History

avutil/vf_scdet_vulkan: add new filter

Carbon copy of vf_scdet.

Niklas Haas committed 1 year ago

Commit 9568e401 parent 0894cfb4 files · +415 −0Browse files

Showing the changes to libavfilter (3 of 4 files). Show all

Unified · Split

3 files
  1. libavfilter/Makefile +1 −0
  2. libavfilter/allfilters.c +1 −0
  3. libavfilter/vf_scdet_vulkan.c +412 −0
libavfilter/Makefile+1−0
@@ -472,6 +472,7 @@ OBJS-$(CONFIG_SCALE_VULKAN_FILTER) += vf_scale_vulkan.o vulkan.o vulka · 471 unmodified lines
472OBJS-$(CONFIG_SCALE2REF_FILTER) += vf_scale.o scale_eval.o framesync.o472OBJS-$(CONFIG_SCALE2REF_FILTER) += vf_scale.o scale_eval.o framesync.o
473OBJS-$(CONFIG_SCALE2REF_NPP_FILTER) += vf_scale_npp.o scale_eval.o473OBJS-$(CONFIG_SCALE2REF_NPP_FILTER) += vf_scale_npp.o scale_eval.o
474OBJS-$(CONFIG_SCDET_FILTER) += vf_scdet.o474OBJS-$(CONFIG_SCDET_FILTER) += vf_scdet.o
475OBJS-$(CONFIG_SCDET_VULKAN_FILTER) += vf_scdet_vulkan.o
475OBJS-$(CONFIG_SCHARR_FILTER) += vf_convolution.o476OBJS-$(CONFIG_SCHARR_FILTER) += vf_convolution.o
476OBJS-$(CONFIG_SCROLL_FILTER) += vf_scroll.o477OBJS-$(CONFIG_SCROLL_FILTER) += vf_scroll.o
477OBJS-$(CONFIG_SEGMENT_FILTER) += f_segment.o478OBJS-$(CONFIG_SEGMENT_FILTER) += f_segment.o
libavfilter/allfilters.c+1−0
@@ -444,6 +444,7 @@ extern const FFFilter ff_vf_scale_vulkan; · 443 unmodified lines
444extern const FFFilter ff_vf_scale2ref;444extern const FFFilter ff_vf_scale2ref;
445extern const FFFilter ff_vf_scale2ref_npp;445extern const FFFilter ff_vf_scale2ref_npp;
446extern const FFFilter ff_vf_scdet;446extern const FFFilter ff_vf_scdet;
447extern const FFFilter ff_vf_scdet_vulkan;
447extern const FFFilter ff_vf_scharr;448extern const FFFilter ff_vf_scharr;
448extern const FFFilter ff_vf_scroll;449extern const FFFilter ff_vf_scroll;
449extern const FFFilter ff_vf_segment;450extern const FFFilter ff_vf_segment;
libavfilter/vf_scdet_vulkan.c+412−0
@@ -0,0 +1,412 @@
1/*
2 * Copyright 2025 (c) Niklas Haas
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 "libavutil/avassert.h"
22#include "libavutil/vulkan_spirv.h"
23#include "libavutil/opt.h"
24#include "libavutil/timestamp.h"
25#include "vulkan_filter.h"
26
27#include "filters.h"
28
29typedef struct SceneDetectVulkanContext {
30 FFVulkanContext vkctx;
31
32 int initialized;
33 FFVkExecPool e;
34 AVVulkanDeviceQueueFamily *qf;
35 FFVulkanShader shd;
36 AVBufferPool *det_buf_pool;
37
38 double threshold;
39 int sc_pass;
40
41 int nb_planes;
42 double prev_mafd;
43 AVFrame *prev;
44 AVFrame *cur;
45} SceneDetectVulkanContext;
46
47typedef struct SceneDetectBuf {
48#define SLICES 16
49 uint32_t frame_sad[SLICES];
50} SceneDetectBuf;
51
52static av_cold int init_filter(AVFilterContext *ctx)
53{
54 int err;
55 uint8_t *spv_data;
56 size_t spv_len;
57 void *spv_opaque = NULL;
58 SceneDetectVulkanContext *s = ctx->priv;
59 FFVulkanContext *vkctx = &s->vkctx;
60 FFVulkanShader *shd;
61 FFVkSPIRVCompiler *spv;
62 FFVulkanDescriptorSetBinding *desc;
63
64 const AVPixFmtDescriptor *pixdesc = av_pix_fmt_desc_get(s->vkctx.input_format);
65 const int lumaonly = !(pixdesc->flags & AV_PIX_FMT_FLAG_RGB) &&
66 (pixdesc->flags & AV_PIX_FMT_FLAG_PLANAR);
67 s->nb_planes = lumaonly ? 1 : av_pix_fmt_count_planes(s->vkctx.input_format);
68
69 spv = ff_vk_spirv_init();
70 if (!spv) {
71 av_log(ctx, AV_LOG_ERROR, "Unable to initialize SPIR-V compiler!\n");
72 return AVERROR_EXTERNAL;
73 }
74
75 s->qf = ff_vk_qf_find(vkctx, VK_QUEUE_COMPUTE_BIT, 0);
76 if (!s->qf) {
77 av_log(ctx, AV_LOG_ERROR, "Device has no compute queues\n");
78 err = AVERROR(ENOTSUP);
79 goto fail;
80 }
81
82 RET(ff_vk_exec_pool_init(vkctx, s->qf, &s->e, s->qf->num*4, 0, 0, 0, NULL));
83 RET(ff_vk_shader_init(vkctx, &s->shd, "scdet",
84 VK_SHADER_STAGE_COMPUTE_BIT,
85 (const char *[]) { "GL_KHR_shader_subgroup_arithmetic" }, 1,
86 32, 32, 1,
87 0));
88 shd = &s->shd;
89
90 desc = (FFVulkanDescriptorSetBinding []) {
91 {
92 .name = "prev_img",
93 .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
94 .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.input_format, FF_VK_REP_UINT),
95 .mem_quali = "readonly",
96 .dimensions = 2,
97 .elems = av_pix_fmt_count_planes(s->vkctx.input_format),
98 .stages = VK_SHADER_STAGE_COMPUTE_BIT,
99 }, {
100 .name = "cur_img",
101 .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
102 .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.input_format, FF_VK_REP_UINT),
103 .mem_quali = "readonly",
104 .dimensions = 2,
105 .elems = av_pix_fmt_count_planes(s->vkctx.input_format),
106 .stages = VK_SHADER_STAGE_COMPUTE_BIT,
107 }, {
108 .name = "sad_buffer",
109 .type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
110 .stages = VK_SHADER_STAGE_COMPUTE_BIT,
111 .buf_content = "uint frame_sad[];",
112 }
113 };
114
115 RET(ff_vk_shader_add_descriptor_set(vkctx, &s->shd, desc, 3, 0, 0));
116
117 GLSLC(0, shared uint wg_sum; );
118 GLSLC(0, void main() );
119 GLSLC(0, { );
120 GLSLF(1, const uint slice = gl_WorkGroupID.x %% %u; ,SLICES);
121 GLSLC(1, const ivec2 pos = ivec2(gl_GlobalInvocationID.xy); );
122 GLSLC(1, wg_sum = 0; );
123 GLSLC(1, barrier(); );
124 for (int i = 0; i < s->nb_planes; i++) {
125 GLSLF(1, if (IS_WITHIN(pos, imageSize(cur_img[%d]))) { ,i);
126 GLSLF(2, uvec4 prev = imageLoad(prev_img[%d], pos); ,i);
127 GLSLF(2, uvec4 cur = imageLoad(cur_img[%d], pos); ,i);
128 GLSLC(2, uvec4 sad = abs(ivec4(cur) - ivec4(prev)); );
129 GLSLC(2, uint sum = subgroupAdd(sad.x + sad.y + sad.z); );
130 GLSLC(2, if (subgroupElect()) );
131 GLSLC(3, atomicAdd(wg_sum, sum); );
132 GLSLC(1, } );
133 }
134 GLSLC(1, barrier(); );
135 GLSLC(1, if (gl_LocalInvocationIndex == 0) );
136 GLSLC(2, atomicAdd(frame_sad[slice], wg_sum); );
137 GLSLC(0, } );
138
139 RET(spv->compile_shader(vkctx, spv, &s->shd, &spv_data, &spv_len, "main",
140 &spv_opaque));
141 RET(ff_vk_shader_link(vkctx, &s->shd, spv_data, spv_len, "main"));
142
143 RET(ff_vk_shader_register_exec(vkctx, &s->e, &s->shd));
144
145 s->initialized = 1;
146
147fail:
148 if (spv_opaque)
149 spv->free_shader(spv, &spv_opaque);
150 if (spv)
151 spv->uninit(&spv);
152
153 return err;
154}
155
156static double evaluate(AVFilterContext *ctx, const SceneDetectBuf *buf)
157{
158 SceneDetectVulkanContext *s = ctx->priv;
159 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(s->vkctx.input_format);
160 const AVFilterLink *inlink = ctx->inputs[0];
161 uint64_t count;
162 double mafd, diff;
163
164 uint64_t sad = 0;
165 for (int i = 0; i < SLICES; i++)
166 sad += buf->frame_sad[i];
167
168 av_assert2(s->nb_planes == 1 || !(desc->log2_chroma_w || desc->log2_chroma_h));
169 count = s->nb_planes * inlink->w * inlink->h;
170 mafd = (double) sad * 100.0 / count / (1ULL << desc->comp[0].depth);
171 diff = fabs(mafd - s->prev_mafd);
172 s->prev_mafd = mafd;
173
174 return av_clipf(FFMIN(mafd, diff), 0.0, 100.0);
175}
176
177static int scdet_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
178{
179 int err;
180 AVFilterContext *ctx = link->dst;
181 SceneDetectVulkanContext *s = ctx->priv;
182 AVFilterLink *outlink = ctx->outputs[0];
183
184 VkImageView prev_views[AV_NUM_DATA_POINTERS];
185 VkImageView cur_views[AV_NUM_DATA_POINTERS];
186 VkImageMemoryBarrier2 img_bar[8];
187 int nb_img_bar = 0;
188
189 FFVulkanContext *vkctx = &s->vkctx;
190 FFVulkanFunctions *vk = &vkctx->vkfn;
191 FFVkExecContext *exec = NULL;
192 AVBufferRef *buf = NULL;
193 FFVkBuffer *buf_vk;
194
195 SceneDetectBuf *sad;
196 double score = 0.0;
197 char str[64];
198
199 if (!s->initialized)
200 RET(init_filter(ctx));
201
202 av_frame_free(&s->prev);
203 s->prev = s->cur;
204 s->cur = av_frame_clone(in);
205 if (!s->prev)
206 goto done;
207
208 RET(ff_vk_get_pooled_buffer(vkctx, &s->det_buf_pool, &buf,
209 VK_BUFFER_USAGE_TRANSFER_DST_BIT |
210 VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
211 NULL,
212 sizeof(SceneDetectBuf),
213 VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
214 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
215 VK_MEMORY_PROPERTY_HOST_COHERENT_BIT));
216 buf_vk = (FFVkBuffer *)buf->data;
217 sad = (SceneDetectBuf *) buf_vk->mapped_mem;
218
219 exec = ff_vk_exec_get(vkctx, &s->e);
220 ff_vk_exec_start(vkctx, exec);
221
222 RET(ff_vk_exec_add_dep_frame(vkctx, exec, s->prev,
223 VK_PIPELINE_STAGE_2_NONE,
224 VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT));
225 RET(ff_vk_create_imageviews(vkctx, exec, prev_views, s->prev, FF_VK_REP_UINT));
226
227 ff_vk_shader_update_img_array(vkctx, exec, &s->shd, s->prev, prev_views, 0, 0,
228 VK_IMAGE_LAYOUT_GENERAL, VK_NULL_HANDLE);
229
230 ff_vk_frame_barrier(vkctx, exec, s->prev, img_bar, &nb_img_bar,
231 VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
232 VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
233 VK_ACCESS_SHADER_READ_BIT,
234 VK_IMAGE_LAYOUT_GENERAL,
235 VK_QUEUE_FAMILY_IGNORED);
236
237 RET(ff_vk_exec_add_dep_frame(vkctx, exec, s->cur,
238 VK_PIPELINE_STAGE_2_NONE,
239 VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT));
240 RET(ff_vk_create_imageviews(vkctx, exec, cur_views, s->cur, FF_VK_REP_UINT));
241
242 ff_vk_shader_update_img_array(vkctx, exec, &s->shd, s->cur, cur_views, 0, 1,
243 VK_IMAGE_LAYOUT_GENERAL, VK_NULL_HANDLE);
244
245 ff_vk_frame_barrier(vkctx, exec, s->cur, img_bar, &nb_img_bar,
246 VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
247 VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
248 VK_ACCESS_SHADER_READ_BIT,
249 VK_IMAGE_LAYOUT_GENERAL,
250 VK_QUEUE_FAMILY_IGNORED);
251
252 /* zero buffer */
253 vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
254 .sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
255 .pBufferMemoryBarriers = &(VkBufferMemoryBarrier2) {
256 .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER_2,
257 .srcStageMask = VK_PIPELINE_STAGE_2_NONE,
258 .dstStageMask = VK_PIPELINE_STAGE_2_TRANSFER_BIT,
259 .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
260 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
261 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
262 .buffer = buf_vk->buf,
263 .size = buf_vk->size,
264 .offset = 0,
265 },
266 .bufferMemoryBarrierCount = 1,
267 });
268
269 vk->CmdFillBuffer(exec->buf, buf_vk->buf, 0, buf_vk->size, 0x0);
270
271 vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
272 .sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
273 .pImageMemoryBarriers = img_bar,
274 .imageMemoryBarrierCount = nb_img_bar,
275 .pBufferMemoryBarriers = &(VkBufferMemoryBarrier2) {
276 .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER_2,
277 .srcStageMask = VK_PIPELINE_STAGE_2_TRANSFER_BIT,
278 .dstStageMask = VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
279 .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
280 .dstAccessMask = VK_ACCESS_2_SHADER_STORAGE_READ_BIT |
281 VK_ACCESS_2_SHADER_STORAGE_WRITE_BIT,
282 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
283 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
284 .buffer = buf_vk->buf,
285 .size = buf_vk->size,
286 .offset = 0,
287 },
288 .bufferMemoryBarrierCount = 1,
289 });
290
291 RET(ff_vk_shader_update_desc_buffer(&s->vkctx, exec, &s->shd, 0, 2, 0,
292 buf_vk, 0, buf_vk->size,
293 VK_FORMAT_UNDEFINED));
294
295 ff_vk_exec_bind_shader(vkctx, exec, &s->shd);
296
297 vk->CmdDispatch(exec->buf,
298 FFALIGN(in->width, s->shd.lg_size[0]) / s->shd.lg_size[0],
299 FFALIGN(in->height, s->shd.lg_size[1]) / s->shd.lg_size[1],
300 s->shd.lg_size[2]);
301
302 vk->CmdPipelineBarrier2(exec->buf, &(VkDependencyInfo) {
303 .sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO,
304 .pBufferMemoryBarriers = &(VkBufferMemoryBarrier2) {
305 .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER_2,
306 .srcStageMask = VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
307 .dstStageMask = VK_PIPELINE_STAGE_2_HOST_BIT,
308 .srcAccessMask = VK_ACCESS_2_SHADER_STORAGE_READ_BIT |
309 VK_ACCESS_2_SHADER_STORAGE_WRITE_BIT,
310 .dstAccessMask = VK_ACCESS_HOST_READ_BIT,
311 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
312 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
313 .buffer = buf_vk->buf,
314 .size = buf_vk->size,
315 .offset = 0,
316 },
317 .bufferMemoryBarrierCount = 1,
318 });
319
320 RET(ff_vk_exec_submit(vkctx, exec));
321 ff_vk_exec_wait(vkctx, exec);
322 score = evaluate(ctx, sad);
323
324done:
325 snprintf(str, sizeof(str), "%0.3f", s->prev_mafd);
326 av_dict_set(&in->metadata, "lavfi.scd.mafd", str, 0);
327 snprintf(str, sizeof(str), "%0.3f", score);
328 av_dict_set(&in->metadata, "lavfi.scd.score", str, 0);
329
330 if (score >= s->threshold) {
331 const char *pts = av_ts2timestr(in->pts, &link->time_base);
332 av_dict_set(&in->metadata, "lavfi.scd.time", pts, 0);
333 av_log(s, AV_LOG_INFO, "lavfi.scd.score: %.3f, lavfi.scd.time: %s\n",
334 score, pts);
335 }
336
337 av_buffer_unref(&buf);
338 if (!s->sc_pass || score >= s->threshold)
339 return ff_filter_frame(outlink, in);
340 else {
341 av_frame_free(&in);
342 return 0;
343 }
344
345fail:
346 if (exec)
347 ff_vk_exec_discard_deps(&s->vkctx, exec);
348 av_frame_free(&in);
349 av_buffer_unref(&buf);
350 return err;
351}
352
353static void scdet_vulkan_uninit(AVFilterContext *avctx)
354{
355 SceneDetectVulkanContext *s = avctx->priv;
356 FFVulkanContext *vkctx = &s->vkctx;
357
358 av_frame_free(&s->prev);
359 av_frame_free(&s->cur);
360
361 ff_vk_exec_pool_free(vkctx, &s->e);
362 ff_vk_shader_free(vkctx, &s->shd);
363
364 av_buffer_pool_uninit(&s->det_buf_pool);
365
366 ff_vk_uninit(&s->vkctx);
367
368 s->initialized = 0;
369}
370
371#define OFFSET(x) offsetof(SceneDetectVulkanContext, x)
372#define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
373static const AVOption scdet_vulkan_options[] = {
374 { "threshold", "set scene change detect threshold", OFFSET(threshold), AV_OPT_TYPE_DOUBLE, {.dbl = 10.}, 0, 100., FLAGS },
375 { "t", "set scene change detect threshold", OFFSET(threshold), AV_OPT_TYPE_DOUBLE, {.dbl = 10.}, 0, 100., FLAGS },
376 { "sc_pass", "Set the flag to pass scene change frames", OFFSET(sc_pass), AV_OPT_TYPE_BOOL, {.i64 = 0 }, 0, 1, FLAGS },
377 { "s", "Set the flag to pass scene change frames", OFFSET(sc_pass), AV_OPT_TYPE_BOOL, {.i64 = 0 }, 0, 1, FLAGS },
378 { NULL }
379};
380
381AVFILTER_DEFINE_CLASS(scdet_vulkan);
382
383static const AVFilterPad scdet_vulkan_inputs[] = {
384 {
385 .name = "default",
386 .type = AVMEDIA_TYPE_VIDEO,
387 .filter_frame = &scdet_vulkan_filter_frame,
388 .config_props = &ff_vk_filter_config_input,
389 },
390};
391
392static const AVFilterPad scdet_vulkan_outputs[] = {
393 {
394 .name = "default",
395 .type = AVMEDIA_TYPE_VIDEO,
396 .config_props = &ff_vk_filter_config_output,
397 },
398};
399
400const FFFilter ff_vf_scdet_vulkan = {
401 .p.name = "scdet_vulkan",
402 .p.description = NULL_IF_CONFIG_SMALL("Detect video scene change"),
403 .p.priv_class = &scdet_vulkan_class,
404 .p.flags = AVFILTER_FLAG_HWDEVICE,
405 .priv_size = sizeof(SceneDetectVulkanContext),
406 .init = &ff_vk_filter_init,
407 .uninit = &scdet_vulkan_uninit,
408 FILTER_INPUTS(scdet_vulkan_inputs),
409 FILTER_OUTPUTS(scdet_vulkan_outputs),
410 FILTER_SINGLE_PIXFMT(AV_PIX_FMT_VULKAN),
411 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
412};

Keyboard shortcuts

General

Show keyboard shortcuts
?
Focus the current search
/

Issues and pull requests

Open filters
f

Issues, pull requests and Inbox

Select next item
j↓
Select previous item
k↑
Open selected item
Enter
Dismiss filters or clear selection
Esc

Repository code

Go to file
t