[FFmpeg-devel] [PATCH 05/11] avcodec: add stride alignment needed for AVX-512

[James Darnley]

On 2017-11-10 02:38, James Almer wrote:
> On 11/9/2017 8:58 AM, James Darnley wrote:
>> ---
>>  configure             | 2 ++
>>  libavcodec/internal.h | 4 +++-
>>  2 files changed, 5 insertions(+), 1 deletion(-)
>>
>> diff --git a/configure b/configure
>> index 146a87324c..fce8030d91 100755
>> --- a/configure
>> +++ b/configure
>> @@ -1886,6 +1886,7 @@ ARCH_FEATURES="
>>      local_aligned
>>      simd_align_16
>>      simd_align_32
>> +    simd_align_64
>>  "
>>  
>>  BUILTIN_LIST="
>> @@ -2385,6 +2386,7 @@ fast_clz_if_any="aarch64 alpha avr32 mips ppc x86"
>>  fast_unaligned_if_any="aarch64 ppc x86"
>>  simd_align_16_if_any="altivec neon sse"
>>  simd_align_32_if_any="avx"
>> +simd_align_64_if_any="avx512"
>>  
>>  # system capabilities
>>  symver_if_any="symver_asm_label symver_gnu_asm"
>> diff --git a/libavcodec/internal.h b/libavcodec/internal.h
>> index 7748f09f54..84070431ed 100644
>> --- a/libavcodec/internal.h
>> +++ b/libavcodec/internal.h
>> @@ -87,7 +87,9 @@
>>  
>>  #define FF_SIGNBIT(x) ((x) >> CHAR_BIT * sizeof(x) - 1)
>>  
>> -#if HAVE_SIMD_ALIGN_32
>> +#if HAVE_SIMD_ALIGN_64
>> +#   define STRIDE_ALIGN 64 /* AVX-512 */
>> +#elif HAVE_SIMD_ALIGN_32
>>  #   define STRIDE_ALIGN 32
>>  #elif HAVE_SIMD_ALIGN_16
>>  #   define STRIDE_ALIGN 16
>>
> 
> LGTM, but I'd really like to find a way to start using
> av_get_cpu_max_align() as it was meant to be done when it was
> introduced, instead of keeping hardcoding alignment based on configure
> time options.

That sounds like a good idea to me.  If time allows I will try to start
a discussion about it.  It sounds ripe for a lot of bikeshedding :)

> Maybe adding a line to the av_force_cpu_flags() and
> av_set_cpu_flags_mask() doxy stating you should reinit all your decoder
> and encoder contexts after calling them or similar.

Did you want me to add that one of my patches?