[Mesa-dev] [PATCH] radeonsi: add basic glClearBufferSubData acceleration

[Marek Olšák]

On Thu, Nov 5, 2015 at 7:58 PM, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
> On 04.11.2015 00:47, Marek Olšák wrote:
>>
>> From: Marek Olšák <marek.olsak at amd.com>
>>
>> ---
>>   src/gallium/drivers/radeonsi/si_blit.c | 55
>> ++++++++++++++++++++++++++++++++++
>>   1 file changed, 55 insertions(+)
>>
>> diff --git a/src/gallium/drivers/radeonsi/si_blit.c
>> b/src/gallium/drivers/radeonsi/si_blit.c
>> index fce014a..e934146 100644
>> --- a/src/gallium/drivers/radeonsi/si_blit.c
>> +++ b/src/gallium/drivers/radeonsi/si_blit.c
>> @@ -731,9 +731,64 @@ static void si_flush_resource(struct pipe_context
>> *ctx,
>>         }
>>   }
>>
>> +static void si_pipe_clear_buffer(struct pipe_context *ctx,
>> +                                struct pipe_resource *dst,
>> +                                unsigned offset, unsigned size,
>> +                                const void *clear_value,
>> +                                int clear_value_size)
>> +{
>> +       struct si_context *sctx = (struct si_context*)ctx;
>> +       const uint32_t *u32 = clear_value;
>> +       unsigned i;
>> +       bool clear_value_fits_dword = true;
>> +       uint8_t *map;
>> +
>> +       if (clear_value_size > 4)
>> +               for (i = 1; i < clear_value_size / 4; i++)
>> +                       if (u32[0] != u32[i]) {
>> +                               clear_value_fits_dword = false;
>> +                               break;
>> +                       }
>> +
>> +       /* Use CP DMA for the simple case. */
>> +       if (offset % 4 == 0 && size % 4 == 0 && clear_value_fits_dword) {
>> +               uint32_t value = u32[0];
>> +
>> +               switch (clear_value_size) {
>> +               case 1:
>> +                       value &= 0xff;
>> +                       value |= (value << 8) | (value << 16) | (value <<
>> 24);
>> +                       break;
>> +               case 2:
>> +                       value &= 0xffff;
>> +                       value |= value << 16;
>> +                       break;
>> +               }
>
>
> To reduce the chance of complaints by valgrind et al:
>
> switch (clear_value_size) {
> case 1:
>         value = *(uint8_t *)u32;
>         value |= (value << 8) | (value << 16) | (value << 24);
>         break;
> case 2:
>         value = *(uint16_t *)u32;
>         value |= value << 16;
>         break;
> default:
>         value = *u32;
>         break;
> }

Thanks.

The preliminary plan is to use transform feedback for fills>=64 bits
(already implemented by u_blitter), and CP DMA should be used for
32-bit fills and any fills that can be lowered to 32-bit.

Unaligned 8-bit and 16-bit fills should fill the largest aligned
subrange using CP DMA. Then, the unaligned beginning and ending dwords
will be filled separately using:
COPY_DATA from mem to reg
REG_RMW to fill the requested bytes
COPY_DATA from reg to mem

Marek