2 * This file is part of FFmpeg.
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.
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.
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
19 #include "libavutil/random_seed.h"
20 #include "libavutil/opt.h"
21 #include "vulkan_filter.h"
22 #include "scale_eval.h"
24 #include "colorspace.h"
26 #define CGROUPS (int [3]){ 32, 32, 1 }
35 typedef struct ScaleVulkanContext {
36 FFVulkanContext vkctx;
38 FFVkQueueFamilyCtx qf;
39 FFVkExecContext *exec;
41 FFVkBuffer params_buf;
43 /* Shader updators, must be in the main filter struct */
44 VkDescriptorImageInfo input_images[3];
45 VkDescriptorImageInfo output_images[3];
46 VkDescriptorBufferInfo params_desc;
48 char *out_format_string;
52 enum ScalerFunc scaler;
53 enum AVColorRange out_range;
58 static const char scale_bilinear[] = {
59 C(0, vec4 scale_bilinear(int idx, ivec2 pos, vec2 crop_range, vec2 crop_off))
61 C(1, vec2 npos = (vec2(pos) + 0.5f) / imageSize(output_img[idx]); )
62 C(1, npos *= crop_range; /* Reduce the range */ )
63 C(1, npos += crop_off; /* Offset the start */ )
64 C(1, return texture(input_img[idx], npos); )
68 static const char rgb2yuv[] = {
69 C(0, vec4 rgb2yuv(vec4 src, int fullrange) )
71 C(1, src *= yuv_matrix; )
72 C(1, if (fullrange == 1) { )
73 C(2, src += vec4(0.0, 0.5, 0.5, 0.0); )
75 C(2, src *= vec4(219.0 / 255.0, 224.0 / 255.0, 224.0 / 255.0, 1.0); )
76 C(2, src += vec4(16.0 / 255.0, 128.0 / 255.0, 128.0 / 255.0, 0.0); )
82 static const char write_nv12[] = {
83 C(0, void write_nv12(vec4 src, ivec2 pos) )
85 C(1, imageStore(output_img[0], pos, vec4(src.r, 0.0, 0.0, 0.0)); )
86 C(1, pos /= ivec2(2); )
87 C(1, imageStore(output_img[1], pos, vec4(src.g, src.b, 0.0, 0.0)); )
91 static const char write_420[] = {
92 C(0, void write_420(vec4 src, ivec2 pos) )
94 C(1, imageStore(output_img[0], pos, vec4(src.r, 0.0, 0.0, 0.0)); )
95 C(1, pos /= ivec2(2); )
96 C(1, imageStore(output_img[1], pos, vec4(src.g, 0.0, 0.0, 0.0)); )
97 C(1, imageStore(output_img[2], pos, vec4(src.b, 0.0, 0.0, 0.0)); )
101 static const char write_444[] = {
102 C(0, void write_444(vec4 src, ivec2 pos) )
104 C(1, imageStore(output_img[0], pos, vec4(src.r, 0.0, 0.0, 0.0)); )
105 C(1, imageStore(output_img[1], pos, vec4(src.g, 0.0, 0.0, 0.0)); )
106 C(1, imageStore(output_img[2], pos, vec4(src.b, 0.0, 0.0, 0.0)); )
110 static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
113 FFVkSampler *sampler;
114 VkFilter sampler_mode;
115 ScaleVulkanContext *s = ctx->priv;
116 FFVulkanContext *vkctx = &s->vkctx;
118 int crop_x = in->crop_left;
119 int crop_y = in->crop_top;
120 int crop_w = in->width - (in->crop_left + in->crop_right);
121 int crop_h = in->height - (in->crop_top + in->crop_bottom);
122 int in_planes = av_pix_fmt_count_planes(s->vkctx.input_format);
124 ff_vk_qf_init(vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT, 0);
128 sampler_mode = VK_FILTER_NEAREST;
131 sampler_mode = VK_FILTER_LINEAR;
135 /* Create a sampler */
136 sampler = ff_vk_init_sampler(vkctx, 0, sampler_mode);
138 return AVERROR_EXTERNAL;
140 s->pl = ff_vk_create_pipeline(vkctx, &s->qf);
142 return AVERROR(ENOMEM);
144 { /* Create the shader */
145 FFVulkanDescriptorSetBinding desc_i[2] = {
148 .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
151 .stages = VK_SHADER_STAGE_COMPUTE_BIT,
152 .updater = s->input_images,
156 .name = "output_img",
157 .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
158 .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format),
159 .mem_quali = "writeonly",
161 .elems = av_pix_fmt_count_planes(s->vkctx.output_format),
162 .stages = VK_SHADER_STAGE_COMPUTE_BIT,
163 .updater = s->output_images,
167 FFVulkanDescriptorSetBinding desc_b = {
169 .type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
170 .mem_quali = "readonly",
171 .mem_layout = "std430",
172 .stages = VK_SHADER_STAGE_COMPUTE_BIT,
173 .updater = &s->params_desc,
174 .buf_content = "mat4 yuv_matrix;",
177 FFVkSPIRVShader *shd = ff_vk_init_shader(s->pl, "scale_compute",
178 VK_SHADER_STAGE_COMPUTE_BIT);
180 return AVERROR(ENOMEM);
182 ff_vk_set_compute_shader_sizes(shd, CGROUPS);
184 RET(ff_vk_add_descriptor_set(vkctx, s->pl, shd, desc_i, FF_ARRAY_ELEMS(desc_i), 0)); /* set 0 */
185 RET(ff_vk_add_descriptor_set(vkctx, s->pl, shd, &desc_b, 1, 0)); /* set 1 */
187 GLSLD( scale_bilinear );
189 if (s->vkctx.output_format != s->vkctx.input_format) {
193 switch (s->vkctx.output_format) {
194 case AV_PIX_FMT_NV12: GLSLD(write_nv12); break;
195 case AV_PIX_FMT_YUV420P: GLSLD( write_420); break;
196 case AV_PIX_FMT_YUV444P: GLSLD( write_444); break;
200 GLSLC(0, void main() );
202 GLSLC(1, ivec2 size; );
203 GLSLC(1, ivec2 pos = ivec2(gl_GlobalInvocationID.xy); );
204 GLSLF(1, vec2 in_d = vec2(%i, %i); ,in->width, in->height);
205 GLSLF(1, vec2 c_r = vec2(%i, %i) / in_d; ,crop_w, crop_h);
206 GLSLF(1, vec2 c_o = vec2(%i, %i) / in_d; ,crop_x,crop_y);
209 if (s->vkctx.output_format == s->vkctx.input_format) {
210 for (int i = 0; i < desc_i[1].elems; i++) {
211 GLSLF(1, size = imageSize(output_img[%i]); ,i);
212 GLSLC(1, if (IS_WITHIN(pos, size)) { );
216 GLSLF(2, vec4 res = scale_bilinear(%i, pos, c_r, c_o); ,i);
217 GLSLF(2, imageStore(output_img[%i], pos, res); ,i);
223 GLSLC(1, vec4 res = scale_bilinear(0, pos, c_r, c_o); );
224 GLSLF(1, res = rgb2yuv(res, %i); ,s->out_range == AVCOL_RANGE_JPEG);
225 switch (s->vkctx.output_format) {
226 case AV_PIX_FMT_NV12: GLSLC(1, write_nv12(res, pos); ); break;
227 case AV_PIX_FMT_YUV420P: GLSLC(1, write_420(res, pos); ); break;
228 case AV_PIX_FMT_YUV444P: GLSLC(1, write_444(res, pos); ); break;
229 default: return AVERROR(EINVAL);
235 RET(ff_vk_compile_shader(vkctx, shd, "main"));
238 RET(ff_vk_init_pipeline_layout(vkctx, s->pl));
239 RET(ff_vk_init_compute_pipeline(vkctx, s->pl));
241 if (s->vkctx.output_format != s->vkctx.input_format) {
242 const AVLumaCoefficients *lcoeffs;
243 double tmp_mat[3][3];
246 float yuv_matrix[4][4];
249 lcoeffs = av_csp_luma_coeffs_from_avcsp(in->colorspace);
251 av_log(ctx, AV_LOG_ERROR, "Unsupported colorspace\n");
252 return AVERROR(EINVAL);
255 RET(ff_vk_create_buf(vkctx, &s->params_buf,
257 VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
258 VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT));
260 RET(ff_vk_map_buffers(vkctx, &s->params_buf, (uint8_t **)&par, 1, 0));
262 ff_fill_rgb2yuv_table(lcoeffs, tmp_mat);
264 memset(par, 0, sizeof(*par));
266 for (int y = 0; y < 3; y++)
267 for (int x = 0; x < 3; x++)
268 par->yuv_matrix[x][y] = tmp_mat[x][y];
270 par->yuv_matrix[3][3] = 1.0;
272 RET(ff_vk_unmap_buffers(vkctx, &s->params_buf, 1, 1));
274 s->params_desc.buffer = s->params_buf.buf;
275 s->params_desc.range = VK_WHOLE_SIZE;
277 ff_vk_update_descriptor_set(vkctx, s->pl, 1);
280 /* Execution context */
281 RET(ff_vk_create_exec_ctx(vkctx, &s->exec, &s->qf));
291 static int process_frames(AVFilterContext *avctx, AVFrame *out_f, AVFrame *in_f)
294 VkCommandBuffer cmd_buf;
295 ScaleVulkanContext *s = avctx->priv;
296 FFVulkanContext *vkctx = &s->vkctx;
297 FFVulkanFunctions *vk = &vkctx->vkfn;
298 AVVkFrame *in = (AVVkFrame *)in_f->data[0];
299 AVVkFrame *out = (AVVkFrame *)out_f->data[0];
300 VkImageMemoryBarrier barriers[AV_NUM_DATA_POINTERS*2];
301 int barrier_count = 0;
302 const int planes = av_pix_fmt_count_planes(s->vkctx.input_format);
303 const VkFormat *input_formats = av_vkfmt_from_pixfmt(s->vkctx.input_format);
304 const VkFormat *output_formats = av_vkfmt_from_pixfmt(s->vkctx.output_format);
306 /* Update descriptors and init the exec context */
307 ff_vk_start_exec_recording(vkctx, s->exec);
308 cmd_buf = ff_vk_get_exec_buf(s->exec);
310 for (int i = 0; i < planes; i++) {
311 RET(ff_vk_create_imageview(vkctx, s->exec,
312 &s->input_images[i].imageView, in->img[i],
314 ff_comp_identity_map));
316 RET(ff_vk_create_imageview(vkctx, s->exec,
317 &s->output_images[i].imageView, out->img[i],
319 ff_comp_identity_map));
321 s->input_images[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
322 s->output_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL;
325 ff_vk_update_descriptor_set(vkctx, s->pl, 0);
327 for (int i = 0; i < planes; i++) {
328 VkImageMemoryBarrier bar = {
329 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
331 .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
332 .oldLayout = in->layout[i],
333 .newLayout = s->input_images[i].imageLayout,
334 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
335 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
337 .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
338 .subresourceRange.levelCount = 1,
339 .subresourceRange.layerCount = 1,
342 memcpy(&barriers[barrier_count++], &bar, sizeof(VkImageMemoryBarrier));
344 in->layout[i] = bar.newLayout;
345 in->access[i] = bar.dstAccessMask;
348 for (int i = 0; i < av_pix_fmt_count_planes(s->vkctx.output_format); i++) {
349 VkImageMemoryBarrier bar = {
350 .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
352 .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
353 .oldLayout = out->layout[i],
354 .newLayout = s->output_images[i].imageLayout,
355 .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
356 .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
357 .image = out->img[i],
358 .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
359 .subresourceRange.levelCount = 1,
360 .subresourceRange.layerCount = 1,
363 memcpy(&barriers[barrier_count++], &bar, sizeof(VkImageMemoryBarrier));
365 out->layout[i] = bar.newLayout;
366 out->access[i] = bar.dstAccessMask;
369 vk->CmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
370 VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0,
371 0, NULL, 0, NULL, barrier_count, barriers);
373 ff_vk_bind_pipeline_exec(vkctx, s->exec, s->pl);
375 vk->CmdDispatch(cmd_buf,
376 FFALIGN(vkctx->output_width, CGROUPS[0])/CGROUPS[0],
377 FFALIGN(vkctx->output_height, CGROUPS[1])/CGROUPS[1], 1);
379 ff_vk_add_exec_dep(vkctx, s->exec, in_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
380 ff_vk_add_exec_dep(vkctx, s->exec, out_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
382 err = ff_vk_submit_exec_queue(vkctx, s->exec);
386 ff_vk_qf_rotate(&s->qf);
391 ff_vk_discard_exec_deps(s->exec);
395 static int scale_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
398 AVFilterContext *ctx = link->dst;
399 ScaleVulkanContext *s = ctx->priv;
400 AVFilterLink *outlink = ctx->outputs[0];
402 AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
404 err = AVERROR(ENOMEM);
409 RET(init_filter(ctx, in));
411 RET(process_frames(ctx, out, in));
413 err = av_frame_copy_props(out, in);
417 if (s->out_range != AVCOL_RANGE_UNSPECIFIED)
418 out->color_range = s->out_range;
419 if (s->vkctx.output_format != s->vkctx.input_format)
420 out->chroma_location = AVCHROMA_LOC_TOPLEFT;
424 return ff_filter_frame(outlink, out);
432 static int scale_vulkan_config_output(AVFilterLink *outlink)
435 AVFilterContext *avctx = outlink->src;
436 ScaleVulkanContext *s = avctx->priv;
437 FFVulkanContext *vkctx = &s->vkctx;
438 AVFilterLink *inlink = outlink->src->inputs[0];
440 err = ff_scale_eval_dimensions(s, s->w_expr, s->h_expr, inlink, outlink,
441 &vkctx->output_width,
442 &vkctx->output_height);
446 if (s->out_format_string) {
447 s->vkctx.output_format = av_get_pix_fmt(s->out_format_string);
448 if (s->vkctx.output_format == AV_PIX_FMT_NONE) {
449 av_log(avctx, AV_LOG_ERROR, "Invalid output format.\n");
450 return AVERROR(EINVAL);
453 s->vkctx.output_format = s->vkctx.input_format;
456 if (s->vkctx.output_format != s->vkctx.input_format) {
457 if (!ff_vk_mt_is_np_rgb(s->vkctx.input_format)) {
458 av_log(avctx, AV_LOG_ERROR, "Unsupported input format for conversion\n");
459 return AVERROR(EINVAL);
461 if (s->vkctx.output_format != AV_PIX_FMT_NV12 &&
462 s->vkctx.output_format != AV_PIX_FMT_YUV420P &&
463 s->vkctx.output_format != AV_PIX_FMT_YUV444P) {
464 av_log(avctx, AV_LOG_ERROR, "Unsupported output format\n");
465 return AVERROR(EINVAL);
467 } else if (s->out_range != AVCOL_RANGE_UNSPECIFIED) {
468 av_log(avctx, AV_LOG_ERROR, "Cannot change range without converting format\n");
469 return AVERROR(EINVAL);
472 return ff_vk_filter_config_output(outlink);
475 static void scale_vulkan_uninit(AVFilterContext *avctx)
477 ScaleVulkanContext *s = avctx->priv;
479 ff_vk_free_buf(&s->vkctx, &s->params_buf);
480 ff_vk_uninit(&s->vkctx);
485 #define OFFSET(x) offsetof(ScaleVulkanContext, x)
486 #define FLAGS (AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM)
487 static const AVOption scale_vulkan_options[] = {
488 { "w", "Output video width", OFFSET(w_expr), AV_OPT_TYPE_STRING, {.str = "iw"}, .flags = FLAGS },
489 { "h", "Output video height", OFFSET(h_expr), AV_OPT_TYPE_STRING, {.str = "ih"}, .flags = FLAGS },
490 { "scaler", "Scaler function", OFFSET(scaler), AV_OPT_TYPE_INT, {.i64 = F_BILINEAR}, 0, F_NB, .flags = FLAGS, "scaler" },
491 { "bilinear", "Bilinear interpolation (fastest)", 0, AV_OPT_TYPE_CONST, {.i64 = F_BILINEAR}, 0, 0, .flags = FLAGS, "scaler" },
492 { "nearest", "Nearest (useful for pixel art)", 0, AV_OPT_TYPE_CONST, {.i64 = F_NEAREST}, 0, 0, .flags = FLAGS, "scaler" },
493 { "format", "Output video format (software format of hardware frames)", OFFSET(out_format_string), AV_OPT_TYPE_STRING, .flags = FLAGS },
494 { "out_range", "Output colour range (from 0 to 2) (default 0)", OFFSET(out_range), AV_OPT_TYPE_INT, {.i64 = AVCOL_RANGE_UNSPECIFIED}, AVCOL_RANGE_UNSPECIFIED, AVCOL_RANGE_JPEG, .flags = FLAGS, "range" },
495 { "full", "Full range", 0, AV_OPT_TYPE_CONST, { .i64 = AVCOL_RANGE_JPEG }, 0, 0, FLAGS, "range" },
496 { "limited", "Limited range", 0, AV_OPT_TYPE_CONST, { .i64 = AVCOL_RANGE_MPEG }, 0, 0, FLAGS, "range" },
497 { "jpeg", "Full range", 0, AV_OPT_TYPE_CONST, { .i64 = AVCOL_RANGE_JPEG }, 0, 0, FLAGS, "range" },
498 { "mpeg", "Limited range", 0, AV_OPT_TYPE_CONST, { .i64 = AVCOL_RANGE_MPEG }, 0, 0, FLAGS, "range" },
499 { "tv", "Limited range", 0, AV_OPT_TYPE_CONST, { .i64 = AVCOL_RANGE_MPEG }, 0, 0, FLAGS, "range" },
500 { "pc", "Full range", 0, AV_OPT_TYPE_CONST, { .i64 = AVCOL_RANGE_JPEG }, 0, 0, FLAGS, "range" },
504 AVFILTER_DEFINE_CLASS(scale_vulkan);
506 static const AVFilterPad scale_vulkan_inputs[] = {
509 .type = AVMEDIA_TYPE_VIDEO,
510 .filter_frame = &scale_vulkan_filter_frame,
511 .config_props = &ff_vk_filter_config_input,
515 static const AVFilterPad scale_vulkan_outputs[] = {
518 .type = AVMEDIA_TYPE_VIDEO,
519 .config_props = &scale_vulkan_config_output,
523 const AVFilter ff_vf_scale_vulkan = {
524 .name = "scale_vulkan",
525 .description = NULL_IF_CONFIG_SMALL("Scale Vulkan frames"),
526 .priv_size = sizeof(ScaleVulkanContext),
527 .init = &ff_vk_filter_init,
528 .uninit = &scale_vulkan_uninit,
529 FILTER_INPUTS(scale_vulkan_inputs),
530 FILTER_OUTPUTS(scale_vulkan_outputs),
531 FILTER_SINGLE_PIXFMT(AV_PIX_FMT_VULKAN),
532 .priv_class = &scale_vulkan_class,
533 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE,