* [FFmpeg-devel] [PATCH] libavfilter/vf_remap_opencl: make kernel handle subsampled chroma
@ 2025-09-05 14:21 Matthias Welwarsky via ffmpeg-devel
0 siblings, 0 replies; only message in thread
From: Matthias Welwarsky via ffmpeg-devel @ 2025-09-05 14:21 UTC (permalink / raw)
To: ffmpeg-devel; +Cc: Matthias Welwarsky
This patch fixes handling YUV frames with subsampled chroma.
Before this patch, the remap kernel was already aware of multiple planes,
but it treated the subsampled chroma incorrectly. The maps were read
based on destination coordinates and the resulting source coordinates
were also not correctly transformed back to the source image.
Also, for both Y and UV, the wrong vector components were used.
With this patch, it is possible to have a decode->remap->encode
chain entirely on the GPU. I'm using it to reframe videos from an
insta360 camera. On a Lenovo T480 with the built-in UHD 620 GPU, it
can achieve about 5x real-time (VAAPI decode -> remap_opencl
-> VAAPI encode).
Signed-off-by: Matthias Welwarsky <ffmpeg@welwarsky.de>
---
libavfilter/opencl/remap.cl | 128 ++++++++++++++++++++++++----------
libavfilter/vf_remap_opencl.c | 43 ++++++++++--
2 files changed, 130 insertions(+), 41 deletions(-)
diff --git a/libavfilter/opencl/remap.cl b/libavfilter/opencl/remap.cl
index fba82d134e..1a18f1afff 100644
--- a/libavfilter/opencl/remap.cl
+++ b/libavfilter/opencl/remap.cl
@@ -22,52 +22,108 @@ const sampler_t linear_sampler =
(CLK_NORMALIZED_COORDS_FALSE |
const sampler_t nearest_sampler = (CLK_NORMALIZED_COORDS_FALSE |
CLK_FILTER_NEAREST);
-__kernel void remap_near(__write_only image2d_t dst,
- __read_only image2d_t src,
- __read_only image2d_t xmapi,
- __read_only image2d_t ymapi,
- float4 fill_color)
+__kernel void remap_near(
+ __write_only image2d_t dst,
+ __read_only image2d_t src,
+ __read_only image2d_t xmapi,
+ __read_only image2d_t ymapi,
+ float4 fill_color,
+ float4 scale,
+ int4 swizzle)
{
int2 p = (int2)(get_global_id(0), get_global_id(1));
- int2 dimi = get_image_dim(src);
- float2 dimf = (float2)(dimi.x, dimi.y);
- float4 val;
- int2 mi;
- float m;
- float4 xmap = read_imagef(xmapi, nearest_sampler, p);
- float4 ymap = read_imagef(ymapi, nearest_sampler, p);
- float2 pos = (float2)(xmap.x, ymap.x);
- pos.xy = pos.xy * 65535.f;
+ /* image dimensions */
+ int2 src_dim = get_image_dim(src);
+ int2 dst_dim = get_image_dim(dst);
+ int2 map_dim = get_image_dim(xmapi);
- mi = ((pos >= (float2)(0.f, 0.f)) * (pos < dimf) * (p <= dimi));
- m = mi.x && mi.y;
- val = mix(fill_color, read_imagef(src, nearest_sampler, pos), m);
+ float2 src_dimf = (float2)(src_dim.x, src_dim.y);
+ float2 dst_dimf = (float2)(dst_dim.x, dst_dim.y);
+ float2 map_dimf = (float2)(map_dim.x, map_dim.y);
+
+ /* compute map scaling to full-res */
+ float2 map_scale = map_dimf / dst_dimf;
+
+ /* scaled position to fetch from the maps */
+ float2 map_p = (float2)(p.x, p.y) * map_scale;
+
+ /* read mapping coordinates from full-res maps */
+ float4 xmap = read_imagef(xmapi, nearest_sampler, map_p);
+ float4 ymap = read_imagef(ymapi, nearest_sampler, map_p);
+ float2 pos = (float2)(xmap.x, ymap.x) * 65535.f;
+
+ pos /= map_scale;
+
+ /* check bounds */
+ int2 mi = ((pos >= (float2)(0.f,0.f)) * (pos < src_dimf));
+ float m = mi.x && mi.y;
+
+ /* read source and apply swizzle + scale */
+ float4 src_val = read_imagef(src, nearest_sampler, pos);
+
+ float tmp[4];
+ vstore4(src_val, 0, tmp);
+ src_val = (float4)(tmp[swizzle.x] * scale.x,
+ tmp[swizzle.y] * scale.y,
+ tmp[swizzle.z] * scale.z,
+ tmp[swizzle.w] * scale.w);
+
+ /* mix with fill color if out-of-bounds */
+ float4 val = mix(fill_color, src_val, m);
write_imagef(dst, p, val);
}
-__kernel void remap_linear(__write_only image2d_t dst,
- __read_only image2d_t src,
- __read_only image2d_t xmapi,
- __read_only image2d_t ymapi,
- float4 fill_color)
+__kernel void remap_linear(
+ __write_only image2d_t dst,
+ __read_only image2d_t src,
+ __read_only image2d_t xmapi,
+ __read_only image2d_t ymapi,
+ float4 fill_color,
+ float4 scale,
+ int4 swizzle)
{
int2 p = (int2)(get_global_id(0), get_global_id(1));
- int2 dimi = get_image_dim(src);
- float2 dimf = (float2)(dimi.x, dimi.y);
- float4 val;
- int2 mi;
- float m;
- float4 xmap = read_imagef(xmapi, nearest_sampler, p);
- float4 ymap = read_imagef(ymapi, nearest_sampler, p);
- float2 pos = (float2)(xmap.x, ymap.x);
- pos.xy = pos.xy * 65535.f;
-
- mi = ((pos >= (float2)(0.f, 0.f)) * (pos < dimf) * (p <= dimi));
- m = mi.x && mi.y;
- val = mix(fill_color, read_imagef(src, linear_sampler, pos), m);
+
+ int2 src_dim = get_image_dim(src);
+ int2 dst_dim = get_image_dim(dst);
+ int2 map_dim = get_image_dim(xmapi);
+
+ float2 src_dimf = (float2)(src_dim.x, src_dim.y);
+ float2 dst_dimf = (float2)(dst_dim.x, dst_dim.y);
+ float2 map_dimf = (float2)(map_dim.x, map_dim.y);
+
+ float2 map_scale = map_dimf / dst_dimf;
+
+ float2 map_p = (float2)(p.x, p.y) * map_scale;
+
+ float4 xmap = read_imagef(xmapi, nearest_sampler, map_p);
+ float4 ymap = read_imagef(ymapi, nearest_sampler, map_p);
+ float2 pos = (float2)(xmap.x, ymap.x) * 65535.f;
+
+ pos /= map_scale;
+
+ int2 mi = ((pos >= (float2)(0.f,0.f)) * (pos < src_dimf));
+ float m = mi.x && mi.y;
+ float4 src_val = read_imagef(src, linear_sampler, pos);
+
+ float tmp[4];
+ vstore4(src_val, 0, tmp);
+ src_val = (float4)(tmp[swizzle.x] * scale.x,
+ tmp[swizzle.y] * scale.y,
+ tmp[swizzle.z] * scale.z,
+ tmp[swizzle.w] * scale.w);
+
+ float4 val = mix(fill_color, src_val, m);
write_imagef(dst, p, val);
}
diff --git a/libavfilter/vf_remap_opencl.c b/libavfilter/vf_remap_opencl.c
index bb83944b2f..f6c5211f73 100644
--- a/libavfilter/vf_remap_opencl.c
+++ b/libavfilter/vf_remap_opencl.c
@@ -29,10 +29,18 @@
#include "opencl_source.h"
#include "video.h"
+#define MAX_PLANES 4
+
+typedef struct PlaneParams {
+ cl_float4 scale;
+ cl_int4 swizzle;
+} PlaneParams;
+
typedef struct RemapOpenCLContext {
OpenCLFilterContext ocf;
int nb_planes;
+ PlaneParams plane[MAX_PLANES];
int interp;
uint8_t fill_rgba[4];
cl_float4 cl_fill_color;
@@ -41,6 +49,7 @@ typedef struct RemapOpenCLContext {
cl_kernel kernel;
cl_command_queue command_queue;
FFFrameSync fs;
} RemapOpenCLContext;
@@ -73,10 +82,9 @@ static int remap_opencl_load(AVFilterContext *avctx,
cl_int cle;
const char *source = ff_source_remap_cl;
const char *kernel = kernels[ctx->interp];
- const AVPixFmtDescriptor *main_desc;
int err, main_planes;
- const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(main_format);
- int is_rgb = !!(desc->flags & AV_PIX_FMT_FLAG_RGB);
+ const AVPixFmtDescriptor *main_desc = av_pix_fmt_desc_get(main_format);
+ int is_rgb = !!(main_desc->flags & AV_PIX_FMT_FLAG_RGB);
const float scale = 1.f / 255.f;
uint8_t rgba_map[4];
@@ -94,8 +102,6 @@ static int remap_opencl_load(AVFilterContext *avctx,
ctx->cl_fill_color.s[3] = ctx->fill_rgba[3] * scale;
}
- main_desc = av_pix_fmt_desc_get(main_format);
-
main_planes = 0;
for (int i = 0; i < main_desc->nb_components; i++)
main_planes = FFMAX(main_planes,
@@ -103,6 +109,26 @@ static int remap_opencl_load(AVFilterContext *avctx,
ctx->nb_planes = main_planes;
+ for (int p = 0; p < ctx->nb_planes; p++) {
+ PlaneParams *pp = &ctx->plane[p];
+
+ if (is_rgb) {
+ // RGB plane (single plane)
+ pp->scale = (cl_float4){{1.0f, 1.0f, 1.0f, 1.0f}};
+ pp->swizzle = (cl_int4){{0, 1, 2, 3}};
+ } else {
+ pp->scale = (cl_float4){{1.0f, 0.0f, 0.0f, 1.0f}};
+ pp->swizzle = (cl_int4){{0, 0, 0, 3}};
+ } else {
+ pp->scale = (cl_float4){{1.0f, 1.0f, 0.0f, 1.0f}};
+ pp->swizzle = (cl_int4){{0, 1, 0, 3}};
+ }
err = ff_opencl_filter_load_program(avctx, &source, 1);
if (err < 0)
goto fail;
@@ -132,6 +158,7 @@ static int remap_opencl_process_frame(FFFrameSync *fs)
AVFilterContext *avctx = fs->parent;
AVFilterLink *outlink = avctx->outputs[0];
RemapOpenCLContext *ctx = avctx->priv;
+
AVFrame *input_main, *input_xmap, *input_ymap;
AVFrame *output;
cl_mem mem;
@@ -199,6 +226,12 @@ static int remap_opencl_process_frame(FFFrameSync *fs)
CL_SET_KERNEL_ARG(ctx->kernel, kernel_arg, cl_float4, &cl_fill_color);
kernel_arg++;
+ CL_SET_KERNEL_ARG(ctx->kernel, kernel_arg, cl_float4, &(ctx-
>plane[plane].scale));
+ kernel_arg++;
+
+ CL_SET_KERNEL_ARG(ctx->kernel, kernel_arg, cl_int4, &(ctx-
>plane[plane].swizzle));
+ kernel_arg++;
+
err = ff_opencl_filter_work_size_from_image(avctx, global_work,
output, plane, 0);
_______________________________________________
ffmpeg-devel mailing list -- ffmpeg-devel@ffmpeg.org
To unsubscribe send an email to ffmpeg-devel-leave@ffmpeg.org
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2025-09-05 14:22 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2025-09-05 14:21 [FFmpeg-devel] [PATCH] libavfilter/vf_remap_opencl: make kernel handle subsampled chroma Matthias Welwarsky via ffmpeg-devel
Git Inbox Mirror of the ffmpeg-devel mailing list - see https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
This inbox may be cloned and mirrored by anyone:
git clone --mirror https://master.gitmailbox.com/ffmpegdev/0 ffmpegdev/git/0.git
# If you have public-inbox 1.1+ installed, you may
# initialize and index your mirror using the following commands:
public-inbox-init -V2 ffmpegdev ffmpegdev/ https://master.gitmailbox.com/ffmpegdev \
ffmpegdev@gitmailbox.com
public-inbox-index ffmpegdev
Example config snippet for mirrors.
AGPL code for this site: git clone https://public-inbox.org/public-inbox.git