har0ke/FFmpeg
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

nlmeans_vulkan: reduce dispatches by parallelizing the planes

Browse Source
This commit is contained in:
Lynne 2023-09-16 01:04:18 +02:00
parent 6bc8ff7d93
commit 658b01b5ee
No known key found for this signature in database
GPG Key ID: A2FEA5F03F034464
1 changed files with 17 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
libavfilter/vf_nlmeans_vulkan.c
View File

@ -538,28 +538,29 @@ static av_cold int init_denoise_pipeline(FFVulkanContext *vkctx, FFVkExecPool *e
GLSLC(0, { ); GLSLC(0, { );
GLSLC(1, ivec2 size; ); GLSLC(1, ivec2 size; );
GLSLC(1, const ivec2 pos = ivec2(gl_GlobalInvocationID.xy); ); GLSLC(1, const ivec2 pos = ivec2(gl_GlobalInvocationID.xy); );
GLSLC(1, const uint plane = uint(gl_WorkGroupID.z); );
GLSLC(0, ); GLSLC(0, );
GLSLC(1, float w_sum; ); GLSLC(1, float w_sum; );
GLSLC(1, float sum; ); GLSLC(1, float sum; );
GLSLC(1, vec4 src; ); GLSLC(1, vec4 src; );
GLSLC(1, vec4 r; ); GLSLC(1, vec4 r; );
GLSLC(0, ); GLSLC(0, );
GLSLC(1, size = imageSize(output_img[plane]); );
for (int i = 0; i < planes; i++) { GLSLC(1, if (!IS_WITHIN(pos, size)) );
GLSLF(1, src = texture(input_img[%i], pos); ,i); GLSLC(2, return; );
GLSLC(0, );
GLSLC(1, src = texture(input_img[plane], pos); );
GLSLC(0, );
for (int c = 0; c < desc->nb_components; c++) { for (int c = 0; c < desc->nb_components; c++) {
if (desc->comp[c].plane == i) {
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);
GLSLF(1, w_sum = weights_%i[pos.y*ws_stride[%i] + pos.x]; ,c, c); GLSLF(1, if (plane == %i) { ,desc->comp[c].plane);
GLSLF(1, sum = sums_%i[pos.y*ws_stride[%i] + pos.x]; ,c, c); GLSLF(2, w_sum = weights_%i[pos.y*ws_stride[%i] + pos.x]; ,c, c);
GLSLF(1, r[%i] = (sum + src[%i]*255) / (1.0 + w_sum) / 255; ,off, off); GLSLF(2, sum = sums_%i[pos.y*ws_stride[%i] + pos.x]; ,c, c);
GLSLF(2, r[%i] = (sum + src[%i]*255) / (1.0 + w_sum) / 255; ,off, off);
GLSLC(1, } );
GLSLC(0, ); GLSLC(0, );
} }
} GLSLC(1, imageStore(output_img[plane], pos, r); );
GLSLF(1, imageStore(output_img[%i], pos, r); ,i);
GLSLC(0, );
}
GLSLC(0, } ); GLSLC(0, } );
RET(spv->compile_shader(spv, vkctx, shd, &spv_data, &spv_len, "main", &spv_opaque)); RET(spv->compile_shader(spv, vkctx, shd, &spv_data, &spv_len, "main", &spv_opaque));
@ -716,7 +717,7 @@ static int denoise_pass(NLMeansVulkanContext *s, FFVkExecContext *exec,
vk->CmdDispatch(exec->buf, vk->CmdDispatch(exec->buf,
FFALIGN(vkctx->output_width, s->pl_denoise.wg_size[0])/s->pl_denoise.wg_size[0], FFALIGN(vkctx->output_width, s->pl_denoise.wg_size[0])/s->pl_denoise.wg_size[0],
FFALIGN(vkctx->output_height, s->pl_denoise.wg_size[1])/s->pl_denoise.wg_size[1], FFALIGN(vkctx->output_height, s->pl_denoise.wg_size[1])/s->pl_denoise.wg_size[1],
1); av_pix_fmt_count_planes(s->vkctx.output_format));
return 0; return 0;
} }