eessppeelllloo/ffmpeg
1
0
Fork 0
mirror of https://git.ffmpeg.org/ffmpeg.git synced 2026-03-29 16:57:46 +02:00
Code Issues Packages Projects Releases Wiki Activity

scale_vulkan: use compile-time SPIR-V generation for debayering

Browse Source
This commit is contained in:
Lynne
2026-01-09 12:51:34 +01:00
parent f2a55af9a4
commit 80229c161b
3 changed files with 107 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files

107
libavfilter/vf_scale_vulkan.c
View File

@@ -27,7 +27,8 @@
#include "colorspace.h"
#include "video.h"
extern const char *ff_source_debayer_comp;
extern const unsigned char ff_debayer_comp_spv_data[];
extern const unsigned int ff_debayer_comp_spv_len;
enum ScalerFunc {
F_BILINEAR = 0,
@@ -195,25 +196,6 @@ static int init_scale_shader(AVFilterContext *ctx, FFVulkanShader *shd,
return 0;
}
static int init_debayer_shader(ScaleVulkanContext *s, FFVulkanShader *shd,
FFVulkanDescriptorSetBinding *desc, AVFrame *in)
{
GLSLD(ff_source_debayer_comp);
GLSLC(0, void main(void));
GLSLC(0, { );
if (s->debayer == DB_BILINEAR)
GLSLC(1, debayer_bilinear(););
else if (s->debayer == DB_BILINEAR_HQ)
GLSLC(1, debayer_bilinear_hq(););
GLSLC(0, } );
shd->lg_size[0] <<= 1;
shd->lg_size[1] <<= 1;
return 0;
}
static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
{
int err;
@@ -227,7 +209,6 @@ static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
FFVkSPIRVCompiler *spv;
FFVulkanDescriptorSetBinding *desc;
int debayer = s->vkctx.input_format == AV_PIX_FMT_BAYER_RGGB16;
int in_planes = av_pix_fmt_count_planes(s->vkctx.input_format);
switch (s->scaler) {
@@ -245,17 +226,9 @@ static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
return AVERROR_EXTERNAL;
}
s->qf = ff_vk_qf_find(vkctx, VK_QUEUE_COMPUTE_BIT, 0);
if (!s->qf) {
av_log(ctx, AV_LOG_ERROR, "Device has no compute queues\n");
err = AVERROR(ENOTSUP);
goto fail;
}
RET(ff_vk_exec_pool_init(vkctx, s->qf, &s->e, s->qf->num*4, 0, 0, 0, NULL));
if (!debayer)
RET(ff_vk_init_sampler(vkctx, &s->sampler, 0, sampler_mode));
RET(ff_vk_init_sampler(vkctx, &s->sampler, 0, sampler_mode));
RET(ff_vk_shader_init(vkctx, &s->shd, "scale",
VK_SHADER_STAGE_COMPUTE_BIT,
@@ -266,12 +239,8 @@ static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
desc = (FFVulkanDescriptorSetBinding []) {
{
.name = "input_img",
.type = debayer ?
VK_DESCRIPTOR_TYPE_STORAGE_IMAGE :
VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
.mem_layout = debayer ?
ff_vk_shader_rep_fmt(s->vkctx.input_format, FF_VK_REP_FLOAT) :
NULL,
.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.mem_layout = NULL,
.mem_quali = "readonly",
.dimensions = 2,
.elems = in_planes,
@@ -303,10 +272,7 @@ static av_cold int init_filter(AVFilterContext *ctx, AVFrame *in)
ff_vk_shader_add_push_const(&s->shd, 0, sizeof(s->opts),
VK_SHADER_STAGE_COMPUTE_BIT);
if (debayer)
err = init_debayer_shader(s, shd, desc, in);
else
err = init_scale_shader(ctx, shd, desc, in);
err = init_scale_shader(ctx, shd, desc, in);
if (err < 0)
goto fail;
@@ -327,6 +293,52 @@ fail:
return err;
}
static av_cold int init_debayer(AVFilterContext *ctx, AVFrame *in)
{
int err;
ScaleVulkanContext *s = ctx->priv;
FFVulkanContext *vkctx = &s->vkctx;
FFVulkanShader *shd = &s->shd;
RET(ff_vk_exec_pool_init(vkctx, s->qf, &s->e, s->qf->num*2, 0, 0, 0, NULL));
SPEC_LIST_CREATE(sl, 1, 1*sizeof(int32_t))
SPEC_LIST_ADD(sl, 0, 32, s->debayer);
ff_vk_shader_load(&s->shd, VK_SHADER_STAGE_COMPUTE_BIT, sl,
(uint32_t []) { 32, 32, 1 }, 0);
ff_vk_shader_add_push_const(&s->shd, 0, sizeof(s->opts),
VK_SHADER_STAGE_COMPUTE_BIT);
const FFVulkanDescriptorSetBinding desc[] = {
{
.name = "src",
.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
},
{
.name = "dst",
.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
},
};
ff_vk_shader_add_descriptor_set(vkctx, &s->shd, desc, 2, 0, 0);
RET(ff_vk_shader_link(vkctx, shd,
ff_debayer_comp_spv_data,
ff_debayer_comp_spv_len, "main"));
RET(ff_vk_shader_register_exec(vkctx, &s->e, &s->shd));
shd->lg_size[0] <<= 1;
shd->lg_size[1] <<= 1;
s->initialized = 1;
fail:
return err;
}
static int scale_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
{
int err;
@@ -340,8 +352,19 @@ static int scale_vulkan_filter_frame(AVFilterLink *link, AVFrame *in)
goto fail;
}
if (!s->initialized)
RET(init_filter(ctx, in));
s->qf = ff_vk_qf_find(&s->vkctx, VK_QUEUE_COMPUTE_BIT, 0);
if (!s->qf) {
av_log(ctx, AV_LOG_ERROR, "Device has no compute queues\n");
err = AVERROR(ENOTSUP);
goto fail;
}
if (!s->initialized) {
if (s->vkctx.input_format == AV_PIX_FMT_BAYER_RGGB16)
RET(init_debayer(ctx, in));
else
RET(init_filter(ctx, in));
}
s->opts.crop_x = in->crop_left;
s->opts.crop_y = in->crop_top;

8
libavfilter/vulkan/Makefile
View File

@@ -3,10 +3,4 @@ clean::
OBJS-$(CONFIG_AVGBLUR_VULKAN_FILTER) += vulkan/avgblur.comp.spv.o
OBJS-$(CONFIG_BWDIF_VULKAN_FILTER) += vulkan/bwdif.comp.spv.o
OBJS-$(CONFIG_SCALE_VULKAN_FILTER) += vulkan/debayer.o
VULKAN = $(subst $(SRC_PATH)/,,$(wildcard $(SRC_PATH)/libavfilter/vulkan/*.comp))
.SECONDARY: $(VULKAN:.comp=.c)
libavfilter/vulkan/%.c: TAG = VULKAN
libavfilter/vulkan/%.c: $(SRC_PATH)/libavfilter/vulkan/%.comp
$(M)$(SRC_PATH)/tools/source2c $< $@
OBJS-$(CONFIG_SCALE_VULKAN_FILTER) += vulkan/debayer.comp.spv.o

56
libavfilter/vulkan/debayer.comp → libavfilter/vulkan/debayer.comp.glsl
View File

@@ -19,46 +19,62 @@
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#version 460
#pragma shader_stage(compute)
#extension GL_EXT_shader_image_load_formatted : require
layout (local_size_x_id = 253, local_size_y_id = 254, local_size_z_id = 255) in;
layout (set = 0, binding = 0) uniform readonly image2D src;
layout (set = 0, binding = 1) uniform writeonly image2D dst;
layout (constant_id = 0) const int debayer_mode = 0;
layout(push_constant, std430) uniform pushConstants {
mat4 yuv_matrix;
int crop_x;
int crop_y;
int crop_w;
int crop_h;
};
#define LD(xo, yo) \
(imageLoad(input_img[0], pos + ivec2((xo), (yo))).r)
(imageLoad(src, pos + ivec2((xo), (yo))).r)
void debayer_bilinear(void)
void debayer_bilinear(ivec2 pos)
{
ivec2 pos = ivec2(gl_GlobalInvocationID.xy) << 1;
/* R basis */
vec4 tl = vec4(LD(0, 0),
(LD(1, 0) + LD(-1, 0) + LD(0, 1) + LD(0, -1)) / 4.0f,
(LD(-1, -1) + LD(1, 1) + LD(-1, 1) + LD(1, -1)) / 4.0f,
1.0f);
imageStore(output_img[0], pos, tl);
imageStore(dst, pos, tl);
/* G1 basis */
vec4 tr = vec4((LD(2, 0) + LD(0, 0)) / 2.0f,
LD(1, 0),
(LD(1, 1) + LD(1, -1)) / 2.0f,
1.0f);
imageStore(output_img[0], pos + ivec2(1, 0), tr);
imageStore(dst, pos + ivec2(1, 0), tr);
/* G2 basis */
vec4 bl = vec4((LD(0, 2) + LD(0, 0)) / 2.0f,
LD(0, 1),
(LD(1, 1) + LD(-1, 1)) / 2.0f,
1.0f);
imageStore(output_img[0], pos + ivec2(0, 1), bl);
imageStore(dst, pos + ivec2(0, 1), bl);
/* B basis */
vec4 br = vec4((LD(0, 0) + LD(2, 2) + LD(0, 2) + LD(2, 0)) / 4.0f,
(LD(2, 1) + LD(0, 1) + LD(1, 2) + LD(1, 0)) / 4.0f,
LD(1, 1),
1.0f);
imageStore(output_img[0], pos + ivec2(1, 1), br);
imageStore(dst, pos + ivec2(1, 1), br);
}
void debayer_bilinear_hq(void)
void debayer_bilinear_hq(ivec2 pos)
{
ivec2 pos = ivec2(gl_GlobalInvocationID.xy) << 1;
/* R basis */
vec4 tl = vec4(LD(0, 0),
(4.0f*LD(0, 0) + 2.0f*(LD(0, -1) + LD(0, 1) + LD(-1, 0) + LD(1, 0)) -
@@ -66,7 +82,7 @@ void debayer_bilinear_hq(void)
(12.0f*LD(0, 0) + 4.0f*(LD(-1, -1) + LD(-1, 1) + LD(1, -1) + LD(1, 1)) -
3.0f*(LD(0, -2) + LD(0, 2) + LD(-2, 0) + LD(2, 0))) / 16.0f,
1.0f);
imageStore(output_img[0], pos, tl);
imageStore(dst, pos, tl);
/* G1 basis */
vec4 tr = vec4((10.0f*LD(1, 0) + 8.0f*(LD(0, 0) + LD(2, 0)) -
@@ -77,7 +93,7 @@ void debayer_bilinear_hq(void)
2.0f*(LD(0, -1) + LD(0, 1) + LD(2, -1) + LD(2, 1) + LD(1, -2) + LD(1, 2)) +
LD(-1, 0) + LD(3, 0)) / 16.0f,
1.0f);
imageStore(output_img[0], pos + ivec2(1, 0), tr);
imageStore(dst, pos + ivec2(1, 0), tr);
/* G2 basis */
@@ -89,7 +105,7 @@ void debayer_bilinear_hq(void)
2.0f*(LD(-1, 0) + LD(1, 2) + LD(-1, 2) + LD(1, 0) + LD(-2, 1) + LD(2, 1)) +
LD(0, -1) + LD(0, 3)) / 16.0f,
1.0f);
imageStore(output_img[0], pos + ivec2(0, 1), bl);
imageStore(dst, pos + ivec2(0, 1), bl);
/* B basis */
vec4 br = vec4((12.0f*LD(1, 1) + 4.0f*(LD(0, 0) + LD(0, 2) + LD(2, 0) + LD(2, 2)) -
@@ -98,5 +114,15 @@ void debayer_bilinear_hq(void)
(LD(1, -1) + LD(1, 3) + LD(-1, 1) + LD(3, 1))) / 8.0f,
LD(1, 1),
1.0f);
imageStore(output_img[0], pos + ivec2(1, 1), br);
imageStore(dst, pos + ivec2(1, 1), br);
}
void main(void)
{
ivec2 pos = ivec2(gl_GlobalInvocationID.xy) << 1;
if (debayer_mode == 0)
debayer_bilinear(pos);
else
debayer_bilinear_hq(pos);
}