[Mesa-dev] [PATCH] mesa: add special case ubyte[4] / BGRA conversion function

[Brian Paul]

On 02/18/2016 03:03 PM, Roland Scheidegger wrote:
> Looks ok to me, albeit the sse2 code I posted a while ago for handling
> swizzled stores probably would be faster still (though only with x86
> sse2...). At least I think it would hit this (if not the next patch in
> the series should...).

It would be great to have those optimizations.  This patch alone made a 
decent improvement for the app I was testing.

Honestly, I miss the old texstore / image conversion code.  It was easy 
to optimize hot spots.  This is kind of bolted on the side.

-Brian

> Matt Turner reviewed the first 2 patches in the series (albeit I had to
> revert the second one due to a crash bug in i965 which was hiding behind
> a __SSE3__ ifdef before...) but the others are unreviewed (honestly, I'm
> not fully convinced of the approach neither...).
>
> https://lists.freedesktop.org/archives/mesa-dev/2016-January/105116.html
>
> Roland
>
>
> Am 18.02.2016 um 16:52 schrieb Brian Paul:
>> This reduces a glTexImage(GL_RGBA, GL_UNSIGNED_BYTE) hot spot in when
>> storing the texture as BGRA.
>> ---
>>   src/mesa/main/format_utils.c | 74 +++++++++++++++++++++++++++++++++++++++++---
>>   1 file changed, 69 insertions(+), 5 deletions(-)
>>
>> diff --git a/src/mesa/main/format_utils.c b/src/mesa/main/format_utils.c
>> index 5fdabd5..d16d69c 100644
>> --- a/src/mesa/main/format_utils.c
>> +++ b/src/mesa/main/format_utils.c
>> @@ -179,6 +179,63 @@ _mesa_compute_rgba2base2rgba_component_mapping(GLenum baseFormat, uint8_t *map)
>>      }
>>   }
>>
>> +
>> +/**
>> + * Special case conversion function to swap r/b channels from the source
>> + * image to the dest image.
>> + */
>> +static void
>> +convert_ubyte_rgba_to_bgra(size_t width, size_t height,
>> +                           const uint8_t *src, size_t src_stride,
>> +                           uint8_t *dst, size_t dst_stride)
>> +{
>> +   int row;
>> +
>> +   if (sizeof(void *) == 8 &&
>> +       src_stride % 8 == 0 &&
>> +       dst_stride % 8 == 0 &&
>> +       (GLsizeiptr) src % 8 == 0 &&
>> +       (GLsizeiptr) dst % 8 == 0) {
>> +      /* use 64-bit word to swizzle two 32-bit pixels.  We need 8-byte
>> +       * alignment for src/dst addresses and strides.
>> +       */
>> +      for (row = 0; row < height; row++) {
>> +         const GLuint64 *s = (const GLuint64 *) src;
>> +         GLuint64 *d = (GLuint64 *) dst;
>> +         int i;
>> +         for (i = 0; i < width/2; i++) {
>> +            d[i] = ( (s[i] & 0xff00ff00ff00ff00) |
>> +                    ((s[i] &       0xff000000ff) << 16) |
>> +                    ((s[i] &   0xff000000ff0000) >> 16));
>> +         }
>> +         if (width & 1) {
>> +            /* handle the case of odd widths */
>> +            const GLuint s = ((const GLuint *) src)[width - 1];
>> +            GLuint *d = (GLuint *) dst + width - 1;
>> +            *d = ( (s & 0xff00ff00) |
>> +                  ((s &       0xff) << 16) |
>> +                  ((s &   0xff0000) >> 16));
>> +         }
>> +         src += src_stride;
>> +         dst += dst_stride;
>> +      }
>> +   } else {
>> +      for (row = 0; row < height; row++) {
>> +         const GLuint *s = (const GLuint *) src;
>> +         GLuint *d = (GLuint *) dst;
>> +         int i;
>> +         for (i = 0; i < width; i++) {
>> +            d[i] = ( (s[i] & 0xff00ff00) |
>> +                    ((s[i] &       0xff) << 16) |
>> +                    ((s[i] &   0xff0000) >> 16));
>> +         }
>> +         src += src_stride;
>> +         dst += dst_stride;
>> +      }
>> +   }
>> +}
>> +
>> +
>>   /**
>>    * This can be used to convert between most color formats.
>>    *
>> @@ -299,11 +356,18 @@ _mesa_format_convert(void *void_dst, uint32_t dst_format, size_t dst_stride,
>>               return;
>>            } else if (src_array_format == RGBA8_UBYTE) {
>>               assert(!_mesa_is_format_integer_color(dst_format));
>> -            for (row = 0; row < height; ++row) {
>> -               _mesa_pack_ubyte_rgba_row(dst_format, width,
>> -                                         (const uint8_t (*)[4])src, dst);
>> -               src += src_stride;
>> -               dst += dst_stride;
>> +
>> +            if (dst_format == MESA_FORMAT_B8G8R8A8_UNORM) {
>> +               convert_ubyte_rgba_to_bgra(width, height, src, src_stride,
>> +                                          dst, dst_stride);
>> +            }
>> +            else {
>> +               for (row = 0; row < height; ++row) {
>> +                  _mesa_pack_ubyte_rgba_row(dst_format, width,
>> +                                            (const uint8_t (*)[4])src, dst);
>> +                  src += src_stride;
>> +                  dst += dst_stride;
>> +               }
>>               }
>>               return;
>>            } else if (src_array_format == RGBA32_UINT &&
>>
>