From d633e23192ef17207f4a6acd3009da3126aab395 Mon Sep 17 00:00:00 2001 From: =?utf8?q?Marek=20Ol=C5=A1=C3=A1k?= Date: Sat, 18 Feb 2017 15:30:25 +0100 Subject: [PATCH] radeonsi: skip LDS stores in TCS if there are no LDS output reads MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit This removes a lot of useless LDS stores. A few games read TESSINNER/OUTER, but not any other outputs. Most games don't read any outputs. The only app doing LDS output reads is UE4 Lightsroom Interior. Reviewed-by: Nicolai Hähnle --- src/gallium/drivers/radeonsi/si_shader.c | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-) diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c index 4075eefe623..b85874ae403 100644 --- a/src/gallium/drivers/radeonsi/si_shader.c +++ b/src/gallium/drivers/radeonsi/si_shader.c @@ -982,10 +982,12 @@ static void store_output_tcs(struct lp_build_tgsi_context *bld_base, struct si_shader_context *ctx = si_shader_context(bld_base); struct gallivm_state *gallivm = bld_base->base.gallivm; const struct tgsi_full_dst_register *reg = &inst->Dst[0]; + const struct tgsi_shader_info *sh_info = &ctx->shader->selector->info; unsigned chan_index; LLVMValueRef dw_addr, stride; LLVMValueRef rw_buffers, buffer, base, buf_addr; LLVMValueRef values[4]; + bool skip_lds_store; /* Only handle per-patch and per-vertex outputs here. * Vectors will be lowered to scalars and this function will be called again. @@ -1000,9 +1002,20 @@ static void store_output_tcs(struct lp_build_tgsi_context *bld_base, stride = unpack_param(ctx, SI_PARAM_TCS_OUT_LAYOUT, 13, 8); dw_addr = get_tcs_out_current_patch_offset(ctx); dw_addr = get_dw_address(ctx, reg, NULL, stride, dw_addr); + skip_lds_store = !sh_info->reads_pervertex_outputs; } else { dw_addr = get_tcs_out_current_patch_data_offset(ctx); dw_addr = get_dw_address(ctx, reg, NULL, NULL, dw_addr); + skip_lds_store = !sh_info->reads_perpatch_outputs; + + if (!reg->Register.Indirect) { + int name = sh_info->output_semantic_name[reg->Register.Index]; + + /* Always write tess factors into LDS for the TCS epilog. */ + if (name == TGSI_SEMANTIC_TESSINNER || + name == TGSI_SEMANTIC_TESSOUTER) + skip_lds_store = false; + } } rw_buffers = LLVMGetParam(ctx->main_fn, @@ -1020,7 +1033,9 @@ static void store_output_tcs(struct lp_build_tgsi_context *bld_base, if (inst->Instruction.Saturate) value = ac_emit_clamp(&ctx->ac, value); - lds_store(bld_base, chan_index, dw_addr, value); + /* Skip LDS stores if there is no LDS read of this output. */ + if (!skip_lds_store) + lds_store(bld_base, chan_index, dw_addr, value); value = LLVMBuildBitCast(gallivm->builder, value, ctx->i32, ""); values[chan_index] = value; -- 2.30.2