From b14473b21aa057181ec85e0ea3bac3e5fa053875 Mon Sep 17 00:00:00 2001 From: Lynne Date: Fri, 17 Feb 2023 03:11:19 +0100 Subject: [PATCH 56/72] avgblur_vulkan: port for the rewrite --- libavfilter/vf_avgblur_vulkan.c | 339 ++++++++++---------------------- 1 file changed, 108 insertions(+), 231 deletions(-) diff --git a/libavfilter/vf_avgblur_vulkan.c b/libavfilter/vf_avgblur_vulkan.c index d118ce802c..17b2167951 100644 --- a/libavfilter/vf_avgblur_vulkan.c +++ b/libavfilter/vf_avgblur_vulkan.c @@ -1,4 +1,6 @@ /* + * Copyright (c) Lynne + * * This file is part of FFmpeg. * * FFmpeg is free software; you can redistribute it and/or @@ -19,23 +21,20 @@ #include "libavutil/random_seed.h" #include "libavutil/opt.h" #include "vulkan_filter.h" +#include "vulkan_spirv.h" #include "internal.h" -#define CGS 32 - typedef struct AvgBlurVulkanContext { FFVulkanContext vkctx; int initialized; + FFVkExecPool e; FFVkQueueFamilyCtx qf; - FFVkExecContext *exec; - FFVulkanPipeline *pl_hor; - FFVulkanPipeline *pl_ver; - - /* Shader updators, must be in the main filter struct */ - VkDescriptorImageInfo input_images[3]; - VkDescriptorImageInfo tmp_images[3]; - VkDescriptorImageInfo output_images[3]; + VkSampler sampler; + FFVulkanPipeline pl_hor; + FFVkSPIRVShader shd_hor; + FFVulkanPipeline pl_ver; + FFVkSPIRVShader shd_ver; int size_x; int size_y; @@ -71,18 +70,41 @@ static const char blur_kernel[] = { static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in) { int err; - FFVkSPIRVShader *shd; + uint8_t *spv_data; + size_t spv_len; + void *spv_opaque; AvgBlurVulkanContext *s = ctx->priv; FFVulkanContext *vkctx = &s->vkctx; const int planes = av_pix_fmt_count_planes(s->vkctx.output_format); + FFVkSPIRVShader *shd; + FFVkSPIRVCompiler *spv; + FFVulkanDescriptorSetBinding *desc; - FFVulkanDescriptorSetBinding desc_i[2] = { + spv = ff_vk_spirv_init(); + if (!spv) { + av_log(ctx, AV_LOG_ERROR, "Unable to initialize SPIR-V compiler!\n"); + return AVERROR_EXTERNAL; + } + + 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_LINEAR)); + RET(ff_vk_shader_init(&s->pl_hor, &s->shd_hor, "avgblur_hor_compute", + VK_SHADER_STAGE_COMPUTE_BIT)); + RET(ff_vk_shader_init(&s->pl_ver, &s->shd_ver, "avgblur_ver_compute", + VK_SHADER_STAGE_COMPUTE_BIT)); + shd = &s->shd_hor; + + ff_vk_shader_set_compute_sizes(shd, 32, 1, 1); + + desc = (FFVulkanDescriptorSetBinding []) { { .name = "input_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", @@ -95,238 +117,79 @@ static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in) }, }; - ff_vk_qf_init(vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT, 0); + RET(ff_vk_pipeline_descriptor_set_add(vkctx, &s->pl_hor, shd, desc, 2, 0, 0)); - desc_i[0].sampler = ff_vk_init_sampler(vkctx, 1, VK_FILTER_LINEAR); - if (!desc_i[0].sampler) - return AVERROR_EXTERNAL; - - { /* Create shader for the horizontal pass */ - desc_i[0].updater = s->input_images; - desc_i[1].updater = s->tmp_images; - - s->pl_hor = ff_vk_create_pipeline(vkctx, &s->qf); - if (!s->pl_hor) - return AVERROR(ENOMEM); - - shd = ff_vk_init_shader(s->pl_hor, "avgblur_compute_hor", - VK_SHADER_STAGE_COMPUTE_BIT); - if (!shd) - return AVERROR(ENOMEM); - - ff_vk_set_compute_shader_sizes(shd, (int [3]){ CGS, 1, 1 }); - - RET(ff_vk_add_descriptor_set(vkctx, s->pl_hor, shd, desc_i, FF_ARRAY_ELEMS(desc_i), 0)); - - GLSLF(0, #define FILTER_RADIUS (%i) ,s->size_x - 1); - GLSLC(0, #define INC(x) (ivec2(x, 0)) ); - GLSLC(0, #define DIR(var) (var.x) ); - GLSLD( blur_kernel ); - GLSLC(0, void main() ); - GLSLC(0, { ); - GLSLC(1, ivec2 size; ); - GLSLC(1, const ivec2 pos = ivec2(gl_GlobalInvocationID.xy); ); - for (int i = 0; i < planes; i++) { - GLSLC(0, ); - GLSLF(1, size = imageSize(output_img[%i]); ,i); - GLSLC(1, if (IS_WITHIN(pos, size)) { ); - if (s->planes & (1 << i)) { - GLSLF(2, distort(pos, %i); ,i); - } else { - GLSLF(2, vec4 res = texture(input_img[%i], pos); ,i); - GLSLF(2, imageStore(output_img[%i], pos, res); ,i); - } - GLSLC(1, } ); + GLSLF(0, #define FILTER_RADIUS (%i) ,s->size_x - 1); + GLSLC(0, #define INC(x) (ivec2(x, 0)) ); + GLSLC(0, #define DIR(var) (var.x) ); + GLSLD( blur_kernel ); + GLSLC(0, void main() ); + GLSLC(0, { ); + GLSLC(1, ivec2 size; ); + GLSLC(1, const ivec2 pos = ivec2(gl_GlobalInvocationID.xy); ); + for (int i = 0; i < planes; i++) { + GLSLC(0, ); + GLSLF(1, size = imageSize(output_img[%i]); ,i); + GLSLC(1, if (IS_WITHIN(pos, size)) { ); + if (s->planes & (1 << i)) { + GLSLF(2, distort(pos, %i); ,i); + } else { + GLSLF(2, vec4 res = texture(input_img[%i], pos); ,i); + GLSLF(2, imageStore(output_img[%i], pos, res); ,i); } - GLSLC(0, } ); - - RET(ff_vk_compile_shader(vkctx, shd, "main")); - - RET(ff_vk_init_pipeline_layout(vkctx, s->pl_hor)); - RET(ff_vk_init_compute_pipeline(vkctx, s->pl_hor)); + GLSLC(1, } ); } - - { /* Create shader for the vertical pass */ - desc_i[0].updater = s->tmp_images; - desc_i[1].updater = s->output_images; - - s->pl_ver = ff_vk_create_pipeline(vkctx, &s->qf); - if (!s->pl_ver) - return AVERROR(ENOMEM); - - shd = ff_vk_init_shader(s->pl_ver, "avgblur_compute_ver", - VK_SHADER_STAGE_COMPUTE_BIT); - if (!shd) - return AVERROR(ENOMEM); - - ff_vk_set_compute_shader_sizes(shd, (int [3]){ 1, CGS, 1 }); - - RET(ff_vk_add_descriptor_set(vkctx, s->pl_ver, shd, desc_i, FF_ARRAY_ELEMS(desc_i), 0)); - - GLSLF(0, #define FILTER_RADIUS (%i) ,s->size_y - 1); - GLSLC(0, #define INC(x) (ivec2(0, x)) ); - GLSLC(0, #define DIR(var) (var.y) ); - GLSLD( blur_kernel ); - GLSLC(0, void main() ); - GLSLC(0, { ); - GLSLC(1, ivec2 size; ); - GLSLC(1, const ivec2 pos = ivec2(gl_GlobalInvocationID.xy); ); - for (int i = 0; i < planes; i++) { - GLSLC(0, ); - GLSLF(1, size = imageSize(output_img[%i]); ,i); - GLSLC(1, if (IS_WITHIN(pos, size)) { ); - if (s->planes & (1 << i)) { - GLSLF(2, distort(pos, %i); ,i); - } else { - GLSLF(2, vec4 res = texture(input_img[%i], pos); ,i); - GLSLF(2, imageStore(output_img[%i], pos, res); ,i); - } - GLSLC(1, } ); + GLSLC(0, } ); + + shd = &s->shd_ver; + ff_vk_shader_set_compute_sizes(shd, 1, 32, 1); + RET(ff_vk_pipeline_descriptor_set_add(vkctx, &s->pl_ver, shd, desc, 2, 0, 0)); + + GLSLF(0, #define FILTER_RADIUS (%i) ,s->size_y - 1); + GLSLC(0, #define INC(x) (ivec2(0, x)) ); + GLSLC(0, #define DIR(var) (var.y) ); + GLSLD( blur_kernel ); + GLSLC(0, void main() ); + GLSLC(0, { ); + GLSLC(1, ivec2 size; ); + GLSLC(1, const ivec2 pos = ivec2(gl_GlobalInvocationID.xy); ); + for (int i = 0; i < planes; i++) { + GLSLC(0, ); + GLSLF(1, size = imageSize(output_img[%i]); ,i); + GLSLC(1, if (IS_WITHIN(pos, size)) { ); + if (s->planes & (1 << i)) { + GLSLF(2, distort(pos, %i); ,i); + } else { + GLSLF(2, vec4 res = texture(input_img[%i], pos); ,i); + GLSLF(2, imageStore(output_img[%i], pos, res); ,i); } - GLSLC(0, } ); - - RET(ff_vk_compile_shader(vkctx, shd, "main")); - - RET(ff_vk_init_pipeline_layout(vkctx, s->pl_ver)); - RET(ff_vk_init_compute_pipeline(vkctx, s->pl_ver)); + GLSLC(1, } ); } + GLSLC(0, } ); + + RET(spv->compile_shader(spv, ctx, &s->shd_hor, &spv_data, &spv_len, "main", + &spv_opaque)); + RET(ff_vk_shader_create(vkctx, &s->shd_hor, spv_data, spv_len, "main")); + RET(spv->compile_shader(spv, ctx, &s->shd_ver, &spv_data, &spv_len, "main", + &spv_opaque)); + RET(ff_vk_shader_create(vkctx, &s->shd_ver, spv_data, spv_len, "main")); - /* Execution context */ - RET(ff_vk_create_exec_ctx(vkctx, &s->exec, &s->qf)); + RET(ff_vk_init_compute_pipeline(vkctx, &s->pl_hor, &s->shd_hor)); + RET(ff_vk_init_compute_pipeline(vkctx, &s->pl_ver, &s->shd_ver)); + RET(ff_vk_exec_pipeline_register(vkctx, &s->e, &s->pl_hor)); + RET(ff_vk_exec_pipeline_register(vkctx, &s->e, &s->pl_ver)); s->initialized = 1; return 0; fail: - return err; -} - -static int process_frames(AVFilterContext *avctx, AVFrame *out_f, AVFrame *tmp_f, AVFrame *in_f) -{ - int err; - VkCommandBuffer cmd_buf; - AvgBlurVulkanContext *s = avctx->priv; - FFVulkanContext *vkctx = &s->vkctx; - FFVulkanFunctions *vk = &vkctx->vkfn; - AVVkFrame *in = (AVVkFrame *)in_f->data[0]; - AVVkFrame *tmp = (AVVkFrame *)tmp_f->data[0]; - AVVkFrame *out = (AVVkFrame *)out_f->data[0]; - - const VkFormat *input_formats = av_vkfmt_from_pixfmt(s->vkctx.input_format); - const VkFormat *output_formats = av_vkfmt_from_pixfmt(s->vkctx.output_format); - - int planes = av_pix_fmt_count_planes(s->vkctx.output_format); - - /* Update descriptors and init the exec context */ - ff_vk_start_exec_recording(vkctx, s->exec); - cmd_buf = ff_vk_get_exec_buf(s->exec); - - for (int i = 0; i < planes; i++) { - RET(ff_vk_create_imageview(vkctx, s->exec, - &s->input_images[i].imageView, in->img[i], - input_formats[i], - ff_comp_identity_map)); - - RET(ff_vk_create_imageview(vkctx, s->exec, - &s->tmp_images[i].imageView, tmp->img[i], - output_formats[i], - ff_comp_identity_map)); - - RET(ff_vk_create_imageview(vkctx, s->exec, - &s->output_images[i].imageView, out->img[i], - output_formats[i], - ff_comp_identity_map)); - - s->input_images[i].imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; - s->tmp_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL; - s->output_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL; - } - - ff_vk_update_descriptor_set(vkctx, s->pl_hor, 0); - ff_vk_update_descriptor_set(vkctx, s->pl_ver, 0); - - for (int i = 0; i < planes; i++) { - VkImageMemoryBarrier bar[] = { - { - .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, - .srcAccessMask = 0, - .dstAccessMask = VK_ACCESS_SHADER_READ_BIT, - .oldLayout = in->layout[i], - .newLayout = s->input_images[i].imageLayout, - .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, - .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, - .image = in->img[i], - .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, - .subresourceRange.levelCount = 1, - .subresourceRange.layerCount = 1, - }, - { - .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, - .srcAccessMask = 0, - .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_SHADER_READ_BIT, - .oldLayout = tmp->layout[i], - .newLayout = s->tmp_images[i].imageLayout, - .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, - .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, - .image = tmp->img[i], - .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, - .subresourceRange.levelCount = 1, - .subresourceRange.layerCount = 1, - }, - { - .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER, - .srcAccessMask = 0, - .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT, - .oldLayout = out->layout[i], - .newLayout = s->output_images[i].imageLayout, - .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, - .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, - .image = out->img[i], - .subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT, - .subresourceRange.levelCount = 1, - .subresourceRange.layerCount = 1, - }, - }; - - vk->CmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, - VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, - 0, NULL, 0, NULL, FF_ARRAY_ELEMS(bar), bar); - - in->layout[i] = bar[0].newLayout; - in->access[i] = bar[0].dstAccessMask; - - tmp->layout[i] = bar[1].newLayout; - tmp->access[i] = bar[1].dstAccessMask; - - out->layout[i] = bar[2].newLayout; - out->access[i] = bar[2].dstAccessMask; - } - - ff_vk_bind_pipeline_exec(vkctx, s->exec, s->pl_hor); - - vk->CmdDispatch(cmd_buf, FFALIGN(s->vkctx.output_width, CGS)/CGS, - s->vkctx.output_height, 1); - - ff_vk_bind_pipeline_exec(vkctx, s->exec, s->pl_ver); - - vk->CmdDispatch(cmd_buf, s->vkctx.output_width, - FFALIGN(s->vkctx.output_height, CGS)/CGS, 1); - - ff_vk_add_exec_dep(vkctx, s->exec, in_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT); - ff_vk_add_exec_dep(vkctx, s->exec, out_f, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT); - - err = ff_vk_submit_exec_queue(vkctx,s->exec); - if (err) - return err; - - ff_vk_qf_rotate(&s->qf); + if (spv_opaque) + spv->free_shader(spv, &spv_opaque); + if (spv) + spv->uninit(&spv); return err; - -fail: - ff_vk_discard_exec_deps(s->exec); - return err; } static int avgblur_vulkan_filter_frame(AVFilterLink *link, AVFrame *in) @@ -352,7 +215,9 @@ static int avgblur_vulkan_filter_frame(AVFilterLink *link, AVFrame *in) if (!s->initialized) RET(init_filter(ctx, in)); - RET(process_frames(ctx, out, tmp, in)); + RET(ff_vk_filter_process_2pass(&s->vkctx, &s->e, + (FFVulkanPipeline *[2]){ &s->pl_hor, &s->pl_ver }, + out, tmp, in, s->sampler, NULL, 0)); err = av_frame_copy_props(out, in); if (err < 0) @@ -373,6 +238,18 @@ fail: static void avgblur_vulkan_uninit(AVFilterContext *avctx) { AvgBlurVulkanContext *s = avctx->priv; + FFVulkanContext *vkctx = &s->vkctx; + FFVulkanFunctions *vk = &vkctx->vkfn; + + ff_vk_exec_pool_free(vkctx, &s->e); + ff_vk_pipeline_free(vkctx, &s->pl_hor); + ff_vk_pipeline_free(vkctx, &s->pl_ver); + ff_vk_shader_free(vkctx, &s->shd_hor); + ff_vk_shader_free(vkctx, &s->shd_ver); + + if (s->sampler) + vk->DestroySampler(vkctx->hwctx->act_dev, s->sampler, + vkctx->hwctx->alloc); ff_vk_uninit(&s->vkctx); -- 2.39.2