virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2025-02-04 06:08:26 +02:00
Code Issues Releases Activity

scale_vulkan: port for the rewrite

Browse Source
This commit is contained in:
Lynne 2023-02-17 03:13:32 +01:00
parent 8e9ceb1efb
commit d4b51b5085
No known key found for this signature in database
GPG Key ID: A2FEA5F03F034464
1 changed files with 130 additions and 246 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

376
libavfilter/vf_scale_vulkan.c
View File

@ -1,4 +1,6 @@
/*
* Copyright (c) Lynne
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
@ -19,12 +21,11 @@
#include "libavutil/random_seed.h"
#include "libavutil/opt.h"
#include "vulkan_filter.h"
#include "vulkan_spirv.h"
#include "scale_eval.h"
#include "internal.h"
#include "colorspace.h"
#define CGROUPS (int [3]){ 32, 32, 1 }
enum ScalerFunc {
F_BILINEAR = 0,
F_NEAREST,
@ -35,15 +36,17 @@ enum ScalerFunc {
typedef struct ScaleVulkanContext {
FFVulkanContext vkctx;
int initialized;
FFVulkanPipeline pl;
FFVkExecPool e;
FFVkQueueFamilyCtx qf;
FFVkExecContext *exec;
FFVulkanPipeline *pl;
FFVkBuffer params_buf;
FFVkSPIRVShader shd;
VkSampler sampler;
/* Shader updators, must be in the main filter struct */
VkDescriptorImageInfo input_images[3];
VkDescriptorImageInfo output_images[3];
VkDescriptorBufferInfo params_desc;
/* Push constants / options */
struct {
float yuv_matrix[4][4];
} opts;
char *out_format_string;
char *w_expr;
@ -51,8 +54,6 @@ typedef struct ScaleVulkanContext {
enum ScalerFunc scaler;
enum AVColorRange out_range;
int initialized;
} ScaleVulkanContext;
static const char scale_bilinear[] = {
@ -110,10 +111,15 @@ static const char write_444[] = {
static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
{
int err;
FFVkSampler *sampler;
uint8_t *spv_data;
size_t spv_len;
void *spv_opaque = NULL;
VkFilter sampler_mode;
ScaleVulkanContext *s = ctx->priv;
FFVulkanContext *vkctx = &s->vkctx;
FFVkSPIRVShader *shd = &s->shd;
FFVkSPIRVCompiler *spv;
FFVulkanDescriptorSetBinding *desc;
int crop_x = in->crop_left;
int crop_y = in->crop_top;
@ -121,8 +127,6 @@ static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
int crop_h = in->height - (in->crop_top + in->crop_bottom);
int in_planes = av_pix_fmt_count_planes(s->vkctx.input_format);
ff_vk_qf_init(vkctx, &s->qf, VK_QUEUE_COMPUTE_BIT, 0);
switch (s->scaler) {
case F_NEAREST:
sampler_mode = VK_FILTER_NEAREST;
@ -132,263 +136,133 @@ static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
break;
};
/* Create a sampler */
sampler = ff_vk_init_sampler(vkctx, 0, sampler_mode);
if (!sampler)
spv = ff_vk_spirv_init();
if (!spv) {
av_log(ctx, AV_LOG_ERROR, "Unable to initialize SPIR-V compiler!\n");
return AVERROR_EXTERNAL;
s->pl = ff_vk_create_pipeline(vkctx, &s->qf);
if (!s->pl)
return AVERROR(ENOMEM);
{ /* Create the shader */
FFVulkanDescriptorSetBinding desc_i[2] = {
{
.name = "input_img",
.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
.dimensions = 2,
.elems = in_planes,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.updater = s->input_images,
.sampler = 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 = av_pix_fmt_count_planes(s->vkctx.output_format),
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.updater = s->output_images,
},
};
FFVulkanDescriptorSetBinding desc_b = {
.name = "params",
.type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
.mem_quali = "readonly",
.mem_layout = "std430",
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.updater = &s->params_desc,
.buf_content = "mat4 yuv_matrix;",
};
FFVkSPIRVShader *shd = ff_vk_init_shader(s->pl, "scale_compute",
VK_SHADER_STAGE_COMPUTE_BIT);
if (!shd)
return AVERROR(ENOMEM);
ff_vk_set_compute_shader_sizes(shd, CGROUPS);
RET(ff_vk_add_descriptor_set(vkctx, s->pl, shd, desc_i, FF_ARRAY_ELEMS(desc_i), 0)); /* set 0 */
RET(ff_vk_add_descriptor_set(vkctx, s->pl, shd, &desc_b, 1, 0)); /* set 1 */
GLSLD( scale_bilinear );
if (s->vkctx.output_format != s->vkctx.input_format) {
GLSLD( rgb2yuv );
}
switch (s->vkctx.output_format) {
case AV_PIX_FMT_NV12: GLSLD(write_nv12); break;
case AV_PIX_FMT_YUV420P: GLSLD( write_420); break;
case AV_PIX_FMT_YUV444P: GLSLD( write_444); break;
default: break;
}
GLSLC(0, void main() );
GLSLC(0, { );
GLSLC(1, ivec2 size; );
GLSLC(1, ivec2 pos = ivec2(gl_GlobalInvocationID.xy); );
GLSLF(1, vec2 in_d = vec2(%i, %i); ,in->width, in->height);
GLSLF(1, vec2 c_r = vec2(%i, %i) / in_d; ,crop_w, crop_h);
GLSLF(1, vec2 c_o = vec2(%i, %i) / in_d; ,crop_x,crop_y);
GLSLC(0, );
if (s->vkctx.output_format == s->vkctx.input_format) {
for (int i = 0; i < desc_i[1].elems; i++) {
GLSLF(1, size = imageSize(output_img[%i]); ,i);
GLSLC(1, if (IS_WITHIN(pos, size)) { );
switch (s->scaler) {
case F_NEAREST:
case F_BILINEAR:
GLSLF(2, vec4 res = scale_bilinear(%i, pos, c_r, c_o); ,i);
GLSLF(2, imageStore(output_img[%i], pos, res); ,i);
break;
};
GLSLC(1, } );
}
} else {
GLSLC(1, vec4 res = scale_bilinear(0, pos, c_r, c_o); );
GLSLF(1, res = rgb2yuv(res, %i); ,s->out_range == AVCOL_RANGE_JPEG);
switch (s->vkctx.output_format) {
case AV_PIX_FMT_NV12: GLSLC(1, write_nv12(res, pos); ); break;
case AV_PIX_FMT_YUV420P: GLSLC(1, write_420(res, pos); ); break;
case AV_PIX_FMT_YUV444P: GLSLC(1, write_444(res, pos); ); break;
default: return AVERROR(EINVAL);
}
}
GLSLC(0, } );
RET(ff_vk_compile_shader(vkctx, shd, "main"));
}
RET(ff_vk_init_pipeline_layout(vkctx, s->pl));
RET(ff_vk_init_compute_pipeline(vkctx, s->pl));
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, 0, sampler_mode));
RET(ff_vk_shader_init(&s->pl, &s->shd, "scale_compute",
VK_SHADER_STAGE_COMPUTE_BIT, 0));
ff_vk_shader_set_compute_sizes(&s->shd, 32, 32, 1);
GLSLC(0, layout(push_constant, std430) uniform pushConstants { );
GLSLC(1, mat4 yuv_matrix; );
GLSLC(0, }; );
GLSLC(0, );
ff_vk_add_push_constant(&s->pl, 0, sizeof(s->opts),
VK_SHADER_STAGE_COMPUTE_BIT);
desc = (FFVulkanDescriptorSetBinding []) {
{
.name = "input_img",
.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
.dimensions = 2,
.elems = in_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 = av_pix_fmt_count_planes(s->vkctx.output_format),
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
},
};
RET(ff_vk_pipeline_descriptor_set_add(vkctx, &s->pl, shd, desc, 2, 0, 0));
GLSLD( scale_bilinear );
if (s->vkctx.output_format != s->vkctx.input_format) {
GLSLD( rgb2yuv );
}
switch (s->vkctx.output_format) {
case AV_PIX_FMT_NV12: GLSLD(write_nv12); break;
case AV_PIX_FMT_YUV420P: GLSLD( write_420); break;
case AV_PIX_FMT_YUV444P: GLSLD( write_444); break;
default: break;
}
GLSLC(0, void main() );
GLSLC(0, { );
GLSLC(1, ivec2 size; );
GLSLC(1, ivec2 pos = ivec2(gl_GlobalInvocationID.xy); );
GLSLF(1, vec2 in_d = vec2(%i, %i); ,in->width, in->height);
GLSLF(1, vec2 c_r = vec2(%i, %i) / in_d; ,crop_w, crop_h);
GLSLF(1, vec2 c_o = vec2(%i, %i) / in_d; ,crop_x,crop_y);
GLSLC(0, );
if (s->vkctx.output_format == s->vkctx.input_format) {
for (int i = 0; i < desc[i].elems; i++) {
GLSLF(1, size = imageSize(output_img[%i]); ,i);
GLSLC(1, if (IS_WITHIN(pos, size)) { );
switch (s->scaler) {
case F_NEAREST:
case F_BILINEAR:
GLSLF(2, vec4 res = scale_bilinear(%i, pos, c_r, c_o); ,i);
GLSLF(2, imageStore(output_img[%i], pos, res); ,i);
break;
};
GLSLC(1, } );
}
} else {
GLSLC(1, vec4 res = scale_bilinear(0, pos, c_r, c_o); );
GLSLF(1, res = rgb2yuv(res, %i); ,s->out_range == AVCOL_RANGE_JPEG);
switch (s->vkctx.output_format) {
case AV_PIX_FMT_NV12: GLSLC(1, write_nv12(res, pos); ); break;
case AV_PIX_FMT_YUV420P: GLSLC(1, write_420(res, pos); ); break;
case AV_PIX_FMT_YUV444P: GLSLC(1, write_444(res, pos); ); break;
default: return AVERROR(EINVAL);
}
}
GLSLC(0, } );
if (s->vkctx.output_format != s->vkctx.input_format) {
const AVLumaCoefficients *lcoeffs;
double tmp_mat[3][3];
struct {
float yuv_matrix[4][4];
} *par;
lcoeffs = av_csp_luma_coeffs_from_avcsp(in->colorspace);
if (!lcoeffs) {
av_log(ctx, AV_LOG_ERROR, "Unsupported colorspace\n");
return AVERROR(EINVAL);
}
RET(ff_vk_create_buf(vkctx, &s->params_buf,
sizeof(*par), NULL, NULL,
VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT));
RET(ff_vk_map_buffers(vkctx, &s->params_buf, (uint8_t **)&par, 1, 0));
ff_fill_rgb2yuv_table(lcoeffs, tmp_mat);
memset(par, 0, sizeof(*par));
for (int y = 0; y < 3; y++)
for (int x = 0; x < 3; x++)
par->yuv_matrix[x][y] = tmp_mat[x][y];
par->yuv_matrix[3][3] = 1.0;
RET(ff_vk_unmap_buffers(vkctx, &s->params_buf, 1, 1));
s->params_desc.buffer = s->params_buf.buf;
s->params_desc.range = VK_WHOLE_SIZE;
ff_vk_update_descriptor_set(vkctx, s->pl, 1);
s->opts.yuv_matrix[x][y] = tmp_mat[x][y];
s->opts.yuv_matrix[3][3] = 1.0;
}
/* Execution context */
RET(ff_vk_create_exec_ctx(vkctx, &s->exec, &s->qf));
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->initialized = 1;
return 0;
fail:
return err;
}
if (spv_opaque)
spv->free_shader(spv, &spv_opaque);
if (spv)
spv->uninit(&spv);
static int process_frames(AVFilterContext *avctx, AVFrame *out_f, AVFrame *in_f)
{
int err = 0;
VkCommandBuffer cmd_buf;
ScaleVulkanContext *s = avctx->priv;
FFVulkanContext *vkctx = &s->vkctx;
FFVulkanFunctions *vk = &vkctx->vkfn;
AVVkFrame *in = (AVVkFrame *)in_f->data[0];
AVVkFrame *out = (AVVkFrame *)out_f->data[0];
VkImageMemoryBarrier barriers[AV_NUM_DATA_POINTERS*2];
int barrier_count = 0;
const int planes = av_pix_fmt_count_planes(s->vkctx.input_format);
const VkFormat *input_formats = av_vkfmt_from_pixfmt(s->vkctx.input_format);
const VkFormat *output_formats = av_vkfmt_from_pixfmt(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->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->output_images[i].imageLayout = VK_IMAGE_LAYOUT_GENERAL;
}
ff_vk_update_descriptor_set(vkctx, s->pl, 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,
};
memcpy(&barriers[barrier_count++], &bar, sizeof(VkImageMemoryBarrier));
in->layout[i] = bar.newLayout;
in->access[i] = bar.dstAccessMask;
}
for (int i = 0; i < av_pix_fmt_count_planes(s->vkctx.output_format); i++) {
VkImageMemoryBarrier bar = {
.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,
};
memcpy(&barriers[barrier_count++], &bar, sizeof(VkImageMemoryBarrier));
out->layout[i] = bar.newLayout;
out->access[i] = bar.dstAccessMask;
}
vk->CmdPipelineBarrier(cmd_buf, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0,
0, NULL, 0, NULL, barrier_count, barriers);
ff_vk_bind_pipeline_exec(vkctx, s->exec, s->pl);
vk->CmdDispatch(cmd_buf,
FFALIGN(vkctx->output_width, CGROUPS[0])/CGROUPS[0],
FFALIGN(vkctx->output_height, CGROUPS[1])/CGROUPS[1], 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);
return err;
fail:
ff_vk_discard_exec_deps(s->exec);
return err;
}
@ -408,7 +282,8 @@ static int scale_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
if (!s->initialized)
RET(init_filter(ctx, in));
RET(process_frames(ctx, out, in));
RET(ff_vk_filter_process_simple(&s->vkctx, &s->e, &s->pl, out, in,
s->sampler, &s->opts, sizeof(s->opts)));
err = av_frame_copy_props(out, in);
if (err < 0)
@ -475,8 +350,17 @@ static int scale_vulkan_config_output(AVFilterLink *outlink)
static void scale_vulkan_uninit(AVFilterContext *avctx)
{
ScaleVulkanContext *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);
ff_vk_shader_free(vkctx, &s->shd);
if (s->sampler)
vk->DestroySampler(vkctx->hwctx->act_dev, s->sampler,
vkctx->hwctx->alloc);
ff_vk_free_buf(&s->vkctx, &s->params_buf);
ff_vk_uninit(&s->vkctx);
s->initialized = 0;