vf_cuda_grid: Phase 2b — delegated scaling to upstream scale_npp
После попытки in-filter NPP scaling обнаружено что nppiResize не имеет _C2R variant для NV12 UV interleaved (только C1R, C3R, C4R). Alternatives: - 2× nppiResize_8u_C1R с split/merge через intermediate buffers - custom CUDA kernel - treat UV pair как 16u (blending artifact на boundaries) Pragmatic decision: cuda_grid делает только composition (same-size memcpy), а scaling делегируется существующему scale_npp filter в filter chain: [0]scale_npp=1280:1080[s0]; [1]scale_npp=640:360[s1]; ... \ [s0][s1]...cuda_grid=layout=main_plus_preview Unix philosophy + leverages production-tested NPP code. Controller (Phase 3) auto-generates filter graph с scale_npp per input. Revert: - #include <nppi.h> - libnpp dependency в configure (cuda_grid_filter_deps="ffnvcodec") - nppiResize* calls в compose path Add: - Error message с примером scale_npp chain pattern - Doc в file header c filter graph пример Phase 2 = full deliverable (2a + 2b). Дальше Phase 3 controller.
This commit is contained in:
gx
@@ -3317,7 +3317,7 @@ thumbnail_cuda_filter_deps_any="cuda_nvcc cuda_llvm"
|
||||
transpose_npp_filter_deps="ffnvcodec libnpp"
|
||||
overlay_cuda_filter_deps="ffnvcodec"
|
||||
overlay_cuda_filter_deps_any="cuda_nvcc cuda_llvm"
|
||||
cuda_grid_filter_deps="ffnvcodec libnpp"
|
||||
cuda_grid_filter_deps="ffnvcodec"
|
||||
sharpen_npp_filter_deps="ffnvcodec libnpp"
|
||||
|
||||
ddagrab_filter_deps="d3d11va IDXGIOutput1 DXGI_OUTDUPL_FRAME_INFO"
|
||||
|
||||
+45
-71
@@ -4,13 +4,25 @@
|
||||
* Принимает N CUDA-frames на входе, выдаёт один composed frame с N cells
|
||||
* в layout. End-to-end CUDA (без CPU round-trip).
|
||||
*
|
||||
* Phase 2a: layout templates (single/dual_h/dual_v/quad/main_plus_preview/
|
||||
* Phase 2: layout templates (single/dual_h/dual_v/quad/main_plus_preview/
|
||||
* six_grid/nine_grid/sixteen_grid/panoramic), dynamic nb_inputs, output size
|
||||
* через option (default 1920×1080). Cell rects = normalized × output size.
|
||||
* **Scaling пока нет** — каждый input должен быть точно cell size (Phase 2b NPP).
|
||||
*
|
||||
* **Scaling delegated to upstream `scale_npp`** filter (Unix philosophy +
|
||||
* production-tested NPP code). NPP не имеет nppiResize_8u_C2R для NV12 UV
|
||||
* interleaved, поэтому in-filter scaling = either two intermediate plane
|
||||
* buffers либо custom CUDA kernel — оба больше work чем filter chain'ить:
|
||||
*
|
||||
* ffmpeg ... -filter_complex \
|
||||
* "[0]scale_npp=1280:1080[s0]; \
|
||||
* [1]scale_npp=640:360[s1]; \
|
||||
* [2]scale_npp=640:360[s2]; \
|
||||
* [3]scale_npp=640:360[s3]; \
|
||||
* [s0][s1][s2][s3]cuda_grid=layout=main_plus_preview[out]"
|
||||
*
|
||||
* Controller (Phase 3) auto-generates filter graph с scale_npp per input.
|
||||
*
|
||||
* Future phases (см. gx/vf-cuda-grid#1):
|
||||
* - Phase 2b: per-cell scaling через libnpp (mixed-size inputs)
|
||||
* - Phase 3: runtime layout switching через process_command (ZMQ)
|
||||
* - Phase 4+: overlay primitives (rect/text/icon/image/dim/graph/chat)
|
||||
*
|
||||
@@ -19,8 +31,6 @@
|
||||
|
||||
#include "config_components.h"
|
||||
|
||||
#include <nppi.h>
|
||||
|
||||
#include "libavutil/avstring.h"
|
||||
#include "libavutil/common.h"
|
||||
#include "libavutil/cuda_check.h"
|
||||
@@ -215,9 +225,6 @@ static int cuda_grid_compose(FFFrameSync *fs)
|
||||
if (ret < 0)
|
||||
goto fail;
|
||||
|
||||
/* NPP в этом thread'е работает в нашем CUDA stream */
|
||||
nppSetStream(s->cu_stream);
|
||||
|
||||
for (i = 0; i < nb; i++) {
|
||||
AVFrame *src = in[i];
|
||||
int cx = s->cell_px[i].x;
|
||||
@@ -225,71 +232,38 @@ static int cuda_grid_compose(FFFrameSync *fs)
|
||||
int cw = s->cell_px[i].w;
|
||||
int ch = s->cell_px[i].h;
|
||||
|
||||
if (src->width == cw && src->height == ch) {
|
||||
/* Fast path: same size — memcpy без NPP overhead */
|
||||
ret = copy_input_plane(ctx,
|
||||
(CUdeviceptr)src->data[0], src->linesize[0],
|
||||
src->width, src->height,
|
||||
(CUdeviceptr)out->data[0], out->linesize[0],
|
||||
cx, cy, 1);
|
||||
if (ret < 0) {
|
||||
CHECK_CU(s->hwctx->internal->cuda_dl->cuCtxPopCurrent(&dummy));
|
||||
goto fail;
|
||||
}
|
||||
ret = copy_input_plane(ctx,
|
||||
(CUdeviceptr)src->data[1], src->linesize[1],
|
||||
src->width / 2, src->height / 2,
|
||||
(CUdeviceptr)out->data[1], out->linesize[1],
|
||||
cx / 2, cy / 2, 2);
|
||||
if (ret < 0) {
|
||||
CHECK_CU(s->hwctx->internal->cuda_dl->cuCtxPopCurrent(&dummy));
|
||||
goto fail;
|
||||
}
|
||||
} else {
|
||||
/* Phase 2b: NPP scaling. Output обязательно chroma-aligned (cell coords
|
||||
* выровнены до 2 в config_output). */
|
||||
NppStatus npp_err;
|
||||
double xfactor_y = (double)cw / src->width;
|
||||
double yfactor_y = (double)ch / src->height;
|
||||
double xfactor_uv = (double)(cw / 2) / (src->width / 2);
|
||||
double yfactor_uv = (double)(ch / 2) / (src->height / 2);
|
||||
if (src->width != cw || src->height != ch) {
|
||||
av_log(ctx, AV_LOG_ERROR,
|
||||
"input %d size %dx%d != cell size %dx%d. "
|
||||
"cuda_grid не делает scaling — используй upstream scale_npp:\n"
|
||||
" [in%d]scale_npp=%d:%d[scaled%d]; [scaled%d]...cuda_grid=layout=%s\n",
|
||||
i, src->width, src->height, cw, ch,
|
||||
i, cw, ch, i, i, s->layout->name);
|
||||
CHECK_CU(s->hwctx->internal->cuda_dl->cuCtxPopCurrent(&dummy));
|
||||
ret = AVERROR(EINVAL);
|
||||
goto fail;
|
||||
}
|
||||
|
||||
uint8_t *dst_y_ptr = out->data[0] + (size_t)cy * out->linesize[0] + cx;
|
||||
uint8_t *dst_uv_ptr = out->data[1] + (size_t)(cy / 2) * out->linesize[1] + cx;
|
||||
/* dst_uv X в bytes = cx (2 bytes per UV-pair × cx/2 = cx bytes) */
|
||||
/* Y plane (1 byte per pixel) */
|
||||
ret = copy_input_plane(ctx,
|
||||
(CUdeviceptr)src->data[0], src->linesize[0],
|
||||
src->width, src->height,
|
||||
(CUdeviceptr)out->data[0], out->linesize[0],
|
||||
cx, cy, 1);
|
||||
if (ret < 0) {
|
||||
CHECK_CU(s->hwctx->internal->cuda_dl->cuCtxPopCurrent(&dummy));
|
||||
goto fail;
|
||||
}
|
||||
|
||||
/* Y plane — 1 channel (luma) */
|
||||
npp_err = nppiResizeSqrPixel_8u_C1R(
|
||||
src->data[0], (NppiSize){src->width, src->height},
|
||||
src->linesize[0], (NppiRect){0, 0, src->width, src->height},
|
||||
dst_y_ptr, out->linesize[0],
|
||||
(NppiRect){0, 0, cw, ch},
|
||||
xfactor_y, yfactor_y, 0.0, 0.0,
|
||||
NPPI_INTER_LINEAR);
|
||||
if (npp_err != NPP_SUCCESS) {
|
||||
av_log(ctx, AV_LOG_ERROR,
|
||||
"input %d Y plane NPP resize %dx%d→%dx%d failed: %d\n",
|
||||
i, src->width, src->height, cw, ch, npp_err);
|
||||
CHECK_CU(s->hwctx->internal->cuda_dl->cuCtxPopCurrent(&dummy));
|
||||
ret = AVERROR_EXTERNAL;
|
||||
goto fail;
|
||||
}
|
||||
|
||||
/* UV plane — 2 channels interleaved (NV12 chroma) */
|
||||
npp_err = nppiResizeSqrPixel_8u_C2R(
|
||||
src->data[1], (NppiSize){src->width / 2, src->height / 2},
|
||||
src->linesize[1], (NppiRect){0, 0, src->width / 2, src->height / 2},
|
||||
dst_uv_ptr, out->linesize[1],
|
||||
(NppiRect){0, 0, cw / 2, ch / 2},
|
||||
xfactor_uv, yfactor_uv, 0.0, 0.0,
|
||||
NPPI_INTER_LINEAR);
|
||||
if (npp_err != NPP_SUCCESS) {
|
||||
av_log(ctx, AV_LOG_ERROR,
|
||||
"input %d UV plane NPP resize failed: %d\n", i, npp_err);
|
||||
CHECK_CU(s->hwctx->internal->cuda_dl->cuCtxPopCurrent(&dummy));
|
||||
ret = AVERROR_EXTERNAL;
|
||||
goto fail;
|
||||
}
|
||||
/* UV plane (NV12: half resolution, 2 bytes per "pixel") */
|
||||
ret = copy_input_plane(ctx,
|
||||
(CUdeviceptr)src->data[1], src->linesize[1],
|
||||
src->width / 2, src->height / 2,
|
||||
(CUdeviceptr)out->data[1], out->linesize[1],
|
||||
cx / 2, cy / 2, 2);
|
||||
if (ret < 0) {
|
||||
CHECK_CU(s->hwctx->internal->cuda_dl->cuCtxPopCurrent(&dummy));
|
||||
goto fail;
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user