[FFmpeg-devel] [PATCH v2 4/5] libswscale: Enable hscale_avx2 for all input sizes.

[Alan Kelly]

Pushing this back up to the top. This is required to enable the previous
patch in this chain. Thanks

On Fri, Apr 22, 2022 at 10:04 AM Alan Kelly <alankelly at google.com> wrote:

> Ping!
>
> On Thu, Feb 17, 2022 at 11:04 AM Alan Kelly <alankelly at google.com> wrote:
>
>> ff_shuffle_filter_coefficients shuffles the tail as required.
>> ---
>>  libswscale/utils.c       | 19 ++++++++++++++++---
>>  libswscale/x86/swscale.c |  6 ++----
>>  2 files changed, 18 insertions(+), 7 deletions(-)
>>
>> diff --git a/libswscale/utils.c b/libswscale/utils.c
>> index 7c8e1bbdde..d818c9ce55 100644
>> --- a/libswscale/utils.c
>> +++ b/libswscale/utils.c
>> @@ -285,8 +285,7 @@ int ff_shuffle_filter_coefficients(SwsContext *c, int
>> *filterPos,
>>  #if ARCH_X86_64
>>      int i, j, k;
>>      int cpu_flags = av_get_cpu_flags();
>> -    // avx2 hscale filter processes 16 pixel blocks.
>> -    if (!filter || dstW % 16 != 0)
>> +    if (!filter)
>>          return 0;
>>      if (EXTERNAL_AVX2_FAST(cpu_flags) && !(cpu_flags &
>> AV_CPU_FLAG_SLOW_GATHER)) {
>>          if ((c->srcBpc == 8) && (c->dstBpc <= 14)) {
>> @@ -298,9 +297,11 @@ int ff_shuffle_filter_coefficients(SwsContext *c,
>> int *filterPos,
>>             }
>>             // Do not swap filterPos for pixels which won't be processed
>> by
>>             // the main loop.
>> -           for (i = 0; i + 8 <= dstW; i += 8) {
>> +           for (i = 0; i + 16 <= dstW; i += 16) {
>>                 FFSWAP(int, filterPos[i + 2], filterPos[i + 4]);
>>                 FFSWAP(int, filterPos[i + 3], filterPos[i + 5]);
>> +               FFSWAP(int, filterPos[i + 10], filterPos[i + 12]);
>> +               FFSWAP(int, filterPos[i + 11], filterPos[i + 13]);
>>             }
>>             if (filterSize > 4) {
>>                 // 16 pixels are processed at a time.
>> @@ -314,6 +315,18 @@ int ff_shuffle_filter_coefficients(SwsContext *c,
>> int *filterPos,
>>                         }
>>                     }
>>                 }
>> +               // 4 pixels are processed at a time in the tail.
>> +               for (; i < dstW; i += 4) {
>> +                   // 4 filter coeffs are processed at a time.
>> +                   int rem = dstW - i >= 4 ? 4 : dstW - i;
>> +                   for (k = 0; k + 4 <= filterSize; k += 4) {
>> +                       for (j = 0; j < rem; ++j) {
>> +                           int from = (i + j) * filterSize + k;
>> +                           int to = i * filterSize + j * 4 + k * 4;
>> +                           memcpy(&filter[to], &filterCopy[from], 4 *
>> sizeof(int16_t));
>> +                       }
>> +                   }
>> +               }
>>             }
>>             av_free(filterCopy);
>>          }
>> diff --git a/libswscale/x86/swscale.c b/libswscale/x86/swscale.c
>> index 73869355b8..76f5a70fc5 100644
>> --- a/libswscale/x86/swscale.c
>> +++ b/libswscale/x86/swscale.c
>> @@ -691,10 +691,8 @@ switch(c->dstBpc){ \
>>
>>      if (EXTERNAL_AVX2_FAST(cpu_flags) && !(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);
>> -            if (c->dstW % 16 == 0)
>> -                ASSIGN_AVX2_SCALE_FUNC(c->hyScale, c->hLumFilterSize);
>> +            ASSIGN_AVX2_SCALE_FUNC(c->hcScale, c->hChrFilterSize);
>> +            ASSIGN_AVX2_SCALE_FUNC(c->hyScale, c->hLumFilterSize);
>>          }
>>      }
>>
>> --
>> 2.35.1.265.g69c8d7142f-goog
>>
>>