[Mesa-dev] [PATCH] radeonsi: support TGSI compute shaders with variable block size

Nicolai Hähnle nhaehnle at gmail.com
Fri Sep 9 08:12:59 UTC 2016


From: Nicolai Hähnle <nicolai.haehnle at amd.com>

Not sure if it's possible to avoid programming the block size twice (once for
the userdata and once for the dispatch).

Since the shaders are compiled with a pessimistic upper limit on the number of
registers, asynchronously compiling variants may be worth considering in the
future if we observe the shaders to be dispatched with small block sizes.
---
I think this is sufficient to support variable group sizes on radeonsi, but
it's completely untested. Do you keep the latest version of your series in a
public repository somewhere?

 src/gallium/drivers/radeonsi/si_compute.c | 10 +++++++++-
 src/gallium/drivers/radeonsi/si_shader.c  | 29 ++++++++++++++++++++---------
 src/gallium/drivers/radeonsi/si_shader.h  |  4 +++-
 3 files changed, 32 insertions(+), 11 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_compute.c b/src/gallium/drivers/radeonsi/si_compute.c
index 5041761..26e096c 100644
--- a/src/gallium/drivers/radeonsi/si_compute.c
+++ b/src/gallium/drivers/radeonsi/si_compute.c
@@ -379,25 +379,33 @@ static void si_setup_tgsi_grid(struct si_context *sctx,
 		for (i = 0; i < 3; ++i) {
 			radeon_emit(cs, PKT3(PKT3_COPY_DATA, 4, 0));
 			radeon_emit(cs, COPY_DATA_SRC_SEL(COPY_DATA_MEM) |
 					COPY_DATA_DST_SEL(COPY_DATA_REG));
 			radeon_emit(cs, (va +  4 * i));
 			radeon_emit(cs, (va + 4 * i) >> 32);
 			radeon_emit(cs, (grid_size_reg >> 2) + i);
 			radeon_emit(cs, 0);
 		}
 	} else {
+		struct si_compute *program = sctx->cs_shader_state.program;
+		bool variable_group_size =
+			program->shader.selector->info.properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] == 0;
 
-		radeon_set_sh_reg_seq(cs, grid_size_reg, 3);
+		radeon_set_sh_reg_seq(cs, grid_size_reg, variable_group_size ? 6 : 3);
 		radeon_emit(cs, info->grid[0]);
 		radeon_emit(cs, info->grid[1]);
 		radeon_emit(cs, info->grid[2]);
+		if (variable_group_size) {
+			radeon_emit(cs, info->block[0]);
+			radeon_emit(cs, info->block[1]);
+			radeon_emit(cs, info->block[2]);
+		}
 	}
 }
 
 static void si_emit_dispatch_packets(struct si_context *sctx,
                                      const struct pipe_grid_info *info)
 {
 	struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
 	bool render_cond_bit = sctx->b.render_cond && !sctx->b.render_cond_force_off;
 	unsigned waves_per_threadgroup =
 		DIV_ROUND_UP(info->block[0] * info->block[1] * info->block[2], 64);
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 0b7de18..730ee21 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -1783,30 +1783,35 @@ static void declare_system_value(
 
 	case TGSI_SEMANTIC_GRID_SIZE:
 		value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_GRID_SIZE);
 		break;
 
 	case TGSI_SEMANTIC_BLOCK_SIZE:
 	{
 		LLVMValueRef values[3];
 		unsigned i;
 		unsigned *properties = ctx->shader->selector->info.properties;
-		unsigned sizes[3] = {
-			properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH],
-			properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT],
-			properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH]
-		};
 
-		for (i = 0; i < 3; ++i)
-			values[i] = lp_build_const_int32(gallivm, sizes[i]);
+		if (properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] != 0) {
+			unsigned sizes[3] = {
+				properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH],
+				properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT],
+				properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH]
+			};
+
+			for (i = 0; i < 3; ++i)
+				values[i] = lp_build_const_int32(gallivm, sizes[i]);
 
-		value = lp_build_gather_values(gallivm, values, 3);
+			value = lp_build_gather_values(gallivm, values, 3);
+		} else {
+			value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_BLOCK_SIZE);
+		}
 		break;
 	}
 
 	case TGSI_SEMANTIC_BLOCK_ID:
 		value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_BLOCK_ID);
 		break;
 
 	case TGSI_SEMANTIC_THREAD_ID:
 		value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_THREAD_ID);
 		break;
@@ -5705,20 +5710,21 @@ static void create_function(struct si_shader_context *ctx)
 
 			for (i = 0; i < num_return_sgprs; i++)
 				returns[i] = ctx->i32;
 			for (; i < num_returns; i++)
 				returns[i] = ctx->f32;
 		}
 		break;
 
 	case PIPE_SHADER_COMPUTE:
 		params[SI_PARAM_GRID_SIZE] = v3i32;
+		params[SI_PARAM_BLOCK_SIZE] = v3i32;
 		params[SI_PARAM_BLOCK_ID] = v3i32;
 		last_sgpr = SI_PARAM_BLOCK_ID;
 
 		params[SI_PARAM_THREAD_ID] = v3i32;
 		num_params = SI_PARAM_THREAD_ID + 1;
 		break;
 	default:
 		assert(0 && "unimplemented shader");
 		return;
 	}
@@ -5741,21 +5747,26 @@ static void create_function(struct si_shader_context *ctx)
 					  S_0286D0_LINEAR_CENTROID_ENA(1) |
 					  S_0286D0_FRONT_FACE_ENA(1) |
 					  S_0286D0_POS_FIXED_PT_ENA(1));
 	} else if (ctx->type == PIPE_SHADER_COMPUTE) {
 		const unsigned *properties = shader->selector->info.properties;
 		unsigned max_work_group_size =
 		               properties[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
 		               properties[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
 		               properties[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH];
 
-		assert(max_work_group_size);
+		if (!max_work_group_size) {
+			/* This is a variable group size compute shader,
+			 * compile it for the maximum possible group size.
+			 */
+			max_work_group_size = 2048;
+		}
 
 		radeon_llvm_add_attribute(ctx->radeon_bld.main_fn,
 		                          "amdgpu-max-work-group-size",
 		                          max_work_group_size);
 	}
 
 	shader->info.num_input_sgprs = 0;
 	shader->info.num_input_vgprs = 0;
 
 	for (i = 0; i <= last_sgpr; ++i)
diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index fc1b22d..afdb3f5 100644
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -120,21 +120,22 @@ enum {
 	/* GS limits */
 	SI_GS_NUM_USER_SGPR = SI_NUM_RESOURCE_SGPRS,
 	SI_GSCOPY_NUM_USER_SGPR = SI_SGPR_RW_BUFFERS_HI + 1,
 
 	/* PS only */
 	SI_SGPR_ALPHA_REF	= SI_NUM_RESOURCE_SGPRS,
 	SI_PS_NUM_USER_SGPR,
 
 	/* CS only */
 	SI_SGPR_GRID_SIZE = SI_NUM_RESOURCE_SGPRS,
-	SI_CS_NUM_USER_SGPR = SI_SGPR_GRID_SIZE + 3
+	SI_SGPR_BLOCK_SIZE = SI_SGPR_GRID_SIZE + 3,
+	SI_CS_NUM_USER_SGPR = SI_SGPR_BLOCK_SIZE + 3
 };
 
 /* LLVM function parameter indices */
 enum {
 	SI_PARAM_RW_BUFFERS,
 	SI_PARAM_CONST_BUFFERS,
 	SI_PARAM_SAMPLERS,
 	SI_PARAM_IMAGES,
 	SI_PARAM_SHADER_BUFFERS,
 	SI_NUM_RESOURCE_PARAMS,
@@ -210,20 +211,21 @@ enum {
 	SI_PARAM_POS_Y_FLOAT,
 	SI_PARAM_POS_Z_FLOAT,
 	SI_PARAM_POS_W_FLOAT,
 	SI_PARAM_FRONT_FACE,
 	SI_PARAM_ANCILLARY,
 	SI_PARAM_SAMPLE_COVERAGE,
 	SI_PARAM_POS_FIXED_PT,
 
 	/* CS only parameters */
 	SI_PARAM_GRID_SIZE = SI_NUM_RESOURCE_PARAMS,
+	SI_PARAM_BLOCK_SIZE,
 	SI_PARAM_BLOCK_ID,
 	SI_PARAM_THREAD_ID,
 
 	SI_NUM_PARAMS = SI_PARAM_POS_FIXED_PT + 9, /* +8 for COLOR[0..1] */
 };
 
 /* SI-specific system values. */
 enum {
 	TGSI_SEMANTIC_DEFAULT_TESSOUTER_SI = TGSI_SEMANTIC_COUNT,
 	TGSI_SEMANTIC_DEFAULT_TESSINNER_SI,
-- 
2.7.4



More information about the mesa-dev mailing list