[Mesa-dev] [PATCH 2/4] i965/tiled_memcpy: Move SSSE3 code back into inline functions.

[Brian Paul]

On 04/11/2016 05:17 PM, Roland Scheidegger wrote:
> Am 11.04.2016 um 21:05 schrieb Matt Turner:
>> This will make adding SSE2 code a lot cleaner.
>> ---
>>   src/mesa/drivers/dri/i965/intel_tiled_memcpy.c | 42 +++++++++++++++-----------
>>   1 file changed, 24 insertions(+), 18 deletions(-)
>>
>> diff --git a/src/mesa/drivers/dri/i965/intel_tiled_memcpy.c b/src/mesa/drivers/dri/i965/intel_tiled_memcpy.c
>> index fa5ec75..5d58530 100644
>> --- a/src/mesa/drivers/dri/i965/intel_tiled_memcpy.c
>> +++ b/src/mesa/drivers/dri/i965/intel_tiled_memcpy.c
>> @@ -85,6 +85,22 @@ rgba8_copy(void *dst, const void *src, size_t bytes)
>>   #ifdef __SSSE3__
>>   static const uint8_t rgba8_permutation[16] =
>>      { 2,1,0,3, 6,5,4,7, 10,9,8,11, 14,13,12,15 };
>> +
>> +static inline void
>> +rgba8_copy_16_aligned_dst(void *dst, const void *src)
>> +{
>> +   _mm_store_si128(dst,
>> +                   _mm_shuffle_epi8(_mm_loadu_si128(src),
>> +                                    *(__m128i *)rgba8_permutation));
>> +}
>> +
>> +static inline void
>> +rgba8_copy_16_aligned_src(void *dst, const void *src)
>> +{
>> +   _mm_storeu_si128(dst,
>> +                    _mm_shuffle_epi8(_mm_load_si128(src),
>> +                                     *(__m128i *)rgba8_permutation));
>> +}
>>   #endif
>>
>>   /**
>> @@ -93,23 +109,18 @@ static const uint8_t rgba8_permutation[16] =
>>   static inline void *
>>   rgba8_copy_aligned_dst(void *dst, const void *src, size_t bytes)
>>   {
>> -   uint8_t *d = dst;
>> -   uint8_t const *s = src;
>> -
>>      assert(bytes == 0 || !(((uintptr_t)dst) & 0xf));
>>
>>   #ifdef __SSSE3__
>>      while (bytes >= 16) {
>> -      _mm_store_si128((__m128i *)d,
>> -                      _mm_shuffle_epi8(_mm_loadu_si128((__m128i *)s),
>> -                                       *(__m128i *) rgba8_permutation));
>> -      s += 16;
>> -      d += 16;
>> +      rgba8_copy_16_aligned_dst(dst, src);
>> +      src += 16;
>> +      dst += 16;
>>         bytes -= 16;
>>      }
>>   #endif
>>
>> -   rgba8_copy(d, s, bytes);
>> +   rgba8_copy(dst, src, bytes);
>>
>>      return dst;
>>   }
>> @@ -120,23 +131,18 @@ rgba8_copy_aligned_dst(void *dst, const void *src, size_t bytes)
>>   static inline void *
>>   rgba8_copy_aligned_src(void *dst, const void *src, size_t bytes)
>>   {
>> -   uint8_t *d = dst;
>> -   uint8_t const *s = src;
>> -
>>      assert(bytes == 0 || !(((uintptr_t)src) & 0xf));
>>
>>   #ifdef __SSSE3__
>>      while (bytes >= 16) {
>> -      _mm_storeu_si128((__m128i *)d,
>> -                       _mm_shuffle_epi8(_mm_load_si128((__m128i *)s),
>> -                                        *(__m128i *) rgba8_permutation));
>> -      s += 16;
>> -      d += 16;
>> +      rgba8_copy_16_aligned_src(dst, src);
>> +      src += 16;
>> +      dst += 16;
>>         bytes -= 16;
>>      }
>>   #endif
>>
>> -   rgba8_copy(d, s, bytes);
>> +   rgba8_copy(dst, src, bytes);
>>
>>      return dst;
>>   }
>>
>
> I thought void ptr arithmetic is illegal (though a gcc extension), so
> maybe should avoid that?

I'm pretty sure MSVC would complain about that, but we don't build this 
driver on Windows, obviously.

-Brian