[Mesa-dev] [PATCH 6/9] radeonsi: don't count fast clears and prefetches into CP DMA stats

[Marek Olšák]

On Mon, Jan 2, 2017 at 11:17 PM, Bas Nieuwenhuizen
<bas at basnieuwenhuizen.nl> wrote:
> On Mon, Jan 2, 2017 at 9:17 PM, Marek Olšák <maraeo at gmail.com> wrote:
>> From: Marek Olšák <marek.olsak at amd.com>
>>
>> ---
>>  src/gallium/drivers/radeonsi/si_cp_dma.c | 8 ++++++--
>>  1 file changed, 6 insertions(+), 2 deletions(-)
>>
>> diff --git a/src/gallium/drivers/radeonsi/si_cp_dma.c b/src/gallium/drivers/radeonsi/si_cp_dma.c
>> index 5809317..fd772c5 100644
>> --- a/src/gallium/drivers/radeonsi/si_cp_dma.c
>> +++ b/src/gallium/drivers/radeonsi/si_cp_dma.c
>> @@ -208,21 +208,23 @@ static void si_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
>>                 /* Emit the clear packet. */
>>                 si_emit_cp_dma(sctx, va, value, byte_count, dma_flags, coher);
>>
>>                 size -= byte_count;
>>                 va += byte_count;
>>         }
>>
>>         if (tc_l2_flag)
>>                 r600_resource(dst)->TC_L2_dirty = true;
>>
>> -       sctx->b.num_cp_dma_calls++;
>> +       /* If it's not a framebuffer fast clear... */
>> +       if (coher == R600_COHERENCY_SHADER)
>> +               sctx->b.num_cp_dma_calls++;
>>  }
>>
>>  /**
>>   * Realign the CP DMA engine. This must be done after a copy with an unaligned
>>   * size.
>>   *
>>   * \param size  Remaining size to the CP DMA alignment.
>>   */
>>  static void si_cp_dma_realign_engine(struct si_context *sctx, unsigned size)
>>  {
>> @@ -334,17 +336,19 @@ void si_copy_buffer(struct si_context *sctx,
>>                                dma_flags, R600_COHERENCY_SHADER);
>>         }
>>
>>         /* Finally, realign the engine if the size wasn't aligned. */
>>         if (realign_size)
>>                 si_cp_dma_realign_engine(sctx, realign_size);
>>
>>         if (tc_l2_flag)
>>                 r600_resource(dst)->TC_L2_dirty = true;
>>
>> -       sctx->b.num_cp_dma_calls++;
>> +       /* If it's not a prefetch... */
>> +       if (dst_offset != src_offset)
>> +               sctx->b.num_cp_dma_calls++;
>
> And src != dst ?

Not needed. The variables contain full GPUVM addresses.

Marek