[Mesa-dev] [PATCH 4/4] radeonsi: add a workaround for inexact SNORM8 blitting again

[Marek Olšák]

On Mon, Jun 19, 2017 at 2:56 PM, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
> On 17.06.2017 15:44, Marek Olšák wrote:
>>
>> From: Marek Olšák <marek.olsak at amd.com>
>>
>> GFX9 is affected.
>> ---
>>   src/gallium/drivers/radeonsi/si_blit.c | 36
>> ++++++++++++++++++++++++++++++++++
>>   1 file changed, 36 insertions(+)
>>
>> diff --git a/src/gallium/drivers/radeonsi/si_blit.c
>> b/src/gallium/drivers/radeonsi/si_blit.c
>> index 1159594..b78fddf 100644
>> --- a/src/gallium/drivers/radeonsi/si_blit.c
>> +++ b/src/gallium/drivers/radeonsi/si_blit.c
>> @@ -1133,20 +1133,56 @@ void si_resource_copy_region(struct pipe_context
>> *ctx,
>>                                 src_templ.format =
>> PIPE_FORMAT_R32G32B32A32_UINT;
>>                                 break;
>>                         default:
>>                                 fprintf(stderr, "Unhandled format %s with
>> blocksize %u\n",
>>
>> util_format_short_name(src->format), blocksize);
>>                                 assert(0);
>>                         }
>>                 }
>>         }
>>   +     /* SNORM8 blitting has precision issues on some chips. Use the
>> SINT
>> +        * equivalent instead, which doesn't force DCC decompression.
>> +        * Note that some chips avoid this issue by using SDMA.
>> +        */
>> +       if (util_format_is_snorm8(dst_templ.format)) {
>> +               switch (dst_templ.format) {
>> +               case PIPE_FORMAT_R8_SNORM:
>> +                       dst_templ.format = src_templ.format =
>> PIPE_FORMAT_R8_SINT;
>> +                       break;
>> +               case PIPE_FORMAT_R8G8_SNORM:
>> +                       dst_templ.format = src_templ.format =
>> PIPE_FORMAT_R8G8_SINT;
>> +                       break;
>> +               case PIPE_FORMAT_R8G8B8X8_SNORM:
>> +                       dst_templ.format = src_templ.format =
>> PIPE_FORMAT_R8G8B8X8_SINT;
>> +                       break;
>> +               case PIPE_FORMAT_R8G8B8A8_SNORM:
>> +               case PIPE_FORMAT_A8B8G8R8_SNORM: /* also swizzle */
>> +               case PIPE_FORMAT_X8B8G8R8_SNORM: /* also swizzle */
>> +                       dst_templ.format = src_templ.format =
>> PIPE_FORMAT_R8G8B8A8_SINT;
>
>
> Why does R8G8B8X8 need a separate case but not X8B8G8R8?

There are no SINT variants for ABGR and XBGR, so we have to use RGBA.

>
> Do we have a test for this?

We only have tests for GL_x_SNORM where x is R8, RG8, RGB8, and RGBA8.

Marek