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

bwdif_vulkan: convert to storage images

Browse Source 
texture() uses bilinear scaling; imageLoad() accesses the image directly.
The reason why texture() was used throughout Vulkan filters is that
back when they were written, they were targetting old Intel hardware,
which had a texel cache only for sampled images.

These days, GPUs have a generic cache that doesn't care what source it
gets populated with. Additionally, bypassing the sampling circuitry saves
us some performance.

Finally, all the old texture() code had an issue where unnormalized
coordinates were used, but an offset of 0.5 was not added, hence each
pixel ended up being interpolated. This fixes this.
This commit is contained in:
Lynne
2025-02-17 18:22:45 +01:00
parent 042ed96d0a
commit 779763181f
2 changed files with 43 additions and 47 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
libavfilter/vf_bwdif_vulkan.c
View File

@@ -34,7 +34,6 @@ typedef struct BWDIFVulkanContext {
int initialized;
FFVkExecPool e;
AVVulkanDeviceQueueFamily *qf;
VkSampler sampler;
FFVulkanShader shd;
} BWDIFVulkanContext;
@@ -73,7 +72,6 @@ static av_cold int init_filter(AVFilterContext *ctx)
}
RET(ff_vk_exec_pool_init(vkctx, s->qf, &s->e, s->qf->num*4, 0, 0, 0, NULL));
RET(ff_vk_init_sampler(vkctx, &s->sampler, 1, VK_FILTER_NEAREST));
RET(ff_vk_shader_init(vkctx, &s->shd, "bwdif",
VK_SHADER_STAGE_COMPUTE_BIT,
@@ -85,27 +83,30 @@ static av_cold int init_filter(AVFilterContext *ctx)
desc = (FFVulkanDescriptorSetBinding []) {
{
.name = "prev",
.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.mem_layout = ff_vk_shader_rep_fmt(s->vkctx.input_format, FF_VK_REP_FLOAT),
.mem_quali = "readonly",
.dimensions = 2,
.elems = planes,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.samplers = DUP_SAMPLER(s->sampler),
},
{
.name = "cur",
.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.mem_layout = ff_vk_shader_rep_fmt(s->vkctx.input_format, FF_VK_REP_FLOAT),
.mem_quali = "readonly",
.dimensions = 2,
.elems = planes,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.samplers = DUP_SAMPLER(s->sampler),
},
{
.name = "next",
.type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
.mem_layout = ff_vk_shader_rep_fmt(s->vkctx.input_format, FF_VK_REP_FLOAT),
.mem_quali = "readonly",
.dimensions = 2,
.elems = planes,
.stages = VK_SHADER_STAGE_COMPUTE_BIT,
.samplers = DUP_SAMPLER(s->sampler),
},
{
.name = "dst",
@@ -166,7 +167,7 @@ static av_cold int init_filter(AVFilterContext *ctx)
GLSLC(2, if (!IS_WITHIN(pos, size)) );
GLSLC(3, return; );
}
GLSLF(2, imageStore(dst[%i], pos, texture(cur[%i], pos)); ,i, i);
GLSLF(2, imageStore(dst[%i], pos, imageLoad(cur[%i], pos)); ,i, i);
}
GLSLC(1, } );
GLSLC(0, } );
@@ -201,7 +202,7 @@ static void bwdif_vulkan_filter_frame(AVFilterContext *ctx, AVFrame *dst,
ff_vk_filter_process_Nin(&s->vkctx, &s->e, &s->shd, dst,
(AVFrame *[]){ y->prev, y->cur, y->next }, 3,
s->sampler, &params, sizeof(params));
VK_NULL_HANDLE, &params, sizeof(params));
if (y->current_field == YADIF_FIELD_END)
y->current_field = YADIF_FIELD_NORMAL;
@@ -211,15 +212,10 @@ static void bwdif_vulkan_uninit(AVFilterContext *avctx)
{
BWDIFVulkanContext *s = avctx->priv;
FFVulkanContext *vkctx = &s->vkctx;
FFVulkanFunctions *vk = &vkctx->vkfn;
ff_vk_exec_pool_free(vkctx, &s->e);
ff_vk_shader_free(vkctx, &s->shd);
if (s->sampler)
vk->DestroySampler(vkctx->hwctx->act_dev, s->sampler,
vkctx->hwctx->alloc);
ff_vk_uninit(&s->vkctx);
ff_yadif_uninit(avctx);

64
libavfilter/vulkan/bwdif.comp
View File

@@ -30,10 +30,10 @@ vec4 process_intra(vec4 cur[4])
void process_plane_intra(int idx, ivec2 pos)
{
vec4 dcur[4];
dcur[0] = texture(cur[idx], pos - ivec2(0, 3));
dcur[1] = texture(cur[idx], pos - ivec2(0, 1));
dcur[2] = texture(cur[idx], pos + ivec2(0, 1));
dcur[3] = texture(cur[idx], pos + ivec2(0, 3));
dcur[0] = imageLoad(cur[idx], pos - ivec2(0, 3));
dcur[1] = imageLoad(cur[idx], pos - ivec2(0, 1));
dcur[2] = imageLoad(cur[idx], pos + ivec2(0, 1));
dcur[3] = imageLoad(cur[idx], pos + ivec2(0, 3));
imageStore(dst[idx], pos, process_intra(dcur));
}
@@ -81,41 +81,41 @@ void process_plane(int idx, const ivec2 pos, bool filter_field,
vec4 prev2[5];
vec4 next2[5];
dcur[0] = texture(cur[idx], pos - ivec2(0, 3));
dcur[1] = texture(cur[idx], pos - ivec2(0, 1));
dcur[2] = texture(cur[idx], pos + ivec2(0, 1));
dcur[3] = texture(cur[idx], pos + ivec2(0, 3));
dcur[0] = imageLoad(cur[idx], pos - ivec2(0, 3));
dcur[1] = imageLoad(cur[idx], pos - ivec2(0, 1));
dcur[2] = imageLoad(cur[idx], pos + ivec2(0, 1));
dcur[3] = imageLoad(cur[idx], pos + ivec2(0, 3));
prev1[0] = texture(prev[idx], pos - ivec2(0, 1));
prev1[1] = texture(prev[idx], pos + ivec2(0, 1));
prev1[0] = imageLoad(prev[idx], pos - ivec2(0, 1));
prev1[1] = imageLoad(prev[idx], pos + ivec2(0, 1));
next1[0] = texture(next[idx], pos - ivec2(0, 1));
next1[1] = texture(next[idx], pos + ivec2(0, 1));
next1[0] = imageLoad(next[idx], pos - ivec2(0, 1));
next1[1] = imageLoad(next[idx], pos + ivec2(0, 1));
if (field_parity) {
prev2[0] = texture(prev[idx], pos - ivec2(0, 4));
prev2[1] = texture(prev[idx], pos - ivec2(0, 2));
prev2[2] = texture(prev[idx], pos);
prev2[3] = texture(prev[idx], pos + ivec2(0, 2));
prev2[4] = texture(prev[idx], pos + ivec2(0, 4));
prev2[0] = imageLoad(prev[idx], pos - ivec2(0, 4));
prev2[1] = imageLoad(prev[idx], pos - ivec2(0, 2));
prev2[2] = imageLoad(prev[idx], pos);
prev2[3] = imageLoad(prev[idx], pos + ivec2(0, 2));
prev2[4] = imageLoad(prev[idx], pos + ivec2(0, 4));
next2[0] = texture(cur[idx], pos - ivec2(0, 4));
next2[1] = texture(cur[idx], pos - ivec2(0, 2));
next2[2] = texture(cur[idx], pos);
next2[3] = texture(cur[idx], pos + ivec2(0, 2));
next2[4] = texture(cur[idx], pos + ivec2(0, 4));
next2[0] = imageLoad(cur[idx], pos - ivec2(0, 4));
next2[1] = imageLoad(cur[idx], pos - ivec2(0, 2));
next2[2] = imageLoad(cur[idx], pos);
next2[3] = imageLoad(cur[idx], pos + ivec2(0, 2));
next2[4] = imageLoad(cur[idx], pos + ivec2(0, 4));
} else {
prev2[0] = texture(cur[idx], pos - ivec2(0, 4));
prev2[1] = texture(cur[idx], pos - ivec2(0, 2));
prev2[2] = texture(cur[idx], pos);
prev2[3] = texture(cur[idx], pos + ivec2(0, 2));
prev2[4] = texture(cur[idx], pos + ivec2(0, 4));
prev2[0] = imageLoad(cur[idx], pos - ivec2(0, 4));
prev2[1] = imageLoad(cur[idx], pos - ivec2(0, 2));
prev2[2] = imageLoad(cur[idx], pos);
prev2[3] = imageLoad(cur[idx], pos + ivec2(0, 2));
prev2[4] = imageLoad(cur[idx], pos + ivec2(0, 4));
next2[0] = texture(next[idx], pos - ivec2(0, 4));
next2[1] = texture(next[idx], pos - ivec2(0, 2));
next2[2] = texture(next[idx], pos);
next2[3] = texture(next[idx], pos + ivec2(0, 2));
next2[4] = texture(next[idx], pos + ivec2(0, 4));
next2[0] = imageLoad(next[idx], pos - ivec2(0, 4));
next2[1] = imageLoad(next[idx], pos - ivec2(0, 2));
next2[2] = imageLoad(next[idx], pos);
next2[3] = imageLoad(next[idx], pos + ivec2(0, 2));
next2[4] = imageLoad(next[idx], pos + ivec2(0, 4));
}
imageStore(dst[idx], pos, process_line(prev2, prev1, dcur, next1, next2));