[Mesa-dev] [PATCH 59/61] radeonsi: pass tessellation ring addresses via user SGPRs

Marek Olšák maraeo at gmail.com
Mon Apr 24 08:45:56 UTC 2017


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

This removes s_load_dword latency for tess rings.

We need just 1 SGPR for the address if we use 64K alignment. The final asm
for recreating the descriptor is:

    // s2 is (address >> 16)
    s_mov_b32 s3, 0
    s_lshl_b64 s[4:5], s[2:3], 16
    s_mov_b32 s6, -1
    s_mov_b32 s7, 0x27fac
---
 src/gallium/drivers/radeonsi/si_pm4.h             |   2 +-
 src/gallium/drivers/radeonsi/si_shader.c          | 102 +++++++++++++---------
 src/gallium/drivers/radeonsi/si_shader.h          |   5 ++
 src/gallium/drivers/radeonsi/si_shader_internal.h |   2 +
 src/gallium/drivers/radeonsi/si_state.h           |   3 -
 src/gallium/drivers/radeonsi/si_state_draw.c      |   3 +-
 src/gallium/drivers/radeonsi/si_state_shaders.c   |  50 ++++++++---
 7 files changed, 111 insertions(+), 56 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_pm4.h b/src/gallium/drivers/radeonsi/si_pm4.h
index 106abe1..189c481 100644
--- a/src/gallium/drivers/radeonsi/si_pm4.h
+++ b/src/gallium/drivers/radeonsi/si_pm4.h
@@ -23,21 +23,21 @@
  * Authors:
  *      Christian König <christian.koenig at amd.com>
  */
 
 #ifndef SI_PM4_H
 #define SI_PM4_H
 
 #include "radeon/radeon_winsys.h"
 
 #define SI_PM4_MAX_DW		176
-#define SI_PM4_MAX_BO		1
+#define SI_PM4_MAX_BO		3
 
 // forward defines
 struct si_context;
 enum chip_class;
 
 struct si_pm4_state
 {
 	/* optional indirect buffer */
 	struct r600_resource	*indirect_buffer;
 
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 974be02..57bcd6b 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -904,20 +904,43 @@ static void lds_store(struct lp_build_tgsi_context *bld_base,
 	struct gallivm_state *gallivm = &ctx->gallivm;
 
 	dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
 			    LLVMConstInt(ctx->i32, dw_offset_imm, 0));
 
 	value = LLVMBuildBitCast(gallivm->builder, value, ctx->i32, "");
 	ac_build_indexed_store(&ctx->ac, ctx->lds,
 			       dw_addr, value);
 }
 
+static LLVMValueRef desc_from_addr_base64k(struct si_shader_context *ctx,
+						  unsigned param)
+{
+	LLVMBuilderRef builder = ctx->gallivm.builder;
+
+	LLVMValueRef addr = LLVMGetParam(ctx->main_fn, param);
+	addr = LLVMBuildZExt(builder, addr, ctx->i64, "");
+	addr = LLVMBuildShl(builder, addr, LLVMConstInt(ctx->i64, 16, 0), "");
+
+	uint64_t desc2 = 0xffffffff;
+	uint64_t desc3 = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
+			 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
+		         S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
+		         S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
+			 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
+		         S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
+	LLVMValueRef hi = LLVMConstInt(ctx->i64, desc2 | (desc3 << 32), 0);
+
+	LLVMValueRef desc = LLVMGetUndef(LLVMVectorType(ctx->i64, 2));
+	desc = LLVMBuildInsertElement(builder, desc, addr, ctx->i32_0, "");
+	return LLVMBuildInsertElement(builder, desc, hi, ctx->i32_1, "");
+}
+
 static LLVMValueRef fetch_input_tcs(
 	struct lp_build_tgsi_context *bld_base,
 	const struct tgsi_full_src_register *reg,
 	enum tgsi_opcode_type type, unsigned swizzle)
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 	LLVMValueRef dw_addr, stride;
 
 	stride = unpack_param(ctx, ctx->param_vs_state_bits, 24, 8);
 	dw_addr = get_tcs_in_current_patch_offset(ctx);
@@ -945,45 +968,42 @@ static LLVMValueRef fetch_output_tcs(
 
 	return lds_load(bld_base, type, swizzle, dw_addr);
 }
 
 static LLVMValueRef fetch_input_tes(
 	struct lp_build_tgsi_context *bld_base,
 	const struct tgsi_full_src_register *reg,
 	enum tgsi_opcode_type type, unsigned swizzle)
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
-	LLVMValueRef rw_buffers, buffer, base, addr;
+	LLVMValueRef buffer, base, addr;
 
-	rw_buffers = LLVMGetParam(ctx->main_fn,
-				  ctx->param_rw_buffers);
-	buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
-			LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
+	buffer = desc_from_addr_base64k(ctx, ctx->param_tcs_offchip_addr_base64k);
 
 	base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
 	addr = get_tcs_tes_buffer_address_from_reg(ctx, NULL, reg);
 
 	return buffer_load(bld_base, type, swizzle, buffer, base, addr, true);
 }
 
 static void store_output_tcs(struct lp_build_tgsi_context *bld_base,
 			     const struct tgsi_full_instruction *inst,
 			     const struct tgsi_opcode_info *info,
 			     LLVMValueRef dst[4])
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 	struct gallivm_state *gallivm = &ctx->gallivm;
 	const struct tgsi_full_dst_register *reg = &inst->Dst[0];
 	const struct tgsi_shader_info *sh_info = &ctx->shader->selector->info;
 	unsigned chan_index;
 	LLVMValueRef dw_addr, stride;
-	LLVMValueRef rw_buffers, buffer, base, buf_addr;
+	LLVMValueRef buffer, base, buf_addr;
 	LLVMValueRef values[4];
 	bool skip_lds_store;
 	bool is_tess_factor = false;
 
 	/* Only handle per-patch and per-vertex outputs here.
 	 * Vectors will be lowered to scalars and this function will be called again.
 	 */
 	if (reg->Register.File != TGSI_FILE_OUTPUT ||
 	    (dst[0] && LLVMGetTypeKind(LLVMTypeOf(dst[0])) == LLVMVectorTypeKind)) {
 		si_llvm_emit_store(bld_base, inst, info, dst);
@@ -1005,24 +1025,21 @@ static void store_output_tcs(struct lp_build_tgsi_context *bld_base,
 
 			/* Always write tess factors into LDS for the TCS epilog. */
 			if (name == TGSI_SEMANTIC_TESSINNER ||
 			    name == TGSI_SEMANTIC_TESSOUTER) {
 				skip_lds_store = false;
 				is_tess_factor = true;
 			}
 		}
 	}
 
-	rw_buffers = LLVMGetParam(ctx->main_fn,
-				  ctx->param_rw_buffers);
-	buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
-			LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
+	buffer = desc_from_addr_base64k(ctx, ctx->param_tcs_offchip_addr_base64k);
 
 	base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
 	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];
 
 		if (inst->Instruction.Saturate)
 			value = ac_build_clamp(&ctx->ac, value);
@@ -1528,27 +1545,24 @@ static void declare_system_value(struct si_shader_context *ctx,
 			value = unpack_param(ctx, ctx->param_tcs_out_lds_layout, 26, 6);
 		else if (ctx->type == PIPE_SHADER_TESS_EVAL)
 			value = unpack_param(ctx, ctx->param_tcs_offchip_layout, 9, 7);
 		else
 			assert(!"invalid shader stage for TGSI_SEMANTIC_VERTICESIN");
 		break;
 
 	case TGSI_SEMANTIC_TESSINNER:
 	case TGSI_SEMANTIC_TESSOUTER:
 	{
-		LLVMValueRef rw_buffers, buffer, base, addr;
+		LLVMValueRef buffer, base, addr;
 		int param = si_shader_io_get_unique_index(decl->Semantic.Name, 0);
 
-		rw_buffers = LLVMGetParam(ctx->main_fn,
-					  ctx->param_rw_buffers);
-		buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
-		        LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
+		buffer = desc_from_addr_base64k(ctx, ctx->param_tcs_offchip_addr_base64k);
 
 		base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
 		addr = get_tcs_tes_buffer_address(ctx, get_rel_patch_id(ctx), NULL,
 		                          LLVMConstInt(ctx->i32, param, 0));
 
 		value = buffer_load(&ctx->bld_base, TGSI_TYPE_FLOAT,
 		                    ~0, buffer, base, addr, true);
 
 		break;
 	}
@@ -2424,30 +2438,26 @@ handle_semantic:
 }
 
 /**
  * Forward all outputs from the vertex shader to the TES. This is only used
  * for the fixed function TCS.
  */
 static void si_copy_tcs_inputs(struct lp_build_tgsi_context *bld_base)
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 	struct gallivm_state *gallivm = &ctx->gallivm;
-	LLVMValueRef invocation_id, rw_buffers, buffer, buffer_offset;
+	LLVMValueRef invocation_id, buffer, buffer_offset;
 	LLVMValueRef lds_vertex_stride, lds_vertex_offset, lds_base;
 	uint64_t inputs;
 
 	invocation_id = unpack_param(ctx, ctx->param_tcs_rel_ids, 8, 5);
-
-	rw_buffers = LLVMGetParam(ctx->main_fn, ctx->param_rw_buffers);
-	buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
-	                LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
-
+	buffer = desc_from_addr_base64k(ctx, ctx->param_tcs_offchip_addr_base64k);
 	buffer_offset = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
 
 	lds_vertex_stride = unpack_param(ctx, ctx->param_vs_state_bits, 24, 8);
 	lds_vertex_offset = LLVMBuildMul(gallivm->builder, invocation_id,
 	                                 lds_vertex_stride, "");
 	lds_base = get_tcs_in_current_patch_offset(ctx);
 	lds_base = LLVMBuildAdd(gallivm->builder, lds_base, lds_vertex_offset, "");
 
 	inputs = ctx->shader->key.mono.ff_tcs_inputs_to_copy;
 	while (inputs) {
@@ -2473,21 +2483,21 @@ static void si_copy_tcs_inputs(struct lp_build_tgsi_context *bld_base)
 static void si_write_tess_factors(struct lp_build_tgsi_context *bld_base,
 				  LLVMValueRef rel_patch_id,
 				  LLVMValueRef invocation_id,
 				  LLVMValueRef tcs_out_current_patch_data_offset)
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 	struct gallivm_state *gallivm = &ctx->gallivm;
 	struct si_shader *shader = ctx->shader;
 	unsigned tess_inner_index, tess_outer_index;
 	LLVMValueRef lds_base, lds_inner, lds_outer, byteoffset, buffer;
-	LLVMValueRef out[6], vec0, vec1, rw_buffers, tf_base, inner[4], outer[4];
+	LLVMValueRef out[6], vec0, vec1, tf_base, inner[4], outer[4];
 	unsigned stride, outer_comps, inner_comps, i, offset;
 	struct lp_build_if_state if_ctx, inner_if_ctx;
 
 	si_llvm_emit_barrier(NULL, bld_base, NULL);
 
 	/* Do this only for invocation 0, because the tess levels are per-patch,
 	 * not per-vertex.
 	 *
 	 * This can't jump, because invocation 0 executes this. It should
 	 * at least mask out the loads and stores for other invocations.
@@ -2555,24 +2565,21 @@ static void si_write_tess_factors(struct lp_build_tgsi_context *bld_base,
 	}
 
 	/* Convert the outputs to vectors for stores. */
 	vec0 = lp_build_gather_values(gallivm, out, MIN2(stride, 4));
 	vec1 = NULL;
 
 	if (stride > 4)
 		vec1 = lp_build_gather_values(gallivm, out+4, stride - 4);
 
 	/* Get the buffer. */
-	rw_buffers = LLVMGetParam(ctx->main_fn,
-				  ctx->param_rw_buffers);
-	buffer = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
-			LLVMConstInt(ctx->i32, SI_HS_RING_TESS_FACTOR, 0));
+	buffer = desc_from_addr_base64k(ctx, ctx->param_tcs_factor_addr_base64k);
 
 	/* Get the offset. */
 	tf_base = LLVMGetParam(ctx->main_fn,
 			       ctx->param_tcs_factor_offset);
 	byteoffset = LLVMBuildMul(gallivm->builder, rel_patch_id,
 				  LLVMConstInt(ctx->i32, 4 * stride, 0), "");
 
 	lp_build_if(&inner_if_ctx, gallivm,
 		    LLVMBuildICmp(gallivm->builder, LLVMIntEQ,
 				  rel_patch_id, ctx->i32_0, ""));
@@ -2598,22 +2605,21 @@ static void si_write_tess_factors(struct lp_build_tgsi_context *bld_base,
 		ac_build_buffer_store_dword(&ctx->ac, buffer, vec1,
 					    stride - 4, byteoffset, tf_base,
 					    offset, 1, 0, true, false);
 
 	/* Store the tess factors into the offchip buffer if TES reads them. */
 	if (shader->key.part.tcs.epilog.tes_reads_tess_factors) {
 		LLVMValueRef buf, base, inner_vec, outer_vec, tf_outer_offset;
 		LLVMValueRef tf_inner_offset;
 		unsigned param_outer, param_inner;
 
-		buf = ac_build_indexed_load_const(&ctx->ac, rw_buffers,
-				LLVMConstInt(ctx->i32, SI_HS_RING_TESS_OFFCHIP, 0));
+		buf = desc_from_addr_base64k(ctx, ctx->param_tcs_offchip_addr_base64k);
 		base = LLVMGetParam(ctx->main_fn, ctx->param_tcs_offchip_offset);
 
 		param_outer = si_shader_io_get_unique_index(
 				      TGSI_SEMANTIC_TESSOUTER, 0);
 		tf_outer_offset = get_tcs_tes_buffer_address(ctx, rel_patch_id, NULL,
 					LLVMConstInt(ctx->i32, param_outer, 0));
 
 		outer_vec = lp_build_gather_values(gallivm, outer,
 						   util_next_power_of_two(outer_comps));
 
@@ -2684,33 +2690,38 @@ static void si_llvm_emit_tcs_epilogue(struct lp_build_tgsi_context *bld_base)
 
 	rel_patch_id = get_rel_patch_id(ctx);
 	invocation_id = unpack_param(ctx, ctx->param_tcs_rel_ids, 8, 5);
 	tf_lds_offset = get_tcs_out_current_patch_data_offset(ctx);
 
 	/* Return epilog parameters from this function. */
 	LLVMBuilderRef builder = ctx->gallivm.builder;
 	LLVMValueRef ret = ctx->return_value;
 	unsigned vgpr;
 
-	ret = si_insert_input_ptr_as_2xi32(ctx, ret,
-					   ctx->param_rw_buffers, 0);
-
 	if (ctx->screen->b.chip_class >= GFX9) {
 		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_offchip_addr_base64k,
+					  8 + GFX9_SGPR_TCS_OFFCHIP_ADDR_BASE64K);
+		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_addr_base64k,
+					  8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K);
 		/* Tess offchip and tess factor offsets are at the beginning. */
 		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_offset, 2);
 		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_offset, 4);
-		vgpr = 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT + 1;
+		vgpr = 8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K + 1;
 	} else {
 		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_layout,
 					  GFX6_SGPR_TCS_OFFCHIP_LAYOUT);
+		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_addr_base64k,
+					  GFX6_SGPR_TCS_OFFCHIP_ADDR_BASE64K);
+		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_addr_base64k,
+					  GFX6_SGPR_TCS_FACTOR_ADDR_BASE64K);
 		/* Tess offchip and tess factor offsets are after user SGPRs. */
 		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_offset,
 					  GFX6_TCS_NUM_USER_SGPR);
 		ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_offset,
 					  GFX6_TCS_NUM_USER_SGPR + 1);
 		vgpr = GFX6_TCS_NUM_USER_SGPR + 2;
 	}
 
 	/* VGPRs */
 	rel_patch_id = bitcast(bld_base, TGSI_TYPE_FLOAT, rel_patch_id);
@@ -2735,22 +2746,26 @@ static void si_set_ls_return_value_for_tcs(struct si_shader_context *ctx)
 	ret = si_insert_input_ret(ctx, ret, ctx->param_merged_scratch_offset, 5);
 
 	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);
+	ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_offchip_addr_base64k,
+				  8 + GFX9_SGPR_TCS_OFFCHIP_ADDR_BASE64K);
+	ret = si_insert_input_ret(ctx, ret, ctx->param_tcs_factor_addr_base64k,
+				  8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K);
 
-	unsigned desc_param = ctx->param_tcs_out_lds_layout + 2;
+	unsigned desc_param = ctx->param_tcs_factor_addr_base64k + 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;
@@ -5914,20 +5929,22 @@ static void create_function(struct si_shader_context *ctx)
 		declare_vs_input_vgprs(ctx, params, &num_params,
 				       &num_prolog_vgprs);
 		break;
 
 	case PIPE_SHADER_TESS_CTRL: /* SI-CI-VI */
 		declare_default_desc_pointers(ctx, params, &num_params);
 		params[ctx->param_tcs_offchip_layout = num_params++] = ctx->i32;
 		params[ctx->param_tcs_out_lds_offsets = num_params++] = ctx->i32;
 		params[ctx->param_tcs_out_lds_layout = num_params++] = ctx->i32;
 		params[ctx->param_vs_state_bits = num_params++] = ctx->i32;
+		params[ctx->param_tcs_offchip_addr_base64k = num_params++] = ctx->i32;
+		params[ctx->param_tcs_factor_addr_base64k = num_params++] = ctx->i32;
 		params[ctx->param_tcs_offchip_offset = num_params++] = ctx->i32;
 		params[ctx->param_tcs_factor_offset = num_params++] = ctx->i32;
 		last_sgpr = num_params - 1;
 
 		/* VGPRs */
 		params[ctx->param_tcs_patch_id = num_params++] = ctx->i32;
 		params[ctx->param_tcs_rel_ids = num_params++] = ctx->i32;
 
 		/* param_tcs_offchip_offset and param_tcs_factor_offset are
 		 * placed after the user SGPRs.
@@ -5951,20 +5968,22 @@ static void create_function(struct si_shader_context *ctx)
 
 		params[num_params++] = ctx->i32; /* unused */
 		params[num_params++] = ctx->i32; /* unused */
 		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;
 		params[ctx->param_tcs_out_lds_offsets = num_params++] = ctx->i32;
 		params[ctx->param_tcs_out_lds_layout = num_params++] = ctx->i32;
+		params[ctx->param_tcs_offchip_addr_base64k = num_params++] = ctx->i32;
+		params[ctx->param_tcs_factor_addr_base64k = num_params++] = ctx->i32;
 		params[num_params++] = ctx->i32; /* unused */
 
 		declare_per_stage_desc_pointers(ctx, params, &num_params,
 						ctx->type == PIPE_SHADER_TESS_CTRL);
 		last_sgpr = num_params - 1;
 
 		/* VGPRs (first TCS, then VS) */
 		params[ctx->param_tcs_patch_id = num_params++] = ctx->i32;
 		params[ctx->param_tcs_rel_ids = num_params++] = ctx->i32;
 
@@ -5977,21 +5996,21 @@ static void create_function(struct si_shader_context *ctx)
 				returns[num_returns++] = ctx->i32; /* SGPRs */
 			for (i = 0; i < 2; i++)
 				returns[num_returns++] = ctx->f32; /* VGPRs */
 		} else {
 			/* TCS return values are inputs to the TCS epilog.
 			 *
 			 * param_tcs_offchip_offset, param_tcs_factor_offset,
 			 * param_tcs_offchip_layout, and param_rw_buffers
 			 * should be passed to the epilog.
 			 */
-			for (i = 0; i <= 8 + GFX9_SGPR_TCS_OFFCHIP_LAYOUT; i++)
+			for (i = 0; i <= 8 + GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K; i++)
 				returns[num_returns++] = ctx->i32; /* SGPRs */
 			for (i = 0; i < 3; i++)
 				returns[num_returns++] = ctx->f32; /* VGPRs */
 		}
 		break;
 
 	case SI_SHADER_MERGED_VERTEX_OR_TESSEVAL_GEOMETRY:
 		/* Merged stages have 8 system SGPRs at the beginning. */
 		params[ctx->param_rw_buffers = num_params++] = /* SPI_SHADER_USER_DATA_ADDR_LO_GS */
 			const_array(ctx->v16i8, SI_NUM_RW_BUFFERS);
@@ -6006,21 +6025,21 @@ static void create_function(struct si_shader_context *ctx)
 		params[num_params++] = ctx->i32; /* unused */
 		declare_per_stage_desc_pointers(ctx, params, &num_params,
 						(ctx->type == PIPE_SHADER_VERTEX ||
 						 ctx->type == PIPE_SHADER_TESS_EVAL));
 		if (ctx->type == PIPE_SHADER_VERTEX) {
 			declare_vs_specific_input_sgprs(ctx, params, &num_params);
 		} else {
 			/* TESS_EVAL (and also GEOMETRY):
 			 * Declare as many input SGPRs as the VS has. */
 			params[ctx->param_tcs_offchip_layout = num_params++] = ctx->i32;
-			params[num_params++] = ctx->i32; /* unused */
+			params[ctx->param_tcs_offchip_addr_base64k = num_params++] = ctx->i32;
 			params[num_params++] = ctx->i32; /* unused */
 			params[num_params++] = ctx->i32; /* unused */
 			params[num_params++] = ctx->i32; /* unused */
 			params[ctx->param_vs_state_bits = num_params++] = ctx->i32; /* unused */
 		}
 
 		declare_per_stage_desc_pointers(ctx, params, &num_params,
 						ctx->type == PIPE_SHADER_GEOMETRY);
 		last_sgpr = num_params - 1;
 
@@ -6044,20 +6063,21 @@ static void create_function(struct si_shader_context *ctx)
 			for (i = 0; i < 8 + GFX9_GS_NUM_USER_SGPR; i++)
 				returns[num_returns++] = ctx->i32; /* SGPRs */
 			for (i = 0; i < 5; i++)
 				returns[num_returns++] = ctx->f32; /* VGPRs */
 		}
 		break;
 
 	case PIPE_SHADER_TESS_EVAL:
 		declare_default_desc_pointers(ctx, params, &num_params);
 		params[ctx->param_tcs_offchip_layout = num_params++] = ctx->i32;
+		params[ctx->param_tcs_offchip_addr_base64k = num_params++] = ctx->i32;
 
 		if (shader->key.as_es) {
 			params[ctx->param_tcs_offchip_offset = num_params++] = ctx->i32;
 			params[num_params++] = ctx->i32;
 			params[ctx->param_es2gs_offset = num_params++] = ctx->i32;
 		} else {
 			params[num_params++] = ctx->i32;
 			declare_streamout_params(ctx, &shader->selector->so,
 						 params, ctx->i32, &num_params);
 			params[ctx->param_tcs_offchip_offset = num_params++] = ctx->i32;
@@ -8578,51 +8598,55 @@ static bool si_shader_select_vs_parts(struct si_screen *sscreen,
  */
 static void si_build_tcs_epilog_function(struct si_shader_context *ctx,
 					 union si_shader_part_key *key)
 {
 	struct gallivm_state *gallivm = &ctx->gallivm;
 	struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
 	LLVMTypeRef params[32];
 	LLVMValueRef func;
 	int last_sgpr, num_params = 0;
 
-	/* Declare inputs. Only RW_BUFFERS and TESS_FACTOR_OFFSET are used. */
-	params[ctx->param_rw_buffers = num_params++] =
-		const_array(ctx->v16i8, SI_NUM_RW_BUFFERS);
-
 	if (ctx->screen->b.chip_class >= GFX9) {
+		params[num_params++] = ctx->i64;
 		params[ctx->param_tcs_offchip_offset = num_params++] = ctx->i32;
 		params[num_params++] = ctx->i32; /* wave info */
 		params[ctx->param_tcs_factor_offset = num_params++] = ctx->i32;
 		params[num_params++] = ctx->i32;
 		params[num_params++] = ctx->i32;
 		params[num_params++] = ctx->i32;
 		params[num_params++] = ctx->i64;
 		params[num_params++] = ctx->i64;
 		params[num_params++] = ctx->i64;
 		params[num_params++] = ctx->i64;
 		params[num_params++] = ctx->i64;
 		params[num_params++] = ctx->i64;
 		params[num_params++] = ctx->i32;
 		params[num_params++] = ctx->i32;
 		params[num_params++] = ctx->i32;
 		params[num_params++] = ctx->i32;
 		params[ctx->param_tcs_offchip_layout = num_params++] = ctx->i32;
+		params[num_params++] = ctx->i32;
+		params[num_params++] = ctx->i32;
+		params[ctx->param_tcs_offchip_addr_base64k = num_params++] = ctx->i32;
+		params[ctx->param_tcs_factor_addr_base64k = num_params++] = ctx->i32;
 	} else {
 		params[num_params++] = ctx->i64;
 		params[num_params++] = ctx->i64;
 		params[num_params++] = ctx->i64;
 		params[num_params++] = ctx->i64;
+		params[num_params++] = ctx->i64;
 		params[ctx->param_tcs_offchip_layout = num_params++] = ctx->i32;
 		params[num_params++] = ctx->i32;
 		params[num_params++] = ctx->i32;
 		params[num_params++] = ctx->i32;
+		params[ctx->param_tcs_offchip_addr_base64k = num_params++] = ctx->i32;
+		params[ctx->param_tcs_factor_addr_base64k = num_params++] = ctx->i32;
 		params[ctx->param_tcs_offchip_offset = num_params++] = ctx->i32;
 		params[ctx->param_tcs_factor_offset = num_params++] = ctx->i32;
 	}
 	last_sgpr = num_params - 1;
 
 	params[num_params++] = ctx->i32; /* patch index within the wave (REL_PATCH_ID) */
 	params[num_params++] = ctx->i32; /* invocation ID within the patch */
 	params[num_params++] = ctx->i32; /* LDS offset where tess factors should be loaded from */
 
 	/* Create the function. */
diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h
index d051715..c9bd904 100644
--- a/src/gallium/drivers/radeonsi/si_shader.h
+++ b/src/gallium/drivers/radeonsi/si_shader.h
@@ -166,33 +166,38 @@ enum {
 	SI_SGPR_VERTEX_BUFFERS	= SI_NUM_RESOURCE_SGPRS,
 	SI_SGPR_VERTEX_BUFFERS_HI,
 	SI_SGPR_BASE_VERTEX,
 	SI_SGPR_START_INSTANCE,
 	SI_SGPR_DRAWID,
 	SI_SGPR_VS_STATE_BITS,
 	SI_VS_NUM_USER_SGPR,
 
 	/* TES */
 	SI_SGPR_TES_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
+	SI_SGPR_TES_OFFCHIP_ADDR_BASE64K,
 	SI_TES_NUM_USER_SGPR,
 
 	/* GFX6-8: TCS only */
 	GFX6_SGPR_TCS_OFFCHIP_LAYOUT = SI_NUM_RESOURCE_SGPRS,
 	GFX6_SGPR_TCS_OUT_OFFSETS,
 	GFX6_SGPR_TCS_OUT_LAYOUT,
 	GFX6_SGPR_TCS_IN_LAYOUT,
+	GFX6_SGPR_TCS_OFFCHIP_ADDR_BASE64K,
+	GFX6_SGPR_TCS_FACTOR_ADDR_BASE64K,
 	GFX6_TCS_NUM_USER_SGPR,
 
 	/* GFX9: Merged LS-HS (VS-TCS) only. */
 	GFX9_SGPR_TCS_OFFCHIP_LAYOUT = SI_VS_NUM_USER_SGPR,
 	GFX9_SGPR_TCS_OUT_OFFSETS,
 	GFX9_SGPR_TCS_OUT_LAYOUT,
+	GFX9_SGPR_TCS_OFFCHIP_ADDR_BASE64K,
+	GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K,
 	GFX9_SGPR_unused_to_align_the_next_pointer,
 	GFX9_SGPR_TCS_CONST_BUFFERS,
 	GFX9_SGPR_TCS_CONST_BUFFERS_HI,
 	GFX9_SGPR_TCS_SAMPLERS,  /* images & sampler states interleaved */
 	GFX9_SGPR_TCS_SAMPLERS_HI,
 	GFX9_SGPR_TCS_IMAGES,
 	GFX9_SGPR_TCS_IMAGES_HI,
 	GFX9_SGPR_TCS_SHADER_BUFFERS,
 	GFX9_SGPR_TCS_SHADER_BUFFERS_HI,
 	GFX9_TCS_NUM_USER_SGPR,
diff --git a/src/gallium/drivers/radeonsi/si_shader_internal.h b/src/gallium/drivers/radeonsi/si_shader_internal.h
index 233a96f..8a9bf10 100644
--- a/src/gallium/drivers/radeonsi/si_shader_internal.h
+++ b/src/gallium/drivers/radeonsi/si_shader_internal.h
@@ -155,20 +155,22 @@ struct si_shader_context {
 	 */
 	int param_tcs_out_lds_offsets;
 	/* Layout of TCS outputs / TES inputs:
 	 *   [0:12] = stride between output patches in DW, num_outputs * num_vertices * 4
 	 *            max = 32*32*4
 	 *   [13:20] = stride between output vertices in DW = num_inputs * 4
 	 *             max = 32*4
 	 *   [26:31] = gl_PatchVerticesIn, max = 32
 	 */
 	int param_tcs_out_lds_layout;
+	int param_tcs_offchip_addr_base64k;
+	int param_tcs_factor_addr_base64k;
 	int param_tcs_offchip_offset;
 	int param_tcs_factor_offset;
 	int param_tcs_patch_id;
 	int param_tcs_rel_ids;
 
 	/* API TES */
 	int param_tes_u;
 	int param_tes_v;
 	int param_tes_rel_patch_id;
 	int param_tes_patch_id;
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index 6257299..629d614 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -157,23 +157,20 @@ union si_state_atoms {
 
 #define SI_NUM_ATOMS (sizeof(union si_state_atoms)/sizeof(struct r600_atom*))
 
 struct si_shader_data {
 	struct r600_atom	atom;
 	uint32_t		sh_base[SI_NUM_SHADERS];
 };
 
 /* Private read-write buffer slots. */
 enum {
-	SI_HS_RING_TESS_FACTOR,
-	SI_HS_RING_TESS_OFFCHIP,
-
 	SI_ES_RING_ESGS,
 	SI_GS_RING_ESGS,
 
 	SI_RING_GSVS,
 
 	SI_VS_STREAMOUT_BUF0,
 	SI_VS_STREAMOUT_BUF1,
 	SI_VS_STREAMOUT_BUF2,
 	SI_VS_STREAMOUT_BUF3,
 
diff --git a/src/gallium/drivers/radeonsi/si_state_draw.c b/src/gallium/drivers/radeonsi/si_state_draw.c
index 393f64f..6a756c4 100644
--- a/src/gallium/drivers/radeonsi/si_state_draw.c
+++ b/src/gallium/drivers/radeonsi/si_state_draw.c
@@ -260,22 +260,23 @@ static void si_emit_derived_tess_state(struct si_context *sctx,
 		/* Set userdata SGPRs for TCS. */
 		radeon_set_sh_reg_seq(cs,
 			R_00B430_SPI_SHADER_USER_DATA_HS_0 + GFX6_SGPR_TCS_OFFCHIP_LAYOUT * 4, 4);
 		radeon_emit(cs, offchip_layout);
 		radeon_emit(cs, tcs_out_offsets);
 		radeon_emit(cs, tcs_out_layout | (num_tcs_input_cp << 26));
 		radeon_emit(cs, tcs_in_layout);
 	}
 
 	/* Set userdata SGPRs for TES. */
-	radeon_set_sh_reg_seq(cs, tes_sh_base + SI_SGPR_TES_OFFCHIP_LAYOUT * 4, 1);
+	radeon_set_sh_reg_seq(cs, tes_sh_base + SI_SGPR_TES_OFFCHIP_LAYOUT * 4, 2);
 	radeon_emit(cs, offchip_layout);
+	radeon_emit(cs, r600_resource(sctx->tess_offchip_ring)->gpu_address >> 16);
 
 	ls_hs_config = S_028B58_NUM_PATCHES(*num_patches) |
 		       S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp) |
 		       S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp);
 
 	if (sctx->b.chip_class >= CIK)
 		radeon_set_context_reg_idx(cs, R_028B58_VGT_LS_HS_CONFIG, 2,
 					   ls_hs_config);
 	else
 		radeon_set_context_reg(cs, R_028B58_VGT_LS_HS_CONFIG,
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index f9cd5c3..278766f 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -2758,80 +2758,106 @@ static void si_init_tess_factor_ring(struct si_context *sctx)
 	case VI:
 	case GFX9:
 		max_offchip_buffers = MIN2(max_offchip_buffers, 508);
 		break;
 	default:
 		assert(0);
 		return;
 	}
 
 	assert(!sctx->tf_ring);
+	/* Use 64K alignment for both rings, so that we can pass the address
+	 * to shaders as one SGPR containing bits [16:47].
+	 */
 	sctx->tf_ring = r600_aligned_buffer_create(sctx->b.b.screen,
 						   R600_RESOURCE_FLAG_UNMAPPABLE,
 						   PIPE_USAGE_DEFAULT,
 						   32768 * sctx->screen->b.info.max_se,
-						   256);
+						   64 * 1024);
 	if (!sctx->tf_ring)
 		return;
 
 	assert(((sctx->tf_ring->width0 / 4) & C_030938_SIZE) == 0);
 
 	sctx->tess_offchip_ring =
 		r600_aligned_buffer_create(sctx->b.b.screen,
 					   R600_RESOURCE_FLAG_UNMAPPABLE,
 					   PIPE_USAGE_DEFAULT,
 					   max_offchip_buffers *
 					   sctx->screen->tess_offchip_block_dw_size * 4,
-					   256);
+					   64 * 1024);
 	if (!sctx->tess_offchip_ring)
 		return;
 
 	si_init_config_add_vgt_flush(sctx);
 
+	uint64_t offchip_va = r600_resource(sctx->tess_offchip_ring)->gpu_address;
+	uint64_t factor_va = r600_resource(sctx->tf_ring)->gpu_address;
+	assert((offchip_va & 0xffff) == 0);
+	assert((factor_va & 0xffff) == 0);
+
+	si_pm4_add_bo(sctx->init_config, r600_resource(sctx->tess_offchip_ring),
+		      RADEON_USAGE_READWRITE, RADEON_PRIO_SHADER_RINGS);
+	si_pm4_add_bo(sctx->init_config, r600_resource(sctx->tf_ring),
+		      RADEON_USAGE_READWRITE, RADEON_PRIO_SHADER_RINGS);
+
 	/* Append these registers to the init config state. */
 	if (sctx->b.chip_class >= CIK) {
 		if (sctx->b.chip_class >= VI)
 			--max_offchip_buffers;
 
 		si_pm4_set_reg(sctx->init_config, R_030938_VGT_TF_RING_SIZE,
 			       S_030938_SIZE(sctx->tf_ring->width0 / 4));
 		si_pm4_set_reg(sctx->init_config, R_030940_VGT_TF_MEMORY_BASE,
-			       r600_resource(sctx->tf_ring)->gpu_address >> 8);
+			       factor_va >> 8);
 		if (sctx->b.chip_class >= GFX9)
 			si_pm4_set_reg(sctx->init_config, R_030944_VGT_TF_MEMORY_BASE_HI,
-				       r600_resource(sctx->tf_ring)->gpu_address >> 40);
+				       factor_va >> 40);
 		si_pm4_set_reg(sctx->init_config, R_03093C_VGT_HS_OFFCHIP_PARAM,
 		             S_03093C_OFFCHIP_BUFFERING(max_offchip_buffers) |
 		             S_03093C_OFFCHIP_GRANULARITY(offchip_granularity));
 	} else {
 		assert(offchip_granularity == V_03093C_X_8K_DWORDS);
 		si_pm4_set_reg(sctx->init_config, R_008988_VGT_TF_RING_SIZE,
 			       S_008988_SIZE(sctx->tf_ring->width0 / 4));
 		si_pm4_set_reg(sctx->init_config, R_0089B8_VGT_TF_MEMORY_BASE,
-			       r600_resource(sctx->tf_ring)->gpu_address >> 8);
+			       factor_va >> 8);
 		si_pm4_set_reg(sctx->init_config, R_0089B0_VGT_HS_OFFCHIP_PARAM,
 		               S_0089B0_OFFCHIP_BUFFERING(max_offchip_buffers));
 	}
 
+	if (sctx->b.chip_class >= GFX9) {
+		si_pm4_set_reg(sctx->init_config,
+			       R_00B430_SPI_SHADER_USER_DATA_LS_0 +
+			       GFX9_SGPR_TCS_OFFCHIP_ADDR_BASE64K * 4,
+			       offchip_va >> 16);
+		si_pm4_set_reg(sctx->init_config,
+			       R_00B430_SPI_SHADER_USER_DATA_LS_0 +
+			       GFX9_SGPR_TCS_FACTOR_ADDR_BASE64K * 4,
+			       factor_va >> 16);
+	} else {
+		si_pm4_set_reg(sctx->init_config,
+			       R_00B430_SPI_SHADER_USER_DATA_HS_0 +
+			       GFX6_SGPR_TCS_OFFCHIP_ADDR_BASE64K * 4,
+			       offchip_va >> 16);
+		si_pm4_set_reg(sctx->init_config,
+			       R_00B430_SPI_SHADER_USER_DATA_HS_0 +
+			       GFX6_SGPR_TCS_FACTOR_ADDR_BASE64K * 4,
+			       factor_va >> 16);
+	}
+
 	/* Flush the context to re-emit the init_config state.
 	 * This is done only once in a lifetime of a context.
 	 */
 	si_pm4_upload_indirect_buffer(sctx, sctx->init_config);
 	sctx->b.initial_gfx_cs_size = 0; /* force flush */
 	si_context_gfx_flush(sctx, RADEON_FLUSH_ASYNC, NULL);
-
-	si_set_ring_buffer(&sctx->b.b, SI_HS_RING_TESS_FACTOR, sctx->tf_ring,
-			   0, sctx->tf_ring->width0, false, false, 0, 0, 0);
-
-	si_set_ring_buffer(&sctx->b.b, SI_HS_RING_TESS_OFFCHIP,
-	                   sctx->tess_offchip_ring, 0,
-	                   sctx->tess_offchip_ring->width0, false, false, 0, 0, 0);
 }
 
 /**
  * This is used when TCS is NULL in the VS->TCS->TES chain. In this case,
  * VS passes its outputs to TES directly, so the fixed-function shader only
  * has to write TESSOUTER and TESSINNER.
  */
 static void si_generate_fixed_func_tcs(struct si_context *sctx)
 {
 	struct ureg_src outer, inner;
-- 
2.7.4



More information about the mesa-dev mailing list