Use OpenCL for much faster diffraction calculation

6 files changed, 344 insertions, 6 deletions

diff --git a/src/Makefile.am b/src/Makefile.am
index 6a40c9a0..824211f8 100644
--- a/src/Makefile.am
+++ b/src/Makefile.am
@@ -10,6 +10,10 @@ AM_CPPFLAGS = -DDATADIR=\""$(datadir)"\"
 pattern_sim_SOURCES = pattern_sim.c diffraction.c utils.c image.c cell.c \
                       hdf5-file.c detector.c sfac.c intensities.c \
                       reflections.c
+if HAVE_OPENCL
+pattern_sim_SOURCES += diffraction-gpu.c
+endif
+
 pattern_sim_LDADD = @LIBS@
 
 process_hkl_SOURCES = process_hkl.c sfac.c statistics.c cell.c utils.c \
diff --git a/src/diffraction-gpu.c b/src/diffraction-gpu.c
new file mode 100644
index 00000000..4172da3f
--- /dev/null
+++ b/src/diffraction-gpu.c
@@ -0,0 +1,286 @@
+/*
+ * diffraction-gpu.c
+ *
+ * Calculate diffraction patterns by Fourier methods (GPU version)
+ *
+ * (c) 2006-2010 Thomas White <taw@physics.org>
+ *
+ * Part of CrystFEL - crystallography with a FEL
+ *
+ */
+
+
+#include <stdlib.h>
+#include <math.h>
+#include <stdio.h>
+#include <string.h>
+#include <complex.h>
+#include <CL/cl.h>
+
+#include "image.h"
+#include "utils.h"
+#include "cell.h"
+#include "diffraction.h"
+#include "sfac.h"
+
+
+#define SAMPLING (5)
+#define BWSAMPLING (10)
+#define BANDWIDTH (1.0 / 100.0)
+
+
+static cl_device_id get_first_dev(cl_context ctx)
+{
+	cl_device_id dev;
+	cl_int r;
+
+	r = clGetContextInfo(ctx, CL_CONTEXT_DEVICES, sizeof(dev), &dev, NULL);
+	if ( r != CL_SUCCESS ) {
+		ERROR("Couldn't get device\n");
+		return 0;
+	}
+
+	return dev;
+}
+
+
+static void show_build_log(cl_program prog, cl_device_id dev)
+{
+	cl_int r;
+	char log[4096];
+	size_t s;
+
+	r = clGetProgramBuildInfo(prog, dev, CL_PROGRAM_BUILD_LOG, 4096, log,
+	                          &s);
+
+	STATUS("%s\n", log);
+}
+
+
+static cl_program load_program(const char *filename, cl_context ctx,
+                               cl_device_id dev, cl_int *err)
+{
+	FILE *fh;
+	cl_program prog;
+	char *source;
+	size_t len;
+	cl_int r;
+
+	fh = fopen(filename, "r");
+	if ( fh == NULL ) {
+		ERROR("Couldn't open '%s'\n", filename);
+		*err = CL_INVALID_PROGRAM;
+		return 0;
+	}
+	source = malloc(16384);
+	len = fread(source, 1, 16383, fh);
+	fclose(fh);
+	source[len] = '\0';
+
+	prog = clCreateProgramWithSource(ctx, 1, (const char **)&source,
+	                                 NULL, err);
+	if ( *err != CL_SUCCESS ) {
+		ERROR("Couldn't load source\n");
+		return 0;
+	}
+
+	r = clBuildProgram(prog, 0, NULL, "-Werror", NULL, NULL);
+	if ( r != CL_SUCCESS ) {
+		ERROR("Couldn't build program\n");
+		show_build_log(prog, dev);
+		*err = r;
+		return 0;
+	}
+
+	*err = CL_SUCCESS;
+	return prog;
+}
+
+
+void get_diffraction_gpu(struct image *image, int na, int nb, int nc,
+                         int no_sfac)
+{
+	cl_uint nplat;
+	cl_platform_id platforms[8];
+	cl_context_properties prop[3];
+	cl_context ctx;
+	cl_int err;
+	cl_command_queue cq;
+	cl_program prog;
+	cl_device_id dev;
+	cl_kernel kern;
+	double ax, ay, az;
+	double bx, by, bz;
+	double cx, cy, cz;
+	double a, b, c, d;
+	float kc;
+	const size_t dims[2] = {1024, 1024};
+
+	cl_mem sfacs;
+	size_t sfac_size;
+	float *sfac_ptr;
+	cl_mem tt;
+	size_t tt_size;
+	float *tt_ptr;
+	int x, y;
+	cl_float16 cell;
+	cl_mem diff;
+	size_t diff_size;
+	float *diff_ptr;
+	int i;
+
+	if ( image->molecule == NULL ) return;
+
+	/* Generate structure factors if required */
+	if ( !no_sfac ) {
+		if ( image->molecule->reflections == NULL ) {
+			get_reflections_cached(image->molecule,
+			                       ph_lambda_to_en(image->lambda));
+		}
+	}
+
+	cell_get_cartesian(image->molecule->cell, &ax, &ay, &az,
+		                                  &bx, &by, &bz,
+		                                  &cx, &cy, &cz);
+	cell[0] = ax;  cell[1] = ay;  cell[2] = az;
+	cell[3] = bx;  cell[4] = by;  cell[5] = bz;
+	cell[6] = cx;  cell[7] = cy;  cell[8] = cz;
+
+	cell_get_parameters(image->molecule->cell,
+	                    &a, &b, &c, &d, &d, &d);
+	STATUS("Particle size = %i x %i x %i (=%5.2f x %5.2f x %5.2f nm)\n",
+	       na, nb, nc, na*a/1.0e-9, nb*b/1.0e-9, nc*c/1.0e-9);
+
+	err = clGetPlatformIDs(8, platforms, &nplat);
+	if ( err != CL_SUCCESS ) {
+		ERROR("Couldn't get platform IDs: %i\n", err);
+		return;
+	}
+	STATUS("%i platforms\n", nplat);
+	prop[0] = CL_CONTEXT_PLATFORM;
+	prop[1] = (cl_context_properties)platforms[0];
+	prop[2] = 0;
+
+	ctx = clCreateContextFromType(prop, CL_DEVICE_TYPE_GPU, NULL, NULL, &err);
+	if ( err != CL_SUCCESS ) {
+		ERROR("Couldn't create OpenCL context: %i\n", err);
+		switch ( err ) {
+		case CL_INVALID_PLATFORM :
+			ERROR("Invalid platform\n");
+			break;
+		case CL_OUT_OF_HOST_MEMORY :
+			ERROR("Out of memory\n");
+			break;
+		}
+		return;
+	}
+
+	dev = get_first_dev(ctx);
+
+	cq = clCreateCommandQueue(ctx, dev, 0, &err);
+	if ( err != CL_SUCCESS ) {
+		ERROR("Couldn't create OpenCL command queue\n");
+		return;
+	}
+
+	/* Create buffer for the picture */
+	diff_size = image->width*image->height*sizeof(cl_float)*2; /* complex */
+	diff = clCreateBuffer(ctx, CL_MEM_READ_WRITE, diff_size, NULL, &err);
+	if ( err != CL_SUCCESS ) {
+		ERROR("Couldn't allocate diffraction memory\n");
+		return;
+	}
+
+	/* Create a single-precision version of the scattering factors */
+	sfac_size = IDIM*IDIM*sizeof(cl_float)*2; /* complex */
+	sfac_ptr = malloc(IDIM*IDIM*sizeof(cl_float)*2);
+	for ( i=0; i<IDIM*IDIM; i++ ) {
+		sfac_ptr[2*i+0] = creal(image->molecule->reflections[i]);
+		sfac_ptr[2*i+1] = cimag(image->molecule->reflections[i]);
+	}
+	sfacs = clCreateBuffer(ctx, CL_MEM_READ_WRITE | CL_MEM_USE_HOST_PTR,
+	                       sfac_size, sfac_ptr, &err);
+	if ( err != CL_SUCCESS ) {
+		ERROR("Couldn't allocate sfac memory\n");
+		return;
+	}
+
+	tt_size = image->width*image->height*sizeof(cl_float);
+	tt = clCreateBuffer(ctx, CL_MEM_READ_WRITE, tt_size, NULL, &err);
+	if ( err != CL_SUCCESS ) {
+		ERROR("Couldn't allocate twotheta memory\n");
+		return;
+	}
+
+	prog = load_program(DATADIR"/crystfel/diffraction.cl", ctx, dev, &err);
+	if ( err != CL_SUCCESS ) {
+		return;
+	}
+
+	kern = clCreateKernel(prog, "diffraction", &err);
+	if ( err != CL_SUCCESS ) {
+		ERROR("Couldn't create kernel\n");
+		return;
+	}
+
+	/* Calculate wavelength */
+	kc = 1.0/image->lambda;  /* Centre value */
+
+	clSetKernelArg(kern, 0, sizeof(cl_mem), &diff);
+	clSetKernelArg(kern, 1, sizeof(cl_mem), &tt);
+	clSetKernelArg(kern, 2, sizeof(cl_float), &kc);
+	clSetKernelArg(kern, 3, sizeof(cl_int), &image->width);
+	clSetKernelArg(kern, 4, sizeof(cl_float), &image->det.panels[0].cx);
+	clSetKernelArg(kern, 5, sizeof(cl_float), &image->det.panels[0].cy);
+	clSetKernelArg(kern, 6, sizeof(cl_float), &image->resolution);
+	clSetKernelArg(kern, 7, sizeof(cl_float), &image->camera_len);
+	clSetKernelArg(kern, 8, sizeof(cl_float16), &cell);
+	clSetKernelArg(kern, 9, sizeof(cl_mem), &sfacs);
+
+	STATUS("Running...\n");
+	err = clEnqueueNDRangeKernel(cq, kern, 2, NULL, dims, NULL,
+	                             0, NULL, NULL);
+	if ( err != CL_SUCCESS ) {
+		ERROR("Couldn't enqueue kernel\n");
+		return;
+	}
+
+	diff_ptr = clEnqueueMapBuffer(cq, diff, CL_TRUE, CL_MAP_READ, 0,
+	                              diff_size, 0, NULL, NULL, &err);
+	if ( err != CL_SUCCESS ) {
+		ERROR("Couldn't map sfac buffer\n");
+		return;
+	}
+	tt_ptr = clEnqueueMapBuffer(cq, tt, CL_TRUE, CL_MAP_READ, 0,
+	                            tt_size, 0, NULL, NULL, &err);
+	if ( err != CL_SUCCESS ) {
+		ERROR("Couldn't map tt buffer\n");
+		return;
+	}
+	STATUS("Done!\n");
+
+	image->sfacs = calloc(image->width * image->height,
+	                      sizeof(double complex));
+	image->twotheta = calloc(image->width * image->height, sizeof(double));
+
+	for ( x=0; x<image->width; x++ ) {
+	for ( y=0; y<image->height; y++ ) {
+
+		float re, im, tt;
+
+		re = diff_ptr[2*(x + image->width*y)+0];
+		im = diff_ptr[2*(x + image->width*y)+1];
+		tt = tt_ptr[x + image->width*y];
+
+		image->sfacs[x + image->width*y] = re + I*im;
+		image->twotheta[x + image->width*y] = tt;
+
+	}
+	}
+
+	clReleaseProgram(prog);
+	clReleaseMemObject(sfacs);
+	clReleaseCommandQueue(cq);
+	clReleaseContext(ctx);
+
+}
diff --git a/src/diffraction-gpu.h b/src/diffraction-gpu.h
new file mode 100644
index 00000000..687fecf3
--- /dev/null
+++ b/src/diffraction-gpu.h
@@ -0,0 +1,34 @@
+/*
+ * diffraction-gpu.h
+ *
+ * Calculate diffraction patterns by Fourier methods (GPU version)
+ *
+ * (c) 2006-2010 Thomas White <taw@physics.org>
+ *
+ * Part of CrystFEL - crystallography with a FEL
+ *
+ */
+
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#ifndef DIFFRACTION_GPU_H
+#define DIFFRACTION_GPU_H
+
+#include "image.h"
+#include "cell.h"
+
+#if HAVE_OPENCL
+extern void get_diffraction_gpu(struct image *image, int na, int nb, int nc,
+                                int nosfac);
+#else
+static void get_diffraction_gpu(struct image *image, int na, int nb, int nc,
+                                int nosfac)
+{
+	/* Do nothing */
+	ERROR("This copy of CrystFEL was not compiled with OpenCL support.\n");
+}
+#endif
+
+#endif	/* DIFFRACTION_GPU_H */
diff --git a/src/diffraction.c b/src/diffraction.c
index 88b50d1f..fb993bf6 100644
--- a/src/diffraction.c
+++ b/src/diffraction.c
@@ -23,9 +23,9 @@
 #include "sfac.h"
 
 
-#define SAMPLING (5)
-#define BWSAMPLING (10)
-#define BANDWIDTH (1.0 / 100.0)
+#define SAMPLING (1)
+#define BWSAMPLING (1)
+#define BANDWIDTH (0.0 / 100.0)
 
 
 static double lattice_factor(struct rvec q, double ax, double ay, double az,
diff --git a/src/image.h b/src/image.h
index 144ebe42..78194433 100644
--- a/src/image.h
+++ b/src/image.h
@@ -88,8 +88,8 @@ struct image {
 	 * If FORMULATION_PIXELSIZE, then pixel_size only is needed.*/
 	FormulationMode		fmode;
 	double			pixel_size;
-	double			camera_len;
-	double			resolution;	/* pixels per metre */
+	float			camera_len;
+	float			resolution;	/* pixels per metre */
 
 	/* Wavelength must always be given */
 	double			lambda;		/* Wavelength in m */
diff --git a/src/pattern_sim.c b/src/pattern_sim.c
index d0aad29b..e3d675d2 100644
--- a/src/pattern_sim.c
+++ b/src/pattern_sim.c
@@ -21,6 +21,7 @@
 
 #include "image.h"
 #include "diffraction.h"
+#include "diffraction-gpu.h"
 #include "cell.h"
 #include "utils.h"
 #include "hdf5-file.h"
@@ -38,6 +39,7 @@ static void show_help(const char *s)
 "\n"
 " -h, --help                Display this help message.\n"
 "     --simulation-details  Show technical details of the simulation.\n"
+"     --gpu                 Use the GPU to speed up the calculation.\n"
 "\n"
 "     --near-bragg          Output h,k,l,I near Bragg conditions.\n"
 " -n, --number=<N>          Generate N images.  Default 1.\n"
@@ -157,6 +159,7 @@ int main(int argc, char *argv[])
 	int config_nonoise = 0;
 	int config_nobloom = 0;
 	int config_nosfac = 0;
+	int config_gpu = 0;
 	int ndone = 0;    /* Number of simulations done (images or not) */
 	int number = 1;   /* Number used for filename of image */
 	int n_images = 1; /* Generate one image by default */
@@ -166,6 +169,7 @@ int main(int argc, char *argv[])
 	const struct option longopts[] = {
 		{"help",               0, NULL,               'h'},
 		{"simulation-details", 0, &config_simdetails,  1},
+		{"gpu",                0, &config_gpu,         1},
 		{"near-bragg",         0, &config_nearbragg,   1},
 		{"random-orientation", 0, NULL,               'r'},
 		{"number",             1, NULL,               'n'},
@@ -274,11 +278,20 @@ int main(int argc, char *argv[])
 		image.twotheta = NULL;
 		image.hdr = NULL;
 
-		get_diffraction(&image, na, nb, nc, config_nosfac);
+		if ( config_gpu ) {
+			get_diffraction_gpu(&image, na, nb, nc, config_nosfac);
+		} else {
+			get_diffraction(&image, na, nb, nc, config_nosfac);
+		}
 		if ( image.molecule == NULL ) {
 			ERROR("Couldn't open molecule.pdb\n");
 			return 1;
 		}
+		if ( image.sfacs == NULL ) {
+			ERROR("Diffraction calculation failed.\n");
+			goto skip;
+		}
+
 		record_image(&image, !config_nowater, !config_nonoise,
 		             !config_nobloom);
 
@@ -305,6 +318,7 @@ int main(int argc, char *argv[])
 		free(image.sfacs);
 		free(image.twotheta);
 
+skip:
 		ndone++;
 
 		if ( n_images && (ndone >= n_images) ) done = 1;