git.sesse.net Git - ffmpeg/blob - libavfilter/vf_avgblur_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
  23 #define CGS 32
  24
  25 typedef struct AvgBlurVulkanContext {
  26     VulkanFilterContext vkctx;
  27
  28     int initialized;
  29     FFVkExecContext *exec;
  30     VulkanPipeline *pl_hor;
  31     VulkanPipeline *pl_ver;
  32
  33     /* Shader updators, must be in the main filter struct */
  34     VkDescriptorImageInfo input_images[3];
  35     VkDescriptorImageInfo tmp_images[3];
  36     VkDescriptorImageInfo output_images[3];
  37
  38     int size_x;
  39     int size_y;
  40     int planes;
  41 } AvgBlurVulkanContext;
  42
  43 static const char blur_kernel[] = {
  44     C(0, shared vec4 cache[DIR(gl_WorkGroupSize) + FILTER_RADIUS*2 + 1];           )
  45     C(0,                                                                           )
  46     C(0, void distort(const ivec2 pos, const int idx)                              )
  47     C(0, {                                                                         )
  48     C(1,     const uint cp = DIR(gl_LocalInvocationID) + FILTER_RADIUS;            )
  49     C(0,                                                                           )
  50     C(1,     cache[cp] = texture(input_img[idx], pos);                             )
  51     C(0,                                                                           )
  52     C(1,     const ivec2 loc_l = pos - INC(FILTER_RADIUS);                         )
  53     C(1,     cache[cp - FILTER_RADIUS] = texture(input_img[idx], loc_l);           )
  54     C(0,                                                                           )
  55     C(1,     const ivec2 loc_h = pos + INC(DIR(gl_WorkGroupSize));                 )
  56     C(1,     cache[cp + DIR(gl_WorkGroupSize)] = texture(input_img[idx], loc_h);   )
  57     C(0,                                                                           )
  58     C(1,     barrier();                                                            )
  59     C(0,                                                                           )
  60     C(1,     vec4 sum = vec4(0);                                                   )
  61     C(1,     for (int p = -FILTER_RADIUS; p <= FILTER_RADIUS; p++)                 )
  62     C(2,         sum += cache[cp + p];                                             )
  63     C(0,                                                                           )
  64     C(1,     sum /= vec4(FILTER_RADIUS*2 + 1);                                     )
  65     C(1,     imageStore(output_img[idx], pos, sum);                                )
  66     C(0, }                                                                         )
  67 };
  68
  69 static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
  70 {
  71     int err;
  72     SPIRVShader *shd;
  73     AvgBlurVulkanContext *s = ctx->priv;
  74     const int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
  75     VkSampler *sampler = ff_vk_init_sampler(ctx, 1, VK_FILTER_LINEAR);
  76
  77     VulkanDescriptorSetBinding desc_i[2] = {
  78         {
  79             .name       = "input_img",
  80             .type       = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
  81             .dimensions = 2,
  82             .elems      = planes,
  83             .stages     = VK_SHADER_STAGE_COMPUTE_BIT,
  84             .samplers   = DUP_SAMPLER_ARRAY4(*sampler),
  85         },
  86         {
  87             .name       = "output_img",
  88             .type       = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
  89             .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format),
  90             .mem_quali  = "writeonly",
  91             .dimensions = 2,
  92             .elems      = planes,
  93             .stages     = VK_SHADER_STAGE_COMPUTE_BIT,
  94         },
  95     };
  96
  97     if (!sampler)
  98         return AVERROR_EXTERNAL;
  99
 100     s->vkctx.queue_family_idx = s->vkctx.hwctx->queue_family_comp_index;
 101     s->vkctx.queue_count = GET_QUEUE_COUNT(s->vkctx.hwctx, 0, 1, 0);
 102     s->vkctx.cur_queue_idx = rand() % s->vkctx.queue_count;
 103
 104     { /* Create shader for the horizontal pass */
 105         desc_i[0].updater = s->input_images;
 106         desc_i[1].updater = s->tmp_images;
 107
 108         s->pl_hor = ff_vk_create_pipeline(ctx);
 109         if (!s->pl_hor)
 110             return AVERROR(ENOMEM);
 111
 112         shd = ff_vk_init_shader(ctx, s->pl_hor, "avgblur_compute_hor",
 113                                 VK_SHADER_STAGE_COMPUTE_BIT);
 114
 115         ff_vk_set_compute_shader_sizes(ctx, shd, (int [3]){ CGS, 1, 1 });
 116
 117         RET(ff_vk_add_descriptor_set(ctx, s->pl_hor, shd, desc_i, 2, 0));
 118
 119         GLSLF(0, #define FILTER_RADIUS (%i)                     ,s->size_x - 1);
 120         GLSLC(0, #define INC(x) (ivec2(x, 0))                                 );
 121         GLSLC(0, #define DIR(var) (var.x)                                     );
 122         GLSLD(   blur_kernel                                                  );
 123         GLSLC(0, void main()                                                  );
 124         GLSLC(0, {                                                            );
 125         GLSLC(1,     ivec2 size;                                              );
 126         GLSLC(1,     const ivec2 pos = ivec2(gl_GlobalInvocationID.xy);       );
 127         for (int i = 0; i < planes; i++) {
 128             GLSLC(0,                                                          );
 129             GLSLF(1,  size = imageSize(output_img[%i]);                     ,i);
 130             GLSLC(1,  if (IS_WITHIN(pos, size)) {                             );
 131             if (s->planes & (1 << i)) {
 132                 GLSLF(2, distort(pos, %i);                                  ,i);
 133             } else {
 134                 GLSLF(2, vec4 res = texture(input_img[%i], pos);            ,i);
 135                 GLSLF(2, imageStore(output_img[%i], pos, res);              ,i);
 136             }
 137             GLSLC(1, }                                                        );
 138         }
 139         GLSLC(0, }                                                            );
 140
 141         RET(ff_vk_compile_shader(ctx, shd, "main"));
 142
 143         RET(ff_vk_init_pipeline_layout(ctx, s->pl_hor));
 144         RET(ff_vk_init_compute_pipeline(ctx, s->pl_hor));
 145     }
 146
 147     { /* Create shader for the vertical pass */
 148         desc_i[0].updater = s->tmp_images;
 149         desc_i[1].updater = s->output_images;
 150
 151         s->pl_ver = ff_vk_create_pipeline(ctx);
 152         if (!s->pl_ver)
 153             return AVERROR(ENOMEM);
 154
 155         shd = ff_vk_init_shader(ctx, s->pl_ver, "avgblur_compute_ver",
 156                                 VK_SHADER_STAGE_COMPUTE_BIT);
 157
 158         ff_vk_set_compute_shader_sizes(ctx, shd, (int [3]){ 1, CGS, 1 });
 159
 160         RET(ff_vk_add_descriptor_set(ctx, s->pl_ver, shd, desc_i, 2, 0));
 161
 162         GLSLF(0, #define FILTER_RADIUS (%i)                     ,s->size_y - 1);
 163         GLSLC(0, #define INC(x) (ivec2(0, x))                                 );
 164         GLSLC(0, #define DIR(var) (var.y)                                     );
 165         GLSLD(   blur_kernel                                                  );
 166         GLSLC(0, void main()                                                  );
 167         GLSLC(0, {                                                            );
 168         GLSLC(1,     ivec2 size;                                              );
 169         GLSLC(1,     const ivec2 pos = ivec2(gl_GlobalInvocationID.xy);       );
 170         for (int i = 0; i < planes; i++) {
 171             GLSLC(0,                                                          );
 172             GLSLF(1,  size = imageSize(output_img[%i]);                     ,i);
 173             GLSLC(1,  if (IS_WITHIN(pos, size)) {                             );
 174             if (s->planes & (1 << i)) {
 175                 GLSLF(2, distort(pos, %i);                                  ,i);
 176             } else {
 177                 GLSLF(2, vec4 res = texture(input_img[%i], pos);            ,i);
 178                 GLSLF(2, imageStore(output_img[%i], pos, res);              ,i);
 179             }
 180             GLSLC(1, }                                                        );
 181         }
 182         GLSLC(0, }                                                            );
 183
 184         RET(ff_vk_compile_shader(ctx, shd, "main"));
 185
 186         RET(ff_vk_init_pipeline_layout(ctx, s->pl_ver));
 187         RET(ff_vk_init_compute_pipeline(ctx, s->pl_ver));
 188     }
 189
 190     /* Execution context */
 191     RET(ff_vk_create_exec_ctx(ctx, &s->exec));
 192
 193     s->initialized = 1;
 194
 195     return 0;
 196
 197 fail:
 198     return err;
 199 }
 200
 201 static int process_frames(AVFilterContext *avctx, AVFrame *out_f, AVFrame *tmp_f, AVFrame *in_f)
 202 {
 203     int err;
 204     VkCommandBuffer cmd_buf;
 205     AvgBlurVulkanContext *s = avctx->priv;
 206     AVVkFrame *in = (AVVkFrame *)in_f->data[0];
 207     AVVkFrame *tmp = (AVVkFrame *)tmp_f->data[0];
 208     AVVkFrame *out = (AVVkFrame *)out_f->data[0];
 209     int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
 210
 211     /* Update descriptors and init the exec context */
 212     ff_vk_start_exec_recording(avctx, s->exec);
 213     cmd_buf = ff_vk_get_exec_buf(avctx, s->exec);
 214
 215     for (int i = 0; i < planes; i++) {
 216         RET(ff_vk_create_imageview(avctx, s->exec, &s->input_images[i].imageView,
 217                                    in->img[i],
 218                                    av_vkfmt_from_pixfmt(s->vkctx.input_format)[i],
 219                                    ff_comp_identity_map));
 220
 221         RET(ff_vk_create_imageview(avctx, s->exec, &s->tmp_images[i].imageView,
 222                                    tmp->img[i],
 223                                    av_vkfmt_from_pixfmt(s->vkctx.output_format)[i],
 224                                    ff_comp_identity_map));
 225
 226         RET(ff_vk_create_imageview(avctx, s->exec, &s->output_images[i].imageView,
 227                                    out->img[i],
 228                                    av_vkfmt_from_pixfmt(s->vkctx.output_format)[i],
 229                                    ff_comp_identity_map));
 230
 231         s->input_images[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
 232         s->tmp_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL;
 233         s->output_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL;
 234     }
 235
 236     ff_vk_update_descriptor_set(avctx, s->pl_hor, 0);
 237     ff_vk_update_descriptor_set(avctx, s->pl_ver, 0);
 238
 239     for (int i = 0; i < planes; i++) {
 240         VkImageMemoryBarrier bar[] = {
 241             {
 242                 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
 243                 .srcAccessMask = 0,
 244                 .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
 245                 .oldLayout = in->layout[i],
 246                 .newLayout = s->input_images[i].imageLayout,
 247                 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 248                 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 249                 .image = in->img[i],
 250                 .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
 251                 .subresourceRange.levelCount = 1,
 252                 .subresourceRange.layerCount = 1,
 253             },
 254             {
 255                 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
 256                 .srcAccessMask = 0,
 257                 .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_SHADER_READ_BIT,
 258                 .oldLayout = tmp->layout[i],
 259                 .newLayout = s->tmp_images[i].imageLayout,
 260                 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 261                 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 262                 .image = tmp->img[i],
 263                 .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
 264                 .subresourceRange.levelCount = 1,
 265                 .subresourceRange.layerCount = 1,
 266             },
 267             {
 268                 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
 269                 .srcAccessMask = 0,
 270                 .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
 271                 .oldLayout = out->layout[i],
 272                 .newLayout = s->output_images[i].imageLayout,
 273                 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 274                 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 275                 .image = out->img[i],
 276                 .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
 277                 .subresourceRange.levelCount = 1,
 278                 .subresourceRange.layerCount = 1,
 279             },
 280         };
 281
 282         vkCmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
 283                              VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0,
 284                              0, NULL, 0, NULL, FF_ARRAY_ELEMS(bar), bar);
 285
 286         in->layout[i]  = bar[0].newLayout;
 287         in->access[i]  = bar[0].dstAccessMask;
 288
 289         tmp->layout[i] = bar[1].newLayout;
 290         tmp->access[i] = bar[1].dstAccessMask;
 291
 292         out->layout[i] = bar[2].newLayout;
 293         out->access[i] = bar[2].dstAccessMask;
 294     }
 295
 296     ff_vk_bind_pipeline_exec(avctx, s->exec, s->pl_hor);
 297
 298     vkCmdDispatch(cmd_buf, FFALIGN(s->vkctx.output_width, CGS)/CGS,
 299                   s->vkctx.output_height, 1);
 300
 301     ff_vk_bind_pipeline_exec(avctx, s->exec, s->pl_ver);
 302
 303     vkCmdDispatch(cmd_buf, s->vkctx.output_width,
 304                   FFALIGN(s->vkctx.output_height, CGS)/CGS, 1);
 305
 306     ff_vk_add_exec_dep(avctx, s->exec, in_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
 307     ff_vk_add_exec_dep(avctx, s->exec, out_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
 308
 309     err = ff_vk_submit_exec_queue(avctx, s->exec);
 310     if (err)
 311         return err;
 312
 313     return err;
 314
 315 fail:
 316     ff_vk_discard_exec_deps(avctx, s->exec);
 317     return err;
 318 }
 319
 320 static int avgblur_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
 321 {
 322     int err;
 323     AVFrame *tmp = NULL, *out = NULL;
 324     AVFilterContext *ctx = link->dst;
 325     AvgBlurVulkanContext *s = ctx->priv;
 326     AVFilterLink *outlink = ctx->outputs[0];
 327
 328     out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
 329     if (!out) {
 330         err = AVERROR(ENOMEM);
 331         goto fail;
 332     }
 333
 334     tmp = ff_get_video_buffer(outlink, outlink->w, outlink->h);
 335     if (!out) {
 336         err = AVERROR(ENOMEM);
 337         goto fail;
 338     }
 339
 340     if (!s->initialized)
 341         RET(init_filter(ctx, in));
 342
 343     RET(process_frames(ctx, out, tmp, in));
 344
 345     err = av_frame_copy_props(out, in);
 346     if (err < 0)
 347         goto fail;
 348
 349     av_frame_free(&in);
 350     av_frame_free(&tmp);
 351
 352     return ff_filter_frame(outlink, out);
 353
 354 fail:
 355     av_frame_free(&in);
 356     av_frame_free(&tmp);
 357     av_frame_free(&out);
 358     return err;
 359 }
 360
 361 static void avgblur_vulkan_uninit(AVFilterContext *avctx)
 362 {
 363     AvgBlurVulkanContext *s = avctx->priv;
 364
 365     ff_vk_filter_uninit(avctx);
 366
 367     s->initialized = 0;
 368 }
 369
 370 #define OFFSET(x) offsetof(AvgBlurVulkanContext, x)
 371 #define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
 372 static const AVOption avgblur_vulkan_options[] = {
 373     { "sizeX",  "Set horizontal radius", OFFSET(size_x), AV_OPT_TYPE_INT, {.i64 = 3}, 1, 32, .flags = FLAGS },
 374     { "planes", "Set planes to filter (bitmask)", OFFSET(planes), AV_OPT_TYPE_INT, {.i64 = 0xF}, 0, 0xF, .flags = FLAGS },
 375     { "sizeY",  "Set vertical radius", OFFSET(size_y), AV_OPT_TYPE_INT, {.i64 = 3}, 1, 32, .flags = FLAGS },
 376     { NULL },
 377 };
 378
 379 AVFILTER_DEFINE_CLASS(avgblur_vulkan);
 380
 381 static const AVFilterPad avgblur_vulkan_inputs[] = {
 382     {
 383         .name         = "default",
 384         .type         = AVMEDIA_TYPE_VIDEO,
 385         .filter_frame = &avgblur_vulkan_filter_frame,
 386         .config_props = &ff_vk_filter_config_input,
 387     },
 388     { NULL }
 389 };
 390
 391 static const AVFilterPad avgblur_vulkan_outputs[] = {
 392     {
 393         .name = "default",
 394         .type = AVMEDIA_TYPE_VIDEO,
 395         .config_props = &ff_vk_filter_config_output,
 396     },
 397     { NULL }
 398 };
 399
 400 AVFilter ff_vf_avgblur_vulkan = {
 401     .name           = "avgblur_vulkan",
 402     .description    = NULL_IF_CONFIG_SMALL("Apply avgblur mask to input video"),
 403     .priv_size      = sizeof(AvgBlurVulkanContext),
 404     .init           = &ff_vk_filter_init,
 405     .uninit         = &avgblur_vulkan_uninit,
 406     .query_formats  = &ff_vk_filter_query_formats,
 407     .inputs         = avgblur_vulkan_inputs,
 408     .outputs        = avgblur_vulkan_outputs,
 409     .priv_class     = &avgblur_vulkan_class,
 410     .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
 411 };