1 /*
2 * Copyright (c) 2018 Danil Iashchenko
3 *
4 * This file is part of FFmpeg.
5 *
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.
10 *
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.
15 *
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
19 */
20
21 #include "config_components.h"
22
23 #include "libavutil/common.h"
24 #include "libavutil/imgutils.h"
25 #include "libavutil/mem.h"
26 #include "libavutil/opt.h"
27 #include "libavutil/pixdesc.h"
28 #include "libavutil/avstring.h"
29
30
31 #include "avfilter.h"
32 #include "internal.h"
33 #include "opencl.h"
34 #include "opencl_source.h"
35 #include "video.h"
36
37 typedef struct ConvolutionOpenCLContext {
38 OpenCLFilterContext ocf;
39
40 int initialised;
41 cl_kernel kernel;
42 cl_command_queue command_queue;
43
44 char *matrix_str[4];
45
46 cl_mem matrix[4];
47 cl_int matrix_sizes[4];
48 cl_int dims[4];
49 cl_float rdivs[4];
50 cl_float biases[4];
51
52 cl_int planes;
53 cl_float scale;
54 cl_float delta;
55
56 } ConvolutionOpenCLContext;
57
convolution_opencl_init(AVFilterContext * avctx)58 static int convolution_opencl_init(AVFilterContext *avctx)
59 {
60 ConvolutionOpenCLContext *ctx = avctx->priv;
61 const char *kernel_name;
62 cl_int cle;
63 int err;
64
65 err = ff_opencl_filter_load_program(avctx, &ff_opencl_source_convolution, 1);
66 if (err < 0)
67 goto fail;
68
69 ctx->command_queue = clCreateCommandQueue(ctx->ocf.hwctx->context,
70 ctx->ocf.hwctx->device_id,
71 0, &cle);
72 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to create OpenCL "
73 "command queue %d.\n", cle);
74
75 if (!strcmp(avctx->filter->name, "convolution_opencl")) {
76 kernel_name = "convolution_global";
77 } else if (!strcmp(avctx->filter->name, "sobel_opencl")) {
78 kernel_name = "sobel_global";
79 } else if (!strcmp(avctx->filter->name, "prewitt_opencl")){
80 kernel_name = "prewitt_global";
81 } else if (!strcmp(avctx->filter->name, "roberts_opencl")){
82 kernel_name = "roberts_global";
83 }
84 ctx->kernel = clCreateKernel(ctx->ocf.program, kernel_name, &cle);
85 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to create "
86 "kernel %d.\n", cle);
87
88 ctx->initialised = 1;
89 return 0;
90
91 fail:
92 if (ctx->command_queue)
93 clReleaseCommandQueue(ctx->command_queue);
94 if (ctx->kernel)
95 clReleaseKernel(ctx->kernel);
96 return err;
97 }
98
99
100
convolution_opencl_make_filter_params(AVFilterContext * avctx)101 static int convolution_opencl_make_filter_params(AVFilterContext *avctx)
102 {
103 ConvolutionOpenCLContext *ctx = avctx->priv;
104 float *matrix = NULL;
105 size_t matrix_bytes;
106 cl_mem buffer;
107 cl_int cle;
108 int i, j;
109 int sscanf_err;
110 char *p, *arg, *saveptr = NULL;
111 float input_matrix[4][49];
112
113 for (i = 0; i < 4; i++) {
114 ctx->biases[i] = ctx->biases[i] / 255.0;
115 }
116
117 for (i = 0; i < 4; i++) {
118 p = ctx->matrix_str[i];
119 while (ctx->matrix_sizes[i] < 49) {
120 arg = av_strtok(p, " ", &saveptr);
121 if (!arg) {
122 break;
123 }
124 p = NULL;
125 sscanf_err = sscanf(arg, "%f", &input_matrix[i][ctx->matrix_sizes[i]]);
126 if (sscanf_err != 1) {
127 av_log(ctx, AV_LOG_ERROR, "Matrix is sequence of 9, 25 or 49 signed numbers\n");
128 return AVERROR(EINVAL);
129 }
130 ctx->matrix_sizes[i]++;
131 }
132 if (ctx->matrix_sizes[i] == 9) {
133 ctx->dims[i] = 3;
134 } else if (ctx->matrix_sizes[i] == 25) {
135 ctx->dims[i] = 5;
136 } else if (ctx->matrix_sizes[i] == 49) {
137 ctx->dims[i] = 7;
138 } else {
139 av_log(ctx, AV_LOG_ERROR, "Invalid matrix size:%d\n", ctx->matrix_sizes[i]);
140 return AVERROR(EINVAL);
141 }
142
143 }
144
145 for (j = 0; j < 4; j++) {
146 matrix_bytes = sizeof(float)*ctx->matrix_sizes[j];
147 matrix = av_malloc(matrix_bytes);
148 if (!matrix) {
149 av_freep(&matrix);
150 return AVERROR(ENOMEM);
151 }
152
153 for (i = 0; i < ctx->matrix_sizes[j]; i++)
154 matrix[i] = input_matrix[j][i];
155
156 buffer = clCreateBuffer(ctx->ocf.hwctx->context,
157 CL_MEM_READ_ONLY |
158 CL_MEM_COPY_HOST_PTR |
159 CL_MEM_HOST_NO_ACCESS,
160 matrix_bytes, matrix, &cle);
161 if (!buffer) {
162 av_log(avctx, AV_LOG_ERROR, "Failed to create matrix buffer: "
163 "%d.\n", cle);
164 av_freep(&matrix);
165 return AVERROR(EIO);
166 }
167 ctx->matrix[j] = buffer;
168 av_freep(&matrix);
169 }
170
171 return 0;
172 }
173
convolution_opencl_filter_frame(AVFilterLink * inlink,AVFrame * input)174 static int convolution_opencl_filter_frame(AVFilterLink *inlink, AVFrame *input)
175 {
176 AVFilterContext *avctx = inlink->dst;
177 AVFilterLink *outlink = avctx->outputs[0];
178 ConvolutionOpenCLContext *ctx = avctx->priv;
179 AVFrame *output = NULL;
180 cl_int cle;
181 size_t global_work[2];
182 cl_mem src, dst;
183 int err, p;
184 size_t origin[3] = {0, 0, 0};
185 size_t region[3] = {0, 0, 1};
186
187 av_log(ctx, AV_LOG_DEBUG, "Filter input: %s, %ux%u (%"PRId64").\n",
188 av_get_pix_fmt_name(input->format),
189 input->width, input->height, input->pts);
190
191 if (!input->hw_frames_ctx)
192 return AVERROR(EINVAL);
193
194 if (!ctx->initialised) {
195 err = convolution_opencl_init(avctx);
196 if (err < 0)
197 goto fail;
198
199 if (!strcmp(avctx->filter->name, "convolution_opencl")) {
200 err = convolution_opencl_make_filter_params(avctx);
201 if (err < 0)
202 goto fail;
203 } else {
204 ctx->delta /= 255.0;
205 }
206
207 }
208
209 output = ff_get_video_buffer(outlink, outlink->w, outlink->h);
210 if (!output) {
211 err = AVERROR(ENOMEM);
212 goto fail;
213 }
214
215 for (p = 0; p < FF_ARRAY_ELEMS(output->data); p++) {
216 src = (cl_mem) input->data[p];
217 dst = (cl_mem)output->data[p];
218
219 if (!dst)
220 break;
221
222 if (!strcmp(avctx->filter->name, "convolution_opencl")) {
223 CL_SET_KERNEL_ARG(ctx->kernel, 0, cl_mem, &dst);
224 CL_SET_KERNEL_ARG(ctx->kernel, 1, cl_mem, &src);
225 CL_SET_KERNEL_ARG(ctx->kernel, 2, cl_int, &ctx->dims[p]);
226 CL_SET_KERNEL_ARG(ctx->kernel, 3, cl_mem, &ctx->matrix[p]);
227 CL_SET_KERNEL_ARG(ctx->kernel, 4, cl_float, &ctx->rdivs[p]);
228 CL_SET_KERNEL_ARG(ctx->kernel, 5, cl_float, &ctx->biases[p]);
229
230 err = ff_opencl_filter_work_size_from_image(avctx, global_work, output, p, 0);
231 if (err < 0)
232 goto fail;
233
234 av_log(avctx, AV_LOG_DEBUG, "Run kernel on plane %d "
235 "(%"SIZE_SPECIFIER"x%"SIZE_SPECIFIER").\n",
236 p, global_work[0], global_work[1]);
237
238 cle = clEnqueueNDRangeKernel(ctx->command_queue, ctx->kernel, 2, NULL,
239 global_work, NULL,
240 0, NULL, NULL);
241 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to enqueue "
242 "kernel: %d.\n", cle);
243 } else {
244 if (!(ctx->planes & (1 << p))) {
245 err = ff_opencl_filter_work_size_from_image(avctx, region, output, p, 0);
246 if (err < 0)
247 goto fail;
248
249 cle = clEnqueueCopyImage(ctx->command_queue, src, dst,
250 origin, origin, region, 0, NULL, NULL);
251 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to copy plane %d: %d.\n",
252 p, cle);
253 } else {
254 CL_SET_KERNEL_ARG(ctx->kernel, 0, cl_mem, &dst);
255 CL_SET_KERNEL_ARG(ctx->kernel, 1, cl_mem, &src);
256 CL_SET_KERNEL_ARG(ctx->kernel, 2, cl_float, &ctx->scale);
257 CL_SET_KERNEL_ARG(ctx->kernel, 3, cl_float, &ctx->delta);
258
259 err = ff_opencl_filter_work_size_from_image(avctx, global_work, output, p, 0);
260 if (err < 0)
261 goto fail;
262
263 av_log(avctx, AV_LOG_DEBUG, "Run kernel on plane %d "
264 "(%"SIZE_SPECIFIER"x%"SIZE_SPECIFIER").\n",
265 p, global_work[0], global_work[1]);
266
267 cle = clEnqueueNDRangeKernel(ctx->command_queue, ctx->kernel, 2, NULL,
268 global_work, NULL,
269 0, NULL, NULL);
270 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to enqueue "
271 "kernel: %d.\n", cle);
272 }
273 }
274 }
275
276 cle = clFinish(ctx->command_queue);
277 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to finish command queue: %d.\n", cle);
278
279 err = av_frame_copy_props(output, input);
280 if (err < 0)
281 goto fail;
282
283 av_frame_free(&input);
284
285 av_log(ctx, AV_LOG_DEBUG, "Filter output: %s, %ux%u (%"PRId64").\n",
286 av_get_pix_fmt_name(output->format),
287 output->width, output->height, output->pts);
288
289 return ff_filter_frame(outlink, output);
290
291 fail:
292 clFinish(ctx->command_queue);
293 av_frame_free(&input);
294 av_frame_free(&output);
295 return err;
296 }
297
convolution_opencl_uninit(AVFilterContext * avctx)298 static av_cold void convolution_opencl_uninit(AVFilterContext *avctx)
299 {
300 ConvolutionOpenCLContext *ctx = avctx->priv;
301 cl_int cle;
302 int i;
303
304 for (i = 0; i < 4; i++) {
305 clReleaseMemObject(ctx->matrix[i]);
306 }
307
308 if (ctx->kernel) {
309 cle = clReleaseKernel(ctx->kernel);
310 if (cle != CL_SUCCESS)
311 av_log(avctx, AV_LOG_ERROR, "Failed to release "
312 "kernel: %d.\n", cle);
313 }
314
315 if (ctx->command_queue) {
316 cle = clReleaseCommandQueue(ctx->command_queue);
317 if (cle != CL_SUCCESS)
318 av_log(avctx, AV_LOG_ERROR, "Failed to release "
319 "command queue: %d.\n", cle);
320 }
321
322 ff_opencl_filter_uninit(avctx);
323 }
324
325 static const AVFilterPad convolution_opencl_inputs[] = {
326 {
327 .name = "default",
328 .type = AVMEDIA_TYPE_VIDEO,
329 .filter_frame = &convolution_opencl_filter_frame,
330 .config_props = &ff_opencl_filter_config_input,
331 },
332 };
333
334 static const AVFilterPad convolution_opencl_outputs[] = {
335 {
336 .name = "default",
337 .type = AVMEDIA_TYPE_VIDEO,
338 .config_props = &ff_opencl_filter_config_output,
339 },
340 };
341
342 #define OFFSET(x) offsetof(ConvolutionOpenCLContext, x)
343 #define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
344
345 #if CONFIG_CONVOLUTION_OPENCL_FILTER
346
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},
360 { NULL }
361 };
362
363 AVFILTER_DEFINE_CLASS(convolution_opencl);
364
365 const 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 FILTER_INPUTS(convolution_opencl_inputs),
373 FILTER_OUTPUTS(convolution_opencl_outputs),
374 FILTER_SINGLE_PIXFMT(AV_PIX_FMT_OPENCL),
375 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
376 };
377
378 #endif /* CONFIG_CONVOLUTION_OPENCL_FILTER */
379
380 #if CONFIG_SOBEL_OPENCL_FILTER
381
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},
386 { NULL }
387 };
388
389 AVFILTER_DEFINE_CLASS(sobel_opencl);
390
391 const 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 FILTER_INPUTS(convolution_opencl_inputs),
399 FILTER_OUTPUTS(convolution_opencl_outputs),
400 FILTER_SINGLE_PIXFMT(AV_PIX_FMT_OPENCL),
401 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
402 };
403
404 #endif /* CONFIG_SOBEL_OPENCL_FILTER */
405
406 #if CONFIG_PREWITT_OPENCL_FILTER
407
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},
412 { NULL }
413 };
414
415 AVFILTER_DEFINE_CLASS(prewitt_opencl);
416
417 const 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 FILTER_INPUTS(convolution_opencl_inputs),
425 FILTER_OUTPUTS(convolution_opencl_outputs),
426 FILTER_SINGLE_PIXFMT(AV_PIX_FMT_OPENCL),
427 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
428 };
429
430 #endif /* CONFIG_PREWITT_OPENCL_FILTER */
431
432 #if CONFIG_ROBERTS_OPENCL_FILTER
433
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},
438 { NULL }
439 };
440
441 AVFILTER_DEFINE_CLASS(roberts_opencl);
442
443 const 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 FILTER_INPUTS(convolution_opencl_inputs),
451 FILTER_OUTPUTS(convolution_opencl_outputs),
452 FILTER_SINGLE_PIXFMT(AV_PIX_FMT_OPENCL),
453 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
454 };
455
456 #endif /* CONFIG_ROBERTS_OPENCL_FILTER */
457