Welcome to mirror list, hosted at ThFree Co, Russian Federation.

github.com/FFmpeg/FFmpeg.git - Unnamed repository; edit this file 'description' to name the repository.
summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorTimo Rothenpieler <timo@rothenpieler.org>2020-11-04 02:19:55 +0300
committerTimo Rothenpieler <timo@rothenpieler.org>2020-11-04 03:36:39 +0300
commit2f3cb5cccab51c53f7440760fbd2914edbea801b (patch)
treee5a22bc9d4eda39c080f0b42d27f6fd22bba765a /libavutil/hwcontext_cuda.c
parent755203c2d270e1334987b462f80151eb88e30121 (diff)
avutil/hwcontext_cuda: query correct alignment from device
Diffstat (limited to 'libavutil/hwcontext_cuda.c')
-rw-r--r--libavutil/hwcontext_cuda.c29
1 files changed, 23 insertions, 6 deletions
diff --git a/libavutil/hwcontext_cuda.c b/libavutil/hwcontext_cuda.c
index 546515ee66..7278bd2988 100644
--- a/libavutil/hwcontext_cuda.c
+++ b/libavutil/hwcontext_cuda.c
@@ -30,10 +30,9 @@
#include "pixfmt.h"
#include "imgutils.h"
-#define CUDA_FRAME_ALIGNMENT 512
-
typedef struct CUDAFramesContext {
int shift_width, shift_height;
+ int tex_alignment;
} CUDAFramesContext;
static const enum AVPixelFormat supported_formats[] = {
@@ -127,8 +126,11 @@ fail:
static int cuda_frames_init(AVHWFramesContext *ctx)
{
- CUDAFramesContext *priv = ctx->internal->priv;
- int i;
+ AVHWDeviceContext *device_ctx = ctx->device_ctx;
+ AVCUDADeviceContext *hwctx = device_ctx->hwctx;
+ CUDAFramesContext *priv = ctx->internal->priv;
+ CudaFunctions *cu = hwctx->internal->cuda_dl;
+ int err, i;
for (i = 0; i < FF_ARRAY_ELEMS(supported_formats); i++) {
if (ctx->sw_format == supported_formats[i])
@@ -140,10 +142,24 @@ static int cuda_frames_init(AVHWFramesContext *ctx)
return AVERROR(ENOSYS);
}
+ err = CHECK_CU(cu->cuDeviceGetAttribute(&priv->tex_alignment,
+ 14 /* CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT */,
+ hwctx->internal->cuda_device));
+ if (err < 0)
+ return err;
+
+ av_log(ctx, AV_LOG_DEBUG, "CUDA texture alignment: %d\n", priv->tex_alignment);
+
+ // YUV420P is a special case.
+ // Since nvenc expects the U/V planes to have half the linesize of the Y plane
+ // alignment has to be doubled to ensure the U/V planes still end up aligned.
+ if (ctx->sw_format == AV_PIX_FMT_YUV420P)
+ priv->tex_alignment *= 2;
+
av_pix_fmt_get_chroma_sub_sample(ctx->sw_format, &priv->shift_width, &priv->shift_height);
if (!ctx->pool) {
- int size = av_image_get_buffer_size(ctx->sw_format, ctx->width, ctx->height, CUDA_FRAME_ALIGNMENT);
+ int size = av_image_get_buffer_size(ctx->sw_format, ctx->width, ctx->height, priv->tex_alignment);
if (size < 0)
return size;
@@ -157,6 +173,7 @@ static int cuda_frames_init(AVHWFramesContext *ctx)
static int cuda_get_buffer(AVHWFramesContext *ctx, AVFrame *frame)
{
+ CUDAFramesContext *priv = ctx->internal->priv;
int res;
frame->buf[0] = av_buffer_pool_get(ctx->pool);
@@ -164,7 +181,7 @@ static int cuda_get_buffer(AVHWFramesContext *ctx, AVFrame *frame)
return AVERROR(ENOMEM);
res = av_image_fill_arrays(frame->data, frame->linesize, frame->buf[0]->data,
- ctx->sw_format, ctx->width, ctx->height, CUDA_FRAME_ALIGNMENT);
+ ctx->sw_format, ctx->width, ctx->height, priv->tex_alignment);
if (res < 0)
return res;