[FFmpeg-devel] [PATCH] avutil/hwcontext_cuda: add AVCUDAFramesContext and AVCUDAFramesContext.flags

[Mark Thompson]

On 07/05/18 22:10, Timo Rothenpieler wrote:
> Frames can be mapped from nvdec/cuvid, not needing any actual memory
> allocation, but all other features of the hw_frames_ctx.
> Hence the dummy-mode, which does not allocate any (notable amounts of)
> memory but otherwise behaves the exact same.
> ---
>  doc/APIchanges             |  3 +++
>  libavutil/hwcontext_cuda.c | 10 ++++++++++
>  libavutil/hwcontext_cuda.h | 18 +++++++++++++++++-
>  libavutil/version.h        |  2 +-
>  4 files changed, 31 insertions(+), 2 deletions(-)
> 
> diff --git a/doc/APIchanges b/doc/APIchanges
> index ede5b186ae..82ec888fd8 100644
> --- a/doc/APIchanges
> +++ b/doc/APIchanges
> @@ -15,6 +15,9 @@ libavutil:     2017-10-21
>  
>  API changes, most recent first:
>  
> +2018-05-xx - xxxxxxxxxx - lavu 56.19.100 - hwcontext.h
> +  Add AVCUDAFramesContext and AVCUDAFramesContext.flags.
> +
>  2018-04-xx - xxxxxxxxxx - lavu 56.18.100 - pixdesc.h
>    Add AV_PIX_FMT_FLAG_ALPHA to AV_PIX_FMT_PAL8.
>  
> diff --git a/libavutil/hwcontext_cuda.c b/libavutil/hwcontext_cuda.c
> index 37827a770c..0d867ef0f5 100644
> --- a/libavutil/hwcontext_cuda.c
> +++ b/libavutil/hwcontext_cuda.c
> @@ -83,6 +83,7 @@ static void cuda_buffer_free(void *opaque, uint8_t *data)
>  static AVBufferRef *cuda_pool_alloc(void *opaque, int size)
>  {
>      AVHWFramesContext     *ctx = opaque;
> +    AVCUDAFramesContext *frctx = ctx->hwctx;
>      AVCUDADeviceContext *hwctx = ctx->device_ctx->hwctx;
>      CudaFunctions          *cu = hwctx->internal->cuda_dl;
>  
> @@ -97,6 +98,10 @@ static AVBufferRef *cuda_pool_alloc(void *opaque, int size)
>          return NULL;
>      }
>  
> +    // A lot of places expect the pointer to be !=NULL, so make minimum allocation instead.
> +    if (frctx->flags & AV_CUDA_HWFRAMES_DUMMY_MODE)
> +        size = 1;

What are those places - can we get rid of them instead?  Allocating a single byte here is rather yucky.

> +
>      err = cu->cuMemAlloc(&data, size);
>      if (err != CUDA_SUCCESS)
>          goto fail;
> @@ -161,6 +166,7 @@ static int cuda_frames_init(AVHWFramesContext *ctx)
>  
>  static int cuda_get_buffer(AVHWFramesContext *ctx, AVFrame *frame)
>  {
> +    AVCUDAFramesContext *frctx = ctx->hwctx;
>      int aligned_width;
>      int width_in_bytes = ctx->width;
>  
> @@ -210,6 +216,9 @@ static int cuda_get_buffer(AVHWFramesContext *ctx, AVFrame *frame)
>      frame->width  = ctx->width;
>      frame->height = ctx->height;
>  
> +    if (frctx->flags & AV_CUDA_HWFRAMES_DUMMY_MODE)
> +        frame->data[0] = frame->data[1] = frame->data[2] = NULL;

Are you intending linesize and buf[0] to have been filled with specific values here?

A comment saying exactly what is set and what isn't might help.  (Maybe the comment goes with the flag itself.)

> +
>      return 0;
>  }
>  
> @@ -402,6 +411,7 @@ const HWContextType ff_hwcontext_type_cuda = {
>      .name                 = "CUDA",
>  
>      .device_hwctx_size    = sizeof(AVCUDADeviceContext),
> +    .frames_hwctx_size    = sizeof(AVCUDAFramesContext),
>      .frames_priv_size     = sizeof(CUDAFramesContext),
>  
>      .device_create        = cuda_device_create,
> diff --git a/libavutil/hwcontext_cuda.h b/libavutil/hwcontext_cuda.h
> index 12dae8449e..388d6f8f1c 100644
> --- a/libavutil/hwcontext_cuda.h
> +++ b/libavutil/hwcontext_cuda.h
> @@ -45,7 +45,23 @@ typedef struct AVCUDADeviceContext {
>  } AVCUDADeviceContext;
>  
>  /**
> - * AVHWFramesContext.hwctx is currently not used
> + * This struct is allocated as AVHWFramesContext.hwctx
>   */
> +typedef struct AVCUDAFramesContext {
> +    /**
> +     * Special implementation-specific flags.
> +     *
> +     * Must be set by the user before calling av_hwframe_ctx_init().
> +     */

This makes it sound like you /must/ write to the field.  Leaving it as zero is also fine.

> +    int flags;
> +} AVCUDAFramesContext;
> +
> +/**
> + * No actual allocation will happen, but otherwise behaves like normal.
> + *
> + * This is to be used if a AVHWFramesContext is required, but the actual
> + * allocation has to happen outside of it.
> + */
> +#define AV_CUDA_HWFRAMES_DUMMY_MODE (1 << 0)
>  
>  #endif /* AVUTIL_HWCONTEXT_CUDA_H */
> diff --git a/libavutil/version.h b/libavutil/version.h
> index 5185454d9b..84409b1d69 100644
> --- a/libavutil/version.h
> +++ b/libavutil/version.h
> @@ -79,7 +79,7 @@
>   */
>  
>  #define LIBAVUTIL_VERSION_MAJOR  56
> -#define LIBAVUTIL_VERSION_MINOR  18
> +#define LIBAVUTIL_VERSION_MINOR  19
>  #define LIBAVUTIL_VERSION_MICRO 100
>  
>  #define LIBAVUTIL_VERSION_INT   AV_VERSION_INT(LIBAVUTIL_VERSION_MAJOR, \
> 

This approach seems ok to me.

Thanks,

- Mark