[Mesa-dev] [PATCH 7/9] radv: start allocating user sgprs

[Dave Airlie]

From: Dave Airlie <airlied at redhat.com>

This adds an initial implementation to allocate the user
sgprs and make sure we don't run out if we try to bind
a bunch of descriptor sets.

This can be enhanced further in the future if we add
support for inlining push constants.

Signed-off-by: Dave Airlie <airlied at redhat.com>
---
 src/amd/common/ac_nir_to_llvm.c | 87 +++++++++++++++++++++++++++++++++++------
 1 file changed, 74 insertions(+), 13 deletions(-)

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 8368a12..8ae69d5 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -556,16 +556,15 @@ static void declare_tess_lds(struct nir_to_llvm_context *ctx)
 		"tess_lds");
 }
 
-static void create_function(struct nir_to_llvm_context *ctx)
+struct user_sgpr_info {
+	bool need_ring_offsets;
+	uint8_t sgpr_count;
+};
+
+static void allocate_user_sgprs(struct nir_to_llvm_context *ctx,
+				struct user_sgpr_info *user_sgpr_info)
 {
-	LLVMTypeRef arg_types[23];
-	unsigned arg_idx = 0;
-	unsigned array_params_mask = 0;
-	unsigned sgpr_count = 0, user_sgpr_count;
-	unsigned i;
-	unsigned num_sets = ctx->options->layout ? ctx->options->layout->num_sets : 0;
-	unsigned user_sgpr_idx;
-	bool need_ring_offsets = false;
+	memset(user_sgpr_info, 0, sizeof(struct user_sgpr_info));
 
 	/* until we sort out scratch/global buffers always assign ring offsets for gs/vs/es */
 	if (ctx->stage == MESA_SHADER_GEOMETRY ||
@@ -573,13 +572,75 @@ static void create_function(struct nir_to_llvm_context *ctx)
 	    ctx->stage == MESA_SHADER_TESS_CTRL ||
 	    ctx->stage == MESA_SHADER_TESS_EVAL ||
 	    ctx->is_gs_copy_shader)
-		need_ring_offsets = true;
+		user_sgpr_info->need_ring_offsets = true;
 
 	if (ctx->stage == MESA_SHADER_FRAGMENT &&
 	    ctx->shader_info->info.ps.needs_sample_positions)
-		need_ring_offsets = true;
+		user_sgpr_info->need_ring_offsets = true;
+
+	/* 2 user sgprs will nearly always be allocated for scratch/rings */
+	if (ctx->options->supports_spill || user_sgpr_info->need_ring_offsets) {
+		user_sgpr_info->sgpr_count += 2;
+	}
+
+	switch (ctx->stage) {
+	case MESA_SHADER_COMPUTE:
+		user_sgpr_info->sgpr_count += ctx->shader_info->info.cs.grid_components_used;
+		break;
+	case MESA_SHADER_FRAGMENT:
+		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 if (ctx->shader_info->info.vs.needs_base_vertex_start_instance) {
+				user_sgpr_info->sgpr_count += 2;
+			}
+		}
+		if (ctx->options->key.vs.as_ls)
+			user_sgpr_info->sgpr_count++;
+		break;
+	case MESA_SHADER_TESS_CTRL:
+		user_sgpr_info->sgpr_count += 4;
+		break;
+	case MESA_SHADER_TESS_EVAL:
+		user_sgpr_info->sgpr_count += 1;
+		break;
+	case MESA_SHADER_GEOMETRY:
+		user_sgpr_info->sgpr_count += 2;
+		break;
+	default:
+		break;
+	}
+
+	if (ctx->shader_info->info.needs_push_constants)
+		user_sgpr_info->sgpr_count += 2;
+
+	uint32_t remaining_sgprs = 16 - user_sgpr_info->sgpr_count;
+	if (remaining_sgprs / 2 < util_bitcount(ctx->shader_info->info.desc_set_used_mask)) {
+		fprintf(stderr, "radv: TODO: add support for indirect sgprs\n");
+		/* need to add support for indirect descriptor sets */
+		assert(0);
+	} else {
+		user_sgpr_info->sgpr_count += util_bitcount(ctx->shader_info->info.desc_set_used_mask) * 2;
+	}
+}
+
+static void create_function(struct nir_to_llvm_context *ctx)
+{
+	LLVMTypeRef arg_types[23];
+	unsigned arg_idx = 0;
+	unsigned array_params_mask = 0;
+	unsigned sgpr_count = 0, user_sgpr_count;
+	unsigned i;
+	unsigned num_sets = ctx->options->layout ? ctx->options->layout->num_sets : 0;
+	unsigned user_sgpr_idx;
+	struct user_sgpr_info user_sgpr_info;
 
-	if (need_ring_offsets && !ctx->options->supports_spill) {
+	allocate_user_sgprs(ctx, &user_sgpr_info);
+	if (user_sgpr_info.need_ring_offsets && !ctx->options->supports_spill) {
 		arg_types[arg_idx++] = const_array(ctx->v16i8, 16); /* address of rings */
 	}
 
@@ -730,7 +791,7 @@ static void create_function(struct nir_to_llvm_context *ctx)
 	arg_idx = 0;
 	user_sgpr_idx = 0;
 
-	if (ctx->options->supports_spill || need_ring_offsets) {
+	if (ctx->options->supports_spill || user_sgpr_info.need_ring_offsets) {
 		set_userdata_location_shader(ctx, AC_UD_SCRATCH_RING_OFFSETS, user_sgpr_idx, 2);
 		user_sgpr_idx += 2;
 		if (ctx->options->supports_spill) {
-- 
2.9.3