[Mesa-dev] [PATCH 1/3] radv: do not allocate one extra SGPR for the draw ID

[Samuel Pitoiset]

We always emit 0 in there, just remove the allocated user SGPR
and return a constant instead. This will free one SGPR for apps
that need the draw ID.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
---
 src/amd/common/ac_nir_to_llvm.c  | 17 +++--------------
 src/amd/vulkan/radv_cmd_buffer.c |  6 ++----
 src/amd/vulkan/radv_pipeline.c   |  4 ----
 src/amd/vulkan/radv_private.h    |  1 -
 4 files changed, 5 insertions(+), 23 deletions(-)

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index b3336ffafe..47f62f42dc 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -571,11 +571,7 @@ static void allocate_user_sgprs(struct nir_to_llvm_context *ctx,
 	case MESA_SHADER_VERTEX:
 		if (!ctx->is_gs_copy_shader) {
 			user_sgpr_info->sgpr_count += ctx->shader_info->info.vs.has_vertex_buffers ? 2 : 0;
-			if (ctx->shader_info->info.vs.needs_draw_id) {
-				user_sgpr_info->sgpr_count += 3;
-			} else {
-				user_sgpr_info->sgpr_count += 2;
-			}
+			user_sgpr_info->sgpr_count += 2;
 		}
 		if (ctx->options->key.vs.as_ls)
 			user_sgpr_info->sgpr_count++;
@@ -661,9 +657,6 @@ declare_vs_specific_input_sgprs(struct nir_to_llvm_context *ctx,
 		}
 		add_arg(args, ARG_SGPR, ctx->ac.i32, &ctx->abi.base_vertex);
 		add_arg(args, ARG_SGPR, ctx->ac.i32, &ctx->abi.start_instance);
-		if (ctx->shader_info->info.vs.needs_draw_id) {
-			add_arg(args, ARG_SGPR, ctx->ac.i32, &ctx->abi.draw_id);
-		}
 	}
 }
 
@@ -749,12 +742,8 @@ set_vs_specific_input_locs(struct nir_to_llvm_context *ctx,
 				       user_sgpr_idx, 2);
 		}
 
-		unsigned vs_num = 2;
-		if (ctx->shader_info->info.vs.needs_draw_id)
-			vs_num++;
-
 		set_loc_shader(ctx, AC_UD_VS_BASE_VERTEX_START_INSTANCE,
-			       user_sgpr_idx, vs_num);
+			       user_sgpr_idx, 2);
 	}
 }
 
@@ -4332,7 +4321,7 @@ static void visit_intrinsic(struct ac_nir_context *ctx,
 		result = ctx->abi->start_instance;
 		break;
 	case nir_intrinsic_load_draw_id:
-		result = ctx->abi->draw_id;
+		result = ctx->abi->draw_id ? ctx->abi->draw_id : ctx->ac.i32_0;
 		break;
 	case nir_intrinsic_load_view_index:
 		result = ctx->nctx->view_index ? ctx->nctx->view_index : ctx->ac.i32_0;
diff --git a/src/amd/vulkan/radv_cmd_buffer.c b/src/amd/vulkan/radv_cmd_buffer.c
index b694174de6..e347c3f563 100644
--- a/src/amd/vulkan/radv_cmd_buffer.c
+++ b/src/amd/vulkan/radv_cmd_buffer.c
@@ -3360,13 +3360,11 @@ radv_emit_draw_packets(struct radv_cmd_buffer *cmd_buffer,
 
 		if (info->vertex_offset != state->last_vertex_offset ||
 		    info->first_instance != state->last_first_instance) {
-			radeon_set_sh_reg_seq(cs, state->pipeline->graphics.vtx_base_sgpr,
-					      state->pipeline->graphics.vtx_emit_num);
+			radeon_set_sh_reg_seq(cs, state->pipeline->graphics.vtx_base_sgpr, 2);
 
 			radeon_emit(cs, info->vertex_offset);
 			radeon_emit(cs, info->first_instance);
-			if (state->pipeline->graphics.vtx_emit_num == 3)
-				radeon_emit(cs, 0);
+
 			state->last_first_instance = info->first_instance;
 			state->last_vertex_offset = info->vertex_offset;
 		}
diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index 589e49a813..a3ba6ee3aa 100644
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -2691,10 +2691,6 @@ radv_pipeline_init(struct radv_pipeline *pipeline,
 	if (loc->sgpr_idx != -1) {
 		pipeline->graphics.vtx_base_sgpr = pipeline->user_data_0[MESA_SHADER_VERTEX];
 		pipeline->graphics.vtx_base_sgpr += loc->sgpr_idx * 4;
-		if (radv_get_vertex_shader(pipeline)->info.info.vs.needs_draw_id)
-			pipeline->graphics.vtx_emit_num = 3;
-		else
-			pipeline->graphics.vtx_emit_num = 2;
 	}
 
 	pipeline->graphics.vtx_reuse_depth = 30;
diff --git a/src/amd/vulkan/radv_private.h b/src/amd/vulkan/radv_private.h
index f650b9a360..7ad0c3baa8 100644
--- a/src/amd/vulkan/radv_private.h
+++ b/src/amd/vulkan/radv_private.h
@@ -1257,7 +1257,6 @@ struct radv_pipeline {
 			bool wd_switch_on_eop;
 			bool ia_switch_on_eoi;
 			bool partial_vs_wave;
-			uint8_t vtx_emit_num;
 			uint32_t vtx_reuse_depth;
 			struct radv_prim_vertex_count prim_vertex_count;
  			bool can_use_guardband;
-- 
2.16.1