git.sesse.net Git - ffmpeg/blob - libavfilter/vf_chromaber_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 CGROUPS (int [3]){ 32, 32, 1 }
  25
  26 typedef struct ChromaticAberrationVulkanContext {
  27     VulkanFilterContext vkctx;
  28
  29     int initialized;
  30     FFVkExecContext *exec;
  31     VulkanPipeline *pl;
  32
  33     /* Shader updators, must be in the main filter struct */
  34     VkDescriptorImageInfo input_images[3];
  35     VkDescriptorImageInfo output_images[3];
  36
  37     /* Push constants / options */
  38     struct {
  39         float dist[2];
  40     } opts;
  41 } ChromaticAberrationVulkanContext;
  42
  43 static const char distort_chroma_kernel[] = {
  44     C(0, void distort_rgb(ivec2 size, ivec2 pos)                               )
  45     C(0, {                                                                     )
  46     C(1,     const vec2 p = ((vec2(pos)/vec2(size)) - 0.5f)*2.0f;              )
  47     C(1,     const vec2 o = p * (dist - 1.0f);                                 )
  48     C(0,                                                                       )
  49     C(1,     vec4 res;                                                         )
  50     C(1,     res.r = texture(input_img[0], ((p - o)/2.0f) + 0.5f).r;           )
  51     C(1,     res.g = texture(input_img[0], ((p    )/2.0f) + 0.5f).g;           )
  52     C(1,     res.b = texture(input_img[0], ((p + o)/2.0f) + 0.5f).b;           )
  53     C(1,     res.a = texture(input_img[0], ((p    )/2.0f) + 0.5f).a;           )
  54     C(1,     imageStore(output_img[0], pos, res);                              )
  55     C(0, }                                                                     )
  56     C(0,                                                                       )
  57     C(0, void distort_chroma(int idx, ivec2 size, ivec2 pos)                   )
  58     C(0, {                                                                     )
  59     C(1,     vec2 p = ((vec2(pos)/vec2(size)) - 0.5f)*2.0f;                    )
  60     C(1,     float d = sqrt(p.x*p.x + p.y*p.y);                                )
  61     C(1,     p *= d / (d*     dist);                                           )
  62     C(1,     vec4 res = texture(input_img[idx], (p/2.0f) + 0.5f);              )
  63     C(1,     imageStore(output_img[idx], pos, res);                            )
  64     C(0, }                                                                     )
  65 };
  66
  67 static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
  68 {
  69     int err;
  70     ChromaticAberrationVulkanContext *s = ctx->priv;
  71
  72     /* Create a sampler */
  73     VkSampler *sampler = ff_vk_init_sampler(ctx, 0, VK_FILTER_LINEAR);
  74     if (!sampler)
  75         return AVERROR_EXTERNAL;
  76
  77     s->vkctx.queue_family_idx = s->vkctx.hwctx->queue_family_comp_index;
  78     s->vkctx.queue_count = GET_QUEUE_COUNT(s->vkctx.hwctx, 0, 1, 0);
  79     s->vkctx.cur_queue_idx = av_get_random_seed() % s->vkctx.queue_count;
  80
  81     s->pl = ff_vk_create_pipeline(ctx);
  82     if (!s->pl)
  83         return AVERROR(ENOMEM);
  84
  85     /* Normalize options */
  86     s->opts.dist[0] = (s->opts.dist[0] / 100.0f) + 1.0f;
  87     s->opts.dist[1] = (s->opts.dist[1] / 100.0f) + 1.0f;
  88
  89     { /* Create the shader */
  90         const int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
  91         VulkanDescriptorSetBinding desc_i[2] = {
  92             {
  93                 .name       = "input_img",
  94                 .type       = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
  95                 .dimensions = 2,
  96                 .elems      = planes,
  97                 .stages     = VK_SHADER_STAGE_COMPUTE_BIT,
  98                 .updater    = s->input_images,
  99                 .samplers   = DUP_SAMPLER_ARRAY4(*sampler),
 100             },
 101             {
 102                 .name       = "output_img",
 103                 .type       = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
 104                 .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format),
 105                 .mem_quali  = "writeonly",
 106                 .dimensions = 2,
 107                 .elems      = planes,
 108                 .stages     = VK_SHADER_STAGE_COMPUTE_BIT,
 109                 .updater    = s->output_images,
 110             },
 111         };
 112
 113         SPIRVShader *shd = ff_vk_init_shader(ctx, s->pl, "chromaber_compute",
 114                                              VK_SHADER_STAGE_COMPUTE_BIT);
 115         if (!shd)
 116             return AVERROR(ENOMEM);
 117
 118         ff_vk_set_compute_shader_sizes(ctx, shd, CGROUPS);
 119
 120         GLSLC(0, layout(push_constant, std430) uniform pushConstants {        );
 121         GLSLC(1,    vec2 dist;                                                );
 122         GLSLC(0, };                                                           );
 123         GLSLC(0,                                                              );
 124
 125         ff_vk_add_push_constant(ctx, s->pl, 0, sizeof(s->opts),
 126                                 VK_SHADER_STAGE_COMPUTE_BIT);
 127
 128         RET(ff_vk_add_descriptor_set(ctx, s->pl, shd, desc_i, 2, 0)); /* set 0 */
 129
 130         GLSLD(   distort_chroma_kernel                                        );
 131         GLSLC(0, void main()                                                  );
 132         GLSLC(0, {                                                            );
 133         GLSLC(1,     ivec2 pos = ivec2(gl_GlobalInvocationID.xy);             );
 134         if (planes == 1) {
 135             GLSLC(1, distort_rgb(imageSize(output_img[0]), pos);              );
 136         } else {
 137             GLSLC(1, ivec2 size = imageSize(output_img[0]);                   );
 138             GLSLC(1, vec2 npos = vec2(pos)/vec2(size);                        );
 139             GLSLC(1, vec4 res = texture(input_img[0], npos);                  );
 140             GLSLC(1, imageStore(output_img[0], pos, res);                     );
 141             for (int i = 1; i < planes; i++) {
 142                 GLSLC(0,                                                      );
 143                 GLSLF(1,  size = imageSize(output_img[%i]);                 ,i);
 144                 GLSLC(1,  if (IS_WITHIN(pos, size)) {                         );
 145                 GLSLF(2,      distort_chroma(%i, size, pos);                ,i);
 146                 GLSLC(1,  } else {                                            );
 147                 GLSLC(2,    npos = vec2(pos)/vec2(size);                      );
 148                 GLSLF(2,    res = texture(input_img[%i], npos);             ,i);
 149                 GLSLF(2,    imageStore(output_img[%i], pos, res);           ,i);
 150                 GLSLC(1, }                                                    );
 151             }
 152         }
 153         GLSLC(0, }                                                            );
 154
 155         RET(ff_vk_compile_shader(ctx, shd, "main"));
 156     }
 157
 158     RET(ff_vk_init_pipeline_layout(ctx, s->pl));
 159     RET(ff_vk_init_compute_pipeline(ctx, s->pl));
 160
 161     /* Execution context */
 162     RET(ff_vk_create_exec_ctx(ctx, &s->exec));
 163
 164     s->initialized = 1;
 165
 166     return 0;
 167
 168 fail:
 169     return err;
 170 }
 171
 172 static int process_frames(AVFilterContext *avctx, AVFrame *out_f, AVFrame *in_f)
 173 {
 174     int err = 0;
 175     VkCommandBuffer cmd_buf;
 176     ChromaticAberrationVulkanContext *s = avctx->priv;
 177     AVVkFrame *in = (AVVkFrame *)in_f->data[0];
 178     AVVkFrame *out = (AVVkFrame *)out_f->data[0];
 179     int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
 180
 181     /* Update descriptors and init the exec context */
 182     ff_vk_start_exec_recording(avctx, s->exec);
 183     cmd_buf = ff_vk_get_exec_buf(avctx, s->exec);
 184
 185     for (int i = 0; i < planes; i++) {
 186         RET(ff_vk_create_imageview(avctx, s->exec, &s->input_images[i].imageView,
 187                                    in->img[i],
 188                                    av_vkfmt_from_pixfmt(s->vkctx.input_format)[i],
 189                                    ff_comp_identity_map));
 190
 191         RET(ff_vk_create_imageview(avctx, s->exec, &s->output_images[i].imageView,
 192                                    out->img[i],
 193                                    av_vkfmt_from_pixfmt(s->vkctx.output_format)[i],
 194                                    ff_comp_identity_map));
 195
 196         s->input_images[i].imageLayout  = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
 197         s->output_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL;
 198     }
 199
 200     ff_vk_update_descriptor_set(avctx, s->pl, 0);
 201
 202     for (int i = 0; i < planes; i++) {
 203         VkImageMemoryBarrier bar[2] = {
 204             {
 205                 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
 206                 .srcAccessMask = 0,
 207                 .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
 208                 .oldLayout = in->layout[i],
 209                 .newLayout = s->input_images[i].imageLayout,
 210                 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 211                 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 212                 .image = in->img[i],
 213                 .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
 214                 .subresourceRange.levelCount = 1,
 215                 .subresourceRange.layerCount = 1,
 216             },
 217             {
 218                 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
 219                 .srcAccessMask = 0,
 220                 .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
 221                 .oldLayout = out->layout[i],
 222                 .newLayout = s->output_images[i].imageLayout,
 223                 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 224                 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
 225                 .image = out->img[i],
 226                 .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
 227                 .subresourceRange.levelCount = 1,
 228                 .subresourceRange.layerCount = 1,
 229             },
 230         };
 231
 232         vkCmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
 233                              VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0,
 234                              0, NULL, 0, NULL, FF_ARRAY_ELEMS(bar), bar);
 235
 236         in->layout[i]  = bar[0].newLayout;
 237         in->access[i]  = bar[0].dstAccessMask;
 238
 239         out->layout[i] = bar[1].newLayout;
 240         out->access[i] = bar[1].dstAccessMask;
 241     }
 242
 243     ff_vk_bind_pipeline_exec(avctx, s->exec, s->pl);
 244
 245     ff_vk_update_push_exec(avctx, s->exec, VK_SHADER_STAGE_COMPUTE_BIT,
 246                            0, sizeof(s->opts), &s->opts);
 247
 248     vkCmdDispatch(cmd_buf,
 249                   FFALIGN(s->vkctx.output_width,  CGROUPS[0])/CGROUPS[0],
 250                   FFALIGN(s->vkctx.output_height, CGROUPS[1])/CGROUPS[1], 1);
 251
 252     ff_vk_add_exec_dep(avctx, s->exec, in_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
 253     ff_vk_add_exec_dep(avctx, s->exec, out_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
 254
 255     err = ff_vk_submit_exec_queue(avctx, s->exec);
 256     if (err)
 257         return err;
 258
 259     return err;
 260
 261 fail:
 262     ff_vk_discard_exec_deps(avctx, s->exec);
 263     return err;
 264 }
 265
 266 static int chromaber_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
 267 {
 268     int err;
 269     AVFilterContext *ctx = link->dst;
 270     ChromaticAberrationVulkanContext *s = ctx->priv;
 271     AVFilterLink *outlink = ctx->outputs[0];
 272
 273     AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
 274     if (!out) {
 275         err = AVERROR(ENOMEM);
 276         goto fail;
 277     }
 278
 279     if (!s->initialized)
 280         RET(init_filter(ctx, in));
 281
 282     RET(process_frames(ctx, out, in));
 283
 284     err = av_frame_copy_props(out, in);
 285     if (err < 0)
 286         goto fail;
 287
 288     av_frame_free(&in);
 289
 290     return ff_filter_frame(outlink, out);
 291
 292 fail:
 293     av_frame_free(&in);
 294     av_frame_free(&out);
 295     return err;
 296 }
 297
 298 static void chromaber_vulkan_uninit(AVFilterContext *avctx)
 299 {
 300     ChromaticAberrationVulkanContext *s = avctx->priv;
 301
 302     ff_vk_filter_uninit(avctx);
 303
 304     s->initialized = 0;
 305 }
 306
 307 #define OFFSET(x) offsetof(ChromaticAberrationVulkanContext, x)
 308 #define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
 309 static const AVOption chromaber_vulkan_options[] = {
 310     { "dist_x", "Set horizontal distortion amount", OFFSET(opts.dist[0]), AV_OPT_TYPE_FLOAT, {.dbl = 0.0f}, -10.0f, 10.0f, .flags = FLAGS },
 311     { "dist_y", "Set vertical distortion amount",   OFFSET(opts.dist[1]), AV_OPT_TYPE_FLOAT, {.dbl = 0.0f}, -10.0f, 10.0f, .flags = FLAGS },
 312     { NULL },
 313 };
 314
 315 AVFILTER_DEFINE_CLASS(chromaber_vulkan);
 316
 317 static const AVFilterPad chromaber_vulkan_inputs[] = {
 318     {
 319         .name         = "default",
 320         .type         = AVMEDIA_TYPE_VIDEO,
 321         .filter_frame = &chromaber_vulkan_filter_frame,
 322         .config_props = &ff_vk_filter_config_input,
 323     },
 324     { NULL }
 325 };
 326
 327 static const AVFilterPad chromaber_vulkan_outputs[] = {
 328     {
 329         .name = "default",
 330         .type = AVMEDIA_TYPE_VIDEO,
 331         .config_props = &ff_vk_filter_config_output,
 332     },
 333     { NULL }
 334 };
 335
 336 const AVFilter ff_vf_chromaber_vulkan = {
 337     .name           = "chromaber_vulkan",
 338     .description    = NULL_IF_CONFIG_SMALL("Offset chroma of input video (chromatic aberration)"),
 339     .priv_size      = sizeof(ChromaticAberrationVulkanContext),
 340     .init           = &ff_vk_filter_init,
 341     .uninit         = &chromaber_vulkan_uninit,
 342     .query_formats  = &ff_vk_filter_query_formats,
 343     .inputs         = chromaber_vulkan_inputs,
 344     .outputs        = chromaber_vulkan_outputs,
 345     .priv_class     = &chromaber_vulkan_class,
 346     .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,
 347 };