mirror of
https://github.com/jellyfin/jellyfin-ffmpeg.git
synced 2026-09-02 21:03:08 +03:00
366 lines
16 KiB
Diff
366 lines
16 KiB
Diff
Index: FFmpeg/libavfilter/vf_overlay_vulkan.c
|
|
===================================================================
|
|
--- FFmpeg.orig/libavfilter/vf_overlay_vulkan.c
|
|
+++ FFmpeg/libavfilter/vf_overlay_vulkan.c
|
|
@@ -33,9 +33,11 @@ typedef struct OverlayVulkanContext {
|
|
|
|
int initialized;
|
|
FFVulkanPipeline pl;
|
|
+ FFVulkanPipeline pl_pass;
|
|
FFVkExecPool e;
|
|
FFVkQueueFamilyCtx qf;
|
|
FFVkSPIRVShader shd;
|
|
+ FFVkSPIRVShader shd_pass;
|
|
VkSampler sampler;
|
|
|
|
/* Push constants / options */
|
|
@@ -48,6 +50,10 @@ typedef struct OverlayVulkanContext {
|
|
int overlay_y;
|
|
int overlay_w;
|
|
int overlay_h;
|
|
+
|
|
+ int opt_repeatlast;
|
|
+ int opt_shortest;
|
|
+ int opt_eof_action;
|
|
} OverlayVulkanContext;
|
|
|
|
static const char overlay_noalpha[] = {
|
|
@@ -84,15 +90,16 @@ static const char overlay_alpha[] = {
|
|
static av_cold int init_filter(AVFilterContext *ctx)
|
|
{
|
|
int err;
|
|
- uint8_t *spv_data;
|
|
+ uint8_t *spv_data, *spv_data_pass;
|
|
size_t spv_len;
|
|
void *spv_opaque = NULL;
|
|
+ void *spv_opaque_pass = NULL;
|
|
OverlayVulkanContext *s = ctx->priv;
|
|
FFVulkanContext *vkctx = &s->vkctx;
|
|
const int planes = av_pix_fmt_count_planes(s->vkctx.output_format);
|
|
const int ialpha = av_pix_fmt_desc_get(s->vkctx.input_format)->flags & AV_PIX_FMT_FLAG_ALPHA;
|
|
const AVPixFmtDescriptor *pix_desc = av_pix_fmt_desc_get(s->vkctx.output_format);
|
|
- FFVkSPIRVShader *shd = &s->shd;
|
|
+ FFVkSPIRVShader *shd;
|
|
FFVkSPIRVCompiler *spv;
|
|
FFVulkanDescriptorSetBinding *desc;
|
|
|
|
@@ -105,90 +112,143 @@ static av_cold int init_filter(AVFilterC
|
|
ff_vk_qf_init(vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT);
|
|
RET(ff_vk_exec_pool_init(vkctx, &s->qf, &s->e, s->qf.nb_queues*4, 0, 0, 0, NULL));
|
|
RET(ff_vk_init_sampler(vkctx, &s->sampler, 1, VK_FILTER_NEAREST));
|
|
+
|
|
+ /* overlay_compute */
|
|
RET(ff_vk_shader_init(&s->pl, &s->shd, "overlay_compute",
|
|
VK_SHADER_STAGE_COMPUTE_BIT, 0));
|
|
+ {
|
|
+ shd = &s->shd;
|
|
+ ff_vk_shader_set_compute_sizes(shd, 32, 32, 1);
|
|
+
|
|
+ GLSLC(0, layout(push_constant, std430) uniform pushConstants { );
|
|
+ GLSLC(1, ivec2 o_offset[3]; );
|
|
+ GLSLC(1, ivec2 o_size[3]; );
|
|
+ GLSLC(0, }; );
|
|
+ GLSLC(0, );
|
|
+
|
|
+ ff_vk_add_push_constant(&s->pl, 0, sizeof(s->opts),
|
|
+ VK_SHADER_STAGE_COMPUTE_BIT);
|
|
+
|
|
+ desc = (FFVulkanDescriptorSetBinding []) {
|
|
+ {
|
|
+ .name = "main_img",
|
|
+ .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
|
|
+ .dimensions = 2,
|
|
+ .elems = planes,
|
|
+ .stages = VK_SHADER_STAGE_COMPUTE_BIT,
|
|
+ .samplers = DUP_SAMPLER(s->sampler),
|
|
+ },
|
|
+ {
|
|
+ .name = "overlay_img",
|
|
+ .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
|
|
+ .dimensions = 2,
|
|
+ .elems = planes,
|
|
+ .stages = VK_SHADER_STAGE_COMPUTE_BIT,
|
|
+ .samplers = DUP_SAMPLER(s->sampler),
|
|
+ },
|
|
+ {
|
|
+ .name = "output_img",
|
|
+ .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
|
|
+ .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format),
|
|
+ .mem_quali = "writeonly",
|
|
+ .dimensions = 2,
|
|
+ .elems = planes,
|
|
+ .stages = VK_SHADER_STAGE_COMPUTE_BIT,
|
|
+ },
|
|
+ };
|
|
+
|
|
+ RET(ff_vk_pipeline_descriptor_set_add(vkctx, &s->pl, shd, desc, 3, 0, 0));
|
|
+
|
|
+ GLSLD( overlay_noalpha );
|
|
+ GLSLD( overlay_alpha );
|
|
+ GLSLC(0, void main() );
|
|
+ GLSLC(0, { );
|
|
+ GLSLC(1, ivec2 pos = ivec2(gl_GlobalInvocationID.xy); );
|
|
+ GLSLF(1, int planes = %i; ,planes);
|
|
+ GLSLC(1, for (int i = 0; i < planes; i++) { );
|
|
+ if (ialpha)
|
|
+ GLSLC(2, overlay_alpha_opaque(i, pos); );
|
|
+ else
|
|
+ GLSLC(2, overlay_noalpha(i, pos); );
|
|
+ GLSLC(1, } );
|
|
+ GLSLC(0, } );
|
|
+
|
|
+ RET(spv->compile_shader(spv, ctx, shd, &spv_data, &spv_len, "main",
|
|
+ &spv_opaque));
|
|
+ RET(ff_vk_shader_create(vkctx, shd, spv_data, spv_len, "main"));
|
|
+
|
|
+ RET(ff_vk_init_compute_pipeline(vkctx, &s->pl, shd));
|
|
+ RET(ff_vk_exec_pipeline_register(vkctx, &s->e, &s->pl));
|
|
+
|
|
+ s->opts.o_offset[0] = s->overlay_x;
|
|
+ s->opts.o_offset[1] = s->overlay_y;
|
|
+ s->opts.o_offset[2] = s->opts.o_offset[0] >> pix_desc->log2_chroma_w;
|
|
+ s->opts.o_offset[3] = s->opts.o_offset[1] >> pix_desc->log2_chroma_h;
|
|
+ s->opts.o_offset[4] = s->opts.o_offset[0] >> pix_desc->log2_chroma_w;
|
|
+ s->opts.o_offset[5] = s->opts.o_offset[1] >> pix_desc->log2_chroma_h;
|
|
+
|
|
+ s->opts.o_size[0] = s->overlay_w;
|
|
+ s->opts.o_size[1] = s->overlay_h;
|
|
+ s->opts.o_size[2] = s->opts.o_size[0] >> pix_desc->log2_chroma_w;
|
|
+ s->opts.o_size[3] = s->opts.o_size[1] >> pix_desc->log2_chroma_h;
|
|
+ s->opts.o_size[4] = s->opts.o_size[0] >> pix_desc->log2_chroma_w;
|
|
+ s->opts.o_size[5] = s->opts.o_size[1] >> pix_desc->log2_chroma_h;
|
|
+ }
|
|
|
|
- ff_vk_shader_set_compute_sizes(&s->shd, 32, 32, 1);
|
|
+ /* overlay_compute_pass */
|
|
+ RET(ff_vk_shader_init(&s->pl_pass, &s->shd_pass, "overlay_compute_pass",
|
|
+ VK_SHADER_STAGE_COMPUTE_BIT, 0));
|
|
+ {
|
|
+ shd = &s->shd_pass;
|
|
+ ff_vk_shader_set_compute_sizes(shd, 32, 32, 1);
|
|
+
|
|
+ desc = (FFVulkanDescriptorSetBinding []) {
|
|
+ {
|
|
+ .name = "main_img",
|
|
+ .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
|
|
+ .dimensions = 2,
|
|
+ .elems = planes,
|
|
+ .stages = VK_SHADER_STAGE_COMPUTE_BIT,
|
|
+ .samplers = DUP_SAMPLER(s->sampler),
|
|
+ },
|
|
+ {
|
|
+ .name = "output_img",
|
|
+ .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
|
|
+ .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format),
|
|
+ .mem_quali = "writeonly",
|
|
+ .dimensions = 2,
|
|
+ .elems = planes,
|
|
+ .stages = VK_SHADER_STAGE_COMPUTE_BIT,
|
|
+ },
|
|
+ };
|
|
+
|
|
+ RET(ff_vk_pipeline_descriptor_set_add(vkctx, &s->pl_pass, shd, desc, 2, 0, 0));
|
|
+
|
|
+ GLSLC(0, void main() );
|
|
+ GLSLC(0, { );
|
|
+ GLSLC(1, ivec2 pos = ivec2(gl_GlobalInvocationID.xy); );
|
|
+ GLSLF(1, int planes = %i; ,planes);
|
|
+ GLSLC(1, for (int i = 0; i < planes; i++) { );
|
|
+ GLSLC(2, vec4 res = texture(main_img[i], pos); );
|
|
+ GLSLC(2, imageStore(output_img[i], pos, res); );
|
|
+ GLSLC(1, } );
|
|
+ GLSLC(0, } );
|
|
+
|
|
+ RET(spv->compile_shader(spv, ctx, shd, &spv_data_pass, &spv_len, "main",
|
|
+ &spv_opaque));
|
|
+ RET(ff_vk_shader_create(vkctx, shd, spv_data_pass, spv_len, "main"));
|
|
|
|
- GLSLC(0, layout(push_constant, std430) uniform pushConstants { );
|
|
- GLSLC(1, ivec2 o_offset[3]; );
|
|
- GLSLC(1, ivec2 o_size[3]; );
|
|
- GLSLC(0, }; );
|
|
- GLSLC(0, );
|
|
-
|
|
- ff_vk_add_push_constant(&s->pl, 0, sizeof(s->opts),
|
|
- VK_SHADER_STAGE_COMPUTE_BIT);
|
|
-
|
|
- desc = (FFVulkanDescriptorSetBinding []) {
|
|
- {
|
|
- .name = "main_img",
|
|
- .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
|
|
- .dimensions = 2,
|
|
- .elems = planes,
|
|
- .stages = VK_SHADER_STAGE_COMPUTE_BIT,
|
|
- .samplers = DUP_SAMPLER(s->sampler),
|
|
- },
|
|
- {
|
|
- .name = "overlay_img",
|
|
- .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
|
|
- .dimensions = 2,
|
|
- .elems = planes,
|
|
- .stages = VK_SHADER_STAGE_COMPUTE_BIT,
|
|
- .samplers = DUP_SAMPLER(s->sampler),
|
|
- },
|
|
- {
|
|
- .name = "output_img",
|
|
- .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
|
|
- .mem_layout = ff_vk_shader_rep_fmt(s->vkctx.output_format),
|
|
- .mem_quali = "writeonly",
|
|
- .dimensions = 2,
|
|
- .elems = planes,
|
|
- .stages = VK_SHADER_STAGE_COMPUTE_BIT,
|
|
- },
|
|
- };
|
|
-
|
|
- RET(ff_vk_pipeline_descriptor_set_add(vkctx, &s->pl, shd, desc, 3, 0, 0));
|
|
-
|
|
- GLSLD( overlay_noalpha );
|
|
- GLSLD( overlay_alpha );
|
|
- GLSLC(0, void main() );
|
|
- GLSLC(0, { );
|
|
- GLSLC(1, ivec2 pos = ivec2(gl_GlobalInvocationID.xy); );
|
|
- GLSLF(1, int planes = %i; ,planes);
|
|
- GLSLC(1, for (int i = 0; i < planes; i++) { );
|
|
- if (ialpha)
|
|
- GLSLC(2, overlay_alpha_opaque(i, pos); );
|
|
- else
|
|
- GLSLC(2, overlay_noalpha(i, pos); );
|
|
- GLSLC(1, } );
|
|
- GLSLC(0, } );
|
|
-
|
|
- RET(spv->compile_shader(spv, ctx, shd, &spv_data, &spv_len, "main",
|
|
- &spv_opaque));
|
|
- RET(ff_vk_shader_create(vkctx, shd, spv_data, spv_len, "main"));
|
|
-
|
|
- RET(ff_vk_init_compute_pipeline(vkctx, &s->pl, shd));
|
|
- RET(ff_vk_exec_pipeline_register(vkctx, &s->e, &s->pl));
|
|
-
|
|
- s->opts.o_offset[0] = s->overlay_x;
|
|
- s->opts.o_offset[1] = s->overlay_y;
|
|
- s->opts.o_offset[2] = s->opts.o_offset[0] >> pix_desc->log2_chroma_w;
|
|
- s->opts.o_offset[3] = s->opts.o_offset[1] >> pix_desc->log2_chroma_h;
|
|
- s->opts.o_offset[4] = s->opts.o_offset[0] >> pix_desc->log2_chroma_w;
|
|
- s->opts.o_offset[5] = s->opts.o_offset[1] >> pix_desc->log2_chroma_h;
|
|
-
|
|
- s->opts.o_size[0] = s->overlay_w;
|
|
- s->opts.o_size[1] = s->overlay_h;
|
|
- s->opts.o_size[2] = s->opts.o_size[0] >> pix_desc->log2_chroma_w;
|
|
- s->opts.o_size[3] = s->opts.o_size[1] >> pix_desc->log2_chroma_h;
|
|
- s->opts.o_size[4] = s->opts.o_size[0] >> pix_desc->log2_chroma_w;
|
|
- s->opts.o_size[5] = s->opts.o_size[1] >> pix_desc->log2_chroma_h;
|
|
+ RET(ff_vk_init_compute_pipeline(vkctx, &s->pl_pass, shd));
|
|
+ RET(ff_vk_exec_pipeline_register(vkctx, &s->e, &s->pl_pass));
|
|
+ }
|
|
|
|
s->initialized = 1;
|
|
|
|
fail:
|
|
if (spv_opaque)
|
|
spv->free_shader(spv, &spv_opaque);
|
|
+ if (spv_opaque_pass)
|
|
+ spv->free_shader(spv, &spv_opaque_pass);
|
|
if (spv)
|
|
spv->uninit(&spv);
|
|
|
|
@@ -210,22 +270,11 @@ static int overlay_vulkan_blend(FFFrameS
|
|
if (err < 0)
|
|
goto fail;
|
|
|
|
- if (!input_main || !input_overlay)
|
|
- return 0;
|
|
-
|
|
- if (!s->initialized) {
|
|
- AVHWFramesContext *main_fc = (AVHWFramesContext*)input_main->hw_frames_ctx->data;
|
|
- AVHWFramesContext *overlay_fc = (AVHWFramesContext*)input_overlay->hw_frames_ctx->data;
|
|
- if (main_fc->sw_format != overlay_fc->sw_format) {
|
|
- av_log(ctx, AV_LOG_ERROR, "Mismatching sw formats!\n");
|
|
- return AVERROR(EINVAL);
|
|
- }
|
|
-
|
|
- s->overlay_w = input_overlay->width;
|
|
- s->overlay_h = input_overlay->height;
|
|
+ if (!input_main)
|
|
+ return AVERROR_BUG;
|
|
|
|
+ if (!s->initialized)
|
|
RET(init_filter(ctx));
|
|
- }
|
|
|
|
out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
|
|
if (!out) {
|
|
@@ -233,9 +282,13 @@ static int overlay_vulkan_blend(FFFrameS
|
|
goto fail;
|
|
}
|
|
|
|
- RET(ff_vk_filter_process_Nin(&s->vkctx, &s->e, &s->pl,
|
|
- out, (AVFrame *[]){ input_main, input_overlay }, 2,
|
|
- s->sampler, &s->opts, sizeof(s->opts)));
|
|
+ if (input_overlay)
|
|
+ RET(ff_vk_filter_process_Nin(&s->vkctx, &s->e, &s->pl,
|
|
+ out, (AVFrame *[]){ input_main, input_overlay }, 2,
|
|
+ s->sampler, &s->opts, sizeof(s->opts)));
|
|
+ else /* passthrough */
|
|
+ RET(ff_vk_filter_process_simple(&s->vkctx, &s->e, &s->pl_pass, out, input_main,
|
|
+ s->sampler, NULL, 0));
|
|
|
|
err = av_frame_copy_props(out, input_main);
|
|
if (err < 0)
|
|
@@ -253,6 +306,20 @@ static int overlay_vulkan_config_output(
|
|
int err;
|
|
AVFilterContext *avctx = outlink->src;
|
|
OverlayVulkanContext *s = avctx->priv;
|
|
+ AVFilterLink *inlink = avctx->inputs[0];
|
|
+ AVFilterLink *inlink_overlay = avctx->inputs[1];
|
|
+ FilterLink *inl = ff_filter_link(inlink);
|
|
+ FilterLink *inl_overlay = ff_filter_link(inlink_overlay);
|
|
+ AVHWFramesContext *main_fc = (AVHWFramesContext*)inl->hw_frames_ctx->data;
|
|
+ AVHWFramesContext *overlay_fc = (AVHWFramesContext*)inl_overlay->hw_frames_ctx->data;
|
|
+
|
|
+ if (main_fc->sw_format != overlay_fc->sw_format) {
|
|
+ av_log(avctx, AV_LOG_ERROR, "Mismatching sw formats!\n");
|
|
+ return AVERROR(EINVAL);
|
|
+ }
|
|
+
|
|
+ s->overlay_w = inlink_overlay->w;
|
|
+ s->overlay_h = inlink_overlay->h;
|
|
|
|
err = ff_vk_filter_config_output(outlink);
|
|
if (err < 0)
|
|
@@ -262,6 +329,11 @@ static int overlay_vulkan_config_output(
|
|
if (err < 0)
|
|
return err;
|
|
|
|
+ s->fs.opt_repeatlast = s->opt_repeatlast;
|
|
+ s->fs.opt_shortest = s->opt_shortest;
|
|
+ s->fs.opt_eof_action = s->opt_eof_action;
|
|
+ s->fs.time_base = outlink->time_base = inlink->time_base;
|
|
+
|
|
return ff_framesync_configure(&s->fs);
|
|
}
|
|
|
|
@@ -289,7 +361,9 @@ static void overlay_vulkan_uninit(AVFilt
|
|
|
|
ff_vk_exec_pool_free(vkctx, &s->e);
|
|
ff_vk_pipeline_free(vkctx, &s->pl);
|
|
+ ff_vk_pipeline_free(vkctx, &s->pl_pass);
|
|
ff_vk_shader_free(vkctx, &s->shd);
|
|
+ ff_vk_shader_free(vkctx, &s->shd_pass);
|
|
|
|
if (s->sampler)
|
|
vk->DestroySampler(vkctx->hwctx->act_dev, s->sampler,
|
|
@@ -306,6 +380,14 @@ static void overlay_vulkan_uninit(AVFilt
|
|
static const AVOption overlay_vulkan_options[] = {
|
|
{ "x", "Set horizontal offset", OFFSET(overlay_x), AV_OPT_TYPE_INT, {.i64 = 0}, 0, INT_MAX, .flags = FLAGS },
|
|
{ "y", "Set vertical offset", OFFSET(overlay_y), AV_OPT_TYPE_INT, {.i64 = 0}, 0, INT_MAX, .flags = FLAGS },
|
|
+ { "eof_action", "Action to take when encountering EOF from secondary input ",
|
|
+ OFFSET(opt_eof_action), AV_OPT_TYPE_INT, { .i64 = EOF_ACTION_REPEAT },
|
|
+ EOF_ACTION_REPEAT, EOF_ACTION_PASS, .flags = FLAGS, .unit = "eof_action" },
|
|
+ { "repeat", "Repeat the previous frame.", 0, AV_OPT_TYPE_CONST, { .i64 = EOF_ACTION_REPEAT }, .flags = FLAGS, .unit = "eof_action" },
|
|
+ { "endall", "End both streams.", 0, AV_OPT_TYPE_CONST, { .i64 = EOF_ACTION_ENDALL }, .flags = FLAGS, .unit = "eof_action" },
|
|
+ { "pass", "Pass through the main input.", 0, AV_OPT_TYPE_CONST, { .i64 = EOF_ACTION_PASS }, .flags = FLAGS, .unit = "eof_action" },
|
|
+ { "shortest", "force termination when the shortest input terminates", OFFSET(opt_shortest), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, FLAGS },
|
|
+ { "repeatlast", "repeat overlay of the last overlay frame", OFFSET(opt_repeatlast), AV_OPT_TYPE_BOOL, { .i64 = 1 }, 0, 1, FLAGS },
|
|
{ NULL },
|
|
};
|
|
|