[Mesa-dev] [PATCH] radv/gfx10: use L2 for DMA copy/fill operations

[Bas Nieuwenhuizen]

r-b

though it sounds like some of our cache flushes might be not ideal.

On Thu, Jul 25, 2019 at 3:35 PM Samuel Pitoiset
<samuel.pitoiset at gmail.com> wrote:
>
> It's coherent and faster. GFX7-GFX9 should also support this but
> for now only uses L2 for GFX10 because it's untested on previous gens.
>
> This fixes dEQP-VK.memory.pipeline_barrier.transfer_*
>
> This also fixes some missing geometry in Dawn Of War III because
> VBOs weren't updated correctly.
>
> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
> ---
>  src/amd/vulkan/si_cmd_buffer.c | 16 ++++++++++++++++
>  1 file changed, 16 insertions(+)
>
> diff --git a/src/amd/vulkan/si_cmd_buffer.c b/src/amd/vulkan/si_cmd_buffer.c
> index 21a90cb2514..94f759139ee 100644
> --- a/src/amd/vulkan/si_cmd_buffer.c
> +++ b/src/amd/vulkan/si_cmd_buffer.c
> @@ -1501,6 +1501,14 @@ void si_cp_dma_buffer_copy(struct radv_cmd_buffer *cmd_buffer,
>                 unsigned dma_flags = 0;
>                 unsigned byte_count = MIN2(size, cp_dma_max_byte_count(cmd_buffer));
>
> +               if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX10) {
> +                       /* DMA operations via L2 are coherent and faster.
> +                        * TODO: GFX7-GFX9 should also support this but it
> +                        * requires tests/benchmarks.
> +                        */
> +                       dma_flags |= CP_DMA_USE_L2;
> +               }
> +
>                 si_cp_dma_prepare(cmd_buffer, byte_count,
>                                   size + skipped_size + realign_size,
>                                   &dma_flags);
> @@ -1545,6 +1553,14 @@ void si_cp_dma_clear_buffer(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
>                 unsigned byte_count = MIN2(size, cp_dma_max_byte_count(cmd_buffer));
>                 unsigned dma_flags = CP_DMA_CLEAR;
>
> +               if (cmd_buffer->device->physical_device->rad_info.chip_class >= GFX10) {
> +                       /* DMA operations via L2 are coherent and faster.
> +                        * TODO: GFX7-GFX9 should also support this but it
> +                        * requires tests/benchmarks.
> +                        */
> +                       dma_flags |= CP_DMA_USE_L2;
> +               }
> +
>                 si_cp_dma_prepare(cmd_buffer, byte_count, size, &dma_flags);
>
>                 /* Emit the clear packet. */
> --
> 2.22.0
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev