[Mesa-dev] [PATCH v3 09/14] radeonsi: Use buffer loads and stores for passing data from TCS to TES.

[Bas Nieuwenhuizen]

We always try to use 4-component loads, as LLVM does not combine loads
and they bypass the L1 cache.

We can't use a similar strategy for stores and this is especially
notable with the tess factors, as they are often set with separate
MOV's per component in the TGSI.

We keep storing to LDS and the LDS space, so we can load the outputs
later, either due to the shader, of for wrting the tess factors.

Signed-off-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle at amd.com>
Reviewed-by: Marek Olšák <marek.olsak at amd.com>
---
 src/gallium/drivers/radeonsi/si_shader.c | 66 ++++++++++++++++++++++++--------
 1 file changed, 50 insertions(+), 16 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index b04d0f7..6694f00 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -1082,18 +1082,18 @@ static LLVMValueRef fetch_input_tes(
 	enum tgsi_opcode_type type, unsigned swizzle)
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
-	LLVMValueRef dw_addr, stride;
+	struct gallivm_state *gallivm = bld_base->base.gallivm;
+	LLVMValueRef rw_buffers, buffer, base, addr;
 
-	if (reg->Register.Dimension) {
-		stride = unpack_param(ctx, SI_PARAM_TCS_OUT_LAYOUT, 13, 8);
-		dw_addr = get_tcs_out_current_patch_offset(ctx);
-		dw_addr = get_dw_address(ctx, NULL, reg, stride, dw_addr);
-	} else {
-		dw_addr = get_tcs_out_current_patch_data_offset(ctx);
-		dw_addr = get_dw_address(ctx, NULL, reg, NULL, dw_addr);
-	}
+	rw_buffers = LLVMGetParam(ctx->radeon_bld.main_fn,
+				  SI_PARAM_RW_BUFFERS);
+	buffer = build_indexed_load_const(ctx, rw_buffers,
+			lp_build_const_int32(gallivm, SI_HS_RING_TESS_OFFCHIP));
 
-	return lds_load(bld_base, type, swizzle, dw_addr);
+	base = LLVMGetParam(ctx->radeon_bld.main_fn, ctx->param_oc_lds);
+	addr = get_tcs_tes_buffer_address_from_reg(ctx, NULL, reg);
+
+	return buffer_load(bld_base, type, swizzle, buffer, base, addr);
 }
 
 static void store_output_tcs(struct lp_build_tgsi_context *bld_base,
@@ -1102,9 +1102,12 @@ static void store_output_tcs(struct lp_build_tgsi_context *bld_base,
 			     LLVMValueRef dst[4])
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
+	struct gallivm_state *gallivm = bld_base->base.gallivm;
 	const struct tgsi_full_dst_register *reg = &inst->Dst[0];
 	unsigned chan_index;
 	LLVMValueRef dw_addr, stride;
+	LLVMValueRef rw_buffers, buffer, base, buf_addr;
+	LLVMValueRef values[4];
 
 	/* Only handle per-patch and per-vertex outputs here.
 	 * Vectors will be lowered to scalars and this function will be called again.
@@ -1124,6 +1127,15 @@ static void store_output_tcs(struct lp_build_tgsi_context *bld_base,
 		dw_addr = get_dw_address(ctx, reg, NULL, NULL, dw_addr);
 	}
 
+	rw_buffers = LLVMGetParam(ctx->radeon_bld.main_fn,
+				  SI_PARAM_RW_BUFFERS);
+	buffer = build_indexed_load_const(ctx, rw_buffers,
+			lp_build_const_int32(gallivm, SI_HS_RING_TESS_OFFCHIP));
+
+	base = LLVMGetParam(ctx->radeon_bld.main_fn, ctx->param_oc_lds);
+	buf_addr = get_tcs_tes_buffer_address_from_reg(ctx, reg, NULL);
+
+
 	TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst, chan_index) {
 		LLVMValueRef value = dst[chan_index];
 
@@ -1131,6 +1143,22 @@ static void store_output_tcs(struct lp_build_tgsi_context *bld_base,
 			value = radeon_llvm_saturate(bld_base, value);
 
 		lds_store(bld_base, chan_index, dw_addr, value);
+
+		value = LLVMBuildBitCast(gallivm->builder, value, ctx->i32, "");
+		values[chan_index] = value;
+
+		if (inst->Dst[0].Register.WriteMask != 0xF) {
+			build_tbuffer_store_dwords(ctx, buffer, value, 1,
+			                           buf_addr, base,
+			                           4 * chan_index);
+		}
+	}
+
+	if (inst->Dst[0].Register.WriteMask == 0xF) {
+		LLVMValueRef value = lp_build_gather_values(bld_base->base.gallivm,
+		                                            values, 4);
+		build_tbuffer_store_dwords(ctx, buffer, value, 4, buf_addr,
+		                           base, 0);
 	}
 }
 
@@ -1641,15 +1669,21 @@ static void declare_system_value(
 	case TGSI_SEMANTIC_TESSINNER:
 	case TGSI_SEMANTIC_TESSOUTER:
 	{
-		LLVMValueRef dw_addr;
+		LLVMValueRef rw_buffers, buffer, base, addr;
 		int param = si_shader_io_get_unique_index(decl->Semantic.Name, 0);
 
-		dw_addr = get_tcs_out_current_patch_data_offset(ctx);
-		dw_addr = LLVMBuildAdd(gallivm->builder, dw_addr,
-				       lp_build_const_int32(gallivm, param * 4), "");
+		rw_buffers = LLVMGetParam(ctx->radeon_bld.main_fn,
+					SI_PARAM_RW_BUFFERS);
+		buffer = build_indexed_load_const(ctx, rw_buffers,
+		        lp_build_const_int32(gallivm, SI_HS_RING_TESS_OFFCHIP));
+
+		base = LLVMGetParam(ctx->radeon_bld.main_fn, ctx->param_oc_lds);
+		addr = get_tcs_tes_buffer_address(ctx, NULL,
+		                          lp_build_const_int32(gallivm, param));
+
+		value = buffer_load(&radeon_bld->soa.bld_base, TGSI_TYPE_FLOAT,
+		                    ~0, buffer, base, addr);
 
-		value = lds_load(&radeon_bld->soa.bld_base, TGSI_TYPE_FLOAT,
-				 ~0, dw_addr);
 		break;
 	}
 
-- 
2.8.3