gx/ffmpeg-patched
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

vf_cuda_grid: Phase 2b — NPP scaling для mixed-size inputs

Browse Source 
- Add libnpp dependency в configure (cuda_grid_filter_deps)
- #include <nppi.h>, nppSetStream(s->cu_stream) перед resize batch
- Smart copy-or-scale в compose path:
  - src.size == cell.size → cuMemcpy2DAsync (fast path, zero overhead)
  - else → nppiResizeSqrPixel_8u_C1R для Y + _C2R для NV12 UV interleaved
- NPPI_INTER_LINEAR interpolation (bilinear — стандартный для video)
- Destination pointer offset через explicit pointer arithmetic
  (NPP не имеет dst_offset параметра, нужно сместить pSrc указатель)

Это unblock'ает mixed-size cameras (parking 1920x1080 + gate_lpr 2688x1520
в одном grid с main_plus_preview layout — big cell scaled до 1280x1080,
small cells scaled до 640x360).

Phase 2 complete (2a + 2b). Phase 3 будет controller sidecar.
This commit is contained in:
gx
2026-05-19 21:20:04 +01:00
parent df476472e2
commit 11f310061a
2 changed files with 69 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
configure
Vendored
+1 -1
View File
@@ -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"
cuda_grid_filter_deps="ffnvcodec libnpp"
sharpen_npp_filter_deps="ffnvcodec libnpp"
ddagrab_filter_deps="d3d11va IDXGIOutput1 DXGI_OUTDUPL_FRAME_INFO"
libavfilter/vf_cuda_grid.c
+68 -28
View File
@@ -19,6 +19,8 @@
#include "config_components.h"
#include <nppi.h>
#include "libavutil/avstring.h"
#include "libavutil/common.h"
#include "libavutil/cuda_check.h"
@@ -213,6 +215,9 @@ 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;
@@ -220,36 +225,71 @@ 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) {
av_log(ctx, AV_LOG_ERROR,
"input %d size %dx%d != cell size %dx%d "
"(Phase 2a: no scaling — Phase 2b добавит NPP resize)\n",
i, src->width, src->height, cw, ch);
CHECK_CU(s->hwctx->internal->cuda_dl->cuCtxPopCurrent(&dummy));
ret = AVERROR(EINVAL);
goto fail;
}
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);
/* Y plane */
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;
}
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) */
/* UV plane (NV12: половинное разрешение, 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;
/* 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;
}
}
}