2 * Copyright (c) 2018 Danil Iashchenko
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 #include "libavutil/common.h"
22 #include "libavutil/imgutils.h"
23 #include "libavutil/mem.h"
24 #include "libavutil/opt.h"
25 #include "libavutil/pixdesc.h"
26 #include "libavutil/avstring.h"
32 #include "opencl_source.h"
35 typedef struct ConvolutionOpenCLContext {
36 OpenCLFilterContext ocf;
40 cl_command_queue command_queue;
45 cl_int matrix_sizes[4];
54 } ConvolutionOpenCLContext;
56 static int convolution_opencl_init(AVFilterContext *avctx)
58 ConvolutionOpenCLContext *ctx = avctx->priv;
59 const char *kernel_name;
63 err = ff_opencl_filter_load_program(avctx, &ff_opencl_source_convolution, 1);
67 ctx->command_queue = clCreateCommandQueue(ctx->ocf.hwctx->context,
68 ctx->ocf.hwctx->device_id,
70 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to create OpenCL "
71 "command queue %d.\n", cle);
73 if (!strcmp(avctx->filter->name, "convolution_opencl")) {
74 kernel_name = "convolution_global";
75 } else if (!strcmp(avctx->filter->name, "sobel_opencl")) {
76 kernel_name = "sobel_global";
77 } else if (!strcmp(avctx->filter->name, "prewitt_opencl")){
78 kernel_name = "prewitt_global";
79 } else if (!strcmp(avctx->filter->name, "roberts_opencl")){
80 kernel_name = "roberts_global";
82 ctx->kernel = clCreateKernel(ctx->ocf.program, kernel_name, &cle);
83 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to create "
90 if (ctx->command_queue)
91 clReleaseCommandQueue(ctx->command_queue);
93 clReleaseKernel(ctx->kernel);
99 static int convolution_opencl_make_filter_params(AVFilterContext *avctx)
101 ConvolutionOpenCLContext *ctx = avctx->priv;
102 float *matrix = NULL;
108 char *p, *arg, *saveptr = NULL;
109 float input_matrix[4][49];
111 for (i = 0; i < 4; i++) {
112 ctx->biases[i] = ctx->biases[i] / 255.0;
115 for (i = 0; i < 4; i++) {
116 p = ctx->matrix_str[i];
117 while (ctx->matrix_sizes[i] < 49) {
118 arg = av_strtok(p, " ", &saveptr);
123 sscanf_err = sscanf(arg, "%f", &input_matrix[i][ctx->matrix_sizes[i]]);
124 if (sscanf_err != 1) {
125 av_log(ctx, AV_LOG_ERROR, "Matrix is sequence of 9, 25 or 49 signed numbers\n");
126 return AVERROR(EINVAL);
128 ctx->matrix_sizes[i]++;
130 if (ctx->matrix_sizes[i] == 9) {
132 } else if (ctx->matrix_sizes[i] == 25) {
134 } else if (ctx->matrix_sizes[i] == 49) {
137 av_log(ctx, AV_LOG_ERROR, "Invalid matrix size:%d\n", ctx->matrix_sizes[i]);
138 return AVERROR(EINVAL);
143 for (j = 0; j < 4; j++) {
144 matrix_bytes = sizeof(float)*ctx->matrix_sizes[j];
145 matrix = av_malloc(matrix_bytes);
148 return AVERROR(ENOMEM);
151 for (i = 0; i < ctx->matrix_sizes[j]; i++)
152 matrix[i] = input_matrix[j][i];
154 buffer = clCreateBuffer(ctx->ocf.hwctx->context,
156 CL_MEM_COPY_HOST_PTR |
157 CL_MEM_HOST_NO_ACCESS,
158 matrix_bytes, matrix, &cle);
160 av_log(avctx, AV_LOG_ERROR, "Failed to create matrix buffer: "
165 ctx->matrix[j] = buffer;
172 static int convolution_opencl_filter_frame(AVFilterLink *inlink, AVFrame *input)
174 AVFilterContext *avctx = inlink->dst;
175 AVFilterLink *outlink = avctx->outputs[0];
176 ConvolutionOpenCLContext *ctx = avctx->priv;
177 AVFrame *output = NULL;
179 size_t global_work[2];
182 size_t origin[3] = {0, 0, 0};
183 size_t region[3] = {0, 0, 1};
185 av_log(ctx, AV_LOG_DEBUG, "Filter input: %s, %ux%u (%"PRId64").\n",
186 av_get_pix_fmt_name(input->format),
187 input->width, input->height, input->pts);
189 if (!input->hw_frames_ctx)
190 return AVERROR(EINVAL);
192 if (!ctx->initialised) {
193 err = convolution_opencl_init(avctx);
197 if (!strcmp(avctx->filter->name, "convolution_opencl")) {
198 err = convolution_opencl_make_filter_params(avctx);
207 output = ff_get_video_buffer(outlink, outlink->w, outlink->h);
209 err = AVERROR(ENOMEM);
213 for (p = 0; p < FF_ARRAY_ELEMS(output->data); p++) {
214 src = (cl_mem) input->data[p];
215 dst = (cl_mem)output->data[p];
220 if (!strcmp(avctx->filter->name, "convolution_opencl")) {
221 CL_SET_KERNEL_ARG(ctx->kernel, 0, cl_mem, &dst);
222 CL_SET_KERNEL_ARG(ctx->kernel, 1, cl_mem, &src);
223 CL_SET_KERNEL_ARG(ctx->kernel, 2, cl_int, &ctx->dims[p]);
224 CL_SET_KERNEL_ARG(ctx->kernel, 3, cl_mem, &ctx->matrix[p]);
225 CL_SET_KERNEL_ARG(ctx->kernel, 4, cl_float, &ctx->rdivs[p]);
226 CL_SET_KERNEL_ARG(ctx->kernel, 5, cl_float, &ctx->biases[p]);
228 err = ff_opencl_filter_work_size_from_image(avctx, global_work, output, p, 0);
232 av_log(avctx, AV_LOG_DEBUG, "Run kernel on plane %d "
233 "(%"SIZE_SPECIFIER"x%"SIZE_SPECIFIER").\n",
234 p, global_work[0], global_work[1]);
236 cle = clEnqueueNDRangeKernel(ctx->command_queue, ctx->kernel, 2, NULL,
239 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to enqueue "
240 "kernel: %d.\n", cle);
242 if (!(ctx->planes & (1 << p))) {
243 err = ff_opencl_filter_work_size_from_image(avctx, region, output, p, 0);
247 cle = clEnqueueCopyImage(ctx->command_queue, src, dst,
248 origin, origin, region, 0, NULL, NULL);
249 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to copy plane %d: %d.\n",
252 CL_SET_KERNEL_ARG(ctx->kernel, 0, cl_mem, &dst);
253 CL_SET_KERNEL_ARG(ctx->kernel, 1, cl_mem, &src);
254 CL_SET_KERNEL_ARG(ctx->kernel, 2, cl_float, &ctx->scale);
255 CL_SET_KERNEL_ARG(ctx->kernel, 3, cl_float, &ctx->delta);
257 err = ff_opencl_filter_work_size_from_image(avctx, global_work, output, p, 0);
261 av_log(avctx, AV_LOG_DEBUG, "Run kernel on plane %d "
262 "(%"SIZE_SPECIFIER"x%"SIZE_SPECIFIER").\n",
263 p, global_work[0], global_work[1]);
265 cle = clEnqueueNDRangeKernel(ctx->command_queue, ctx->kernel, 2, NULL,
268 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to enqueue "
269 "kernel: %d.\n", cle);
274 cle = clFinish(ctx->command_queue);
275 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to finish command queue: %d.\n", cle);
277 err = av_frame_copy_props(output, input);
281 av_frame_free(&input);
283 av_log(ctx, AV_LOG_DEBUG, "Filter output: %s, %ux%u (%"PRId64").\n",
284 av_get_pix_fmt_name(output->format),
285 output->width, output->height, output->pts);
287 return ff_filter_frame(outlink, output);
290 clFinish(ctx->command_queue);
291 av_frame_free(&input);
292 av_frame_free(&output);
296 static av_cold void convolution_opencl_uninit(AVFilterContext *avctx)
298 ConvolutionOpenCLContext *ctx = avctx->priv;
302 for (i = 0; i < 4; i++) {
303 clReleaseMemObject(ctx->matrix[i]);
307 cle = clReleaseKernel(ctx->kernel);
308 if (cle != CL_SUCCESS)
309 av_log(avctx, AV_LOG_ERROR, "Failed to release "
310 "kernel: %d.\n", cle);
313 if (ctx->command_queue) {
314 cle = clReleaseCommandQueue(ctx->command_queue);
315 if (cle != CL_SUCCESS)
316 av_log(avctx, AV_LOG_ERROR, "Failed to release "
317 "command queue: %d.\n", cle);
320 ff_opencl_filter_uninit(avctx);
323 static const AVFilterPad convolution_opencl_inputs[] = {
326 .type = AVMEDIA_TYPE_VIDEO,
327 .filter_frame = &convolution_opencl_filter_frame,
328 .config_props = &ff_opencl_filter_config_input,
333 static const AVFilterPad convolution_opencl_outputs[] = {
336 .type = AVMEDIA_TYPE_VIDEO,
337 .config_props = &ff_opencl_filter_config_output,
342 #define OFFSET(x) offsetof(ConvolutionOpenCLContext, x)
343 #define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
345 #if CONFIG_CONVOLUTION_OPENCL_FILTER
347 static const AVOption convolution_opencl_options[] = {
348 { "0m", "set matrix for 2nd plane", OFFSET(matrix_str[0]), AV_OPT_TYPE_STRING, {.str="0 0 0 0 1 0 0 0 0"}, 0, 0, FLAGS },
349 { "1m", "set matrix for 2nd plane", OFFSET(matrix_str[1]), AV_OPT_TYPE_STRING, {.str="0 0 0 0 1 0 0 0 0"}, 0, 0, FLAGS },
350 { "2m", "set matrix for 3rd plane", OFFSET(matrix_str[2]), AV_OPT_TYPE_STRING, {.str="0 0 0 0 1 0 0 0 0"}, 0, 0, FLAGS },
351 { "3m", "set matrix for 4th plane", OFFSET(matrix_str[3]), AV_OPT_TYPE_STRING, {.str="0 0 0 0 1 0 0 0 0"}, 0, 0, FLAGS },
352 { "0rdiv", "set rdiv for 1nd plane", OFFSET(rdivs[0]), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, INT_MAX, FLAGS},
353 { "1rdiv", "set rdiv for 2nd plane", OFFSET(rdivs[1]), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, INT_MAX, FLAGS},
354 { "2rdiv", "set rdiv for 3rd plane", OFFSET(rdivs[2]), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, INT_MAX, FLAGS},
355 { "3rdiv", "set rdiv for 4th plane", OFFSET(rdivs[3]), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, INT_MAX, FLAGS},
356 { "0bias", "set bias for 1st plane", OFFSET(biases[0]), AV_OPT_TYPE_FLOAT, {.dbl=0.0}, 0.0, INT_MAX, FLAGS},
357 { "1bias", "set bias for 2nd plane", OFFSET(biases[1]), AV_OPT_TYPE_FLOAT, {.dbl=0.0}, 0.0, INT_MAX, FLAGS},
358 { "2bias", "set bias for 3rd plane", OFFSET(biases[2]), AV_OPT_TYPE_FLOAT, {.dbl=0.0}, 0.0, INT_MAX, FLAGS},
359 { "3bias", "set bias for 4th plane", OFFSET(biases[3]), AV_OPT_TYPE_FLOAT, {.dbl=0.0}, 0.0, INT_MAX, FLAGS},
363 AVFILTER_DEFINE_CLASS(convolution_opencl);
365 AVFilter ff_vf_convolution_opencl = {
366 .name = "convolution_opencl",
367 .description = NULL_IF_CONFIG_SMALL("Apply convolution mask to input video"),
368 .priv_size = sizeof(ConvolutionOpenCLContext),
369 .priv_class = &convolution_opencl_class,
370 .init = &ff_opencl_filter_init,
371 .uninit = &convolution_opencl_uninit,
372 .query_formats = &ff_opencl_filter_query_formats,
373 .inputs = convolution_opencl_inputs,
374 .outputs = convolution_opencl_outputs,
375 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
378 #endif /* CONFIG_CONVOLUTION_OPENCL_FILTER */
380 #if CONFIG_SOBEL_OPENCL_FILTER
382 static const AVOption sobel_opencl_options[] = {
383 { "planes", "set planes to filter", OFFSET(planes), AV_OPT_TYPE_INT, {.i64=15}, 0, 15, FLAGS},
384 { "scale", "set scale", OFFSET(scale), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, 65535, FLAGS},
385 { "delta", "set delta", OFFSET(delta), AV_OPT_TYPE_FLOAT, {.dbl=0}, -65535, 65535, FLAGS},
389 AVFILTER_DEFINE_CLASS(sobel_opencl);
391 AVFilter ff_vf_sobel_opencl = {
392 .name = "sobel_opencl",
393 .description = NULL_IF_CONFIG_SMALL("Apply sobel operator"),
394 .priv_size = sizeof(ConvolutionOpenCLContext),
395 .priv_class = &sobel_opencl_class,
396 .init = &ff_opencl_filter_init,
397 .uninit = &convolution_opencl_uninit,
398 .query_formats = &ff_opencl_filter_query_formats,
399 .inputs = convolution_opencl_inputs,
400 .outputs = convolution_opencl_outputs,
401 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
404 #endif /* CONFIG_SOBEL_OPENCL_FILTER */
406 #if CONFIG_PREWITT_OPENCL_FILTER
408 static const AVOption prewitt_opencl_options[] = {
409 { "planes", "set planes to filter", OFFSET(planes), AV_OPT_TYPE_INT, {.i64=15}, 0, 15, FLAGS},
410 { "scale", "set scale", OFFSET(scale), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, 65535, FLAGS},
411 { "delta", "set delta", OFFSET(delta), AV_OPT_TYPE_FLOAT, {.dbl=0}, -65535, 65535, FLAGS},
415 AVFILTER_DEFINE_CLASS(prewitt_opencl);
417 AVFilter ff_vf_prewitt_opencl = {
418 .name = "prewitt_opencl",
419 .description = NULL_IF_CONFIG_SMALL("Apply prewitt operator"),
420 .priv_size = sizeof(ConvolutionOpenCLContext),
421 .priv_class = &prewitt_opencl_class,
422 .init = &ff_opencl_filter_init,
423 .uninit = &convolution_opencl_uninit,
424 .query_formats = &ff_opencl_filter_query_formats,
425 .inputs = convolution_opencl_inputs,
426 .outputs = convolution_opencl_outputs,
427 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
430 #endif /* CONFIG_PREWITT_OPENCL_FILTER */
432 #if CONFIG_ROBERTS_OPENCL_FILTER
434 static const AVOption roberts_opencl_options[] = {
435 { "planes", "set planes to filter", OFFSET(planes), AV_OPT_TYPE_INT, {.i64=15}, 0, 15, FLAGS},
436 { "scale", "set scale", OFFSET(scale), AV_OPT_TYPE_FLOAT, {.dbl=1.0}, 0.0, 65535, FLAGS},
437 { "delta", "set delta", OFFSET(delta), AV_OPT_TYPE_FLOAT, {.dbl=0}, -65535, 65535, FLAGS},
441 AVFILTER_DEFINE_CLASS(roberts_opencl);
443 AVFilter ff_vf_roberts_opencl = {
444 .name = "roberts_opencl",
445 .description = NULL_IF_CONFIG_SMALL("Apply roberts operator"),
446 .priv_size = sizeof(ConvolutionOpenCLContext),
447 .priv_class = &roberts_opencl_class,
448 .init = &ff_opencl_filter_init,
449 .uninit = &convolution_opencl_uninit,
450 .query_formats = &ff_opencl_filter_query_formats,
451 .inputs = convolution_opencl_inputs,
452 .outputs = convolution_opencl_outputs,
453 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
456 #endif /* CONFIG_ROBERTS_OPENCL_FILTER */