virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2025-08-10 06:10:52 +02:00
Code Issues Releases Activity

nlmeans_vulkan: switch to imageLoad()

Browse Source
This commit is contained in:
Lynne
2025-02-17 18:42:04 +01:00
parent 9c4a26d9b0
commit a535639620
1 changed files with 24 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

56
libavfilter/vf_nlmeans_vulkan.c
View File

@@ -37,7 +37,6 @@ typedef struct NLMeansVulkanContext {
int initialized; int initialized;
FFVkExecPool e; FFVkExecPool e;
AVVulkanDeviceQueueFamily *qf; AVVulkanDeviceQueueFamily *qf;
VkSampler sampler;
AVBufferPool *integral_buf_pool; AVBufferPool *integral_buf_pool;
AVBufferPool *ws_buf_pool; AVBufferPool *ws_buf_pool;
@@ -66,16 +65,16 @@ typedef struct NLMeansVulkanContext {
static void insert_first(FFVulkanShader *shd, int r, const char *off, int horiz, int plane, int comp) static void insert_first(FFVulkanShader *shd, int r, const char *off, int horiz, int plane, int comp)
{ {
GLSLF(4, s1 = texture(input_img[%i], pos + ivec2(%i + %s, %i + %s))[%i]; GLSLF(4, s1 = imageLoad(input_img[%i], pos + ivec2(%i + %s, %i + %s))[%i];
,plane, horiz ? r : 0, horiz ? off : "0", !horiz ? r : 0, !horiz ? off : "0", comp); ,plane, horiz ? r : 0, horiz ? off : "0", !horiz ? r : 0, !horiz ? off : "0", comp);
GLSLF(4, s2[0] = texture(input_img[%i], pos + offs[0] + ivec2(%i + %s, %i + %s))[%i]; GLSLF(4, s2[0] = imageLoad(input_img[%i], pos + offs[0] + ivec2(%i + %s, %i + %s))[%i];
,plane, horiz ? r : 0, horiz ? off : "0", !horiz ? r : 0, !horiz ? off : "0", comp); ,plane, horiz ? r : 0, horiz ? off : "0", !horiz ? r : 0, !horiz ? off : "0", comp);
GLSLF(4, s2[1] = texture(input_img[%i], pos + offs[1] + ivec2(%i + %s, %i + %s))[%i]; GLSLF(4, s2[1] = imageLoad(input_img[%i], pos + offs[1] + ivec2(%i + %s, %i + %s))[%i];
,plane, horiz ? r : 0, horiz ? off : "0", !horiz ? r : 0, !horiz ? off : "0", comp); ,plane, horiz ? r : 0, horiz ? off : "0", !horiz ? r : 0, !horiz ? off : "0", comp);
GLSLF(4, s2[2] = texture(input_img[%i], pos + offs[2] + ivec2(%i + %s, %i + %s))[%i]; GLSLF(4, s2[2] = imageLoad(input_img[%i], pos + offs[2] + ivec2(%i + %s, %i + %s))[%i];
,plane, horiz ? r : 0, horiz ? off : "0", !horiz ? r : 0, !horiz ? off : "0", comp); ,plane, horiz ? r : 0, horiz ? off : "0", !horiz ? r : 0, !horiz ? off : "0", comp);
GLSLF(4, s2[3] = texture(input_img[%i], pos + offs[3] + ivec2(%i + %s, %i + %s))[%i]; GLSLF(4, s2[3] = imageLoad(input_img[%i], pos + offs[3] + ivec2(%i + %s, %i + %s))[%i];
,plane, horiz ? r : 0, horiz ? off : "0", !horiz ? r : 0, !horiz ? off : "0", comp); ,plane, horiz ? r : 0, horiz ? off : "0", !horiz ? r : 0, !horiz ? off : "0", comp);
GLSLC(4, s2 = (s1 - s2) * (s1 - s2); ); GLSLC(4, s2 = (s1 - s2) * (s1 - s2); );
@@ -164,10 +163,10 @@ static void insert_weights_pass(FFVulkanShader *shd, int nb_rows, int vert,
GLSLC(0, ); GLSLC(0, );
GLSLC(3, lt = ((pos.x - p) < 0) || ((pos.y - p) < 0); ); GLSLC(3, lt = ((pos.x - p) < 0) || ((pos.y - p) < 0); );
GLSLC(0, ); GLSLC(0, );
GLSLF(3, src[0] = texture(input_img[%i], pos + offs[0])[%i]; ,plane, comp); GLSLF(3, src[0] = imageLoad(input_img[%i], pos + offs[0])[%i]; ,plane, comp);
GLSLF(3, src[1] = texture(input_img[%i], pos + offs[1])[%i]; ,plane, comp); GLSLF(3, src[1] = imageLoad(input_img[%i], pos + offs[1])[%i]; ,plane, comp);
GLSLF(3, src[2] = texture(input_img[%i], pos + offs[2])[%i]; ,plane, comp); GLSLF(3, src[2] = imageLoad(input_img[%i], pos + offs[2])[%i]; ,plane, comp);
GLSLF(3, src[3] = texture(input_img[%i], pos + offs[3])[%i]; ,plane, comp); GLSLF(3, src[3] = imageLoad(input_img[%i], pos + offs[3])[%i]; ,plane, comp);
GLSLC(0, ); GLSLC(0, );
GLSLC(3, if (lt == false) { ); GLSLC(3, if (lt == false) { );
GLSLC(3, offset = int_stride * uint64_t(pos.y - p); ); GLSLC(3, offset = int_stride * uint64_t(pos.y - p); );
@@ -210,7 +209,7 @@ typedef struct HorizontalPushData {
static av_cold int init_weights_pipeline(FFVulkanContext *vkctx, FFVkExecPool *exec, static av_cold int init_weights_pipeline(FFVulkanContext *vkctx, FFVkExecPool *exec,
FFVulkanShader *shd, FFVulkanShader *shd,
VkSampler sampler, FFVkSPIRVCompiler *spv, FFVkSPIRVCompiler *spv,
int width, int height, int t, int width, int height, int t,
const AVPixFmtDescriptor *desc, const AVPixFmtDescriptor *desc,
int planes, int *nb_rows) int planes, int *nb_rows)
@@ -275,11 +274,12 @@ static av_cold int init_weights_pipeline(FFVulkanContext *vkctx, FFVkExecPool *e
desc_set = (FFVulkanDescriptorSetBinding []) { desc_set = (FFVulkanDescriptorSetBinding []) {
{ {
.name = "input_img", .name = "input_img",
.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.mem_layout = ff_vk_shader_rep_fmt(vkctx->input_format, FF_VK_REP_FLOAT),
.mem_quali = "readonly",
.dimensions = 2, .dimensions = 2,
.elems = planes, .elems = planes,
.stages = VK_SHADER_STAGE_COMPUTE_BIT, .stages = VK_SHADER_STAGE_COMPUTE_BIT,
.samplers = DUP_SAMPLER(sampler),
}, },
{ {
.name = "weights_buffer_0", .name = "weights_buffer_0",
@@ -418,8 +418,7 @@ typedef struct DenoisePushData {
} DenoisePushData; } DenoisePushData;
static av_cold int init_denoise_pipeline(FFVulkanContext *vkctx, FFVkExecPool *exec, static av_cold int init_denoise_pipeline(FFVulkanContext *vkctx, FFVkExecPool *exec,
FFVulkanShader *shd, FFVulkanShader *shd, FFVkSPIRVCompiler *spv,
VkSampler sampler, FFVkSPIRVCompiler *spv,
const AVPixFmtDescriptor *desc, int planes) const AVPixFmtDescriptor *desc, int planes)
{ {
int err; int err;
@@ -445,11 +444,12 @@ static av_cold int init_denoise_pipeline(FFVulkanContext *vkctx, FFVkExecPool *e
desc_set = (FFVulkanDescriptorSetBinding []) { desc_set = (FFVulkanDescriptorSetBinding []) {
{ {
.name = "input_img", .name = "input_img",
.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, .type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.mem_layout = ff_vk_shader_rep_fmt(vkctx->input_format, FF_VK_REP_FLOAT),
.mem_quali = "readonly",
.dimensions = 2, .dimensions = 2,
.elems = planes, .elems = planes,
.stages = VK_SHADER_STAGE_COMPUTE_BIT, .stages = VK_SHADER_STAGE_COMPUTE_BIT,
.samplers = DUP_SAMPLER(sampler),
}, },
{ {
.name = "output_img", .name = "output_img",
@@ -539,7 +539,7 @@ static av_cold int init_denoise_pipeline(FFVulkanContext *vkctx, FFVkExecPool *e
GLSLC(1, if (!IS_WITHIN(pos, size)) ); GLSLC(1, if (!IS_WITHIN(pos, size)) );
GLSLC(2, return; ); GLSLC(2, return; );
GLSLC(0, ); GLSLC(0, );
GLSLC(1, src = texture(input_img[plane], pos); ); GLSLC(1, src = imageLoad(input_img[plane], pos); );
GLSLC(0, ); GLSLC(0, );
for (int c = 0; c < desc->nb_components; c++) { for (int c = 0; c < desc->nb_components; c++) {
int off = desc->comp[c].offset / (FFALIGN(desc->comp[c].depth, 8)/8); int off = desc->comp[c].offset / (FFALIGN(desc->comp[c].depth, 8)/8);
@@ -660,13 +660,12 @@ static av_cold int init_filter(AVFilterContext *ctx)
} }
RET(ff_vk_exec_pool_init(vkctx, s->qf, &s->e, 1, 0, 0, 0, NULL)); RET(ff_vk_exec_pool_init(vkctx, s->qf, &s->e, 1, 0, 0, 0, NULL));
RET(ff_vk_init_sampler(vkctx, &s->sampler, 1, VK_FILTER_NEAREST));
RET(init_weights_pipeline(vkctx, &s->e, &s->shd_weights, s->sampler, RET(init_weights_pipeline(vkctx, &s->e, &s->shd_weights,
spv, s->vkctx.output_width, s->vkctx.output_height, spv, s->vkctx.output_width, s->vkctx.output_height,
s->opts.t, desc, planes, &s->pl_weights_rows)); s->opts.t, desc, planes, &s->pl_weights_rows));
RET(init_denoise_pipeline(vkctx, &s->e, &s->shd_denoise, s->sampler, RET(init_denoise_pipeline(vkctx, &s->e, &s->shd_denoise,
spv, desc, planes)); spv, desc, planes));
RET(ff_vk_shader_update_desc_buffer(vkctx, &s->e.contexts[0], &s->shd_weights, RET(ff_vk_shader_update_desc_buffer(vkctx, &s->e.contexts[0], &s->shd_weights,
@@ -865,7 +864,7 @@ static int nlmeans_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_SHADER_READ_BIT, VK_ACCESS_SHADER_READ_BIT,
VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_LAYOUT_GENERAL,
VK_QUEUE_FAMILY_IGNORED); VK_QUEUE_FAMILY_IGNORED);
/* Output frame prep */ /* Output frame prep */
@@ -944,8 +943,7 @@ static int nlmeans_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
/* Update weights descriptors */ /* Update weights descriptors */
ff_vk_shader_update_img_array(vkctx, exec, &s->shd_weights, in, in_views, 0, 0, ff_vk_shader_update_img_array(vkctx, exec, &s->shd_weights, in, in_views, 0, 0,
VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_LAYOUT_GENERAL, VK_NULL_HANDLE);
s->sampler);
for (int i = 0; i < desc->nb_components; i++) { for (int i = 0; i < desc->nb_components; i++) {
RET(ff_vk_shader_update_desc_buffer(&s->vkctx, exec, &s->shd_weights, 0, 1 + i*2 + 0, 0, RET(ff_vk_shader_update_desc_buffer(&s->vkctx, exec, &s->shd_weights, 0, 1 + i*2 + 0, 0,
ws_vk, weights_offs[i], ws_size[i], ws_vk, weights_offs[i], ws_size[i],
@@ -957,10 +955,9 @@ static int nlmeans_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
/* Update denoise descriptors */ /* Update denoise descriptors */
ff_vk_shader_update_img_array(vkctx, exec, &s->shd_denoise, in, in_views, 0, 0, ff_vk_shader_update_img_array(vkctx, exec, &s->shd_denoise, in, in_views, 0, 0,
VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_LAYOUT_GENERAL, VK_NULL_HANDLE);
s->sampler);
ff_vk_shader_update_img_array(vkctx, exec, &s->shd_denoise, out, out_views, 0, 1, ff_vk_shader_update_img_array(vkctx, exec, &s->shd_denoise, out, out_views, 0, 1,
VK_IMAGE_LAYOUT_GENERAL, s->sampler); VK_IMAGE_LAYOUT_GENERAL, VK_NULL_HANDLE);
for (int i = 0; i < desc->nb_components; i++) { for (int i = 0; i < desc->nb_components; i++) {
RET(ff_vk_shader_update_desc_buffer(&s->vkctx, exec, &s->shd_denoise, 1, i*2 + 0, 0, RET(ff_vk_shader_update_desc_buffer(&s->vkctx, exec, &s->shd_denoise, 1, i*2 + 0, 0,
ws_vk, weights_offs[i], ws_size[i], ws_vk, weights_offs[i], ws_size[i],
@@ -1052,7 +1049,6 @@ static void nlmeans_vulkan_uninit(AVFilterContext *avctx)
{ {
NLMeansVulkanContext *s = avctx->priv; NLMeansVulkanContext *s = avctx->priv;
FFVulkanContext *vkctx = &s->vkctx; FFVulkanContext *vkctx = &s->vkctx;
FFVulkanFunctions *vk = &vkctx->vkfn;
ff_vk_exec_pool_free(vkctx, &s->e); ff_vk_exec_pool_free(vkctx, &s->e);
ff_vk_shader_free(vkctx, &s->shd_weights); ff_vk_shader_free(vkctx, &s->shd_weights);
@@ -1061,10 +1057,6 @@ static void nlmeans_vulkan_uninit(AVFilterContext *avctx)
av_buffer_pool_uninit(&s->integral_buf_pool); av_buffer_pool_uninit(&s->integral_buf_pool);
av_buffer_pool_uninit(&s->ws_buf_pool); av_buffer_pool_uninit(&s->ws_buf_pool);
if (s->sampler)
vk->DestroySampler(vkctx->hwctx->act_dev, s->sampler,
vkctx->hwctx->alloc);
ff_vk_uninit(&s->vkctx); ff_vk_uninit(&s->vkctx);
av_freep(&s->xoffsets); av_freep(&s->xoffsets);