]> git.sesse.net Git - ffmpeg/blob - libavfilter/vf_overlay_vulkan.c
lavfi: add an overlay_vulkan filter
[ffmpeg] / libavfilter / vf_overlay_vulkan.c
1 /*
2  * This file is part of FFmpeg.
3  *
4  * FFmpeg is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU Lesser General Public
6  * License as published by the Free Software Foundation; either
7  * version 2.1 of the License, or (at your option) any later version.
8  *
9  * FFmpeg is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
12  * Lesser General Public License for more details.
13  *
14  * You should have received a copy of the GNU Lesser General Public
15  * License along with FFmpeg; if not, write to the Free Software
16  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17  */
18
19 #include "libavutil/opt.h"
20 #include "vulkan.h"
21 #include "internal.h"
22 #include "framesync.h"
23
24 #define CGROUPS (int [3]){ 32, 32, 1 }
25
26 typedef struct OverlayVulkanContext {
27     VulkanFilterContext vkctx;
28
29     int initialized;
30     VulkanPipeline *pl;
31     FFVkExecContext *exec;
32     FFFrameSync fs;
33     FFVkBuffer params_buf;
34
35     /* Shader updators, must be in the main filter struct */
36     VkDescriptorImageInfo main_images[3];
37     VkDescriptorImageInfo overlay_images[3];
38     VkDescriptorImageInfo output_images[3];
39     VkDescriptorBufferInfo params_desc;
40
41     int overlay_x;
42     int overlay_y;
43     int overlay_w;
44     int overlay_h;
45 } OverlayVulkanContext;
46
47 static const char overlay_noalpha[] = {
48     C(0, void overlay_noalpha(int i, ivec2 pos)                                )
49     C(0, {                                                                     )
50     C(1,     if ((o_offset[i].x <= pos.x) && (o_offset[i].y <= pos.y) &&
51                  (pos.x < (o_offset[i].x + o_size[i].x)) &&
52                  (pos.y < (o_offset[i].y + o_size[i].y))) {                    )
53     C(2,         vec4 res = texture(overlay_img[i], pos - o_offset[i]);        )
54     C(2,         imageStore(output_img[i], pos, res);                          )
55     C(1,     } else {                                                          )
56     C(2,         vec4 res = texture(main_img[i], pos);                         )
57     C(2,         imageStore(output_img[i], pos, res);                          )
58     C(1,     }                                                                 )
59     C(0, }                                                                     )
60 };
61
62 static av_cold int init_filter(AVFilterContext *ctx)
63 {
64     int err;
65     OverlayVulkanContext *s = ctx->priv;
66     VkSampler *sampler = ff_vk_init_sampler(ctx, 1, VK_FILTER_LINEAR);
67     if (!sampler)
68         return AVERROR_EXTERNAL;
69
70     s->pl = ff_vk_create_pipeline(ctx);
71     if (!s->pl)
72         return AVERROR(ENOMEM);
73
74     { /* Create the shader */
75         const int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
76
77         VulkanDescriptorSetBinding desc_i[3] = {
78             {
79                 .name       = "main_img",
80                 .type       = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
81                 .dimensions = 2,
82                 .elems      = planes,
83                 .stages     = VK_SHADER_STAGE_COMPUTE_BIT,
84                 .updater    = s->main_images,
85                 .samplers   = DUP_SAMPLER_ARRAY4(*sampler),
86             },
87             {
88                 .name       = "overlay_img",
89                 .type       = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
90                 .dimensions = 2,
91                 .elems      = planes,
92                 .stages     = VK_SHADER_STAGE_COMPUTE_BIT,
93                 .updater    = s->overlay_images,
94                 .samplers   = DUP_SAMPLER_ARRAY4(*sampler),
95             },
96             {
97                 .name       = "output_img",
98                 .type       = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
99                 .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format),
100                 .mem_quali  = "writeonly",
101                 .dimensions = 2,
102                 .elems      = planes,
103                 .stages     = VK_SHADER_STAGE_COMPUTE_BIT,
104                 .updater    = s->output_images,
105             },
106         };
107
108         VulkanDescriptorSetBinding desc_b = {
109             .name        = "params",
110             .type        = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
111             .mem_quali   = "readonly",
112             .mem_layout  = "std430",
113             .stages      = VK_SHADER_STAGE_COMPUTE_BIT,
114             .updater     = &s->params_desc,
115             .buf_content = "ivec2 o_offset[3], o_size[3];",
116         };
117
118         SPIRVShader *shd = ff_vk_init_shader(ctx, s->pl, "overlay_compute",
119                                              VK_SHADER_STAGE_COMPUTE_BIT);
120         if (!shd)
121             return AVERROR(ENOMEM);
122
123         ff_vk_set_compute_shader_sizes(ctx, shd, CGROUPS);
124
125         RET(ff_vk_add_descriptor_set(ctx, s->pl, shd,  desc_i, 3, 0)); /* set 0 */
126         RET(ff_vk_add_descriptor_set(ctx, s->pl, shd, &desc_b, 1, 0)); /* set 1 */
127
128         GLSLD(   overlay_noalpha                                              );
129         GLSLC(0, void main()                                                  );
130         GLSLC(0, {                                                            );
131         GLSLC(1,     ivec2 pos = ivec2(gl_GlobalInvocationID.xy);             );
132         GLSLF(1,     int planes = %i;                                  ,planes);
133         GLSLC(1,     for (int i = 0; i < planes; i++) {                       );
134         GLSLC(2,         overlay_noalpha(i, pos);                             );
135         GLSLC(1,     }                                                        );
136         GLSLC(0, }                                                            );
137
138         RET(ff_vk_compile_shader(ctx, shd, "main"));
139     }
140
141     RET(ff_vk_init_pipeline_layout(ctx, s->pl));
142     RET(ff_vk_init_compute_pipeline(ctx, s->pl));
143
144     { /* Create and update buffer */
145         const AVPixFmtDescriptor *desc;
146
147         /* NOTE: std430 requires the same identical struct layout, padding and
148          * alignment as C, so we're allowed to do this, as this will map
149          * exactly to what the shader recieves */
150         struct {
151             int32_t o_offset[2*3];
152             int32_t o_size[2*3];
153         } *par;
154
155         err = ff_vk_create_buf(ctx, &s->params_buf,
156                                sizeof(*par),
157                                VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
158                                VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT);
159         if (err)
160             return err;
161
162         err = ff_vk_map_buffers(ctx, &s->params_buf, (uint8_t **)&par, 1, 0);
163         if (err)
164             return err;
165
166         desc = av_pix_fmt_desc_get(s->vkctx.output_format);
167
168         par->o_offset[0] = s->overlay_x;
169         par->o_offset[1] = s->overlay_y;
170         par->o_offset[2] = par->o_offset[0] >> desc->log2_chroma_w;
171         par->o_offset[3] = par->o_offset[1] >> desc->log2_chroma_h;
172         par->o_offset[4] = par->o_offset[0] >> desc->log2_chroma_w;
173         par->o_offset[5] = par->o_offset[1] >> desc->log2_chroma_h;
174
175         par->o_size[0] = s->overlay_w;
176         par->o_size[1] = s->overlay_h;
177         par->o_size[2] = par->o_size[0] >> desc->log2_chroma_w;
178         par->o_size[3] = par->o_size[1] >> desc->log2_chroma_h;
179         par->o_size[4] = par->o_size[0] >> desc->log2_chroma_w;
180         par->o_size[5] = par->o_size[1] >> desc->log2_chroma_h;
181
182         err = ff_vk_unmap_buffers(ctx, &s->params_buf, 1, 1);
183         if (err)
184             return err;
185
186         s->params_desc.buffer = s->params_buf.buf;
187         s->params_desc.range  = VK_WHOLE_SIZE;
188
189         ff_vk_update_descriptor_set(ctx, s->pl, 1);
190     }
191
192     /* Execution context */
193     RET(ff_vk_create_exec_ctx(ctx, &s->exec,
194                               s->vkctx.hwctx->queue_family_comp_index));
195
196     s->initialized = 1;
197
198     return 0;
199
200 fail:
201     return err;
202 }
203
204 static int process_frames(AVFilterContext *avctx, AVFrame *out_f,
205                           AVFrame *main_f, AVFrame *overlay_f)
206 {
207     int err;
208     OverlayVulkanContext *s = avctx->priv;
209     int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
210
211     AVVkFrame *out     = (AVVkFrame *)out_f->data[0];
212     AVVkFrame *main    = (AVVkFrame *)main_f->data[0];
213     AVVkFrame *overlay = (AVVkFrame *)overlay_f->data[0];
214
215     AVHWFramesContext *main_fc = (AVHWFramesContext*)main_f->hw_frames_ctx->data;
216     AVHWFramesContext *overlay_fc = (AVHWFramesContext*)overlay_f->hw_frames_ctx->data;
217
218     for (int i = 0; i < planes; i++) {
219         RET(ff_vk_create_imageview(avctx, &s->main_images[i].imageView, main->img[i],
220                                    av_vkfmt_from_pixfmt(main_fc->sw_format)[i],
221                                    ff_comp_identity_map));
222
223         RET(ff_vk_create_imageview(avctx, &s->overlay_images[i].imageView, overlay->img[i],
224                                    av_vkfmt_from_pixfmt(overlay_fc->sw_format)[i],
225                                    ff_comp_identity_map));
226
227         RET(ff_vk_create_imageview(avctx, &s->output_images[i].imageView, out->img[i],
228                                    av_vkfmt_from_pixfmt(s->vkctx.output_format)[i],
229                                    ff_comp_identity_map));
230
231         s->main_images[i].imageLayout    = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
232         s->overlay_images[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
233         s->output_images[i].imageLayout  = VK_IMAGE_LAYOUT_GENERAL;
234     }
235
236     ff_vk_update_descriptor_set(avctx, s->pl, 0);
237
238     ff_vk_start_exec_recording(avctx, s->exec);
239
240     for (int i = 0; i < planes; i++) {
241         VkImageMemoryBarrier bar[3] = {
242             {
243                 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
244                 .srcAccessMask = 0,
245                 .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
246                 .oldLayout = main->layout[i],
247                 .newLayout = s->main_images[i].imageLayout,
248                 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
249                 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
250                 .image = main->img[i],
251                 .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
252                 .subresourceRange.levelCount = 1,
253                 .subresourceRange.layerCount = 1,
254             },
255             {
256                 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
257                 .srcAccessMask = 0,
258                 .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
259                 .oldLayout = overlay->layout[i],
260                 .newLayout = s->overlay_images[i].imageLayout,
261                 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
262                 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
263                 .image = overlay->img[i],
264                 .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
265                 .subresourceRange.levelCount = 1,
266                 .subresourceRange.layerCount = 1,
267             },
268             {
269                 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
270                 .srcAccessMask = 0,
271                 .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
272                 .oldLayout = out->layout[i],
273                 .newLayout = s->output_images[i].imageLayout,
274                 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
275                 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
276                 .image = out->img[i],
277                 .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
278                 .subresourceRange.levelCount = 1,
279                 .subresourceRange.layerCount = 1,
280             },
281         };
282
283         vkCmdPipelineBarrier(s->exec->buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
284                              VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0,
285                              0, NULL, 0, NULL, FF_ARRAY_ELEMS(bar), bar);
286
287         main->layout[i]    = bar[0].newLayout;
288         main->access[i]    = bar[0].dstAccessMask;
289
290         overlay->layout[i] = bar[1].newLayout;
291         overlay->access[i] = bar[1].dstAccessMask;
292
293         out->layout[i]     = bar[2].newLayout;
294         out->access[i]     = bar[2].dstAccessMask;
295     }
296
297     ff_vk_bind_pipeline_exec(avctx, s->exec, s->pl);
298
299     vkCmdDispatch(s->exec->buf,
300                   FFALIGN(s->vkctx.output_width,  CGROUPS[0])/CGROUPS[0],
301                   FFALIGN(s->vkctx.output_height, CGROUPS[1])/CGROUPS[1], 1);
302
303     ff_vk_add_exec_dep(avctx, s->exec, main_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
304     ff_vk_add_exec_dep(avctx, s->exec, overlay_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
305     ff_vk_add_exec_dep(avctx, s->exec, out_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
306
307     err = ff_vk_submit_exec_queue(avctx, s->exec);
308     if (err)
309         return err;
310
311 fail:
312
313     for (int i = 0; i < planes; i++) {
314         ff_vk_destroy_imageview(avctx, &s->main_images[i].imageView);
315         ff_vk_destroy_imageview(avctx, &s->overlay_images[i].imageView);
316         ff_vk_destroy_imageview(avctx, &s->output_images[i].imageView);
317     }
318
319     return err;
320 }
321
322 static int overlay_vulkan_blend(FFFrameSync *fs)
323 {
324     int err;
325     AVFilterContext *ctx = fs->parent;
326     OverlayVulkanContext *s = ctx->priv;
327     AVFilterLink *outlink = ctx->outputs[0];
328     AVFrame *input_main, *input_overlay, *out;
329
330     err = ff_framesync_get_frame(fs, 0, &input_main, 0);
331     if (err < 0)
332         goto fail;
333     err = ff_framesync_get_frame(fs, 1, &input_overlay, 0);
334     if (err < 0)
335         goto fail;
336
337     if (!input_main || !input_overlay)
338         return 0;
339
340     if (!s->initialized) {
341         AVHWFramesContext *main_fc = (AVHWFramesContext*)input_main->hw_frames_ctx->data;
342         AVHWFramesContext *overlay_fc = (AVHWFramesContext*)input_overlay->hw_frames_ctx->data;
343         if (main_fc->sw_format != overlay_fc->sw_format) {
344             av_log(ctx, AV_LOG_ERROR, "Mismatching sw formats!\n");
345             return AVERROR(EINVAL);
346         }
347
348         s->overlay_w = input_overlay->width;
349         s->overlay_h = input_overlay->height;
350
351         RET(init_filter(ctx));
352     }
353
354     out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
355     if (!out) {
356         err = AVERROR(ENOMEM);
357         goto fail;
358     }
359
360     RET(process_frames(ctx, out, input_main, input_overlay));
361
362     err = av_frame_copy_props(out, input_main);
363     if (err < 0)
364         goto fail;
365
366     return ff_filter_frame(outlink, out);
367
368 fail:
369     av_frame_free(&out);
370     return err;
371 }
372
373 static int overlay_vulkan_config_output(AVFilterLink *outlink)
374 {
375     int err;
376     AVFilterContext *avctx = outlink->src;
377     OverlayVulkanContext *s = avctx->priv;
378
379     err = ff_vk_filter_config_output(outlink);
380     if (err < 0)
381         return err;
382
383     err = ff_framesync_init_dualinput(&s->fs, avctx);
384     if (err < 0)
385         return err;
386
387     return ff_framesync_configure(&s->fs);
388 }
389
390 static int overlay_vulkan_activate(AVFilterContext *avctx)
391 {
392     OverlayVulkanContext *s = avctx->priv;
393
394     return ff_framesync_activate(&s->fs);
395 }
396
397 static av_cold int overlay_vulkan_init(AVFilterContext *avctx)
398 {
399     OverlayVulkanContext *s = avctx->priv;
400
401     s->fs.on_event = &overlay_vulkan_blend;
402
403     return ff_vk_filter_init(avctx);
404 }
405
406 static void overlay_vulkan_uninit(AVFilterContext *avctx)
407 {
408     OverlayVulkanContext *s = avctx->priv;
409
410     ff_vk_filter_uninit(avctx);
411     ff_framesync_uninit(&s->fs);
412
413     ff_vk_free_buf(avctx, &s->params_buf);
414
415     s->initialized = 0;
416 }
417
418 #define OFFSET(x) offsetof(OverlayVulkanContext, x)
419 #define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
420 static const AVOption overlay_vulkan_options[] = {
421     { "x", "Set horizontal offset", OFFSET(overlay_x), AV_OPT_TYPE_INT, {.i64 = 0}, 0, INT_MAX, .flags = FLAGS },
422     { "y", "Set vertical offset",   OFFSET(overlay_y), AV_OPT_TYPE_INT, {.i64 = 0}, 0, INT_MAX, .flags = FLAGS },
423     { NULL },
424 };
425
426 AVFILTER_DEFINE_CLASS(overlay_vulkan);
427
428 static const AVFilterPad overlay_vulkan_inputs[] = {
429     {
430         .name         = "main",
431         .type         = AVMEDIA_TYPE_VIDEO,
432         .config_props = &ff_vk_filter_config_input,
433     },
434     {
435         .name         = "overlay",
436         .type         = AVMEDIA_TYPE_VIDEO,
437         .config_props = &ff_vk_filter_config_input,
438     },
439     { NULL }
440 };
441
442 static const AVFilterPad overlay_vulkan_outputs[] = {
443     {
444         .name = "default",
445         .type = AVMEDIA_TYPE_VIDEO,
446         .config_props = &overlay_vulkan_config_output,
447     },
448     { NULL }
449 };
450
451 AVFilter ff_vf_overlay_vulkan = {
452     .name           = "overlay_vulkan",
453     .description    = NULL_IF_CONFIG_SMALL("Overlay a source on top of another"),
454     .priv_size      = sizeof(OverlayVulkanContext),
455     .init           = &overlay_vulkan_init,
456     .uninit         = &overlay_vulkan_uninit,
457     .query_formats  = &ff_vk_filter_query_formats,
458     .activate       = &overlay_vulkan_activate,
459     .inputs         = overlay_vulkan_inputs,
460     .outputs        = overlay_vulkan_outputs,
461     .priv_class     = &overlay_vulkan_class,
462     .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
463 };