablePixels && !*questionablePixels) {
+        PS_ASSERT_ARRAY_NON_NULL(*questionablePixels, NULL);
+        PS_ASSERT_ARRAY_SIZE(*questionablePixels, numImages, 0);
+    }
+    for (int i = 1; i < images->n; i++) {
+        psImage *image = images->data[i]; // Image of interest
+        PS_ASSERT_IMAGE_TYPE(image, PS_TYPE_F32, NULL);
+        PS_ASSERT_IMAGE_SIZE(image, numCols, numRows, NULL);
+    }
+    if (errors) {
+        PS_ASSERT_ARRAYS_SIZE_EQUAL(images, errors, NULL);
+        for (int i = 0; i < images->n; i++) {
+            psImage *error = errors->data[i];
+            PS_ASSERT_IMAGE_SIZE(error, numCols, numRows, NULL);
+            PS_ASSERT_IMAGE_TYPE(error, PS_TYPE_F32, NULL);
+        }
+    }
+    if (masks) {
+        PS_ASSERT_ARRAYS_SIZE_EQUAL(images, masks, NULL);
+        for (int i = 0; i < images->n; i++) {
+            psImage *mask  = masks->data[i];
+            PS_ASSERT_IMAGE_SIZE(mask, numCols, numRows, NULL);
+            PS_ASSERT_IMAGE_TYPE(mask, PS_TYPE_MASK, NULL);
+        }
+    }
+    PS_ASSERT_INT_POSITIVE(numIter, NULL);
+    PS_ASSERT_FLOAT_LARGER_THAN(sigmaClip, 0.0, NULL);
+
+    // Allocate and initialize the combined image, if necessary.
+    if (!combine) {
+        combine = psImageAlloc(numCols, numRows, PS_TYPE_F32);
+        if (pixels) {
+            // Set everything we're not working on to NAN
+            psImageInit(combine, NAN);
+        }
+    }
+
+    //
+    // Allocate the questionablePixels psArray, if necesssary, then create a psPixels
+    // struct for each image.
+    //
+    if (questionablePixels) {
+        if (*questionablePixels == NULL) {
+            *questionablePixels = psArrayAlloc(numImages);
+        } else if ((*questionablePixels)->n != numImages) {
+            *questionablePixels = psArrayRealloc(*questionablePixels, numImages);
+        }
+        for (int i = 0; i < numImages; i++) {
+            psFree((*questionablePixels)->data[i]);
+            (*questionablePixels)->data[i] = psPixelsAlloc(PIXEL_LIST_BUFFER);
+        }
+    }
+
+    combineBuffer *buffer = combineBufferAlloc(numImages); // Buffer for combination
+
+    if (pixels) {
+        // Only those specified pixels should be combined.
+
+        for (int p = 0; p < pixels->n; p++) {
+            int x = pixels->data[p].x; // Column of interest
+            int y = pixels->data[p].y; // Row of interest
+
+            if (!combinePixels(combine, questionablePixels ? *questionablePixels : NULL, x, y,
+                               images, errors, masks, maskVal, numIter, sigmaClip, buffer)) {
+                // Bad pixel --- no big deal
+                psErrorClear();
+            }
+        }
+    } else {
+        //
+        // We get here if there is a NULL list of pixels to combine.
+        // Therefore, we combine all pixels in all images.
+        //
+
+        //
+        // Loop over all pixels in all images, set the appropriate data, mask,
+        // error vectors, call psVectorStats(), and set the result in the
+        // combine image.
+        //
+        for (int y = 0; y < numRows; y++) {
+            for (int x = 0; x < numCols; x++) {
+                if (!combinePixels(combine, questionablePixels ? *questionablePixels : NULL, x, y,
+                                   images, errors, masks, maskVal, numIter, sigmaClip, buffer)) {
+                    // Bad pixel --- no big deal
+                    psErrorClear();
+                }
+            }
+        }
+    }
+
+    psFree(buffer);
+
+    psTrace("psModules.imcombine", 3, "Exiting pmCombineImages(%ld)\n", images->n);
+    return combine;
+}
+
+
+/******************************************************************************
+XXX: Directly from Paul Price
+ *****************************************************************************/
+static psF32 CalcGradient(
+    psImage *image,
+    psImage *imageMask,
+    psS32 x,
+    psS32 y
+)
+{
+    psTrace("psModules.imcombine", 4, "Calling CalcGradient(%d, %d)\n", x, y);
+    int num = 0;
+    psVector *pixels = psVectorAlloc(8, PS_TYPE_F32); // Array of pixels
+    psVector *mask = psVectorAlloc(8, PS_TYPE_U8); // Corresponding mask
+
+    // Get limits
+    int xMin = PS_MAX(x - 1, 0);
+    int xMax = PS_MIN(x + 1, image->numCols - 1);
+    int yMin = PS_MAX(y - 1, 0);
+    int yMax = PS_MIN(y + 1, image->numRows - 1);
+    if (imageMask != NULL) {
+        for (int j = yMin; j <= yMax; j++) {
+            for (int i = xMin; i <= xMax; i++) {
+                if ((i != x) && (j != y) && (0 == imageMask->data.U8[j][i])) {
+                    pixels->data.F32[num] = image->data.F32[j][i];
+                    mask->data.U8[num] = 0;
+                    num++;
+                } else {
+                    mask->data.U8[num] = 1;
+                }
+            }
+        }
+    } else {
+        //
+        // This code is simply the previous loop without the imageMask.
+        // XXX: Consider restructuring this.
+        //
+        for (int j = yMin; j <= yMax; j++) {
+            for (int i = xMin; i <= xMax; i++) {
+                if ((i != x) && (j != y)) {
+                    pixels->data.F32[num] = image->data.F32[j][i];
+                    mask->data.U8[num] = 0;
+                    num++;
+                } else {
+                    mask->data.U8[num] = 1;
+                }
+            }
+        }
+    }
+
+    pixels->n = num;
+    mask->n = num;
+
+    // Get the median
+    psStats *stats = psStatsAlloc(PS_STAT_SAMPLE_MEDIAN);
+    psVectorStats(stats, pixels, NULL, mask, 1);
+    float median = stats->sampleMedian;
+    psFree(stats);
+    psFree(pixels);
+    psFree(mask);
+
+    psTrace("psModules.imcombine", 4, "Exiting CalcGradient(%d, %d)\n", x, y);
+    return(median / image->data.F32[y][x]);
+}
+
+/******************************************************************************
+DetermineRegion(image, myOutToIn): for a psImage and a psPlaneTransform to that
+image, this routine determines the size of the input image which maps to that
+image, and returns the result in a psRegion struct.
+
+XXX: Basically, this routine is only guaranteed to work if the transform is
+linear.
+
+XXX: Shouldn't this functionality be part of psImageTransform()?
+ *****************************************************************************/
+static psRegion DetermineRegion(psImage *image,
+                                psPlaneTransform *myOutToIn)
+{
+    psTrace("psModules.imcombine", 4, "Calling DetermineRegion()\n");
+    psRegion myRegion;
+    myRegion.x0 = PS_MAX_F32;
+    myRegion.x1 = PS_MIN_F32;
+    myRegion.y0 = PS_MAX_F32;
+    myRegion.y1 = PS_MIN_F32;
+    psPlane in;
+    psPlane out;
+
+    in.x = 0.0;
+    in.y = 0.0;
+
+    psPlaneTransformApply(&out, myOutToIn, &in);
+    if (out.x < myRegion.x0) {
+        myRegion.x0 = out.x;
+    }
+    if (out.x > myRegion.x1) {
+        myRegion.x1 = out.x;
+    }
+    if (out.y < myRegion.y0) {
+        myRegion.y0 = out.y;
+    }
+    if (out.y > myRegion.y1) {
+        myRegion.y1 = out.y;
+    }
+
+    in.x = (psF32) (image->numCols);
+    in.y = 0.0;
+    psPlaneTransformApply(&out, myOutToIn, &in);
+    if (out.x < myRegion.x0) {
+        myRegion.x0 = out.x;
+    }
+    if (out.x > myRegion.x1) {
+        myRegion.x1 = out.x;
+    }
+    if (out.y < myRegion.y0) {
+        myRegion.y0 = out.y;
+    }
+    if (out.y > myRegion.y1) {
+        myRegion.y1 = out.y;
+    }
+
+    in.x = (psF32) (image->numCols);
+    ;
+    in.y = 0.0;
+    psPlaneTransformApply(&out, myOutToIn, &in);
+    if (out.x < myRegion.x0) {
+        myRegion.x0 = out.x;
+    }
+    if (out.x > myRegion.x1) {
+        myRegion.x1 = out.x;
+    }
+    if (out.y < myRegion.y0) {
+        myRegion.y0 = out.y;
+    }
+    if (out.y > myRegion.y1) {
+        myRegion.y1 = out.y;
+    }
+
+    in.x = (psF32) (image->numCols);
+    in.y = (psF32) (image->numRows);
+    psPlaneTransformApply(&out, myOutToIn, &in);
+    if (out.x < myRegion.x0) {
+        myRegion.x0 = out.x;
+    }
+    if (out.x > myRegion.x1) {
+        myRegion.x1 = out.x;
+    }
+    if (out.y < myRegion.y0) {
+        myRegion.y0 = out.y;
+    }
+    if (out.y > myRegion.y1) {
+        myRegion.y1 = out.y;
+    }
+
+    psTrace("psModules.imcombine", 4, "Exiting DetermineRegion()\n");
+    return(myRegion);
+}
+
+/******************************************************************************
+XXX: Don't we have a psLib function for this?
+ *****************************************************************************/
+static psImage *ImageConvertF32(psImage *image)
+{
+    psTrace("psModules.imcombine", 4, "Calling ImageConvertF32()\n");
+    psImage *imgF32 = psImageAlloc(image->numCols, image->numRows, PS_TYPE_F32);
+
+    for (psS32 i = 0 ; i < image->numRows ; i++) {
+        for (psS32 j = 0 ; j < image->numCols ; j++) {
+            imgF32->data.F32[i][j] = (psF32) image->data.U8[i][j];
+        }
+    }
+
+    psTrace("psModules.imcombine", 4, "Exiting ImageConvertF32()\n");
+    return(imgF32);
+}
+
+
+//
+// The following macros define how big the initial pixel list will be, and
+// how much it should be incremented when realloc'ed.
+//
+#define PS_REJECT_PIXEL_INITIAL_PIXEL_LIST_LENGTH 100
+#define PS_REJECT_PIXEL_INITIAL_PIXEL_LIST_LENGTH_INC 100
+/******************************************************************************
+pmRejectPixels(images, errors, inToOut, outToIn, rejThreshold,
+gradLimit)
+
+XXX: Optimization: we don't need to transform the entire mask image.
+XXX: The inToOut and outToIn transforms are confusing.  Verify that what
+     I think they mean syncs with PWP.
+ *****************************************************************************/
+psArray *pmRejectPixels(
+    const psArray *images,              ///< Array of input images
+    const psArray *masks,               ///< Array of input image masks
+    const psArray *errors,              ///< The pixels which were rejected in the combination
+    const psArray *inToOut,             ///< Transformation from input to output system
+    const psArray *outToIn,             ///< Transformation from output to input system
+    psF32 rejThreshold,                 ///< Rejection threshold
+    psF32 gradLimit                     ///< Gradient limit
+)
+{
+    psTrace("psModules.imcombine", 3, "Calling pmRejectPixels()\n");
+    PS_ASSERT_PTR_NON_NULL(images, NULL);
+    for (psS32 im = 0 ; im < images->n ; im++) {
+        psImage *tmpImage = (psImage *) images->data[im];
+        PS_ASSERT_IMAGE_NON_NULL(tmpImage, NULL);
+        PS_ASSERT_IMAGE_NON_EMPTY(tmpImage, NULL);
+        PS_ASSERT_IMAGE_TYPE(tmpImage, PS_TYPE_F32, NULL);
+        if (masks != NULL) {
+            PS_ASSERT_INT_EQUAL(images->n, masks->n, NULL);
+            psImage *tmpMask = (psImage *) masks->data[im];
+            PS_ASSERT_IMAGE_NON_NULL(tmpMask, NULL);
+            PS_ASSERT_IMAGE_NON_EMPTY(tmpMask, NULL);
+            PS_ASSERT_IMAGE_TYPE(tmpMask, PS_TYPE_F32, NULL);
+            PS_ASSERT_IMAGES_SIZE_EQUAL(tmpImage, tmpMask, NULL);
+        }
+        PS_ASSERT_IMAGES_SIZE_EQUAL(((psImage *) images->data[0]), tmpImage, NULL);
+    }
+    PS_ASSERT_PTR_NON_NULL(errors, NULL);
+    PS_ASSERT_PTR_NON_NULL(inToOut, NULL);
+    PS_ASSERT_PTR_NON_NULL(outToIn, NULL);
+    // Ensure that the psArray parameters have an element for each image.
+    psS32 numImages = images->n;
+    PS_ASSERT_INT_EQUAL(numImages, errors->n, NULL);
+    PS_ASSERT_INT_EQUAL(numImages, inToOut->n, NULL);
+    PS_ASSERT_INT_EQUAL(numImages, outToIn->n, NULL);
+
+    //
+    // Create the psArray of psPixelLists, one for each image, for rejected pixels.
+    //
+    psArray *rejects = psArrayAlloc(numImages);
+    for (psS32 im = 0 ; im < numImages ; im++) {
+        rejects->data[im] = (psPtr *) psPixelsAlloc(PS_REJECT_PIXEL_INITIAL_PIXEL_LIST_LENGTH);
+        ((psPixels *)(rejects->data[im]))->n = ((psPixels *)(rejects->data[im]))->nalloc;
+        psPixels *pixels = (psPixels *) rejects->data[im];
+        pixels->n = 0;
+    }
+    //
+    // rPtr is used to maintain a count of the questionable pixels for each image.
+    //
+    psVector *rPtr = psVectorAlloc(numImages, PS_TYPE_S32);
+    psVectorInit(rPtr, 0);
+
+    psS32 numCols = ((psImage *) images->data[0])->numCols;
+    psS32 numRows = ((psImage *) images->data[0])->numRows;
+    psRegion myRegion = psRegionSet(0, numCols-1, 0, numRows-1);
+    psU32 maskVal = 1;  // XXX: Is this appropriate?
+
+    psPlane *inCoords = psAlloc(sizeof(psPlane));
+    psPlane *outCoords = psAlloc(sizeof(psPlane));
+
+    for (psS32 im = 0 ; im < numImages ; im++) {
+        //
+        // Extract data from psArrays.
+        //
+        psPixels *pixelList = (psPixels *) errors->data[im];
+
+        psImage *currImage = (psImage *) images->data[im];
+        myRegion.x0 = 0;
+        myRegion.x1 = currImage->numCols;
+        myRegion.y0 = 0;
+        myRegion.y1 = currImage->numRows;
+        psPlaneTransform *myInToOut = (psPlaneTransform *) inToOut->data[im];
+        psPlaneTransform *myOutToIn = (psPlaneTransform *) outToIn->data[im];
+
+        //
+        // Create a psU8 mask image from the list of cosmic pixels.
+        //
+        psImage *maskImage = NULL;
+        maskImage = psPixelsToMask(maskImage, pixelList, myRegion, maskVal);
+        psImage *maskImageF32 = ImageConvertF32(maskImage);
+
+        //
+        // Transform that mask image into detector coordinate space
+        //
+        psRegion myRegionXForm = DetermineRegion(maskImageF32, myOutToIn);
+        psImage *transformedImage = psImageTransform(NULL, NULL, maskImageF32, NULL,
+                                    0, myOutToIn, myRegionXForm, NULL,
+                                    PS_INTERPOLATE_BILINEAR, 0);
+
+        //
+        // Loop over all cosmic pixels.  Transform their coords to detector space.
+        // If the value of the transformed mask is larger than rejThreshold, then
+        // this might be a cosmic ray pixel.  We then calculate the mean gradient
+        // in other images.
+        //
+
+        psImageInterpolateOptions *interp = psImageInterpolateOptionsAlloc(PS_INTERPOLATE_BILINEAR,
+                                                                           transformedImage, NULL, NULL,
+                                                                           0, 0.0, 0.0, 0, 0, 0.0);
+
+        for (psS32 p = 0 ; p < pixelList->n ; p++) {
+            inCoords->x = 0.5 + (psF32) (pixelList->data[p]).x;
+            inCoords->y = 0.5 + (psF32) (pixelList->data[p]).y;
+            psPlaneTransformApply(outCoords, myInToOut, inCoords);
+            double maskVal;
+            if (!psImageInterpolate(&maskVal, NULL, NULL, outCoords->x, outCoords->y, interp)) {
+                psError(PS_ERR_UNKNOWN, false, "Unable to interpolate image.");
+                psFree(interp);
+                psFree(maskImage);
+                psFree(maskImageF32);
+                psFree(transformedImage);
+                psFree(inCoords);
+                psFree(outCoords);
+                psFree(rejects);
+                return NULL;
+            }
+            if (maskVal > rejThreshold) {
+
+                // This is a possible cosmic array pixel.  We must calculate the gradient
+                // at this location in all input images.
+                psF32 meanGrads = 0.0;
+                psS32 numGrads = 0;
+                //
+                // Loop through all other images, calculate their mean gradient.
+                //
+                for (psS32 otherImg = 0 ; otherImg < numImages ; otherImg++) {
+                    if (im != otherImg) {
+                        // Map the outCoords to inCoords that for otherImg space.
+                        psImage *tmpMask = NULL;
+                        if (masks != NULL) {
+                            tmpMask = masks->data[otherImg];
+                        }
+                        psPlaneTransformApply(inCoords,
+                                              (psPlaneTransform * )outToIn->data[otherImg],
+                                              outCoords);
+                        psS32 xPix = (int)(inCoords->x + 0.5);
+                        psS32 yPix = (int)(inCoords->y + 0.5);
+                        if ((xPix >= 0) && (xPix <= ((psImage*)(images->data[otherImg]))->numCols - 1) &&
+                                (yPix >= 0) && (yPix <= ((psImage*)(images->data[otherImg]))->numRows - 1)) {
+                            meanGrads += CalcGradient(images->data[otherImg], tmpMask, xPix, yPix);
+                            numGrads++;
+                        }
+                    }
+                }
+                if (numGrads > 0) {
+                    meanGrads /= (psF32) numGrads;
+                } else {
+                    // XXX: my idea.  Verify with PWP:
+                    meanGrads = 1.0 + gradLimit;
+                }
+
+                // XXX: The SDRS and the prototype code differ significantly here:
+                // if (CalcGradient(inputs->data[i], pixelList->data.x, pixelList->data.y) < (gradLimit * meanGrads)) {
+                if (meanGrads < gradLimit) {
+                    //
+                    // Add this to the list of questionable pixels.  We must ensure that the
+                    // pixelList is large enough; if not, we realloc()
+                    //
+                    psS32 ptr = rPtr->data.S32[im];
+                    psPixels *pixelListPtr = (psPixels *) rejects->data[im];
+                    if (ptr >= pixelListPtr->nalloc) {
+                        rejects->data[im] = (psPtr *) psPixelsRealloc(((psPixels *) rejects->data[im]),
+                                            ((((psPixels *) rejects->data[im])->nalloc) + PS_REJECT_PIXEL_INITIAL_PIXEL_LIST_LENGTH_INC));
+                        // XXX: Can the realloc() fail?  Must we check for NULL?
+                    }
+
+                    ((psPixels *) rejects->data[im])->data[ptr].x = (pixelList->data[p]).x;
+                    ((psPixels *) rejects->data[im])->data[ptr].y = (pixelList->data[p]).y;
+                    (rPtr->data.S32[im])++;
+                    // XXX: this pixel ->n increment is wierd
+                    (((psPixels *) rejects->data[im])->n)++;
+                }
+            }
+        }
+
+        psFree(interp);
+        psFree(maskImage);
+        psFree(maskImageF32);
+        psFree(transformedImage);
+    }
+
+    psFree(inCoords);
+    psFree(outCoords);
+    psTrace("psModules.imcombine", 3, "Exiting pmRejectPixels()\n");
+    return(rejects);
+}
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmImageCombine.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmImageCombine.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmImageCombine.h	(revision 19481)
@@ -0,0 +1,44 @@
+/* @file  pmImageCombine.h
+ *
+ * This file will perform image combination of several images of the
+ * same field, produce a list of questionable pixels, then tag some
+ * of those pixels as cosmic rays.
+ *
+ * @author Paul Price, IfA (original prototype)
+ * @author GLG, MHPCC
+ *
+ * @version $Revision: 1.6 $ $Name: not supported by cvs2svn $
+ * @date $Date: 2007-03-30 21:12:56 $
+ * Copyright 2004 Maui High Performance Computing Center, University of Hawaii
+ */
+
+#ifndef PM_IMAGE_COMBINE_H
+#define PM_IMAGE_COMBINE_H
+
+/// @addtogroup imcombine Image Combinations
+/// @{
+
+psImage *pmCombineImages(
+    psImage *combine,                   ///< Combined image (output)
+    psArray **questionablePixels,       ///< Array of rejection masks
+    const psArray *images,              ///< Array of input images
+    const psArray *errors,              ///< Array of input error images
+    const psArray *masks,               ///< Array of input masks
+    psU32 maskVal,                      ///< Mask value
+    const psPixels *pixels,             ///< Pixels to combine
+    psS32 numIter,                      ///< Number of rejection iterations
+    psF32 sigmaClip                     ///< Number of standard deviations at which to reject
+);
+
+psArray *pmRejectPixels(
+    const psArray *images,              ///< Array of input images
+    const psArray *masks,               ///< Array of input image masks
+    const psArray *errors,              ///< The pixels which were rejected in the combination
+    const psArray *inToOut,             ///< Transformation from input to output system
+    const psArray *outToIn,             ///< Transformation from output to input system
+    psF32 rejThreshold,                 ///< Rejection threshold
+    psF32 gradLimit                     ///< Gradient limit
+);
+
+/// @}
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmPSFEnvelope.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmPSFEnvelope.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmPSFEnvelope.c	(revision 19481)
@@ -0,0 +1,326 @@
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+
+#include <stdio.h>
+#include <string.h>
+#include <pslib.h>
+
+#include "pmHDU.h"
+#include "pmFPA.h"
+#include "pmReadoutFake.h"
+
+#include "pmMoments.h"
+#include "pmResiduals.h"
+#include "pmGrowthCurve.h"
+#include "pmTrend2D.h"
+#include "pmPSF.h"
+#include "pmModel.h"
+#include "pmModelClass.h"
+#include "pmModelUtils.h"
+#include "pmPeaks.h"
+#include "pmSource.h"
+#include "pmSourceUtils.h"
+#include "pmSourceFitModel.h"
+#include "pmPSFtry.h"
+
+
+#include "pmPSFEnvelope.h"
+
+
+
+//#define TESTING                         // Enable test output
+#define PEAK_FLUX 1.0e4                 // Peak flux for each source
+#define SKY_VALUE 0.0e0                 // Sky value for fake image
+#define WEIGHT_VAL 3.0                  // Weighting for image
+#define WEIGHT_FACTOR 10.0              // Factor to multiply image by to get weighting
+#define PSF_STATS PS_STAT_SAMPLE_MEDIAN | PS_STAT_SAMPLE_STDEV // Statistics options for measuring PSF
+#define SOURCE_FIT_ITERATIONS 100       // Number of iterations for source fitting
+
+
+// XXX To do:
+//
+// * PSF variation when only a portion of the image is present (e.g., the edge of an FPA overlapping a
+// skycell) may mean a disastrously weird PSF in the missing regions.  To counter this, get a region of
+// validity for each PSF (perhaps from an associated mask, or have the user work it out), and taper the PSF
+// when outside this region (perhaps multiply the peak flux by a Gaussian whose arguments are the distance
+// from the valid region, and a width that the user supplies).
+
+
+// We deliberately do not include the calculation of and storing of residuals (data - model) for the PSF
+// model, because (1) there is no code in psModules to do this, and we're not going to implement it here; and
+// (2) this is intended to generate "nice" or "ideal" PSFs to feed into pmSubtraction (PSF matching code), so
+// any residuals will hopefully be dealt with by that.
+
+
+pmPSF *pmPSFEnvelope(int numCols, int numRows, // Size of original image
+                     const psArray *inputs, // Input PSF models
+                     int instances, // Number of instances per dimension
+                     int radius,        // Radius of each PSF
+                     const char *modelName,// Name of PSF model to use
+                     int xOrder, int yOrder // Order for PSF variation fit
+                     )
+{
+    PS_ASSERT_INT_POSITIVE(numCols, NULL);
+    PS_ASSERT_INT_POSITIVE(numRows, NULL);
+    PS_ASSERT_ARRAY_NON_NULL(inputs, NULL);
+    PS_ASSERT_INT_POSITIVE(instances, NULL);
+    PS_ASSERT_INT_POSITIVE(radius, NULL);
+    PS_ASSERT_STRING_NON_EMPTY(modelName, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(xOrder, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(yOrder, NULL);
+
+    float xOrigSpacing = (float)(numCols - 2 * radius) / (float)(instances - 1); // Spacing between instances
+    float yOrigSpacing = (float)(numRows - 2 * radius) / (float)(instances - 1); // Spacing between instances
+    int fakeSpacing = 2 * radius + 1;   // Spacing between instances (x and y) in the fake image
+    int fakeSize = instances * fakeSpacing; // Size of fake image
+
+    // Generate list of fake sources (instances of the PSF)
+    int numFakes = PS_SQR(instances);   // Number of fake sources
+    psArray *fakes = psArrayAlloc(numFakes); // Fake sources
+    psVector *xOffset = psVectorAlloc(numFakes, PS_TYPE_S32); // X offset from fake position to image
+    psVector *yOffset = psVectorAlloc(numFakes, PS_TYPE_S32); // Y offset from fake position to image
+    for (int j = 0, index = 0; j < instances; j++) {
+        float yOrig = j * yOrigSpacing + radius; // Source position in original image
+        float yFake = j * fakeSpacing + radius; // Position in fake image
+        int dy = yFake - yOrig;         // Difference between fake and original position
+
+        for (int i = 0; i < instances; i++, index++) {
+            float xOrig = i * xOrigSpacing + radius; // Source position in original image
+            float xFake = i * fakeSpacing + radius; // Position in fake image
+            int dx = xFake - xOrig;     // Difference between fake and original position
+
+            pmSource *fake = pmSourceAlloc(); // Fake source
+            fake->peak = pmPeakAlloc(xFake - dx, yFake - dy, PEAK_FLUX, PM_PEAK_LONE);
+            fake->type = PM_SOURCE_TYPE_STAR;
+            fake->psfMag = -2.5 * log10(PEAK_FLUX);
+
+            psTrace("psModules.imcombine", 5, "Source %d: %.2f,%.2f\n",
+                    index, xOrig, yOrig);
+
+            fakes->data[index] = fake;
+            xOffset->data.S32[index] = dx;
+            yOffset->data.S32[index] = dy;
+        }
+    }
+
+    // Generate fake images with each PSF, and take the envelope
+    psImage *envelope = psImageAlloc(fakeSize, fakeSize, PS_TYPE_F32); // Image with envelope of PSFs
+    psImageInit(envelope, SKY_VALUE);
+    pmReadout *fakeRO = pmReadoutAlloc(NULL); // Fake readout
+    float maxRadius = 0.0;              // Maximum radius of sources
+    for (int i = 0; i < inputs->n; i++) {
+        pmPSF *psf = inputs->data[i];   // PSF of interest
+        pmResiduals *resid = psf->residuals;// PSF residuals
+        psf->residuals = NULL;
+        if (!pmReadoutFakeFromSources(fakeRO, fakeSize, fakeSize, fakes, xOffset, yOffset, psf,
+                                      NAN, radius, true)) {
+            psError(PS_ERR_UNKNOWN, false, "Unable to generate fake readout.");
+            psFree(envelope);
+            psFree(yOffset);
+            psFree(xOffset);
+            psFree(fakes);
+            psf->residuals = resid;
+            return NULL;
+        }
+        psf->residuals = resid;
+
+        // Need to renormalise sources so they all have the same peak.  You would think they do have the same
+        // peak already, but it seems that the residual map messes things up by adding extra flux
+        for (int j = 0; j < numFakes; j++) {
+            pmSource *source = fakes->data[j]; // Fake source
+            float x = source->peak->xf + xOffset->data.S32[j]; // x coordinate of source
+            float y = source->peak->yf + yOffset->data.S32[j]; // y coordinate of source
+
+            double flux = fakeRO->image->data.F32[(int)y][(int)x];
+            float norm = PEAK_FLUX / flux; // Normalisation for source
+            psRegion region = psRegionSet(x - radius, x + radius, y - radius, y + radius); // PSF region
+            psImage *subImage = psImageSubset(fakeRO->image, region); // Subimage of fake PSF
+            psImage *subEnv = psImageSubset(envelope, region); // Subimage of envelope
+            psBinaryOp(subImage, subImage, "*", psScalarAlloc(norm, PS_TYPE_F32));
+            psBinaryOp(subEnv, subEnv, "MAX", subImage);
+            psFree(subImage);
+            psFree(subEnv);
+
+            // Get the radius
+            pmModel *model = pmModelFromPSFforXY(psf, x, y, PEAK_FLUX); // Model for source
+            float srcRadius = model->modelRadius(model->params, PS_SQR(WEIGHT_VAL)); // Radius for source
+            if (srcRadius > maxRadius) {
+                maxRadius = srcRadius;
+            }
+            psFree(model);
+        }
+
+#ifdef TESTING
+        {
+            // Write out the PSF field
+            psString name = NULL;
+            psStringAppend(&name, "psf_field_%03d.fits", i);
+            psFits *fits = psFitsOpen(name, "w");
+            psFitsWriteImage(fits, NULL, fakeRO->image, 0, NULL);
+            psFitsClose(fits);
+            psFree(name);
+        }
+#endif
+
+    }
+    psFree(fakeRO);
+
+    if (maxRadius > radius) {
+        maxRadius = radius;
+    }
+
+#ifdef TESTING
+    {
+        // Write out the envelope
+        psFits *fits = psFitsOpen("psf_field_envelope.fits", "w");
+        psFitsWriteImage(fits, NULL, envelope, 0, NULL);
+        psFitsClose(fits);
+    }
+#endif
+
+    // Put the fake sources onto a full-size image
+    pmReadout *readout = pmReadoutAlloc(NULL); // Readout to contain envelope pixels
+    readout->image = psImageAlloc(numCols, numRows, PS_TYPE_F32);
+    psImageInit(readout->image, 0.0);
+    for (int i = 0; i < numFakes; i++) {
+        pmSource *source = fakes->data[i]; // Fake source
+        // Position of source on fake image
+        int xFake = source->peak->x + xOffset->data.S32[i];
+        int yFake = source->peak->y + yOffset->data.S32[i];
+        psRegion region = psRegionSet(xFake - radius, xFake + radius,
+                                      yFake - radius, yFake + radius); // PSF region
+        psImage *subImage = psImageSubset(envelope, region); // Subimage of fake PSF
+
+        // Position of source on "real" image
+        int x0 = source->peak->x - radius;
+        int y0 = source->peak->y - radius;
+
+        if (!psImageOverlaySection(readout->image, subImage, x0, y0, "=")) {
+            psError(PS_ERR_UNKNOWN, false, "Unable to overlay PSF");
+            psFree(subImage);
+            psFree(readout);
+            psFree(xOffset);
+            psFree(yOffset);
+            psFree(fakes);
+            return NULL;
+        }
+        psFree(subImage);
+    }
+    psFree(xOffset);
+    psFree(yOffset);
+    psFree(envelope);
+
+    // XXX Setting the weight seems to be an art
+    // Can't set it too high so that pixels are rejected as insignificant
+    // Can't set it too low so that it's hard to get to the minimum
+    // Have also tried:
+    // *** readout->weight = (psImage*)psBinaryOp(NULL, readout->image, "*", readout->image);
+    // *** readout->weight = (psImage*)psBinaryOp(NULL, readout->image, "*", psScalarAlloc(WEIGHT_FACTOR, PS_TYPE_F32));
+    readout->weight = (psImage*)psBinaryOp(NULL, readout->image, "+", psScalarAlloc(WEIGHT_VAL, PS_TYPE_F32));
+    readout->mask = psImageAlloc(numCols, numRows, PS_TYPE_MASK);
+    psImageInit(readout->mask, 0);
+
+#ifdef TESTING
+    {
+        // Write out the envelope
+        psFits *fits = psFitsOpen("psf_field_full.fits", "w");
+        psFitsWriteImage(fits, NULL, readout->image, 0, NULL);
+        psFitsClose(fits);
+    }
+#endif
+
+    // Reset the sources to point to the new pixels, and measure the moments in preparation for PSF fitting
+    for (int i = 0; i < numFakes; i++) {
+        pmSource *source = fakes->data[i]; // Fake source
+        float x = source->peak->xf;     // x coordinates of source
+        float y = source->peak->yf;     // y coordinates of source
+
+        psFree(source->pixels);
+        psFree(source->weight);
+        psFree(source->maskView);
+        psFree(source->maskObj);
+        source->pixels = NULL;
+        source->weight = NULL;
+        source->maskView = NULL;
+        source->maskObj = NULL;
+
+        if (!pmSourceDefinePixels(source, readout, x, y, maxRadius)) {
+            psError(PS_ERR_UNKNOWN, false, "Unable to define pixels for source.");
+            psFree(readout);
+            psFree(fakes);
+            return NULL;
+        }
+
+        if (!pmSourceMoments(source, maxRadius)) {
+            psError(PS_ERR_UNKNOWN, false, "Unable to measure moments for source.");
+            psFree(readout);
+            psFree(fakes);
+            return NULL;
+        }
+    }
+
+    // Don't assume Poisson errors
+    pmPSFOptions *options = pmPSFOptionsAlloc(); // Options for fitting a PSF
+    options->poissonErrorsPhotLMM = true;
+    options->poissonErrorsPhotLin = false;
+    options->poissonErrorsParams = true;
+    options->stats = psStatsAlloc(PSF_STATS);
+    options->radius = maxRadius;
+    options->psfTrendMode = PM_TREND_MAP;
+    options->psfTrendNx = xOrder;
+    options->psfTrendNy = yOrder;
+    options->psfFieldNx = numCols;
+    options->psfFieldNy = numRows;
+    options->psfFieldXo = 0;
+    options->psfFieldYo = 0;
+
+    pmSourceFitModelInit(SOURCE_FIT_ITERATIONS, 0.01, WEIGHT_VAL, true);
+
+    pmPSFtry *try = pmPSFtryModel(fakes, modelName, options, 0, 0xff);
+    psFree(options);
+    if (!try) {
+        psError(PS_ERR_UNKNOWN, false, "Unable to fit PSF model to PSF envelope.");
+        psFree(readout);
+        psFree(fakes);
+        return NULL;
+    }
+
+    pmPSF *psf = psMemIncrRefCounter(try->psf); // Output PSF
+    psFree(try);
+
+#ifdef TESTING
+    {
+        // Need to translate peak flux --> integrated flux
+        pmModel *fakeModel = pmModelFromPSFforXY(psf, (float)numCols / 2.0, (float)numRows / 2.0,
+                                                 1.0); // Fake model, with central intensity of 1.0
+        float flux0 = fakeModel->modelFlux(fakeModel->params); // Flux for central intensity of 1.0
+        for (int i = 0; i < numFakes; i++) {
+            pmSource *source = fakes->data[i]; // Fake source
+            source->psfMag -= 2.5 * log10(flux0);
+        }
+
+        pmReadout *generated = pmReadoutAlloc(NULL); // Generated image
+        pmReadoutFakeFromSources(generated, numCols, numRows, fakes, NULL, NULL, psf, NAN, radius,
+                                 false);
+        {
+            psFits *fits = psFitsOpen("psf_field_model.fits", "w");
+            psFitsWriteImage(fits, NULL, generated->image, 0, NULL);
+            psFitsClose(fits);
+        }
+        psBinaryOp(generated->image, generated->image, "-", readout->image);
+        {
+            psFits *fits = psFitsOpen("psf_field_resid.fits", "w");
+            psFitsWriteImage(fits, NULL, generated->image, 0, NULL);
+            psFitsClose(fits);
+        }
+        psFree(generated);
+    }
+#endif
+
+    psFree(fakes);
+    psFree(readout);
+
+    return psf;
+}
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmPSFEnvelope.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmPSFEnvelope.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmPSFEnvelope.h	(revision 19481)
@@ -0,0 +1,22 @@
+#ifndef PM_PSF_ENVELOPE_H
+#define PM_PSF_ENVELOPE_H
+
+#include <pslib.h>
+#include <pmMoments.h>
+#include <pmResiduals.h>
+#include <pmGrowthCurve.h>
+#include <pmTrend2D.h>
+#include <pmPSF.h>
+
+/// Generate a PSF which is the envelope of an array of PSFs
+///
+/// Generates multiple instances of the PSFs (distributed over an image)
+pmPSF *pmPSFEnvelope(int numCols, int numRows, // Size of original image
+                     const psArray *inputs, // Input PSF models
+                     int instances,     // Number of instances per dimension
+                     int radius,        // Radius of each PSF
+                     const char *modelName, // Name of PSF model to use
+                     int xOrder, int yOrder // Order for PSF variation
+    );
+
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmReadoutCombine.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmReadoutCombine.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmReadoutCombine.c	(revision 19481)
@@ -0,0 +1,422 @@
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <stdio.h>
+#include <string.h>
+#include <assert.h>
+#include <pslib.h>
+
+#include "pmHDU.h"
+#include "pmFPA.h"
+#include "pmHDUUtils.h"
+#include "pmFPAMaskWeight.h"
+#include "pmConceptsAverage.h"
+#include "pmReadoutStack.h"
+
+#include "pmReadoutCombine.h"
+
+//#define SHOW_BUSY 1                   // Show that the function is busy
+
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Public functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+// Allocator for pmCombineParams
+pmCombineParams *pmCombineParamsAlloc(psStatsOptions combine)
+{
+    pmCombineParams *params = psAlloc(sizeof(pmCombineParams));
+
+    params->combine = combine;
+    params->maskVal = 0;
+    params->blank = 0;
+    params->nKeep = 0;
+    params->fracHigh = 0.0;
+    params->fracHigh = 0.0;
+    params->iter = 1;
+    params->rej = INFINITY;
+    params->weights = false;
+
+    return params;
+}
+
+// check the input parameters and set up the output images
+bool pmReadoutCombinePrepare(pmReadout *output, const psArray *inputs, const pmCombineParams *params)
+{
+    // Check inputs
+    PS_ASSERT_PTR_NON_NULL(output, false);
+    PS_ASSERT_ARRAY_NON_NULL(inputs, false);
+    PS_ASSERT_PTR_NON_NULL(params, false);
+    PS_ASSERT_FLOAT_WITHIN_RANGE(params->fracLow, 0.0, 1.0, false);
+    PS_ASSERT_FLOAT_WITHIN_RANGE(params->fracHigh, 0.0, 1.0, false);
+
+    // valid combintion statistic?
+    bool valid = false;
+    valid |= (params->combine == PS_STAT_SAMPLE_MEAN);
+    valid |= (params->combine == PS_STAT_SAMPLE_MEDIAN);
+    valid |= (params->combine == PS_STAT_ROBUST_MEDIAN);
+    valid |= (params->combine == PS_STAT_FITTED_MEAN);
+    valid |= (params->combine == PS_STAT_CLIPPED_MEAN);
+    if (!valid) {
+        psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Combination method is not SAMPLE_MEAN, SAMPLE_MEDIAN, "
+                "ROBUST_MEDIAN, FITTED_MEAN or CLIPPED_MEAN.\n");
+        return false;
+    }
+
+    pmHDU *hdu = pmHDUFromReadout(output); // Output HDU
+    if (!hdu) {
+        psError(PS_ERR_UNEXPECTED_NULL, false, "Unable to find HDU for readout.\n");
+        return false;
+    }
+
+    //  set the output header metadata
+    psString comment = NULL;        // Comment to add to header
+    psStringAppend(&comment, "Combining using statistic: %x", params->combine);
+    if (!hdu->header) {
+	hdu->header = psMetadataAlloc();
+    }
+    psMetadataAddStr(hdu->header, PS_LIST_TAIL, "HISTORY", PS_META_DUPLICATE_OK, comment, "");
+    psFree(comment);
+
+    // note the clipping parameters, if used
+    if (params->combine == PS_STAT_CLIPPED_MEAN) {
+	psString comment = NULL;    // Comment to add to header
+	psStringAppend(&comment, "Combination clipping: %d iterations, rejection at %f sigma", params->iter, params->rej);
+	psMetadataAddStr(hdu->header, PS_LIST_TAIL, "HISTORY", PS_META_DUPLICATE_OK, comment, "");
+	psFree(comment);
+    }
+
+    // note the use of weights
+    if (params->weights) {
+        psMetadataAddStr(hdu->header, PS_LIST_TAIL, "HISTORY", PS_META_DUPLICATE_OK,
+                         "Using input weights to combine images", "");
+    }
+
+    // note the rejection fraction
+    float keepFrac = 1.0 - params->fracLow - params->fracHigh; // Fraction of pixels to keep
+    if (keepFrac != 1.0) {
+        psString comment = NULL;        // Comment to add to header
+        psStringAppend(&comment, "Min/max rejection: %f high, %f low, keep %d",
+                       params->fracHigh, params->fracLow, params->nKeep);
+        psMetadataAddStr(hdu->header, PS_LIST_TAIL, "HISTORY", PS_META_DUPLICATE_OK, comment, "");
+        psFree(comment);
+    }
+
+    // note the mask value actually used
+    psMaskType maskVal = params->maskVal; // The mask value
+    if (maskVal) {
+        psString comment = NULL;        // Comment to add to header
+        psStringAppend(&comment, "Mask for combination: %x", maskVal);
+        psMetadataAddStr(hdu->header, PS_LIST_TAIL, "HISTORY", PS_META_DUPLICATE_OK, comment, "");
+        psFree(comment);
+    }
+
+    // determine the output image size based on the input images
+    int row0, col0, numCols, numRows;
+    if (!pmReadoutStackSetOutputSize(&col0, &row0, &numCols, &numRows, inputs)) {
+        psError(PS_ERR_UNKNOWN, false, "problem setting output readout size.");
+        return false;
+    }
+
+    // generate the required output images based on the specified sizes
+    pmReadoutStackDefineOutput(output, col0, row0, numCols, numRows, true, params->weights, params->blank);
+    psTrace("psModules.imcombine", 7, "Output minimum: %d,%d\n", output->col0, output->row0);
+
+    // these calls allocate and save the requested images on the output analysis metadata
+    psImage *counts = pmReadoutSetAnalysisImage(output, PM_READOUT_STACK_ANALYSIS_COUNT, numCols, numRows, PS_TYPE_U16, 0);
+    if (!counts) {
+        return false;
+    }
+    psImage *sigma = pmReadoutSetAnalysisImage(output, PM_READOUT_STACK_ANALYSIS_SIGMA, numCols, numRows, PS_TYPE_F32, NAN);
+    if (!sigma) {
+        return false;
+    }
+
+    // Update the "concepts"
+    psList *inputCells = psListAlloc(NULL); // List of cells
+    for (long i = 0; i < inputs->n; i++) {
+        pmReadout *readout = inputs->data[i]; // Readout of interest
+        psListAdd(inputCells, PS_LIST_TAIL, readout->parent);
+    }
+    bool success = pmConceptsAverageCells(output->parent, inputCells, NULL, NULL, true);
+    psFree(inputCells);
+
+    // set these even though the values are not yet set
+    output->data_exists = true;
+    output->parent->data_exists = true;
+    output->parent->parent->data_exists = true;
+
+    return success;
+}
+
+// XXX: Maybe add support for S16 and S32 types.  Currently, only F32 supported.
+bool pmReadoutCombine(pmReadout *output, const psArray *inputs, const psVector *zero, const psVector *scale,
+                      const pmCombineParams *params)
+{
+    // Check inputs
+    PS_ASSERT_PTR_NON_NULL(output, false);
+    PS_ASSERT_ARRAY_NON_NULL(inputs, false);
+    PS_ASSERT_PTR_NON_NULL(params, false);
+    if (zero) {
+        PS_ASSERT_VECTOR_TYPE(zero, PS_TYPE_F32, false);
+        PS_ASSERT_VECTOR_SIZE(zero, inputs->n, false);
+    }
+    if (scale) {
+        PS_ASSERT_VECTOR_TYPE(scale, PS_TYPE_F32, false);
+        PS_ASSERT_VECTOR_SIZE(scale, inputs->n, false);
+    }
+    PS_ASSERT_FLOAT_WITHIN_RANGE(params->fracLow, 0.0, 1.0, false);
+    PS_ASSERT_FLOAT_WITHIN_RANGE(params->fracHigh, 0.0, 1.0, false);
+
+    // does required/desired data exist?
+    for (int i = 0; i < inputs->n; i++) {
+        pmReadout *readout = inputs->data[i]; // Readout of interest
+        if (!readout->image) {
+            psError(PS_ERR_UNEXPECTED_NULL, true, "Image data is missing for image %d.\n", i);
+            return false;
+        }
+        if (params->weights && !readout->weight) {
+            psError(PS_ERR_UNEXPECTED_NULL, true,
+                    "Rejection based on weights requested, but no weights supplied for image %d.\n", i);
+            return false;
+        }
+    }
+
+    pmHDU *hdu = pmHDUFromReadout(output); // Output HDU
+    if (!hdu) {
+        psError(PS_ERR_UNEXPECTED_NULL, false, "Unable to find HDU for readout.\n");
+        return false;
+    }
+
+    // pthread_t id = pthread_self();
+    // char name[64];
+    // sprintf (name, "%x", (unsigned int) id);
+    // psTimerStart (name);
+
+    psStatsOptions combineStdev = 0; // Statistics option for weights
+    switch (params->combine) {
+      case PS_STAT_SAMPLE_MEAN:
+      case PS_STAT_SAMPLE_MEDIAN:
+        combineStdev = PS_STAT_SAMPLE_STDEV;
+        break;
+      case PS_STAT_ROBUST_MEDIAN:
+        combineStdev = PS_STAT_ROBUST_STDEV;
+        break;
+      case PS_STAT_FITTED_MEAN:
+        combineStdev = PS_STAT_FITTED_STDEV;
+        break;
+      case PS_STAT_CLIPPED_MEAN:
+        combineStdev = PS_STAT_CLIPPED_STDEV;
+        break;
+      default:
+        psAbort("Should never get here --- checked params->combine before.\n");
+    }
+
+    psStats *stats = psStatsAlloc(params->combine | combineStdev); // The statistics to use in the combination
+    if (params->combine == PS_STAT_CLIPPED_MEAN) {
+        stats->clipSigma = params->rej;
+        stats->clipIter = params->iter;
+    }
+
+    psImage *counts = pmReadoutGetAnalysisImage(output, PM_READOUT_STACK_ANALYSIS_COUNT);
+    if (!counts) {
+        return false;
+    }
+    psImage *sigma = pmReadoutGetAnalysisImage(output, PM_READOUT_STACK_ANALYSIS_SIGMA);
+    if (!sigma) {
+        return false;
+    }
+
+    stats->options |= combineStdev;
+
+    int minInputCols, maxInputCols, minInputRows, maxInputRows; // Smallest and largest values to combine
+    int xSize, ySize;                   // Size of the output image
+    if (!pmReadoutStackValidate(&minInputCols, &maxInputCols, &minInputRows, &maxInputRows, &xSize, &ySize, inputs)) {
+        psError(PS_ERR_UNKNOWN, false, "No valid input readouts.");
+        return false;
+    }
+
+    // We loop through each pixel in the output image.  We loop through each input readout.  We determine if
+    // that output pixel is contained in the image from that readout.  If so, we save it in psVector pixels.
+    // If not, we set a mask for that element in pixels.  Then, we mask off pixels not between fracLow and
+    // fracHigh.  Then we call the vector stats routine on those pixels/mask.  Then we set the output pixel
+    // value to the result of the stats call.
+
+    psVector *pixels = psVectorAlloc(inputs->n, PS_TYPE_F32); // Stack of pixels
+    psF32 *pixelsData = pixels->data.F32; // Dereference pixels
+
+    psVector *mask   = psVectorAlloc(inputs->n, PS_TYPE_U8); // Mask for stack
+    psU8 *maskData = mask->data.U8;     // Dereference mask
+
+    psVector *weights = NULL;           // Stack of weights
+    psVector *errors = NULL;            // Stack of errors (sqrt of variance/weights), for psVectorStats
+    psF32 *weightsData = NULL;          // Dereference weights
+    if (params->weights) {
+        weights = psVectorAlloc(inputs->n, PS_TYPE_F32); // Stack of weights
+        weightsData = weights->data.F32;
+    }
+    psVector *index = NULL;             // The indices to sort the pixels
+
+    float keepFrac = 1.0 - params->fracLow - params->fracHigh; // Fraction of pixels to keep
+    psMaskType maskVal = params->maskVal; // The mask value
+
+    #ifndef PS_NO_TRACE
+    psTrace("psModules.imcombine", 3, "Iterating output: %d --> %d, %d --> %d\n",
+            minInputCols - output->col0, maxInputCols - output->col0,
+            minInputRows - output->row0, maxInputRows - output->row0);
+    if (psTraceGetLevel("psModules.imcombine") >= 3) {
+        for (int r = 0; r < inputs->n; r++) {
+            pmReadout *readout = inputs->data[r]; // Input readout
+            psTrace("psModules.imcombine", 3, "Iterating input %d: %d --> %d, %d --> %d\n", r,
+                    minInputCols - readout->col0, maxInputCols - readout->col0,
+                    minInputRows - readout->row0, maxInputRows - readout->row0);
+        }
+    }
+    #endif
+
+    // Dereference output products
+    psF32 **outputImage  = output->image->data.F32; // Output image
+    psU8  **outputMask   = output->mask->data.U8; // Output mask
+    psF32 **outputWeight = NULL; // Output weight map
+    if (output->weight) {
+        outputWeight = output->weight->data.F32;
+    }
+
+    psVector *invScale = NULL;          // Inverse scale; pre-calculated for efficiency
+    if (scale) {
+        invScale = (psVector*)psBinaryOp(NULL, psScalarAlloc(1.0, PS_TYPE_F32), "/", (const psPtr)scale);
+    }
+
+    for (int i = minInputRows; i < maxInputRows; i++) {
+        int yOut = i - output->row0; // y position on output readout
+
+        #ifdef SHOW_BUSY
+        if (psTraceGetLevel("psModules.imcombine") > 9) {
+            printf("Processing row %d\r", i);
+            fflush(stdout);
+        }
+        #endif
+
+        for (int j = minInputCols; j < maxInputCols; j++) {
+            int xOut = j - output->col0; // x position on output readout
+
+            int numValid = 0;           // Number of valid pixels in the stack
+            memset(maskData, 0, mask->n * sizeof(psU8)); // Reset the mask
+            for (int r = 0; r < inputs->n; r++) {
+                pmReadout *readout = inputs->data[r]; // Input readout
+                int yIn = i - readout->row0; // y position on input readout
+                int xIn = j - readout->col0; // x position on input readout
+                psImage *image = readout->image; // The readout image
+
+                pixelsData[r] = image->data.F32[yIn][xIn];
+                if (!isfinite(pixelsData[r])) {
+                    maskData[r] = 1;
+                    continue;
+                }
+
+                // Check mask
+                psImage *roMask = readout->mask; // The mask image
+                if (roMask && roMask->data.U8[yIn][xIn] & maskVal) {
+                    maskData[r] = 1;
+                    continue;
+                }
+
+                if (params->weights) {
+                    weightsData[r] = readout->weight->data.F32[yIn][xIn];
+                }
+
+                if (zero) {
+                    pixelsData[r] -= zero->data.F32[r];
+                }
+                if (scale) {
+                    pixelsData[r] *= invScale->data.F32[r];
+                    if (params->weights) {
+                        weightsData[r] *= invScale->data.F32[r] * invScale->data.F32[r];
+                    }
+                }
+
+                numValid++;
+            }
+
+            if (numValid == 0) {
+                outputMask[yOut][xOut] = params->blank;
+                outputImage[yOut][xOut] = NAN;
+                counts->data.U16[yOut][xOut] = 0;
+                sigma->data.F32[yOut][xOut] = NAN;
+                continue;
+            }
+
+            // Apply fracLow,fracHigh if there are enough pixels
+            if (numValid * keepFrac >= params->nKeep && keepFrac != 1.0) {
+                index = psVectorSortIndex(index, pixels);
+                int numLow = numValid * params->fracLow; // Number of low pixels to clip
+                int numHigh = numValid * params->fracHigh; // Number of high pixels to clip
+                // Low pixels
+                psS32 *indexData = index->data.S32; // Dereference index
+                for (int k = 0, numMasked = 0; numMasked < numLow && k < index->n; k++) {
+                    // Don't count the ones that are already masked
+                    if (!maskData[indexData[k]]) {
+                        maskData[indexData[k]] = 1;
+                        numMasked++;
+                        numValid--;
+                    }
+                }
+                // High pixels
+                for (int k = pixels->n - 1, numMasked = 0; numMasked < numHigh && k >= 0; k--) {
+                    // Don't count the ones that are already masked
+                    if (!maskData[indexData[k]]) {
+                        maskData[indexData[k]] = 1;
+                        numMasked++;
+                        numValid--;
+                    }
+                }
+            }
+            counts->data.U16[yOut][xOut] = numValid;
+
+            // XXXXX this step probably is very expensive : convert errors to variance everywhere?
+            if (params->weights) {
+                errors = (psVector*)psUnaryOp(errors, weights, "sqrt");
+            }
+
+            // Combination
+            if (!psVectorStats(stats, pixels, errors, mask, 1)) {
+                // Can't do much about it, but it's not worth worrying about
+                psErrorClear();
+                outputImage[yOut][xOut] = NAN;
+                outputMask[yOut][xOut] = params->blank;
+                if (params->weights) {
+                    outputWeight[yOut][xOut] = NAN;
+                }
+                sigma->data.F32[yOut][xOut] = NAN;
+            } else {
+                outputImage[yOut][xOut] = psStatsGetValue(stats, params->combine);
+                outputMask[yOut][xOut] = isfinite(outputImage[yOut][xOut]) ? 0 : params->blank;
+                if (params->weights) {
+                    float stdev = psStatsGetValue(stats, combineStdev);
+                    outputWeight[yOut][xOut] = PS_SQR(stdev); // Variance
+                    // XXXX this is not the correct formal error.
+                    // also, the weighted mean is not obviously the correct thing here
+                }
+                sigma->data.F32[yOut][xOut] = psStatsGetValue(stats, combineStdev);
+            }
+        }
+    }
+    #ifdef SHOW_BUSY
+    if (psTraceGetLevel("psModules.imcombine") > 9) {
+        printf("\n");
+    }
+    #endif
+
+    psFree(index);
+    psFree(pixels);
+    psFree(mask);
+    psFree(weights);
+    psFree(errors);
+    psFree(stats);
+    psFree(invScale);
+
+    // fprintf (stderr, "done with combine %x : %f sec\n", (unsigned int) id, psTimerMark (name));
+    return true;
+}
+
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmReadoutCombine.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmReadoutCombine.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmReadoutCombine.h	(revision 19481)
@@ -0,0 +1,50 @@
+/* @file  pmReadoutCombine.h
+ * @brief Combine multiple readouts
+ *
+ * @author George Gusciora, MHPCC
+ * @author Paul Price, IfA
+ *
+ * @version $Revision: 1.14 $ $Name: not supported by cvs2svn $
+ * @date $Date: 2008-08-01 00:01:26 $
+ * Copyright 2004-2006 Institute for Astronomy, University of Hawaii
+ */
+
+#ifndef PM_READOUT_COMBINE_H
+#define PM_READOUT_COMBINE_H
+
+/// @addtogroup imcombine Image Combinations
+/// @{
+
+/// Combination parameters for pmReadoutCombine.
+///
+/// These values define how the combination is performed, and should not vary by detector, so that it can be
+/// re-used for multiple combinations.
+typedef struct {
+    psStatsOptions combine;             ///< Statistic to use when performing the combination
+    psMaskType maskVal;                 ///< Mask value
+    psMaskType blank;                   ///< Mask value to give blank (i.e., no data) pixels
+    int nKeep;                          ///< Mimimum number of pixels to keep
+    float fracHigh;                     ///< Fraction of high pixels to immediately throw
+    float fracLow;                      ///< Fraction of low pixels to immediately throw
+    int iter;                           ///< Number of iterations for clipping (for CLIPPED_MEAN only)
+    float rej;                          ///< Rejection threshould for clipping (for CLIPPED_MEAN only)
+    bool weights;                       ///< Use the supplied weights (instead of calculated stdev)?
+} pmCombineParams;
+
+// Allocator for pmCombineParams
+pmCombineParams *pmCombineParamsAlloc(psStatsOptions statsOptions ///< Statistic to use for combination
+                                     );
+
+// check the input parameters and set up the output images
+bool pmReadoutCombinePrepare(pmReadout *output, const psArray *inputs, const pmCombineParams *params);
+
+/// Combine multiple readouts, applying zero and scale, with optional minmax clipping
+bool pmReadoutCombine(pmReadout *output,///< Output readout; altered and returned
+                      const psArray *inputs,  ///< Array of input readouts (F32 image and weight, U8 mask)
+                      const psVector *zero, ///< Zero corrections to subtract from input, or NULL
+                      const psVector *scale, ///< Scale corrections to divide into input, or NULL
+                      const pmCombineParams *params ///< Combination parameters
+                     );
+
+/// @}
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmStack.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmStack.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmStack.c	(revision 19481)
@@ -0,0 +1,712 @@
+/** @file  pmStack.c
+ *
+ *  This file will perform image combination of several images of the
+ *  same field, produce a list of questionable pixels, then tag some
+ *  of those pixels as cosmic rays.
+ *
+ *  @author Paul Price, IfA
+ *  @author GLG, MHPCC
+ *
+ *  @version $Revision: 1.38.2.5 $ $Name: not supported by cvs2svn $
+ *  @date $Date: 2008-09-11 03:36:08 $
+ *  Copyright 2004-2007 Institute for Astronomy, University of Hawaii
+ *
+ */
+
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <stdio.h>
+#include <pslib.h>
+
+#include "pmHDU.h"
+#include "pmFPA.h"
+#include "pmReadoutStack.h"
+#include "pmConceptsAverage.h"
+
+#include "pmStack.h"
+
+#define PIXEL_LIST_BUFFER 100           // Number of entries to add to pixel list at a time
+#define PIXEL_MAP_BUFFER 2              // Number of entries to add to pixel map at a time
+
+#define NUM_DIRECT_STDEV 5              // For less than this number of values, measure stdev directly
+
+
+// Data structure for use as a buffer when combining pixels
+// Use of this structure means we don't have to do an allocation in the combination function for each pixel
+typedef struct {
+    psVector *pixels;                   // Pixel values
+    psVector *masks;                    // Pixel masks
+    psVector *variances;                // Pixel variances
+    psVector *weights;                  // Pixel weightings
+    psVector *sources;                  // Pixel sources (which image did they come from?)
+    psVector *sort;                     // Buffer for sorting (to get a robust estimator of the standard dev)
+} combineBuffer;
+
+static void combineBufferFree(combineBuffer *buffer)
+{
+    psFree(buffer->pixels);
+    psFree(buffer->masks);
+    psFree(buffer->variances);
+    psFree(buffer->weights);
+    psFree(buffer->sources);
+    psFree(buffer->sort);
+    return;
+}
+
+static combineBuffer *combineBufferAlloc(long numImages // Number of images that will be combined
+    )
+{
+    combineBuffer *buffer = psAlloc(sizeof(combineBuffer));
+    psMemSetDeallocator(buffer, (psFreeFunc)combineBufferFree);
+
+    buffer->pixels = psVectorAlloc(numImages, PS_TYPE_F32);
+    buffer->masks = psVectorAlloc(numImages, PS_TYPE_MASK);
+    buffer->variances = psVectorAlloc(numImages, PS_TYPE_F32);
+    buffer->weights = psVectorAlloc(numImages, PS_TYPE_F32);
+    buffer->sources = psVectorAlloc(numImages, PS_TYPE_U16);
+    buffer->sort = psVectorAlloc(numImages, PS_TYPE_F32);
+    return buffer;
+}
+
+
+// Deallocator for the stack data
+static void stackDataFree(pmStackData *data)
+{
+    psFree(data->readout);
+    psFree(data->reject);
+    psFree(data->inspect);
+    return;
+}
+
+
+// Determine a mean value and variance for the combination
+// Not using psVectorStats because it assumes that the "weights" are errors, and weights by 1/error^2
+static bool combinationMeanVariance(float *mean, // Mean value, to return
+                                    float *var, // Variance value, to return
+                                    const psVector *values, // Values to combine
+                                    const psVector *variances, // Pixel variances to combine
+                                    const psVector *weights // Weights to apply
+                                    )
+{
+    assert(mean);
+    assert(values && weights);
+    assert(values->n == weights->n);
+    assert((var && variances) || !var);
+    assert(!variances || variances->n == values->n);
+    assert(values->type.type == PS_TYPE_F32);
+    assert(!values || values->type.type == PS_TYPE_F32);
+    assert(weights->type.type == PS_TYPE_F32);
+
+    // We're not using the input pixel variances to generate a weighted average for the pixel flux (because
+    // that introduces systematic biases), so the variance of the output pixel value should simply be:
+    //     simga^2 = sum(weight_i^2 * sigma_i^2) / (sum(weight_i))^2
+    // This reduces, when the weights are all identically unity, to:
+    //     variance_combination = sum(variance_i) / N^2
+    // and if the variances are all equal:
+    //     variance_combination = variance_individual / N
+    // which makes sense --- the standard deviation of the combination is reduced by a factor of sqrt(N).
+
+    float sumValueWeight = 0.0;         // Sum of the value multiplied by the weight
+    float sumVarianceWeight = 0.0;     // Sum of the pixel variances multiplied by the global weights
+    float sumWeight = 0.0;              // Sum of the image weights
+    for (int i = 0; i < values->n; i++) {
+        sumValueWeight += values->data.F32[i] * weights->data.F32[i];
+        sumWeight += weights->data.F32[i];
+        if (variances) {
+            sumVarianceWeight += variances->data.F32[i] * PS_SQR(weights->data.F32[i]);
+        }
+    }
+
+    if (sumWeight <= 0) {
+        return false;
+    }
+
+    *mean = sumValueWeight / sumWeight;
+    if (var) {
+        *var = sumVarianceWeight / PS_SQR(sumWeight);
+    }
+    return true;
+}
+
+// Return the median and standard deviation for the pixels
+// Not using psVectorStats because it has additional allocations which slow things down
+static bool combinationMedianStdev(float *median, // Median value, to return
+                                   float *stdev, // Standard deviation value, to return
+                                   const psVector *values, // Values to combine
+                                   const psVector *masks, // Mask to apply
+                                   psVector *sortBuffer // Buffer for sorting
+                                   )
+{
+    assert(values);
+    assert(!masks || values->n == masks->n);
+    assert(values->type.type == PS_TYPE_F32);
+    assert(!masks || masks->type.type == PS_TYPE_MASK);
+    assert(sortBuffer && sortBuffer->nalloc >= values->n && sortBuffer->type.type == PS_TYPE_F32);
+
+    // Need to filter out clipped values
+    int num = 0;            // Number of valid values
+    for (int i = 0; i < values->n; i++) {
+        if (!masks || !masks->data.PS_TYPE_MASK_DATA[i]) {
+            sortBuffer->data.F32[num++] = values->data.F32[i];
+        }
+    }
+    sortBuffer->n = num;
+    if (!psVectorSortInPlace(sortBuffer)) {
+        return false;
+    }
+
+    if (num == 3) {
+        // Attempt to measure standard deviation with only three values (and one of those possibly corrupted)
+        *median = sortBuffer->data.F32[1];
+        if (stdev) {
+            float diff1 = sortBuffer->data.F32[0] - *median;
+            float diff2 = sortBuffer->data.F32[2] - *median;
+            // This factor of sqrt(2) might not be exact, but it's about right
+            *stdev = M_SQRT2 * PS_MIN(fabsf(diff1), fabsf(diff2));
+        }
+    } else {
+        *median = num % 2 ? sortBuffer->data.F32[num / 2] :
+            (sortBuffer->data.F32[num / 2 - 1] + sortBuffer->data.F32[num / 2]) / 2.0;
+        if (stdev) {
+            if (num <= NUM_DIRECT_STDEV) {
+                // If there are not many values, the direct standard deviation is better
+                double sum = 0.0;
+                for (int i = 0; i < num; i++) {
+                    sum += PS_SQR(sortBuffer->data.F32[i] - *median);
+                }
+                *stdev = sqrt(sum / (double)(num - 1));
+            } else {
+                // Standard deviation from the interquartile range
+                *stdev = 0.74 * (sortBuffer->data.F32[(int)(0.75 * num)] -
+                                 sortBuffer->data.F32[(int)(0.25 * num)]);
+            }
+        }
+    }
+
+    return true;
+}
+
+
+// Mark a pixel for inspection
+static inline void combineInspect(const psArray *inputs, // Stack data
+                                  int x, int y, // Pixel
+                                  int source // Source image index
+                                  )
+{
+    pmStackData *data = inputs->data[source]; // Stack data of interest
+    if (!data) {
+        psWarning("Can't find input data for source %d", source);
+        return;
+    }
+    data->inspect = psPixelsAdd(data->inspect, data->inspect->nalloc, x, y);
+    return;
+}
+
+// Given a stack of images, combine with optional rejection.
+// Pixels in the stack that are rejected are marked for subsequent inspection
+static void combinePixels(psImage *image, // Combined image, for output
+                          psImage *mask, // Combined mask, for output
+                          psImage *variance, // Combined variance map, for output
+                          const psArray *inputs, // Stack data
+                          const psVector *weights, // Global (single value) weights for data, or NULL
+                          const psVector *varFactors, // Variance factors for data
+                          const psVector *reject, // Indices of pixels to reject, or NULL
+                          int x, int y, // Coordinates of interest; frame of output image
+                          psMaskType maskVal, // Value to mask
+                          psMaskType bad, // Value to give bad pixels
+                          int numIter, // Number of rejection iterations
+                          float rej, // Number of standard deviations at which to reject
+                          bool useVariance, // Use variance for rejection when combining?
+                          bool safe,    // Combine safely?
+                          combineBuffer *buffer // Buffer for combination; to avoid multiple allocations
+                         )
+{
+    // Rudimentary error checking
+    assert(image);
+    assert(mask);
+    assert(inputs);
+    assert(numIter >= 0);
+    assert(buffer);
+    assert(varFactors);
+    assert((useVariance && variance) || !useVariance);
+
+    psVector *pixelData = buffer->pixels; // Values for the pixel of interest
+    psVector *pixelMasks = buffer->masks; // Masks for the pixel of interest
+    psVector *pixelVariances = variance ? buffer->variances : NULL; // Variances for the pixel of interest
+    psVector *pixelWeights = buffer->weights; // Image weights for the pixel of interest
+    psVector *pixelSources = buffer->sources; // Sources for the pixel of interest
+    psVector *sort = buffer->sort;      // Sort buffer
+
+    // Extract the pixel and mask data
+    int num = 0;                        // Number of good images
+    for (int i = 0, j = 0; i < inputs->n; i++) {
+        // Check if this pixel has been rejected.  Assumes that the rejection pixel list is sorted --- it
+        // should be because of how pixelMapGenerate works
+        if (reject && reject->data.U16[j] == i) {
+            j++;
+            continue;
+        }
+
+        pmStackData *data = inputs->data[i]; // Stack data of interest
+        if (!data) {
+            continue;
+        }
+
+        int xIn = x - data->readout->col0, yIn = y - data->readout->row0; // Coordinates on input readout
+        psImage *mask = data->readout->mask; // Mask of interest
+        if (mask->data.PS_TYPE_MASK_DATA[yIn][xIn] & maskVal) {
+            continue;
+        }
+
+        psImage *image = data->readout->image; // Image of interest
+        psImage *variance = data->readout->weight; // Variance ("weight") map of interest
+        pixelData->data.F32[num] = image->data.F32[yIn][xIn];
+        if (variance) {
+            pixelVariances->data.F32[num] = variance->data.F32[yIn][xIn];
+        }
+        pixelWeights->data.F32[num] = data->weight;
+        pixelSources->data.U16[num] = i;
+        num++;
+    }
+    pixelData->n = num;
+    if (variance) {
+        pixelVariances->n = num;
+    }
+    pixelWeights->n = num;
+    pixelSources->n = num;
+
+
+    // The sensible thing to do varies according to how many good pixels there are.
+    // Default option is that the pixel is bad
+    float imageValue = NAN, varianceValue = NAN; // Value for combined image and variance map
+    psMaskType maskValue = bad;         // Value for combined mask
+    switch (num) {
+      case 0:
+        // Nothing to combine: it's bad
+        break;
+      case 1: {
+          // Accept the single pixel unless we have to be safe
+          if (!safe) {
+              imageValue = pixelData->data.F32[0];
+              if (variance) {
+                  varianceValue = pixelVariances->data.F32[0];
+              }
+              maskValue = 0;
+          }
+          break;
+      }
+      case 2: {
+          // Accept the mean of the pixels only if we're going to reject based on the variance, or we're not
+          // playing safe
+          if (useVariance || !safe) {
+              float mean, var;   // Mean and variance from combination
+              if (combinationMeanVariance(&mean, &var, pixelData, pixelVariances, pixelWeights)) {
+                  imageValue = mean;
+                  varianceValue = var;
+                  maskValue = 0;
+              }
+          }
+          if (useVariance && numIter > 0) {
+              // Use variance to check that the two are consistent
+              float diff = pixelData->data.F32[0] - pixelData->data.F32[1];
+              float sigma2 = pixelVariances->data.F32[0] * varFactors->data.F32[pixelSources->data.U16[0]] +
+                  pixelVariances->data.F32[1] * varFactors->data.F32[pixelSources->data.U16[1]];
+              if (PS_SQR(diff) > PS_SQR(rej) * sigma2) {
+                  // Not consistent: mark both for inspection
+                  combineInspect(inputs, x, y, pixelSources->data.U16[0]);
+                  combineInspect(inputs, x, y, pixelSources->data.U16[1]);
+              }
+          }
+          break;
+      }
+      default: {
+          // Record the value derived with no clipping, because pixels rejected using the harsh clipping
+          // applied in the first pass might later be accepted.
+          float mean, var;           // Mean and variance of the combination
+          if (!combinationMeanVariance(&mean, &var, pixelData, pixelVariances, pixelWeights)) {
+              break;
+          }
+          imageValue = mean;
+          varianceValue = var;
+          maskValue = 0;
+
+          // Prepare for clipping iteration
+          if (numIter > 0) {
+              pixelMasks->n = num;
+              psVectorInit(pixelMasks, 0);
+              if (useVariance) {
+                  // Convert to rejection limits --- saves doing it later.
+                  // Using squared rejection limit because it's cheaper than sqrts
+                  float rej2 = PS_SQR(rej); // Rejection level squared
+                  for (int i = 0; i < num; i++) {
+                      pixelVariances->data.F32[i] *= rej2 * varFactors->data.F32[pixelSources->data.U16[i]];
+                  }
+              }
+          }
+
+          // The clipping that follows is solely to identify suspect pixels.
+          // These suspect pixels will be inspected in more detail by other functions.
+          int numClipped = INT_MAX;     // Number of pixels clipped per iteration
+          int totalClipped = 0;         // Total number of pixels clipped
+          for (int i = 0; i < numIter && numClipped > 0 && num - totalClipped > 2; i++) {
+              numClipped = 0;
+              float median, stdev;    // Median and stdev of the combination, for rejection
+
+              if (!combinationMedianStdev(&median, useVariance ? NULL : &stdev,
+                                          pixelData, pixelMasks, sort)) {
+                  psWarning("Bad median/stdev at %d,%d", x, y);
+                  break;
+              }
+
+              float limit = rej * stdev; // Rejection limit, when rejecting based on data properties
+
+// Mask a pixel for inspection
+#define MASK_PIXEL_FOR_INSPECTION() \
+    pixelMasks->data.PS_TYPE_MASK_DATA[j] = 0xff; \
+    combineInspect(inputs, x, y, pixelSources->data.U16[j]); \
+    numClipped++; \
+    totalClipped++;
+
+              for (int j = 0; j < num; j++) {
+                  if (pixelMasks->data.PS_TYPE_MASK_DATA[j]) {
+                      continue;
+                  }
+                  float diff = pixelData->data.F32[j] - median; // Difference from expected
+                  if (useVariance) {
+                      // Comparing squares --- cheaper than lots of sqrts
+                      // pixelVariances includes the variance factor and the rejection limit, from above
+                      if (PS_SQR(diff) > pixelVariances->data.F32[j]) {
+                          MASK_PIXEL_FOR_INSPECTION();
+                      }
+                  } else if (fabsf(diff) > limit) {
+                      MASK_PIXEL_FOR_INSPECTION();
+                  }
+              }
+          }
+          break;
+      }
+    }
+
+    image->data.F32[y][x] = imageValue;
+    mask->data.PS_TYPE_MASK_DATA[y][x] = maskValue;
+    if (variance) {
+        variance->data.F32[y][x] = varianceValue;
+    }
+
+    return;
+}
+
+
+// Ensure the input array of pmStackData is valid, and get some details out of it
+static bool validateInputData(bool *haveVariances, // Do we have variance maps in the sky images?
+                              bool *haveRejects, // Do we have lists of rejected pixels?
+                              int *num,    // Number of inputs
+                              int *numCols, int *numRows, // Size of (sky) images
+                              psArray *input // Input array of pmStackData to validate
+    )
+{
+    PS_ASSERT_ARRAY_NON_NULL(input, false);
+    *num = input->n;
+
+    pmStackData *data = NULL;           // First image off the rank, used as a template
+    for (int i = 0; !data && i < input->n; i++) {
+        data = input->data[i];
+    }
+    PS_ASSERT_PTR_NON_NULL(data, false);
+    assert(psMemGetDeallocator(data) == (psFreeFunc)stackDataFree); // Ensure it's the right type
+    *haveVariances = false;
+    PS_ASSERT_IMAGE_NON_NULL(data->readout->image, false);
+    PS_ASSERT_IMAGE_TYPE(data->readout->image, PS_TYPE_F32, false);
+    PS_ASSERT_IMAGE_NON_NULL(data->readout->mask, false);
+    PS_ASSERT_IMAGE_TYPE(data->readout->mask, PS_TYPE_MASK, false);
+    PS_ASSERT_IMAGES_SIZE_EQUAL(data->readout->image, data->readout->mask, false);
+    *numCols = data->readout->image->numCols;
+    *numRows = data->readout->image->numRows;
+    if (data->readout->weight) {
+        *haveVariances = true;
+        PS_ASSERT_IMAGE_NON_NULL(data->readout->weight, false);
+        PS_ASSERT_IMAGES_SIZE_EQUAL(data->readout->image, data->readout->weight, false);
+        PS_ASSERT_IMAGE_TYPE(data->readout->weight, PS_TYPE_F32, false);
+    }
+    *haveRejects = (data->reject != NULL);
+
+    // Make sure the rest correspond with the first
+    for (int i = 1; i < *num; i++) {
+        pmStackData *data = input->data[i]; // Stack data for this input
+        if (!data) {
+            continue;
+        }
+        assert(psMemGetDeallocator(data) == (psFreeFunc)stackDataFree); // Ensure it's the right type
+        if (!data->readout) {
+            psError(PS_ERR_UNEXPECTED_NULL, true, "The readout is specified in some but not all inputs.");
+            return false;
+        }
+        if ((*haveRejects && !data->reject) || (data->reject && !*haveRejects)) {
+            psError(PS_ERR_UNEXPECTED_NULL, true,
+                    "The rejected pixels are specified in some but not all inputs.");
+            return false;
+        }
+        PS_ASSERT_IMAGE_NON_NULL(data->readout->image, false);
+        PS_ASSERT_IMAGE_NON_NULL(data->readout->mask, false);
+        PS_ASSERT_IMAGE_TYPE(data->readout->image, PS_TYPE_F32, false);
+        PS_ASSERT_IMAGE_TYPE(data->readout->mask, PS_TYPE_MASK, false);
+        PS_ASSERT_IMAGE_SIZE(data->readout->image, *numCols, *numRows, false);
+        PS_ASSERT_IMAGES_SIZE_EQUAL(data->readout->image, data->readout->mask, false);
+        if (*haveVariances) {
+            PS_ASSERT_IMAGE_NON_NULL(data->readout->weight, false);
+            PS_ASSERT_IMAGES_SIZE_EQUAL(data->readout->image, data->readout->weight, false);
+            PS_ASSERT_IMAGE_TYPE(data->readout->weight, PS_TYPE_F32, false);
+        }
+    }
+
+    return true;
+}
+
+
+// Generate a "pixel map".
+//
+// A "pixel map" is an image-like structure containing a vector that contains the indices of images.  The idea
+// is to provide a reverse lookup for an array of pixel lists, so that the image for which a pixel is flagged
+// can be identified easily.
+static psArray *pixelMapGenerate(const psArray *input, // Data to stack
+                                 int minCols, int maxCols, int minRows, int maxRows // Bounds of interest
+    )
+{
+    int numCols = maxCols - minCols + 1, numRows = maxRows - minRows + 1; // Size of map
+
+    psArray *map = psArrayAlloc(numRows); // The pixel map
+    for (int y = 0; y < numRows; y++) {
+        map->data[y] = psArrayAlloc(numCols);
+    }
+
+    for (int i = 0; i < input->n; i++) {
+        pmStackData *data = input->data[i];
+        if (!data) {
+            continue;
+        }
+        assert(data->reject);
+        psPixels *pixels = data->reject; // The rejected pixels
+        for (int j = 0; j < pixels->n; j++) {
+            int x = pixels->data[j].x - minCols, y = pixels->data[j].y - minRows; // Coordinates of interest
+            if (x < 0 || x >= numCols || y < 0 || y >= numRows) {
+                continue;
+            }
+            psArray *columns = map->data[y]; // The columns for that row
+            psVector *images = columns->data[x]; // The images for that column
+            if (!images) {
+                images = columns->data[x] = psVectorAllocEmpty(PIXEL_MAP_BUFFER, PS_TYPE_U16);
+            }
+            int size = images->n;       // Element number at which to add
+            columns->data[x] = psVectorExtend(images, PIXEL_MAP_BUFFER, 1);
+            images->data.U16[size] = i;
+        }
+    }
+
+    return map;
+}
+
+// Query a "pixel map", by returning the list of image indices for a particular pixel.
+static psVector *pixelMapQuery(const psArray *map, // Pixel map
+                               int x0, int y0, // Offset into map
+                               int x, int y // Coordinates of interest
+    )
+{
+    // Adjust for offset
+    x -= x0;
+    y -= y0;
+
+    assert(y >= 0 && y < map->n);
+    psArray *colMap = map->data[y];     // Columns for that row
+    assert(x >= 0 && x < colMap->n);
+    return colMap->data[x];
+}
+
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Public functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+/// Constructor
+pmStackData *pmStackDataAlloc(pmReadout *readout, float weight)
+{
+    pmStackData *data = psAlloc(sizeof(pmStackData)); // Stack data, to return
+    psMemSetDeallocator(data, (psFreeFunc)stackDataFree);
+
+    data->readout = psMemIncrRefCounter(readout);
+    data->reject = NULL;
+    data->inspect = NULL;
+    data->weight = weight;
+
+    return data;
+}
+
+/// Stack input images
+bool pmStackCombine(pmReadout *combined, psArray *input, psMaskType maskVal, psMaskType bad,
+                    int kernelSize, int numIter, float rej, bool entire, bool useVariance, bool safe)
+{
+    PS_ASSERT_PTR_NON_NULL(combined, false);
+    bool haveVariances;                 // Do we have the variance maps?
+    bool haveRejects;                   // Do we have lists of rejected pixels?
+    int num;                            // Number of inputs
+    int numCols, numRows;               // Size of (sky) images
+    if (!validateInputData(&haveVariances, &haveRejects, &num, &numCols, &numRows, input)) {
+        return false;
+    }
+    PS_ASSERT_INT_NONNEGATIVE(kernelSize, false);
+    PS_ASSERT_INT_POSITIVE(bad, false);
+    PS_ASSERT_INT_NONNEGATIVE(numIter, false);
+    if (isnan(rej)) {
+        PS_ASSERT_INT_EQUAL(numIter, 0, false);
+    } else {
+        PS_ASSERT_FLOAT_LARGER_THAN(rej, 0.0, false);
+    }
+    if (haveRejects) {
+        // This is a subsequent combination, so expect that the image and mask already exist
+        PS_ASSERT_IMAGE_NON_NULL(combined->image, false);
+        PS_ASSERT_IMAGE_TYPE(combined->image, PS_TYPE_F32, false);
+        PS_ASSERT_IMAGE_NON_NULL(combined->mask, false);
+        PS_ASSERT_IMAGE_TYPE(combined->mask, PS_TYPE_MASK, false);
+        PS_ASSERT_IMAGES_SIZE_EQUAL(combined->image, combined->mask, false);
+    }
+    if (useVariance && !haveVariances) {
+        psWarning("Unable to use variance in rejection if no variance maps supplied --- option turned off");
+        useVariance = false;
+    }
+
+    psVector *varFactors = psVectorAlloc(num, PS_TYPE_F32); // Variance factors for each image
+    psVector *weights = psVectorAlloc(num, PS_TYPE_F32); // Relative weighting for each image
+    psArray *stack = psArrayAlloc(num); // Stack of readouts
+    for (int i = 0; i < num; i++) {
+        pmStackData *data = input->data[i]; // Stack data for this input
+        if (!data) {
+            weights->data.F32[i] = 0.0;
+            continue;
+        }
+        weights->data.F32[i] = data->weight;
+        stack->data[i] = psMemIncrRefCounter(data->readout);
+        // Variance factor
+        float vf = psMetadataLookupF32(NULL, data->readout->parent->concepts, "CELL.VARFACTOR"); // Var factor
+        if (!isfinite(vf)) {
+            psWarning("Non-finite CELL.VARFACTOR for image %d --- setting to unity.", i);
+            vf = 1.0;
+        }
+        varFactors->data.F32[i] = vf;
+        if (!haveRejects && !data->inspect) {
+            data->inspect = psPixelsAllocEmpty(PIXEL_LIST_BUFFER);
+        }
+    }
+
+    int minInputCols, maxInputCols, minInputRows, maxInputRows; // Smallest and largest values to combine
+    int xSize, ySize;                   // Size of the output image
+    if (!pmReadoutStackValidate(&minInputCols, &maxInputCols, &minInputRows, &maxInputRows, &xSize, &ySize,
+                                stack)) {
+        psError(PS_ERR_UNKNOWN, false, "Input stack is not valid.");
+        psFree(stack);
+        return false;
+    }
+    psFree(stack);
+    pmReadoutUpdateSize(combined, minInputCols, minInputRows, xSize, ySize, true, true, bad);
+    psTrace("psModules.imcombine", 1, "Have for combination [%d:%d,%d:%d] (%dx%d)\n",
+            minInputCols, maxInputCols, minInputRows, maxInputRows, xSize, ySize);
+
+    // Reduce combination area by the size of the kernel
+    minInputCols += kernelSize;
+    maxInputCols -= kernelSize;
+    minInputRows += kernelSize;
+    maxInputRows -= kernelSize;
+    psTrace("psModules.imcombine", 1, "Combining on [%d:%d,%d:%d]\n",
+            minInputCols, maxInputCols, minInputRows, maxInputRows);
+
+
+    // Buffer for combination
+    combineBuffer *buffer = combineBufferAlloc(num);
+
+    if (haveRejects) {
+        psImage *combinedImage = combined->image; // Combined image
+        psImage *combinedMask = combined->mask; // Combined mask
+        psImage *combinedVariance = combined->weight; // Combined variance map
+
+        psArray *pixelMap = pixelMapGenerate(input, minInputCols, maxInputCols,
+                                             minInputRows, maxInputRows); // Map of pixels to source
+        psPixels *pixels = NULL;            // Total list of pixels, with no duplicates
+        for (int i = 0; i < num; i++) {
+            pmStackData *data = input->data[i]; // Stacking data; contains the list of pixels
+            if (!data) {
+                continue;
+            }
+            pixels = psPixelsConcatenate(pixels, data->reject);
+        }
+
+        if (entire) {
+            // Combine entire image
+            for (int y = minInputRows; y < maxInputRows; y++) {
+                for (int x = minInputCols; x < maxInputCols; x++) {
+                    psVector *reject = pixelMapQuery(pixelMap, minInputCols, minInputRows,
+                                                     x, y); // Reject these images
+                    combinePixels(combinedImage, combinedMask, combinedVariance, input, weights, varFactors,
+                                  reject, x, y, maskVal, bad, numIter, rej, useVariance, safe, buffer);
+                }
+            }
+        } else {
+            // Only combine previously rejected pixels
+            for (int i = 0; i < pixels->n; i++) {
+                // Pixel coordinates are in the frame of the original image
+                int x = pixels->data[i].x, y = pixels->data[i].y; // Coordinates of interest
+                if (x < minInputCols || x >= maxInputCols || y < minInputRows || y >= maxInputRows) {
+                    continue;
+                }
+                psVector *reject = pixelMapQuery(pixelMap, minInputCols, minInputRows,
+                                                 x, y); // Reject these images
+                combinePixels(combinedImage, combinedMask, combinedVariance, input, weights, varFactors,
+                              reject, x, y, maskVal, bad, numIter, rej, useVariance, safe, buffer);
+            }
+        }
+        psFree(pixels);
+        psFree(pixelMap);
+    } else {
+        // Pull the products out, allocate if necessary
+        psImage *combinedImage = combined->image; // Combined image
+        if (!combinedImage) {
+            combined->image = psImageAlloc(numCols, numRows, PS_TYPE_F32);
+            combinedImage = combined->image;
+        }
+        psImage *combinedMask = combined->mask; // Combined mask
+        if (!combinedMask) {
+            combined->mask = psImageAlloc(numCols, numRows, PS_TYPE_MASK);
+            combinedMask = combined->mask;
+        }
+
+        psImage *combinedVariance = combined->weight; // Combined variance map
+        if (haveVariances && !combinedVariance) {
+            combined->weight = psImageAlloc(numCols, numRows, PS_TYPE_F32);
+            combinedVariance = combined->weight;
+        }
+
+        for (int y = minInputRows; y < maxInputRows; y++) {
+            for (int x = minInputCols; x < maxInputCols; x++) {
+                combinePixels(combinedImage, combinedMask, combinedVariance, input, weights, varFactors,
+                              NULL, x, y, maskVal, bad, numIter, rej, useVariance, safe, buffer);
+            }
+        }
+
+#ifndef PS_NO_TRACE
+        if (psTraceGetLevel("psModules.imcombine") >= 5) {
+            for (int i = 0; i < num; i++) {
+                pmStackData *data = input->data[i]; // Stack data for this input
+                if (!data || !data->inspect) {
+                    continue;
+                }
+                psTrace("psModules.imcombine", 5, "Image %d: %ld pixels to inspect.\n", i, data->inspect->n);
+            }
+        }
+#endif
+    }
+
+    psFree(weights);
+    psFree(buffer);
+
+    return true;
+}
+
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmStack.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmStack.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmStack.h	(revision 19481)
@@ -0,0 +1,54 @@
+/* @file  pmStack.h
+ *
+ * This file will perform image combination of several images of the
+ * same field, produce a list of questionable pixels, then tag some
+ * of those pixels as defects.
+ *
+ * @author Paul Price, IfA
+ * @author GLG, MHPCC
+ *
+ * @version $Revision: 1.7 $ $Name: not supported by cvs2svn $
+ * @date $Date: 2008-03-17 21:38:43 $
+ *
+ * Copyright 2004-2007 Institute for Astronomy, University of Hawaii
+ */
+
+#ifndef PM_STACK_H
+#define PM_STACK_H
+
+#include <pslib.h>
+#include <pmHDU.h>
+#include <pmFPA.h>
+
+/// @addtogroup imcombine Image Combinations
+/// @{
+
+
+/// Container for input image
+typedef struct {
+    pmReadout *readout;                 ///< Warped readout (sky cell)
+    psPixels *reject;                   ///< Pixels to reject
+    psPixels *inspect;                  ///< Pixels to inspect
+    float weight;                       ///< Relative weighting for image
+} pmStackData;
+
+/// Constructor
+pmStackData *pmStackDataAlloc(pmReadout *readout, ///< Warped readout (sky cell)
+                              float weight ///< Weight to apply
+    );
+
+/// Stack input images
+bool pmStackCombine(pmReadout *combined,///< Combined readout (output)
+                    psArray *input,     ///< Input array of pmStackData
+                    psMaskType maskVal, ///< Mask value of bad pixels
+                    psMaskType bad,     ///< Mask value to give rejected pixels
+                    int kernelSize,     ///< Half-size of the convolution kernel
+                    int numIter,        ///< Number of iterations
+                    float rej,          ///< Rejection limit (standard deviations)
+                    bool entire,        ///< Combine entire image even if rejection lists provided?
+                    bool useVariance,   ///< Use variance values for rejection?
+                    bool safe           ///< Play safe with small numbers of input pixels (mask if N <= 2)?
+    );
+
+/// @}
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmStackReject.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmStackReject.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmStackReject.c	(revision 19481)
@@ -0,0 +1,311 @@
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <stdio.h>
+#include <string.h>
+#include <pslib.h>
+
+#include "pmSubtraction.h"
+#include "pmSubtractionThreads.h"
+#include "pmSubtractionKernels.h"
+
+#define PIXEL_LIST_BUFFER 100           // Number of pixels to add to list at a time
+
+//#define TESTING                         // Testing output
+
+// Mask values
+typedef enum {
+    PM_STACK_MASK_BAD      = 0x01,      // Bad pixel
+    PM_STACK_MASK_CONVOLVE = 0x02,      // Touching a bad pixel
+    PM_STACK_MASK_ALL      = 0xff,      // All mask bits
+} pmStackMask;
+
+static bool threaded = false;           // Running threaded?
+
+
+// Grow the rejection mask
+static inline bool stackRejectGrow(psImage *target,   // Target mask image (product)
+                                   psImage *source, // Source mask image (to be grown)
+                                   const pmSubtractionKernels *kernels, // Subtraction kernels
+                                   int numCols, int numRows, // Size of image
+                                   int xMin, int xMax, int yMin, int yMax, // Bounds of convolution
+                                   float poorFrac       // Fraction for "poor"
+    )
+{
+    int size = kernels->size;           // Half-size of convolution kernel
+    psImage *polyValues = p_pmSubtractionPolynomialFromCoords(NULL, kernels, numCols, numRows,
+                                                              xMin + size + 1, yMin + size + 1); // Polynomial
+    int box = p_pmSubtractionBadRadius(NULL, kernels, polyValues, false, poorFrac); // Radius of bad box
+    psFree(polyValues);
+
+    if (box > 0) {
+        // Convolve a subimage, then stick it in the target
+        if (threaded) {
+            psMutexLock(source);
+        }
+        psImage *mask = psImageSubset(source, psRegionSet(xMin - box, xMax + box,
+                                                          yMin - box, yMax + box)); // Mask to convolve
+        if (threaded) {
+            psMutexUnlock(source);
+        }
+        psImage *convolved = psImageConvolveMask(NULL, mask, PM_STACK_MASK_BAD, PM_STACK_MASK_CONVOLVE,
+                                                 -box, box, -box, box); // Convolved mask
+        if (threaded) {
+            psMutexLock(source);
+        }
+        psFree(mask);
+        if (threaded) {
+            psMutexUnlock(source);
+        }
+
+        int numBytes = (xMax - xMin) * PSELEMTYPE_SIZEOF(PS_TYPE_MASK); // Number of bytes to copy
+        psAssert(convolved->numCols - 2 * box == xMax - xMin, "Bad number of columns");
+        psAssert(convolved->numRows - 2 * box == yMax - yMin, "Bad number of rows");
+
+        for (int yTarget = yMin, ySource = box; yTarget < yMax; yTarget++, ySource++) {
+            memcpy(&target->data.PS_TYPE_MASK_DATA[yTarget][xMin],
+                   &convolved->data.PS_TYPE_MASK_DATA[ySource][box], numBytes);
+        }
+        psFree(convolved);
+    }
+    return true;
+}
+
+// Thread entry for stackRejectGrow
+static bool stackRejectGrowThread(psThreadJob *job // Job to execute
+    )
+{
+    PS_ASSERT_THREAD_JOB_NON_NULL(job, false);
+
+    psArray *args = job->args;          // Job arguments
+    psImage *target = args->data[0];    // Target mask image
+    psImage *source = args->data[1];    // Source mask image
+    const pmSubtractionKernels *kernels = args->data[2]; // Subtraction kernels
+    int numCols = PS_SCALAR_VALUE(args->data[3], S32); // Number of columns
+    int numRows = PS_SCALAR_VALUE(args->data[4], S32); // Number of rows
+    int xMin = PS_SCALAR_VALUE(args->data[5], S32); // Minimum x value
+    int xMax = PS_SCALAR_VALUE(args->data[6], S32); // Maximum x value
+    int yMin = PS_SCALAR_VALUE(args->data[7], S32); // Minimum y value
+    int yMax = PS_SCALAR_VALUE(args->data[8], S32); // Maximum y value
+    float poorFrac = PS_SCALAR_VALUE(args->data[9], F32); // Fraction for "poor"
+
+    return stackRejectGrow(target, source, kernels, numCols, numRows, xMin, xMax, yMin, yMax, poorFrac);
+}
+
+bool pmStackRejectThreadsInit(void)
+{
+    if (threaded) {
+        psAbort("Already running threaded.");
+    }
+    threaded = true;
+
+    if (!pmSubtractionThreaded()) {
+        pmSubtractionThreadsInit(NULL, NULL);
+    }
+
+    {
+        psThreadTask *task = psThreadTaskAlloc("PSMODULES_STACK_REJECT_GROW", 10);
+        task->function = &stackRejectGrowThread;
+        psThreadTaskAdd(task);
+        psFree(task);
+    }
+
+    return true;
+}
+
+
+psPixels *pmStackReject(const psPixels *in, int numCols, int numRows, float threshold, float poorFrac,
+                        const psArray *subRegions, const psArray *subKernels)
+{
+    PS_ASSERT_PIXELS_NON_NULL(in, NULL);
+    PS_ASSERT_FLOAT_LARGER_THAN_OR_EQUAL(threshold, 0.0, NULL);
+    PS_ASSERT_FLOAT_LESS_THAN_OR_EQUAL(threshold, 1.0, NULL);
+    PS_ASSERT_ARRAY_NON_NULL(subRegions, NULL);
+    PS_ASSERT_ARRAY_NON_NULL(subKernels, NULL);
+    PS_ASSERT_ARRAYS_SIZE_EQUAL(subRegions, subKernels, NULL);
+
+    // Trivial case
+    if (in->n == 0) {
+        return psPixelsAllocEmpty(0);
+    }
+
+    // Check consistency of kernels
+    int numRegions = subRegions->n;     // Number of regions
+    int size = 0;                       // Size of kernel
+    for (int i = 0; i < numRegions; i++) {
+        pmSubtractionKernels *kernels = subKernels->data[i]; // Kernel of interest
+        if (size == 0) {
+            size = kernels->size;
+        } else if (kernels->size != size) {
+            psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Kernel sizes are not identical: %d vs %d",
+                    size, kernels->size);
+            return NULL;
+        }
+    }
+
+    psImage *mask = psPixelsToMask(NULL, in, psRegionSet(0, numCols - 1, 0, numRows - 1), 1); // Mask
+    psImage *image = psImageCopy(NULL, mask, PS_TYPE_F32); // Floating-point version, so we can convolve
+    psFree(mask);
+
+    // Convolve the image with the kernel --- we're basically applying a matched filter and then thresholding
+    pmReadout *convRO = pmReadoutAlloc(NULL); // Readout with convolved image
+    pmReadout *inRO = pmReadoutAlloc(NULL); // Readout with input image
+    inRO->image = image;
+    if (threaded) {
+        psMutexInit(image);
+    }
+    for (int i = 0; i < numRegions; i++) {
+        psRegion *region = subRegions->data[i]; // Region of interest
+        pmSubtractionKernels *kernels = subKernels->data[i]; // Kernel of interest
+        if (!pmSubtractionConvolve(convRO, NULL, inRO, NULL, NULL, 0, 0, 1.0, region, kernels, false, true)) {
+            psError(PS_ERR_UNKNOWN, false, "Unable to convolve mask image in region %d.", i);
+            psFree(convRO);
+            psFree(inRO);
+            return NULL;
+        }
+
+        // Need to adjust the thresholding level for the normalisation of the kernel --- the application of
+        // the kernel may scale the unit level that we've inserted.
+
+        // Image of the kernel at the centre of the region
+        float xNorm = (region->x0 + 0.5 * (region->x1 - region->x0) - kernels->numCols/2.0) /
+            (float)kernels->numCols;
+        float yNorm = (region->y0 + 0.5 * (region->y1 - region->y0) - kernels->numRows/2.0) /
+            (float)kernels->numRows;
+        psImage *kernel = pmSubtractionKernelImage(kernels, xNorm, yNorm, false);
+        if (!kernel) {
+            psError(PS_ERR_UNKNOWN, false, "Unable to generate kernel image.");
+            psFree(convRO);
+            psFree(inRO);
+            return NULL;
+        }
+        float sum = 0.0;
+        for (int y = 0; y < kernel->numRows; y++) {
+            for (int x = 0; x < kernel->numCols; x++) {
+                sum += kernel->data.F32[y][x];
+            }
+        }
+        psFree(kernel);
+
+        // Range for normalisation
+        int xMin = PS_MAX(0, region->x0), xMax = PS_MIN(numCols - 1, region->x1);
+        int yMin = PS_MAX(0, region->y0), yMax = PS_MIN(numRows - 1, region->y1);
+        psTrace("psModules.imcombine", 2, "Normalising convolved mask image by %f over %d:%d,%d:%d\n",
+                sum, xMin, xMax, yMin, yMax);
+        sum = 1.0 / sum;
+        for (int y = yMin; y <= yMax; y++) {
+            for (int x = xMin; x <= xMax; x++) {
+                convRO->image->data.F32[y][x] *= sum;
+            }
+        }
+    }
+    if (threaded) {
+        psMutexDestroy(image);
+    }
+    psFree(inRO);
+    psImage *convolved = psMemIncrRefCounter(convRO->image);
+    psFree(convRO);
+
+#ifdef TESTING
+    {
+        static int seqNum = 0;          // Sequence number
+        psString name = NULL;           // Name of image
+        psStringAppend(&name, "inspect_conv_%02d.fits", seqNum);
+        seqNum++;
+        psFits *fits = psFitsOpen(name, "w"); // FITS file pointer
+        psFree(name);
+        psFitsWriteImage(fits, NULL, convolved, 0, NULL);
+        psFitsClose(fits);
+    }
+#endif
+
+    // Threshold the convolved image
+    psPixels *bad = psPixelsAllocEmpty(PIXEL_LIST_BUFFER); // List of pixels that should be masked
+    for (int y = size; y < convolved->numRows - size; y++) {
+        for (int x = size; x < convolved->numCols - size; x++) {
+            if (convolved->data.F32[y][x] > threshold) {
+                bad = psPixelsAdd(bad, bad->nalloc, x, y);
+            }
+        }
+    }
+
+    // Now, grow the mask to include everything that touches a bad pixel in the convolution
+    psImage *source = psPixelsToMask(NULL, bad, psRegionSet(0, numCols - 1, 0, numRows - 1),
+                                     PM_STACK_MASK_BAD); // Mask image to grow
+    psImage *target = psImageRecycle(convolved, numCols, numRows, PS_TYPE_MASK); // Grown image
+    psImageInit(target, 0);
+    if (threaded) {
+        psMutexInit(source);
+    }
+    for (int i = 0; i < subRegions->n; i++) {
+        psRegion *region = subRegions->data[i]; // Subtraction region
+        pmSubtractionKernels *kernels = subKernels->data[i]; // Subtraction kernel
+
+        int size = kernels->size;           // Half-size of kernel
+        int fullSize = 2 * size + 1;        // Full size of kernel
+
+        // Get region for convolution: [xMin:xMax,yMin:yMax]
+        int xMin = PS_MAX(region->x0, size), xMax = PS_MIN(region->x1, numCols - size);
+        int yMin = PS_MAX(region->y0, size), yMax = PS_MIN(region->y1, numRows - size);
+
+        for (int j = yMin; j < yMax; j += fullSize) {
+            int ySubMax = PS_MIN(j + fullSize, yMax); // Range for subregion of interest
+            for (int i = xMin; i < xMax; i += fullSize) {
+                int xSubMax = PS_MIN(i + fullSize, xMax); // Range for subregion of interest
+
+                if (threaded) {
+                    psThreadJob *job = psThreadJobAlloc("PSMODULES_STACK_REJECT_GROW"); // Job to execute
+                    psArray *args = job->args; // Job arguments
+                    psArrayAdd(args, 1, target);
+                    psMutexLock(source);
+                    psArrayAdd(args, 1, source);
+                    psMutexUnlock(source);
+                    psArrayAdd(args, 1, kernels);
+                    PS_ARRAY_ADD_SCALAR(args, numCols, PS_TYPE_S32);
+                    PS_ARRAY_ADD_SCALAR(args, numRows, PS_TYPE_S32);
+                    PS_ARRAY_ADD_SCALAR(args, i, PS_TYPE_S32);
+                    PS_ARRAY_ADD_SCALAR(args, xSubMax, PS_TYPE_S32);
+                    PS_ARRAY_ADD_SCALAR(args, j, PS_TYPE_S32);
+                    PS_ARRAY_ADD_SCALAR(args, ySubMax, PS_TYPE_S32);
+                    PS_ARRAY_ADD_SCALAR(args, poorFrac, PS_TYPE_F32);
+                    if (!psThreadJobAddPending(job)) {
+                        psFree(job);
+                        psFree(source);
+                        psFree(target);
+                        return false;
+                    }
+                    psFree(job);
+                } else if (!stackRejectGrow(target, source, kernels, numCols, numRows,
+                                            i, xSubMax, j, ySubMax, poorFrac)) {
+                    psError(PS_ERR_UNKNOWN, false, "Unable to grow bad pixels.");
+                    psFree(source);
+                    psFree(target);
+                    return NULL;
+                }
+            }
+        }
+    }
+
+    if (!psThreadPoolWait(false)) {
+        psError(PS_ERR_UNKNOWN, false, "Unable to grow bad pixels.");
+        psFree(source);
+        psFree(target);
+        return NULL;
+    }
+
+    // Harvest the jobs
+    if (threaded) {
+        psThreadJob *job;                   // Job to destroy
+        while ((job = psThreadJobGetDone())) {
+            psFree(job);
+        }
+
+        psMutexDestroy(source);
+    }
+
+    psFree(source);
+    bad = psPixelsFromMask(bad, target, PM_STACK_MASK_ALL);
+
+    return bad;
+}
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmStackReject.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmStackReject.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmStackReject.h	(revision 19481)
@@ -0,0 +1,22 @@
+#ifndef PM_STACK_REJECT_H
+#define PM_STACK_REJECT_H
+
+#include <pslib.h>
+#include <pmSubtractionKernels.h>
+
+/// Given a list of pixels from the convolved image, find the corresponding (smaller subset of) pixels in the
+/// original image, and then convolve those to get the list of all pixels which should be rejected
+///
+/// We apply a matched filter to the corresponding mask image, and threshold to find the original pixels
+psPixels *pmStackReject(const psPixels *in, ///< List of pixels in the convolved image
+                        int numCols, int numRows, ///< Size of image of interest
+                        float threshold, ///< Threshold on convolved image, 0..1
+                        float poorFrac, ///< Fraction for "poor"
+                        const psArray *regions, ///< Array of image regions for image
+                        const psArray *kernels ///< Array of kernel parameters for each region
+    );
+
+/// Initialise threads for stack rejection
+bool pmStackRejectThreadsInit(void);
+
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtraction.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtraction.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtraction.c	(revision 19481)
@@ -0,0 +1,1211 @@
+/** @file pmSubtraction.c
+ *
+ *  @author Paul Price, IfA
+ *  @author GLG, MHPCC
+ *
+ *  Copyright 2004-2007 Institute for Astronomy, University of Hawaii
+ *
+ */
+
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <stdio.h>
+#include <math.h>
+#include <string.h>
+#include <pslib.h>
+
+#include "pmHDU.h"                      // Required for pmFPA.h
+#include "pmFPA.h"
+#include "pmSubtractionStamps.h"
+#include "pmSubtractionEquation.h"
+#include "pmSubtractionThreads.h"
+
+#include "pmSubtraction.h"
+
+//#define TESTING
+
+#define PIXEL_LIST_BUFFER 100           // Number of entries to add to pixel list at a time
+#define MIN_SAMPLE_STATS    7           // Minimum number to use sample statistics; otherwise use quartiles
+
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Private (file-static) functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+// Generate the kernel to apply to the variance from the normal kernel
+static psKernel *varianceKernel(psKernel *out, // Output kernel
+                                psKernel *normalKernel // Normal kernel
+                                )
+{
+    // Kernel range
+    int xMin = normalKernel->xMin, xMax = normalKernel->xMax;
+    int yMin = normalKernel->yMin, yMax = normalKernel->yMax;
+
+    if (!out) {
+        out = psKernelAlloc(xMin, xMax, yMin, yMax);
+    }
+
+    // Take the square of the normal kernel
+    double sumNormal = 0.0, sumVariance = 0.0; // Sum of the normal and variance kernels
+    for (int v = yMin; v <= yMax; v++) {
+        for (int u = xMin; u <= xMax; u++) {
+            float value = normalKernel->kernel[v][u]; // Value of interest
+            float value2 = PS_SQR(value); // Value squared
+            sumNormal += value;
+            sumVariance += value2;
+            out->kernel[v][u] = value2;
+        }
+    }
+
+    // Normalise so that the sum of the variance kernel is the square of the sum of the normal kernel
+    // This is required to keep the relative scaling between the image and the weight map
+    psBinaryOp(out->image, out->image, "*", psScalarAlloc(PS_SQR(sumNormal) / sumVariance, PS_TYPE_F32));
+
+    return out;
+}
+
+// Generate an image of the solved kernel
+static psKernel *solvedKernel(psKernel *kernel, // Kernel, to return
+                              const pmSubtractionKernels *kernels, // Kernel basis functions
+                              const psImage *polyValues, // Spatial polynomial values
+                              bool wantDual // Want the dual (second) kernel?
+                              )
+{
+    assert(kernels);
+    assert(polyValues);
+
+    int numKernels = kernels->num;      // Number of kernel basis functions
+    int size = kernels->size;           // Kernel half-size
+    if (!kernel) {
+        kernel = psKernelAlloc(-size, size, -size, size);
+    }
+    psImageInit(kernel->image, 0.0);
+
+    for (int i = 0; i < numKernels; i++) {
+        double value = p_pmSubtractionSolutionCoeff(kernels, polyValues, i, wantDual); // Polynomial value
+
+        switch (kernels->type) {
+          case PM_SUBTRACTION_KERNEL_POIS: {
+              int u = kernels->u->data.S32[i]; // Offset in x
+              int v = kernels->v->data.S32[i]; // Offset in y
+              kernel->kernel[v][u] += value;
+              kernel->kernel[0][0] -= value;
+              break;
+          }
+          /* SPAM and FRIES use the same method */
+          case PM_SUBTRACTION_KERNEL_SPAM:
+          case PM_SUBTRACTION_KERNEL_FRIES: {
+              int uStart = kernels->u->data.S32[i];
+              int uStop = kernels->uStop->data.S32[i];
+              int vStart = kernels->v->data.S32[i];
+              int vStop = kernels->vStop->data.S32[i];
+
+              // Normalising sum of kernel component to unity
+              float norm = 1.0 / (float)((uStop - uStart + 1) * (vStop - vStart + 1));
+
+              for (int v = vStart; v <= vStop; v++) {
+                  for (int u = uStart; u <= uStop; u++) {
+                      kernel->kernel[v][u] += norm * value;
+                      kernel->kernel[0][0] -= value;
+                  }
+              }
+              break;
+          }
+          case PM_SUBTRACTION_KERNEL_GUNK: {
+              if (i < kernels->inner) {
+                  // Using pre-calculated function
+                  psKernel *preCalc = kernels->preCalc->data[i]; // Precalculated values
+                  // Iterating over the kernel
+                  for (int v = -size; v <= size; v++) {
+                      for (int u = -size; u <= size; u++) {
+                          kernel->kernel[v][u] += preCalc->kernel[v][u] * value;
+                          // Photometric scaling is built into the preCalc kernel --- no subtraction!
+                      }
+                  }
+              } else {
+                  // Using delta function
+                  int u = kernels->u->data.S32[i]; // Offset in x
+                  int v = kernels->v->data.S32[i]; // Offset in y
+                  kernel->kernel[v][u] += value;
+                  kernel->kernel[0][0] -= value;
+              }
+              break;
+          }
+          case PM_SUBTRACTION_KERNEL_ISIS: {
+              psKernel *preCalc = kernels->preCalc->data[i]; // Precalculated values
+              // Iterating over the kernel
+              for (int v = -size; v <= size; v++) {
+                  for (int u = -size; u <= size; u++) {
+                      kernel->kernel[v][u] += preCalc->kernel[v][u] * value;
+                      // Photometric scaling is built into the preCalc kernel --- no subtraction!
+                  }
+              }
+              break;
+          }
+          case PM_SUBTRACTION_KERNEL_RINGS: {
+              psArray *preCalc = kernels->preCalc->data[i]; // Precalculated data
+              psVector *uCoords = preCalc->data[0]; // u coordinates
+              psVector *vCoords = preCalc->data[1]; // v coordinates
+              psVector *poly = preCalc->data[2]; // Polynomial values
+              int num = uCoords->n;     // Number of pixels
+
+              for (int j = 0; j < num; j++) {
+                  int u = uCoords->data.S32[j], v = vCoords->data.S32[j]; // Kernel coordinates
+                  kernel->kernel[v][u] += poly->data.F32[j] * value;
+              }
+              // Photometric scaling is built into the kernel --- no subtraction!
+              break;
+          }
+          default:
+            psAbort("Should never get here.");
+        }
+    }
+
+    // Put in the normalisation component
+    kernel->kernel[0][0] += (wantDual ? 1.0 : p_pmSubtractionSolutionNorm(kernels));
+
+    return kernel;
+}
+
+// Subtract the (0,0) element to preserve photometric scaling
+static void convolveSub(psKernel *convolved, // Convolved image
+                        const psKernel *image, // Image being convolved
+                        int footprint  // Size of region of interest
+                        )
+{
+    // Can't use psBinaryOp because the images are of different size
+    for (int y = -footprint; y <= footprint; y++) {
+        for (int x = -footprint; x <= footprint; x++) {
+            convolved->kernel[y][x] -= image->kernel[y][x];
+        }
+    }
+    return;
+}
+
+// Generate the convolution given some offset
+static psKernel *convolveOffset(const psKernel *image, // Image to convolve (a kernel for convenience)
+                                int u, int v, // Offset to apply
+                                int footprint // Size of region of interest
+                                )
+{
+    psKernel *convolved = psKernelAlloc(-footprint, footprint, -footprint, footprint); // Convolved image
+    int numBytes = (2 * footprint + 1) * PSELEMTYPE_SIZEOF(PS_TYPE_F32); // Number of bytes to copy
+    for (int y = -footprint; y <= footprint; y++) {
+        // Convolution with a delta function is just the value specified by the offset
+        memcpy(&convolved->kernel[y][-footprint], &image->kernel[y - v][-footprint - u], numBytes);
+    }
+    return convolved;
+}
+
+
+// Convolve an image using FFT
+static void convolveFFT(psImage *target,// Place the result in here
+                        psImage *image, // Image to convolve
+                        psImage *mask, // Mask image
+                        psMaskType maskVal, // Value to mask
+                        const psKernel *kernel, // Kernel by which to convolve
+                        psRegion region,// Region of interest
+                        float background, // Background to add
+                        int size        // Size of (square) kernel
+                        )
+{
+    psRegion border = psRegionSet(region.x0 - size, region.x1 + size,
+                                  region.y0 - size, region.y1 + size); // Add a border
+
+    bool threaded = pmSubtractionThreaded(); // Are we running threaded?
+
+    if (threaded) {
+        psMutexLock(image);
+    }
+    psImage *subImage = psImageSubset(image, border); // Subimage to convolve
+    if (threaded) {
+        psMutexUnlock(image);
+    }
+
+    psImage *subMask;                   // Subimage mask
+    if (mask) {
+        if (threaded) {
+            psMutexLock(mask);
+        }
+        subMask = psImageSubset(mask, border);
+        if (threaded) {
+            psMutexUnlock(mask);
+        }
+    } else {
+        subMask = NULL;
+    }
+
+    psImage *convolved = psImageConvolveFFT(NULL, subImage, subMask, maskVal, kernel); // Convolution
+
+    if (threaded) {
+        psMutexLock(image);
+    }
+    psFree(subImage);
+    if (threaded) {
+        psMutexUnlock(image);
+    }
+
+    if (mask) {
+        if (threaded) {
+            psMutexLock(mask);
+        }
+        psFree(subMask);
+        if (threaded) {
+            psMutexUnlock(mask);
+        }
+    }
+
+    // Now, we have to stick it in where it belongs
+    int xMin = region.x0, xMax = region.x1, yMin = region.y0, yMax = region.y1; // Bounds of region
+    if (background != 0.0) {
+        for (int yTarget = yMin, ySource = size; yTarget < yMax; yTarget++, ySource++) {
+            for (int xTarget = xMin, xSource = size; xTarget < xMax; xTarget++, xSource++) {
+                target->data.F32[yTarget][xTarget] = convolved->data.F32[ySource][xSource] + background;
+            }
+        }
+    } else {
+        int numBytes = (xMax - xMin) * PSELEMTYPE_SIZEOF(PS_TYPE_F32); // Number of bytes to copy
+        for (int yTarget = yMin, ySource = size; yTarget < yMax; yTarget++, ySource++) {
+            memcpy(&target->data.F32[yTarget][xMin], &convolved->data.F32[ySource][size], numBytes);
+        }
+    }
+    psFree(convolved);
+
+    return;
+}
+
+// Convolve an image directly
+static void convolveDirect(psImage *target, // Put the result here
+                           const psImage *image, // Image to convolve
+                           const psKernel *kernel, // Kernel by which to convolve
+                           psRegion region,// Region of interest
+                           float background, // Background to add
+                           int size        // Size of (square) kernel
+                           )
+{
+    for (int y = region.y0; y < region.y1; y++) {
+        for (int x = region.x0; x < region.x1; x++) {
+            target->data.F32[y][x] = background;
+            for (int v = -size; v <= size; v++) {
+                for (int u = -size; u <= size; u++) {
+                    target->data.F32[y][x] += kernel->kernel[v][u] * image->data.F32[y - v][x - u];
+                }
+            }
+        }
+    }
+    return;
+}
+
+// Convolve a region of an image
+static inline void convolveRegion(psImage *convImage, // Convolved image (output)
+                                  psImage *convWeight, // Convolved weight map (output), or NULL
+                                  psImage *convMask, // Convolve mask (output), or NULL
+                                  psKernel **kernelImage, // Convolution kernel for the image
+                                  psKernel **kernelWeight, // Convolution kernel for the weight map, or NULL
+                                  psImage *image, // Image to convolve
+                                  psImage *weight, // Weight map to convolve, or NULL
+                                  psImage *subMask, // Subtraction mask
+                                  const pmSubtractionKernels *kernels, // Kernels
+                                  const psImage *polyValues, // Polynomial values
+                                  float background, // Background value to apply
+                                  psRegion region, // Region to convolve
+                                  psMaskType maskBad, // Value to give bad pixels
+                                  psMaskType maskPoor, // Value to give poor pixels
+                                  float poorFrac, // Fraction for "poor"
+                                  bool useFFT,  // Use FFT to convolve?
+                                  bool wantDual // Want the dual convolution?
+    )
+{
+    *kernelImage = solvedKernel(*kernelImage, kernels, polyValues, wantDual);
+    if (weight || subMask) {
+        *kernelWeight = varianceKernel(*kernelWeight, *kernelImage);
+    }
+
+    psMaskType subBad;                  // Bad pixels in subtraction mask
+    psMaskType subConvBad;              // Bad pixels in subtraction mask when convolving
+    psMaskType subConvPoor;             // Poor pixels in subtraction mask when convolving
+    if (kernels->mode == PM_SUBTRACTION_MODE_1 || (kernels->mode == PM_SUBTRACTION_MODE_DUAL && !wantDual)) {
+        subBad = PM_SUBTRACTION_MASK_BAD_1;
+        subConvBad = PM_SUBTRACTION_MASK_CONVOLVE_BAD_1;
+        subConvPoor = PM_SUBTRACTION_MASK_CONVOLVE_1;
+    } else {
+        subBad = PM_SUBTRACTION_MASK_BAD_2;
+        subConvBad = PM_SUBTRACTION_MASK_CONVOLVE_BAD_2;
+        subConvPoor = PM_SUBTRACTION_MASK_CONVOLVE_2;
+    }
+
+    // Convolve the image and weight
+    if (useFFT) {
+        // Use Fast Fourier Transform to do the convolution
+        // This provides a big speed-up for large kernels
+        convolveFFT(convImage, image, subMask, subBad, *kernelImage, region, background, kernels->size);
+        if (weight) {
+            convolveFFT(convWeight, weight, subMask, subBad, *kernelWeight, region, 0.0, kernels->size);
+        }
+    } else {
+        // XXX Direct convolution doesn't account for bad pixels yet
+        convolveDirect(convImage, image, *kernelImage, region, background, kernels->size);
+        if (weight) {
+            convolveDirect(convWeight, weight, *kernelWeight, region, 0.0, kernels->size);
+        }
+    }
+
+    // Convolve the mask for bad pixels
+    if (subMask && convMask) {
+        int box = p_pmSubtractionBadRadius(*kernelImage, kernels, polyValues,
+                                           wantDual, poorFrac); // Size of bad box
+        if (box > 0) {
+            int colMin = region.x0, colMax = region.x1, rowMin = region.y0, rowMax = region.y1; // Bounds
+            bool threaded = pmSubtractionThreaded(); // Are we running threaded?
+
+            if (threaded) {
+                psMutexLock(subMask);
+            }
+            psImage *image = psImageSubset(subMask, psRegionSet(colMin - box, colMax + box,
+                                                                rowMin - box, rowMax + box)); // Mask
+            if (threaded) {
+                psMutexUnlock(subMask);
+            }
+
+            psImage *convolved = psImageConvolveMask(NULL, image, subBad, subConvBad,
+                                                     -box, box, -box, box); // Convolved subtraction mask
+
+            if (threaded) {
+                psMutexLock(subMask);
+            }
+            psFree(image);
+            if (threaded) {
+                psMutexUnlock(subMask);
+            }
+
+            psAssert(convolved->numCols - 2 * box == colMax - colMin, "Bad number of columns");
+            psAssert(convolved->numRows - 2 * box == rowMax - rowMin, "Bad number of rows");
+
+            for (int yTarget = rowMin, ySource = box; yTarget < rowMax; yTarget++, ySource++) {
+                // Dereference images
+                psMaskType *target = &convMask->data.PS_TYPE_MASK_DATA[yTarget][colMin]; // Target values
+                psMaskType *source = &convolved->data.PS_TYPE_MASK_DATA[ySource][box]; // Source values
+                for (int xTarget = colMin; xTarget < colMax; xTarget++, target++, source++) {
+                    if (*source & subConvBad) {
+                        *target |= maskBad;
+                    } else if (*source & subConvPoor) {
+                        *target |= maskPoor;
+                    }
+                }
+            }
+
+            // No need to lock: we own this
+            psFree(convolved);
+        }
+    }
+
+    return;
+}
+
+
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Semi-public functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+psKernel *p_pmSubtractionConvolveStampPrecalc(const psKernel *image, const psKernel *kernel)
+{
+    PS_ASSERT_KERNEL_NON_NULL(image, NULL);
+    PS_ASSERT_KERNEL_NON_NULL(kernel, NULL);
+
+    psImage *conv = psImageConvolveFFT(NULL, image->image, NULL, 0, kernel); // Convolved image
+    int x0 = - image->xMin, y0 = - image->yMin; // Position of centre of convolved image
+    psKernel *convolved = psKernelAllocFromImage(conv, x0, y0); // Kernel version
+    psFree(conv);
+    return convolved;
+}
+
+int p_pmSubtractionBadRadius(psKernel *preKernel, const pmSubtractionKernels *kernels,
+                             const psImage *polyValues, bool wantDual, float poorFrac)
+{
+    psKernel *kernel;                   // Kernel to use
+    if (!preKernel) {
+        kernel = solvedKernel(NULL, kernels, polyValues, wantDual);
+    } else {
+        kernel = psMemIncrRefCounter(preKernel);
+    }
+    PS_ASSERT_IMAGE_NON_NULL(polyValues, -1);
+
+    int xMin = kernel->xMin, xMax = kernel->xMax, yMin = kernel->yMin, yMax = kernel->yMax; // Bounds
+
+    // Determine the threshold between bad and poor
+    double sumKernel2 = 0.0;            // Sum of the kernel-squared
+    for (int y = yMin; y <= yMax; y++) {
+        for (int x = xMin; x <= xMax; x++) {
+            sumKernel2 += PS_SQR(kernel->kernel[y][x]);
+        }
+    }
+    float threshold = sumKernel2 * poorFrac; // Threshold between poor and bad
+
+    // Get bounds of threshold region
+    // Start with the entire kernel, and keep reducing the size of the box until it sum goes above threshold
+    int box = kernels->size;            // Size of box with bad pixels
+    for (double sumBox = 0.0; sumBox < threshold && box > 0; box--) {
+        for (int x = -box; x <= box; x++) {
+            sumBox += PS_SQR(kernel->kernel[-box][x]) + PS_SQR(kernel->kernel[box][x]);
+        }
+        for (int y = -box + 1; y <= box - 1; y++) {
+            // Note: not doing corners
+            sumBox += PS_SQR(kernel->kernel[y][-box]) + PS_SQR(kernel->kernel[y][box]);
+        }
+    }
+
+    psFree(kernel);
+
+    return box;
+}
+
+psImage *p_pmSubtractionPolynomialFromCoords(psImage *output, const pmSubtractionKernels *kernels,
+                                             int numCols, int numRows, int x, int y)
+{
+    assert(kernels);
+    assert(numCols > 0 && numRows > 0);
+
+    // Size to use when calculating normalised coordinates (different from actual size when convolving
+    // subimage)
+    int xNormSize = (kernels->numCols > 0 ? kernels->numCols : numCols);
+    int yNormSize = (kernels->numRows > 0 ? kernels->numRows : numRows);
+
+    // Normalised coordinates
+    float yNorm = 2.0 * (float)(y - yNormSize/2.0) / (float)yNormSize;
+    float xNorm = 2.0 * (float)(x - xNormSize/2.0) / (float)xNormSize;
+
+    return p_pmSubtractionPolynomial(output, kernels->spatialOrder, xNorm, yNorm);
+}
+
+psImage *p_pmSubtractionPolynomial(psImage *output, int spatialOrder, float x, float y)
+{
+    assert(spatialOrder >= 0);
+    assert(x >= -1 && x <= 1);
+    assert(y >= -1 && y <= 1);
+
+    output = psImageRecycle(output, spatialOrder + 1, spatialOrder + 1, PS_TYPE_F64);
+    output->data.F64[0][0] = 1.0;
+
+    double value = 1.0;
+    for (int i = 1; i <= spatialOrder; i++) {
+        value *= x;
+        output->data.F64[0][i] = value;
+    }
+
+    value = 1.0;
+    for (int j = 1; j <= spatialOrder; j++) {
+        value *= y;
+        output->data.F64[j][0] = value;
+    }
+
+    for (int j = 1; j <= spatialOrder; j++) {
+        for (int i = 1; i <= spatialOrder - j; i++) {
+            output->data.F64[j][i] = output->data.F64[j][0] * output->data.F64[0][i];
+        }
+    }
+
+    return output;
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Public functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+
+// Convolve a stamp by a single kernel basis function
+static psKernel *convolveStampSingle(const pmSubtractionKernels *kernels, // Kernel basis functions
+                                     int index, // Kernel basis function index
+                                     const psKernel *image, // Image to convolve (a kernel for convenience)
+                                     int footprint // Size of region of interest
+    )
+{
+#if 0
+    if (index == kernels->num) {
+        // Normalisation component
+        return convolveOffset(image, 0, 0, footprint);
+    }
+#endif
+
+    switch (kernels->type) {
+      case PM_SUBTRACTION_KERNEL_POIS: {
+          int u = kernels->u->data.S32[index]; // Offset in x
+          int v = kernels->v->data.S32[index]; // Offset in y
+          psKernel *convolved = convolveOffset(image, u, v, footprint); // Convolved image
+          convolveSub(convolved, image, footprint);
+          return convolved;
+      }
+        // Method for SPAM and FRIES is the same
+      case PM_SUBTRACTION_KERNEL_SPAM:
+      case PM_SUBTRACTION_KERNEL_FRIES: {
+          psKernel *convolved = psKernelAlloc(-footprint, footprint,
+                                              -footprint, footprint); // Convolved image
+          int uStart = kernels->u->data.S32[index];
+          int uStop = kernels->uStop->data.S32[index];
+          int vStart = kernels->v->data.S32[index];
+          int vStop = kernels->vStop->data.S32[index];
+          float norm = 1.0 / (uStop - uStart + 1) * (vStop - vStart + 1); // Normalisation
+          for (int y = -footprint; y <= footprint; y++) {
+              for (int x = -footprint; x <= footprint; x++) {
+                  double sum = 0.0;
+                  for (int v = vStart; v <= vStop; v++) {
+                      for (int u = uStart; u <= uStop; u++) {
+                          sum += image->kernel[y - v][x - u];
+                      }
+                  }
+                  convolved->kernel[y][x] = norm * sum;
+              }
+          }
+          convolveSub(convolved, image, footprint);
+          return convolved;
+      }
+      case PM_SUBTRACTION_KERNEL_GUNK: {
+          if (index < kernels->inner) {
+              // Photometric scaling is already built in to the precalculated kernel
+              return p_pmSubtractionConvolveStampPrecalc(image, kernels->preCalc->data[index]);
+          }
+          // Using delta function
+          int u = kernels->u->data.S32[index]; // Offset in x
+          int v = kernels->v->data.S32[index]; // Offset in y
+          psKernel *convolved = convolveOffset(image, u, v, footprint); // Convolved image
+          convolveSub(convolved, image, footprint);
+          return convolved;
+      }
+      case PM_SUBTRACTION_KERNEL_ISIS: {
+          // Photometric scaling is already built in to the precalculated kernel
+          return p_pmSubtractionConvolveStampPrecalc(image, kernels->preCalc->data[index]);
+      }
+      case PM_SUBTRACTION_KERNEL_RINGS: {
+          psKernel *convolved = psKernelAlloc(-footprint, footprint,
+                                              -footprint, footprint); // Convolved image
+          psArray *preCalc = kernels->preCalc->data[index]; // Precalculated data
+          psVector *uCoords = preCalc->data[0]; // u coordinates
+          psVector *vCoords = preCalc->data[1]; // v coordinates
+          psVector *poly = preCalc->data[2]; // Polynomial values
+          int num = uCoords->n;         // Number of pixels
+          for (int y = -footprint; y <= footprint; y++) {
+              for (int x = -footprint; x <= footprint; x++) {
+                  double sum = 0.0;             // Accumulated sum from convolution
+                  for (int j = 0; j < num; j++) {
+                      int u = uCoords->data.S32[j], v = vCoords->data.S32[j]; // Kernel coordinates
+                      sum += image->kernel[y - v][x - u] * poly->data.F32[j];
+                  }
+                  convolved->kernel[y][x] = sum;
+                  // Photometric scaling is built into the kernel --- no subtraction!
+              }
+          }
+          return convolved;
+      }
+      default:
+        psAbort("Should never get here.");
+    }
+    return NULL;
+}
+
+// Convolve the stamp by each of the kernel basis functions
+static psArray *convolveStamp(psArray *convolutions, // The convolutions
+                              const psKernel *image, // Image to convolve
+                              const pmSubtractionKernels *kernels, // Kernel basis functions
+                              int footprint // Stamp half-size
+    )
+{
+    assert(image);
+    assert(kernels);
+    assert(footprint >= 0);
+
+    if (convolutions) {
+        // Already done
+        return convolutions;
+    }
+
+    int numKernels = kernels->num;      // Number of kernels
+    convolutions = psArrayAlloc(numKernels);
+
+    for (int i = 0; i < numKernels; i++) {
+        convolutions->data[i] = convolveStampSingle(kernels, i, image, footprint);
+    }
+
+    return convolutions;
+}
+
+
+bool pmSubtractionConvolveStamp(pmSubtractionStamp *stamp, const pmSubtractionKernels *kernels, int footprint)
+{
+    PS_ASSERT_PTR_NON_NULL(stamp, false);
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, false);
+    PS_ASSERT_INT_NONNEGATIVE(footprint, false);
+
+    if (stamp->status != PM_SUBTRACTION_STAMP_CALCULATE) {
+        psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Stamp not marked for calculation.");
+        return false;
+    }
+
+    switch (kernels->mode) {
+      case PM_SUBTRACTION_MODE_1:
+        stamp->convolutions1 = convolveStamp(stamp->convolutions1, stamp->image1, kernels, footprint);
+        break;
+      case PM_SUBTRACTION_MODE_2:
+        stamp->convolutions2 = convolveStamp(stamp->convolutions2, stamp->image2, kernels, footprint);
+        break;
+      case PM_SUBTRACTION_MODE_UNSURE:
+      case PM_SUBTRACTION_MODE_DUAL:
+        stamp->convolutions1 = convolveStamp(stamp->convolutions1, stamp->image1, kernels, footprint);
+        stamp->convolutions2 = convolveStamp(stamp->convolutions2, stamp->image2, kernels, footprint);
+        break;
+      default:
+        psAbort("Unsupported subtraction mode: %x", kernels->mode);
+    }
+
+    return true;
+}
+
+
+
+
+int pmSubtractionRejectStamps(float *rmsPtr, int *numPtr, pmSubtractionStampList *stamps,
+                              const psVector *deviations, psImage *subMask, float sigmaRej,
+                              int footprint)
+{
+    PM_ASSERT_SUBTRACTION_STAMP_LIST_NON_NULL(stamps, -1);
+    PS_ASSERT_VECTOR_NON_NULL(deviations, -1);
+    PS_ASSERT_VECTOR_TYPE(deviations, PS_TYPE_F32, -1);
+    PS_ASSERT_IMAGE_NON_EMPTY(subMask, -1);
+    PS_ASSERT_IMAGE_TYPE(subMask, PS_TYPE_MASK, -1);
+
+    // I used to measure the rms deviation about zero, and use that as the sigma against which to clip, but
+    // the distribution is actually something like a chi^2 or Student's t, both of which become Gaussian-like
+    // with large N.  Therefore, let's just treat this as a Gaussian distribution.
+
+    int numStamps = 0;                  // Number of used stamps
+    psVector *mask = psVectorAlloc(stamps->num, PS_TYPE_MASK); // Mask, for statistics
+    psVectorInit(mask, 0);
+    for (int i = 0; i < stamps->num; i++) {
+        pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+        if (stamp->status != PM_SUBTRACTION_STAMP_USED) {
+            mask->data.PS_TYPE_MASK_DATA[i] = 0xff;
+            continue;
+        }
+        numStamps++;
+    }
+    psTrace("psModules.imcombine", 1, "Number of good stamps: %d\n", numStamps);
+
+    if (numStamps == 0) {
+        psError(PS_ERR_UNKNOWN, true, "No good stamps found.");
+        psFree(mask);
+        return -1;
+    }
+
+    psStats *stats = psStatsAlloc(PS_STAT_SAMPLE_MEAN | PS_STAT_SAMPLE_STDEV |
+                                  PS_STAT_SAMPLE_MEDIAN | PS_STAT_SAMPLE_QUARTILE); // Statistics for deviatns
+    if (!psVectorStats(stats, deviations, NULL, mask, 0xff)) {
+        psError(PS_ERR_UNKNOWN, false, "Unable to measure statistics for deviations.");
+        psFree(stats);
+        psFree(mask);
+        return -1;
+    }
+    psFree(mask);
+
+    double mean, rms;                 // Mean and RMS of deviations
+    if (numStamps < MIN_SAMPLE_STATS) {
+        mean = stats->sampleMean;
+        rms = stats->sampleStdev;
+    } else {
+        mean = stats->sampleMedian;
+        rms = 0.74 * (stats->sampleUQ - stats->sampleLQ);
+    }
+    psFree(stats);
+
+    psTrace("psModules.imcombine", 1, "Mean: %f\n", mean);
+    psTrace("psModules.imcombine", 1, "RMS deviation: %f\n", rms);
+
+    if (rmsPtr) {
+        *rmsPtr = rms;
+    }
+    if (numPtr) {
+        *numPtr = numStamps;
+    }
+
+    psLogMsg("psModules.imcombine", PS_LOG_INFO, "RMS deviation from %d stamps: %lf", numStamps, rms);
+
+    if (!isfinite(sigmaRej) || sigmaRej <= 0.0) {
+        // User just wanted to calculate and record the RMS for posterity
+        return 0;
+    }
+
+    float limit = sigmaRej * rms; // Limit on maximum deviation
+    psTrace("psModules.imcombine", 1, "Deviation limit: %f\n", limit);
+
+    int numRejected = 0;                // Number of stamps rejected
+    int numGood = 0;                    // Number of good stamps
+    double newMean = 0.0;               // New mean
+    for (int i = 0; i < stamps->num; i++) {
+        pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+        if (stamp->status == PM_SUBTRACTION_STAMP_USED) {
+            // Should we reject stars with low deviation?  Well, if this is really a Gaussian-like
+            // distribution and they're low, then we have the right to ask why.  Isn't it suspicious that
+            // they're anomalously low, compared to the rest of the population which (we hope) is indicative
+            // of normality?  Besides, the standard deviation is going to be blown up by stars that didn't
+            // subtract well, in which case very few (if any) stars will be legitimately rejected for being
+            // low.
+            if (fabsf(deviations->data.F32[i] - mean) > limit) {
+                // Mask out the stamp in the image so you it's not found again
+                psTrace("psModules.imcombine", 3, "Rejecting stamp %d (%d,%d)\n", i,
+                        (int)(stamp->x + 0.5), (int)(stamp->y + 0.5));
+                numRejected++;
+                for (int y = stamp->y - footprint; y <= stamp->y + footprint; y++) {
+                    for (int x = stamp->x - footprint; x <= stamp->x + footprint; x++) {
+                        subMask->data.PS_TYPE_MASK_DATA[y][x] |= PM_SUBTRACTION_MASK_REJ;
+                    }
+                }
+
+                // Set stamp for replacement
+                stamp->x = 0;
+                stamp->y = 0;
+                stamp->xNorm = NAN;
+                stamp->yNorm = NAN;
+                stamp->status = PM_SUBTRACTION_STAMP_REJECTED;
+                // Recalculate convolutions
+                psFree(stamp->convolutions1);
+                psFree(stamp->convolutions2);
+                stamp->convolutions1 = stamp->convolutions2 = NULL;
+                psFree(stamp->image1);
+                psFree(stamp->image2);
+                psFree(stamp->weight);
+                stamp->image1 = stamp->image2 = stamp->weight = NULL;
+                psFree(stamp->matrix1);
+                psFree(stamp->matrix2);
+                psFree(stamp->matrixX);
+                stamp->matrix1 = stamp->matrix2 = stamp->matrixX = NULL;
+                psFree(stamp->vector1);
+                psFree(stamp->vector2);
+                stamp->vector1 = stamp->vector2 = NULL;
+            } else {
+                numGood++;
+                newMean += deviations->data.F32[i];
+            }
+        }
+    }
+    newMean /= numGood;
+
+    if (numRejected > 0) {
+        psLogMsg("psModules.imcombine", PS_LOG_INFO,
+                 "%d good stamps; %d rejected.\nMean deviation: %lf --> %lf\n",
+                 numGood, numRejected, mean, newMean);
+    } else {
+        psLogMsg("psModules.imcombine", PS_LOG_INFO,
+                 "%d good stamps; 0 rejected.\nMean deviation: %lf\n",
+                 numGood, mean);
+    }
+
+    return numRejected;
+}
+
+psImage *pmSubtractionKernelImage(const pmSubtractionKernels *kernels, float x, float y, bool wantDual)
+{
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, NULL);
+    PM_ASSERT_SUBTRACTION_KERNELS_SOLUTION(kernels, NULL);
+    PS_ASSERT_FLOAT_WITHIN_RANGE(x, -1.0, 1.0, NULL);
+    PS_ASSERT_FLOAT_WITHIN_RANGE(y, -1.0, 1.0, NULL);
+
+    // Precalulate polynomial values
+    psImage *polyValues = p_pmSubtractionPolynomial(NULL, kernels->spatialOrder, x, y);
+
+    // The appropriate kernel
+    psKernel *kernel = solvedKernel(NULL, kernels, polyValues, wantDual);
+
+    psFree(polyValues);
+
+    psImage *image = psMemIncrRefCounter(kernel->image); // Image of the kernel
+    psFree(kernel);
+
+    return image;
+}
+
+
+float pmSubtractionVarianceFactor(const pmSubtractionKernels *kernels, float x, float y, bool wantDual)
+{
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, NAN);
+    PM_ASSERT_SUBTRACTION_KERNELS_SOLUTION(kernels, NAN);
+    PS_ASSERT_FLOAT_WITHIN_RANGE(x, -1.0, 1.0, NAN);
+    PS_ASSERT_FLOAT_WITHIN_RANGE(y, -1.0, 1.0, NAN);
+
+    // Precalulate polynomial values
+    psImage *polyValues = p_pmSubtractionPolynomial(NULL, kernels->spatialOrder, x, y);
+
+    psKernel *kernel = solvedKernel(NULL, kernels, polyValues, wantDual); // The appropriate kernel
+    psFree(polyValues);
+
+    double sumKernel2 = 0.0;            // Sum of the kernel squared
+    double sumKernel = 0.0;             // Sum of the kernel
+    for (int y = kernel->yMin; y <= kernel->yMax; y++) {
+        for (int x = kernel->xMin; x <= kernel->xMax; x++) {
+            sumKernel += kernel->kernel[y][x];
+            sumKernel2 += PS_SQR(kernel->kernel[y][x]);
+        }
+    }
+
+    psFree(kernel);
+
+    return sumKernel2 / PS_SQR(sumKernel);
+}
+
+#if 0
+psArray *pmSubtractionKernelSolutions(const pmSubtractionKernels *kernels, float x, float y, bool wantDual)
+{
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, NULL);
+    PM_ASSERT_SUBTRACTION_KERNELS_SOLUTION(kernels, NULL);
+    PS_ASSERT_FLOAT_WITHIN_RANGE(x, -1.0, 1.0, NULL);
+    PS_ASSERT_FLOAT_WITHIN_RANGE(y, -1.0, 1.0, NULL);
+
+    psArray *images = psArrayAlloc(solution->n - 1); // Images of each kernel to return
+    psVector *fakeSolution = psVectorAlloc(solution->n, PS_TYPE_F64); // Fake solution vector
+    psVectorInit(fakeSolution, 0.0);
+
+    for (int i = 0; i < solution->n - 1; i++) {
+        fakeSolution->data.F64[i] = solution->data.F64[i];
+        images->data[i] = pmSubtractionKernelImage(kernels, x, y, wantDual);
+        fakeSolution->data.F64[i] = 0.0;
+    }
+
+    psFree(fakeSolution);
+
+    return images;
+}
+#endif
+
+
+// XXX Put kernelImage, kernelWeight and polyValues on thread-dependent data
+static bool subtractionConvolvePatch(int numCols, int numRows, // Size of image
+                                     int x0, int y0, // Offsets for image
+                                     pmReadout *out1, pmReadout *out2, // Output readouts
+                                     psImage *convMask, // Output convolved mask
+                                     const pmReadout *ro1, const pmReadout *ro2, // Input readouts
+                                     psImage *subMask, // Input subtraction mask
+                                     psMaskType maskBad, // Mask value to give bad pixels
+                                     psMaskType maskPoor, // Mask value to give poor pixels
+                                     float poorFrac, // Fraction for "poor"
+                                     const psRegion *region, // Patch to convolve
+                                     const pmSubtractionKernels *kernels, // Kernels
+                                     bool doBG, // Add in background when convolving?
+                                     bool useFFT // Use FFT to do the convolution?
+    )
+{
+    int size = kernels->size;           // Half-size of kernel
+    int xMin = region->x0, xMax = region->x1, yMin = region->y0, yMax = region->y1; // Bounds of patch
+
+    psKernel *kernelImage = NULL;       // Kernel for the images
+    psKernel *kernelWeight = NULL;      // Kernel for the weight maps
+
+    // Only generate polynomial values every kernel footprint, since we have already assumed
+    // (with the stamps) that it does not vary rapidly on this scale.
+    psImage *polyValues = p_pmSubtractionPolynomialFromCoords(NULL, kernels, numCols, numRows,
+                                                              xMin + x0 + size + 1,
+                                                              yMin + y0 + size + 1);
+    float background = doBG ? p_pmSubtractionSolutionBackground(kernels, polyValues) : 0.0; // Background term
+
+    if (kernels->mode == PM_SUBTRACTION_MODE_1 || kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+        convolveRegion(out1->image, out1->weight, convMask, &kernelImage, &kernelWeight,
+                       ro1->image, ro1->weight, subMask, kernels, polyValues, background, *region,
+                       maskBad, maskPoor, poorFrac, useFFT, false);
+    }
+    if (kernels->mode == PM_SUBTRACTION_MODE_2 || kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+        convolveRegion(out2->image, out2->weight, convMask, &kernelImage, &kernelWeight,
+                       ro2->image, ro2->weight, subMask, kernels, polyValues, background, *region,
+                       maskBad, maskPoor, poorFrac, useFFT, kernels->mode == PM_SUBTRACTION_MODE_DUAL);
+    }
+
+    psFree(kernelImage);
+    psFree(kernelWeight);
+    psFree(polyValues);
+
+    if ((kernels->mode == PM_SUBTRACTION_MODE_1 || kernels->mode == PM_SUBTRACTION_MODE_DUAL) && ro1->mask) {
+        psMaskType **target = convMask->data.PS_TYPE_MASK_DATA; // Target mask
+        psMaskType **source = ro1->mask->data.PS_TYPE_MASK_DATA; // Source mask
+
+        for (int y = yMin; y < yMax; y++) {
+            for (int x = xMin; x < xMax; x++) {
+                target[y][x] |= source[y][x];
+            }
+        }
+    }
+    if ((kernels->mode == PM_SUBTRACTION_MODE_2 || kernels->mode == PM_SUBTRACTION_MODE_DUAL) && ro2->mask) {
+        psMaskType **target = convMask->data.PS_TYPE_MASK_DATA; // Target mask
+        psMaskType **source = ro2->mask->data.PS_TYPE_MASK_DATA; // Source mask
+
+        for (int y = yMin; y < yMax; y++) {
+            for (int x = xMin; x < xMax; x++) {
+                target[y][x] |= source[y][x];
+            }
+        }
+    }
+
+    return true;
+}
+
+bool pmSubtractionConvolveThread(psThreadJob *job)
+{
+    PS_ASSERT_THREAD_JOB_NON_NULL(job, false);
+
+    psArray *args = job->args;          // Arguments
+    int numCols = PS_SCALAR_VALUE(args->data[0], S32); // Number of columns
+    int numRows = PS_SCALAR_VALUE(args->data[1], S32); // Number of rows
+    int x0 = PS_SCALAR_VALUE(args->data[2], S32); // Offset in x
+    int y0 = PS_SCALAR_VALUE(args->data[3], S32); // Offset in x
+    pmReadout *out1 = args->data[4];    // Output readout 1
+    pmReadout *out2 = args->data[5];    // Output readout 2
+    psImage *convMask = args->data[6];  // Output convolved mask
+    const pmReadout *ro1 = args->data[7]; // Input readout 1
+    const pmReadout *ro2 = args->data[8]; // Input readout 2
+    psImage *subMask = args->data[9]; // Subtraction mask
+    psMaskType maskBad = PS_SCALAR_VALUE(args->data[10], U8); // Output mask value for bad pixels
+    psMaskType maskPoor = PS_SCALAR_VALUE(args->data[11], U8); // Output mask value for poor pixels
+    float poorFrac = PS_SCALAR_VALUE(args->data[12], F32); // Fraction for "poor"
+    const psRegion *region = args->data[13]; // Region to convolve
+    const pmSubtractionKernels *kernels = args->data[14]; // Kernels
+    bool doBG = PS_SCALAR_VALUE(args->data[15], U8); // Do background subtraction?
+    bool useFFT = PS_SCALAR_VALUE(args->data[16], U8); // Use FFT for convolution?
+
+    return subtractionConvolvePatch(numCols, numRows, x0, y0, out1, out2, convMask, ro1, ro2, subMask,
+                                    maskBad, maskPoor, poorFrac, region, kernels, doBG, useFFT);
+}
+
+bool pmSubtractionConvolve(pmReadout *out1, pmReadout *out2, const pmReadout *ro1, const pmReadout *ro2,
+                           psImage *subMask, psMaskType maskBad, psMaskType maskPoor, float poorFrac,
+                           const psRegion *region, const pmSubtractionKernels *kernels,
+                           bool doBG, bool useFFT)
+{
+    int numCols = 0, numRows = 0;       // Image dimensions
+    int x0 = 0, y0 = 0;                 // Image offset
+    if (kernels->mode == PM_SUBTRACTION_MODE_1 || kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+        PM_ASSERT_READOUT_NON_NULL(out1, false);
+        PM_ASSERT_READOUT_NON_NULL(ro1, false);
+        PM_ASSERT_READOUT_IMAGE(ro1, false);
+        numCols = ro1->image->numCols;
+        numRows = ro1->image->numRows;
+        x0 = ro1->col0;
+        y0 = ro1->row0;
+    }
+    if (kernels->mode == PM_SUBTRACTION_MODE_2 || kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+        PM_ASSERT_READOUT_NON_NULL(out2, false);
+        PM_ASSERT_READOUT_NON_NULL(ro2, false);
+        PM_ASSERT_READOUT_IMAGE(ro2, false);
+        if (numCols == 0 && numRows == 0) {
+            numCols = ro2->image->numCols;
+            numRows = ro2->image->numRows;
+            x0 = ro2->col0;
+            y0 = ro2->row0;
+        }
+    }
+    if (kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+        PS_ASSERT_IMAGES_SIZE_EQUAL(ro1->image, ro2->image, false);
+    }
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, false);
+    PM_ASSERT_SUBTRACTION_KERNELS_SOLUTION(kernels, false);
+    if (subMask) {
+        PS_ASSERT_IMAGE_NON_NULL(subMask, false);
+        PS_ASSERT_IMAGE_TYPE(subMask, PS_TYPE_MASK, false);
+        PS_ASSERT_IMAGE_SIZE(subMask, numCols, numRows, false);
+    }
+    if (region && psRegionIsNaN(*region)) {
+        psString string = psRegionToString(*region);
+        psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Input region (%s) contains NAN values", string);
+        psFree(string);
+        return false;
+    }
+
+    // Inputs
+    psImage *image1 = NULL, *weight1 = NULL; // Image and weight map from input 1
+    if (ro1) {
+        image1 = ro1->image;
+        weight1 = ro1->weight;
+    }
+    psImage *image2 = NULL, *weight2 = NULL; // Image and weight map from input 2
+    if (ro2) {
+        image2 = ro2->image;
+        weight2 = ro2->weight;
+    }
+
+    bool threaded = pmSubtractionThreaded(); // Running threaded?
+
+    // Outputs
+    psImage *convImage1 = NULL, *convWeight1 = NULL; // Convolved image and weight from input 1
+    if (kernels->mode == PM_SUBTRACTION_MODE_1 || kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+        convImage1 = out1->image;
+        if (!convImage1) {
+            convImage1 = out1->image = psImageAlloc(numCols, numRows, PS_TYPE_F32);
+            if (threaded) {
+                psMutexInit(convImage1);
+            }
+        }
+        if (weight1) {
+            if (!out1->weight) {
+                out1->weight = psImageAlloc(numCols, numRows, PS_TYPE_F32);
+                if (threaded) {
+                    psMutexInit(out1->weight);
+                }
+            }
+            convWeight1 = out1->weight;
+            psImageInit(convWeight1, 0.0);
+        }
+    }
+    psImage *convImage2 = NULL, *convWeight2 = NULL; // Convolved image and weight from input 2
+    if (kernels->mode == PM_SUBTRACTION_MODE_2 || kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+        convImage2 = out2->image;
+        if (!convImage2) {
+            convImage2 = out2->image = psImageAlloc(numCols, numRows, PS_TYPE_F32);
+            if (threaded) {
+                psMutexInit(convImage2);
+            }
+        }
+        if (weight2) {
+            if (!out2->weight) {
+                out2->weight = psImageAlloc(numCols, numRows, PS_TYPE_F32);
+                if (threaded) {
+                    psMutexInit(out2->weight);
+                }
+            }
+            convWeight2 = out2->weight;
+            psImageInit(convWeight2, 0.0);
+        }
+    }
+    psImage *convMask = NULL;           // Convolved mask image (common to inputs 1 and 2)
+    if (subMask) {
+        if (threaded) {
+            psMutexInit(subMask);
+        }
+        if (kernels->mode == PM_SUBTRACTION_MODE_1 || kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+            if (!out1->mask) {
+                out1->mask = psImageAlloc(numCols, numRows, PS_TYPE_MASK);
+            }
+            convMask = out1->mask;
+        }
+        if (kernels->mode == PM_SUBTRACTION_MODE_2 || kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+            if (convMask) {
+                if (out2->mask) {
+                    psFree(out2->mask);
+                }
+                out2->mask = psMemIncrRefCounter(convMask);
+            } else {
+                if (!out2->mask) {
+                    out2->mask = psImageAlloc(numCols, numRows, PS_TYPE_MASK);
+                }
+                convMask = out2->mask;
+            }
+        }
+        psImageInit(convMask, 0);
+    }
+
+
+    int size = kernels->size;           // Half-size of kernel
+    int fullSize = 2 * size + 1;        // Full size of kernel
+
+    // Get region for convolution: [xMin:xMax,yMin:yMax]
+    int xMin = size, xMax = numCols - size;
+    int yMin = size, yMax = numRows - size;
+    if (region) {
+        xMin = PS_MAX(region->x0, xMin);
+        xMax = PS_MIN(region->x1, xMax);
+        yMin = PS_MAX(region->y0, yMin);
+        yMax = PS_MIN(region->y1, yMax);
+    }
+
+#if 0
+    // XXX Use thread-specific data to store these
+    psImage *polyValues = NULL;         // Pre-calculated polynomial values
+    psKernel *kernelImage = NULL;       // Kernel for the images
+    psKernel *kernelWeight = NULL;      // Kernel for the weight maps
+#endif
+
+    for (int j = yMin; j < yMax; j += fullSize) {
+        int ySubMax = PS_MIN(j + fullSize, yMax); // Range for subregion of interest
+        for (int i = xMin; i < xMax; i += fullSize) {
+            int xSubMax = PS_MIN(i + fullSize, xMax); // Range for subregion of interest
+
+            psRegion *subRegion = psRegionAlloc(i, xSubMax, j, ySubMax); // Bounds of subtraction
+            if (threaded) {
+                psThreadJob *job = psThreadJobAlloc("PSMODULES_SUBTRACTION_CONVOLVE");
+                psArray *args = job->args;
+                PS_ARRAY_ADD_SCALAR(args, numCols, PS_TYPE_S32);
+                PS_ARRAY_ADD_SCALAR(args, numRows, PS_TYPE_S32);
+                PS_ARRAY_ADD_SCALAR(args, x0, PS_TYPE_S32);
+                PS_ARRAY_ADD_SCALAR(args, y0, PS_TYPE_S32);
+                psArrayAdd(args, 1, out1);
+                psArrayAdd(args, 1, out2);
+                psArrayAdd(args, 1, convMask);
+                psArrayAdd(args, 1, (pmReadout*)ro1); // Casting away const
+                psArrayAdd(args, 1, (pmReadout*)ro2); // Casting away const
+                // Since adding to the array can impact the reference count, we need to lock
+                if (subMask) {
+                    psMutexLock(subMask);
+                }
+                psArrayAdd(args, 1, subMask);
+                if (subMask) {
+                    psMutexUnlock(subMask);
+                }
+                PS_ARRAY_ADD_SCALAR(args, maskBad, PS_TYPE_U8);
+                PS_ARRAY_ADD_SCALAR(args, maskPoor, PS_TYPE_U8);
+                PS_ARRAY_ADD_SCALAR(args, poorFrac, PS_TYPE_F32);
+                psArrayAdd(args, 1, subRegion);
+                psArrayAdd(args, 1, (pmSubtractionKernels*)kernels); // Casting away const
+                PS_ARRAY_ADD_SCALAR(args, doBG, PS_TYPE_U8);
+                PS_ARRAY_ADD_SCALAR(args, useFFT, PS_TYPE_U8);
+
+                if (!psThreadJobAddPending(job)) {
+                    psFree(job);
+                    return false;
+                }
+                psFree(job);
+            } else {
+                subtractionConvolvePatch(numCols, numRows, x0, y0, out1, out2, convMask, ro1, ro2, subMask,
+                                         maskBad, maskPoor, poorFrac, subRegion, kernels, doBG, useFFT);
+            }
+            psFree(subRegion);
+        }
+    }
+
+    if (!psThreadPoolWait(false)) {
+        psError(PS_ERR_UNKNOWN, false, "Error waiting for threads.");
+        return false;
+    }
+
+    // We don't rely on psThreadPoolWait to harvest the jobs because the job contains a reference to the
+    // subMask, which is being changed on a thread, and psThreadPoolWait doesn't know that it needs to be
+    // locked before freeing.  After psThreadPoolWait, however, the jobs are completed, the threads are idle,
+    // and so there's no need to lock the subMask when we're blowing away the jobs.
+    if (threaded) {
+        psThreadJob *job;               // Completed job
+        while ((job = psThreadJobGetDone())) {
+            psFree(job);
+        }
+
+        if (subMask) {
+            psMutexDestroy(subMask);
+        }
+    }
+
+    // Copy anything that wasn't convolved
+    switch (kernels->mode) {
+      case PM_SUBTRACTION_MODE_1:
+        if (out2) {
+            out2->image = psMemIncrRefCounter(ro2->image);
+            out2->weight = psMemIncrRefCounter(ro2->weight);
+            out2->mask = psMemIncrRefCounter(ro2->mask);
+        }
+        break;
+      case PM_SUBTRACTION_MODE_2:
+        if (out1) {
+            out1->image = psMemIncrRefCounter(ro1->image);
+            out1->weight = psMemIncrRefCounter(ro1->weight);
+            out1->mask = psMemIncrRefCounter(ro1->mask);
+        }
+        break;
+      case PM_SUBTRACTION_MODE_DUAL:
+        break;
+      default:
+        psAbort("Should never get here.");
+    }
+
+    return true;
+}
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtraction.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtraction.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtraction.h	(revision 19481)
@@ -0,0 +1,146 @@
+/* @file pmSubtraction.h
+ *
+ * PSF-matched image subtraction, based on the Alard & Lupton (1998) and Alard (2000) methods.
+ *
+ * @author Paul Price, IfA
+ * @author GLG, MHPCC
+ *
+ * @version $Revision: 1.31 $ $Name: not supported by cvs2svn $
+ * @date $Date: 2008-08-30 02:28:07 $
+ * Copyright 2004-207 Institute for Astronomy, University of Hawaii
+ */
+
+#ifndef PM_SUBTRACTION_H
+#define PM_SUBTRACTION_H
+
+#include <pslib.h>
+
+#include <pmHDU.h>
+#include <pmFPA.h>
+#include <pmSubtractionKernels.h>
+#include <pmSubtractionStamps.h>
+
+/// @addtogroup imcombine Image Combinations
+/// @{
+
+/// Mask values for the subtraction mask
+typedef enum {
+    PM_SUBTRACTION_MASK_CLEAR          = 0x00, // No masking
+    PM_SUBTRACTION_MASK_BAD_1          = 0x01, // Image 1 is bad
+    PM_SUBTRACTION_MASK_BAD_2          = 0x02, // Image 2 is bad
+    PM_SUBTRACTION_MASK_CONVOLVE_1     = 0x04, // If image 1 is convolved, would be poor or bad
+    PM_SUBTRACTION_MASK_CONVOLVE_2     = 0x08, // If image 2 is convolved, would be poor or bad
+    PM_SUBTRACTION_MASK_CONVOLVE_BAD_1 = 0x10, // If image 1 is convolved, would be bad
+    PM_SUBTRACTION_MASK_CONVOLVE_BAD_2 = 0x20, // If image 2 is convolved, would be bad
+    PM_SUBTRACTION_MASK_BORDER         = 0x40, // Image border
+    PM_SUBTRACTION_MASK_REJ            = 0x80, // Previously tried as a stamp, and rejected
+} pmSubtractionMasks;
+
+
+/// Number of terms in a polynomial
+#define PM_SUBTRACTION_POLYTERMS(ORDER) (((ORDER) + 1) * ((ORDER) + 2) / 2)
+
+/// Set the indices for the normalisation and background terms
+#define PM_SUBTRACTION_INDICES(NORM,BG,KERNELS) { \
+    int numSpatial = PM_SUBTRACTION_POLYTERMS((KERNELS)->spatialOrder); /* Number of spatial terms */ \
+    NORM = (KERNELS)->num * numSpatial; \
+    BG = NORM + 1; \
+}
+
+/// Return the index for the start of the normalisation terms
+#define PM_SUBTRACTION_INDEX_NORM(KERNELS) \
+    ((KERNELS)->num * PM_SUBTRACTION_POLYTERMS((KERNELS)->spatialOrder))
+
+/// Return the index for the start of the background terms
+#define PM_SUBTRACTION_INDEX_BG(KERNELS) \
+    (((KERNELS)->num * PM_SUBTRACTION_POLYTERMS((KERNELS)->spatialOrder)) + 1)
+
+
+/// Convolve the reference stamp with the kernel components
+bool pmSubtractionConvolveStamp(pmSubtractionStamp *stamp, ///< Stamp to convolve
+                                const pmSubtractionKernels *kernels, ///< Kernel parameters
+                                int footprint ///< Half-size of region over which to calculate equation
+    );
+
+/// Reject stamps
+int pmSubtractionRejectStamps(float *rms, ///< RMS deviation, to return
+                              int *num, ///< Number of good stamps, to return
+                              pmSubtractionStampList *stamps, ///< Stamps
+                              const psVector *deviations, ///< Deviations for each stamp
+                              psImage *subMask, ///< Subtraction mask
+                              float sigmaRej, ///< Number of RMS deviations above zero at which to reject
+                              int footprint ///< Half-size of stamp
+    );
+
+/// Generate an image of the convolution kernel
+psImage *pmSubtractionKernelImage(const pmSubtractionKernels *kernels, ///< Kernel parameters
+                                  float x, float y,///< Normalised position [-1,1] for which to generate image
+                                  bool wantDual ///< Calculate for the dual kernel?
+                                  );
+
+/// Return the variance factor for a kernel
+///
+/// The variance factor allows conversion from the large-scale variance (which is what is calculated by
+/// pmSubtractionConvolve) and the small-scale (pixel-to-pixel) variance.
+float pmSubtractionVarianceFactor(const pmSubtractionKernels *kernels, ///< Kernel parameters
+                                  float x, float y, ///< Normalised position [-1,1]
+                                  bool wantDual ///< Calculate for the dual kernel?
+    );
+
+/// Generate images of the convolution kernel elements
+psArray *pmSubtractionKernelSolutions(const psVector *solution, ///< Solution vector
+                                      const pmSubtractionKernels *kernels, ///< Kernel parameters
+                                      float x, float y ///< Normalised position [-1,1] for images
+    );
+
+/// Execute a thread job to convolve a patch of the image
+bool pmSubtractionConvolveThread(psThreadJob *job ///< Job to execute
+    );
+
+/// Convolve image in preparation for subtraction
+bool pmSubtractionConvolve(pmReadout *out1, ///< Output image 1
+                           pmReadout *out2, ///< Output image 2 (DUAL mode only)
+                           const pmReadout *ro1, // Input image 1
+                           const pmReadout *ro2, // Input image 2
+                           psImage *subMask, ///< Subtraction mask (or NULL)
+                           psMaskType maskBad, ///< Mask value to give bad pixels
+                           psMaskType maskPoor, ///< Mask value to give poor pixels
+                           float poorFrac, ///< Fraction for "poor"
+                           const psRegion *region, ///< Region to convolve (or NULL)
+                           const pmSubtractionKernels *kernels, ///< Kernel parameters
+                           bool doBG,   ///< Apply background term?
+                           bool useFFT  ///< Use Fast Fourier Transform for the convolution?
+    );
+
+/// Generate the convolution of an image, given a precalculated kernel
+///
+/// The 'image' is a kernel for convenience --- intended to be a stamp
+psKernel *p_pmSubtractionConvolveStampPrecalc(const psKernel *image, ///< Image to convolve
+                                              const psKernel *kernel ///< Kernel by which to convolve
+    );
+
+/// Given (normalised) coordinates (x,y), generate a matrix where the elements (i,j) are x^i * y^j
+psImage *p_pmSubtractionPolynomial(psImage *output, ///< Output matrix, or NULL
+                                   int spatialOrder, ///< Maximum spatial polynomial order
+                                   float x, float y ///< Normalised position of interest, [-1,1]
+    );
+
+/// Given pixel coordinates (x,y), generate a matrix where the elements (i,j) are x^i * y^j
+///
+/// Same as p_pmSubtractionPolynomial except that the normalisation is applied
+psImage *p_pmSubtractionPolynomialFromCoords(psImage *output, ///< Output matrix, or NULL
+                                             const pmSubtractionKernels *kernels, ///< Kernel parameters
+                                             int numCols, int numRows, ///< Size of image of interest
+                                             int x, int y ///< Position of interest
+    );
+
+/// Return the radius from the centre of the convolution kernel that distinguishes "bad" and "poor" pixels
+int p_pmSubtractionBadRadius(psKernel *preKernel, ///< Pre-calculated convolution kernel
+                             const pmSubtractionKernels *kernels, ///< Kernel parameters
+                             const psImage *polyValues, ///< Polynomial values
+                             bool wantDual, ///< Calculate for the dual kernel?
+                             float poorFrac ///< Fraction for "poor"
+    );
+
+/// @}
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionEquation.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionEquation.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionEquation.c	(revision 19481)
@@ -0,0 +1,1138 @@
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <stdio.h>
+#include <string.h>
+#include <pslib.h>
+
+#include "pmSubtraction.h"
+#include "pmSubtractionKernels.h"
+#include "pmSubtractionStamps.h"
+#include "pmSubtractionThreads.h"
+
+#include "pmSubtractionEquation.h"
+
+
+//#define TESTING
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Private (file-static) functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+// Calculate the sum over a stamp product
+static inline double calculateSumProduct(const psKernel *image1, // First image in multiplication
+                                         const psKernel *image2, // Second image in multiplication
+                                         const psKernel *weight, // Weight image
+                                         int footprint // (Half-)Size of stamp
+    )
+{
+    double sum = 0.0;                   // Sum of the image products
+    for (int y = - footprint; y <= footprint; y++) {
+        for (int x = - footprint; x <= footprint; x++) {
+            sum += image1->kernel[y][x] * image2->kernel[y][x] / 1.0; // weight->kernel[y][x];
+        }
+    }
+    return sum;
+}
+
+// Calculate a single element of the least-squares matrix, with the polynomial expansions in one direction
+static inline bool calculateMatrixElement1(psImage *matrix, // Matrix to calculate
+                                           int i, int j, // Coordinates of element
+                                           const psKernel *image1, // First image in multiplication
+                                           const psKernel *image2, // Second image in multiplication
+                                           const psKernel *weight, // Weight image
+                                           const psImage *polyValues, // Spatial polynomial values
+                                           int numKernels, // Number of kernel basis functions
+                                           int footprint, // (Half-)Size of stamp
+                                           int spatialOrder, // Maximum order of spatial variation
+                                           bool symmetric // Is the matrix symmetric?
+    )
+{
+    double sum = calculateSumProduct(image1, image2, weight, footprint); // Sum of the image products
+    if (!isfinite(sum)) {
+        return false;
+    }
+
+    // Generate the pseudo-convolutions from the spatial polynomial terms
+    for (int iyOrder = 0, iIndex = i; iyOrder <= spatialOrder; iyOrder++) {
+        for (int ixOrder = 0; ixOrder <= spatialOrder - iyOrder; ixOrder++, iIndex += numKernels) {
+            double convPoly = sum * polyValues->data.F64[iyOrder][ixOrder];
+
+            assert(iIndex < matrix->numRows && j < matrix->numCols);
+
+            matrix->data.F64[iIndex][j] = convPoly;
+            if (symmetric) {
+
+                assert(iIndex < matrix->numCols && j < matrix->numRows);
+
+                matrix->data.F64[j][iIndex] = convPoly;
+            }
+        }
+    }
+    return true;
+}
+
+// Calculate a single element of the least-squares matrix, with the polynomial expansions in both directions
+static inline bool calculateMatrixElement2(psImage *matrix, // Matrix to calculate
+                                           int i, int j, // Coordinates of element
+                                           const psKernel *image1, // First image in multiplication
+                                           const psKernel *image2, // Second image in multiplication
+                                           const psKernel *weight, // Weight image
+                                           const psImage *polyValues, // Spatial polynomial values
+                                           int numKernels, // Number of kernel basis functions
+                                           int footprint, // (Half-)Size of stamp
+                                           int spatialOrder, // Maximum order of spatial variation
+                                           bool symmetric // Is the matrix symmetric?
+    )
+{
+    double sum = calculateSumProduct(image1, image2, weight, footprint); // Sum of the image products
+    if (!isfinite(sum)) {
+        return false;
+    }
+
+    // Generate the pseudo-convolutions from the spatial polynomial terms
+    for (int iyOrder = 0, iIndex = i; iyOrder <= spatialOrder; iyOrder++) {
+        for (int ixOrder = 0; ixOrder <= spatialOrder - iyOrder; ixOrder++, iIndex += numKernels) {
+            double iPoly = polyValues->data.F64[iyOrder][ixOrder]; // Value of polynomial
+            for (int jyOrder = 0, jIndex = j; jyOrder <= spatialOrder; jyOrder++) {
+                for (int jxOrder = 0; jxOrder <= spatialOrder - jyOrder; jxOrder++, jIndex += numKernels) {
+                    double convPoly = sum * iPoly * polyValues->data.F64[jyOrder][jxOrder];
+
+                    assert(iIndex < matrix->numRows && jIndex < matrix->numCols);
+
+                    matrix->data.F64[iIndex][jIndex] = convPoly;
+                    if (symmetric) {
+
+                        assert(iIndex < matrix->numCols && jIndex < matrix->numRows);
+
+                        matrix->data.F64[jIndex][iIndex] = convPoly;
+                    }
+                }
+            }
+        }
+    }
+    return true;
+}
+
+// Calculate the square part of the matrix derived from multiplying convolutions
+static bool calculateMatrixSquare(psImage *matrix, // Matrix to calculate
+                                  const psArray *convolutions1, // Convolutions for element 1
+                                  const psArray *convolutions2, // Convolutions for element 2
+                                  const psKernel *weight, // Weight image
+                                  const psImage *polyValues, // Polynomial values
+                                  int numKernels, // Number of kernel basis functions
+                                  int spatialOrder, // Order of spatial variation
+                                  int footprint // Half-size of stamp
+                                  )
+{
+    bool symmetric = (convolutions1 == convolutions2 ? true : false); // Is matrix symmetric?
+
+    for (int i = 0; i < numKernels; i++) {
+        psKernel *iConv = convolutions1->data[i]; // Convolution for i-th element
+
+        for (int j = (symmetric ? i : 0); j < numKernels; j++) {
+            psKernel *jConv = convolutions2->data[j]; // Convolution for j-th element
+
+            if (!calculateMatrixElement2(matrix, i, j, iConv, jConv, weight, polyValues, numKernels,
+                                         footprint, spatialOrder, symmetric)) {
+                psTrace("psModules.imcombine", 2, "Bad sumCC at %d, %d", i, j);
+                return false;
+            }
+        }
+    }
+
+    return true;
+}
+
+// Calculate least-squares matrix and vector
+static bool calculateMatrix(psImage *matrix, // Matrix to calculate
+                            const pmSubtractionKernels *kernels, // Kernel components
+                            const psArray *convolutions, // Convolutions of source with kernels
+                            const psKernel *input, // Input stamp, or NULL
+                            const psKernel *weight, // Weight stamp
+                            const psImage *polyValues, // Spatial polynomial values
+                            int footprint, // (Half-)Size of stamp
+                            bool normAndBG // Calculate normalisation and background terms?
+    )
+{
+    int numKernels = kernels->num;      // Number of kernel components
+    int spatialOrder = kernels->spatialOrder; // Maximum order of spatial variation
+    int numSpatial = PM_SUBTRACTION_POLYTERMS(spatialOrder); // Number of spatial variation terms
+    int bgOrder = kernels->bgOrder;     // Maximum order of background fit
+    int numBackground = normAndBG ? PM_SUBTRACTION_POLYTERMS(bgOrder) : 0; // Number of background terms
+    int numTerms = numKernels * numSpatial + (normAndBG ? 1 + numBackground : 0); // Total number of terms
+    assert(matrix);
+    assert(matrix->numCols == matrix->numRows);
+    assert(matrix->numCols == numTerms);
+    assert(convolutions && convolutions->n == numKernels);
+    assert(polyValues);
+    assert(!normAndBG || input);        // If we want the normalisation and BG, then we need the input image
+
+    // Square part of the matrix (convolution-convolution products)
+    if (!calculateMatrixSquare(matrix, convolutions, convolutions, weight, polyValues, numKernels,
+                               spatialOrder, footprint)) {
+        return false;
+    }
+
+    // XXX To support higher-order background model than simply constant, the below code needs to be updated.
+    if (normAndBG) {
+        int normIndex = PM_SUBTRACTION_INDEX_NORM(kernels); // Index for normalisation
+        int bgIndex = PM_SUBTRACTION_INDEX_BG(kernels); // Index in matrix for background
+
+        for (int i = 0; i < numKernels; i++) {
+            psKernel *conv = convolutions->data[i]; // Convolution for i-th element
+
+            // Normalisation-convolution terms
+            if (!calculateMatrixElement1(matrix, i, normIndex, conv, input, weight, polyValues, numKernels,
+                                         footprint, spatialOrder, true)) {
+                psTrace("psModules.imcombine", 2, "Bad sumIC at %d", i);
+                return false;
+            }
+
+            // Background-convolution terms
+            double sumC = 0.0;          // Sum of the convolution
+            for (int y = - footprint; y <= footprint; y++) {
+                for (int x = - footprint; x <= footprint; x++) {
+                    sumC += conv->kernel[y][x] / 1.0; // weight->kernel[y][x];
+                }
+            }
+            if (!isfinite(sumC)) {
+                psTrace("psModules.imcombine", 2, "Bad sumC at %d", i);
+                return false;
+            }
+
+            for (int yOrder = 0, index = i; yOrder <= spatialOrder; yOrder++) {
+                for (int xOrder = 0; xOrder <= spatialOrder - yOrder; xOrder++, index += numKernels) {
+                    double value = sumC * polyValues->data.F64[yOrder][xOrder];
+                    matrix->data.F64[index][bgIndex] = value;
+                    matrix->data.F64[bgIndex][index] = value;
+                }
+            }
+        }
+
+        // Background only, normalisation only, and background-normalisation terms
+        double sum1 = 0.0;              // Sum of the weighting
+        double sumI = 0.0;              // Sum of the input
+        double sumII = 0.0;             // Sum of the input squared
+        for (int y = - footprint; y <= footprint; y++) {
+            for (int x = - footprint; x <= footprint; x++) {
+                double invNoise2 = 1.0 / 1.0; // weight->kernel[y][x];
+                double value = input->kernel[y][x] * invNoise2;
+                sumI += value;
+                sumII += value * input->kernel[y][x];
+                sum1 += invNoise2;
+            }
+        }
+        if (!isfinite(sumI)) {
+            psTrace("psModules.imcombine", 2, "Bad sumI detected");
+            return false;
+        }
+        if (!isfinite(sumII)) {
+            psTrace("psModules.imcombine", 2, "Bad sumII detected");
+            return false;
+        }
+        if (!isfinite(sum1)) {
+            psTrace("psModules.imcombine", 2, "Bad sum1 detected");
+            return false;
+        }
+        matrix->data.F64[normIndex][normIndex] = sumII;
+        matrix->data.F64[bgIndex][bgIndex] = sum1;
+        matrix->data.F64[normIndex][bgIndex] = sumI;
+        matrix->data.F64[bgIndex][normIndex] = sumI;
+    }
+
+    return true;
+}
+
+
+// Calculate least-squares matrix and vector
+static bool calculateVector(psVector *vector, // Vector to calculate, or NULL
+                            const pmSubtractionKernels *kernels, // Kernel components
+                            const psArray *convolutions, // Convolutions of source with kernels
+                            const psKernel *input, // Input stamp, or NULL if !normAndBG
+                            const psKernel *target, // Target stamp
+                            const psKernel *weight, // Weight stamp
+                            const psImage *polyValues, // Spatial polynomial values
+                            int footprint, // (Half-)Size of stamp
+                            bool normAndBG // Calculate normalisation and background terms?
+    )
+{
+    int numKernels = kernels->num;      // Number of kernel components
+    int spatialOrder = kernels->spatialOrder; // Maximum order of spatial variation
+    int numSpatial = PM_SUBTRACTION_POLYTERMS(spatialOrder); // Number of spatial variation terms
+    int bgOrder = kernels->bgOrder;     // Maximum order of background fit
+    int numBackground = normAndBG ? PM_SUBTRACTION_POLYTERMS(bgOrder) : 0; // Number of background terms
+    int numTerms = numKernels * numSpatial + (normAndBG ? 1 + numBackground : 0); // Total number of terms
+    assert(vector && vector->n == numTerms);
+    assert(convolutions && convolutions->n == numKernels);
+    assert(target);
+    assert(polyValues);
+    assert(!normAndBG || input);       // If we want the normalisation and BG, then we need the input image
+
+    // Convolution terms
+    for (int i = 0; i < numKernels; i++) {
+        psKernel *conv = convolutions->data[i]; // Convolution for i-th element
+        double sumTC = 0.0;          // Sum of the target and convolution
+        for (int y = - footprint; y <= footprint; y++) {
+            for (int x = - footprint; x <= footprint; x++) {
+                sumTC += target->kernel[y][x] * conv->kernel[y][x] / 1.0; // weight->kernel[y][x];
+            }
+        }
+        if (!isfinite(sumTC)) {
+            psTrace("psModules.imcombine", 2, "Bad sumTC at %d", i);
+            return false;
+        }
+        for (int yOrder = 0, index = i; yOrder <= spatialOrder; yOrder++) {
+            for (int xOrder = 0; xOrder <= spatialOrder - yOrder; xOrder++, index += numKernels) {
+                vector->data.F64[index] = sumTC * polyValues->data.F64[yOrder][xOrder];
+            }
+        }
+    }
+
+    if (normAndBG) {
+        // Background terms
+        double sumT = 0.0;              // Sum of the target
+        double sumIT = 0.0;             // Sum of the input-target product
+        for (int y = - footprint; y <= footprint; y++) {
+            for (int x = - footprint; x <= footprint; x++) {
+                float value = target->kernel[y][x] / 1.0; // weight->kernel[y][x];
+                sumIT += value * input->kernel[y][x];
+                sumT += value;
+            }
+        }
+        if (!isfinite(sumT)) {
+            psTrace("psModules.imcombine", 2, "Bad sumI detected");
+            return false;
+        }
+        if (!isfinite(sumIT)) {
+            psTrace("psModules.imcombine", 2, "Bad sumIT detected");
+            return false;
+        }
+
+        int normIndex = PM_SUBTRACTION_INDEX_NORM(kernels); // Index for normalisation term
+        vector->data.F64[normIndex] = sumIT;
+        int bgIndex = PM_SUBTRACTION_INDEX_BG(kernels); // Index for background term
+        vector->data.F64[bgIndex] = sumT;
+    }
+
+    return true;
+}
+
+
+
+// Calculate the cross-matrix, composed of convolutions of each image
+// Note that the cross-matrix is NOT square
+static bool calculateMatrixCross(psImage *matrix, // Matrix to calculate
+                                 const pmSubtractionKernels *kernels, // Kernel components
+                                 const psArray *convolutions1, // Convolutions of image 1
+                                 const psArray *convolutions2, // Convolutions of image 2
+                                 const psKernel *image1, // Image 1 stamp
+                                 const psKernel *weight, // Weight stamp
+                                 const psImage *polyValues, // Spatial polynomial values
+                                 int footprint // (Half-)Size of stamp
+                                 )
+{
+    assert(matrix);
+    int numKernels = kernels->num;      // Number of kernel components
+    int spatialOrder = kernels->spatialOrder; // Maximum order of spatial variation
+    int numSpatial = PM_SUBTRACTION_POLYTERMS(spatialOrder); // Number of spatial polynomial terms
+    int numBackground = PM_SUBTRACTION_POLYTERMS(kernels->bgOrder); // Number of background terms
+    int numCols = numKernels * numSpatial + 1 + numBackground; // Number of columns
+    int numRows = numKernels * numSpatial; // Number of rows
+    assert(matrix->numCols == numCols && matrix->numRows == numRows);
+    assert(convolutions1 && convolutions1->n == numKernels);
+    assert(convolutions2 && convolutions2->n == numKernels);
+
+    int normIndex, bgIndex;             // Indices in matrix for normalisation and background terms
+    PM_SUBTRACTION_INDICES(normIndex, bgIndex, kernels);
+
+    if (!calculateMatrixSquare(matrix, convolutions1, convolutions2, weight, polyValues, numKernels,
+                               spatialOrder, footprint)) {
+        return false;
+    }
+
+    for (int i = 0; i < numKernels; i++) {
+        // Normalisation
+        psKernel *conv = convolutions2->data[i]; // Convolution
+        if (!calculateMatrixElement1(matrix, i, normIndex, conv, image1, weight, polyValues, numKernels,
+                                     footprint, spatialOrder, false)) {
+            psTrace("psModules.imcombine", 2, "Bad sumIC at %d", i);
+            return false;
+        }
+
+        // Background
+        double sumC = 0.0;              // Sum of the weighting
+        for (int y = - footprint; y <= footprint; y++) {
+            for (int x = - footprint; x <= footprint; x++) {
+                sumC += conv->kernel[y][x] / 1.0; // weight->kernel[y][x];
+            }
+        }
+        if (!isfinite(sumC)) {
+            psTrace("psModules.imcombine", 2, "Bad sumC detected at %d", i);
+            return false;
+        }
+        for (int yOrder = 0, index = i; yOrder <= spatialOrder; yOrder++) {
+            for (int xOrder = 0; xOrder <= spatialOrder - yOrder; xOrder++, index += numKernels) {
+                matrix->data.F64[index][bgIndex] = sumC * polyValues->data.F64[yOrder][xOrder];
+            }
+        }
+    }
+
+    return true;
+}
+
+
+// Add in penalty term to least-squares vector
+static bool calculatePenalty(psVector *vector, // Vector to which to add in penalty term
+                             const pmSubtractionKernels *kernels // Kernel parameters
+    )
+{
+    if (kernels->penalty == 0.0) {
+        return true;
+    }
+
+    psVector *penalties = kernels->penalties; // Penalties for each kernel component
+    int spatialOrder = kernels->spatialOrder; // Order of spatial variations
+    int numKernels = kernels->num; // Number of kernel components
+    for (int i = 0; i < numKernels; i++) {
+        for (int yOrder = 0, index = i; yOrder <= spatialOrder; yOrder++) {
+            for (int xOrder = 0; xOrder <= spatialOrder - yOrder; xOrder++, index += numKernels) {
+                vector->data.F64[index] -= penalties->data.F32[i];
+            }
+        }
+    }
+
+    return true;
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Semi-public functions
+// XXX We might like to define these functions as "extern inline" but gcc currently doesn't handle this in c99
+// mode.  See http://gcc.gnu.org/ml/gcc/2006-11/msg00006.html
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+// Calculate the value of a polynomial, specified by coefficients and polynomial values
+double p_pmSubtractionCalculatePolynomial(const psVector *coeff, // Coefficients
+                                                 const psImage *polyValues, // Polynomial values
+                                                 int order, // Order of polynomials
+                                                 int index, // Index at which to begin
+                                                 int step // Step between subsequent indices
+                                                 )
+{
+    double sum = 0.0;                   // Value of the polynomial sum
+    for (int yOrder = 0; yOrder <= order; yOrder++) {
+        for (int xOrder = 0; xOrder <= order - yOrder; xOrder++, index += step) {
+
+            assert(index < coeff->n);
+
+            sum += coeff->data.F64[index] * polyValues->data.F64[yOrder][xOrder];
+        }
+    }
+    return sum;
+}
+
+double p_pmSubtractionSolutionCoeff(const pmSubtractionKernels *kernels, const psImage *polyValues,
+                                           int index, bool wantDual)
+{
+#if 0
+    // This is probably in a tight loop, so don't check inputs
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, NAN);
+    PM_ASSERT_SUBTRACTION_KERNELS_SOLUTION(kernels, NAN);
+    PS_ASSERT_IMAGE_NON_NULL(polyValues, NAN);
+    PS_ASSERT_INT_POSITIVE(index, NAN);
+#endif
+
+    psVector *solution = wantDual ? kernels->solution2 : kernels->solution1; // Solution vector
+    return p_pmSubtractionCalculatePolynomial(solution, polyValues, kernels->spatialOrder, index,
+                                              kernels->num);
+}
+
+double p_pmSubtractionSolutionNorm(const pmSubtractionKernels *kernels)
+{
+#if 0
+    // This is probably in a tight loop, so don't check inputs
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, NAN);
+    PM_ASSERT_SUBTRACTION_KERNELS_SOLUTION(kernels, NAN);
+    PS_ASSERT_IMAGE_NON_NULL(polyValues, NAN);
+#endif
+
+    int normIndex = PM_SUBTRACTION_INDEX_NORM(kernels); // Index for normalisation
+    return kernels->solution1->data.F64[normIndex];
+}
+
+double p_pmSubtractionSolutionBackground(const pmSubtractionKernels *kernels,
+                                                const psImage *polyValues)
+{
+#if 0
+    // This is probably in a tight loop, so don't check inputs
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, NAN);
+    PM_ASSERT_SUBTRACTION_KERNELS_SOLUTION(kernels, NAN);
+    PS_ASSERT_IMAGE_NON_NULL(polyValues, NAN);
+#endif
+
+    int bgIndex = PM_SUBTRACTION_INDEX_BG(kernels); // Index for background
+    return p_pmSubtractionCalculatePolynomial(kernels->solution1, polyValues, kernels->bgOrder, bgIndex, 1);
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Public functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+bool pmSubtractionCalculateEquationThread(psThreadJob *job)
+{
+    PS_ASSERT_THREAD_JOB_NON_NULL(job, false);
+
+    pmSubtractionStampList *stamps = job->args->data[0]; // List of stamps
+    const pmSubtractionKernels *kernels = job->args->data[1]; // Kernels
+    int index = PS_SCALAR_VALUE(job->args->data[2], S32); // Stamp index
+
+    return pmSubtractionCalculateEquationStamp(stamps, kernels, index);
+}
+
+bool pmSubtractionCalculateEquationStamp(pmSubtractionStampList *stamps, const pmSubtractionKernels *kernels,
+                                         int index)
+{
+    PM_ASSERT_SUBTRACTION_STAMP_LIST_NON_NULL(stamps, false);
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, false);
+    PS_ASSERT_INT_NONNEGATIVE(index, false);
+    PS_ASSERT_INT_LESS_THAN(index, stamps->num, false);
+
+    int footprint = stamps->footprint;  // Half-size of stamps
+    int spatialOrder = kernels->spatialOrder; // Maximum order of spatial variation
+    int numKernels = kernels->num;      // Number of kernel basis functions
+    int numSpatial = PM_SUBTRACTION_POLYTERMS(spatialOrder); // Number of spatial variations
+    int numBackground = PM_SUBTRACTION_POLYTERMS(kernels->bgOrder); // Number of background terms
+
+    // Total number of parameters to solve for: coefficient of each kernel basis function, multipled by the
+    // number of coefficients for the spatial polynomial, normalisation and a constant background offset.
+    int numParams = numKernels * numSpatial + 1 + numBackground;
+
+    pmSubtractionStamp *stamp = stamps->stamps->data[index]; // Stamp of interest
+    psAssert(stamp->status == PM_SUBTRACTION_STAMP_CALCULATE, "We only operate on stamps with this state.");
+
+    // Generate convolutions
+    if (!pmSubtractionConvolveStamp(stamp, kernels, footprint)) {
+        psError(PS_ERR_UNKNOWN, false, "Unable to convolve stamp %d.", index);
+        return NULL;
+    }
+
+#ifdef TESTING
+    for (int j = 0; j < numKernels; j++) {
+        if (stamp->convolutions1) {
+            psString convName = NULL;
+            psStringAppend(&convName, "conv1_%03d_%03d.fits", index, j);
+            psFits *fits = psFitsOpen(convName, "w");
+            psFree(convName);
+            psKernel *conv = stamp->convolutions1->data[j];
+            psFitsWriteImage(fits, NULL, conv->image, 0, NULL);
+            psFitsClose(fits);
+        }
+
+        if (stamp->convolutions2) {
+            psString convName = NULL;
+            psStringAppend(&convName, "conv2_%03d_%03d.fits", index, j);
+            psFits *fits = psFitsOpen(convName, "w");
+            psFree(convName);
+            psKernel *conv = stamp->convolutions2->data[j];
+            psFitsWriteImage(fits, NULL, conv->image, 0, NULL);
+            psFitsClose(fits);
+        }
+    }
+#endif
+
+    psImage *polyValues = p_pmSubtractionPolynomial(NULL, spatialOrder,
+                                                    stamp->xNorm, stamp->yNorm); // Polynomial terms
+
+    bool new = stamp->vector1 ? false : true; // Is this a new run?
+    if (new) {
+        stamp->matrix1 = psImageAlloc(numParams, numParams, PS_TYPE_F64);
+        stamp->vector1 = psVectorAlloc(numParams, PS_TYPE_F64);
+    }
+#ifdef TESTING
+    psImageInit(stamp->matrix1, NAN);
+    psVectorInit(stamp->vector1, NAN);
+#endif
+
+    bool status;                    // Status of least-squares matrix/vector calculation
+    switch (kernels->mode) {
+      case PM_SUBTRACTION_MODE_1:
+        status = calculateMatrix(stamp->matrix1, kernels, stamp->convolutions1, stamp->image1,
+                                 stamp->weight, polyValues, footprint, true);
+        status &= calculateVector(stamp->vector1, kernels, stamp->convolutions1, stamp->image1,
+                                  stamp->image2, stamp->weight, polyValues, footprint, true);
+        break;
+      case PM_SUBTRACTION_MODE_2:
+        status = calculateMatrix(stamp->matrix1, kernels, stamp->convolutions2, stamp->image2,
+                                 stamp->weight, polyValues, footprint, true);
+        status &= calculateVector(stamp->vector1, kernels, stamp->convolutions2, stamp->image2,
+                                  stamp->image1, stamp->weight, polyValues, footprint, true);
+        break;
+      case PM_SUBTRACTION_MODE_DUAL:
+        if (new) {
+            stamp->matrix2 = psImageAlloc(numKernels * numSpatial, numKernels * numSpatial, PS_TYPE_F64);
+            stamp->matrixX = psImageAlloc(numParams, numKernels * numSpatial, PS_TYPE_F64);
+            stamp->vector2 = psVectorAlloc(numKernels * numSpatial, PS_TYPE_F64);
+        }
+#ifdef TESTING
+        psImageInit(stamp->matrix2, NAN);
+        psImageInit(stamp->matrixX, NAN);
+        psVectorInit(stamp->vector2, NAN);
+#endif
+        status  = calculateMatrix(stamp->matrix1, kernels, stamp->convolutions1, stamp->image1,
+                                  stamp->weight, polyValues, footprint, true);
+        status &= calculateMatrix(stamp->matrix2, kernels, stamp->convolutions2, NULL,
+                                  stamp->weight, polyValues, footprint, false);
+        status &= calculateMatrixCross(stamp->matrixX, kernels, stamp->convolutions1,
+                                       stamp->convolutions2, stamp->image1, stamp->weight, polyValues,
+                                       footprint);
+        status &= calculateVector(stamp->vector1, kernels, stamp->convolutions1, stamp->image1,
+                                  stamp->image2, stamp->weight, polyValues, footprint, true);
+        status &= calculateVector(stamp->vector2, kernels, stamp->convolutions2, NULL,
+                                  stamp->image2, stamp->weight, polyValues, footprint, false);
+        break;
+      default:
+        psAbort("Unsupported subtraction mode: %x", kernels->mode);
+    }
+
+    if (!status) {
+        stamp->status = PM_SUBTRACTION_STAMP_REJECTED;
+        psWarning("Rejecting stamp %d (%d,%d) because of bad equation",
+                  index, (int)(stamp->x + 0.5), (int)(stamp->y + 0.5));
+    } else {
+        stamp->status = PM_SUBTRACTION_STAMP_USED;
+    }
+
+#ifdef TESTING
+    if (psTraceGetLevel("psModules.imcombine.equation") >= 10) {
+        psString matrixName = NULL;
+        psStringAppend(&matrixName, "matrix1_%d.fits", index);
+        psFits *matrixFile = psFitsOpen(matrixName, "w");
+        psFree(matrixName);
+        psFitsWriteImage(matrixFile, NULL, stamp->matrix1, 0, NULL);
+        psFitsClose(matrixFile);
+
+        matrixName = NULL;
+        psStringAppend(&matrixName, "vector1_%d.fits", index);
+        psImage *dummy = psImageAlloc(stamp->vector1->n, 1, PS_TYPE_F64);
+        memcpy(dummy->data.F64[0], stamp->vector1->data.F64,
+               PSELEMTYPE_SIZEOF(PS_TYPE_F64) * stamp->vector1->n);
+        matrixFile = psFitsOpen(matrixName, "w");
+        psFree(matrixName);
+        psFitsWriteImage(matrixFile, NULL, dummy, 0, NULL);
+        psFree(dummy);
+        psFitsClose(matrixFile);
+
+        if (stamp->vector2) {
+            matrixName = NULL;
+            psStringAppend(&matrixName, "vector2_%d.fits", index);
+            dummy = psImageAlloc(stamp->vector2->n, 1, PS_TYPE_F64);
+            memcpy(dummy->data.F64[0], stamp->vector2->data.F64,
+                   PSELEMTYPE_SIZEOF(PS_TYPE_F64) * stamp->vector2->n);
+            matrixFile = psFitsOpen(matrixName, "w");
+            psFree(matrixName);
+            psFitsWriteImage(matrixFile, NULL, dummy, 0, NULL);
+            psFree(dummy);
+            psFitsClose(matrixFile);
+        }
+
+        if (kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+            matrixName = NULL;
+            psStringAppend(&matrixName, "matrix2_%d.fits", index);
+            matrixFile = psFitsOpen(matrixName, "w");
+            psFree(matrixName);
+            psFitsWriteImage(matrixFile, NULL, stamp->matrix2, 0, NULL);
+            psFitsClose(matrixFile);
+
+            matrixName = NULL;
+            psStringAppend(&matrixName, "matrixX_%d.fits", index);
+            matrixFile = psFitsOpen(matrixName, "w");
+            psFree(matrixName);
+            psFitsWriteImage(matrixFile, NULL, stamp->matrixX, 0, NULL);
+            psFitsClose(matrixFile);
+        }
+    }
+#endif
+
+    psFree(polyValues);
+
+    return true;
+}
+
+bool pmSubtractionCalculateEquation(pmSubtractionStampList *stamps, const pmSubtractionKernels *kernels)
+{
+    PM_ASSERT_SUBTRACTION_STAMP_LIST_NON_NULL(stamps, false);
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, false);
+
+    // We iterate over each stamp, allocate the matrix and vectors if
+    // necessary, and then calculate those matrix/vectors.
+    for (int i = 0; i < stamps->num; i++) {
+        pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+        if (stamp->status != PM_SUBTRACTION_STAMP_CALCULATE) {
+            continue;
+        }
+
+        if (pmSubtractionThreaded()) {
+            psThreadJob *job = psThreadJobAlloc("PSMODULES_SUBTRACTION_CALCULATE_EQUATION");
+            psArrayAdd(job->args, 1, stamps);
+            psArrayAdd(job->args, 1, (pmSubtractionKernels*)kernels); // Casting away const to put on array
+            PS_ARRAY_ADD_SCALAR(job->args, i, PS_TYPE_S32);
+            if (!psThreadJobAddPending(job)) {
+                psFree(job);
+                return false;
+            }
+            psFree(job);
+        } else {
+            pmSubtractionCalculateEquationStamp(stamps, kernels, i);
+        }
+    }
+
+    if (!psThreadPoolWait(true)) {
+        psError(PS_ERR_UNKNOWN, false, "Error waiting for threads.");
+        return false;
+    }
+
+    return true;
+}
+
+bool pmSubtractionSolveEquation(pmSubtractionKernels *kernels, const pmSubtractionStampList *stamps)
+{
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, false);
+    PM_ASSERT_SUBTRACTION_STAMP_LIST_NON_NULL(stamps, false);
+
+    // Check inputs
+    int numParams = -1;                // Number of parameters
+    int numParams2 = 0;                // Number of parameters for part solution (DUAL mode)
+    for (int i = 0; i < stamps->num; i++) {
+        pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+        PS_ASSERT_PTR_NON_NULL(stamp, false);
+        if (stamp->status != PM_SUBTRACTION_STAMP_USED) {
+            continue;
+        }
+
+        PS_ASSERT_VECTOR_NON_NULL(stamp->vector1, false);
+        if (numParams == -1) {
+            numParams = stamp->vector1->n;
+        }
+        PS_ASSERT_VECTOR_SIZE(stamp->vector1, (long)numParams, false);
+        PS_ASSERT_VECTOR_TYPE(stamp->vector1, PS_TYPE_F64, false);
+        PS_ASSERT_IMAGE_NON_NULL(stamp->matrix1, false);
+        PS_ASSERT_IMAGE_SIZE(stamp->matrix1, numParams, numParams, false);
+        PS_ASSERT_IMAGE_TYPE(stamp->matrix1, PS_TYPE_F64, false);
+
+        if (kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+            PS_ASSERT_IMAGE_NON_NULL(stamp->matrix2, false);
+            PS_ASSERT_IMAGE_NON_NULL(stamp->matrixX, false);
+            if (numParams2 == 0) {
+                numParams2 = stamp->matrix2->numCols;
+            }
+            PS_ASSERT_IMAGE_SIZE(stamp->matrix2, numParams2, numParams2, false);
+            PS_ASSERT_IMAGE_SIZE(stamp->matrixX, numParams, numParams2, false);
+            PS_ASSERT_IMAGE_TYPE(stamp->matrix2, PS_TYPE_F64, false);
+            PS_ASSERT_IMAGE_TYPE(stamp->matrixX, PS_TYPE_F64, false);
+            PS_ASSERT_VECTOR_NON_NULL(stamp->vector2, false);
+            PS_ASSERT_VECTOR_SIZE(stamp->vector2, (long)numParams2, false);
+            PS_ASSERT_VECTOR_TYPE(stamp->vector2, PS_TYPE_F64, false);
+        }
+    }
+    if (numParams == -1) {
+        psError(PS_ERR_BAD_PARAMETER_VALUE, true, "No suitable stamps found.");
+        return NULL;
+    }
+
+    if (kernels->mode != PM_SUBTRACTION_MODE_DUAL) {
+        // Accumulate the least-squares matricies and vectors
+        psImage *sumMatrix = psImageAlloc(numParams, numParams, PS_TYPE_F64); // Combined matrix
+        psVector *sumVector = psVectorAlloc(numParams, PS_TYPE_F64); // Combined vector
+        psVectorInit(sumVector, 0.0);
+        psImageInit(sumMatrix, 0.0);
+        int numStamps = 0;              // Number of good stamps
+        for (int i = 0; i < stamps->num; i++) {
+            pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+            if (stamp->status == PM_SUBTRACTION_STAMP_USED) {
+                (void)psBinaryOp(sumMatrix, sumMatrix, "+", stamp->matrix1);
+                (void)psBinaryOp(sumVector, sumVector, "+", stamp->vector1);
+                numStamps++;
+            }
+        }
+        calculatePenalty(sumVector, kernels);
+
+        psVector *permutation = NULL;       // Permutation vector, required for LU decomposition
+        psImage *luMatrix = psMatrixLUD(NULL, &permutation, sumMatrix);
+        psFree(sumMatrix);
+        if (!luMatrix) {
+            psError(PS_ERR_UNKNOWN, true, "LU Decomposition of least-squares matrix failed.\n");
+            psFree(sumVector);
+            psFree(luMatrix);
+            psFree(permutation);
+            return NULL;
+        }
+        kernels->solution1 = psMatrixLUSolve(kernels->solution1, luMatrix, sumVector, permutation);
+        psFree(sumVector);
+        psFree(luMatrix);
+        psFree(permutation);
+        if (!kernels->solution1) {
+            psError(PS_ERR_UNKNOWN, true, "Failed to solve the least-squares system.\n");
+            return NULL;
+        }
+    } else {
+        // Dual convolution solution
+
+        // Accumulation of stamp matrices/vectors
+        psImage *sumMatrix1 = psImageAlloc(numParams, numParams, PS_TYPE_F64);
+        psImage *sumMatrix2 = psImageAlloc(numParams2, numParams2, PS_TYPE_F64);
+        psImage *sumMatrixX = psImageAlloc(numParams, numParams2, PS_TYPE_F64);
+        psVector *sumVector1 = psVectorAlloc(numParams, PS_TYPE_F64);
+        psVector *sumVector2 = psVectorAlloc(numParams, PS_TYPE_F64);
+        psImageInit(sumMatrix1, 0.0);
+        psImageInit(sumMatrix2, 0.0);
+        psImageInit(sumMatrixX, 0.0);
+        psVectorInit(sumVector1, 0.0);
+        psVectorInit(sumVector2, 0.0);
+
+        int numStamps = 0;              // Number of good stamps
+        for (int i = 0; i < stamps->num; i++) {
+            pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+            if (stamp->status == PM_SUBTRACTION_STAMP_USED) {
+                (void)psBinaryOp(sumMatrix1, sumMatrix1, "+", stamp->matrix1);
+                (void)psBinaryOp(sumMatrix2, sumMatrix2, "+", stamp->matrix2);
+                (void)psBinaryOp(sumMatrixX, sumMatrixX, "+", stamp->matrixX);
+                (void)psBinaryOp(sumVector1, sumVector1, "+", stamp->vector1);
+                (void)psBinaryOp(sumVector2, sumVector2, "+", stamp->vector2);
+                numStamps++;
+            }
+        }
+        calculatePenalty(sumVector1, kernels);
+        calculatePenalty(sumVector2, kernels);
+
+#if 0
+        // Apply weighting to maximise the difference between solution coefficients for the same functions
+        double fudge = PS_SQR(2 * stamps->footprint + 1); // Fudge factor
+        for (int i = 0; i < kernels->num; i++) {
+            sumMatrix1->data.F64[i][i] -= fudge;
+            sumMatrix2->data.F64[i][i] += fudge;
+        }
+#endif
+
+        // Pure matrix operations
+
+        // A * a = Ct * b + d
+        // C * a = B  * b + e
+        //
+        // a = (Ct * Bi * C - A)i (Ct * Bi * e - d)
+        // b = Bi * (C * a - e)
+        psVector *a = psVectorRecycle(kernels->solution1, numParams, PS_TYPE_F64);
+        psVector *b = psVectorRecycle(kernels->solution2, numParams2, PS_TYPE_F64);
+#ifdef TESTING
+        psVectorInit(a, NAN);
+        psVectorInit(b, NAN);
+#endif
+        psImage *A = sumMatrix1;
+        psImage *B = sumMatrix2;
+        psImage *C = sumMatrixX;
+        psVector *d = sumVector1;
+        psVector *e = sumVector2;
+
+        assert(a->n == numParams);
+        assert(b->n == numParams2);
+        assert(A->numRows == numParams && A->numCols == numParams);
+        assert(B->numRows == numParams2 && B->numCols == numParams2);
+        assert(C->numRows == numParams2 && C->numCols == numParams);
+        assert(d->n == numParams);
+        assert(e->n == numParams2);
+
+        psImage *Bi = psMatrixInvert(NULL, B, NULL);
+        assert(Bi->numRows == numParams2 && Bi->numCols == numParams2);
+        psImage *Ct = psMatrixTranspose(NULL, C);
+        assert(Ct->numRows == numParams && Ct->numCols == numParams2);
+
+        psImage *BiC = psMatrixMultiply(NULL, Bi, C);
+        assert(BiC->numRows == numParams2 && BiC->numCols == numParams);
+        psImage *CtBi = psMatrixMultiply(NULL, Ct, Bi);
+        assert(CtBi->numRows == numParams && CtBi->numCols == numParams2);
+
+        psImage *CtBiC = psMatrixMultiply(NULL, Ct, BiC);
+        assert(CtBiC->numRows == numParams && CtBiC->numCols == numParams);
+
+        psImage *F = (psImage*)psBinaryOp(NULL, CtBiC, "-", A);
+        assert(F->numRows == numParams && F->numCols == numParams);
+        float det = 0.0;
+        psImage *Fi = psMatrixInvert(NULL, F, &det);
+        assert(Fi->numRows == numParams && Fi->numCols == numParams);
+        psTrace("psModules.imcombine", 4, "Determinant of F: %f\n", det);
+
+        psVector *g = psVectorAlloc(numParams, PS_TYPE_F64);
+#ifdef TESTING
+        psVectorInit(g, NAN);
+#endif
+        assert(CtBi->numRows == numParams && CtBi->numCols == numParams2);
+        assert(e->n == numParams2);
+        assert(d->n == numParams);
+        for (int i = 0; i < numParams; i++) {
+            double value = 0.0;
+            for (int j = 0; j < numParams2; j++) {
+                value += CtBi->data.F64[i][j] * e->data.F64[j];
+            }
+            g->data.F64[i] = value - d->data.F64[i];
+        }
+
+        assert(Fi->numRows == numParams && Fi->numCols == numParams);
+        assert(g->n == numParams);
+        for (int i = 0; i < numParams; i++) {
+            double value = 0.0;
+            for (int j = 0; j < numParams; j++) {
+                value += Fi->data.F64[i][j] * g->data.F64[j];
+            }
+            a->data.F64[i] = value;
+        }
+
+        psVector *h = psVectorAlloc(numParams2, PS_TYPE_F64);
+#ifdef TESTING
+        psVectorInit(h, NAN);
+#endif
+        assert(C->numRows == numParams2 && C->numCols == numParams);
+        assert(a->n == numParams);
+        assert(e->n == numParams2);
+        for (int i = 0; i < numParams2; i++) {
+            double value = 0.0;
+            for (int j = 0; j < numParams; j++) {
+                value += C->data.F64[i][j] * a->data.F64[j];
+            }
+            h->data.F64[i] = value - e->data.F64[i];
+        }
+
+        assert(Bi->numRows == numParams2 && Bi->numCols == numParams2);
+        assert(h->n == numParams2);
+        for (int i = 0; i < numParams2; i++) {
+            double value = 0.0;
+            for (int j = 0; j < numParams2; j++) {
+                value += Bi->data.F64[i][j] * h->data.F64[j];
+            }
+            b->data.F64[i] = value;
+        }
+
+
+#if 0
+        for (int i = 0; i < numParams; i++) {
+            double aVal1 = 0.0, bVal1 = 0.0;
+            for (int j = 0; j < numParams2; j++) {
+                aVal1 += A->data.F64[i][j] * a->data.F64[j];
+                bVal1 += Ct->data.F64[i][j] * b->data.F64[j];
+            }
+            bVal1 += d->data.F64[i];
+            for (int j = numParams2; j < numParams; j++) {
+                aVal1 += A->data.F64[i][j] * a->data.F64[j];
+            }
+            printf("%d: %lf\n", i, aVal1 - bVal1);
+        }
+
+        for (int i = 0; i < numParams2; i++) {
+            double aVal2 = 0.0, bVal2 = 0.0;
+            for (int j = 0; j < numParams2; j++) {
+                aVal2 += C->data.F64[i][j] * a->data.F64[j];
+                bVal2 += B->data.F64[i][j] * b->data.F64[j];
+            }
+            bVal2 += e->data.F64[i];
+            for (int j = numParams2; j < numParams; j++) {
+                aVal2 += C->data.F64[i][j] * a->data.F64[j];
+            }
+            printf("%d: %lf\n", i, aVal2 - bVal2);
+        }
+#endif
+
+        {
+            psFits *fits = psFitsOpen("sumMatrix1.fits", "w");
+            psFitsWriteImage(fits, NULL, sumMatrix1, 0, NULL);
+            psFitsClose(fits);
+        }
+        {
+            psFits *fits = psFitsOpen("sumMatrix2.fits", "w");
+            psFitsWriteImage(fits, NULL, sumMatrix2, 0, NULL);
+            psFitsClose(fits);
+        }
+        {
+            psFits *fits = psFitsOpen("sumMatrixX.fits", "w");
+            psFitsWriteImage(fits, NULL, sumMatrixX, 0, NULL);
+            psFitsClose(fits);
+        }
+        {
+            psFits *fits = psFitsOpen("sumFinverse.fits", "w");
+            psFitsWriteImage(fits, NULL, Fi, 0, NULL);
+            psFitsClose(fits);
+        }
+
+
+        kernels->solution1 = a;
+        kernels->solution2 = b;
+
+        // XXXXX Free temporary matrices and vectors
+
+    }
+
+    if (psTraceGetLevel("psModules.imcombine") >= 7) {
+        for (int i = 0; i < kernels->solution1->n; i++) {
+            psTrace("psModules.imcombine", 7, "Solution 1 %d: %f\n", i, kernels->solution1->data.F64[i]);
+        }
+        if (kernels->mode == PM_SUBTRACTION_MODE_DUAL) {
+            for (int i = 0; i < kernels->solution2->n; i++) {
+                psTrace("psModules.imcombine", 7, "Solution 2 %d: %f\n", i, kernels->solution2->data.F64[i]);
+            }
+        }
+     }
+
+    return true;
+}
+
+psVector *pmSubtractionCalculateDeviations(pmSubtractionStampList *stamps,
+                                           const pmSubtractionKernels *kernels)
+{
+    PM_ASSERT_SUBTRACTION_STAMP_LIST_NON_NULL(stamps, NULL);
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, NULL);
+    PM_ASSERT_SUBTRACTION_KERNELS_SOLUTION(kernels, NULL);
+
+    psVector *deviations = psVectorAlloc(stamps->num, PS_TYPE_F32); // Mean deviation for stamps
+    int footprint = stamps->footprint; // Half-size of stamps
+    long numPixels = PS_SQR(2 * footprint + 1); // Number of pixels in footprint
+    double devNorm = 1.0 / (double)numPixels; // Normalisation for deviations
+    int numKernels = kernels->num;      // Number of kernels
+
+    psImage *polyValues = NULL;         // Polynomial values
+    psKernel *residual = psKernelAlloc(-footprint, footprint, -footprint, footprint); // Residual image
+
+    for (int i = 0; i < stamps->num; i++) {
+        pmSubtractionStamp *stamp = stamps->stamps->data[i]; // The stamp of interest
+        if (stamp->status != PM_SUBTRACTION_STAMP_USED) {
+            deviations->data.F32[i] = NAN;
+            continue;
+        }
+
+        // Calculate coefficients of the kernel basis functions
+        polyValues = p_pmSubtractionPolynomial(polyValues, kernels->spatialOrder, stamp->xNorm, stamp->yNorm);
+        double norm = p_pmSubtractionSolutionNorm(kernels); // Normalisation
+        double background = p_pmSubtractionSolutionBackground(kernels, polyValues);// Difference in background
+
+        // Calculate residuals
+        psKernel *weight = stamp->weight; // Weight postage stamp
+        psImageInit(residual->image, 0.0);
+        if (kernels->mode != PM_SUBTRACTION_MODE_DUAL) {
+            psKernel *target;           // Target postage stamp
+            psKernel *source;           // Source postage stamp
+            psArray *convolutions;      // Convolution postage stamps for each kernel basis function
+            switch (kernels->mode) {
+              case PM_SUBTRACTION_MODE_1:
+                target = stamp->image2;
+                source = stamp->image1;
+                convolutions = stamp->convolutions1;
+                break;
+              case PM_SUBTRACTION_MODE_2:
+                target = stamp->image1;
+                source = stamp->image2;
+                convolutions = stamp->convolutions2;
+                break;
+              default:
+                psAbort("Unsupported subtraction mode: %x", kernels->mode);
+            }
+
+            for (int j = 0; j < numKernels; j++) {
+                psKernel *convolution = convolutions->data[j]; // Convolution
+                double coefficient = p_pmSubtractionSolutionCoeff(kernels, polyValues, j,
+                                                                  false); // Coefficient
+                for (int y = - footprint; y <= footprint; y++) {
+                    for (int x = - footprint; x <= footprint; x++) {
+                        residual->kernel[y][x] -= convolution->kernel[y][x] * coefficient;
+                    }
+                }
+            }
+            for (int y = - footprint; y <= footprint; y++) {
+                for (int x = - footprint; x <= footprint; x++) {
+                    residual->kernel[y][x] += target->kernel[y][x] - background - source->kernel[y][x] * norm;
+                }
+            }
+        } else {
+            // Dual convolution
+            psArray *convolutions1 = stamp->convolutions1; // Convolutions of the first image
+            psArray *convolutions2 = stamp->convolutions2; // Convolutions of the second image
+            psKernel *image1 = stamp->image1; // The first image
+            psKernel *image2 = stamp->image2; // The second image
+
+            for (int j = 0; j < numKernels; j++) {
+                psKernel *conv1 = convolutions1->data[j]; // Convolution of first image
+                psKernel *conv2 = convolutions2->data[j]; // Convolution of second image
+                double coeff1 = p_pmSubtractionSolutionCoeff(kernels, polyValues, j, false); // Coefficient 1
+                double coeff2 = p_pmSubtractionSolutionCoeff(kernels, polyValues, j, true); // Coefficient 2
+
+                for (int y = - footprint; y <= footprint; y++) {
+                    for (int x = - footprint; x <= footprint; x++) {
+                        residual->kernel[y][x] += conv2->kernel[y][x] * coeff2 - conv1->kernel[y][x] * coeff1;
+                    }
+                }
+            }
+            for (int y = - footprint; y <= footprint; y++) {
+                for (int x = - footprint; x <= footprint; x++) {
+                    residual->kernel[y][x] += image2->kernel[y][x] - background - image1->kernel[y][x] * norm;
+                }
+            }
+        }
+
+        double deviation = 0.0;         // Sum of differences
+        for (int y = - footprint; y <= footprint; y++) {
+            for (int x = - footprint; x <= footprint; x++) {
+                double dev = PS_SQR(residual->kernel[y][x]) / weight->kernel[y][x];
+                deviation += dev;
+#ifdef TESTING
+                residual->kernel[y][x] = dev;
+#endif
+            }
+        }
+        deviations->data.F32[i] = sqrtf(devNorm * deviation);
+        psTrace("psModules.imcombine", 5, "Deviation for stamp %d (%d,%d): %f\n",
+                i, (int)(stamp->x + 0.5), (int)(stamp->y + 0.5), deviations->data.F32[i]);
+        if (!isfinite(deviations->data.F32[i])) {
+            stamp->status = PM_SUBTRACTION_STAMP_REJECTED;
+            psTrace("psModules.imcombine", 5,
+                    "Rejecting stamp %d (%d,%d) because of non-finite deviation\n",
+                    i, (int)(stamp->x + 0.5), (int)(stamp->y + 0.5));
+            continue;
+        }
+
+#ifdef TESTING
+        {
+            psString filename = NULL;
+            psStringAppend(&filename, "resid_%03d.fits", i);
+            psFits *fits = psFitsOpen(filename, "w");
+            psFree(filename);
+            psFitsWriteImage(fits, NULL, residual->image, 0, NULL);
+            psFitsClose(fits);
+        }
+        if (stamp->image1) {
+            psString filename = NULL;
+            psStringAppend(&filename, "stamp_image1_%03d.fits", i);
+            psFits *fits = psFitsOpen(filename, "w");
+            psFree(filename);
+            psFitsWriteImage(fits, NULL, stamp->image1->image, 0, NULL);
+            psFitsClose(fits);
+        }
+        if (stamp->image2) {
+            psString filename = NULL;
+            psStringAppend(&filename, "stamp_image2_%03d.fits", i);
+            psFits *fits = psFitsOpen(filename, "w");
+            psFree(filename);
+            psFitsWriteImage(fits, NULL, stamp->image2->image, 0, NULL);
+            psFitsClose(fits);
+        }
+        if (stamp->weight) {
+            psString filename = NULL;
+            psStringAppend(&filename, "stamp_weight_%03d.fits", i);
+            psFits *fits = psFitsOpen(filename, "w");
+            psFree(filename);
+            psFitsWriteImage(fits, NULL, stamp->weight->image, 0, NULL);
+            psFitsClose(fits);
+        }
+#endif
+
+    }
+    psFree(residual);
+    psFree(polyValues);
+
+    return deviations;
+}
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionEquation.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionEquation.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionEquation.h	(revision 19481)
@@ -0,0 +1,56 @@
+#ifndef PM_SUBTRACTION_EQUATION_H
+#define PM_SUBTRACTION_EQUATION_H
+
+#include "pmSubtractionStamps.h"
+#include "pmSubtractionKernels.h"
+
+/// Execute a thread job to calculate the least-squares equation for a stamp
+bool pmSubtractionCalculateEquationThread(psThreadJob *job ///< Job to execute
+    );
+
+/// Calculate the least-squares equation to match the image quality for a single stamp
+bool pmSubtractionCalculateEquationStamp(pmSubtractionStampList *stamps, ///< Stamps
+                                         const pmSubtractionKernels *kernels, ///< Kernel parameters
+                                         int index ///< Index of stamp
+                                    );
+
+/// Calculate the least-squares equation to match the image quality
+bool pmSubtractionCalculateEquation(pmSubtractionStampList *stamps, ///< Stamps
+                                    const pmSubtractionKernels *kernels ///< Kernel parameters
+                                    );
+
+/// Solve the least-squares equation to match the image quality
+bool pmSubtractionSolveEquation(pmSubtractionKernels *kernels, ///< Kernel parameters
+                                const pmSubtractionStampList *stamps ///< Stamps
+    );
+
+/// Calculate deviations
+psVector *pmSubtractionCalculateDeviations(pmSubtractionStampList *stamps, ///< Stamps
+                                           const pmSubtractionKernels *kernels ///< Kernel parameters
+    );
+
+/// Calculate the value of a polynomial, specified by coefficients and polynomial values
+double p_pmSubtractionCalculatePolynomial(const psVector *coeff, ///< Coefficients
+                                          const psImage *polyValues, ///< Polynomial values
+                                          int order, ///< Order of polynomials
+                                          int index, ///< Index at which to begin
+                                          int step ///< Step between subsequent indices
+    );
+
+/// Return the specified coefficient in the solution
+double p_pmSubtractionSolutionCoeff(const pmSubtractionKernels *kernels, ///< Kernel parameters
+                                    const psImage *polyValues, ///< Polynomial values
+                                    int index, ///< Coefficient index to calculate
+                                    bool wantDual ///< Calculate the coefficient for the dual solution?
+    );
+
+/// Return the normalisation in the solution
+double p_pmSubtractionSolutionNorm(const pmSubtractionKernels *kernels ///< Kernel parameters
+    );
+
+/// Return the background (difference) in the solution
+double p_pmSubtractionSolutionBackground(const pmSubtractionKernels *kernels, ///< Kernel parameters
+                                         const psImage *polyValues ///< Polynomial values
+    );
+
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionIO.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionIO.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionIO.c	(revision 19481)
@@ -0,0 +1,618 @@
+#include <stdio.h>
+#include <pslib.h>
+#include <string.h>
+
+#include "pmHDU.h"
+#include "pmFPA.h"
+#include "pmFPAview.h"
+#include "pmFPAfile.h"
+#include "pmFPAfileFitsIO.h"
+#include "pmConcepts.h"
+
+#include "pmSubtraction.h"
+#include "pmSubtractionKernels.h"
+#include "pmSubtractionMatch.h"
+
+#define ARRAY_BUFFER 16                 // Number to add to array at a time
+
+// Names of FITS table columns
+#define NAME_XMIN "XMIN"                // Region of applicability: minimum x value
+#define NAME_XMAX "XMAX"                // Region of applicability: maximum x value
+#define NAME_YMIN "YMIN"                // Region of applicability: minimum y value
+#define NAME_YMAX "YMAX"                // Region of applicability: maximum y value
+#define NAME_KERNEL "KERNEL"            // Kernel description
+#define NAME_TYPE "TYPE"                // Kernel type
+#define NAME_SIZE "SIZE"                // Kernel half-size
+#define NAME_INNER "INNER"              // Size of inner region (only applicable for some kernel types)
+#define NAME_SPATIAL "SPATIAL_ORDER"    // Order of spatial polynomial
+#define NAME_BG "BG_ORDER"              // Order of background polynomial
+#define NAME_MODE "MODE"                // Matching mode
+#define NAME_COLS "COLUMNS"             // Number of columns
+#define NAME_ROWS "ROWS"                // Number of rows
+#define NAME_SOL1 "SOLUTION_1"          // Solution for convolving image 1
+#define NAME_SOL2 "SOLUTION_2"          // Solution for convolving image 2
+
+#define FALLBACK_EXTNAME "SUBTRACTION_KERNEL" // Fallback extension name
+
+// Generate an extension name for the subtraction kernel
+// In the case that there is no chip/cell (the readout is orphaned), the fallback extension name is used
+static psString subtractionKernelExtname(const pmReadout *ro // Readout containing kernel
+                                         )
+{
+    psAssert(ro, "Required parameter");
+
+    psString extname = NULL;            // Name of FITS extension
+    pmCell *cell = ro->parent;     // Parent cell
+    if (cell && cell->parent) {
+        pmChip *chip = cell->parent;        // Parent chip
+        const char *chipName = psMetadataLookupStr(NULL, chip->concepts, "CHIP.NAME"); // Name of chip
+        const char *cellName = psMetadataLookupStr(NULL, cell->concepts, "CELL.NAME"); // Name of cell
+        psAssert(chipName && cellName, "Chip and cell names should be defined in concepts.");
+        int roNum = -1;                 // Index of readout
+        for (int i = 0; i < cell->readouts->n && roNum == -1; i++) {
+            if (cell->readouts->data[i] == ro) {
+                roNum = i;
+            }
+        }
+        if (roNum != -1) {
+            psStringAppend(&extname, "%s_%s_%d", chipName, cellName, roNum);
+        } else {
+            psWarning("Unable to find readout number");
+        }
+    }
+
+    if (!extname) {
+        extname = psStringCopy(FALLBACK_EXTNAME);
+    }
+
+    return extname;
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+bool pmReadoutWriteSubtractionKernels(pmReadout *ro, psFits *fits)
+{
+    PM_ASSERT_READOUT_NON_NULL(ro, false);
+    PS_ASSERT_FITS_NON_NULL(fits, false);
+
+    // Extract the regions and solutions used in the image matching
+    psArray *regions = psArrayAllocEmpty(ARRAY_BUFFER); // Array of regions
+    {
+        psString regex = NULL;          // Regular expression
+        psStringAppend(&regex, "^%s$", PM_SUBTRACTION_ANALYSIS_REGION);
+        psMetadataIterator *iter = psMetadataIteratorAlloc(ro->analysis, PS_LIST_HEAD, regex); // Iterator
+        psFree(regex);
+        psMetadataItem *item = NULL;// Item from iteration
+        while ((item = psMetadataGetAndIncrement(iter))) {
+            assert(item->type == PS_DATA_REGION);
+            regions = psArrayAdd(regions, ARRAY_BUFFER, item->data.V);
+        }
+        psFree(iter);
+    }
+    if (regions->n == 0) {
+        psError(PS_ERR_BAD_PARAMETER_VALUE, true, "No subtraction regions found.");
+        psFree(regions);
+        return false;
+    }
+
+    psArray *kernels = psArrayAllocEmpty(ARRAY_BUFFER); // Array of kernels
+    {
+        psString regex = NULL;          // Regular expression
+        psStringAppend(&regex, "^%s$", PM_SUBTRACTION_ANALYSIS_KERNEL);
+        psMetadataIterator *iter = psMetadataIteratorAlloc(ro->analysis, PS_LIST_HEAD, regex); // Iterator
+        psFree(regex);
+        psMetadataItem *item = NULL;// Item from iteration
+        while ((item = psMetadataGetAndIncrement(iter))) {
+            assert(item->type == PS_DATA_UNKNOWN);
+            // Set the normalisation dimensions, since these will be otherwise unavailable when reading the
+            // images by scans.
+            pmSubtractionKernels *kernel = item->data.V; // Kernel used in subtraction
+            kernel->numCols = ro->image->numCols;
+            kernel->numRows = ro->image->numRows;
+
+            kernels = psArrayAdd(kernels, ARRAY_BUFFER, kernel);
+        }
+        psFree(iter);
+    }
+
+    if (regions->n != kernels->n) {
+        psError(PS_ERR_BAD_PARAMETER_SIZE, true, "Number of regions (%ld) and kernels (%ld) don't match.\n",
+                regions->n, kernels->n);
+        psFree(regions);
+        psFree(kernels);
+        return false;
+    }
+
+    // Format for writing a table
+    int num = regions->n;              // Number of regions and kernels
+    psArray *rows = psArrayAlloc(num); // Array of FITS table rows
+    for (int i = 0; i < num; i++) {
+        psMetadata *row = psMetadataAlloc(); // Row of interest
+        rows->data[i] = psMemIncrRefCounter(row);
+
+        psRegion *region = regions->data[i]; // Region of interest
+        psMetadataAddS32(row, PS_LIST_TAIL, NAME_XMIN,  0, "Applicability minimum x", region->x0);
+        psMetadataAddS32(row, PS_LIST_TAIL, NAME_XMAX,  0, "Applicability maximum x", region->x1);
+        psMetadataAddS32(row, PS_LIST_TAIL, NAME_YMIN,  0, "Applicability minimum y", region->y0);
+        psMetadataAddS32(row, PS_LIST_TAIL, NAME_YMAX,  0, "Applicability maximum y", region->y1);
+
+        pmSubtractionKernels *kernel = kernels->data[i]; // Kernel
+        psMetadataAddStr(row, PS_LIST_TAIL, NAME_KERNEL, 0, "Kernel description", kernel->description);
+
+#if 0
+        psMetadataAddS32(row, PS_LIST_TAIL, NAME_TYPE,  0, "Kernel type (enum)", kernel->type);
+        psMetadataAddS32(row, PS_LIST_TAIL, NAME_SIZE,  0, "Kernel half-size", kernel->size);
+        psMetadataAddS32(row, PS_LIST_TAIL, NAME_INNER, 0, "Size of inner region", kernel->inner);
+        psMetadataAddS32(row, PS_LIST_TAIL, NAME_SPATIAL, 0, "Polynomial order for spatial variations",
+                         kernel->spatialOrder);
+#endif
+
+        psMetadataAddS32(row, PS_LIST_TAIL, NAME_BG,  0, "Polynomial order for background fitting",
+                         kernel->bgOrder);
+        psMetadataAddS32(row, PS_LIST_TAIL, NAME_MODE,  0, "Matching mode (enum)", kernel->mode);
+        psMetadataAddS32(row, PS_LIST_TAIL, NAME_COLS,  0, "Number of columns", kernel->numCols);
+        psMetadataAddS32(row, PS_LIST_TAIL, NAME_ROWS,  0, "Number of rows", kernel->numRows);
+        if (kernel->mode == PM_SUBTRACTION_MODE_1 || kernel->mode == PM_SUBTRACTION_MODE_DUAL) {
+            psMetadataAddVector(row, PS_LIST_TAIL, NAME_SOL1, 0, "Solution vector 1", kernel->solution1);
+        }
+        if (kernel->mode == PM_SUBTRACTION_MODE_2 || kernel->mode == PM_SUBTRACTION_MODE_DUAL) {
+            psMetadataAddVector(row, PS_LIST_TAIL, NAME_SOL2, 0, "Solution vector 2", kernel->solution2);
+        }
+    }
+    psFree(regions);
+    psFree(kernels);
+
+    psMetadata *header = psMetadataAlloc(); // Header for FITS file
+
+    // CVS tags, used to identify the version of this file (in case incompatibilities are introduced)
+    psString cvsFile = psStringCopy("$RCSfile: pmSubtractionIO.c,v $");
+    psString cvsRev  = psStringCopy("$Revision: 1.3 $");
+    psString cvsDate = psStringCopy("$Date: 2008-08-27 02:54:44 $");
+    psStringSubstitute(&cvsFile, NULL, "RCSfile: ");
+    psStringSubstitute(&cvsRev,  NULL, "Revision: ");
+    psStringSubstitute(&cvsDate, NULL, "Date: ");
+
+    psString version = NULL;            // Version information, for header
+    psStringAppend(&version, "%s %s %s", cvsFile, cvsRev, cvsDate);
+    psFree(cvsFile);
+    psFree(cvsRev);
+    psFree(cvsDate);
+    psStringSubstitute(&version, NULL, "$");
+    psMetadataAddStr(header, PS_LIST_TAIL, "PSVERSION", 0, "S/W version", version);
+    psFree(version);
+
+    psString extname = subtractionKernelExtname(ro); // Extension name
+
+    if (!psFitsWriteTable(fits, header, rows, extname)) {
+        psError(PS_ERR_IO, false, "Unable to write subtraction kernel to FITS table.");
+        psFree(header);
+        psFree(rows);
+        psFree(extname);
+        return false;
+    }
+
+    psFree(header);
+    psFree(rows);
+    psFree(extname);
+
+    return true;
+}
+
+static bool pmCellWriteSubtractionKernels(pmCell *cell, const pmFPAview *view,
+                                          pmFPAfile *file, pmConfig *config)
+{
+    PS_ASSERT_PTR_NON_NULL(cell, false);
+    PS_ASSERT_PTR_NON_NULL(cell->readouts, false);
+    PS_ASSERT_PTR_NON_NULL(view, false);
+
+    pmFPAview *thisView = pmFPAviewAlloc(view->nRows); // Copy of input view
+    *thisView = *view;
+
+    for (int i = 0; i < cell->readouts->n; i++) {
+        pmReadout *readout = cell->readouts->data[i];
+        thisView->readout = i;
+        if (!pmReadoutWriteSubtractionKernels(readout, file->fits)) {
+            psError(PS_ERR_IO, false, "Failed to write %dth readout", i);
+            psFree(thisView);
+            return false;
+        }
+    }
+    psFree(thisView);
+    return true;
+}
+
+static bool pmChipWriteSubtractionKernels(pmChip *chip, const pmFPAview *view,
+                                          pmFPAfile *file, pmConfig *config)
+{
+    PS_ASSERT_PTR_NON_NULL(chip, false);
+    PS_ASSERT_PTR_NON_NULL(chip->cells, false);
+    PS_ASSERT_PTR_NON_NULL(view, false);
+
+    pmFPAview *thisView = pmFPAviewAlloc(view->nRows); // Copy of input view
+    *thisView = *view;
+
+    for (int i = 0; i < chip->cells->n; i++) {
+        pmCell *cell = chip->cells->data[i];
+        thisView->cell = i;
+        if (!pmCellWriteSubtractionKernels(cell, thisView, file, config)) {
+            psError(PS_ERR_IO, false, "Failed to write %dth cell", i);
+            psFree(thisView);
+            return false;
+        }
+    }
+    psFree(thisView);
+    return true;
+}
+
+static bool pmFPAWriteSubtractionKernels(pmFPA *fpa, const pmFPAview *view,
+                                         pmFPAfile *file, pmConfig *config)
+{
+    PS_ASSERT_PTR_NON_NULL(view, false);
+    PS_ASSERT_PTR_NON_NULL(fpa, false);
+    PS_ASSERT_PTR_NON_NULL(fpa->chips, false);
+
+    pmFPAview *thisView = pmFPAviewAlloc(view->nRows); // Copy of input view
+    *thisView = *view;
+
+    for (int i = 0; i < fpa->chips->n; i++) {
+        pmChip *chip = fpa->chips->data[i];
+        thisView->chip = i;
+        if (!pmChipWriteSubtractionKernels(chip, thisView, file, config)) {
+            psError(PS_ERR_IO, false, "Failed to write %dth chip", i);
+            psFree(thisView);
+            return false;
+        }
+    }
+    psFree(thisView);
+    return true;
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+
+bool pmReadoutReadSubtractionKernels(pmReadout *ro, psFits *fits)
+{
+    PM_ASSERT_READOUT_NON_NULL(ro, false);
+    PS_ASSERT_FITS_NON_NULL(fits, false);
+
+    psString extname = subtractionKernelExtname(ro); // Extension name
+    if (!psFitsMoveExtName(fits, extname)) {
+        if (strcmp(extname, FALLBACK_EXTNAME) != 0) {
+            psErrorClear();
+            if (!psFitsMoveExtName(fits, FALLBACK_EXTNAME)) {
+                psError(PS_ERR_IO, false, "Unable to move to subtraction kernel table.");
+                psFree(extname);
+                return false;
+            }
+        } else {
+            psError(PS_ERR_IO, false, "Unable to move to subtraction kernel table.");
+            psFree(extname);
+            return false;
+        }
+    }
+    psFree(extname);
+
+    psArray *table = psFitsReadTable(fits); // Table of interest
+    if (!table) {
+        psError(PS_ERR_IO, false, "Unable to read FITS table");
+        return false;
+    }
+
+    // Look up a column value for a row
+#define TABLE_LOOKUP(TYPE, SUFFIX, TARGET, NAME) \
+    TYPE TARGET; \
+    { \
+        bool mdok; \
+        TARGET = psMetadataLookup##SUFFIX(&mdok, row, NAME); \
+        if (!mdok) { \
+            psError(PS_ERR_UNKNOWN, false, "Unable to find column %s in subtraction kernel table.", NAME); \
+            psFree(table); \
+            return false; \
+        } \
+    }
+
+    for (int i = 0; i < table->n; i++) {
+        psMetadata *row = table->data[i]; // Table row
+
+        TABLE_LOOKUP(int, S32, xMin, NAME_XMIN);
+        TABLE_LOOKUP(int, S32, xMax, NAME_XMAX);
+        TABLE_LOOKUP(int, S32, yMin, NAME_YMIN);
+        TABLE_LOOKUP(int, S32, yMax, NAME_YMAX);
+
+        psRegion *region = psRegionAlloc(xMin, xMax, yMin, yMax); // Region of applicability
+        psMetadataAddPtr(ro->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_REGION, PS_DATA_REGION,
+                         "Subtraction region", region);
+        psFree(region);
+
+        TABLE_LOOKUP(const char *, Str, description, NAME_KERNEL);
+        TABLE_LOOKUP(pmSubtractionMode, S32, mode,    NAME_MODE);
+
+#if 0
+        TABLE_LOOKUP(int, S32, size,    NAME_SIZE);
+        TABLE_LOOKUP(int, S32, inner,   NAME_INNER);
+        TABLE_LOOKUP(int, S32, spatial, NAME_SPATIAL);
+#endif
+
+        TABLE_LOOKUP(int, S32, bg,      NAME_BG);
+        TABLE_LOOKUP(int, S32, numCols, NAME_COLS);
+        TABLE_LOOKUP(int, S32, numRows, NAME_ROWS);
+
+        pmSubtractionKernels *kernels = pmSubtractionKernelsFromDescription(description, bg, mode);
+        kernels->numCols = numCols;
+        kernels->numRows = numRows;
+
+        bool mdok;                      // Status of MD lookup
+        if (mode == PM_SUBTRACTION_MODE_1 || mode == PM_SUBTRACTION_MODE_DUAL) {
+            kernels->solution1 = psMemIncrRefCounter(psMetadataLookupPtr(&mdok, row, NAME_SOL1));
+            if (!mdok) {
+                psError(PS_ERR_UNKNOWN, false, "Unable to find column %s in subtraction kernel table.",
+                        NAME_SOL1);
+                psFree(kernels);
+                psFree(table);
+                return false;
+            }
+        }
+        if (mode == PM_SUBTRACTION_MODE_2 || mode == PM_SUBTRACTION_MODE_DUAL) {
+            kernels->solution2 = psMemIncrRefCounter(psMetadataLookupPtr(&mdok, row, NAME_SOL2));
+            if (!mdok) {
+                psError(PS_ERR_UNKNOWN, false, "Unable to find column %s in subtraction kernel table.",
+                        NAME_SOL2);
+                psFree(kernels);
+                psFree(table);
+                return false;
+            }
+        }
+
+        psMetadataAddPtr(ro->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_KERNEL, PS_DATA_UNKNOWN,
+                         "Subtraction kernels", kernels);
+        psFree(kernels);
+    }
+    psFree(table);
+    return true;
+}
+
+static bool pmCellReadSubtractionKernels(pmCell *cell, const pmFPAview *view,
+                                         pmFPAfile *file, pmConfig *config)
+{
+    PS_ASSERT_PTR_NON_NULL(view, false);
+    PS_ASSERT_PTR_NON_NULL(cell, false);
+    PS_ASSERT_PTR_NON_NULL(cell->readouts, false);
+
+    pmFPAview *thisView = pmFPAviewAlloc(view->nRows); // Copy of input view
+    *thisView = *view;
+
+    // Create a readout if none exists
+    if (!cell->readouts || cell->readouts->n == 0) {
+        pmReadout *readout = pmReadoutAlloc(cell); // New readout
+        psFree(readout);                // Drop reference
+    }
+
+    cell->data_exists = false;
+    for (int i = 0; i < cell->readouts->n; i++) {
+        pmReadout *readout = cell->readouts->data[i];
+        thisView->readout = i;
+        pmReadoutReadSubtractionKernels(readout, file->fits);
+        if (!readout->data_exists) {
+            continue;
+        }
+
+        // load in the concept information for this cell
+        if (!pmConceptsReadCell(cell, PM_CONCEPT_SOURCE_HEADER, true, NULL)) {
+            psErrorClear();
+            psWarning("Difficulty reading concepts for cell; attempting to proceed.");
+        }
+        cell->data_exists = true;
+    }
+    psFree(thisView);
+
+    return true;
+}
+
+static bool pmChipReadSubtractionKernels(pmChip *chip, const pmFPAview *view,
+                                         pmFPAfile *file, pmConfig *config)
+{
+    PS_ASSERT_PTR_NON_NULL(view, false);
+    PS_ASSERT_PTR_NON_NULL(chip, false);
+    PS_ASSERT_PTR_NON_NULL(chip->cells, false);
+
+    pmFPAview *thisView = pmFPAviewAlloc(view->nRows); // Copy of input view
+    *thisView = *view;
+
+    chip->data_exists = false;
+    for (int i = 0; i < chip->cells->n; i++) {
+        pmCell *cell = chip->cells->data[i];
+        thisView->cell = i;
+        pmCellReadSubtractionKernels(cell, thisView, file, config);
+        if (!cell->data_exists) {
+            continue;
+        }
+        chip->data_exists = true;
+    }
+    psFree(thisView);
+
+    if (!pmConceptsReadChip(chip, PM_CONCEPT_SOURCE_HEADER, true, true, NULL)) {
+        psError(PS_ERR_IO, false, "Failed to read concepts for chip.\n");
+        return false;
+    }
+
+    return true;
+}
+
+static bool pmFPAReadSubtractionKernels(pmFPA *fpa, const pmFPAview *view,
+                                        pmFPAfile *file, pmConfig *config)
+{
+    PS_ASSERT_PTR_NON_NULL(view, false);
+    PS_ASSERT_PTR_NON_NULL(file, false);
+    PS_ASSERT_PTR_NON_NULL(file->fpa, false);
+    PS_ASSERT_PTR_NON_NULL(file->fpa->chips, false);
+
+    pmFPAview *thisView = pmFPAviewAlloc(view->nRows); // Copy of input view
+    *thisView = *view;
+
+    for (int i = 0; i < fpa->chips->n; i++) {
+        pmChip *chip = fpa->chips->data[i];
+        thisView->chip = i;
+        pmChipReadSubtractionKernels(chip, thisView, file, config);
+    }
+    psFree(thisView);
+
+    if (!pmConceptsReadFPA(fpa, PM_CONCEPT_SOURCE_HEADER, true, NULL)) {
+        psError(PS_ERR_IO, false, "Failed to read concepts for fpa.\n");
+        return false;
+    }
+
+    return true;
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+bool pmSubtractionWriteKernels(const pmFPAview *view, pmFPAfile *file, pmConfig *config)
+{
+    PS_ASSERT_PTR_NON_NULL(view, false);
+    PS_ASSERT_PTR_NON_NULL(file, false);
+    PS_ASSERT_PTR_NON_NULL(file->fpa, false);
+
+    pmFPA *fpa = pmFPAfileSuitableFPA(file, view, config, false); // Suitable FPA for writing
+
+    if (view->chip == -1) {
+        if (!pmFPAWriteSubtractionKernels(fpa, view, file, config)) {
+            psError(PS_ERR_IO, false, "Failed to write subtraction kernels from fpa");
+            psFree(fpa);
+            return false;
+        }
+        psFree(fpa);
+        return true;
+    }
+
+    if (view->chip >= fpa->chips->n) {
+        psError(PS_ERR_UNKNOWN, false, "Writing chip == %d (>= chips->n == %ld)", view->chip, fpa->chips->n);
+        psFree(fpa);
+        return false;
+    }
+    pmChip *chip = fpa->chips->data[view->chip];
+
+    if (view->cell == -1) {
+        if (!pmChipWriteSubtractionKernels(chip, view, file, config)) {
+            psError(PS_ERR_IO, false, "Failed to write objects from chip");
+            psFree(fpa);
+            return false;
+        }
+        psFree(fpa);
+        return true;
+    }
+
+    if (view->cell >= chip->cells->n) {
+        psError(PS_ERR_UNKNOWN, false, "Writing cell == %d (>= cells->n == %ld)",
+                view->cell, chip->cells->n);
+        psFree(fpa);
+        return false;
+    }
+    pmCell *cell = chip->cells->data[view->cell];
+
+    if (view->readout == -1) {
+        if (!pmCellWriteSubtractionKernels(cell, view, file, config)) {
+            psError(PS_ERR_IO, false, "Failed to write objects from cell");
+            psFree(fpa);
+            return false;
+        }
+        psFree(fpa);
+        return true;
+    }
+
+    if (view->readout >= cell->readouts->n) {
+        psError(PS_ERR_UNKNOWN, false, "Writing readout == %d (>= readouts->n == %ld)",
+                view->readout, cell->readouts->n);
+        psFree(fpa);
+        return false;
+    }
+    pmReadout *readout = cell->readouts->data[view->readout];
+
+    if (!pmReadoutWriteSubtractionKernels(readout, file->fits)) {
+        psError(PS_ERR_IO, false, "Failed to write objects from readout %d", view->readout);
+        psFree(fpa);
+        return false;
+    }
+
+    psFree(fpa);
+    return true;
+}
+
+bool pmSubtractionWritePHU(const pmFPAview *view, pmFPAfile *file, pmConfig *config)
+{
+    PS_ASSERT_PTR_NON_NULL(view, false);
+    PS_ASSERT_PTR_NON_NULL(file, false);
+
+    if (file->wrote_phu) {
+        return true;
+    }
+    if (file->fileLevel != PM_FPA_LEVEL_FPA) {
+        return true;
+    }
+    if (file->fpa->chips->n == 1) {
+        return true;
+    }
+
+    // find the FPA phu
+    pmFPA *fpa = pmFPAfileSuitableFPA(file, view, config, false); // Suitable FPA for writing
+    pmHDU *phu = psMemIncrRefCounter(pmFPAviewThisPHU(view, fpa));
+    psFree(fpa);
+
+    // if there is no PHU, this is a single header+image (extension-less) file. This could be the case for an
+    // input SPLIT set of files being written out as a MEF.  if there is a PHU, write it out as a 'blank'
+    psMetadata *outhead = psMetadataAlloc();
+    if (phu) {
+        psMetadataCopy (outhead, phu->header);
+    }
+    psFree(phu);
+
+    pmConfigConformHeader(outhead, file->format);
+
+    psFitsWriteBlank(file->fits, outhead, "");
+    file->wrote_phu = true;
+
+    psTrace("pmFPAfile", 5, "wrote phu %s (type: %d)\n", file->filename, file->type);
+    psFree(outhead);
+
+    return true;
+}
+
+bool pmSubtractionReadKernels(const pmFPAview *view, pmFPAfile *file, pmConfig *config)
+{
+    PS_ASSERT_PTR_NON_NULL(view, false);
+    PS_ASSERT_PTR_NON_NULL(file, false);
+    PS_ASSERT_PTR_NON_NULL(file->fpa, false);
+
+    pmFPA *fpa = file->fpa;
+
+    if (view->chip == -1) {
+        pmFPAReadSubtractionKernels(fpa, view, file, config);
+        return true;
+    }
+
+    if (view->chip >= fpa->chips->n) {
+        return false;
+    }
+    pmChip *chip = fpa->chips->data[view->chip];
+
+    if (view->cell == -1) {
+        pmChipReadSubtractionKernels(chip, view, file, config);
+        return true;
+    }
+
+    if (view->cell >= chip->cells->n) {
+        return false;
+    }
+    pmCell *cell = chip->cells->data[view->cell];
+
+    if (view->readout == -1) {
+        pmCellReadSubtractionKernels(cell, view, file, config);
+        return true;
+    }
+
+    if (view->readout >= cell->readouts->n) {
+        return false;
+    }
+    pmReadout *readout = cell->readouts->data[view->readout];
+
+    return pmReadoutReadSubtractionKernels(readout, file->fits);
+}
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionIO.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionIO.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionIO.h	(revision 19481)
@@ -0,0 +1,37 @@
+#ifndef PM_SUBTRACTION_IO_H
+#define PM_SUBTRACTION_IO_H
+
+#include <pslib.h>
+
+#include <pmHDU.h>
+#include <pmFPA.h>
+
+/// Write subtraction kernels within a readout to a FITS file
+bool pmReadoutWriteSubtractionKernels(
+    pmReadout *readout,                 ///< Readout for which to write subtraction kernels (in analysis MD)
+    psFits *fits                        ///< FITS file to which to write
+    );
+
+
+bool pmReadoutReadSubtractionKernels(
+    pmReadout *readout,                 ///< Readout for which to read subtraction kernels (into analysis MD)
+    psFits *fits                        ///< FITS file to which to write
+    );
+
+
+bool pmSubtractionReadKernels(const pmFPAview *view, ///< View into which to read
+                              pmFPAfile *file, ///< File from which to read
+                              pmConfig *config ///< Configuration
+                              );
+
+bool pmSubtractionWriteKernels(const pmFPAview *view, ///< View from which to write
+                               pmFPAfile *file, ///< File to which to write
+                               pmConfig *config ///< Configuration
+                               );
+
+bool pmSubtractionWritePHU(const pmFPAview *view, // View to PHU
+                           pmFPAfile *file, ///< File to which to write
+                           pmConfig *config ///< Configuration
+                           );
+
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionKernels.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionKernels.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionKernels.c	(revision 19481)
@@ -0,0 +1,762 @@
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <stdio.h>
+#include <string.h>
+#include <strings.h>
+#include <pslib.h>
+
+#include "pmSubtraction.h"
+#include "pmSubtractionKernels.h"
+
+#define RINGS_BUFFER 10                 // Buffer size for RINGS data
+
+
+// Free function for pmSubtractionKernels
+static void subtractionKernelsFree(pmSubtractionKernels *kernels)
+{
+    psFree(kernels->description);
+    psFree(kernels->u);
+    psFree(kernels->v);
+    psFree(kernels->widths);
+    psFree(kernels->uStop);
+    psFree(kernels->vStop);
+    psFree(kernels->preCalc);
+    psFree(kernels->penalties);
+    psFree(kernels->solution1);
+    psFree(kernels->solution2);
+}
+
+// Raise an integer to an integer power
+static inline long power(int value,     // Value
+                         int exp        // Exponent
+    )
+{
+    if (exp == 0) {
+        return 1.0;
+    }
+    long result = value;               // Result to return
+    for (int i = 2; i <= exp; i++) {
+        result *= value;
+    }
+    return result;
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Semi-public functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+bool p_pmSubtractionKernelsAddGrid(pmSubtractionKernels *kernels, int start, int size)
+{
+    PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(kernels, false);
+    PS_ASSERT_INT_NONNEGATIVE(start, false);
+    PS_ASSERT_INT_NONNEGATIVE(size, false);
+
+    int numNew = PS_SQR(2 * size + 1) - 1;  // Number of new kernel parameters to add
+
+    // Ensure the sizes match
+    kernels->widths = psVectorRealloc(kernels->widths, start + numNew);
+    kernels->u = psVectorRealloc(kernels->u, start + numNew);
+    kernels->v = psVectorRealloc(kernels->v, start + numNew);
+    kernels->preCalc = psArrayRealloc(kernels->preCalc, start + numNew);
+    kernels->penalties = psVectorRealloc(kernels->penalties, start + numNew);
+    kernels->inner = start;
+
+    // Generate a set of kernels for each (u,v)
+    for (int v = - size, index = start; v <= size; v++) {
+        for (int u = - size; u <= size; u++, index++) {
+            if (v == 0 && u == 0) {
+                // Skip normalisation component: added explicitly
+                index--;
+                continue;
+            }
+            kernels->widths->data.F32[index] = NAN;
+            kernels->u->data.S32[index] = u;
+            kernels->v->data.S32[index] = v;
+            kernels->preCalc->data[index] = NULL;
+            kernels->penalties->data.F32[index] = kernels->penalty * (PS_SQR(u) + PS_SQR(v));
+
+            psTrace("psModules.imcombine", 7, "Kernel %d: %d %d\n", index, u, v);
+        }
+    }
+
+    return true;
+}
+
+pmSubtractionKernels *p_pmSubtractionKernelsRawISIS(int size, int spatialOrder,
+                                                    const psVector *fwhms, const psVector *orders,
+                                                    float penalty, pmSubtractionMode mode)
+{
+    PS_ASSERT_VECTOR_NON_NULL(fwhms, NULL);
+    PS_ASSERT_VECTOR_TYPE(fwhms, PS_TYPE_F32, NULL);
+    PS_ASSERT_VECTOR_NON_NULL(orders, NULL);
+    PS_ASSERT_VECTOR_TYPE(orders, PS_TYPE_S32, NULL);
+    PS_ASSERT_VECTORS_SIZE_EQUAL(fwhms, orders, NULL);
+    PS_ASSERT_INT_POSITIVE(size, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(spatialOrder, NULL);
+
+    int numGaussians = fwhms->n;       // Number of Gaussians
+
+    int num = 0;                        // Number of basis functions
+    psString params = NULL;             // List of parameters
+    for (int i = 0; i < numGaussians; i++) {
+        int gaussOrder = orders->data.S32[i]; // Polynomial order to apply to Gaussian
+        psStringAppend(&params, "(%.1f,%d)", fwhms->data.F32[i], orders->data.S32[i]);
+        num += (gaussOrder + 1) * (gaussOrder + 2) / 2;
+    }
+
+    pmSubtractionKernels *kernels = pmSubtractionKernelsAlloc(num, PM_SUBTRACTION_KERNEL_ISIS, size,
+                                                              spatialOrder, penalty, mode); // The kernels
+    psStringAppend(&kernels->description, "ISIS(%d,%s,%d,%.2e)", size, params, spatialOrder, penalty);
+
+    psLogMsg("psModules.imcombine", PS_LOG_INFO, "ISIS kernel: %s,%d --> %d elements",
+             params, spatialOrder, num);
+    psFree(params);
+
+    // Set the kernel parameters
+    for (int i = 0, index = 0; i < numGaussians; i++) {
+        float sigma = fwhms->data.F32[i] / (2.0 * sqrtf(2.0 * logf(2.0))); // Gaussian sigma
+        float norm = 1.0 / (M_2_PI * sqrtf(sigma)); // Normalisation for Gaussian
+        // Iterate over (u,v) order
+        for (int uOrder = 0; uOrder <= orders->data.S32[i]; uOrder++) {
+            for (int vOrder = 0; vOrder <= orders->data.S32[i] - uOrder; vOrder++, index++) {
+                // Set the pre-calculated kernel
+                psKernel *preCalc = psKernelAlloc(-size, size, -size, size);
+                double sum = 0.0;       // Normalisation
+                double moment = 0.0;    // Moment, for penalty
+                for (int v = -size; v <= size; v++) {
+                    for (int u = -size; u <= size; u++) {
+                        sum += preCalc->kernel[v][u] = norm * power(u, uOrder) * power(v, vOrder) *
+                            expf(-0.5 * (PS_SQR(u) + PS_SQR(v)) / PS_SQR(sigma));
+                        moment += preCalc->kernel[v][u] * (PS_SQR(u) + PS_SQR(v));
+                    }
+                }
+
+                // Normalise sum of kernel component to unity for even functions
+                if (uOrder % 2 == 0 && vOrder % 2 == 0) {
+                    for (int v = -size; v <= size; v++) {
+                        for (int u = -size; u <= size; u++) {
+                            preCalc->kernel[v][u] = preCalc->kernel[v][u] / sum;
+                        }
+                    }
+                    preCalc->kernel[0][0] -= 1.0;
+                }
+
+                kernels->widths->data.F32[index] = fwhms->data.F32[i];
+                kernels->u->data.S32[index] = uOrder;
+                kernels->v->data.S32[index] = vOrder;
+                if (kernels->preCalc->data[index]) {
+                    psFree(kernels->preCalc->data[index]);
+                }
+                kernels->preCalc->data[index] = preCalc;
+                kernels->penalties->data.F32[index] = kernels->penalty * fabsf(moment);
+
+                psTrace("psModules.imcombine", 7, "Kernel %d: %f %d %d %f\n", index,
+                        fwhms->data.F32[i], uOrder, vOrder, fabsf(moment));
+            }
+        }
+    }
+
+    return kernels;
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Public functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+pmSubtractionKernels *pmSubtractionKernelsAlloc(int numBasisFunctions, pmSubtractionKernelsType type,
+                                                int size, int spatialOrder, float penalty,
+                                                pmSubtractionMode mode)
+{
+    pmSubtractionKernels *kernels = psAlloc(sizeof(pmSubtractionKernels)); // Kernels, to return
+    psMemSetDeallocator(kernels, (psFreeFunc)subtractionKernelsFree);
+
+    kernels->type = type;
+    kernels->description = NULL;
+    kernels->num = numBasisFunctions;
+    kernels->u = psVectorAlloc(numBasisFunctions, PS_TYPE_S32);
+    kernels->v = psVectorAlloc(numBasisFunctions, PS_TYPE_S32);
+    kernels->widths = psVectorAlloc(numBasisFunctions, PS_TYPE_F32);
+    kernels->preCalc = psArrayAlloc(numBasisFunctions);
+    kernels->penalty = penalty;
+    kernels->penalties = psVectorAlloc(numBasisFunctions, PS_TYPE_F32);
+    kernels->uStop = NULL;
+    kernels->vStop = NULL;
+    kernels->size = size;
+    kernels->inner = 0;
+    kernels->spatialOrder = spatialOrder;
+    kernels->bgOrder = 0;
+    kernels->mode = mode;
+    kernels->numCols = 0;
+    kernels->numRows = 0;
+    kernels->solution1 = NULL;
+    kernels->solution2 = NULL;
+
+    return kernels;
+}
+
+pmSubtractionKernels *pmSubtractionKernelsPOIS(int size, int spatialOrder, float penalty,
+                                               pmSubtractionMode mode)
+{
+    PS_ASSERT_INT_POSITIVE(size, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(spatialOrder, NULL);
+
+    int num = PS_SQR(2 * size + 1) - 1; // Number of basis functions
+
+    pmSubtractionKernels *kernels = pmSubtractionKernelsAlloc(num, PM_SUBTRACTION_KERNEL_POIS, size,
+                                                              spatialOrder, penalty, mode); // The kernels
+    psStringAppend(&kernels->description, "POIS(%d,%d,%.2e)", size, spatialOrder, penalty);
+    psLogMsg("psModules.imcombine", PS_LOG_INFO, "POIS kernel: %d,%d --> %d elements",
+             size, spatialOrder, num);
+
+    if (!p_pmSubtractionKernelsAddGrid(kernels, 0, size)) {
+        psAbort("Should never get here.");
+    }
+
+    return kernels;
+}
+
+
+pmSubtractionKernels *pmSubtractionKernelsISIS(int size, int spatialOrder,
+                                               const psVector *fwhms, const psVector *orders,
+                                               float penalty, pmSubtractionMode mode)
+{
+    pmSubtractionKernels *kernels = p_pmSubtractionKernelsRawISIS(size, spatialOrder, fwhms, orders,
+                                                                  penalty, mode); // Kernels
+    if (!kernels) {
+        return NULL;
+    }
+
+    if (psTraceGetLevel("psModules.imcombine.kernel") >= 10) {
+        for (int i = 0; i < kernels->num; i++) {
+            psKernel *kernel = kernels->preCalc->data[i]; // Kernel of interest
+            psString kernelName = NULL;
+            psStringAppend(&kernelName, "kernel%03d.fits", i);
+            psFits *kernelFile = psFitsOpen(kernelName, "w");
+            psFree(kernelName);
+            psFitsWriteImage(kernelFile, NULL, kernel->image, 0, NULL);
+            psFitsClose(kernelFile);
+            double sum = 0.0;
+            for (int y = 0; y < kernel->image->numRows; y++) {
+                for (int x = 0; x < kernel->image->numCols; x++) {
+                    sum += kernel->image->data.F32[y][x];
+                }
+            }
+            psTrace("psModules.imcombine.kernel", 10, "Kernel %d sum: %le\n", i, sum);
+        }
+    }
+
+    return kernels;
+}
+
+pmSubtractionKernels *pmSubtractionKernelsSPAM(int size, int spatialOrder, int inner, int binning,
+                                               float penalty, pmSubtractionMode mode)
+{
+    PS_ASSERT_INT_POSITIVE(size, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(spatialOrder, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(inner, NULL);
+    PS_ASSERT_INT_LARGER_THAN(size, inner, NULL);
+    PS_ASSERT_INT_POSITIVE(binning, NULL);
+
+    // The outer region should be divisible by the "binning"; otherwise allocate remainder to the inner region
+    int numOuter = (size - inner) / binning; // Number of summed pixels in the outer region
+    int numInner = inner + (size - inner) % binning; // Number of pixels in the inner region
+    assert(numOuter * binning + numInner == size);
+    int numTotal = numOuter + numInner; // Total number of summed pixels
+
+    psTrace("psModules.imcombine", 3, "Inner: %d Outer: %d\n", numInner, numOuter);
+
+    int num = PS_SQR(2 * numTotal + 1) - 1; // Number of basis functions
+
+    psTrace("psModules.imcombine", 3, "Number of basis functions: %d\n", num);
+
+    pmSubtractionKernels *kernels = pmSubtractionKernelsAlloc(num, PM_SUBTRACTION_KERNEL_SPAM, size,
+                                                              spatialOrder, penalty, mode); // The kernels
+    kernels->inner = inner;
+    psStringAppend(&kernels->description, "SPAM(%d,%d,%d,%d,%.2e)", size, inner, binning, spatialOrder,
+                   penalty);
+
+    psLogMsg("psModules.imcombine", PS_LOG_INFO, "SPAM kernel: %d,%d,%d,%d --> %d elements",
+             size, inner, binning, spatialOrder, num);
+
+    kernels->uStop = psVectorAlloc(num, PS_TYPE_S32);
+    kernels->vStop = psVectorAlloc(num, PS_TYPE_S32);
+
+    psVector *locations = psVectorAlloc(2 * numTotal + 1, PS_TYPE_S32); // Locations for each kernel element
+    psVector *widths = psVectorAlloc(2 * numTotal + 1, PS_TYPE_S32); // Widths for each kernel element
+    locations->data.S32[numTotal] = 0;
+    widths->data.S32[numTotal] = 0;
+    for (int i = 1; i <= numInner; i++) {
+        locations->data.S32[numTotal + i] = i;
+        widths->data.S32[numTotal + i] = 0;
+        locations->data.S32[numTotal - i] = - i;
+        widths->data.S32[numTotal - i] = 0;
+    }
+    for (int i = numInner + 1; i <= numTotal; i++) {
+        locations->data.S32[numTotal + i] = locations->data.S32[numTotal + i - 1] +
+            widths->data.S32[numTotal + i - 1] + 1;
+        widths->data.S32[numTotal + i] = binning - 1;
+        locations->data.S32[numTotal - i] = locations->data.S32[numTotal - i + 1] - binning;
+        widths->data.S32[numTotal - i] = binning - 1;
+    }
+
+    if (psTraceGetLevel("psModules.imcombine") >= 10) {
+        for (int i = 0; i < 2 * numTotal + 1; i++) {
+            psTrace("psModules.imcombine", 10, "%d: %d -> %d\n", i, locations->data.S32[i],
+                    locations->data.S32[i] + widths->data.S32[i]);
+        }
+    }
+
+    // Set the kernel parameters
+    for (int i = - numTotal, index = 0; i <= numTotal; i++) {
+        int u = locations->data.S32[numTotal + i]; // Location of pixel
+        int uStop = u + widths->data.S32[numTotal + i]; // Width of pixel
+
+        for (int j = - numTotal; j <= numTotal; j++, index++) {
+            if (i == 0 && j == 0) {
+                // Skip normalisation component: added explicitly
+                index--;
+                continue;
+            }
+            int v = locations->data.S32[numTotal + j]; // Location of pixel
+            int vStop = v + widths->data.S32[numTotal + j]; // Width of pixel
+
+            kernels->u->data.S32[index] = u;
+            kernels->v->data.S32[index] = v;
+            kernels->uStop->data.S32[index] = uStop;
+            kernels->vStop->data.S32[index] = vStop;
+
+            psTrace("psModules.imcombine", 7, "Kernel %d: %d %d %d %d\n", index,
+                    u, uStop, v, vStop);
+        }
+    }
+
+    psFree(locations);
+    psFree(widths);
+
+    psWarning("Kernel penalty for dual-convolution is not configured for SPAM kernels.");
+    psVectorInit(kernels->penalties, 0.0);
+
+    return kernels;
+}
+
+
+pmSubtractionKernels *pmSubtractionKernelsFRIES(int size, int spatialOrder, int inner, float penalty,
+                                                pmSubtractionMode mode)
+{
+    PS_ASSERT_INT_POSITIVE(size, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(spatialOrder, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(inner, NULL);
+    PS_ASSERT_INT_LARGER_THAN(size, inner, NULL);
+
+    int fibNum = 0;                     // Number of Fibonacci values
+    int fibLast = 1, fibTotal = 2;      // Fibonacci sequence
+    while (fibTotal < size - inner) {
+        int temp = fibTotal;
+        fibTotal += fibLast;
+        fibLast = temp;
+        fibNum++;
+    }
+
+    int numInner = inner;               // Number of pixels in the inner region
+    int numOuter = fibNum;              // Number of summed pixels in the outer region
+    int numTotal = numOuter + numInner; // Total number of summed pixels
+
+    psTrace("psModules.imcombine", 3, "Inner: %d Outer: %d\n", numInner, numOuter);
+
+    int num = PS_SQR(2 * numTotal + 1) - 1; // Number of basis functions
+
+    psTrace("psModules.imcombine", 3, "Number of basis functions: %d\n", num);
+
+    pmSubtractionKernels *kernels = pmSubtractionKernelsAlloc(num, PM_SUBTRACTION_KERNEL_FRIES, size,
+                                                              spatialOrder, penalty, mode); // The kernels
+    kernels->inner = inner;
+    psStringAppend(&kernels->description, "FRIES(%d,%d,%d,%.2e)", size, inner, spatialOrder, penalty);
+
+    psLogMsg("psModules.imcombine", PS_LOG_INFO, "FRIES kernel: %d,%d,%d --> %d elements",
+             size, inner, spatialOrder, num);
+
+    kernels->uStop = psVectorAlloc(num, PS_TYPE_S32);
+    kernels->vStop = psVectorAlloc(num, PS_TYPE_S32);
+
+    psVector *start = psVectorAlloc(2 * numTotal + 1, PS_TYPE_S32);
+    psVector *stop = psVectorAlloc(2 * numTotal + 1, PS_TYPE_S32);
+    start->data.S32[numTotal] = 0;
+    stop->data.S32[numTotal] = 0;
+    for (int i = 1; i <= numInner; i++) {
+        start->data.S32[numTotal + i] = i;
+        stop->data.S32[numTotal + i] = i;
+        start->data.S32[numTotal - i] = -i;
+        stop->data.S32[numTotal - i] = -i;
+    }
+    for (int i = numInner + 1, fibLast = 1, fib = 2, temp; i <= numTotal;
+         i++, fib = (temp = fib) + fibLast, fibLast = temp) {
+        start->data.S32[numTotal + i] = stop->data.S32[numTotal + i - 1] + 1;
+        stop->data.S32[numTotal + i] = PS_MIN(start->data.S32[numTotal + i] + fib - 1, size);
+        start->data.S32[numTotal - i] = - stop->data.S32[numTotal + i];
+        stop->data.S32[numTotal - i] = - start->data.S32[numTotal + i];
+    }
+
+    if (psTraceGetLevel("psModules.imcombine") >= 10) {
+        for (int i = 0; i < 2 * numTotal + 1; i++) {
+            psTrace("psModules.imcombine", 10, "%d: %d -> %d\n", i, start->data.S32[i], stop->data.S32[i]);
+        }
+    }
+
+    // Set the kernel parameters
+    for (int i = - numTotal, index = 0; i <= numTotal; i++) {
+        int u = start->data.S32[numTotal + i]; // Location of pixel
+        int uStop = stop->data.S32[numTotal + i]; // Width of pixel
+        for (int j = - numTotal; j <= numTotal; j++, index++) {
+            if (i == 0 && j == 0) {
+                // Skip normalisation component: added explicitly
+                index--;
+                continue;
+            }
+            int v = start->data.S32[numTotal + j]; // Location of pixel
+            int vStop = stop->data.S32[numTotal + j]; // Width of pixel
+
+            kernels->u->data.S32[index] = u;
+            kernels->v->data.S32[index] = v;
+            kernels->uStop->data.S32[index] = uStop;
+            kernels->vStop->data.S32[index] = vStop;
+
+            psTrace("psModules.imcombine", 7, "Kernel %d: %d %d %d %d\n", index,
+                    u, uStop, v, vStop);
+        }
+    }
+
+    psFree(start);
+    psFree(stop);
+
+    psWarning("Kernel penalty for dual-convolution is not configured for FRIES kernels.");
+    psVectorInit(kernels->penalties, 0.0);
+
+    return kernels;
+}
+
+// Grid United with Normal Kernel
+pmSubtractionKernels *pmSubtractionKernelsGUNK(int size, int spatialOrder, const psVector *fwhms,
+                                               const psVector *orders, int inner, float penalty,
+                                               pmSubtractionMode mode)
+{
+    PS_ASSERT_INT_POSITIVE(size, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(spatialOrder, NULL);
+    PS_ASSERT_VECTOR_NON_NULL(fwhms, NULL);
+    PS_ASSERT_VECTOR_TYPE(fwhms, PS_TYPE_F32, NULL);
+    PS_ASSERT_VECTOR_NON_NULL(orders, NULL);
+    PS_ASSERT_VECTOR_TYPE(orders, PS_TYPE_S32, NULL);
+    PS_ASSERT_VECTORS_SIZE_EQUAL(fwhms, orders, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(inner, NULL);
+    PS_ASSERT_INT_LESS_THAN(inner, size, NULL);
+
+    // XXX GUNK doesn't seem to work --- doesn't add the POIS components, or at least, they're not noticed
+
+    pmSubtractionKernels *kernels = p_pmSubtractionKernelsRawISIS(size, spatialOrder, fwhms, orders,
+                                                                  penalty, mode); // Kernels
+    psStringPrepend(&kernels->description, "GUNK=");
+    psStringAppend(&kernels->description, "+POIS(%d,%d)", inner, spatialOrder);
+
+    int numISIS = kernels->num;         // Number of ISIS kernels
+
+    if (!p_pmSubtractionKernelsAddGrid(kernels, numISIS, inner)) {
+        psAbort("Should never get here.");
+    }
+
+    return kernels;
+}
+
+// RINGS --- just what it says
+pmSubtractionKernels *pmSubtractionKernelsRINGS(int size, int spatialOrder, int inner, int ringsOrder,
+                                                float penalty, pmSubtractionMode mode)
+{
+    PS_ASSERT_INT_POSITIVE(size, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(spatialOrder, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(inner, NULL);
+    PS_ASSERT_INT_LESS_THAN_OR_EQUAL(inner, size, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(ringsOrder, NULL);
+
+    int fibNum = 0;                     // Number of Fibonacci values
+    {
+        int fibIndex = 1, fibIndexMinus1 = 0; // Fibonnacci parameters
+        int radius = inner;
+        while (radius + fibIndex < size) {
+            radius++;
+            int fibNew = fibIndex + fibIndexMinus1;
+            fibIndexMinus1 = fibIndex;
+            fibIndex = fibNew;
+            radius += fibIndex;
+            fibNum++;
+        }
+    }
+
+    int numInner = inner - 1;           // Number of pixels in the inner region
+    int numOuter = fibNum;              // Number of summed pixels in the outer region
+
+    int numRings = numOuter + numInner; // Number of rings (not including the central pixel)
+    int numPoly = PM_SUBTRACTION_POLYTERMS(ringsOrder); // Number of polynomial variants of each ring
+
+    int num = numRings * numPoly; // Total number of basis functions
+
+    pmSubtractionKernels *kernels = pmSubtractionKernelsAlloc(num, PM_SUBTRACTION_KERNEL_RINGS, size,
+                                                              spatialOrder, penalty, mode); // The kernels
+    kernels->inner = inner;
+    psStringAppend(&kernels->description, "RINGS(%d,%d,%d,%d,%.2e)", size, inner, ringsOrder, spatialOrder,
+                   penalty);
+
+    psLogMsg("psModules.imcombine", PS_LOG_INFO, "RINGS kernel: %d,%d,%d,%d --> %d elements",
+             size, inner, ringsOrder, spatialOrder, num);
+
+    // Set the Gaussian kernel parameters
+    int fibIndex = 1, fibIndexMinus1 = 0; // Fibonnacci parameters
+    int radiusLast = 1;                 // Last radius
+    for (int i = 1, index = 0; i < numRings + 1; i++) {
+        float lower2;                   // Lower limit of radius^2
+        float upper2;                   // Upper limit of radius^2
+        if (i <= inner) {
+            // A ring every pixel width
+            float radius = i;
+            lower2 = PS_SQR(radius - 0.5);
+            upper2 = PS_SQR(radius + 0.5);
+            radiusLast = i;
+        } else {
+            // Rings Fibonacci distributed (2, 3, 5...)
+            int fibNew = fibIndex + fibIndexMinus1;
+            fibIndexMinus1 = fibIndex;
+            fibIndex = fibNew;
+
+            float radiusLower = radiusLast + 1;
+            radiusLast = radiusLower + fibIndex;
+            float radiusUpper = radiusLast;
+
+            lower2 = PS_SQR(radiusLower - 0.5);
+            upper2 = PS_SQR(radiusUpper + 0.5);
+        }
+
+        psTrace("psModules.imcombine", 8, "Radius limits: %f --> %f\n", sqrtf(lower2), sqrtf(upper2));
+
+        // Iterate over (u,v) order
+        for (int uOrder = 0; uOrder <= (i == 0 ? 0 : ringsOrder); uOrder++) {
+            for (int vOrder = 0; vOrder <= (i == 0 ? 0 : ringsOrder - uOrder); vOrder++, index++) {
+
+                psArray *data = psArrayAlloc(3); // Container for data
+                psVector *uCoords = data->data[0] = psVectorAllocEmpty(RINGS_BUFFER, PS_TYPE_S32); // u coords
+                psVector *vCoords = data->data[1] = psVectorAllocEmpty(RINGS_BUFFER, PS_TYPE_S32); // v coords
+                psVector *poly = data->data[2] = psVectorAllocEmpty(RINGS_BUFFER, PS_TYPE_F32); // Polynomial
+                double moment = 0.0;    // Moment, for penalty
+
+                if (i == 0) {
+                    // Central pixel is easy
+                    uCoords->data.S32[0] = vCoords->data.S32[0] = 0;
+                    poly->data.F32[0] = 1.0;
+                    uCoords->n = vCoords->n = poly->n = 1;
+                    radiusLast = 0;
+                    moment = 0.0;
+                } else {
+                    int j = 0;          // Index for data
+                    double norm = 0.0;  // Normalisation
+                    for (int v = -size; v <= size; v++) {
+                        int v2 = PS_SQR(v);   // Square of v
+                        float vPoly = powf(v/(float)size, vOrder); // Value of v^vOrder
+
+                        for (int u = -size; u <= size; u++) {
+                            int u2 = PS_SQR(u); // Square of u
+                            int distance2 = u2 + v2; // Distance from the centre
+                            if (distance2 > lower2 && distance2 < upper2) {
+                                float uPoly = powf(u/(float)size, uOrder); // Value of u^uOrder
+
+                                float polyVal = uPoly * vPoly; // Value of polynomial
+                                if (polyVal != 0) { // No point adding it otherwise
+                                    uCoords->data.S32[j] = u;
+                                    vCoords->data.S32[j] = v;
+                                    poly->data.F32[j] = polyVal;
+                                    norm += polyVal;
+                                    moment += polyVal * (PS_SQR(u) + PS_SQR(v));
+
+                                    psVectorExtend(uCoords, RINGS_BUFFER, 1);
+                                    psVectorExtend(vCoords, RINGS_BUFFER, 1);
+                                    psVectorExtend(poly, RINGS_BUFFER, 1);
+                                    psTrace("psModules.imcombine", 9, "u = %d, v = %d, poly = %f\n",
+                                            u, v, poly->data.F32[j]);
+                                    j++;
+                                }
+                            }
+                        }
+                    }
+                    // Normalise kernel component to unit sum
+                    if (uOrder % 2 == 0 && vOrder % 2 == 0) {
+                        psBinaryOp(poly, poly, "*", psScalarAlloc(1.0 / norm, PS_TYPE_F32));
+                        // Add subtraction of 0,0 component to preserve photometric scaling
+                        uCoords->data.S32[j] = 0;
+                        vCoords->data.S32[j] = 0;
+                        poly->data.F32[j] = -1.0;
+                        psVectorExtend(uCoords, RINGS_BUFFER, 1);
+                        psVectorExtend(vCoords, RINGS_BUFFER, 1);
+                        psVectorExtend(poly, RINGS_BUFFER, 1);
+                    } else {
+                        norm = powf(size, uOrder) * powf(size, vOrder);
+                        psBinaryOp(poly, poly, "*", psScalarAlloc(1.0 / norm, PS_TYPE_F32));
+                    }
+//                    moment /= norm;
+                }
+
+                psTrace("psModules.imcombine", 8, "%ld pixels in kernel\n", uCoords->n);
+
+                kernels->preCalc->data[index] = data;
+                kernels->u->data.S32[index] = uOrder;
+                kernels->v->data.S32[index] = vOrder;
+                kernels->penalties->data.F32[index] = kernels->penalty * fabsf(moment);
+
+                psTrace("psModules.imcombine", 7, "Kernel %d: %d %d %d\n", index,
+                        i, uOrder, vOrder);
+            }
+        }
+    }
+
+    return kernels;
+}
+
+pmSubtractionKernels *pmSubtractionKernelsGenerate(pmSubtractionKernelsType type, int size, int spatialOrder,
+                                                   const psVector *fwhms, const psVector *orders, int inner,
+                                                   int binning, int ringsOrder, float penalty,
+                                                   pmSubtractionMode mode)
+{
+    switch (type) {
+      case PM_SUBTRACTION_KERNEL_POIS:
+        return pmSubtractionKernelsPOIS(size, spatialOrder, penalty, mode);
+      case PM_SUBTRACTION_KERNEL_ISIS:
+        return pmSubtractionKernelsISIS(size, spatialOrder, fwhms, orders, penalty, mode);
+      case PM_SUBTRACTION_KERNEL_SPAM:
+        return pmSubtractionKernelsSPAM(size, spatialOrder, inner, binning, penalty, mode);
+      case PM_SUBTRACTION_KERNEL_FRIES:
+        return pmSubtractionKernelsFRIES(size, spatialOrder, inner, penalty, mode);
+      case PM_SUBTRACTION_KERNEL_GUNK:
+        return pmSubtractionKernelsGUNK(size, spatialOrder, fwhms, orders, inner, penalty, mode);
+      case PM_SUBTRACTION_KERNEL_RINGS:
+        return pmSubtractionKernelsRINGS(size, spatialOrder, inner, ringsOrder, penalty, mode);
+      default:
+        psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Unknown kernel type: %x", type);
+        return NULL;
+    }
+}
+
+
+// Intermediate string parsing functions required because of different APIs for strtol and strtof
+static inline int parseStringInt(const char *string)
+{
+    return strtol(string, NULL, 10);
+}
+static inline float parseStringFloat(const char *string)
+{
+    return strtof(string, NULL);
+}
+
+
+// Parse a string of a number, up to some delimiter, and advance past the delimiter
+#define PARSE_STRING_NUMBER(TARGET, STRING, DELIM, PARSEFUNC) { \
+    char *start = STRING;               /* Start of string */ \
+    char *end = strchr(STRING, DELIM);  /* End of string */ \
+    if (!end) { \
+        psAbort("End of string encountered"); \
+    } \
+    int stringSize = end - STRING;      /* Size of string with value, NOT including \0 */ \
+    char value[stringSize + 1];         /* String to parse */ \
+    strncpy(value, start, stringSize); \
+    value[stringSize] = '\0'; \
+    TARGET = PARSEFUNC(value); \
+    STRING += stringSize + 1;           /* Advance past delimiter */ \
+}
+
+
+pmSubtractionKernels *pmSubtractionKernelsFromDescription(const char *description, int bgOrder,
+                                                          pmSubtractionMode mode)
+{
+    PS_ASSERT_STRING_NON_EMPTY(description, NULL);
+
+    if (bgOrder != 0) {
+        psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Background order %d is not yet supported.", bgOrder);
+        return false;
+    }
+
+    pmSubtractionKernelsType type = PM_SUBTRACTION_KERNEL_NONE; // Type of kernel
+    int size = 0;                       // Half-size of kernel
+    int spatialOrder = 0;               // Order of spatial variations
+    const psVector *fwhms = NULL;       // FWHM of Gaussians
+    const psVector *orders = NULL;      // Polynomial order for each FWHM
+    int inner = 0;                      // Size of inner region
+    int binning = 0;                    // Binning to use
+    int ringsOrder = 0;                 // Polynomial order for rings
+    float penalty = 0.0;                // Penalty for wideness
+
+    if (strncmp(description, "ISIS", 4) == 0) {
+        // XXX Support for GUNK
+        if (strstr(description, "+POIS")) {
+            type = PM_SUBTRACTION_KERNEL_GUNK;
+            psAbort("Deciphering GUNK kernels (%s) is not currently supported.", description);
+        } else {
+            type = PM_SUBTRACTION_KERNEL_ISIS;
+            char *ptr = (char*)description + 5;    // Eat "ISIS("
+            PARSE_STRING_NUMBER(size, ptr, ',', parseStringInt);
+
+            // Count the number of Gaussians
+            int numGauss = 0;
+            for (char *string = ptr; string; string = strchr(string, '(')) {
+                numGauss++;
+            }
+
+            fwhms = psVectorAlloc(numGauss, PS_TYPE_F32);
+            orders = psVectorAlloc(numGauss, PS_TYPE_S32);
+
+            for (int i = 0; i < numGauss; i++) {
+                ptr++;                  // Eat the '('
+                PARSE_STRING_NUMBER(fwhms->data.F32[i], ptr, ',', parseStringFloat); // Eat "1.234,"
+                PARSE_STRING_NUMBER(orders->data.S32[i], ptr, ')', parseStringInt); // Eat "3)"
+            }
+
+            ptr++;                      // Eat ','
+            PARSE_STRING_NUMBER(spatialOrder, ptr, ',', parseStringInt);
+            penalty = parseStringFloat(ptr);
+        }
+    } else if (strncmp(description, "RINGS", 5) == 0) {
+        type = PM_SUBTRACTION_KERNEL_RINGS;
+        char *ptr = (char*)description + 6;
+        PARSE_STRING_NUMBER(size, ptr, ',', parseStringInt);
+        PARSE_STRING_NUMBER(inner, ptr, ',', parseStringInt);
+        PARSE_STRING_NUMBER(ringsOrder, ptr, ',', parseStringInt);
+        PARSE_STRING_NUMBER(spatialOrder, ptr, ',', parseStringInt);
+        PARSE_STRING_NUMBER(penalty, ptr, ')', parseStringInt);
+    } else {
+        psAbort("Deciphering kernels other than ISIS and RINGS is not currently supported.");
+    }
+
+
+    return pmSubtractionKernelsGenerate(type, size, spatialOrder, fwhms, orders,
+                                        inner, binning, ringsOrder, penalty, mode);
+}
+
+
+pmSubtractionKernelsType pmSubtractionKernelsTypeFromString(const char *type)
+{
+    if (strcasecmp(type, "POIS") == 0) {
+        return PM_SUBTRACTION_KERNEL_POIS;
+    }
+    if (strcasecmp(type, "ISIS") == 0) {
+        return PM_SUBTRACTION_KERNEL_ISIS;
+    }
+    if (strcasecmp(type, "SPAM") == 0) {
+        return PM_SUBTRACTION_KERNEL_SPAM;
+    }
+    if (strcasecmp(type, "FRIES") == 0) {
+        return PM_SUBTRACTION_KERNEL_FRIES;
+    }
+    if (strcasecmp(type, "GUNK") == 0) {
+        return PM_SUBTRACTION_KERNEL_GUNK;
+    }
+    if (strcasecmp(type, "RINGS") == 0) {
+        return PM_SUBTRACTION_KERNEL_RINGS;
+    }
+
+    psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Unrecognised kernel type: %s", type);
+    return PM_SUBTRACTION_KERNEL_NONE;
+}
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionKernels.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionKernels.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionKernels.h	(revision 19481)
@@ -0,0 +1,202 @@
+#ifndef PM_SUBTRACTION_KERNELS_H
+#define PM_SUBTRACTION_KERNELS_H
+
+#include <pslib.h>
+
+/// Type of subtraction kernel
+typedef enum {
+    PM_SUBTRACTION_KERNEL_NONE,         ///< Nothing --- an error
+    PM_SUBTRACTION_KERNEL_POIS,         ///< Pan-STARRS Optimal Image Subtraction --- delta functions
+    PM_SUBTRACTION_KERNEL_ISIS,         ///< Traditional kernel --- gaussians modified by polynomials
+    PM_SUBTRACTION_KERNEL_SPAM,         ///< Summed Pixels for Advanced Matching --- summed delta functions
+    PM_SUBTRACTION_KERNEL_FRIES,        ///< Fibonacci Radius Increases Excellence of Subtraction
+    PM_SUBTRACTION_KERNEL_GUNK,         ///< Grid United with Normal Kernel --- POIS and ISIS hybrid
+    PM_SUBTRACTION_KERNEL_RINGS,        ///< Rings Instead of the Normal Gaussian Subtraction
+} pmSubtractionKernelsType;
+
+/// Modes --- specifies which image to convolve
+typedef enum {
+    PM_SUBTRACTION_MODE_ERR,            // Error in the mode
+    PM_SUBTRACTION_MODE_1,              // Convolve image 1
+    PM_SUBTRACTION_MODE_2,              // Convolve image 2
+    PM_SUBTRACTION_MODE_UNSURE,         // Not sure yet which image to convolve so try to satisfy both
+    PM_SUBTRACTION_MODE_DUAL,           // Dual convolution
+} pmSubtractionMode;
+
+/// Kernels specification
+typedef struct {
+    pmSubtractionKernelsType type;      ///< Type of kernels --- allowing the use of multiple kernels
+    psString description;               ///< Description of the kernel parameters
+    long num;                           ///< Number of kernel components (not including the spatial ones)
+    psVector *u, *v;                    ///< Offset (for POIS) or polynomial order (for ISIS)
+    psVector *widths;                   ///< Gaussian FWHMs (ISIS)
+    psVector *uStop, *vStop;            ///< Width of kernel element (SPAM,FRIES only)
+    psArray *preCalc;                   ///< Array of images containing pre-calculated kernel (for ISIS)
+    float penalty;                      ///< Penalty for wideness
+    psVector *penalties;                ///< Penalty for each kernel component
+    int size;                           ///< The half-size of the kernel
+    int inner;                          ///< The size of an inner region
+    int spatialOrder;                   ///< The spatial order of the kernels
+    int bgOrder;                        ///< The order for the background fitting
+    pmSubtractionMode mode;             ///< Mode for subtraction
+    int numCols, numRows;               ///< Size of image (for normalisation), or zero to use image provided
+    psVector *solution1, *solution2;    ///< Solution for the PSF matching
+} pmSubtractionKernels;
+
+// Assertion to check pmSubtractionKernels
+#define PM_ASSERT_SUBTRACTION_KERNELS_NON_NULL(KERNELS, RETURNVALUE) { \
+    PS_ASSERT_PTR_NON_NULL(KERNELS, RETURNVALUE); \
+    PS_ASSERT_STRING_NON_EMPTY((KERNELS)->description, RETURNVALUE); \
+    PS_ASSERT_INT_POSITIVE((KERNELS)->num, RETURNVALUE); \
+    PS_ASSERT_VECTOR_NON_NULL((KERNELS)->u, RETURNVALUE); \
+    PS_ASSERT_VECTOR_NON_NULL((KERNELS)->v, RETURNVALUE); \
+    PS_ASSERT_VECTOR_TYPE((KERNELS)->u, PS_TYPE_S32, RETURNVALUE); \
+    PS_ASSERT_VECTOR_TYPE((KERNELS)->v, PS_TYPE_S32, RETURNVALUE); \
+    PS_ASSERT_VECTOR_SIZE((KERNELS)->u, (KERNELS)->num, RETURNVALUE); \
+    PS_ASSERT_VECTOR_SIZE((KERNELS)->v, (KERNELS)->num, RETURNVALUE); \
+    if ((KERNELS)->type == PM_SUBTRACTION_KERNEL_ISIS) { \
+        PS_ASSERT_VECTOR_NON_NULL((KERNELS)->widths, RETURNVALUE); \
+        PS_ASSERT_VECTOR_TYPE((KERNELS)->widths, PS_TYPE_F32, RETURNVALUE); \
+        PS_ASSERT_VECTOR_SIZE((KERNELS)->widths, (KERNELS)->num, RETURNVALUE); \
+    } \
+    if ((KERNELS)->uStop || (KERNELS)->vStop) { \
+        PS_ASSERT_VECTOR_NON_NULL((KERNELS)->uStop, RETURNVALUE); \
+        PS_ASSERT_VECTOR_NON_NULL((KERNELS)->vStop, RETURNVALUE); \
+        PS_ASSERT_VECTOR_TYPE((KERNELS)->uStop, PS_TYPE_S32, RETURNVALUE); \
+        PS_ASSERT_VECTOR_TYPE((KERNELS)->vStop, PS_TYPE_S32, RETURNVALUE); \
+        PS_ASSERT_VECTOR_SIZE((KERNELS)->uStop, (KERNELS)->num, RETURNVALUE); \
+        PS_ASSERT_VECTOR_SIZE((KERNELS)->vStop, (KERNELS)->num, RETURNVALUE); \
+    } \
+    if ((KERNELS)->preCalc) { \
+        PS_ASSERT_ARRAY_NON_NULL((KERNELS)->preCalc, RETURNVALUE); \
+        PS_ASSERT_ARRAY_SIZE((KERNELS)->preCalc, (KERNELS)->num, RETURNVALUE); \
+    } \
+    PS_ASSERT_INT_NONNEGATIVE((KERNELS)->size, RETURNVALUE); \
+    PS_ASSERT_INT_NONNEGATIVE((KERNELS)->inner, RETURNVALUE); \
+    PS_ASSERT_INT_NONNEGATIVE((KERNELS)->spatialOrder, RETURNVALUE); \
+    PS_ASSERT_INT_NONNEGATIVE((KERNELS)->bgOrder, RETURNVALUE); \
+}
+
+// Assertion to check that the solution is attached
+#define PM_ASSERT_SUBTRACTION_KERNELS_SOLUTION(KERNELS, RETURNVALUE) { \
+    PS_ASSERT_VECTOR_NON_NULL((KERNELS)->solution1, RETURNVALUE); \
+    PS_ASSERT_VECTOR_TYPE((KERNELS)->solution1, PS_TYPE_F64, RETURNVALUE); \
+    PS_ASSERT_VECTOR_SIZE((KERNELS)->solution1, \
+                          (KERNELS)->num * PM_SUBTRACTION_POLYTERMS((KERNELS)->spatialOrder) + 1 + \
+                              PM_SUBTRACTION_POLYTERMS((KERNELS)->bgOrder), \
+                          RETURNVALUE); \
+    if (kernels->mode == PM_SUBTRACTION_MODE_DUAL) { \
+        PS_ASSERT_VECTOR_NON_NULL(kernels->solution2, RETURNVALUE); \
+        PS_ASSERT_VECTOR_TYPE((KERNELS)->solution2, PS_TYPE_F64, RETURNVALUE); \
+        PS_ASSERT_VECTOR_SIZE((KERNELS)->solution2, \
+                              (KERNELS)->num * PM_SUBTRACTION_POLYTERMS((KERNELS)->spatialOrder), \
+                               RETURNVALUE); \
+    } \
+}
+
+/// Generate a delta-function grid for subtraction kernels (like the POIS kernel)
+bool p_pmSubtractionKernelsAddGrid(pmSubtractionKernels *kernels, ///< The subtraction kernels to append to
+                                   int start, ///< Index at which to start appending
+                                   int size ///< Half-size of the grid
+    );
+
+/// General allocator for pmSubtractionKernels
+///
+/// Unlike the functions for the specific kernel type, this function does not set up the basis functions, but
+/// merely allocates space for their storage.
+pmSubtractionKernels *pmSubtractionKernelsAlloc(int numBasisFunctions, ///< Number of basis functions
+                                                pmSubtractionKernelsType type, ///< Kernel type
+                                                int size, ///< Half-size of kernel
+                                                int spatialOrder, ///< Order of spatial variations
+                                                float penalty, ///< Penalty for wideness
+                                                pmSubtractionMode mode ///< Mode for subtraction
+    );
+
+/// Generate POIS kernels
+pmSubtractionKernels *pmSubtractionKernelsPOIS(int size, ///< Half-size of the kernel (in both dims)
+                                               int spatialOrder, ///< Order of spatial variations
+                                               float penalty, ///< Penalty for wideness
+                                               pmSubtractionMode mode ///< Mode for subtraction
+    );
+
+/// Generate ISIS kernels without the flux scaling built in
+pmSubtractionKernels *p_pmSubtractionKernelsRawISIS(int size, ///< Half-size of the kernel
+                                                    int spatialOrder, ///< Order of spatial variations
+                                                    const psVector *fwhms, ///< Gaussian FWHMs
+                                                    const psVector *orders, ///< Polynomial order of gaussians
+                                                    float penalty, ///< Penalty for wideness
+                                                    pmSubtractionMode mode ///< Mode for subtraction
+    );
+
+/// Generate ISIS kernels
+pmSubtractionKernels *pmSubtractionKernelsISIS(int size, ///< Half-size of the kernel
+                                               int spatialOrder, ///< Order of spatial variations
+                                               const psVector *fwhms, ///< Gaussian FWHMs
+                                               const psVector *orders, ///< Polynomial order of gaussians
+                                               float penalty, ///< Penalty for wideness
+                                               pmSubtractionMode mode ///< Mode for subtraction
+                                               );
+
+/// Generate SPAM kernels
+pmSubtractionKernels *pmSubtractionKernelsSPAM(int size, ///< Half-size of the kernel
+                                               int spatialOrder, ///< Order of spatial variations
+                                               int inner, ///< Inner radius to preserve unbinned
+                                               int binning, ///< Kernel binning factor
+                                               float penalty, ///< Penalty for wideness
+                                               pmSubtractionMode mode ///< Mode for subtraction
+    );
+
+/// Generate FRIES kernels
+pmSubtractionKernels *pmSubtractionKernelsFRIES(int size, ///< Half-size of the kernel
+                                                int spatialOrder, ///< Order of spatial variations
+                                                int inner, ///< Inner radius to preserve unbinned
+                                                float penalty, ///< Penalty for wideness
+                                                pmSubtractionMode mode ///< Mode for subtraction
+    );
+
+/// Generate GUNK kernels
+pmSubtractionKernels *pmSubtractionKernelsGUNK(int size, ///< Half-size of the kernel
+                                               int spatialOrder, ///< Order of spatial variations
+                                               const psVector *fwhms, ///< Gaussian FWHMs
+                                               const psVector *orders, ///< Polynomial order of gaussians
+                                               int inner, ///< Inner radius containing grid of delta functions
+                                               float penalty, ///< Penalty for wideness
+                                               pmSubtractionMode mode ///< Mode for subtraction
+    );
+
+/// Generate RINGS kernels
+pmSubtractionKernels *pmSubtractionKernelsRINGS(int size, ///< Half-size of the kernel
+                                                int spatialOrder, ///< Order of spatial variations
+                                                int inner, ///< Inner radius to preserve unbinned
+                                                int ringsOrder, ///< Polynomial order
+                                                float penalty, ///< Penalty for wideness
+                                                pmSubtractionMode mode ///< Mode for subtraction
+    );
+
+
+/// Generate a kernel of a specified type
+pmSubtractionKernels *pmSubtractionKernelsGenerate(pmSubtractionKernelsType type, ///< Kernel type
+                                                   int size, ///< Half-size of the kernel
+                                                   int spatialOrder, ///< Order of spatial variations
+                                                   const psVector *fwhms, ///< Gaussian FWHMs
+                                                   const psVector *orders, ///< Polynomial order of gaussians
+                                                   int inner, ///< Inner radius to preserve unbinned
+                                                   int binning, ///< Kernel binning factor
+                                                   int ringsOrder, ///< Polynomial order for RINGS
+                                                   float penalty, ///< Penalty for wideness
+                                                   pmSubtractionMode mode ///< Mode for subtraction
+    );
+
+/// Generate a kernel using the description
+pmSubtractionKernels *pmSubtractionKernelsFromDescription(
+    const char *description,            ///< Description of kernel
+    int bgOrder,                        ///< Polynomial order for background fitting
+    pmSubtractionMode mode              ///< Mode for subtraction
+    );
+
+/// Return the appropriate type from a string
+pmSubtractionKernelsType pmSubtractionKernelsTypeFromString(const char *string // String name for kernel type
+    );
+
+
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionMask.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionMask.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionMask.c	(revision 19481)
@@ -0,0 +1,245 @@
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <stdio.h>
+#include <pslib.h>
+
+#include "pmSubtraction.h"
+#include "pmSubtractionKernels.h"
+
+#include "pmSubtractionMask.h"
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Private (file-static) functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+// Mark a pixel as blank in the image, mask and weight
+static inline void markBlank(psImage *image, // Image to mark as blank
+                             psImage *mask, // Mask to mark as blank (or NULL)
+                             psImage *weight, // Weight map to mark as blank (or NULL)
+                             int x, int y, // Coordinates to mark blank
+                             psMaskType blank // Blank mask value
+    )
+{
+    image->data.F32[y][x] = NAN;
+    if (mask) {
+        mask->data.PS_TYPE_MASK_DATA[y][x] |= blank;
+    }
+    if (weight) {
+        weight->data.F32[y][x] = NAN;
+    }
+    return;
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Public functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+psImage *pmSubtractionMask(const psImage *mask1, const psImage *mask2, psMaskType maskVal,
+                           int size, int footprint, float badFrac, bool useFFT)
+{
+    PS_ASSERT_IMAGE_NON_NULL(mask1, NULL);
+    PS_ASSERT_IMAGE_TYPE(mask1, PS_TYPE_MASK, NULL);
+    if (mask2) {
+        PS_ASSERT_IMAGE_NON_NULL(mask2, NULL);
+        PS_ASSERT_IMAGE_TYPE(mask2, PS_TYPE_MASK, NULL);
+        PS_ASSERT_IMAGES_SIZE_EQUAL(mask2, mask1, NULL);
+    }
+    PS_ASSERT_INT_NONNEGATIVE(size, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(footprint, NULL);
+    if (isfinite(badFrac)) {
+        PS_ASSERT_FLOAT_LARGER_THAN(badFrac, 0.0, NULL);
+        PS_ASSERT_FLOAT_LESS_THAN_OR_EQUAL(badFrac, 1.0, NULL);
+    }
+
+    int numCols = mask1->numCols, numRows = mask1->numRows; // Size of the images
+
+    // Dereference inputs for convenience
+    psMaskType **data1 = mask1->data.PS_TYPE_MASK_DATA;
+    psMaskType **data2 = NULL;
+    if (mask2) {
+        data2 = mask2->data.PS_TYPE_MASK_DATA;
+    }
+
+    // First, a pass through to determine the fraction of bad pixels
+    if (isfinite(badFrac) && badFrac != 1.0) {
+        int numBad = 0;                 // Number of bad pixels
+        for (int y = 0; y < numRows; y++) {
+            for (int x = 0; x < numCols; x++) {
+                if (data1[y][x] & maskVal) {
+                    numBad++;
+                    continue;
+                }
+                if (data2 && data2[y][x] & maskVal) {
+                    numBad++;
+                }
+            }
+        }
+        if (numBad > badFrac * numCols * numRows) {
+            psError(PS_ERR_BAD_PARAMETER_VALUE, true,
+                    "Fraction of bad pixels (%d/%d=%f) exceeds limit (%f)\n",
+                    numBad, numCols * numRows, (float)numBad/(float)(numCols * numRows), badFrac);
+            return NULL;
+        }
+    }
+
+    // Worried about the masks for bad pixels and bad stamps colliding, so make our own mask
+    psImage *mask = psImageAlloc(numCols, numRows, PS_TYPE_MASK); // The global mask
+    psImageInit(mask, 0);
+    psMaskType **maskData = mask->data.PS_TYPE_MASK_DATA; // Dereference for convenience
+
+    // Block out a border around the edge of the image
+
+    // Bottom stripe
+    for (int y = 0; y < PS_MIN(size + footprint, numRows); y++) {
+        for (int x = 0; x < numCols; x++) {
+            maskData[y][x] |= PM_SUBTRACTION_MASK_BORDER;
+        }
+    }
+    // Either side
+    for (int y = PS_MIN(size + footprint, numRows); y < numRows - size - footprint; y++) {
+        for (int x = 0; x < PS_MIN(size + footprint, numCols); x++) {
+            maskData[y][x] |= PM_SUBTRACTION_MASK_BORDER;
+        }
+        for (int x = PS_MAX(numCols - size - footprint, 0); x < numCols; x++) {
+            maskData[y][x] |= PM_SUBTRACTION_MASK_BORDER;
+        }
+    }
+    // Top stripe
+    for (int y = PS_MAX(numRows - size - footprint, 0); y < numRows; y++) {
+        for (int x = 0; x < numCols; x++) {
+            maskData[y][x] |= PM_SUBTRACTION_MASK_BORDER;
+        }
+    }
+
+    // XXX Could do something smarter here --- we will get images that are predominantly masked (where the
+    // skycell isn't overlapped by a large fraction by the observation), so that convolving around every bad
+    // pixel is wasting time.  As a first cut, I've put in a check on the fraction of bad pixels, but we could
+    // imagine looking for the edge of big regions and convolving just at the edge.  As a second cut, allow
+    // use of FFT convolution.
+
+    for (int y = 0; y < numRows; y++) {
+        for (int x = 0; x < numCols; x++) {
+            if (data1[y][x] & maskVal) {
+                maskData[y][x] |= PM_SUBTRACTION_MASK_BAD_1;
+            }
+            if (data2 && data2[y][x] & maskVal) {
+                maskData[y][x] |= PM_SUBTRACTION_MASK_BAD_2;
+            }
+        }
+    }
+
+    // Block out the entire stamp footprint around bad input pixels.
+
+    // We want to block out with the CONVOLVE mask anything that would be bad if we convolved with a bad
+    // reference pixel (within 'size').  Then we want to block out with the FOOTPRINT mask everything within a
+    // footprint's distance of those (within 'footprint').
+
+    if (!psImageConvolveMask(mask, mask, PM_SUBTRACTION_MASK_BAD_1,
+                             PM_SUBTRACTION_MASK_CONVOLVE_1,
+                             -size, size, -size, size)) {
+        psError(PS_ERR_UNKNOWN, false, "Unable to convolve bad pixels from mask 1.");
+        psFree(mask);
+        return NULL;
+    }
+    if (!psImageConvolveMask(mask, mask, PM_SUBTRACTION_MASK_BAD_2,
+                             PM_SUBTRACTION_MASK_CONVOLVE_2,
+                             -size, size, -size, size)) {
+        psError(PS_ERR_UNKNOWN, false, "Unable to convolve bad pixels from mask 2.");
+        psFree(mask);
+        return NULL;
+    }
+
+    return mask;
+}
+
+
+bool pmSubtractionBorder(psImage *image, psImage *weight, psImage *mask,
+                         int size, psMaskType blank)
+{
+    PS_ASSERT_IMAGE_NON_NULL(image, false);
+    PS_ASSERT_IMAGE_TYPE(image, PS_TYPE_F32, false);
+    if (mask) {
+        PS_ASSERT_IMAGE_NON_NULL(mask, false);
+        PS_ASSERT_IMAGES_SIZE_EQUAL(mask, image, false);
+        PS_ASSERT_IMAGE_TYPE(mask, PS_TYPE_MASK, false);
+    }
+    if (weight) {
+        PS_ASSERT_IMAGE_NON_NULL(weight, false);
+        PS_ASSERT_IMAGES_SIZE_EQUAL(weight, image, false);
+        PS_ASSERT_IMAGE_TYPE(weight, PS_TYPE_F32, false);
+    }
+
+    int numCols = image->numCols, numRows = image->numRows; // Image dimensions
+
+    for (int y = size; y < numRows - size; y++) {
+        for (int x = 0; x < size; x++) {
+            markBlank(image, mask, weight, x, y, blank);
+        }
+        for (int x = numCols - size; x < numCols; x++) {
+            markBlank(image, mask, weight, x, y, blank);
+        }
+    }
+    for (int y = 0; y < size; y++) {
+        for (int x = 0; x < numCols; x++) {
+            markBlank(image, mask, weight, x, y, blank);
+        }
+    }
+    for (int y = numRows - size; y < numRows; y++) {
+        for (int x = 0; x < numCols; x++) {
+            markBlank(image, mask, weight, x, y, blank);
+        }
+    }
+
+    return true;
+}
+
+
+bool pmSubtractionMaskApply(psImage *image, psImage *weight, const psImage *mask, pmSubtractionMode mode)
+{
+    PS_ASSERT_IMAGE_NON_NULL(image, false);
+    PS_ASSERT_IMAGE_TYPE(image, PS_TYPE_F32, false);
+    if (weight) {
+        PS_ASSERT_IMAGE_NON_NULL(weight, false);
+        PS_ASSERT_IMAGE_TYPE(weight, PS_TYPE_F32, false);
+        PS_ASSERT_IMAGES_SIZE_EQUAL(weight, image, false);
+    }
+    PS_ASSERT_IMAGE_NON_NULL(mask, false);
+    PS_ASSERT_IMAGE_TYPE(mask, PS_TYPE_MASK, false);
+    PS_ASSERT_IMAGES_SIZE_EQUAL(mask, image, false);
+
+    bool maskVal = PM_SUBTRACTION_MASK_BORDER; // Value to mask
+    switch (mode) {
+      case PM_SUBTRACTION_MODE_1:
+        maskVal |= PM_SUBTRACTION_MASK_CONVOLVE_1;
+        break;
+      case PM_SUBTRACTION_MODE_2:
+        maskVal |= PM_SUBTRACTION_MASK_CONVOLVE_2;
+        break;
+      case PM_SUBTRACTION_MODE_DUAL:
+        maskVal |= PM_SUBTRACTION_MASK_CONVOLVE_2 | PM_SUBTRACTION_MASK_CONVOLVE_2;
+        break;
+      case PM_SUBTRACTION_MODE_ERR:
+      case PM_SUBTRACTION_MODE_UNSURE:
+      default:
+        psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Unsuppored subtraction mode: %x", mode);
+        return false;
+    }
+
+    int numCols = image->numCols, numRows = image->numRows; // Size of image
+    psMaskType **maskData = mask->data.PS_TYPE_MASK_DATA; // Dereference mask
+
+    for (int y = 0; y < numRows; y++) {
+        for (int x = 0; x < numCols; x++) {
+            if (maskData[y][x] & maskVal) {
+                image->data.F32[y][x] = NAN;
+                if (weight) {
+                    weight->data.F32[y][x] = NAN;
+                }
+            }
+        }
+    }
+
+    return true;
+}
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionMask.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionMask.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionMask.h	(revision 19481)
@@ -0,0 +1,36 @@
+#ifndef PM_SUBTRACTION_MASK_H
+#define PM_SUBTRACTION_MASK_H
+
+#include <pslib.h>
+
+/// Generate a mask for use in the subtraction process
+psImage *pmSubtractionMask(const psImage *refMask, ///< Mask for the reference image (will be convolved)
+                           const psImage *inMask, ///< Mask for the input image, or NULL
+                           psMaskType maskVal, ///< Value to mask out
+                           int size, ///< Half-size of the kernel (pmSubtractionKernels.size)
+                           int footprint, ///< Half-size of the kernel footprint
+                           float badFrac, ///< Maximum fraction of bad input pixels to accept
+                           bool useFFT  ///< Use FFT to do convolution?
+    );
+
+/// Mark the non-convolved part of the image as blank
+bool pmSubtractionBorder(psImage *image,///< Image
+                         psImage *weight, ///< Weight map (or NULL)
+                         psImage *mask, ///< Mask (or NULL)
+                         int size,      ///< Kernel half-size
+                         psMaskType blank ///< Mask value for blank regions
+    );
+
+/// Apply the subtraction mask to an image and weight.
+///
+/// Unfortunately, image subtraction may result in a bi-modal image in masked areas, which can upset image
+/// statistics (very important for quantising images so that a product can be written out!).  This function
+/// sets masked areas to NAN in the image and weight.
+bool pmSubtractionMaskApply(psImage *image, ///< Image to which to apply mask
+                            psImage *weight, ///< Weight map to which to apply mask (or NULL)
+                            const psImage *mask, ///< Subtraction mask
+                            pmSubtractionMode mode ///< Subtraction mode
+    );
+
+
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionMatch.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionMatch.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionMatch.c	(revision 19481)
@@ -0,0 +1,828 @@
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <stdio.h>
+#include <unistd.h>
+#include <string.h>
+#include <pslib.h>
+
+#include "pmHDU.h"
+#include "pmFPA.h"
+#include "pmSubtractionParams.h"
+#include "pmSubtractionKernels.h"
+#include "pmSubtractionStamps.h"
+#include "pmSubtractionEquation.h"
+#include "pmSubtraction.h"
+#include "pmSubtractionMask.h"
+#include "pmSubtractionThreads.h"
+#include "pmSubtractionMatch.h"
+
+
+#define KERNEL_MOSAIC 2                 // Half-number of kernel instances in the mosaic image
+#define BG_STAT PS_STAT_ROBUST_MEDIAN   // Statistic to use for background
+
+static bool useFFT = true;              // Do convolutions using FFT
+
+
+//#define TESTING
+//#define TESTING_MEMORY
+
+// Output memory usage information
+static void memCheck(const char *where)
+{
+#ifdef TESTING_MEMORY
+    psMemBlock **leaks = NULL;
+    int numLeaks = psMemCheckLeaks(0, &leaks, NULL, true);
+    size_t largestSize = 0;
+    psMemId largest = 0;
+    size_t totalSize = 0;
+    for (int i = 0; i < numLeaks; i++) {
+        psMemBlock *mb = leaks[i];
+        totalSize += mb->userMemorySize;
+        if (mb->userMemorySize > largestSize) {
+            largestSize = mb->userMemorySize;
+            largest = mb->id;
+        }
+    }
+    psFree(leaks);
+    fprintf(stderr, "%s:\n", where);
+    fprintf(stderr, "    Memory in use: %zd\n", totalSize);
+    fprintf(stderr, "    Largest block: %ld\n", largest);
+    fprintf(stderr, "    sbrk(): %zd\n", (size_t)sbrk(0));
+#endif
+    return;
+}
+
+
+static bool getStamps(pmSubtractionStampList **stamps, // Stamps to read
+                      const pmReadout *ro1, // Readout 1
+                      const pmReadout *ro2, // Readout 2
+                      const psImage *subMask, // Mask for subtraction, or NULL
+                      psImage *weight,  // Weight map
+                      const psRegion *region, // Region of interest, or NULL
+                      float threshold,  // Threshold for stamp finding
+                      float stampSpacing, // Spacing between stamps
+                      int size,         // Kernel half-size
+                      int footprint,     // Convolution footprint for stamps
+                      pmSubtractionMode mode // Mode for subtraction
+    )
+{
+    psTrace("psModules.imcombine", 3, "Finding stamps...\n");
+    *stamps = pmSubtractionStampsFind(*stamps, ro1->image, subMask, region, threshold, footprint,
+                                      stampSpacing, mode);
+    if (!*stamps) {
+        psError(PS_ERR_UNKNOWN, false, "Unable to find stamps.");
+        return false;
+    }
+
+    memCheck("  find stamps");
+
+    psTrace("psModules.imcombine", 3, "Extracting stamps...\n");
+    if (!pmSubtractionStampsExtract(*stamps, ro1->image, ro2 ? ro2->image : NULL, weight, size)) {
+        psError(PS_ERR_UNKNOWN, false, "Unable to extract stamps.");
+        return false;
+    }
+
+    memCheck("   extract stamps");
+
+    return true;
+}
+
+
+// Record the variance factor in a readout
+static bool recordVarianceFactor(pmReadout *ro, // Readout of interest
+                                 float varFactor, // Variance factor
+                                 const psRegion *region // Region of interest
+                                 )
+{
+    if (region) {
+        varFactor *= (region->x1 - region->x0 + 1) * (region->y1 - region->y0 + 1) /
+            (ro->image->numCols * ro->image->numRows);
+    }
+
+    psMetadataItem *vfItem = psMetadataLookup(ro->analysis, PM_SUBTRACTION_ANALYSIS_VARFACTOR);
+    if (vfItem) {
+        psAssert(vfItem->type == PS_TYPE_F32, "Should be the type we said.");
+        vfItem->data.F32 += varFactor;
+    } else {
+        psMetadataAddF32(ro->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_VARFACTOR, 0,
+                         "Variance factor weighted by the area", varFactor);
+    }
+
+    return true;
+}
+
+bool pmSubtractionMatch(pmReadout *conv1, pmReadout *conv2, const pmReadout *ro1, const pmReadout *ro2,
+                        int footprint, float regionSize, float stampSpacing, float threshold,
+                        const psArray *sources, const char *stampsName,
+                        pmSubtractionKernelsType type, int size, int spatialOrder,
+                        const psVector *isisWidths, const psVector *isisOrders,
+                        int inner, int ringsOrder, int binning, float penalty,
+                        bool optimum, const psVector *optFWHMs, int optOrder, float optThreshold,
+                        int iter, float rej, psMaskType maskVal, psMaskType maskBad, psMaskType maskPoor,
+                        float poorFrac, float badFrac, pmSubtractionMode subMode)
+{
+    if (subMode != PM_SUBTRACTION_MODE_2) {
+        PM_ASSERT_READOUT_NON_NULL(conv1, false);
+        if (conv1->image) {
+            psFree(conv1->image);
+            conv1->image = NULL;
+        }
+        if (conv1->mask) {
+            psFree(conv1->mask);
+            conv1->mask = NULL;
+        }
+        if (conv1->weight) {
+            psFree(conv1->weight);
+            conv1->weight = NULL;
+        }
+    }
+    if (subMode != PM_SUBTRACTION_MODE_1) {
+        PM_ASSERT_READOUT_NON_NULL(conv2, false);
+        if (conv2->image) {
+            psFree(conv2->image);
+            conv2->image = NULL;
+        }
+        if (conv2->mask) {
+            psFree(conv2->mask);
+            conv2->mask = NULL;
+        }
+        if (conv2->weight) {
+            psFree(conv2->weight);
+            conv2->weight = NULL;
+        }
+    }
+
+    PM_ASSERT_READOUT_NON_NULL(ro1, false);
+    PM_ASSERT_READOUT_NON_NULL(ro2, false);
+    PM_ASSERT_READOUT_IMAGE(ro1, false);
+    PM_ASSERT_READOUT_IMAGE(ro2, false);
+    PS_ASSERT_IMAGES_SIZE_EQUAL(ro1->image, ro2->image, false);
+
+    PS_ASSERT_INT_NONNEGATIVE(footprint, false);
+    // regionSize can be just about anything (except maybe negative, but it can be NAN)
+    PS_ASSERT_FLOAT_LARGER_THAN(stampSpacing, 0.0, false);
+    // Don't care what threshold is
+    if (sources) {
+        PS_ASSERT_ARRAY_NON_NULL(sources, false);
+    }
+    // stampsName may be anything
+    // We'll check kernel type when we allocate the kernels
+    PS_ASSERT_INT_POSITIVE(size, false);
+    PS_ASSERT_INT_NONNEGATIVE(spatialOrder, false);
+    if (isisWidths || isisOrders) {
+        PS_ASSERT_VECTOR_NON_NULL(isisWidths, false);
+        PS_ASSERT_VECTOR_TYPE(isisWidths, PS_TYPE_F32, false);
+        PS_ASSERT_VECTOR_NON_NULL(isisOrders, false);
+        PS_ASSERT_VECTOR_TYPE(isisOrders, PS_TYPE_S32, false);
+        PS_ASSERT_VECTORS_SIZE_EQUAL(isisWidths, isisOrders, false);
+    }
+    PS_ASSERT_INT_NONNEGATIVE(inner, false);
+    PS_ASSERT_INT_NONNEGATIVE(ringsOrder, false);
+    PS_ASSERT_INT_POSITIVE(binning, false);
+    if (optimum) {
+        PS_ASSERT_VECTOR_NON_NULL(optFWHMs, false);
+        PS_ASSERT_INT_NONNEGATIVE(optOrder, false);
+        PS_ASSERT_FLOAT_LARGER_THAN(optThreshold, 0.0, false);
+        PS_ASSERT_FLOAT_LESS_THAN_OR_EQUAL(optThreshold, 1.0, false);
+    }
+    PS_ASSERT_INT_POSITIVE(iter, false);
+    PS_ASSERT_FLOAT_LARGER_THAN(rej, 0.0, false);
+    // Don't care about maskVal
+    // Don't care about maskBad
+    // Don't care about maskPoor
+    PS_ASSERT_FLOAT_LARGER_THAN(poorFrac, 0.0, NULL);
+    PS_ASSERT_FLOAT_LESS_THAN_OR_EQUAL(poorFrac, 1.0, NULL);
+    if (isfinite(badFrac)) {
+        PS_ASSERT_FLOAT_LARGER_THAN(badFrac, 0.0, NULL);
+        PS_ASSERT_FLOAT_LESS_THAN_OR_EQUAL(badFrac, 1.0, NULL);
+    }
+
+    // If the stamp footprint is smaller than the kernel size, then we won't get much signal in the outer
+    // parts of the kernel, which can result in bad matching artifacts.
+    if (footprint < size) {
+        psError(PS_ERR_BAD_PARAMETER_VALUE, true,
+                "Stamp footprint (%d) should be larger than or equal to the kernel size (%d)",
+                footprint, size);
+        return false;
+    }
+
+    // Where does our weight map come from?
+    // Getting the weight exactly right is not necessary --- it's just used for weighting.
+    psImage *weight = NULL;             // Weight image to use
+    if (ro1->weight && ro2->weight) {
+        weight = (psImage*)psBinaryOp(NULL, ro1->weight, "+", ro2->weight);
+    } else if (ro1->weight) {
+        weight = psMemIncrRefCounter(ro1->weight);
+    } else if (ro2->weight) {
+        weight = psMemIncrRefCounter(ro2->weight);
+    } else {
+        weight = (psImage*)psBinaryOp(NULL, ro1->image, "+", ro2->image);
+    }
+
+    // Putting important variable declarations here, since they are freed after a "goto" if there is an error.
+    psImage *subMask = NULL;            // Mask for subtraction
+    psRegion *region = NULL;            // Iso-kernel region
+    psString regionString = NULL;       // String for region
+    pmSubtractionStampList *stamps = NULL; // Stamps for matching PSF
+    pmSubtractionKernels *kernels = NULL; // Kernel basis functions
+
+    int numCols = ro1->image->numCols, numRows = ro1->image->numRows; // Image dimensions
+
+    memCheck("start");
+
+    subMask = pmSubtractionMask(ro1->mask, ro2 ? ro2->mask : NULL, maskVal, size, footprint,
+                                badFrac, useFFT);
+    if (!subMask) {
+        psError(PS_ERR_UNKNOWN, false, "Unable to generate subtraction mask.");
+        psFree(weight);
+        return false;
+    }
+
+    memCheck("mask");
+
+    // Get region of interest
+    int xRegions = 1, yRegions = 1;     // Number of iso-kernel regions
+    float xRegionSize = 0, yRegionSize = 0; // Size of iso-kernel regions
+    if (isfinite(regionSize) && regionSize != 0.0) {
+        xRegions = numCols / regionSize + 1;
+        yRegions = numRows / regionSize + 1;
+        xRegionSize = (float)numCols / (float)xRegions;
+        yRegionSize = (float)numRows / (float)yRegions;
+        region = psRegionAlloc(NAN, NAN, NAN, NAN);
+    }
+
+    // Iterate over iso-kernel regions
+    for (int j = 0; j < yRegions; j++) {
+        for (int i = 0; i < xRegions; i++) {
+            psTrace("psModules.imcombine", 1, "Subtracting region %d of %d...\n",
+                    j * xRegions + i + 1, xRegions * yRegions);
+            if (region) {
+                *region = psRegionSet((int)(i * xRegionSize), (int)((i + 1) * xRegionSize),
+                                      (int)(j * yRegionSize), (int)((j + 1) * yRegionSize));
+                psFree(regionString);
+                regionString = psRegionToString(*region);
+                psTrace("psModules.imcombine", 3, "Iso-kernel region: %s out of %d,%d\n",
+                        regionString, numCols, numRows);
+            }
+
+            if (sources) {
+                stamps = pmSubtractionStampsSetFromSources(sources, subMask, region, footprint,
+                                                           stampSpacing, subMode);
+            } else if (stampsName && strlen(stampsName) > 0) {
+                stamps = pmSubtractionStampsSetFromFile(stampsName, ro1->image, subMask, region, footprint,
+                                                        stampSpacing, subMode);
+            }
+
+            // We get the stamps here; we will also attempt to get stamps at the first iteration, but it
+            // doesn't matter.
+            if (!getStamps(&stamps, ro1, ro2, subMask, weight, NULL, threshold, stampSpacing,
+                           size, footprint, subMode)) {
+                goto MATCH_ERROR;
+            }
+
+            if (subMode == PM_SUBTRACTION_MODE_UNSURE) {
+                // Get backgrounds
+                psStats *bgStats = psStatsAlloc(BG_STAT); // Statistics for background
+                psRandom *rng = psRandomAlloc(PS_RANDOM_TAUS, 0); // Random number generator
+                psVector *buffer = NULL;// Buffer for stats
+                if (!psImageBackground(bgStats, &buffer, ro1->image, ro1->mask, maskVal, rng)) {
+                    psError(PS_ERR_UNKNOWN, false, "Unable to measure background of image 1.");
+                    psFree(bgStats);
+                    psFree(rng);
+                    psFree(buffer);
+                    goto MATCH_ERROR;
+                }
+                float bg1 = psStatsGetValue(bgStats, BG_STAT); // Background for image 1
+                if (!psImageBackground(bgStats, &buffer, ro2->image, ro2->mask, maskVal, rng)) {
+                    psError(PS_ERR_UNKNOWN, false, "Unable to measure background of image 2.");
+                    psFree(bgStats);
+                    psFree(rng);
+                    psFree(buffer);
+                    goto MATCH_ERROR;
+                }
+                float bg2 = psStatsGetValue(bgStats, BG_STAT); // Background for image 2
+                psFree(bgStats);
+                psFree(rng);
+                psFree(buffer);
+
+                pmSubtractionMode newMode = pmSubtractionOrder(stamps, bg1, bg2); // Subtraction mode to use
+                switch (newMode) {
+                  case PM_SUBTRACTION_MODE_1:
+                    psLogMsg("psModules.imcombine", PS_LOG_INFO, "Convolving image 1 to match image 2.");
+                    break;
+                  case PM_SUBTRACTION_MODE_2:
+                    psLogMsg("psModules.imcombine", PS_LOG_INFO, "Convolving image 2 to match image 1.");
+                    break;
+                  default:
+                    psError(PS_ERR_UNKNOWN, false, "Unable to determine subtraction order.");
+                    goto MATCH_ERROR;
+                }
+                subMode = newMode;
+            }
+
+            // Define kernel basis functions
+            if (optimum && (type == PM_SUBTRACTION_KERNEL_ISIS || type == PM_SUBTRACTION_KERNEL_GUNK)) {
+                kernels = pmSubtractionKernelsOptimumISIS(type, size, inner, spatialOrder, optFWHMs, optOrder,
+                                                          stamps, footprint, optThreshold, penalty, subMode);
+                if (!kernels) {
+                    psErrorClear();
+                    psWarning("Unable to derive optimum ISIS kernel --- switching to default.");
+                }
+            }
+            if (kernels == NULL) {
+                // Not an ISIS/GUNK kernel, or the optimum kernel search failed
+                kernels = pmSubtractionKernelsGenerate(type, size, spatialOrder, isisWidths, isisOrders,
+                                                       inner, binning, ringsOrder, penalty, subMode);
+            }
+
+            // Add analysis metadata
+            {
+                psRegion *subRegion;    // Region over which subtraction was performed
+                if (region) {
+                    subRegion = psMemIncrRefCounter(region);
+                } else {
+                    subRegion = psRegionAlloc(0, numCols, 0, numRows);
+                }
+
+                if (subMode == PM_SUBTRACTION_MODE_1 || subMode == PM_SUBTRACTION_MODE_DUAL) {
+                    psMetadataAddPtr(conv1->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_KERNEL,
+                                     PS_DATA_UNKNOWN | PS_META_DUPLICATE_OK, "Subtraction kernels", kernels);
+                    psMetadataAddS32(conv1->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_MODE,
+                                     PS_META_DUPLICATE_OK, "Subtraction kernels", subMode);
+                    psMetadataAddPtr(conv1->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_REGION,
+                                     PS_DATA_REGION | PS_META_DUPLICATE_OK,
+                                     "Region over which subtraction was performed", subRegion);
+                }
+                if (subMode == PM_SUBTRACTION_MODE_2 || subMode == PM_SUBTRACTION_MODE_DUAL) {
+                    psMetadataAddPtr(conv2->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_KERNEL,
+                                     PS_DATA_UNKNOWN | PS_META_DUPLICATE_OK, "Subtraction kernels", kernels);
+                    psMetadataAddS32(conv2->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_MODE,
+                                     PS_META_DUPLICATE_OK, "Subtraction kernels", subMode);
+                    psMetadataAddPtr(conv2->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_REGION,
+                                     PS_DATA_REGION | PS_META_DUPLICATE_OK,
+                                     "Region over which subtraction was performed", subRegion);
+                }
+                psFree(subRegion);
+            }
+
+            memCheck("kernels");
+
+            float rmsStamps = NAN;      // RMS for stamps
+            int numStamps = 0;          // Number of good stamps
+            int numRejected = -1;       // Number of rejected stamps in each iteration
+            for (int k = 0; k < iter && numRejected != 0; k++) {
+                psLogMsg("psModules.imcombine", PS_LOG_INFO, "Iteration %d.", k);
+
+                if (!getStamps(&stamps, ro1, ro2, subMask, weight, region, threshold, stampSpacing,
+                               size, footprint, subMode)) {
+                    goto MATCH_ERROR;
+                }
+
+                psTrace("psModules.imcombine", 3, "Calculating equation...\n");
+                if (!pmSubtractionCalculateEquation(stamps, kernels)) {
+                    psError(PS_ERR_UNKNOWN, false, "Unable to calculate least-squares equation.");
+                    goto MATCH_ERROR;
+                }
+
+                memCheck("  calculate equation");
+
+                psTrace("psModules.imcombine", 3, "Solving equation...\n");
+
+                if (!pmSubtractionSolveEquation(kernels, stamps)) {
+                    psError(PS_ERR_UNKNOWN, false, "Unable to calculate least-squares equation.");
+                    goto MATCH_ERROR;
+                }
+
+                memCheck("  solve equation");
+
+                psVector *deviations = pmSubtractionCalculateDeviations(stamps, kernels); // Stamp deviations
+                if (!deviations) {
+                    psError(PS_ERR_UNKNOWN, false, "Unable to calculate deviations.");
+                    goto MATCH_ERROR;
+                }
+
+                memCheck("   calculate deviations");
+
+                psTrace("psModules.imcombine", 3, "Rejecting stamps...\n");
+                numRejected = pmSubtractionRejectStamps(&rmsStamps, &numStamps, stamps, deviations,
+                                                        subMask, rej, footprint);
+                if (numRejected < 0) {
+                    psError(PS_ERR_UNKNOWN, false, "Unable to reject stamps.");
+                    psFree(deviations);
+                    goto MATCH_ERROR;
+                }
+                psFree(deviations);
+
+                memCheck("  reject stamps");
+            }
+
+            if (numRejected > 0) {
+                psTrace("psModules.imcombine", 3, "Solving equation...\n");
+                if (!pmSubtractionSolveEquation(kernels, stamps)) {
+                    psError(PS_ERR_UNKNOWN, false, "Unable to calculate least-squares equation.");
+                    goto MATCH_ERROR;
+                }
+                psVector *deviations = pmSubtractionCalculateDeviations(stamps, kernels); // Stamp deviations
+                if (!deviations) {
+                    psError(PS_ERR_UNKNOWN, false, "Unable to calculate deviations.");
+                    goto MATCH_ERROR;
+                }
+                pmSubtractionRejectStamps(&rmsStamps, &numStamps, stamps, deviations,
+                                          subMask, NAN, footprint);
+                psFree(deviations);
+            }
+            psFree(stamps);
+            stamps = NULL;
+
+            if (subMode == PM_SUBTRACTION_MODE_1 || subMode == PM_SUBTRACTION_MODE_DUAL) {
+                psMetadataAddS32(conv1->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_STAMPS_NUM,
+                                 PS_META_DUPLICATE_OK, "Number of good stamps", numStamps);
+                psMetadataAddF32(conv1->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_STAMPS_RMS,
+                                 PS_META_DUPLICATE_OK, "RMS deviation of stamps", rmsStamps);
+            }
+            if (subMode == PM_SUBTRACTION_MODE_2 || subMode == PM_SUBTRACTION_MODE_DUAL) {
+                psMetadataAddS32(conv2->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_STAMPS_NUM,
+                                 PS_META_DUPLICATE_OK, "Number of good stamps", numStamps);
+                psMetadataAddF32(conv2->analysis, PS_LIST_TAIL, PM_SUBTRACTION_ANALYSIS_STAMPS_RMS,
+                                 PS_META_DUPLICATE_OK, "RMS deviation of stamps", rmsStamps);
+            }
+
+            memCheck("solution");
+
+            {
+                psTrace("psModules.imcombine", 2, "Generating diagnostics...\n");
+                // Generate image with convolution kernels
+                int fullSize = 2 * size + 1 + 1;    // Full size of kernel
+                int imageSize = (2 * KERNEL_MOSAIC + 1) * fullSize;
+                psImage *convKernels = psImageAlloc((subMode == PM_SUBTRACTION_MODE_DUAL ? 2 : 1) *
+                                                    imageSize - 1 +
+                                                    (subMode == PM_SUBTRACTION_MODE_DUAL ? 4 : 0),
+                                                    imageSize - 1, PS_TYPE_F32);
+                psImageInit(convKernels, NAN);
+                for (int j = -KERNEL_MOSAIC; j <= KERNEL_MOSAIC; j++) {
+                    for (int i = -KERNEL_MOSAIC; i <= KERNEL_MOSAIC; i++) {
+                        psImage *kernel = pmSubtractionKernelImage(kernels, (float)i / (float)KERNEL_MOSAIC,
+                                                                   (float)j / (float)KERNEL_MOSAIC,
+                                                                   false); // Image of the kernel
+                        if (!kernel) {
+                            psError(PS_ERR_UNKNOWN, false, "Unable to generate kernel image.");
+                            psFree(convKernels);
+                            goto MATCH_ERROR;
+                        }
+
+                        if (psImageOverlaySection(convKernels, kernel, (i + KERNEL_MOSAIC) * fullSize,
+                                                  (j + KERNEL_MOSAIC) * fullSize, "=") == 0) {
+                            psError(PS_ERR_UNKNOWN, false, "Unable to overlay kernel image.");
+                            psFree(kernel);
+                            psFree(convKernels);
+                            goto MATCH_ERROR;
+                        }
+                        psFree(kernel);
+
+                        if (subMode == PM_SUBTRACTION_MODE_DUAL) {
+                            kernel = pmSubtractionKernelImage(kernels, (float)i / (float)KERNEL_MOSAIC,
+                                                              (float)j / (float)KERNEL_MOSAIC,
+                                                              true); // Image of the kernel
+                            if (!kernel) {
+                                psError(PS_ERR_UNKNOWN, false, "Unable to generate kernel image.");
+                                psFree(convKernels);
+                                goto MATCH_ERROR;
+                            }
+
+                            if (psImageOverlaySection(convKernels, kernel,
+                                                      (2 * KERNEL_MOSAIC + 1 + i + KERNEL_MOSAIC) * fullSize +
+                                                      4,
+                                                      (j + KERNEL_MOSAIC) * fullSize, "=") == 0) {
+                                psError(PS_ERR_UNKNOWN, false, "Unable to overlay kernel image.");
+                                psFree(kernel);
+                                psFree(convKernels);
+                                goto MATCH_ERROR;
+                            }
+                            psFree(kernel);
+                        }
+
+                    }
+                }
+
+                psString comment = NULL; // Comment for metadata
+                psStringAppend(&comment, "Subtraction kernel for region %s", regionString);
+                psMetadataAddImage(conv1->analysis, PS_LIST_TAIL, "SUBTRACTION.KERNEL.IMAGE",
+                                   PS_META_DUPLICATE_OK, comment, convKernels);
+                psFree(comment);
+                psFree(convKernels);
+            }
+
+#if 0
+            {
+                // Generate images of the kernel components
+                psMetadata *header = psMetadataAlloc(); // Header
+                for (int i = 0; i < solution->n; i++) {
+                    psString name = NULL;       // Header keyword
+                    psStringAppend(&name, "SOLN%04d", i);
+                    psMetadataAddF64(header, PS_LIST_TAIL, name, 0, NULL, solution->data.F64[i]);
+                    psFree(name);
+                }
+                psArray *kernelImages = pmSubtractionKernelSolutions(solution, kernels, 0.0, 0.0);
+                psFits *kernelFile = psFitsOpen("kernels.fits", "w");
+                (void)psFitsWriteImageCube(kernelFile, header, kernelImages, NULL);
+                psFitsClose(kernelFile);
+                psFree(kernelImages);
+                psFree(header);
+            }
+#endif
+
+            memCheck("diag outputs");
+
+            psTrace("psModules.imcombine", 2, "Convolving...\n");
+            if (!pmSubtractionConvolve(conv1, conv2, ro1, ro2, subMask, maskBad, maskPoor, poorFrac,
+                                       region, kernels, true, useFFT)) {
+                psError(PS_ERR_UNKNOWN, false, "Unable to convolve image.");
+                goto MATCH_ERROR;
+            }
+
+            // Set the variance factors
+            switch (subMode) {
+              case PM_SUBTRACTION_MODE_1: {
+                  recordVarianceFactor(conv1, pmSubtractionVarianceFactor(kernels, 0.5, 0.5, false),
+                                       region);
+                  break;
+              }
+              case PM_SUBTRACTION_MODE_2:
+                recordVarianceFactor(conv1, pmSubtractionVarianceFactor(kernels, 0.5, 0.5, false),
+                                     region);
+                break;
+              case PM_SUBTRACTION_MODE_DUAL:
+                recordVarianceFactor(conv1, pmSubtractionVarianceFactor(kernels, 0.5, 0.5, false),
+                                     region);
+                recordVarianceFactor(conv2, pmSubtractionVarianceFactor(kernels, 0.5, 0.5, true),
+                                     region);
+                break;
+              default:
+                psAbort("Should never reach here.");
+            }
+
+            psFree(kernels);
+            kernels = NULL;
+
+            // There is data in the readout now
+            if (subMode == PM_SUBTRACTION_MODE_1 || subMode == PM_SUBTRACTION_MODE_DUAL) {
+                conv1->data_exists = true;
+                if (conv1->parent) {
+                    conv1->parent->data_exists = true;
+                    conv1->parent->parent->data_exists = true;
+                }
+            }
+            if (subMode == PM_SUBTRACTION_MODE_2 || subMode == PM_SUBTRACTION_MODE_DUAL) {
+                conv2->data_exists = true;
+                if (conv2->parent) {
+                    conv2->parent->data_exists = true;
+                    conv2->parent->parent->data_exists = true;
+                }
+            }
+        }
+    }
+    psFree(region);
+    region = NULL;
+    psFree(regionString);
+    regionString = NULL;
+    psFree(subMask);
+    subMask = NULL;
+    psFree(weight);
+    weight = NULL;
+
+    if (!pmSubtractionBorder(conv1->image, conv1->weight, conv1->mask, size, maskBad)) {
+        psError(PS_ERR_UNKNOWN, false, "Unable to set border of convolved image.");
+        goto MATCH_ERROR;
+    }
+
+    memCheck("convolution");
+
+
+#ifdef TESTING
+    {
+        if (subMode == PM_SUBTRACTION_MODE_1 || subMode == PM_SUBTRACTION_MODE_DUAL) {
+            psFits *fits = psFitsOpen("convolved1.fits", "w");
+            psFitsWriteImage(fits, NULL, conv1->image, 0, NULL);
+            psFitsClose(fits);
+        }
+
+        if (subMode == PM_SUBTRACTION_MODE_2 || subMode == PM_SUBTRACTION_MODE_DUAL) {
+            psFits *fits = psFitsOpen("convolved2.fits", "w");
+            psFitsWriteImage(fits, NULL, conv2->image, 0, NULL);
+            psFitsClose(fits);
+        }
+    }
+#endif
+
+    return true;
+
+MATCH_ERROR:
+    psFree(region);
+    psFree(regionString);
+    psFree(subMask);
+    psFree(kernels);
+    psFree(stamps);
+    psFree(weight);
+    return false;
+}
+
+
+// Determine a rough width (integer value) of the star in the image
+// XXX Could improve this by using a user-provided list of floating-point widths (or an end point and
+// increment).
+static int subtractionOrderWidth(const psKernel *kernel, // Image
+                                 float bg, // Background in image
+                                 int size, // Maximum size
+                                 const psArray *models, // Buffer of models
+                                 const psVector *modelSums // Buffer of model sums
+    )
+{
+    assert(kernel);
+    assert(models);
+    assert(modelSums);
+
+    int xMin = -size, xMax = size; // Bounds in x
+    int yMin = -size, yMax = size; // Bounds in y
+
+    // Fit gaussians of varying widths to the image, record the chi^2
+    psVector *chi2 = psVectorAlloc(size, PS_TYPE_F32); // chi^2 as a function of radius
+    for (int sigma = 0; sigma < size; sigma++) {
+        double sumFG = 0.0; // Sum for calculating the normalisation of the Gaussian
+        psKernel *model = models->data[sigma]; // Model of interest
+        for (int y = yMin; y <= yMax; y++) {
+            for (int x = xMin; x <= xMax; x++) {
+                sumFG += model->kernel[y][x] * (kernel->kernel[y][x] - bg);
+            }
+        }
+        float norm = sumFG * modelSums->data.F64[sigma]; // Normalisation for Gaussian
+        double sumDev2 = 0.0;           // Sum of square deviations
+        for (int y = yMin; y <= yMax; y++) {
+            for (int x = xMin; x <= xMax; x++) {
+                float dev = kernel->kernel[y][x] - bg - norm * model->kernel[y][x]; // Deviation
+                sumDev2 += PS_SQR(dev);
+            }
+        }
+        chi2->data.F32[sigma] = sumDev2;
+    }
+
+    // Find the minimum chi^2
+    int bestIndex = -1;                 // Index of best chi^2
+    float bestChi2 = INFINITY;          // Best chi^2
+    for (int i = 0; i < size; i++) {
+        if (chi2->data.F32[i] < bestChi2) {
+            bestChi2 = chi2->data.F32[i];
+            bestIndex = i;
+        }
+    }
+    psFree(chi2);
+
+    return bestIndex + 1;
+}
+
+
+bool pmSubtractionOrderStamp(psVector *ratios, psVector *mask, const pmSubtractionStampList *stamps,
+                             const psArray *models, const psVector *modelSums,
+                             int index, float bg1, float bg2)
+{
+    PS_ASSERT_VECTOR_NON_NULL(ratios, false);
+    PS_ASSERT_VECTOR_NON_NULL(mask, false);
+    PS_ASSERT_VECTORS_SIZE_EQUAL(ratios, mask, false);
+    PM_ASSERT_SUBTRACTION_STAMP_LIST_NON_NULL(stamps, false);
+    PS_ASSERT_INT_NONNEGATIVE(index, false);
+    PS_ASSERT_INT_LESS_THAN(index, stamps->num, false);
+    PS_ASSERT_ARRAY_NON_NULL(models, false);
+    PS_ASSERT_VECTOR_NON_NULL(modelSums, false);
+    PS_ASSERT_VECTOR_SIZE(modelSums, models->n, false);
+
+    pmSubtractionStamp *stamp = stamps->stamps->data[index]; // Stamp of interest
+    psAssert(stamp->status == PM_SUBTRACTION_STAMP_CALCULATE || stamp->status == PM_SUBTRACTION_STAMP_USED,
+             "We checked this earlier.");
+
+    // Widths of stars
+    int width1 = subtractionOrderWidth(stamp->image1, bg1, stamps->footprint, models, modelSums);
+    int width2 = subtractionOrderWidth(stamp->image2, bg2, stamps->footprint, models, modelSums);
+
+    if (width1 == 0 || width2 == 0) {
+        ratios->data.F32[index] = NAN;
+        mask->data.PS_TYPE_MASK_DATA[index] = 0xff;
+    } else {
+        ratios->data.F32[index] = (float)width1 / (float)width2;
+        mask->data.PS_TYPE_MASK_DATA[index] = 0;
+    }
+
+    return true;
+}
+
+bool pmSubtractionOrderThread(psThreadJob *job)
+{
+    PS_ASSERT_THREAD_JOB_NON_NULL(job, false);
+
+    psVector *ratios = job->args->data[0]; // Ratios of widths
+    psVector *mask = job->args->data[1]; // Mask for ratios
+    const pmSubtractionStampList *stamps = job->args->data[2]; // List of stamps
+    const psArray *models = job->args->data[3]; // Gaussian models
+    const psVector *modelSums = job->args->data[4]; // Gaussian model sums
+    int index = PS_SCALAR_VALUE(job->args->data[5], S32); // Stamp index
+    float bg1 = PS_SCALAR_VALUE(job->args->data[6], F32); // Background of image 1
+    float bg2 = PS_SCALAR_VALUE(job->args->data[7], F32); // Background of image 2
+
+    return pmSubtractionOrderStamp(ratios, mask, stamps, models, modelSums, index, bg1, bg2);
+}
+
+pmSubtractionMode pmSubtractionOrder(pmSubtractionStampList *stamps, float bg1, float bg2)
+{
+    PM_ASSERT_SUBTRACTION_STAMP_LIST_NON_NULL(stamps, PM_SUBTRACTION_MODE_ERR);
+
+    psVector *mask = psVectorAlloc(stamps->num, PS_TYPE_MASK); // Mask for stamps
+    psVector *ratios = psVectorAlloc(stamps->num, PS_TYPE_F32); // Ratios of widths
+
+    // Generate models
+    int size = stamps->footprint;       // Maximum size
+    psArray *models = psArrayAlloc(size); // Gaussian models
+    psVector *modelSums = psVectorAlloc(size, PS_TYPE_F64); // Gaussian model sums
+    for (int sigma = 0; sigma < size; sigma++) {
+        psKernel *model = psKernelAlloc(-size, size, -size, size); // Gaussian model
+        float invSigma2 = 1.0 / (float)PS_SQR(1 + sigma); // Inverse sigma squared
+        double sumGG = 0.0;         // Sum of square of Gaussian
+        for (int y = -size; y <= size; y++) {
+            int y2 = PS_SQR(y);     // y squared
+            for (int x = -size; x <= size; x++) {
+                float rad2 = PS_SQR(x) + y2; // Radius squared
+                float value = expf(-rad2 * invSigma2); // Model value
+                model->kernel[y][x] = value;
+                sumGG += PS_SQR(value);
+            }
+        }
+        models->data[sigma] = model;
+        modelSums->data.F64[sigma] = 1.0 / sumGG;
+    }
+
+    // Fit models to stamps
+    for (int i = 0; i < stamps->num; i++) {
+        pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+        if (stamp->status != PM_SUBTRACTION_STAMP_CALCULATE && stamp->status != PM_SUBTRACTION_STAMP_USED) {
+            mask->data.PS_TYPE_MASK_DATA[i] = 0xff;
+            continue;
+        }
+
+        if (pmSubtractionThreaded()) {
+            psThreadJob *job = psThreadJobAlloc("PSMODULES_SUBTRACTION_ORDER");
+            psArrayAdd(job->args, 1, ratios);
+            psArrayAdd(job->args, 1, mask);
+            psArrayAdd(job->args, 1, stamps);
+            psArrayAdd(job->args, 1, models);
+            psArrayAdd(job->args, 1, modelSums);
+            PS_ARRAY_ADD_SCALAR(job->args, i, PS_TYPE_S32);
+            PS_ARRAY_ADD_SCALAR(job->args, bg1, PS_TYPE_F32);
+            PS_ARRAY_ADD_SCALAR(job->args, bg2, PS_TYPE_F32);
+            if (!psThreadJobAddPending(job)) {
+                psFree(job);
+                return false;
+            }
+            psFree(job);
+        } else {
+            if (!pmSubtractionOrderStamp(ratios, mask, stamps, models, modelSums, i, bg1, bg2)) {
+                psError(PS_ERR_UNKNOWN, false, "Unable to measure PSF width for stamp %d", i);
+                psFree(models);
+                psFree(modelSums);
+                psFree(ratios);
+                psFree(mask);
+                return false;
+            }
+        }
+    }
+
+    if (!psThreadPoolWait(true)) {
+        psError(PS_ERR_UNKNOWN, false, "Error waiting for threads.");
+        psFree(models);
+        psFree(modelSums);
+        psFree(ratios);
+        psFree(mask);
+            return false;
+    }
+
+    psFree(models);
+    psFree(modelSums);
+
+    psStats *stats = psStatsAlloc(PS_STAT_ROBUST_MEDIAN);
+    if (!psVectorStats(stats, ratios, NULL, mask, 0xff)) {
+        psError(PS_ERR_UNKNOWN, false, "Unable to calculate statistics for moments ratio.");
+        psFree(mask);
+        psFree(ratios);
+        psFree(stats);
+        return PM_SUBTRACTION_MODE_ERR;
+    }
+    psFree(ratios);
+    psFree(mask);
+
+    psLogMsg("psModules.imcombine", PS_LOG_INFO, "Median width ratio: %lf", stats->robustMedian);
+    pmSubtractionMode mode = (stats->robustMedian <= 1.0 ? PM_SUBTRACTION_MODE_1 : PM_SUBTRACTION_MODE_2);
+    psFree(stats);
+
+    return mode;
+}
+
+
+
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionMatch.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionMatch.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionMatch.h	(revision 19481)
@@ -0,0 +1,77 @@
+#ifndef PM_SUBTRACTION_MATCH_H
+#define PM_SUBTRACTION_MATCH_H
+
+#include <pslib.h>
+
+#include <pmHDU.h>
+#include <pmFPA.h>
+#include <pmSubtractionKernels.h>
+#include <pmSubtractionStamps.h>
+
+#define PM_SUBTRACTION_ANALYSIS_KERNEL "SUBTRACTION.KERNEL" // Name of kernel in the analysis metadata
+#define PM_SUBTRACTION_ANALYSIS_MODE "SUBTRACTION.MODE" // Name of subtraction mode in the analysis metadata
+#define PM_SUBTRACTION_ANALYSIS_REGION "SUBTRACTION.REGION" // Name of subtraction region in the analysis MD
+#define PM_SUBTRACTION_ANALYSIS_STAMPS_RMS "SUBTRACTION.RMS" // Name of stamp rms in the analysis metadata
+#define PM_SUBTRACTION_ANALYSIS_STAMPS_NUM "SUBTRACTION.NUM"// Name of the number of stamps in the analysis MD
+#define PM_SUBTRACTION_ANALYSIS_VARFACTOR "SUBTRACTION.VARFACTOR"// Name of variance factor in the analysis MD
+
+
+/// Match two images
+bool pmSubtractionMatch(pmReadout *conv1, ///< Output convolved data for image 1
+                        pmReadout *conv2, ///< Output convolved data for image 2
+                        const pmReadout *ro1, ///< Image 1
+                        const pmReadout *ro2, ///< Image 2
+                        // Stamp parameters
+                        int footprint,  ///< Stamp half-size
+                        float regionSize, ///< Typical size of iso-kernel regions
+                        float stampSpacing, ///< Typical spacing between stamps
+                        float threshold, ///< Threshold for stamps
+                        const psArray *sources, ///< Sources for stamps
+                        const char *stampsName, ///< Filename for stamps
+                        // Kernel parameters
+                        pmSubtractionKernelsType type, ///< Kernel type
+                        int size,       ///< Kernel half-size
+                        int order,      ///< Spatial polynomial order
+                        const psVector *widths, ///< ISIS Gaussian widths
+                        const psVector *orders, ///< ISIS Polynomial orders
+                        int inner,      ///< Inner radius for various kernel types
+                        int ringsOrder, ///< RINGS polynomial order
+                        int binning,    ///< SPAM kernel binning
+                        float penalty,  ///< Penalty for wideness
+                        bool optimum,   ///< Search for optimum ISIS kernel?
+                        const psVector *optFWHMs, ///< FWHMs for optimum search
+                        int optOrder,   ///< Maximum order for optimum search
+                        float optThreshold, ///< Threshold for optimum search (0..1)
+                        // Operational parameters
+                        int iter,       ///< Rejection iterations
+                        float rej,      ///< Rejection threshold
+                        psMaskType maskVal, ///< Value to mask for input
+                        psMaskType maskBad, ///< Mask for output bad pixels
+                        psMaskType maskPoor, ///< Mask for output poor pixels
+                        float poorFrac, ///< Fraction for "poor"
+                        float badFrac,   ///< Maximum fraction of bad input pixels to accept
+                        pmSubtractionMode mode ///< Mode of subtraction; may be modified
+    );
+
+/// Execute a thread job to measure the PSF width ratios
+bool pmSubtractionOrderThread(psThreadJob *job ///< Job to execute
+    );
+
+/// Measure the PSF width ratio for a single stamp
+bool pmSubtractionOrderStamp(psVector *ratios, ///< PSF width ratios
+                             psVector *mask, ///< Mask for PSF width ratios
+                             const pmSubtractionStampList *stamps, ///< List of stamps
+                             const psArray *models, ///< Pre-calculated gaussian models
+                             const psVector *modelSums, ///< Pre-calculated gaussian model sums
+                             int index, ///< Index of stamp
+                             float bg1, ///< Background for image 1
+                             float bg2  ///< Background for image 2
+    );
+
+/// Determine which image to convolve
+pmSubtractionMode pmSubtractionOrder(pmSubtractionStampList *stamps, ///< Stamps that have been extracted
+                                     float bg1, float bg2 ///< Background for each image
+    );
+
+
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionParams.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionParams.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionParams.c	(revision 19481)
@@ -0,0 +1,510 @@
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <stdio.h>
+#include <string.h>
+#include <math.h>
+#include <pslib.h>
+
+#include "pmSubtractionStamps.h"
+#include "pmSubtraction.h"
+#include "pmSubtractionParams.h"
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+#if 0
+// Convolve the reference stamp by the kernel
+static psKernel *convolveStamp(const pmSubtractionStamp *stamp, // Stamp to be convolved
+                               const psKernel *kernel, // Kernel by which to convolve
+                               int footprint // Size of area to be convolved
+    )
+{
+    psKernel *convolution = psKernelAlloc(-footprint, footprint, -footprint, footprint); // Result
+    psKernel *reference = stamp->reference; // Reference stamp, to be convolved
+
+    // Range of kernel
+    int uMin = kernel->xMin;
+    int uMax = kernel->xMax;
+    int vMin = kernel->yMin;
+    int vMax = kernel->yMax;
+
+    for (int y = -footprint; y <= footprint; y++) {
+        psF32 *conv = &convolution->kernel[y][-footprint]; // Dereference convolution
+        for (int x = -footprint; x <= footprint; x++, conv++) {
+            *conv = 0.0;
+
+            int xStart = x + uMin;      // Start index for convolution
+            for (int v = vMin; v <= vMax; v++) {
+                psF32 *ref = &reference->kernel[y + v][xStart]; // Dereference reference image
+                psF32 *krnl = &kernel->kernel[v][uMin]; // Dereference kernel in x
+                for (int u = uMin; u <= uMax; u++, ref++, krnl++) {
+                    *conv += *ref * *krnl;
+                }
+            }
+        }
+    }
+
+    return convolution;
+}
+#endif
+
+/// Select the appropriate convolution, given the kernel basis function and subtraction mode
+static inline psKernel *selectConvolution(const pmSubtractionStamp *stamp, // Stamp
+                                          int kernelIndex, // Index for kernel component
+                                          pmSubtractionMode mode // Mode of subtraction
+    )
+{
+    switch (mode) {
+      case PM_SUBTRACTION_MODE_1:
+        return stamp->convolutions1->data[kernelIndex];
+      case PM_SUBTRACTION_MODE_2:
+        return stamp->convolutions2->data[kernelIndex];
+      default:
+        psAbort("Unsupported subtraction mode: %x", mode);
+    }
+    return NULL;                        // Unreached
+}
+
+// Accumulate cross-term sums for a stamp
+static void accumulateCross(double *sumI, // Sum of I(x)/sigma(x)^2
+                            double *sumII, // Sum of I(x)^2/sigma(x)^2
+                            double *sumIC, // Sum of I(x)conv(x)/sigma(x)^2
+                            const pmSubtractionStamp *stamp, // Stamp with weight
+                            const psKernel *target, // Target stamp
+                            int kernelIndex, // Index for kernel component
+                            int footprint, // Size of region of interest
+                            pmSubtractionMode mode // Mode of subtraction
+    )
+{
+    psKernel *weight = stamp->weight;   // Weight, sigma(x)^2
+    psKernel *convolution = selectConvolution(stamp, kernelIndex, mode); // Convolution of interest
+
+    for (int y = -footprint; y <= footprint; y++) {
+        psF32 *in = &target->kernel[y][-footprint]; // Dereference input
+        psF32 *wt = &weight->kernel[y][-footprint]; // Dereference weight
+        psF32 *conv = &convolution->kernel[y][-footprint]; // Dereference convolution
+        for (int x = -footprint; x <= footprint; x++, in++, wt++, conv++) {
+            double temp = *in / *wt; // Temporary product
+            *sumI += temp;
+            *sumII += *in * temp;
+            *sumIC += *conv * temp;
+        }
+    }
+    return;
+}
+
+// Accumulate convolution sums for a stamp
+static void accumulateConvolutions(double *sumC, // Sum of conv(x)/sigma(x)^2
+                                   double *sumCC, // Sum of conv(x)^2/sigma(x)^2
+                                   const pmSubtractionStamp *stamp, // Stamp with input and weight
+                                   int kernelIndex, // Index for kernel component
+                                   int footprint, // Size of region of interest
+                                   pmSubtractionMode mode // Mode of subtraction
+    )
+{
+    psKernel *weight = stamp->weight;   // Weight, sigma(x)^2
+    psKernel *convolution = selectConvolution(stamp, kernelIndex, mode); // Convolution of interest
+
+    for (int y = -footprint; y <= footprint; y++) {
+        psF32 *wt = &weight->kernel[y][-footprint]; // Dereference weight
+        psF32 *conv = &convolution->kernel[y][-footprint]; // Dereference convolution
+        for (int x = -footprint; x <= footprint; x++, wt++, conv++) {
+            double convNoise = *conv / *wt; // Temporary product
+            *sumC += convNoise;
+            *sumCC += *conv * convNoise;
+        }
+    }
+    return;
+}
+
+static double accumulateChi2(const psKernel *target, // Target stamp
+                             pmSubtractionStamp *stamp, // Stamp with weight
+                             int kernelIndex, // Index for kernel component
+                             double coeff, // Coefficient of convolution
+                             double bg,  // Background term
+                             int footprint, // Size of region of interest
+                             pmSubtractionMode mode // Mode of subtraction
+    )
+{
+    double chi2 = 0.0;
+    psKernel *weight = stamp->weight;   // Weight, sigma(x)^2
+    psKernel *convolution = selectConvolution(stamp, kernelIndex, mode); // Convolution of interest
+
+    for (int y = -footprint; y <= footprint; y++) {
+        psF32 *in = &target->kernel[y][-footprint]; // Dereference input
+        psF32 *wt = &weight->kernel[y][-footprint]; // Dereference weight
+        psF32 *conv = &convolution->kernel[y][-footprint]; // Dereference convolution
+        for (int x = -footprint; x <= footprint; x++, in++, wt++, conv++) {
+            chi2 += PS_SQR(*in - bg - coeff * *conv) / *wt;
+        }
+    }
+
+    return chi2;
+}
+
+// Return the initial value of chi^2
+static double initialChi2(const psKernel *target, // Target stamp
+                          const pmSubtractionStamp *stamp, // Stamp with weight
+                          int footprint, // Size of convolution
+                          pmSubtractionMode mode // Mode of subtraction
+    )
+{
+    psKernel *weight = stamp->weight;   // Weight map
+    psKernel *source;                   // Source stamp
+    switch (mode) {
+      case PM_SUBTRACTION_MODE_1:
+        source = stamp->image1;
+        break;
+      case PM_SUBTRACTION_MODE_2:
+        source = stamp->image2;
+        break;
+      default:
+        psAbort("Unsupported subtraction mode: %x", mode);
+    }
+
+    double chi2 = 0.0;                  // Chi^2
+    for (int y = -footprint; y <= footprint; y++) {
+        psF32 *in = &target->kernel[y][-footprint]; // Dereference input
+        psF32 *wt = &weight->kernel[y][-footprint]; // Dereference weight
+        psF32 *ref = &source->kernel[y][-footprint]; // Derference reference
+        for (int x = -footprint; x <= footprint; x++, in++, wt++, ref++) {
+            float diff = *in - *ref;    // Temporary value
+            chi2 += PS_SQR(diff) / *wt;
+        }
+    }
+
+    return chi2;
+}
+
+// Subtract a convolution from the input
+static void subtractConvolution(psKernel *target, // Target stamp
+                                const pmSubtractionStamp *stamp, // Stamp with weight
+                                int kernelIndex, // Index for kernel component
+                                float coeff, // Coefficient of subtraction
+                                float bg, // Background term
+                                int footprint, // Size of region of interest
+                                pmSubtractionMode mode // Mode of subtraction
+    )
+{
+    psKernel *convolution = selectConvolution(stamp, kernelIndex, mode); // Convolution of interest
+    for (int y = -footprint; y <= footprint; y++) {
+        psF32 *in = &target->kernel[y][-footprint]; // Dereference input
+        psF32 *conv = &convolution->kernel[y][-footprint]; // Dereference convolution
+        for (int x = -footprint; x <= footprint; x++, in++, conv++) {
+            *in -= *conv * coeff + bg;
+        }
+    }
+
+    return;
+}
+
+
+pmSubtractionKernels *pmSubtractionKernelsOptimumISIS(pmSubtractionKernelsType type, int size, int inner,
+                                                      int spatialOrder, const psVector *fwhms, int maxOrder,
+                                                      const pmSubtractionStampList *stamps, int footprint,
+                                                      float tolerance, float penalty, pmSubtractionMode mode)
+{
+    if (type != PM_SUBTRACTION_KERNEL_ISIS && type != PM_SUBTRACTION_KERNEL_GUNK) {
+        psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Invalid kernel type: %x\n", type);
+        return NULL;
+    }
+    PS_ASSERT_INT_NONNEGATIVE(size, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(inner, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(spatialOrder, NULL);
+    PS_ASSERT_VECTOR_NON_NULL(fwhms, NULL);
+    PS_ASSERT_VECTOR_TYPE(fwhms, PS_TYPE_F32, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(maxOrder, NULL);
+    PM_ASSERT_SUBTRACTION_STAMP_LIST_NON_NULL(stamps, NULL);
+    PS_ASSERT_INT_NONNEGATIVE(footprint, NULL);
+    PS_ASSERT_FLOAT_LARGER_THAN(tolerance, 0.0, NULL);
+
+    // Generate the kernels to test
+    int numGaussians = fwhms->n;       // Number of Gaussians
+    int numKernels = numGaussians * (maxOrder + 1) * (maxOrder + 2) / 2; // Number of kernel components
+    psString params = NULL;             // Parameter, for description
+    for (int i = 0; i < numGaussians; i++) {
+        psStringAppend(&params, "%.2f,", fwhms->data.F32[i]);
+    }
+    params[strlen(params) - 1] = '\0';
+
+    psVector *orders = psVectorAlloc(numGaussians, PS_TYPE_S32); // Polynomial orders
+    psVectorInit(orders, maxOrder);
+    pmSubtractionKernels *kernels = p_pmSubtractionKernelsRawISIS(size, spatialOrder, fwhms, orders,
+                                                                  penalty, mode); // Kernels
+    psFree(orders);
+    psFree(kernels->description);
+    kernels->description = NULL;
+    psStringAppend(&kernels->description, "OptISIS(%d,(%s),%d)", size, params, spatialOrder);
+    psFree(params);
+
+    // Need to save the stamp inputs --- we're changing the values!
+    int numStamps = stamps->num;        // Number of stamps
+    psArray *targets = psArrayAlloc(numStamps); // Deep copies of the targets
+    psVector *badStamps = psVectorAlloc(numStamps, PS_TYPE_U8); // Mark the bad stamps
+    psVectorInit(badStamps, 0);
+    for (int i = 0; i < numStamps; i++) {
+        pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+        if (stamp->status != PM_SUBTRACTION_STAMP_CALCULATE && stamp->status != PM_SUBTRACTION_STAMP_USED) {
+            badStamps->data.U8[i] = 0xff;
+            continue;
+        }
+        psKernel *target;               // Target image of interest
+        switch (mode) {
+          case PM_SUBTRACTION_MODE_1:
+            target = stamp->image2;
+            break;
+          case PM_SUBTRACTION_MODE_2:
+            target = stamp->image1;
+            break;
+          default:
+            psAbort("Unsupported subtraction mode: %x", mode);
+        }
+        psImage *copy = psImageCopy(NULL, target->image, PS_TYPE_F32); // Copy of the image
+        targets->data[i] = psKernelAllocFromImage(copy, size + footprint, size + footprint);
+        psFree(copy);                   // Drop reference
+    }
+
+    // Generate the convolutions, accumulate sums, and measure initial chi^2
+    double sum1 = 0.0;                  // sum of 1/sigma(x,y)^2
+    psVector *sumC = psVectorAlloc(numKernels, PS_TYPE_F64); // sum of R(x)*k(u)/sigma(x)^2
+    psVector *sumCC = psVectorAlloc(numKernels, PS_TYPE_F64); // sum of [R(x)*k(u)]^2/sigma(x)^2
+    psVectorInit(sumC, 0.0);
+    psVectorInit(sumCC, 0.0);
+    double lastChi2 = 0.0;              // Chi^2 from last iteration
+    int numPixels = 0;                  // Number of pixels contributing to chi^2
+    for (int i = 0; i < numStamps; i++) {
+        pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+        if (badStamps->data.U8[i]) {
+            continue;
+        }
+        if (!pmSubtractionConvolveStamp(stamp, kernels, footprint)) {
+            psError(PS_ERR_UNKNOWN, false, "Unable to convolve stamp %d.", i);
+            psFree(targets);
+            psFree(kernels);
+            psFree(badStamps);
+            return NULL;
+        }
+
+        // This sum is invariant to the kernel
+        psKernel *weight = stamp->weight; // Weight map for stamp
+        for (int v = -footprint; v <= footprint; v++) {
+            psF32 *wt = &weight->kernel[v][-footprint]; // Dereference weight map
+            for (int u = -footprint; u <= footprint; u++, wt++) {
+                sum1 += 1.0 / *wt;
+            }
+        }
+        if (!isfinite(sum1)) {
+            psError(PS_ERR_BAD_PARAMETER_VALUE, true,
+                    "Sum of 1/sigma^2 is non-finite for stamp %d (%d,%d)\n",
+                    i, (int)stamp->x, (int)stamp->y);
+            psFree(targets);
+            psFree(kernels);
+            psFree(badStamps);
+            return NULL;
+        }
+
+        for (int j = 0; j < numKernels; j++) {
+            accumulateConvolutions(&sumC->data.F64[j], &sumCC->data.F64[j], stamp, j, footprint, mode);
+        }
+
+        lastChi2 += initialChi2(targets->data[i], stamp, footprint, mode);
+        numPixels += PS_SQR(2 * footprint + 1);
+    }
+    lastChi2 /= numPixels;
+
+    // Rank the kernel components
+    psVector *ranking = psVectorAlloc(numKernels, PS_TYPE_S32); // Ranking of the kernel components
+    psVectorInit(ranking, -1);
+    int cutIndex = -1;                  // Index at which to cut off kernels
+    for (int iter = 0; iter < numKernels; iter++) {
+        int bestIndex = -1;             // Index of best kernel component
+        double bestChi2 = INFINITY;     // Value of best chi^2
+        double bestCoeff = 0;           // Value of best coefficient
+        double bestBG = 0;              // Value of best background
+
+        for (int i = 0; i < numKernels; i++) {
+            if (ranking->data.S32[i] >= 0) {
+                continue;
+            }
+
+            double sumI = 0.0;          // sum of I(x)/sigma(x)^2
+            double sumII = 0.0;         // sum of I(x)^2/sigma(x)^2
+            double sumIC = 0.0;         // sum of I(x)C(x)/sigma(x)^2
+
+            for (int j = 0; j < numStamps; j++) {
+                if (badStamps->data.U8[j]) {
+                    continue;
+                }
+                pmSubtractionStamp *stamp = stamps->stamps->data[j]; // Stamp of interest
+                accumulateCross(&sumI, &sumII, &sumIC, stamp, targets->data[j], i, footprint, mode);
+            }
+
+            double invDet = 1.0 / (sum1 * sumCC->data.F64[i] - PS_SQR(sumC->data.F64[i])); // Determinant^-1
+            double coeff = invDet * (sum1 * sumIC - sumC->data.F64[i] * sumI); // Coefficient for kernel
+            double bg = invDet * (sumCC->data.F64[i] * sumI - sumC->data.F64[i] * sumIC); // Background
+
+            double chi2 = 0.0;          // Chi^2
+            for (int j = 0; j < numStamps; j++) {
+                if (badStamps->data.U8[j]) {
+                    continue;
+                }
+                pmSubtractionStamp *stamp = stamps->stamps->data[j]; // Stamp of interest
+                chi2 += accumulateChi2(targets->data[j], stamp, i, coeff, bg, footprint, mode);
+            }
+
+            if (chi2 < bestChi2) {
+                bestIndex = i;
+                bestCoeff = coeff;
+                bestChi2 = chi2;
+                bestBG = bg;
+            }
+
+            psTrace("psModules.imcombine", 8, "%d: %lf %lf %lf %lf %lf %lf\n", i, sum1, sumI, sumII, sumIC,
+                    sumC->data.F64[i], sumCC->data.F64[i]);
+            psTrace("psModules.imcombine", 6, "%d: %lf %lf %lf\n", i, coeff, bg, chi2);
+        }
+        bestChi2 /= numPixels;
+
+        if (bestIndex == -1) {
+            psError(PS_ERR_UNKNOWN, false, "Unable to find best kernel component in round %d.", iter);
+            psFree(targets);
+            psFree(sumC);
+            psFree(sumCC);
+            psFree(ranking);
+            psFree(kernels);
+            psFree(badStamps);
+            return NULL;
+        }
+
+        // And the winner is....
+        ranking->data.S32[bestIndex] = iter;
+        // Remove its contribution, and don't include it in the future.
+        for (int j = 0; j < numStamps; j++) {
+            if (badStamps->data.U8[j]) {
+                continue;
+            }
+            pmSubtractionStamp *stamp = stamps->stamps->data[j]; // Stamp of interest
+            subtractConvolution(targets->data[j], stamp, bestIndex, bestCoeff, bestBG, footprint, mode);
+        }
+
+        double diff = lastChi2 - bestChi2; // Difference in chi^2 between iterations
+
+        psTrace("psModules.imcombine", 3, "The winner of round %d is %d (%f,%d,%d): %lf (%lf) --> %lf\n",
+                iter, bestIndex, kernels->widths->data.F32[bestIndex], kernels->u->data.S32[bestIndex],
+                kernels->v->data.S32[bestIndex], bestCoeff, diff, bestChi2);
+
+        if (fabsf(diff) < tolerance) {
+            cutIndex = iter;
+            break;
+        }
+
+        lastChi2 = bestChi2;
+    }
+    psFree(targets);
+    psFree(sumC);
+    psFree(sumCC);
+
+    if (cutIndex < 0) {
+        psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Unable to converge to tolerance %g\n", tolerance);
+        psFree(ranking);
+        psFree(kernels);
+        psFree(badStamps);
+        return NULL;
+    }
+
+    int newSize = cutIndex + 1;         // Size of new kernel basis set
+    psTrace("psModules.imcombine", 2, "Accepting %d kernels.\n", newSize);
+    psVector *uNew = psVectorAlloc(newSize, PS_TYPE_S32);
+    psVector *vNew = psVectorAlloc(newSize, PS_TYPE_S32);
+    psVector *widthsNew = psVectorAlloc(newSize, PS_TYPE_F32);
+    psArray *preCalcNew = psArrayAlloc(newSize);
+    psArray *convNew = psArrayAlloc(numStamps);
+    for (int i = 0; i < numStamps; i++) {
+        if (badStamps->data.U8[i]) {
+            continue;
+        }
+        convNew->data[i] = psArrayAlloc(newSize);
+    }
+
+    for (int i = 0; i < numKernels; i++) {
+        int rank = ranking->data.S32[i]; // This kernel component's ranking
+        if (rank >= 0 && rank < newSize) {
+            uNew->data.S32[rank] = kernels->u->data.S32[i];
+            vNew->data.S32[rank] = kernels->v->data.S32[i];
+            widthsNew->data.F32[rank] = kernels->widths->data.F32[i];
+            preCalcNew->data[rank] = psMemIncrRefCounter(kernels->preCalc->data[i]);
+
+            for (int j = 0; j < numStamps; j++) {
+                if (badStamps->data.U8[j]) {
+                    continue;
+                }
+                pmSubtractionStamp *stamp = stamps->stamps->data[j]; // Stamp of interest
+                psArray *convolutions = convNew->data[j]; // Convolutions for this stamp
+                convolutions->data[rank] = psMemIncrRefCounter(selectConvolution(stamp, i, mode));
+            }
+        }
+    }
+    psFree(kernels->u);
+    psFree(kernels->v);
+    psFree(kernels->widths);
+    psFree(kernels->preCalc);
+    kernels->u = uNew;
+    kernels->v = vNew;
+    kernels->widths = widthsNew;
+    kernels->preCalc = preCalcNew;
+    kernels->num = newSize;
+
+    for (int i = 0; i < numStamps; i++) {
+        if (badStamps->data.U8[i]) {
+            continue;
+        }
+        pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+        psFree(stamp->convolutions1);
+        psFree(stamp->convolutions2);
+        switch (mode) {
+          case PM_SUBTRACTION_MODE_1:
+            stamp->convolutions1 = convNew->data[i];
+            stamp->convolutions2 = NULL;
+            break;
+          case PM_SUBTRACTION_MODE_2:
+            stamp->convolutions1 = NULL;
+            stamp->convolutions2 = convNew->data[i];
+            break;
+          default:
+            psAbort("Unsupported subtraction mode: %x", mode);
+        }
+    }
+
+    psFree(badStamps);
+    psFree(ranking);
+
+    // Maintain photometric scaling
+    if (type == PM_SUBTRACTION_KERNEL_ISIS) {
+        psKernel *subtract = kernels->preCalc->data[0]; // Kernel to subtract from the rest
+        for (int i = 1; i < newSize; i++) {
+            if (kernels->u->data.S32[i] % 2 == 0 && kernels->v->data.S32[i] % 2 == 0) {
+                psKernel *kernel = kernels->preCalc->data[i]; // Kernel of interest
+                psBinaryOp(kernel->image, kernel->image, "-", subtract->image);
+            }
+        }
+    } else if (type == PM_SUBTRACTION_KERNEL_GUNK) {
+        psStringPrepend(&kernels->description, "GUNK=");
+        psStringAppend(&kernels->description, "+POIS(%d,%d)", inner, spatialOrder);
+
+        for (int i = 0; i < newSize; i++) {
+            if (kernels->u->data.S32[i] % 2 == 0 && kernels->v->data.S32[i] % 2 == 0) {
+                psKernel *kernel = kernels->preCalc->data[i]; // Kernel of interest
+                kernel->kernel[0][0] -= 1.0;
+            }
+        }
+
+        if (!p_pmSubtractionKernelsAddGrid(kernels, numGaussians, inner)) {
+            psAbort("Should never get here.");
+        }
+
+        kernels->type = PM_SUBTRACTION_KERNEL_GUNK;
+    }
+
+    return kernels;
+}
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionParams.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionParams.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionParams.h	(revision 19481)
@@ -0,0 +1,22 @@
+#ifndef PM_SUBTRACTION_PARAMS_H
+#define PM_SUBTRACTION_PARAMS_H
+
+#include <pslib.h>
+#include <pmSubtractionKernels.h>
+#include <pmSubtractionStamps.h>
+
+/// Generate a set of optimum kernels for ISIS (or GUNK)
+pmSubtractionKernels *pmSubtractionKernelsOptimumISIS(pmSubtractionKernelsType type, ///< Kernel type
+                                                      int size, ///< Half-size of kernel
+                                                      int inner, ///< Inner radius for GUNK
+                                                      int spatialOrder, ///< Spatial polynomial order
+                                                      const psVector *fwhms, ///< Gaussian FWHMs to try
+                                                      int maxOrder, ///< Maximum polynomial order
+                                                      const pmSubtractionStampList *stamps, ///< Stamps
+                                                      int footprint, ///< Convolution footprint for stamps
+                                                      float tolerance, ///< Maximum difference in chi^2
+                                                      float penalty, ///< Penalty for wideness
+                                                      pmSubtractionMode mode // Mode for subtraction
+    );
+
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionStamps.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionStamps.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionStamps.c	(revision 19481)
@@ -0,0 +1,655 @@
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <stdio.h>
+#include <string.h>
+#include <pslib.h>
+
+// All these includes required to get stamps out of an array of pmSources
+#include "pmMoments.h"
+#include "pmPeaks.h"
+#include "pmResiduals.h"
+#include "pmHDU.h"
+#include "pmFPA.h"
+#include "pmGrowthCurve.h"
+#include "pmTrend2D.h"
+#include "pmPSF.h"
+#include "pmModel.h"
+#include "pmSource.h"
+
+
+#include "pmSubtraction.h"
+#include "pmSubtractionStamps.h"
+
+#define STAMP_LIST_BUFFER 20            // Number of stamps to add to list at a time
+
+#define SOURCE_MASK (PM_SOURCE_MODE_FAIL | PM_SOURCE_MODE_SATSTAR | PM_SOURCE_MODE_BLEND | \
+                     PM_SOURCE_MODE_BADPSF | PM_SOURCE_MODE_DEFECT | PM_SOURCE_MODE_SATURATED | \
+                     PM_SOURCE_MODE_CR_LIMIT) // Mask for bad sources
+#define SOURCE_FAINTEST 50.0            // Faintest magnitude to consider
+
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Private (file-static) functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+// Free function for pmSubtractionStampList
+static void subtractionStampListFree(pmSubtractionStampList *list // Stamp list to free
+                                     )
+{
+    psFree(list->stamps);
+    psFree(list->regions);
+    psFree(list->x);
+    psFree(list->y);
+    psFree(list->flux);
+}
+
+// Free function for pmSubtractionStamp
+static void subtractionStampFree(pmSubtractionStamp *stamp // Stamp to free
+                                 )
+{
+    psFree(stamp->image1);
+    psFree(stamp->image2);
+    psFree(stamp->weight);
+    psFree(stamp->convolutions1);
+    psFree(stamp->convolutions2);
+
+    psFree(stamp->matrix1);
+    psFree(stamp->matrix2);
+    psFree(stamp->matrixX);
+    psFree(stamp->vector1);
+    psFree(stamp->vector2);
+
+}
+
+// Is this region OK?
+static bool checkStampRegion(int x, int y, // Coordinates of stamp
+                             const psRegion *region // Region of interest
+                             )
+{
+    if (!region) {
+        return true;
+    }
+    return (x < region->x0 || x > region->x1 || y < region->y0 || y > region->y1) ?
+        false : true;
+}
+
+// Is this position unmasked?
+static bool checkStampMask(int x, int y, // Coordinates of stamp
+                           const psImage *mask, // Mask
+                           pmSubtractionMode mode, // Mode for subtraction
+                           int footprint // Footprint to check for Bad Stuff
+                           )
+{
+    if (!mask) {
+        return true;
+    }
+
+    bool clean = true;                  // Is the footprint clean?
+    int numCols = mask->numCols, numRows = mask->numRows; // Size of image
+
+    // Check the footprint bounds
+    if (x < footprint || x >= numCols - footprint || y < footprint || y >= numRows - footprint) {
+        clean = false;
+    }
+
+    // Determine mask value
+    psMaskType maskVal = PM_SUBTRACTION_MASK_BORDER | PM_SUBTRACTION_MASK_BAD_1 | PM_SUBTRACTION_MASK_BAD_2;
+    switch (mode) {
+      case PM_SUBTRACTION_MODE_1:
+        maskVal |= PM_SUBTRACTION_MASK_CONVOLVE_1;
+        break;
+      case PM_SUBTRACTION_MODE_2:
+        maskVal |= PM_SUBTRACTION_MASK_CONVOLVE_2;
+        break;
+      case PM_SUBTRACTION_MODE_UNSURE:
+      case PM_SUBTRACTION_MODE_DUAL:
+        maskVal |= PM_SUBTRACTION_MASK_CONVOLVE_1 | PM_SUBTRACTION_MASK_CONVOLVE_2;
+        break;
+      default:
+        psAbort("Unsupported subtraction mode: %x", mode);
+    }
+
+    // Check the immediate pixel
+    if (clean && (mask->data.PS_TYPE_MASK_DATA[y][x] & (maskVal | PM_SUBTRACTION_MASK_REJ))) {
+        clean = false;
+    }
+
+    int xMin = PS_MAX(x - footprint, 0), xMax = PS_MIN(x + footprint, numCols - 1); // Bounds in x
+    int yMin = PS_MAX(y - footprint, 0), yMax = PS_MIN(y + footprint, numRows - 1); // Bounds in y
+
+    // Check the footprint
+    if (clean) {
+        for (int j = yMin; j <= yMax; j++) {
+            for (int i = xMin; i <= xMax; i++) {
+                if (mask->data.PS_TYPE_MASK_DATA[j][i] & maskVal) {
+                    clean = false;
+                    goto CHECK_STAMP_MASK_DONE;
+                }
+            }
+        }
+    }
+CHECK_STAMP_MASK_DONE:
+
+    if (!clean) {
+        // Mask the footprint, so we don't select something near it again
+        for (int j = yMin; j <= yMax; j++) {
+            for (int i = xMin; i <= xMax; i++) {
+                mask->data.PS_TYPE_MASK_DATA[j][i] |= PM_SUBTRACTION_MASK_REJ;
+            }
+        }
+    }
+
+    return clean;
+}
+
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// Public functions
+//////////////////////////////////////////////////////////////////////////////////////////////////////////////
+
+pmSubtractionStampList *pmSubtractionStampListAlloc(int numCols, int numRows, const psRegion *region,
+                                                    int footprint, float spacing)
+{
+    pmSubtractionStampList *list = psAlloc(sizeof(pmSubtractionStampList)); // Stamp list to return
+    psMemSetDeallocator(list, (psFreeFunc)subtractionStampListFree);
+
+    // Get region in which to find stamps: [xMin:xMax,yMin:yMax]
+    int xMin = 0, xMax = numCols, yMin = 0, yMax = numRows;
+    if (region) {
+        xMin = PS_MAX(region->x0, xMin);
+        xMax = PS_MIN(region->x1, xMax);
+        yMin = PS_MAX(region->y0, yMin);
+        yMax = PS_MIN(region->y1, yMax);
+    }
+    int xSize = xMax - xMin, ySize = yMax - yMin; // Size of region of interest
+    int xStamps = (float)xSize / spacing + 1, yStamps = (float)ySize / spacing + 1; // Number of stamps
+
+    list->num = xStamps * yStamps;
+    list->stamps = psArrayAlloc(list->num);
+    list->regions = psArrayAlloc(list->num);
+
+    for (int y = 0, index = 0; y < yStamps; y++) {
+        int yStart = yMin + y * ((float)ySize / (float)(yStamps)); // Subregion starts here
+        int yStop = yMin + (y + 1) * ((float)ySize / (float)(yStamps)) - 1; // Subregion stops here
+        assert(yStart >= yMin && yStop < yMax);
+
+        for (int x = 0; x < xStamps; x++, index++) {
+            int xStart = xMin + x * ((float)xSize / (float)(xStamps)); // Subregion starts here
+            int xStop = xMin + (x + 1) * ((float)xSize / (float)(xStamps)) - 1; // Subregion stops here
+            assert(xStart >= xMin && xStop < xMax);
+
+            list->stamps->data[index] = pmSubtractionStampAlloc();
+            psTrace("psModules.imcombine", 6, "Stamp region %d: [%d:%d,%d:%d]\n",
+                    index, xStart, xStop, yStart, yStop);
+            list->regions->data[index] = psRegionAlloc(xStart, xStop, yStart, yStop);
+        }
+    }
+
+    list->x = NULL;
+    list->y = NULL;
+    list->flux = NULL;
+    list->footprint = footprint;
+
+    return list;
+}
+
+pmSubtractionStamp *pmSubtractionStampAlloc(void)
+{
+    pmSubtractionStamp *stamp = psAlloc(sizeof(pmSubtractionStamp)); // Stamp to return
+    psMemSetDeallocator(stamp, (psFreeFunc)subtractionStampFree);
+
+    stamp->x = NAN;
+    stamp->y = NAN;
+    stamp->flux = NAN;
+    stamp->xNorm = NAN;
+    stamp->yNorm = NAN;
+    stamp->status = PM_SUBTRACTION_STAMP_INIT;
+
+    stamp->image1 = NULL;
+    stamp->image2 = NULL;
+    stamp->weight = NULL;
+    stamp->convolutions1 = NULL;
+    stamp->convolutions2 = NULL;
+
+    stamp->matrix1 = NULL;
+    stamp->matrix2 = NULL;
+    stamp->matrixX = NULL;
+    stamp->vector1 = NULL;
+    stamp->vector2 = NULL;
+
+    return stamp;
+}
+
+
+pmSubtractionStampList *pmSubtractionStampsFind(pmSubtractionStampList *stamps, const psImage *image,
+                                                const psImage *subMask, const psRegion *region,
+                                                float threshold, int footprint, float spacing,
+                                                pmSubtractionMode mode)
+{
+    PS_ASSERT_IMAGE_NON_NULL(image, NULL);
+    PS_ASSERT_IMAGE_TYPE(image, PS_TYPE_F32, NULL);
+    if (subMask) {
+        PS_ASSERT_IMAGE_NON_NULL(subMask, NULL);
+        PS_ASSERT_IMAGES_SIZE_EQUAL(image, subMask, NULL);
+        PS_ASSERT_IMAGE_TYPE(subMask, PS_TYPE_MASK, NULL);
+    }
+    PS_ASSERT_INT_NONNEGATIVE(footprint, NULL);
+    PS_ASSERT_FLOAT_LARGER_THAN(spacing, 0.0, NULL);
+    if (region) {
+        if (psRegionIsNaN(*region)) {
+            psString string = psRegionToString(*region);
+            psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Input region (%s) contains NAN values", string);
+            psFree(string);
+            return false;
+        }
+        if (region->x0 < 0 || region->x1 > image->numCols ||
+            region->y0 < 0 || region->y1 > image->numRows) {
+            psString string = psRegionToString(*region);
+            psError(PS_ERR_BAD_PARAMETER_VALUE, true, "Input region (%s) does not fit in image (%dx%d)",
+                    string, image->numCols, image->numRows);
+            psFree(string);
+            return false;
+        }
+    }
+
+    int numRows = image->numRows, numCols = image->numCols; // Size of image
+
+    if (!stamps) {
+        stamps = pmSubtractionStampListAlloc(numCols, numRows, region, footprint, spacing);
+    }
+
+    int numStamps = stamps->num;        // Number of stamp regions
+    int numFound = 0;                   // Number of stamps found
+    int numValid = 0;                   // Number of valid regions
+    for (int i = 0; i < numStamps; i++) {
+        pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+
+        // Only find a new stamp if we need to
+        if (stamp->status != PM_SUBTRACTION_STAMP_REJECTED && stamp->status != PM_SUBTRACTION_STAMP_INIT) {
+            continue;
+        }
+        numValid++;
+
+        float xStamp = 0, yStamp = 0;   // Coordinates of stamp
+        float fluxStamp = NAN;          // Flux of stamp
+        bool goodStamp = false;         // Found a good stamp?
+
+        // A couple different ways of finding stamps:
+        if (stamps->x && stamps->y) {
+            // Get the next stamp from the list
+            psVector *xList = stamps->x->data[i], *yList = stamps->y->data[i]; // Coordinate lists
+            psVector *fluxList = stamps->flux->data[i]; // List of stamp fluxes
+
+            // Take stamp off the top of the (sorted) list
+            if (xList->n > 0) {
+                int index = xList->n - 1; // Index of new stamp
+                xStamp = xList->data.F32[index];
+                yStamp = yList->data.F32[index];
+                fluxStamp = fluxList->data.F32[index];
+
+                // Chop off the top of the list
+                xList->n = index;
+                yList->n = index;
+                fluxList->n = index;
+
+                goodStamp = true;
+            }
+        } else {
+            // Use a simple method of automatically finding stamps --- take the highest pixel in the subregion
+            fluxStamp = threshold;
+            psRegion *subRegion = stamps->regions->data[i]; // Sub-region of interest
+            for (int y = subRegion->y0; y <= subRegion->y1; y++) {
+                for (int x = subRegion->x0; x <= subRegion->x1; x++) {
+                    if (checkStampMask(x, y, subMask, mode, footprint) && image->data.F32[y][x] > fluxStamp) {
+                        fluxStamp = image->data.F32[y][x];
+                        xStamp = x;
+                        yStamp = y;
+                        goodStamp = true;
+                    }
+                }
+            }
+        }
+
+        if (goodStamp) {
+            stamp->x = xStamp;
+            stamp->y = yStamp;
+            stamp->flux = fluxStamp;
+
+            // Reset the postage stamps since we're making a new stamp
+            psFree(stamp->image1);
+            psFree(stamp->image2);
+            psFree(stamp->weight);
+            psFree(stamp->convolutions1);
+            psFree(stamp->convolutions2);
+            stamp->image1 = stamp->image2 = stamp->weight = NULL;
+            stamp->convolutions1 = stamp->convolutions2 = NULL;
+
+            stamp->status = PM_SUBTRACTION_STAMP_FOUND;
+            numFound++;
+            psTrace("psModules.imcombine", 5, "Found stamp in subregion %d: %d,%d\n",
+                    i, (int)stamp->x, (int)stamp->y);
+        } else {
+            stamp->status = PM_SUBTRACTION_STAMP_NONE;
+        }
+    }
+
+    if (numValid > 0) {
+        psLogMsg("psModules.imcombine", PS_LOG_INFO, "Found %d stamps", numFound);
+    }
+
+    return stamps;
+}
+
+
+pmSubtractionStampList *pmSubtractionStampsSet(const psVector *x, const psVector *y, const psVector *flux,
+                                               const psImage *image, const psImage *subMask,
+                                               const psRegion *region, int footprint, float spacing,
+                                               pmSubtractionMode mode)
+
+{
+    PS_ASSERT_VECTOR_NON_NULL(x, NULL);
+    PS_ASSERT_VECTOR_TYPE(x, PS_TYPE_F32, NULL);
+    PS_ASSERT_VECTOR_NON_NULL(y, NULL);
+    PS_ASSERT_VECTOR_TYPE(y, PS_TYPE_F32, NULL);
+    PS_ASSERT_VECTORS_SIZE_EQUAL(y, x, NULL);
+    if (flux) {
+        PS_ASSERT_VECTOR_NON_NULL(flux, NULL);
+        PS_ASSERT_VECTOR_TYPE(flux, PS_TYPE_F32, NULL);
+        PS_ASSERT_VECTORS_SIZE_EQUAL(flux, x, NULL);
+    } else {
+        PS_ASSERT_IMAGE_NON_NULL(image, NULL);
+    }
+    if (subMask) {
+        PS_ASSERT_IMAGE_NON_NULL(subMask, NULL);
+        PS_ASSERT_IMAGE_TYPE(subMask, PS_TYPE_MASK, NULL);
+        if (image) {
+            PS_ASSERT_IMAGE_NON_NULL(image, NULL);
+            PS_ASSERT_IMAGES_SIZE_EQUAL(image, subMask, NULL);
+        }
+    }
+    PS_ASSERT_FLOAT_LARGER_THAN(spacing, 0.0, NULL);
+
+    int numStars = x->n;                // Number of stars
+    pmSubtractionStampList *stamps = pmSubtractionStampListAlloc(subMask->numCols, subMask->numRows,
+                                                                 region, footprint, spacing); // Stamp list
+    int numStamps = stamps->num;        // Number of stamps
+
+    // Initialise the lists
+    stamps->x = psArrayAlloc(numStamps);
+    stamps->y = psArrayAlloc(numStamps);
+    stamps->flux = psArrayAlloc(numStamps);
+    for (int i = 0; i < numStamps; i++) {
+        stamps->x->data[i] = psVectorAllocEmpty(STAMP_LIST_BUFFER, PS_TYPE_F32);
+        stamps->y->data[i] = psVectorAllocEmpty(STAMP_LIST_BUFFER, PS_TYPE_F32);
+        stamps->flux->data[i] = psVectorAllocEmpty(STAMP_LIST_BUFFER, PS_TYPE_F32);
+    }
+
+    // Put the stars into their appropriate subregions
+    for (int i = 0; i < numStars; i++) {
+        float xStamp = x->data.F32[i], yStamp = y->data.F32[i]; // Coordinates of stamp
+        int xPix = xStamp + 0.5, yPix = yStamp + 0.5; // Pixel coordinate of stamp
+        if (!checkStampRegion(xPix, yPix, region)) {
+            // It's not in the big region
+            psTrace("psModules.imcombine", 9, "Rejecting input stamp (%d,%d) because outside region",
+                    xPix, yPix);
+            continue;
+        }
+        if (!checkStampMask(xPix, yPix, subMask, mode, footprint)) {
+            // Not a good stamp
+            psTrace("psModules.imcombine", 9, "Rejecting input stamp (%d,%d) because bad mask",
+                    xPix, yPix);
+            continue;
+        }
+
+        bool found = false;
+        for (int j = 0; j < numStamps && !found; j++) {
+            psRegion *subRegion = stamps->regions->data[j]; // Subregion of interest
+            if (checkStampRegion(xPix, yPix, subRegion)) {
+                psVector *xList = stamps->x->data[j], *yList = stamps->y->data[j]; // Pixel lists
+                psVector *fluxList = stamps->flux->data[j]; // Flux list
+
+                int index = xList->n;   // Index of new stamp candidate
+
+                psVectorExtend(xList, STAMP_LIST_BUFFER, 1);
+                psVectorExtend(yList, STAMP_LIST_BUFFER, 1);
+                psVectorExtend(fluxList, STAMP_LIST_BUFFER, 1);
+
+                xList->data.F32[index] = xStamp;
+                yList->data.F32[index] = yStamp;
+
+                if (flux) {
+                    fluxList->data.F32[index] = flux->data.F32[i];
+                } else {
+                    fluxList->data.F32[index] = image->data.F32[yPix][xPix];
+                }
+
+                found = true;
+                psTrace("psModules.imcombine", 9, "Putting input stamp (%d,%d) into subregion %d",
+                        xPix, yPix, j);
+            }
+        }
+
+        if (!found) {
+            psTrace("psModules.imcombine", 9, "Unable to find subregion for stamp (%d,%d)",
+                    xPix, yPix);
+        }
+    }
+
+    // Sort the list by flux, with the brightest last
+    for (int i = 0; i < numStamps; i++) {
+        psVector *xList = stamps->x->data[i], *yList = stamps->y->data[i]; // Pixel lists
+        psVector *fluxList = stamps->flux->data[i]; // Flux list
+
+        psVector *indexes = psVectorSortIndex(NULL, fluxList); // Indices to sort flux
+        int num = indexes->n;           // Number of candidate stamps in this subregion
+
+        psVector *xSorted = psVectorAlloc(num, PS_TYPE_F32); // Sorted version of x list
+        psVector *ySorted = psVectorAlloc(num, PS_TYPE_F32); // Sorted version of y list
+        psVector *fluxSorted = psVectorAlloc(num, PS_TYPE_F32); // Sorted version of flux list
+        for (int j = 0; j < num; j++) {
+            int k = indexes->data.S32[j]; // Sorted index
+            xSorted->data.F32[j] = xList->data.F32[k];
+            ySorted->data.F32[j] = yList->data.F32[k];
+            fluxSorted->data.F32[j] = fluxList->data.F32[k];
+        }
+        psFree(indexes);
+
+        psFree(stamps->x->data[i]);
+        psFree(stamps->y->data[i]);
+        psFree(stamps->flux->data[i]);
+
+        stamps->x->data[i] = xSorted;
+        stamps->y->data[i] = ySorted;
+        stamps->flux->data[i] = fluxSorted;
+    }
+
+
+    return stamps;
+}
+
+
+bool pmSubtractionStampsExtract(pmSubtractionStampList *stamps, psImage *image1, psImage *image2,
+                                psImage *weight, int kernelSize)
+{
+    PM_ASSERT_SUBTRACTION_STAMP_LIST_NON_NULL(stamps, false);
+    PS_ASSERT_IMAGE_NON_NULL(image1, false);
+    PS_ASSERT_IMAGE_TYPE(image1, PS_TYPE_F32, false);
+    if (image2) {
+        PS_ASSERT_IMAGE_NON_NULL(image2, false);
+        PS_ASSERT_IMAGES_SIZE_EQUAL(image2, image1, false);
+        PS_ASSERT_IMAGE_TYPE(image2, PS_TYPE_F32, false);
+    }
+    PS_ASSERT_IMAGE_NON_NULL(weight, false);
+    PS_ASSERT_IMAGES_SIZE_EQUAL(weight, image1, false);
+    PS_ASSERT_IMAGE_TYPE(weight, PS_TYPE_F32, false);
+    PS_ASSERT_INT_NONNEGATIVE(kernelSize, false);
+
+    int numCols = image1->numCols, numRows = image1->numRows; // Size of images
+    int size = kernelSize + stamps->footprint; // Size of postage stamps
+
+    for (int i = 0; i < stamps->num; i++) {
+        pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+        if (!stamp || stamp->status != PM_SUBTRACTION_STAMP_FOUND) {
+            continue;
+        }
+
+        if (isnan(stamp->xNorm)) {
+            stamp->xNorm = 2.0 * (stamp->x - (float)numCols/2.0) / (float)numCols;
+        }
+        if (isnan(stamp->yNorm)) {
+            stamp->yNorm = 2.0 * (stamp->y - (float)numRows/2.0) / (float)numRows;
+        }
+
+        int x = stamp->x + 0.5, y = stamp->y + 0.5; // Stamp coordinates
+        if (x < size || x > numCols - size || y < size || y > numRows - size) {
+            psError(PS_ERR_UNKNOWN, false, "Stamp %d (%d,%d) is within the image border.\n", i, x, y);
+            return false;
+        }
+
+        // Catch memory leaks --- these should have been freed and NULLed before
+        assert(stamp->image1 == NULL);
+        assert(stamp->image2 == NULL);
+        assert(stamp->weight == NULL);
+
+        psRegion region = psRegionSet(x - size, x + size + 1, y - size, y + size + 1); // Region of interest
+
+        psImage *sub1 = psImageSubset(image1, region); // Subimage with stamp
+        stamp->image1 = psKernelAllocFromImage(sub1, size, size);
+        psFree(sub1);                   // Drop reference
+
+        if (image2) {
+            psImage *sub2 = psImageSubset(image2, region); // Subimage with stamp
+            stamp->image2 = psKernelAllocFromImage(sub2, size, size);
+            psFree(sub2);               // Drop reference
+        }
+
+        psImage *wtSub = psImageSubset(weight, region); // Subimage with stamp
+        stamp->weight = psKernelAllocFromImage(wtSub, size, size);
+        psFree(wtSub);                  // Drop reference
+
+        stamp->status = PM_SUBTRACTION_STAMP_CALCULATE;
+    }
+
+    return true;
+}
+
+#if 0
+bool pmSubtractionStampsGenerate(pmSubtractionStampList *stamps, float fwhm, int kernelSize)
+{
+    PM_ASSERT_SUBTRACTION_STAMP_LIST_NON_NULL(stamps, false);
+    PS_ASSERT_FLOAT_LARGER_THAN(fwhm, 0.0, false);
+    PS_ASSERT_INT_NONNEGATIVE(kernelSize, false);
+
+    int size = kernelSize + stamps->footprint; // Size of postage stamps
+    int num = stamps->num;              // Number of stamps
+    float sigma = fwhm / (2.0 * sqrtf(2.0 * log(2.0))); // Gaussian sigma
+
+    for (int i = 0; i < num; i++) {
+        pmSubtractionStamp *stamp = stamps->stamps->data[i]; // Stamp of interest
+        if (!(stamp->status & PM_SUBTRACTION_STAMP_CALCULATE)) {
+            continue;
+        }
+
+        float x = stamp->x, y = stamp->y; // Coordinates of stamp
+        float flux = stamp->flux; // Flux of star
+        if (!isfinite(flux)) {
+            psWarning("Unable to generate PSF for stamp %d --- bad flux.", i);
+            stamp->status = PM_SUBTRACTION_STAMP_REJECTED;
+            continue;
+        }
+
+        float xStamp = x - (int)(x + 0.5); // x coordinate of star in stamp frame
+        float yStamp = y - (int)(y + 0.5); // y coordinate of star in stamp frame
+
+        psFree(stamp->image2);
+        stamp->image2 = psKernelAlloc(-size, size, -size, size);
+        psKernel *target = stamp->image2; // Target stamp
+
+        // Put in a Waussian, just for fun!
+        for (int v = -size; v <= size; v++) {
+            for (int u = -size; u <= size; u++) {
+                float z = (PS_SQR(u + xStamp) + PS_SQR(v + yStamp)) / (2.0 * PS_SQR(sigma));
+                target->kernel[v][u] = flux / sigma * 0.5 * M_2_SQRTPI * M_SQRT1_2 / (1.0 + z + PS_SQR(z));
+            }
+        }
+
+    }
+
+    return true;
+}
+#endif
+
+pmSubtractionStampList *pmSubtractionStampsSetFromSources(const psArray *sources, const psImage *subMask,
+                                                          const psRegion *region, int footprint,
+                                                          float spacing, pmSubtractionMode mode)
+{
+    PS_ASSERT_ARRAY_NON_NULL(sources, NULL);
+    // Let pmSubtractionStampsSet take care of the rest of the assertions
+
+    int numIn = sources->n;             // Number of sources in input list
+
+    psVector *x = psVectorAllocEmpty(numIn, PS_TYPE_F32); // x coordinates
+    psVector *y = psVectorAllocEmpty(numIn, PS_TYPE_F32); // y coordinates
+    psVector *flux = psVectorAllocEmpty(numIn, PS_TYPE_F32); // Fluxes
+
+    int numOut = 0;                     // Number of sources in output list
+    for (int i = 0; i < numIn; i++) {
+        pmSource *source = sources->data[i]; // Source of interest
+        if (!source || (source->mode & SOURCE_MASK) || !isfinite(source->psfMag) ||
+            source->psfMag > SOURCE_FAINTEST) {
+            continue;
+        }
+        if (source->modelPSF) {
+            x->data.F32[numOut] = source->modelPSF->params->data.F32[PM_PAR_XPOS];
+            y->data.F32[numOut] = source->modelPSF->params->data.F32[PM_PAR_YPOS];
+        } else {
+            x->data.F32[numOut] = source->peak->xf;
+            y->data.F32[numOut] = source->peak->yf;
+        }
+        flux->data.F32[numOut] = powf(10.0, -0.4 * source->psfMag);
+        numOut++;
+    }
+    x->n = numOut;
+    y->n = numOut;
+    flux->n = numOut;
+
+    pmSubtractionStampList *stamps = pmSubtractionStampsSet(x, y, flux, NULL, subMask, region,
+                                                            footprint, spacing, mode); // Stamps to return
+    psFree(x);
+    psFree(y);
+    psFree(flux);
+
+    if (!stamps) {
+        psError(PS_ERR_UNKNOWN, false, "Unable to set stamps from sources.");
+    }
+
+    return stamps;
+}
+
+
+pmSubtractionStampList *pmSubtractionStampsSetFromFile(const char *filename, const psImage *image,
+                                                       const psImage *subMask, const psRegion *region,
+                                                       int footprint, float spacing, pmSubtractionMode mode)
+{
+    PS_ASSERT_STRING_NON_EMPTY(filename, NULL);
+    // Let pmSubtractionStampsSet take care of the rest of the assertions
+
+    psArray *data = psVectorsReadFromFile(filename, "%f %f");
+    if (!data) {
+        psError(PS_ERR_IO, false, "Unable to read stamps file %s", filename);
+        return NULL;
+    }
+    psVector *x = data->data[0], *y = data->data[1]; // Stamp positions
+
+    // Correct for IRAF/FITS (unit-offset) positions to C (zero-offset) positions
+    psBinaryOp(x, x, "-", psScalarAlloc(1.0, PS_TYPE_F32));
+    psBinaryOp(y, y, "-", psScalarAlloc(1.0, PS_TYPE_F32));
+
+    pmSubtractionStampList *stamps = pmSubtractionStampsSet(x, y, NULL, image, subMask, region, footprint,
+                                                            spacing, mode);
+    psFree(data);
+
+    return stamps;
+
+}
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionStamps.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionStamps.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionStamps.h	(revision 19481)
@@ -0,0 +1,126 @@
+#ifndef PM_SUBTRACTION_STAMPS_H
+#define PM_SUBTRACTION_STAMPS_H
+
+#include <pslib.h>
+
+#include "pmSubtractionKernels.h"
+
+/// Status of stamp
+typedef enum {
+    PM_SUBTRACTION_STAMP_INIT,          ///< Initial state
+    PM_SUBTRACTION_STAMP_FOUND,         ///< Found a suitable source for this stamp
+    PM_SUBTRACTION_STAMP_CALCULATE,     ///< Calculate matrix and vector values for this stamp
+    PM_SUBTRACTION_STAMP_USED,          ///< Use this stamp
+    PM_SUBTRACTION_STAMP_REJECTED,      ///< This stamp has been rejected
+    PM_SUBTRACTION_STAMP_NONE           ///< No stamp in this region
+} pmSubtractionStampStatus;
+
+/// A list of stamps
+typedef struct {
+    long num;                           ///< Number of stamps
+    psArray *stamps;                    ///< The stamps
+    psArray *regions;                   ///< Regions for each stamp
+    psArray *x, *y;                     ///< Coordinates for possible stamps (or NULL)
+    psArray *flux;                      ///< Fluxes for possible stamps (or NULL)
+    int footprint;                      ///< Half-size of stamps
+} pmSubtractionStampList;
+
+/// Allocate a list of stamps
+pmSubtractionStampList *pmSubtractionStampListAlloc(int numCols, // Number of columns in image
+                                                    int numRows, // Number of rows in image
+                                                    const psRegion *region, // Region for stamps, or NULL
+                                                    int footprint, // Half-size of stamps
+                                                    float spacing // Rough average spacing between stamps
+    );
+
+/// Assertion for stamp list to be valid
+#define PM_ASSERT_SUBTRACTION_STAMP_LIST_NON_NULL(LIST, RETURNVALUE) { \
+    PS_ASSERT_PTR_NON_NULL(LIST, RETURNVALUE); \
+    PS_ASSERT_ARRAY_NON_NULL((LIST)->stamps, RETURNVALUE); \
+    PS_ASSERT_ARRAY_NON_NULL((LIST)->regions, RETURNVALUE); \
+    PS_ASSERT_INT_POSITIVE((LIST)->num, RETURNVALUE); \
+    PS_ASSERT_ARRAY_SIZE((LIST)->stamps, (LIST)->num, RETURNVALUE); \
+    PS_ASSERT_ARRAY_SIZE((LIST)->regions, (LIST)->num, RETURNVALUE); \
+    PS_ASSERT_INT_NONNEGATIVE((LIST)->footprint, RETURNVALUE); \
+    if ((LIST)->x || (LIST)->y || (LIST)->flux) { \
+        PS_ASSERT_ARRAY_NON_NULL((LIST)->x, RETURNVALUE); \
+        PS_ASSERT_ARRAY_NON_NULL((LIST)->y, RETURNVALUE); \
+        PS_ASSERT_ARRAY_NON_NULL((LIST)->flux, RETURNVALUE); \
+        PS_ASSERT_ARRAY_SIZE((LIST)->x, (LIST)->num, RETURNVALUE); \
+        PS_ASSERT_ARRAY_SIZE((LIST)->y, (LIST)->num, RETURNVALUE); \
+        PS_ASSERT_ARRAY_SIZE((LIST)->flux, (LIST)->num, RETURNVALUE); \
+    } \
+}
+
+/// A stamp for image subtraction
+typedef struct {
+    float x, y;                         ///< Position
+    float flux;                         ///< Flux
+    float xNorm, yNorm;                 ///< Normalised position
+    psKernel *image1;                   ///< Reference image postage stamp
+    psKernel *image2;                   ///< Input image postage stamp
+    psKernel *weight;                   ///< Weight image postage stamp, or NULL
+    psArray *convolutions1;             ///< Convolutions of image 1 for each kernel component, or NULL
+    psArray *convolutions2;             ///< Convolutions of image 2 for each kernel component, or NULL
+    psImage *matrix1, *matrix2;         ///< Least-squares matrices for each image, or NULL
+    psImage *matrixX;                   ///< Cross-matrix (for mode DUAL), or NULL
+    psVector *vector1, *vector2;        ///< Least-squares vectors for each image, or NULL
+    pmSubtractionStampStatus status;    ///< Status of stamp
+} pmSubtractionStamp;
+
+/// Allocate a stamp
+pmSubtractionStamp *pmSubtractionStampAlloc(void);
+
+/// Find stamps on an image
+pmSubtractionStampList *pmSubtractionStampsFind(pmSubtractionStampList *stamps, ///< Output stamps, or NULL
+                                                const psImage *image, ///< Image for which to find stamps
+                                                const psImage *mask, ///< Mask, or NULL
+                                                const psRegion *region, ///< Region to search, or NULL
+                                                float threshold, ///< Threshold for stamps in the image
+                                                int footprint, ///< Half-size for stamps
+                                                float spacing, ///< Rough spacing for stamps
+                                                pmSubtractionMode mode ///< Mode for subtraction
+    );
+
+/// Set stamps based on a list of x,y
+pmSubtractionStampList *pmSubtractionStampsSet(const psVector *x, ///< x coordinates for each stamp
+                                               const psVector *y, ///< y coordinates for each stamp
+                                               const psVector *flux, ///< Flux for each stamp, or NULL
+                                               const psImage *image, ///< Image for flux of stamp
+                                               const psImage *mask, ///< Mask, or NULL
+                                               const psRegion *region, ///< Region to search, or NULL
+                                               int footprint, ///< Half-size for stamps
+                                               float spacing, ///< Rough spacing for stamps
+                                               pmSubtractionMode mode ///< Mode for subtraction
+    );
+
+/// Set stamps based on a list of sources
+pmSubtractionStampList *pmSubtractionStampsSetFromSources(
+    const psArray *sources,             ///< Sources for each stamp
+    const psImage *subMask,             ///< Mask, or NULL
+    const psRegion *region,             ///< Region to search, or NULL
+    int footprint,                      ///< Half-size for stamps
+    float spacing,                      ///< Rough spacing for stamps
+    pmSubtractionMode mode              ///< Mode for subtraction
+    );
+
+/// Set stamps based on values in a file
+pmSubtractionStampList *pmSubtractionStampsSetFromFile(
+    const char *filename,               ///< Filename of file containing x,y (or x,y,flux) on each line
+    const psImage *image,               ///< Image for flux of stamp
+    const psImage *subMask,             ///< Mask, or NULL
+    const psRegion *region,             ///< Region to search, or NULL
+    int footprint,                      ///< Half-size for stamps
+    float spacing,                      ///< Rough spacing for stamps
+    pmSubtractionMode mode              ///< Mode for subtraction
+    );
+
+/// Extract stamps from the images
+bool pmSubtractionStampsExtract(pmSubtractionStampList *stamps, ///< Stamps
+                                psImage *image1, ///< Reference image
+                                psImage *image2, ///< Input image (or NULL)
+                                psImage *weight, ///< Weight (variance) map
+                                int kernelSize ///< Kernel half-size
+    );
+
+#endif
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionThreads.c
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionThreads.c	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionThreads.c	(revision 19481)
@@ -0,0 +1,102 @@
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <stdio.h>
+#include <pslib.h>
+
+#include "pmSubtractionMatch.h"
+#include "pmSubtractionEquation.h"
+#include "pmSubtraction.h"
+
+bool threaded = false;                  // Run with threads?
+
+bool pmSubtractionThreaded(void)
+{
+    return threaded;
+}
+
+// Initialise a mutex in each of the input
+static void subtractionMutexInit(pmReadout *ro)
+{
+    if (!ro) {
+        return;
+    }
+
+    if (ro->image) {
+        psMutexInit(ro->image);
+    }
+    if (ro->weight) {
+        psMutexInit(ro->weight);
+    }
+
+    return;
+}
+
+static void subtractionMutexDestroy(pmReadout *ro)
+{
+    if (!ro) {
+        return;
+    }
+
+    if (ro->image) {
+        psMutexDestroy(ro->image);
+    }
+    if (ro->weight) {
+        psMutexDestroy(ro->weight);
+    }
+
+    return;
+}
+
+void pmSubtractionThreadsInit(pmReadout *in1, pmReadout *in2)
+{
+    if (threaded) {
+        psAbort("Already running threaded.");
+    }
+
+    threaded = true;
+
+    subtractionMutexInit(in1);
+    subtractionMutexInit(in2);
+
+    {
+        psThreadTask *task = psThreadTaskAlloc("PSMODULES_SUBTRACTION_ORDER", 8);
+        task->function = &pmSubtractionOrderThread;
+        psThreadTaskAdd(task);
+        psFree(task);
+    }
+
+    {
+        psThreadTask *task = psThreadTaskAlloc("PSMODULES_SUBTRACTION_CALCULATE_EQUATION", 3);
+        task->function = &pmSubtractionCalculateEquationThread;
+        psThreadTaskAdd(task);
+        psFree(task);
+    }
+
+    {
+        psThreadTask *task = psThreadTaskAlloc("PSMODULES_SUBTRACTION_CONVOLVE", 17);
+        task->function = &pmSubtractionConvolveThread;
+        psThreadTaskAdd(task);
+        psFree(task);
+    }
+
+    return;
+}
+
+
+void pmSubtractionThreadsFinalize(pmReadout *in1, pmReadout *in2)
+{
+    if (!threaded) {
+        return;
+    }
+
+    threaded = false;
+    psThreadTaskRemove("PSMODULES_SUBTRACTION_ORDER");
+    psThreadTaskRemove("PSMODULES_SUBTRACTION_CALCULATE_EQUATION");
+    psThreadTaskRemove("PSMODULES_SUBTRACTION_CONVOLVE");
+
+    subtractionMutexDestroy(in1);
+    subtractionMutexDestroy(in1);
+    return;
+}
Index: /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionThreads.h
===================================================================
--- /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionThreads.h	(revision 19481)
+++ /tags/pap_branch_080910/psModules/src/imcombine/pmSubtractionThreads.h	(revision 19481)
@@ -0,0 +1,19 @@
+#ifndef PM_SUBTRACTION_THREADS_H
+#define PM_SUBTRACTION_THREADS_H
+
+/// Return whether theads have been activated
+bool pmSubtractionThreaded(void);
+
+/// Set up threading for image matching
+///
+/// Sets up thread tasks, and initialises mutexes in readouts
+void pmSubtractionThreadsInit(pmReadout *in1, pmReadout *in2 // Input readouts
+    );
+
+
+/// Take down threading for image matching
+///
+/// Destroys thread tasks, and initialises mutexes in readouts
+void pmSubtractionThreadsFinalize(pmReadout *in1, pmReadout *in2);
+
+#endif
