[Mesa-dev] [PATCH] ac/nir: fix user sgpr allocations on gfx9

[Dave Airlie]

From: Dave Airlie <airlied at redhat.com>

The code to decide how to allocate sgprs didn't take into
account the merged shaders on gfx9.

This updates the code to count the correct number of requires
sgprs, which means we should be able to enable push const
inlining easier.

Signed-off-by: Dave Airlie <airlied at redhat.com>
---
 src/amd/common/ac_nir_to_llvm.c | 38 +++++++++++++++++++++++++++-----------
 1 file changed, 27 insertions(+), 11 deletions(-)

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 02a46dab4db..7560b61e99b 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -565,8 +565,21 @@ static bool needs_view_index_sgpr(struct nir_to_llvm_context *ctx,
 	return false;
 }
 
+static int add_vertex_user_sgprs(const struct ac_shader_info *shader_info)
+{
+	int sgpr_count = 0;
+	sgpr_count += shader_info->vs.has_vertex_buffers ? 2 : 0;
+	if (shader_info->vs.needs_draw_id) {
+		sgpr_count += 3;
+	} else {
+		sgpr_count += 2;
+	}
+	return sgpr_count;
+}
+
 static void allocate_user_sgprs(struct nir_to_llvm_context *ctx,
 				gl_shader_stage stage,
+				bool has_previous_stage, gl_shader_stage previous_stage,
 				bool needs_view_index,
 				struct user_sgpr_info *user_sgpr_info)
 {
@@ -589,7 +602,6 @@ static void allocate_user_sgprs(struct nir_to_llvm_context *ctx,
 		user_sgpr_info->sgpr_count += 2;
 	}
 
-	/* FIXME: fix the number of user sgprs for merged shaders on GFX9 */
 	switch (stage) {
 	case MESA_SHADER_COMPUTE:
 		if (ctx->shader_info->info.cs.uses_grid_size)
@@ -599,24 +611,28 @@ static void allocate_user_sgprs(struct nir_to_llvm_context *ctx,
 		user_sgpr_info->sgpr_count += ctx->shader_info->info.ps.needs_sample_positions;
 		break;
 	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;
-			}
-		}
+		if (!ctx->is_gs_copy_shader)
+			user_sgpr_info->sgpr_count += add_vertex_user_sgprs(&ctx->shader_info->info);
 		if (ctx->options->key.vs.as_ls)
 			user_sgpr_info->sgpr_count++;
 		break;
 	case MESA_SHADER_TESS_CTRL:
-		user_sgpr_info->sgpr_count += 4;
+		if (has_previous_stage) {
+			user_sgpr_info->sgpr_count += add_vertex_user_sgprs(&ctx->shader_info->info);
+			user_sgpr_info->sgpr_count += 5;
+		} else
+			user_sgpr_info->sgpr_count += 4;
 		break;
 	case MESA_SHADER_TESS_EVAL:
 		user_sgpr_info->sgpr_count += 1;
 		break;
 	case MESA_SHADER_GEOMETRY:
+		if (has_previous_stage) {
+			if (previous_stage == MESA_SHADER_VERTEX)
+				user_sgpr_info->sgpr_count += add_vertex_user_sgprs(&ctx->shader_info->info);
+			else if (previous_stage == MESA_SHADER_TESS_EVAL)
+				user_sgpr_info->sgpr_count += 1;
+		}
 		user_sgpr_info->sgpr_count += 2;
 		break;
 	default:
@@ -798,7 +814,7 @@ static void create_function(struct nir_to_llvm_context *ctx,
 	struct arg_info args = {};
 	LLVMValueRef desc_sets;
 	bool needs_view_index = needs_view_index_sgpr(ctx, stage);
-	allocate_user_sgprs(ctx, stage, needs_view_index, &user_sgpr_info);
+	allocate_user_sgprs(ctx, stage, has_previous_stage, stage, needs_view_index, &user_sgpr_info);
 
 	if (user_sgpr_info.need_ring_offsets && !ctx->options->supports_spill) {
 		add_arg(&args, ARG_SGPR, const_array(ctx->ac.v4i32, 16),
-- 
2.14.3