[FFmpeg-devel] [PATCH] avfilter/vf_program_opencl: make it possible to pass plane as argument to kernel

[Paul B Mahol]

On 2/16/20, Mark Thompson <sw at jkqxz.net> wrote:
> On 14/02/2020 10:41, Paul B Mahol wrote:
>> Fixes #7190
>>
>> Signed-off-by: Paul B Mahol <onemda at gmail.com>
>> ---
>>  doc/filters.texi                |  4 ++++
>>  libavfilter/vf_program_opencl.c | 17 +++++++++++++++++
>>  2 files changed, 21 insertions(+)
>>
>> diff --git a/doc/filters.texi b/doc/filters.texi
>> index b957b9302e..5bd0e44b67 100644
>> --- a/doc/filters.texi
>> +++ b/doc/filters.texi
>> @@ -21297,6 +21297,8 @@ Number of inputs to the filter.  Defaults to 1.
>>  @item size, s
>>  Size of output frames.  Defaults to the same as the first input.
>>
>> + at item plane
>> +Currently processed plane number is set as last kernel argument.
>>  @end table
>>
>>  The program source file must contain a kernel function with the given
>> name,
>> @@ -22483,6 +22485,8 @@ Pixel format to use for the generated frames.
>> This must be set.
>>  @item rate, r
>>  Number of frames generated every second.  Default value is '25'.
>>
>> + at item plane
>> +Currently processed plane number is set as last kernel argument.
>>  @end table
>>
>>  For details of how the program loading works, see the
>> @ref{program_opencl}
>> diff --git a/libavfilter/vf_program_opencl.c
>> b/libavfilter/vf_program_opencl.c
>> index ec25e931f5..3cf93464ba 100644
>> --- a/libavfilter/vf_program_opencl.c
>> +++ b/libavfilter/vf_program_opencl.c
>> @@ -42,6 +42,7 @@ typedef struct ProgramOpenCLContext {
>>      const char         *source_file;
>>      const char         *kernel_name;
>>      int                 nb_inputs;
>> +    int                 plane_is_arg;
>>      int                 width, height;
>>      enum AVPixelFormat  source_format;
>>      AVRational          source_rate;
>> @@ -138,6 +139,16 @@ static int program_opencl_run(AVFilterContext *avctx)
>>              }
>>          }
>>
>> +        if (ctx->plane_is_arg) {
>> +            cle = clSetKernelArg(ctx->kernel, 2 + ctx->nb_inputs,
>> sizeof(cl_int), &plane);
>> +            if (cle != CL_SUCCESS) {
>> +                av_log(avctx, AV_LOG_ERROR, "Failed to set kernel "
>> +                       "plane argument %d: %d.\n", plane, cle);
>> +                err = AVERROR_UNKNOWN;
>> +                goto fail;
>> +            }
>> +        }
>> +
>>          err = ff_opencl_filter_work_size_from_image(avctx, global_work,
>>                                                      output, plane, 0);
>>          if (err < 0)
>> @@ -348,6 +359,9 @@ static const AVOption program_opencl_options[] = {
>>      { "s",      "Video size",       OFFSET(width),
>>        AV_OPT_TYPE_IMAGE_SIZE,       { .str = NULL }, 0, 0, FLAGS },
>>
>> +    { "plane",  "Pass currently processed plane as argument",
>> OFFSET(plane_is_arg),
>> +      AV_OPT_TYPE_BOOL,       { .i64 = 0 }, 0, 1, FLAGS },
>> +
>>      { NULL },
>>  };
>>
>> @@ -400,6 +414,9 @@ static const AVOption openclsrc_options[] = {
>>      { "r",      "Video frame rate", OFFSET(source_rate),
>>        AV_OPT_TYPE_VIDEO_RATE,       { .str = "25" }, 0, INT_MAX, FLAGS },
>>
>> +    { "plane",  "Pass currently processed plane as argument",
>> OFFSET(plane_is_arg),
>> +      AV_OPT_TYPE_BOOL,       { .i64 = 0 }, 0, 1, FLAGS },
>> +
>>      { NULL },
>>  };
>>
>>
>
> Intuitively this feels like the wrong way to do it to me - that's going to
> become a conditional inside the kernel, which is always going to go one way
> or the other on any particular run.  (Do you have a specific example in
> mind?)
>
> As an alternative to solve the same problem, suppose you could instead
> provide a list of kernel names which then get executed one for each plane?
> Extending the "kernel" option could do that - provide a comma-separated list
> of kernels, and then run the nth kernel in the list on the nth plane.
>

Agreed, seems better solution.