[Mesa-dev] [PATCH 26/61] radeonsi/gfx9: pass inputs from LS to TCS

Marek Olšák maraeo at gmail.com
Mon Apr 24 08:45:23 UTC 2017


From: Marek Olšák <marek.olsak at amd.com>

---
 src/gallium/drivers/radeonsi/si_shader.c          | 68 ++++++++++++++++++++++-
 src/gallium/drivers/radeonsi/si_shader_internal.h |  3 +
 2 files changed, 69 insertions(+), 2 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 907bc9c..1df9614 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -2605,20 +2605,41 @@ static void si_write_tess_factors(struct lp_build_tgsi_context *bld_base,
 			ac_build_buffer_store_dword(&ctx->ac, buf, inner_vec,
 						    inner_comps, tf_inner_offset,
 						    base, 0, 1, 0, true, false);
 		}
 	}
 
 	lp_build_endif(&if_ctx);
 }
 
 static LLVMValueRef
+si_insert_input_ret(struct si_shader_context *ctx, LLVMValueRef ret,
+		    unsigned param, unsigned return_index)
+{
+	return LLVMBuildInsertValue(ctx->gallivm.builder, ret,
+				    LLVMGetParam(ctx->main_fn, param),
+				    return_index, "");
+}
+
+static LLVMValueRef
+si_insert_input_ret_float(struct si_shader_context *ctx, LLVMValueRef ret,
+			  unsigned param, unsigned return_index)
+{
+	LLVMBuilderRef builder = ctx->gallivm.builder;
+	LLVMValueRef p = LLVMGetParam(ctx->main_fn, param);
+
+	return LLVMBuildInsertValue(builder, ret,
+				    LLVMBuildBitCast(builder, p, ctx->f32, ""),
+				    return_index, "");
+}
+
+static LLVMValueRef
 si_insert_input_ptr_as_2xi32(struct si_shader_context *ctx, LLVMValueRef ret,
 			     unsigned param, unsigned return_index)
 {
 	LLVMBuilderRef builder = ctx->gallivm.builder;
 	LLVMValueRef ptr, lo, hi;
 
 	ptr = LLVMGetParam(ctx->main_fn, param);
 	ptr = LLVMBuildPtrToInt(builder, ptr, ctx->i64, "");
 	ptr = LLVMBuildBitCast(builder, ptr, ctx->v2i32, "");
 	lo = LLVMBuildExtractElement(builder, ptr, ctx->i32_0, "");
@@ -2679,20 +2700,60 @@ static void si_llvm_emit_tcs_epilogue(struct lp_build_tgsi_context *bld_base)
 	rel_patch_id = bitcast(bld_base, TGSI_TYPE_FLOAT, rel_patch_id);
 	invocation_id = bitcast(bld_base, TGSI_TYPE_FLOAT, invocation_id);
 	tf_lds_offset = bitcast(bld_base, TGSI_TYPE_FLOAT, tf_lds_offset);
 
 	ret = LLVMBuildInsertValue(builder, ret, rel_patch_id, vgpr++, "");
 	ret = LLVMBuildInsertValue(builder, ret, invocation_id, vgpr++, "");
 	ret = LLVMBuildInsertValue(builder, ret, tf_lds_offset, vgpr++, "");
 	ctx->return_value = ret;
 }
 
+/* Pass TCS inputs from LS to TCS on GFX9. */
+static void si_set_ls_return_value_for_tcs(struct si_shader_context *ctx)
+{
+	LLVMValueRef ret = ctx->return_value;
+
+	ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_offset, 2);
+	ret = si_insert_input_ret(ctx, ret, ctx->param_merged_wave_info, 3);
+	ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_offset, 4);
+	ret = si_insert_input_ret(ctx, ret, ctx->param_merged_scratch_offset, 5);
+
+	ret = si_insert_input_ptr_as_2xi32(ctx, ret, ctx->param_rw_buffers,
+					   8 + SI_SGPR_RW_BUFFERS);
+
+	ret = si_insert_input_ret(ctx, ret, ctx->param_vs_state_bits,
+				  8 + SI_SGPR_VS_STATE_BITS);
+	ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_layout,
+				  8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT);
+	ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_out_lds_offsets,
+				  8 + GFX9_SGPR_TCS_OUT_OFFSETS);
+	ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_out_lds_layout,
+				  8 + GFX9_SGPR_TCS_OUT_LAYOUT);
+
+	unsigned desc_param = ctx->param_tcs_out_lds_layout + 2;
+	ret = si_insert_input_ptr_as_2xi32(ctx, ret, desc_param,
+					   8 + GFX9_SGPR_TCS_CONST_BUFFERS);
+	ret = si_insert_input_ptr_as_2xi32(ctx, ret, desc_param + 1,
+					   8 + GFX9_SGPR_TCS_SAMPLERS);
+	ret = si_insert_input_ptr_as_2xi32(ctx, ret, desc_param + 2,
+					   8 + GFX9_SGPR_TCS_IMAGES);
+	ret = si_insert_input_ptr_as_2xi32(ctx, ret, desc_param + 3,
+					   8 + GFX9_SGPR_TCS_SHADER_BUFFERS);
+
+	unsigned vgpr = 8 + GFX9_TCS_NUM_USER_SGPR;
+	ret = si_insert_input_ret_float(ctx, ret,
+					ctx->param_tcs_patch_id, vgpr++);
+	ret = si_insert_input_ret_float(ctx, ret,
+					ctx->param_tcs_rel_ids, vgpr++);
+	ctx->return_value = ret;
+}
+
 static void si_llvm_emit_ls_epilogue(struct lp_build_tgsi_context *bld_base)
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 	struct si_shader *shader = ctx->shader;
 	struct tgsi_shader_info *info = &shader->selector->info;
 	struct gallivm_state *gallivm = &ctx->gallivm;
 	unsigned i, chan;
 	LLVMValueRef vertex_id = LLVMGetParam(ctx->main_fn,
 					      ctx->param_rel_auto_id);
 	LLVMValueRef vertex_dw_stride =
@@ -2728,20 +2789,23 @@ static void si_llvm_emit_ls_epilogue(struct lp_build_tgsi_context *bld_base)
 
 		int param = si_shader_io_get_unique_index(name, index);
 		LLVMValueRef dw_addr = LLVMBuildAdd(gallivm->builder, base_dw_addr,
 					LLVMConstInt(ctx->i32, param * 4, 0), "");
 
 		for (chan = 0; chan < 4; chan++) {
 			lds_store(bld_base, chan, dw_addr,
 				  LLVMBuildLoad(gallivm->builder, out_ptr[chan], ""));
 		}
 	}
+
+	if (ctx->screen->b.chip_class >= GFX9)
+		si_set_ls_return_value_for_tcs(ctx);
 }
 
 static void si_llvm_emit_es_epilogue(struct lp_build_tgsi_context *bld_base)
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 	struct gallivm_state *gallivm = &ctx->gallivm;
 	struct si_shader *es = ctx->shader;
 	struct tgsi_shader_info *info = &es->selector->info;
 	LLVMValueRef soffset = LLVMGetParam(ctx->main_fn,
 					    ctx->param_es2gs_offset);
@@ -5792,23 +5856,23 @@ static void create_function(struct si_shader_context *ctx)
 			returns[num_returns++] = ctx->i32; /* SGPRs */
 		for (i = 0; i < 3; i++)
 			returns[num_returns++] = ctx->f32; /* VGPRs */
 		break;
 
 	case SI_SHADER_MERGED_VERTEX_TESSCTRL:
 		/* Merged stages have 8 system SGPRs at the beginning. */
 		params[num_params++] = ctx->i32; /* unused */
 		params[num_params++] = ctx->i32; /* unused */
 		params[ctx->param_tcs_offchip_offset = num_params++] = ctx->i32;
-		params[num_params++] = ctx->i32; /* wave thread counts for LS and HS */
+		params[ctx->param_merged_wave_info = num_params++] = ctx->i32;
 		params[ctx->param_tcs_factor_offset = num_params++] = ctx->i32;
-		params[num_params++] = ctx->i32; /* scratch wave offset */
+		params[ctx->param_merged_scratch_offset = num_params++] = ctx->i32;
 		params[num_params++] = ctx->i32; /* unused */
 		params[num_params++] = ctx->i32; /* unused */
 
 		params[ctx->param_rw_buffers = num_params++] =
 			const_array(ctx->v16i8, SI_NUM_RW_BUFFERS);
 		declare_per_stage_desc_pointers(ctx, params, &num_params,
 						ctx->type == PIPE_SHADER_VERTEX);
 		declare_vs_specific_input_sgprs(ctx, params, &num_params);
 
 		params[ctx->param_tcs_offchip_layout = num_params++] = ctx->i32;
diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h b/src/gallium/drivers/radeonsi/si_shader_internal.h
index 812472f..c344c16 100644
--- a/src/gallium/drivers/radeonsi/si_shader_internal.h
+++ b/src/gallium/drivers/radeonsi/si_shader_internal.h
@@ -106,20 +106,23 @@ struct si_shader_context {
 
 	LLVMValueRef main_fn;
 	LLVMTypeRef return_type;
 
 	/* Parameter indices for LLVMGetParam. */
 	int param_rw_buffers;
 	int param_const_buffers;
 	int param_samplers;
 	int param_images;
 	int param_shader_buffers;
+	/* Common inputs for merged shaders. */
+	int param_merged_wave_info;
+	int param_merged_scratch_offset;
 	/* API VS */
 	int param_vertex_buffers;
 	int param_base_vertex;
 	int param_start_instance;
 	int param_draw_id;
 	int param_vertex_id;
 	int param_rel_auto_id;
 	int param_vs_prim_id;
 	int param_instance_id;
 	int param_vertex_index0;
-- 
2.7.4



More information about the mesa-dev mailing list