[FFmpeg-devel] [PATCH 2/2] libswscale: Test AV_CPU_FLAG_SLOW_GATHER for hscale functions.

[James Almer]

On 12/20/2021 10:57 AM, Alan Kelly wrote:
> This is instead of EXTERNAL_AVX2_FAST so that the avx2 hscale functions
> are only used where they are faster.
> ---
>   libswscale/utils.c        | 2 +-
>   libswscale/x86/swscale.c  | 2 +-
>   tests/checkasm/sw_scale.c | 2 +-
>   3 files changed, 3 insertions(+), 3 deletions(-)
> 
> diff --git a/libswscale/utils.c b/libswscale/utils.c
> index d4a72d3ce1..9a69b45afe 100644
> --- a/libswscale/utils.c
> +++ b/libswscale/utils.c
> @@ -282,7 +282,7 @@ void ff_shuffle_filter_coefficients(SwsContext *c, int *filterPos, int filterSiz
>   #if ARCH_X86_64
>       int i, j, k, l;
>       int cpu_flags = av_get_cpu_flags();
> -    if (EXTERNAL_AVX2_FAST(cpu_flags)){
> +    if (cpu_flags & AV_CPU_FLAG_SLOW_GATHER) {

The other way around. Like this you're enabling these functions only on 
CPUs with slow gather.

The check needs to be if (EXTERNAL_AVX2_FAST(cpu_flags) && !(cpu_flags & 
AV_CPU_FLAG_SLOW_GATHER))

To ensure it's run on AVX2 enabled CPUs with both fast 256bits ops and 
fast gather.

Same below.

>           if ((c->srcBpc == 8) && (c->dstBpc <= 14)){
>               if (dstW % 16 == 0){
>                   if (filter != NULL){
> diff --git a/libswscale/x86/swscale.c b/libswscale/x86/swscale.c
> index c49a05c37b..eb5334a2be 100644
> --- a/libswscale/x86/swscale.c
> +++ b/libswscale/x86/swscale.c
> @@ -578,7 +578,7 @@ switch(c->dstBpc){ \
>                break; \
>       }
>   
> -    if (EXTERNAL_AVX2_FAST(cpu_flags)) {
> +    if (cpu_flags & AV_CPU_FLAG_SLOW_GATHER) {
>           if ((c->srcBpc == 8) && (c->dstBpc <= 14)) {
>               if (c->chrDstW % 16 == 0)
>                   ASSIGN_AVX2_SCALE_FUNC(c->hcScale, c->hChrFilterSize);
> diff --git a/tests/checkasm/sw_scale.c b/tests/checkasm/sw_scale.c
> index f4912e6c2c..680562af08 100644
> --- a/tests/checkasm/sw_scale.c
> +++ b/tests/checkasm/sw_scale.c
> @@ -217,7 +217,7 @@ static void check_hscale(void)
>               }
>               ff_sws_init_scale(ctx);
>               memcpy(filterAvx2, filter, sizeof(uint16_t) * (SRC_PIXELS * MAX_FILTER_WIDTH + MAX_FILTER_WIDTH));
> -            if (cpu_flags & AV_CPU_FLAG_AVX2)
> +            if (cpu_flags & AV_CPU_FLAG_SLOW_GATHER)
>                   ff_shuffle_filter_coefficients(ctx, filterPosAvx, width, filterAvx2, SRC_PIXELS);
>   
>               if (check_func(ctx->hcScale, "hscale_%d_to_%d_width%d", ctx->srcBpc, ctx->dstBpc + 1, width)) {