[Mesa-dev] [PATCH 2/2] radv: do not emit unnecessary ES output stores
Bas Nieuwenhuizen
bas at basnieuwenhuizen.nl
Wed May 16 15:56:26 UTC 2018
Thanks!
Reviewed-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>
for the series.
On Wed, May 16, 2018 at 5:43 PM, Samuel Pitoiset
<samuel.pitoiset at gmail.com> wrote:
> GFX9:
> Totals from affected shaders:
> SGPRS: 472 -> 464 (-1.69 %)
> VGPRS: 576 -> 584 (1.39 %)
> Code Size: 45432 -> 44324 (-2.44 %) bytes
> Max Waves: 40 -> 40 (0.00 %)
>
> VI:
> SGPRS: 720 -> 720 (0.00 %)
> VGPRS: 728 -> 728 (0.00 %)
> Code Size: 45348 -> 43992 (-2.99 %) bytes
> Max Waves: 120 -> 120 (0.00 %)
>
> This affects Rise of Tomb Raider and the three Vulkan demos
> that use a geometry shader (geometryshader, deferredshadows
> and viewportarray).
>
> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
> ---
> src/amd/vulkan/radv_nir_to_llvm.c | 18 +++++++++++++++---
> 1 file changed, 15 insertions(+), 3 deletions(-)
>
> diff --git a/src/amd/vulkan/radv_nir_to_llvm.c b/src/amd/vulkan/radv_nir_to_llvm.c
> index bfae2a489e..2c3abc2109 100644
> --- a/src/amd/vulkan/radv_nir_to_llvm.c
> +++ b/src/amd/vulkan/radv_nir_to_llvm.c
> @@ -2580,6 +2580,8 @@ handle_es_outputs_post(struct radv_shader_context *ctx,
> }
>
> for (unsigned i = 0; i < AC_LLVM_MAX_OUTPUTS; ++i) {
> + unsigned output_usage_mask =
> + ctx->shader_info->info.vs.output_usage_mask[i];
> LLVMValueRef dw_addr = NULL;
> LLVMValueRef *out_ptr = &ctx->abi.outputs[i * 4];
> int param_index;
> @@ -2588,8 +2590,10 @@ handle_es_outputs_post(struct radv_shader_context *ctx,
> if (!(ctx->output_mask & (1ull << i)))
> continue;
>
> - if (i == VARYING_SLOT_CLIP_DIST0)
> + if (i == VARYING_SLOT_CLIP_DIST0) {
> length = ctx->num_output_clips + ctx->num_output_culls;
> + output_usage_mask = (1 << length) - 1;
> + }
>
> param_index = shader_io_get_unique_index(i);
>
> @@ -2598,14 +2602,22 @@ handle_es_outputs_post(struct radv_shader_context *ctx,
> LLVMConstInt(ctx->ac.i32, param_index * 4, false),
> "");
> }
> +
> for (j = 0; j < length; j++) {
> + if (!(output_usage_mask & (1 << j)))
> + continue;
> +
> LLVMValueRef out_val = LLVMBuildLoad(ctx->ac.builder, out_ptr[j], "");
> out_val = LLVMBuildBitCast(ctx->ac.builder, out_val, ctx->ac.i32, "");
>
> if (ctx->ac.chip_class >= GFX9) {
> - ac_lds_store(&ctx->ac, dw_addr,
> + LLVMValueRef dw_addr_offset =
> + LLVMBuildAdd(ctx->ac.builder, dw_addr,
> + LLVMConstInt(ctx->ac.i32,
> + j, false), "");
> +
> + ac_lds_store(&ctx->ac, dw_addr_offset,
> LLVMBuildLoad(ctx->ac.builder, out_ptr[j], ""));
> - dw_addr = LLVMBuildAdd(ctx->ac.builder, dw_addr, ctx->ac.i32_1, "");
> } else {
> ac_build_buffer_store_dword(&ctx->ac,
> ctx->esgs_ring,
> --
> 2.17.0
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
More information about the mesa-dev
mailing list