[gegl] Add OpenCL support for gegl:c2g
- From: Victor Matheus de Araujo Oliveira <vmaolive src gnome org>
- To: commits-list gnome org
- Cc:
- Subject: [gegl] Add OpenCL support for gegl:c2g
- Date: Sun, 1 Apr 2012 18:02:50 +0000 (UTC)
commit 969feed68f04e6e1daa2f8ef97f12410f944866f
Author: Zhang Peixuan <zhangpeixuan cn gmail com>
Date: Thu Mar 15 12:18:44 2012 -0300
Add OpenCL support for gegl:c2g
operations/common/c2g.c | 310 ++++++++++++++++++++++++++++++++++++++++++++++-
1 files changed, 305 insertions(+), 5 deletions(-)
---
diff --git a/operations/common/c2g.c b/operations/common/c2g.c
index 0317977..52962ce 100644
--- a/operations/common/c2g.c
+++ b/operations/common/c2g.c
@@ -95,18 +95,18 @@ static void c2g (GeglBuffer *src,
*/
gfloat nominator = 0;
- gfloat denominator = 0;
+ gfloat denominator = 0;
gint c;
- for (c=0; c<3; c++)
- {
+ for (c=0; c<3; c++)
+ {
nominator += (pixel[c] - min[c]) * (pixel[c] - min[c]);
denominator += (pixel[c] - max[c]) * (pixel[c] - max[c]);
- }
+ }
nominator = sqrt (nominator);
denominator = sqrt (denominator);
denominator = nominator + denominator;
-
+
if (denominator>0.000)
{
dst_buf[dst_offset+0] = nominator/denominator;
@@ -133,6 +133,8 @@ static void prepare (GeglOperation *operation)
GeglOperationAreaFilter *area = GEGL_OPERATION_AREA_FILTER (operation);
area->left = area->right = area->top = area->bottom =
ceil (GEGL_CHANT_PROPERTIES (operation)->radius);
+
+ gegl_operation_set_format (operation, "input", babl_format ("RGBA float"));
gegl_operation_set_format (operation, "output", babl_format ("YA float"));
}
@@ -147,6 +149,297 @@ get_bounding_box (GeglOperation *operation)
return *in_rect;
}
+#include "opencl/gegl-cl.h"
+#include "buffer/gegl-buffer-cl-iterator.h"
+
+static const char* kernel_source =
+"#define TRUE true \n"
+" \n"
+"#define FALSE false \n"
+"#define ANGLE_PRIME 95273 \n"
+"#define RADIUS_PRIME 29537 \n"
+" \n"
+"void sample_min_max(const __global float4 *src_buf, \n"
+" int src_width, \n"
+" int src_height, \n"
+" const __global float *radiuses, \n"
+" const __global float *lut_cos, \n"
+" const __global float *lut_sin, \n"
+" int x, \n"
+" int y, \n"
+" int radius, \n"
+" int samples, \n"
+" float4 *min, \n"
+" float4 *max, \n"
+" int j, \n"
+" int iterations) \n"
+"{ \n"
+" float4 best_min; \n"
+" float4 best_max; \n"
+" float4 center_pix = *(src_buf + src_width * y + x); \n"
+" int i; \n"
+" \n"
+" best_min = center_pix; \n"
+" best_max = center_pix; \n"
+" \n"
+" int angle_no = (src_width * y + x) * (iterations) * \n"
+" samples + j * samples; \n"
+" int radius_no = angle_no; \n"
+" angle_no %= ANGLE_PRIME; \n"
+" radius_no %= RADIUS_PRIME; \n"
+" for(i=0; i<samples; i++) \n"
+" { \n"
+" int angle; \n"
+" float rmag; \n"
+" /* if we've sampled outside the valid image \n"
+" area, we grab another sample instead, this \n"
+" should potentially work better than mirroring \n"
+" or extending the image */ \n"
+" \n"
+" angle = angle_no++; \n"
+" rmag = radiuses[radius_no++] * radius; \n"
+" \n"
+" if( angle_no >= ANGLE_PRIME) \n"
+" angle_no = 0; \n"
+" if( radius_no >= RADIUS_PRIME) \n"
+" radius_no = 0; \n"
+" \n"
+" int u = x + rmag * lut_cos[angle]; \n"
+" int v = y + rmag * lut_sin[angle]; \n"
+" \n"
+" if(u>=src_width || u <0 || v>=src_height || v<0) \n"
+" { \n"
+" //--i; \n"
+" continue; \n"
+" } \n"
+" float4 pixel = *(src_buf + (src_width * v + u)); \n"
+" if(pixel.w<=0.0f) \n"
+" { \n"
+" //--i; \n"
+" continue; \n"
+" } \n"
+" \n"
+" best_min = pixel < best_min ? pixel : best_min; \n"
+" best_max = pixel > best_max ? pixel : best_max; \n"
+" } \n"
+" \n"
+" (*min).xyz = best_min.xyz; \n"
+" (*max).xyz = best_max.xyz; \n"
+"} \n"
+" \n"
+"void compute_envelopes_CL(const __global float4 *src_buf, \n"
+" int src_width, \n"
+" int src_height, \n"
+" const __global float *radiuses, \n"
+" const __global float *lut_cos, \n"
+" const __global float *lut_sin, \n"
+" int x, \n"
+" int y, \n"
+" int radius, \n"
+" int samples, \n"
+" int iterations, \n"
+" float4 *min_envelope, \n"
+" float4 *max_envelope) \n"
+"{ \n"
+" float4 range_sum = 0; \n"
+" float4 relative_brightness_sum = 0; \n"
+" float4 pixel = *(src_buf + src_width * y + x); \n"
+" \n"
+" int i; \n"
+" for(i =0; i<iterations; i++) \n"
+" { \n"
+" float4 min,max; \n"
+" float4 range, relative_brightness; \n"
+" \n"
+" sample_min_max(src_buf, src_width, src_height, \n"
+" radiuses, lut_cos, lut_sin, x, y, \n"
+" radius,samples,&min,&max,i,iterations); \n"
+" range = max - min; \n"
+" relative_brightness = range <= 0.0f ? \n"
+" 0.5f : (pixel - min) / range; \n"
+" relative_brightness_sum += relative_brightness; \n"
+" range_sum += range; \n"
+" } \n"
+" \n"
+" float4 relative_brightness = relative_brightness_sum / iterations;\n"
+" float4 range = range_sum / iterations; \n"
+" \n"
+" if(max_envelope) \n"
+" *max_envelope = pixel + (1.0f - relative_brightness) * range; \n"
+" \n"
+" if(min_envelope) \n"
+" *min_envelope = pixel - relative_brightness * range; \n"
+"} \n"
+" \n"
+"__kernel void C2g_CL(const __global float4 *src_buf, \n"
+" int src_width, \n"
+" int src_height, \n"
+" const __global float *radiuses, \n"
+" const __global float *lut_cos, \n"
+" const __global float *lut_sin, \n"
+" __global float2 *dst_buf, \n"
+" int radius, \n"
+" int samples, \n"
+" int iterations) \n"
+"{ \n"
+" int gidx = get_global_id(0); \n"
+" int gidy = get_global_id(1); \n"
+" \n"
+" int x = gidx + radius; \n"
+" int y = gidy + radius; \n"
+" \n"
+" int src_offset = (src_width * y + x); \n"
+" int dst_offset = gidx + get_global_size(0) * gidy; \n"
+" float4 min,max; \n"
+" \n"
+" compute_envelopes_CL(src_buf, src_width, src_height, \n"
+" radiuses, lut_cos, lut_sin, x, y, \n"
+" radius, samples, iterations, &min, &max); \n"
+" \n"
+" float4 pixel = *(src_buf + src_offset); \n"
+" \n"
+" float nominator=0, denominator=0; \n"
+" float4 t1 = (pixel - min) * (pixel - min); \n"
+" float4 t2 = (pixel - max) * (pixel - max); \n"
+" \n"
+" nominator = t1.x + t1.y + t1.z; \n"
+" denominator = t2.x + t2.y + t2.z; \n"
+" \n"
+" nominator = sqrt(nominator); \n"
+" denominator = sqrt(denominator); \n"
+" denominator+= nominator + denominator; \n"
+" \n"
+" dst_buf[dst_offset].x = (denominator > 0.000f) \n"
+" ? (nominator / denominator) : 0.5f; \n"
+" dst_buf[dst_offset].y = src_buf[src_offset].w; \n"
+"} \n"
+" \n";
+
+static gegl_cl_run_data *cl_data = NULL;
+
+static cl_int
+cl_c2g (cl_mem in_tex,
+ cl_mem out_tex,
+ size_t global_worksize,
+ const GeglRectangle *src_roi,
+ const GeglRectangle *roi,
+ gint radius,
+ gint samples,
+ gint iterations,
+ gdouble rgamma)
+{
+ cl_int cl_err = 0;
+ cl_mem cl_lut_cos, cl_lut_sin, cl_radiuses;
+ const size_t gbl_size[2] = {roi->width, roi->height};
+
+ if (!cl_data)
+ {
+ const char *kernel_name[] ={"C2g_CL", NULL};
+ cl_data = gegl_cl_compile_and_build(kernel_source, kernel_name);
+ }
+ if (!cl_data) return 0;
+
+ compute_luts(rgamma);
+
+ cl_lut_cos = gegl_clCreateBuffer(gegl_cl_get_context(),
+ CL_MEM_READ_ONLY,
+ ANGLE_PRIME * sizeof(cl_float), NULL, &cl_err);
+
+ cl_err |= gegl_clEnqueueWriteBuffer(gegl_cl_get_command_queue(), cl_lut_cos,
+ CL_TRUE, 0, ANGLE_PRIME * sizeof(cl_float), lut_cos,
+ 0, NULL, NULL);
+ if (CL_SUCCESS != cl_err) return cl_err;
+
+ cl_lut_sin = gegl_clCreateBuffer(gegl_cl_get_context(),
+ CL_MEM_READ_ONLY,
+ ANGLE_PRIME * sizeof(cl_float), NULL, &cl_err);
+
+ cl_err |= gegl_clEnqueueWriteBuffer(gegl_cl_get_command_queue(), cl_lut_sin,
+ CL_TRUE, 0, ANGLE_PRIME * sizeof(cl_float), lut_sin,
+ 0, NULL, NULL);
+ if (CL_SUCCESS != cl_err) return cl_err;
+
+ cl_radiuses = gegl_clCreateBuffer(gegl_cl_get_context(),
+ CL_MEM_READ_ONLY,
+ RADIUS_PRIME * sizeof(cl_float), NULL, &cl_err);
+
+ cl_err |= gegl_clEnqueueWriteBuffer(gegl_cl_get_command_queue(), cl_radiuses,
+ CL_TRUE, 0, RADIUS_PRIME * sizeof(cl_float), radiuses,
+ 0, NULL, NULL);
+ if (CL_SUCCESS != cl_err) return cl_err;
+
+ {
+ cl_int cl_src_width = src_roi->width;
+ cl_int cl_src_height = src_roi->height;
+ cl_int cl_radius = radius;
+ cl_int cl_samples = samples;
+ cl_int cl_iterations = iterations;
+
+ cl_err |= gegl_clSetKernelArg(cl_data->kernel[0], 0, sizeof(cl_mem), (void*)&in_tex);
+ cl_err |= gegl_clSetKernelArg(cl_data->kernel[0], 1, sizeof(cl_int), (void*)&cl_src_width);
+ cl_err |= gegl_clSetKernelArg(cl_data->kernel[0], 2, sizeof(cl_int), (void*)&cl_src_height);
+ cl_err |= gegl_clSetKernelArg(cl_data->kernel[0], 3, sizeof(cl_mem), (void*)&cl_radiuses);
+ cl_err |= gegl_clSetKernelArg(cl_data->kernel[0], 4, sizeof(cl_mem), (void*)&cl_lut_cos);
+ cl_err |= gegl_clSetKernelArg(cl_data->kernel[0], 5, sizeof(cl_mem), (void*)&cl_lut_sin);
+ cl_err |= gegl_clSetKernelArg(cl_data->kernel[0], 6, sizeof(cl_mem), (void*)&out_tex);
+ cl_err |= gegl_clSetKernelArg(cl_data->kernel[0], 7, sizeof(cl_int), (void*)&cl_radius);
+ cl_err |= gegl_clSetKernelArg(cl_data->kernel[0], 8, sizeof(cl_int), (void*)&cl_samples);
+ cl_err |= gegl_clSetKernelArg(cl_data->kernel[0], 9, sizeof(cl_int), (void*)&cl_iterations);
+ if (cl_err != CL_SUCCESS) return cl_err;
+ }
+
+ cl_err = gegl_clEnqueueNDRangeKernel(gegl_cl_get_command_queue(), cl_data->kernel[0],
+ 2, NULL, gbl_size, NULL,
+ 0, NULL, NULL);
+ if (cl_err != CL_SUCCESS) return cl_err;
+
+ cl_err = gegl_clEnqueueBarrier(gegl_cl_get_command_queue());
+ if (CL_SUCCESS != cl_err) return cl_err;
+
+ gegl_clFinish(gegl_cl_get_command_queue ());
+
+ gegl_clReleaseMemObject(cl_radiuses);
+ gegl_clReleaseMemObject(cl_lut_cos);
+ gegl_clReleaseMemObject(cl_lut_sin);
+
+ return cl_err;
+}
+
+static gboolean
+cl_process (GeglOperation *operation,
+ GeglBuffer *input,
+ GeglBuffer *output,
+ const GeglRectangle *result)
+{
+ const Babl *in_format = babl_format("RGBA float");
+ const Babl *out_format = gegl_operation_get_format (operation, "output");
+ gint err;
+ gint j;
+ cl_int cl_err;
+
+ GeglOperationAreaFilter *op_area = GEGL_OPERATION_AREA_FILTER (operation);
+ GeglChantO *o = GEGL_CHANT_PROPERTIES (operation);
+
+ GeglBufferClIterator *i = gegl_buffer_cl_iterator_new (output,result, out_format, GEGL_CL_BUFFER_WRITE, GEGL_ABYSS_NONE);
+ gint read = gegl_buffer_cl_iterator_add_2 (i, input, result, in_format, GEGL_CL_BUFFER_READ,
+ op_area->left, op_area->right, op_area->top, op_area->bottom, GEGL_ABYSS_NONE);
+ while (gegl_buffer_cl_iterator_next (i, &err))
+ {
+ if (err) return FALSE;
+ for (j=0; j < i->n; j++)
+ {
+ cl_err = cl_c2g(i->tex[read][j], i->tex[0][j],i->size[0][j], &i->roi[read][j],&i->roi[0][j],
+ o->radius,o->samples,o->iterations,RGAMMA);
+ if (cl_err != CL_SUCCESS)
+ {
+ g_warning("[OpenCL] Error in gegl:c2g: %s", gegl_cl_errstring(cl_err));
+ return FALSE;
+ }
+ }
+ }
+ return TRUE;
+}
+
static gboolean
process (GeglOperation *operation,
GeglBuffer *input,
@@ -158,6 +451,10 @@ process (GeglOperation *operation,
GeglRectangle compute;
compute = gegl_operation_get_required_for_output (operation, "input",result);
+ if (o->radius < 500 && gegl_cl_is_accelerated ())
+ if(cl_process(operation, input, output, result))
+ return TRUE;
+
c2g (input, &compute, output, result,
o->radius,
o->samples,
@@ -179,6 +476,7 @@ gegl_chant_class_init (GeglChantClass *klass)
filter_class->process = process;
operation_class->prepare = prepare;
+
/* we override defined region to avoid growing the size of what is defined
* by the filter. This also allows the tricks used to treat alpha==0 pixels
* in the image as source data not to be skipped by the stochastic sampling
@@ -186,6 +484,8 @@ gegl_chant_class_init (GeglChantClass *klass)
*/
operation_class->get_bounding_box = get_bounding_box;
+ operation_class->opencl_support = TRUE;
+
gegl_operation_class_set_keys (operation_class,
"name", "gegl:c2g",
"categories", "enhance",
[
Date Prev][
Date Next] [
Thread Prev][
Thread Next]
[
Thread Index]
[
Date Index]
[
Author Index]
