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/random_seed.h"
  20 #include "libavutil/opt.h"
  21 #include "vulkan.h"
  22 #include "internal.h"
  23
  24 #define CGS 32
  25
  26 typedef struct AvgBlurVulkanContext {
  27     VulkanFilterContext vkctx;
  28
  29     int initialized;
  30     FFVkExecContext *exec;
  31     VulkanPipeline *pl_hor;
  32     VulkanPipeline *pl_ver;
  33
  34     /* Shader updators, must be in the main filter struct */
  35     VkDescriptorImageInfo input_images[3];
  36     VkDescriptorImageInfo tmp_images[3];
  37     VkDescriptorImageInfo output_images[3];
  38
  39     int size_x;
  40     int size_y;
  41     int planes;
  42 } AvgBlurVulkanContext;
  43
  44 static const char blur_kernel[] = {
  45     C(0, shared vec4 cache[DIR(gl_WorkGroupSize) + FILTER_RADIUS*2 + 1];           )
  46     C(0,                                                                           )
  47     C(0, void distort(const ivec2 pos, const int idx)                              )
  48     C(0, {                                                                         )
  49     C(1,     const uint cp = DIR(gl_LocalInvocationID) + FILTER_RADIUS;            )
  50     C(0,                                                                           )
  51     C(1,     cache[cp] = texture(input_img[idx], pos);                             )
  52     C(0,                                                                           )
  53     C(1,     const ivec2 loc_l = pos - INC(FILTER_RADIUS);                         )
  54     C(1,     cache[cp - FILTER_RADIUS] = texture(input_img[idx], loc_l);           )
  55     C(0,                                                                           )
  56     C(1,     const ivec2 loc_h = pos + INC(DIR(gl_WorkGroupSize));                 )
  57     C(1,     cache[cp + DIR(gl_WorkGroupSize)] = texture(input_img[idx], loc_h);   )
  58     C(0,                                                                           )
  59     C(1,     barrier();                                                            )
  60     C(0,                                                                           )
  61     C(1,     vec4 sum = vec4(0);                                                   )
  62     C(1,     for (int p = -FILTER_RADIUS; p <= FILTER_RADIUS; p++)                 )
  63     C(2,         sum += cache[cp + p];                                             )
  64     C(0,                                                                           )
  65     C(1,     sum /= vec4(FILTER_RADIUS*2 + 1);                                     )
  66     C(1,     imageStore(output_img[idx], pos, sum);                                )
  67     C(0, }                                                                         )
  68 };
  69
  70 static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
  71 {
  72     int err;
  73     SPIRVShader *shd;
  74     AvgBlurVulkanContext *s = ctx->priv;
  75     const int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
  76     VkSampler *sampler = ff_vk_init_sampler(ctx, 1, VK_FILTER_LINEAR);
  77
  78     VulkanDescriptorSetBinding desc_i[2] = {
  79         {
  80             .name       = "input_img",
  81             .type       = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
  82             .dimensions = 2,
  83             .elems      = planes,
  84             .stages     = VK_SHADER_STAGE_COMPUTE_BIT,
  85             .samplers   = DUP_SAMPLER_ARRAY4(*sampler),
  86         },
  87         {
  88             .name       = "output_img",
  89             .type       = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
  90             .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format),
  91             .mem_quali  = "writeonly",
  92             .dimensions = 2,
  93             .elems      = planes,
  94             .stages     = VK_SHADER_STAGE_COMPUTE_BIT,
  95         },
  96     };
  97
  98     if (!sampler)
  99         return AVERROR_EXTERNAL;
 100
 101     s->vkctx.queue_family_idx = s->vkctx.hwctx->queue_family_comp_index;
 102     s->vkctx.queue_count = GET_QUEUE_COUNT(s->vkctx.hwctx, 0, 1, 0);
 103     s->vkctx.cur_queue_idx = av_get_random_seed() % s->vkctx.queue_count;
 104
 105     { /* Create shader for the horizontal pass */
 106         desc_i[0].updater = s->input_images;
 107         desc_i[1].updater = s->tmp_images;
 108
 109         s->pl_hor = ff_vk_create_pipeline(ctx);
 110         if (!s->pl_hor)
 111             return AVERROR(ENOMEM);
 112
 113         shd = ff_vk_init_shader(ctx, s->pl_hor, "avgblur_compute_hor",
 114                                 VK_SHADER_STAGE_COMPUTE_BIT);
 115
 116         ff_vk_set_compute_shader_sizes(ctx, shd, (int [3]){ CGS, 1, 1 });
 117
 118         RET(ff_vk_add_descriptor_set(ctx, s->pl_hor, shd, desc_i, 2, 0));
 119
 120         GLSLF(0, #define FILTER_RADIUS (%i)                     ,s->size_x - 1);
 121         GLSLC(0, #define INC(x) (ivec2(x, 0))                                 );
 122         GLSLC(0, #define DIR(var) (var.x)                                     );
 123         GLSLD(   blur_kernel                                                  );
 124         GLSLC(0, void main()                                                  );
 125         GLSLC(0, {                                                            );
 126         GLSLC(1,     ivec2 size;                                              );
 127         GLSLC(1,     const ivec2 pos = ivec2(gl_GlobalInvocationID.xy);       );
 128         for (int i = 0; i < planes; i++) {
 129             GLSLC(0,                                                          );
 130             GLSLF(1,  size = imageSize(output_img[%i]);                     ,i);
 131             GLSLC(1,  if (IS_WITHIN(pos, size)) {                             );
 132             if (s->planes & (1 << i)) {
 133                 GLSLF(2, distort(pos, %i);                                  ,i);
 134             } else {
 135                 GLSLF(2, vec4 res = texture(input_img[%i], pos);            ,i);
 136                 GLSLF(2, imageStore(output_img[%i], pos, res);              ,i);
 137             }
 138             GLSLC(1, }                                                        );
 139         }
 140         GLSLC(0, }                                                            );
 141
 142         RET(ff_vk_compile_shader(ctx, shd, "main"));
 143
 144         RET(ff_vk_init_pipeline_layout(ctx, s->pl_hor));
 145         RET(ff_vk_init_compute_pipeline(ctx, s->pl_hor));
 146     }
 147
 148     { /* Create shader for the vertical pass */
 149         desc_i[0].updater = s->tmp_images;
 150         desc_i[1].updater = s->output_images;
 151
 152         s->pl_ver = ff_vk_create_pipeline(ctx);
 153         if (!s->pl_ver)
 154             return AVERROR(ENOMEM);
 155
 156         shd = ff_vk_init_shader(ctx, s->pl_ver, "avgblur_compute_ver",
 157                                 VK_SHADER_STAGE_COMPUTE_BIT);
 158
 159         ff_vk_set_compute_shader_sizes(ctx, shd, (int [3]){ 1, CGS, 1 });
 160
 161         RET(ff_vk_add_descriptor_set(ctx, s->pl_ver, shd, desc_i, 2, 0));
 162
 163         GLSLF(0, #define FILTER_RADIUS (%i)                     ,s->size_y - 1);
 164         GLSLC(0, #define INC(x) (ivec2(0, x))                                 );
 165         GLSLC(0, #define DIR(var) (var.y)                                     );
 166         GLSLD(   blur_kernel                                                  );
 167         GLSLC(0, void main()                                                  );
 168         GLSLC(0, {                                                            );
 169         GLSLC(1,     ivec2 size;                                              );
 170         GLSLC(1,     const ivec2 pos = ivec2(gl_GlobalInvocationID.xy);       );
 171         for (int i = 0; i < planes; i++) {
 172             GLSLC(0,                                                          );
 173             GLSLF(1,  size = imageSize(output_img[%i]);                     ,i);
 174             GLSLC(1,  if (IS_WITHIN(pos, size)) {                             );
 175             if (s->planes & (1 << i)) {
 176                 GLSLF(2, distort(pos, %i);                                  ,i);
 177             } else {
 178                 GLSLF(2, vec4 res = texture(input_img[%i], pos);            ,i);
 179                 GLSLF(2, imageStore(output_img[%i], pos, res);              ,i);
 180             }
 181             GLSLC(1, }                                                        );
 182         }
 183         GLSLC(0, }                                                            );
 184
 185         RET(ff_vk_compile_shader(ctx, shd, "main"));
 186
 187         RET(ff_vk_init_pipeline_layout(ctx, s->pl_ver));
 188         RET(ff_vk_init_compute_pipeline(ctx, s->pl_ver));
 189     }
 190
 191     /* Execution context */
 192     RET(ff_vk_create_exec_ctx(ctx, &s->exec));
 193
 194     s->initialized = 1;
 195
 196     return 0;
 197
 198 fail:
 199     return err;
 200 }
 201
 202 static int process_frames(AVFilterContext *avctx, AVFrame *out_f, AVFrame *tmp_f, AVFrame *in_f)
 203 {
 204     int err;
 205     VkCommandBuffer cmd_buf;
 206     AvgBlurVulkanContext *s = avctx->priv;
 207     AVVkFrame *in = (AVVkFrame *)in_f->data[0];
 208     AVVkFrame *tmp = (AVVkFrame *)tmp_f->data[0];
 209     AVVkFrame *out = (AVVkFrame *)out_f->data[0];
 210     int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
 211
 212     /* Update descriptors and init the exec context */
 213     ff_vk_start_exec_recording(avctx, s->exec);
 214     cmd_buf = ff_vk_get_exec_buf(avctx, s->exec);
 215
 216     for (int i = 0; i < planes; i++) {
 217         RET(ff_vk_create_imageview(avctx, s->exec, &s->input_images[i].imageView,
 218                                    in->img[i],
 219                                    av_vkfmt_from_pixfmt(s->vkctx.input_format)[i],
 220                                    ff_comp_identity_map));
 221
 222         RET(ff_vk_create_imageview(avctx, s->exec, &s->tmp_images[i].imageView,
 223                                    tmp->img[i],
 224                                    av_vkfmt_from_pixfmt(s->vkctx.output_format)[i],
 225                                    ff_comp_identity_map));
 226
 227         RET(ff_vk_create_imageview(avctx, s->exec, &s->output_images[i].imageView,
 228                                    out->img[i],
 229                                    av_vkfmt_from_pixfmt(s->vkctx.output_format)[i],
 230                                    ff_comp_identity_map));
 231
 232         s->input_images[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
 233         s->tmp_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL;
 234         s->output_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL;
 235     }
 236
 237     ff_vk_update_descriptor_set(avctx, s->pl_hor, 0);
 238     ff_vk_update_descriptor_set(avctx, s->pl_ver, 0);
 239
 240     for (int i = 0; i < planes; i++) {
 241         VkImageMemoryBarrier bar[] = {
 242             {
 243                 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
 244                 .srcAccessMask = 0,
 245                 .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
 246                 .oldLayout = in->layout[i],
 247                 .newLayout = s->input_images[i].imageLayout,
 248                 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 249                 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 250                 .image = in->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_WRITE_BIT | VK_ACCESS_SHADER_READ_BIT,
 259                 .oldLayout = tmp->layout[i],
 260                 .newLayout = s->tmp_images[i].imageLayout,
 261                 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 262                 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 263                 .image = tmp->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(cmd_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         in->layout[i]  = bar[0].newLayout;
 288         in->access[i]  = bar[0].dstAccessMask;
 289
 290         tmp->layout[i] = bar[1].newLayout;
 291         tmp->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_hor);
 298
 299     vkCmdDispatch(cmd_buf, FFALIGN(s->vkctx.output_width, CGS)/CGS,
 300                   s->vkctx.output_height, 1);
 301
 302     ff_vk_bind_pipeline_exec(avctx, s->exec, s->pl_ver);
 303
 304     vkCmdDispatch(cmd_buf, s->vkctx.output_width,
 305                   FFALIGN(s->vkctx.output_height, CGS)/CGS, 1);
 306
 307     ff_vk_add_exec_dep(avctx, s->exec, in_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
 308     ff_vk_add_exec_dep(avctx, s->exec, out_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
 309
 310     err = ff_vk_submit_exec_queue(avctx, s->exec);
 311     if (err)
 312         return err;
 313
 314     return err;
 315
 316 fail:
 317     ff_vk_discard_exec_deps(avctx, s->exec);
 318     return err;
 319 }
 320
 321 static int avgblur_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
 322 {
 323     int err;
 324     AVFrame *tmp = NULL, *out = NULL;
 325     AVFilterContext *ctx = link->dst;
 326     AvgBlurVulkanContext *s = ctx->priv;
 327     AVFilterLink *outlink = ctx->outputs[0];
 328
 329     out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
 330     if (!out) {
 331         err = AVERROR(ENOMEM);
 332         goto fail;
 333     }
 334
 335     tmp = ff_get_video_buffer(outlink, outlink->w, outlink->h);
 336     if (!out) {
 337         err = AVERROR(ENOMEM);
 338         goto fail;
 339     }
 340
 341     if (!s->initialized)
 342         RET(init_filter(ctx, in));
 343
 344     RET(process_frames(ctx, out, tmp, in));
 345
 346     err = av_frame_copy_props(out, in);
 347     if (err < 0)
 348         goto fail;
 349
 350     av_frame_free(&in);
 351     av_frame_free(&tmp);
 352
 353     return ff_filter_frame(outlink, out);
 354
 355 fail:
 356     av_frame_free(&in);
 357     av_frame_free(&tmp);
 358     av_frame_free(&out);
 359     return err;
 360 }
 361
 362 static void avgblur_vulkan_uninit(AVFilterContext *avctx)
 363 {
 364     AvgBlurVulkanContext *s = avctx->priv;
 365
 366     ff_vk_filter_uninit(avctx);
 367
 368     s->initialized = 0;
 369 }
 370
 371 #define OFFSET(x) offsetof(AvgBlurVulkanContext, x)
 372 #define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
 373 static const AVOption avgblur_vulkan_options[] = {
 374     { "sizeX",  "Set horizontal radius", OFFSET(size_x), AV_OPT_TYPE_INT, {.i64 = 3}, 1, 32, .flags = FLAGS },
 375     { "planes", "Set planes to filter (bitmask)", OFFSET(planes), AV_OPT_TYPE_INT, {.i64 = 0xF}, 0, 0xF, .flags = FLAGS },
 376     { "sizeY",  "Set vertical radius", OFFSET(size_y), AV_OPT_TYPE_INT, {.i64 = 3}, 1, 32, .flags = FLAGS },
 377     { NULL },
 378 };
 379
 380 AVFILTER_DEFINE_CLASS(avgblur_vulkan);
 381
 382 static const AVFilterPad avgblur_vulkan_inputs[] = {
 383     {
 384         .name         = "default",
 385         .type         = AVMEDIA_TYPE_VIDEO,
 386         .filter_frame = &avgblur_vulkan_filter_frame,
 387         .config_props = &ff_vk_filter_config_input,
 388     },
 389     { NULL }
 390 };
 391
 392 static const AVFilterPad avgblur_vulkan_outputs[] = {
 393     {
 394         .name = "default",
 395         .type = AVMEDIA_TYPE_VIDEO,
 396         .config_props = &ff_vk_filter_config_output,
 397     },
 398     { NULL }
 399 };
 400
 401 const AVFilter ff_vf_avgblur_vulkan = {
 402     .name           = "avgblur_vulkan",
 403     .description    = NULL_IF_CONFIG_SMALL("Apply avgblur mask to input video"),
 404     .priv_size      = sizeof(AvgBlurVulkanContext),
 405     .init           = &ff_vk_filter_init,
 406     .uninit         = &avgblur_vulkan_uninit,
 407     .query_formats  = &ff_vk_filter_query_formats,
 408     .inputs         = avgblur_vulkan_inputs,
 409     .outputs        = avgblur_vulkan_outputs,
 410     .priv_class     = &avgblur_vulkan_class,
 411     .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
 412 };