radeonsi: support ARB_compute_variable_group_size

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

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>

7 files changed, 55 insertions, 19 deletions

diff --git a/docs/features.txt b/docs/features.txt
index e91ef6c..533971f 100644
--- a/docs/features.txt
+++ b/docs/features.txt
@@ -279,7 +279,7 @@ Khronos, ARB, and OES extensions that are not part of any OpenGL or OpenGL ES ve
 
   GL_ARB_bindless_texture                               started (airlied)
   GL_ARB_cl_event                                       not started
-  GL_ARB_compute_variable_group_size                    DONE (nvc0)
+  GL_ARB_compute_variable_group_size                    DONE (nvc0, radeonsi)
   GL_ARB_ES3_2_compatibility                            DONE (i965/gen8+)
   GL_ARB_fragment_shader_interlock                      not started
   GL_ARB_gl_spirv                                       not started
diff --git a/docs/relnotes/12.1.0.html b/docs/relnotes/12.1.0.html
index 9ddd99c..2e4b669 100644
--- a/docs/relnotes/12.1.0.html
+++ b/docs/relnotes/12.1.0.html
@@ -49,7 +49,7 @@ Note: some of the new features are only available with certain drivers.
 <li>GL_ARB_ES3_1_compatibility on i965</li>
 <li>GL_ARB_ES3_2_compatibility on i965/gen8+</li>
 <li>GL_ARB_clear_texture on r600, radeonsi</li>
-<li>GL_ARB_compute_variable_group_size on nvc0</li>
+<li>GL_ARB_compute_variable_group_size on nvc0, radeonsi</li>
 <li>GL_ARB_cull_distance on radeonsi</li>
 <li>GL_ARB_enhanced_layouts on i965</li>
 <li>GL_ARB_indirect_parameters on radeonsi</li>
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
index 44863ee..3dbcbc6 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.c
+++ b/src/gallium/drivers/radeon/r600_pipe_common.c
@@ -1037,7 +1037,15 @@ static int r600_get_compute_param(struct pipe_screen *screen,
 		}
 		return sizeof(uint32_t);
 	case PIPE_COMPUTE_CAP_MAX_VARIABLE_THREADS_PER_BLOCK:
-		return 0;
+		if (ret) {
+			uint64_t *max_variable_threads_per_block = ret;
+			if (rscreen->chip_class >= SI && HAVE_LLVM >= 0x309 &&
+			    ir_type == PIPE_SHADER_IR_TGSI)
+				*max_variable_threads_per_block = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
+			else
+				*max_variable_threads_per_block = 0;
+		}
+		return sizeof(uint64_t);
 	}
 
         fprintf(stderr, "unknown PIPE_COMPUTE_CAP %d\n", param);
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.h b/src/gallium/drivers/radeon/r600_pipe_common.h
index 54991e8..290b228 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.h
+++ b/src/gallium/drivers/radeon/r600_pipe_common.h
@@ -106,6 +106,8 @@
 #define R600_MAP_BUFFER_ALIGNMENT 64
 #define R600_MAX_VIEWPORTS        16
 
+#define SI_MAX_VARIABLE_THREADS_PER_BLOCK 1024
+
 enum r600_coherency {
 	R600_COHERENCY_NONE, /* no cache flushes needed */
 	R600_COHERENCY_SHADER,
diff --git a/src/gallium/drivers/radeonsi/si_compute.c b/src/gallium/drivers/radeonsi/si_compute.c
index 1d1df2f..e59bafe 100644
--- a/src/gallium/drivers/radeonsi/si_compute.c
+++ b/src/gallium/drivers/radeonsi/si_compute.c
@@ -601,11 +601,19 @@ static void si_setup_tgsi_grid(struct si_context *sctx,
 			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]);
+		}
 	}
 }
 
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index ff51c8b..49d4121 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -1770,16 +1770,21 @@ static void declare_system_value(
 		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;
 	}
 
@@ -5680,6 +5685,7 @@ static void create_function(struct si_shader_context *ctx)
 
 	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;
 
@@ -5716,7 +5722,12 @@ static void create_function(struct si_shader_context *ctx)
 		               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 = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
+		}
 
 		radeon_llvm_add_attribute(ctx->radeon_bld.main_fn,
 		                          "amdgpu-max-work-group-size",
@@ -6653,11 +6664,16 @@ int si_compile_tgsi_shader(struct si_screen *sscreen,
 		unsigned max_vgprs = 256;
 		unsigned max_sgprs = sscreen->b.chip_class >= VI ? 800 : 512;
 		unsigned max_sgprs_per_wave = 128;
-		unsigned min_waves_per_cu =
-			DIV_ROUND_UP(props[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
-				     props[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
-				     props[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH],
-				     wave_size);
+		unsigned max_block_threads;
+
+		if (props[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH])
+			max_block_threads = props[TGSI_PROPERTY_CS_FIXED_BLOCK_WIDTH] *
+					    props[TGSI_PROPERTY_CS_FIXED_BLOCK_HEIGHT] *
+					    props[TGSI_PROPERTY_CS_FIXED_BLOCK_DEPTH];
+		else
+			max_block_threads = SI_MAX_VARIABLE_THREADS_PER_BLOCK;
+
+		unsigned min_waves_per_cu = DIV_ROUND_UP(max_block_threads, wave_size);
 		unsigned min_waves_per_simd = DIV_ROUND_UP(min_waves_per_cu, 4);
 
 		max_vgprs = max_vgprs / min_waves_per_simd;
diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index 67cb67d2..f2618ac 100644
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -129,7 +129,8 @@ enum {
 
 	/* 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 */
@@ -219,6 +220,7 @@ enum {
 
 	/* CS only parameters */
 	SI_PARAM_GRID_SIZE = SI_NUM_RESOURCE_PARAMS,
+	SI_PARAM_BLOCK_SIZE,
 	SI_PARAM_BLOCK_ID,
 	SI_PARAM_THREAD_ID,