[Mesa-dev] [PATCH 3/9] radv: optimise compute shader grid size emission.

[Dave Airlie]

From: Dave Airlie <airlied at redhat.com>

Signed-off-by: Dave Airlie <airlied at redhat.com>
---
 src/amd/common/ac_nir_to_llvm.c  | 13 ++++++++-----
 src/amd/common/ac_shader_info.c  |  3 +++
 src/amd/common/ac_shader_info.h  |  3 +++
 src/amd/vulkan/radv_cmd_buffer.c | 23 +++++++++++++++--------
 4 files changed, 29 insertions(+), 13 deletions(-)

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 7161caf..7d19a78 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -604,7 +604,8 @@ static void create_function(struct nir_to_llvm_context *ctx)
 
 	switch (ctx->stage) {
 	case MESA_SHADER_COMPUTE:
-		arg_types[arg_idx++] = LLVMVectorType(ctx->i32, 3); /* grid size */
+		if (ctx->shader_info->info.cs.grid_components_used)
+			arg_types[arg_idx++] = LLVMVectorType(ctx->i32, ctx->shader_info->info.cs.grid_components_used); /* grid size */
 		user_sgpr_count = arg_idx;
 		arg_types[arg_idx++] = LLVMVectorType(ctx->i32, 3);
 		arg_types[arg_idx++] = ctx->i32;
@@ -765,10 +766,12 @@ static void create_function(struct nir_to_llvm_context *ctx)
 
 	switch (ctx->stage) {
 	case MESA_SHADER_COMPUTE:
-		set_userdata_location_shader(ctx, AC_UD_CS_GRID_SIZE, user_sgpr_idx, 3);
-		user_sgpr_idx += 3;
-		ctx->num_work_groups =
-		    LLVMGetParam(ctx->main_function, arg_idx++);
+		if (ctx->shader_info->info.cs.grid_components_used) {
+			set_userdata_location_shader(ctx, AC_UD_CS_GRID_SIZE, user_sgpr_idx, ctx->shader_info->info.cs.grid_components_used);
+			user_sgpr_idx += ctx->shader_info->info.cs.grid_components_used;
+			ctx->num_work_groups =
+				LLVMGetParam(ctx->main_function, arg_idx++);
+		}
 		ctx->workgroup_ids =
 		    LLVMGetParam(ctx->main_function, arg_idx++);
 		ctx->tg_size =
diff --git a/src/amd/common/ac_shader_info.c b/src/amd/common/ac_shader_info.c
index 5061860..6be66a9 100644
--- a/src/amd/common/ac_shader_info.c
+++ b/src/amd/common/ac_shader_info.c
@@ -37,6 +37,9 @@ gather_intrinsic_info(nir_intrinsic_instr *instr, struct ac_shader_info *info)
 	case nir_intrinsic_load_draw_id:
 		info->vs.needs_draw_id = true;
 		break;
+	case nir_intrinsic_load_num_work_groups:
+		info->cs.grid_components_used = instr->num_components;
+		break;
 	default:
 		break;
 	}
diff --git a/src/amd/common/ac_shader_info.h b/src/amd/common/ac_shader_info.h
index 9d43637..5acfb21 100644
--- a/src/amd/common/ac_shader_info.h
+++ b/src/amd/common/ac_shader_info.h
@@ -36,6 +36,9 @@ struct ac_shader_info {
 	struct {
 		bool needs_sample_positions;
 	} ps;
+	struct {
+		uint8_t grid_components_used;
+	} cs;
 };
 
 void
diff --git a/src/amd/vulkan/radv_cmd_buffer.c b/src/amd/vulkan/radv_cmd_buffer.c
index 22ae0c4..343e043 100644
--- a/src/amd/vulkan/radv_cmd_buffer.c
+++ b/src/amd/vulkan/radv_cmd_buffer.c
@@ -2791,11 +2791,14 @@ void radv_CmdDispatch(
 							     MESA_SHADER_COMPUTE, AC_UD_CS_GRID_SIZE);
 	if (loc->sgpr_idx != -1) {
 		assert(!loc->indirect);
-		assert(loc->num_sgprs == 3);
-		radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B900_COMPUTE_USER_DATA_0 + loc->sgpr_idx * 4, 3);
+		uint8_t grid_used = cmd_buffer->state.pipeline->shaders[MESA_SHADER_COMPUTE]->info.info.cs.grid_components_used;
+		assert(loc->num_sgprs == grid_used);
+		radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B900_COMPUTE_USER_DATA_0 + loc->sgpr_idx * 4, grid_used);
 		radeon_emit(cmd_buffer->cs, x);
-		radeon_emit(cmd_buffer->cs, y);
-		radeon_emit(cmd_buffer->cs, z);
+		if (grid_used > 1)
+			radeon_emit(cmd_buffer->cs, y);
+		if (grid_used > 2)
+			radeon_emit(cmd_buffer->cs, z);
 	}
 
 	radeon_emit(cmd_buffer->cs, PKT3(PKT3_DISPATCH_DIRECT, 3, 0) |
@@ -2827,7 +2830,8 @@ void radv_CmdDispatchIndirect(
 	struct ac_userdata_info *loc = radv_lookup_user_sgpr(cmd_buffer->state.compute_pipeline,
 							     MESA_SHADER_COMPUTE, AC_UD_CS_GRID_SIZE);
 	if (loc->sgpr_idx != -1) {
-		for (unsigned i = 0; i < 3; ++i) {
+		uint8_t grid_used = cmd_buffer->state.pipeline->shaders[MESA_SHADER_COMPUTE]->info.info.cs.grid_components_used;
+		for (unsigned i = 0; i < grid_used; ++i) {
 			radeon_emit(cmd_buffer->cs, PKT3(PKT3_COPY_DATA, 4, 0));
 			radeon_emit(cmd_buffer->cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
 				    COPY_DATA_DST_SEL(COPY_DATA_REG));
@@ -2898,10 +2902,13 @@ void radv_unaligned_dispatch(
 	struct ac_userdata_info *loc = radv_lookup_user_sgpr(cmd_buffer->state.compute_pipeline,
 							     MESA_SHADER_COMPUTE, AC_UD_CS_GRID_SIZE);
 	if (loc->sgpr_idx != -1) {
-		radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B900_COMPUTE_USER_DATA_0 + loc->sgpr_idx * 4, 3);
+		uint8_t grid_used = cmd_buffer->state.pipeline->shaders[MESA_SHADER_COMPUTE]->info.info.cs.grid_components_used;
+		radeon_set_sh_reg_seq(cmd_buffer->cs, R_00B900_COMPUTE_USER_DATA_0 + loc->sgpr_idx * 4, grid_used);
 		radeon_emit(cmd_buffer->cs, blocks[0]);
-		radeon_emit(cmd_buffer->cs, blocks[1]);
-		radeon_emit(cmd_buffer->cs, blocks[2]);
+		if (grid_used > 1)
+			radeon_emit(cmd_buffer->cs, blocks[1]);
+		if (grid_used > 2)
+			radeon_emit(cmd_buffer->cs, blocks[2]);
 	}
 	radeon_emit(cmd_buffer->cs, PKT3(PKT3_DISPATCH_DIRECT, 3, 0) |
 		    PKT3_SHADER_TYPE_S(1));
-- 
2.9.3