[Mesa-dev] [PATCH 10/10] radeonsi: don't fetch 8 dwords for samplerBuffer and imageBuffer

Marek Olšák maraeo at gmail.com
Mon Nov 28 11:17:03 UTC 2016


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

The compiler doesn't shrink s_load_dwordx8, so we always wasted 4 SGPRs.
Also, the extraction of the descriptor created some really ugly asm code
with lots of VALU bitwise ops and v_readfirstlane.

Totals from *affected* shaders:
SGPRS: 13880 -> 13253 (-4.52 %)
VGPRS: 15200 -> 15088 (-0.74 %)
Code Size: 499864 -> 459816 (-8.01 %) bytes
Max Waves: 1554 -> 1564 (0.64 %)
---
 src/gallium/drivers/radeonsi/si_shader.c | 94 +++++++++++++++-----------------
 1 file changed, 43 insertions(+), 51 deletions(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index f4c6e9c..bb57e78 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -3107,30 +3107,30 @@ static void si_llvm_return_fs_outputs(struct lp_build_tgsi_context *bld_base)
  */
 static LLVMValueRef get_buffer_size(
 	struct lp_build_tgsi_context *bld_base,
 	LLVMValueRef descriptor)
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
 	struct gallivm_state *gallivm = bld_base->base.gallivm;
 	LLVMBuilderRef builder = gallivm->builder;
 	LLVMValueRef size =
 		LLVMBuildExtractElement(builder, descriptor,
-					lp_build_const_int32(gallivm, 6), "");
+					lp_build_const_int32(gallivm, 2), "");
 
 	if (ctx->screen->b.chip_class >= VI) {
 		/* On VI, the descriptor contains the size in bytes,
 		 * but TXQ must return the size in elements.
 		 * The stride is always non-zero for resources using TXQ.
 		 */
 		LLVMValueRef stride =
 			LLVMBuildExtractElement(builder, descriptor,
-						lp_build_const_int32(gallivm, 5), "");
+						lp_build_const_int32(gallivm, 1), "");
 		stride = LLVMBuildLShr(builder, stride,
 				       lp_build_const_int32(gallivm, 16), "");
 		stride = LLVMBuildAnd(builder, stride,
 				      lp_build_const_int32(gallivm, 0x3FFF), "");
 
 		size = LLVMBuildUDiv(builder, size, stride, "");
 	}
 
 	return size;
 }
@@ -3271,20 +3271,26 @@ static LLVMValueRef force_dcc_off(struct si_shader_context *ctx,
 		LLVMValueRef i32_6 = LLVMConstInt(ctx->i32, 6, 0);
 		LLVMValueRef i32_C = LLVMConstInt(ctx->i32, C_008F28_COMPRESSION_EN, 0);
 		LLVMValueRef tmp;
 
 		tmp = LLVMBuildExtractElement(builder, rsrc, i32_6, "");
 		tmp = LLVMBuildAnd(builder, tmp, i32_C, "");
 		return LLVMBuildInsertElement(builder, rsrc, tmp, i32_6, "");
 	}
 }
 
+static LLVMTypeRef const_array(LLVMTypeRef elem_type, int num_elements)
+{
+	return LLVMPointerType(LLVMArrayType(elem_type, num_elements),
+			       CONST_ADDR_SPACE);
+}
+
 /**
  * Load the resource descriptor for \p image.
  */
 static void
 image_fetch_rsrc(
 	struct lp_build_tgsi_context *bld_base,
 	const struct tgsi_full_src_register *image,
 	bool is_store, unsigned target,
 	LLVMValueRef *rsrc)
 {
@@ -3312,20 +3318,33 @@ image_fetch_rsrc(
 		 *    and if the index used to select an individual element is
 		 *    negative or greater than or equal to the size of the
 		 *    array, the results of the operation are undefined but may
 		 *    not lead to termination.
 		 */
 		index = get_bounded_indirect_index(ctx, &image->Indirect,
 						   image->Register.Index,
 						   SI_NUM_IMAGES);
 	}
 
+	if (target == TGSI_TEXTURE_BUFFER) {
+		LLVMBuilderRef builder = ctx->gallivm.builder;
+
+		rsrc_ptr = LLVMBuildPointerCast(builder, rsrc_ptr,
+						const_array(ctx->v4i32, 0), "");
+		index = LLVMBuildMul(builder, index,
+				     LLVMConstInt(ctx->i32, 2, 0), "");
+		index = LLVMBuildAdd(builder, index,
+				     LLVMConstInt(ctx->i32, 1, 0), "");
+		*rsrc = build_indexed_load_const(ctx, rsrc_ptr, index);
+		return;
+	}
+
 	tmp = build_indexed_load_const(ctx, rsrc_ptr, index);
 	if (dcc_off)
 		tmp = force_dcc_off(ctx, tmp);
 	*rsrc = tmp;
 }
 
 static LLVMValueRef image_fetch_coords(
 		struct lp_build_tgsi_context *bld_base,
 		const struct tgsi_full_instruction *inst,
 		unsigned src)
@@ -3387,39 +3406,20 @@ static void image_append_args(
 	}
 
 	/* HAVE_LLVM >= 0x0400 */
 	emit_data->args[emit_data->arg_count++] = glc;
 	emit_data->args[emit_data->arg_count++] = slc;
 	emit_data->args[emit_data->arg_count++] = lwe;
 	emit_data->args[emit_data->arg_count++] = da;
 }
 
 /**
- * Given a 256 bit resource, extract the top half (which stores the buffer
- * resource in the case of textures and images).
- */
-static LLVMValueRef extract_rsrc_top_half(
-		struct si_shader_context *ctx,
-		LLVMValueRef rsrc)
-{
-	struct gallivm_state *gallivm = &ctx->gallivm;
-	struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
-	LLVMTypeRef v2i128 = LLVMVectorType(ctx->i128, 2);
-
-	rsrc = LLVMBuildBitCast(gallivm->builder, rsrc, v2i128, "");
-	rsrc = LLVMBuildExtractElement(gallivm->builder, rsrc, bld_base->uint_bld.one, "");
-	rsrc = LLVMBuildBitCast(gallivm->builder, rsrc, ctx->v4i32, "");
-
-	return rsrc;
-}
-
-/**
  * Append the resource and indexing arguments for buffer intrinsics.
  *
  * \param rsrc the v4i32 buffer resource
  * \param index index into the buffer (stride-based)
  * \param offset byte offset into the buffer
  */
 static void buffer_append_args(
 		struct si_shader_context *ctx,
 		struct lp_build_emit_data *emit_data,
 		LLVMValueRef rsrc,
@@ -3466,21 +3466,20 @@ static void load_fetch_args(
 
 		buffer_append_args(ctx, emit_data, rsrc, bld_base->uint_bld.zero,
 				   offset, false);
 	} else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE) {
 		LLVMValueRef coords;
 
 		image_fetch_rsrc(bld_base, &inst->Src[0], false, target, &rsrc);
 		coords = image_fetch_coords(bld_base, inst, 1);
 
 		if (target == TGSI_TEXTURE_BUFFER) {
-			rsrc = extract_rsrc_top_half(ctx, rsrc);
 			buffer_append_args(ctx, emit_data, rsrc, coords,
 					bld_base->uint_bld.zero, false);
 		} else {
 			emit_data->args[0] = coords;
 			emit_data->args[1] = rsrc;
 			emit_data->args[2] = lp_build_const_int32(gallivm, 15); /* dmask */
 			emit_data->arg_count = 3;
 
 			image_append_args(ctx, emit_data, target, false);
 		}
@@ -3674,22 +3673,20 @@ static void store_fetch_args(
 		buffer_append_args(ctx, emit_data, rsrc, bld_base->uint_bld.zero,
 				   offset, false);
 	} else if (inst->Dst[0].Register.File == TGSI_FILE_IMAGE) {
 		unsigned target = inst->Memory.Texture;
 		LLVMValueRef coords;
 
 		coords = image_fetch_coords(bld_base, inst, 0);
 
 		if (target == TGSI_TEXTURE_BUFFER) {
 			image_fetch_rsrc(bld_base, &memory, true, target, &rsrc);
-
-			rsrc = extract_rsrc_top_half(ctx, rsrc);
 			buffer_append_args(ctx, emit_data, rsrc, coords,
 					bld_base->uint_bld.zero, false);
 		} else {
 			emit_data->args[1] = coords;
 			image_fetch_rsrc(bld_base, &memory, true, target,
 					 &emit_data->args[2]);
 			emit_data->args[3] = lp_build_const_int32(gallivm, 15); /* dmask */
 			emit_data->arg_count = 4;
 
 			image_append_args(ctx, emit_data, target, false);
@@ -3878,21 +3875,20 @@ static void atomic_fetch_args(
 		buffer_append_args(ctx, emit_data, rsrc, bld_base->uint_bld.zero,
 				   offset, true);
 	} else if (inst->Src[0].Register.File == TGSI_FILE_IMAGE) {
 		unsigned target = inst->Memory.Texture;
 		LLVMValueRef coords;
 
 		image_fetch_rsrc(bld_base, &inst->Src[0], true, target, &rsrc);
 		coords = image_fetch_coords(bld_base, inst, 1);
 
 		if (target == TGSI_TEXTURE_BUFFER) {
-			rsrc = extract_rsrc_top_half(ctx, rsrc);
 			buffer_append_args(ctx, emit_data, rsrc, coords,
 					   bld_base->uint_bld.zero, true);
 		} else {
 			emit_data->args[emit_data->arg_count++] = coords;
 			emit_data->args[emit_data->arg_count++] = rsrc;
 
 			image_append_args(ctx, emit_data, target, true);
 		}
 	}
 }
@@ -4122,45 +4118,47 @@ static void set_tex_fetch_args(struct si_shader_context *ctx,
 	emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* tfe */
 	emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* lwe */
 
 	emit_data->arg_count = num_args;
 }
 
 static const struct lp_build_tgsi_action tex_action;
 
 enum desc_type {
 	DESC_IMAGE,
+	DESC_BUFFER,
 	DESC_FMASK,
-	DESC_SAMPLER
+	DESC_SAMPLER,
 };
 
-static LLVMTypeRef const_array(LLVMTypeRef elem_type, int num_elements)
-{
-	return LLVMPointerType(LLVMArrayType(elem_type, num_elements),
-			       CONST_ADDR_SPACE);
-}
-
 /**
  * Load an image view, fmask view. or sampler state descriptor.
  */
 static LLVMValueRef load_sampler_desc_custom(struct si_shader_context *ctx,
 					     LLVMValueRef list, LLVMValueRef index,
 					     enum desc_type type)
 {
 	struct gallivm_state *gallivm = &ctx->gallivm;
 	LLVMBuilderRef builder = gallivm->builder;
 
 	switch (type) {
 	case DESC_IMAGE:
 		/* The image is at [0:7]. */
 		index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 2, 0), "");
 		break;
+	case DESC_BUFFER:
+		/* The buffer is in [4:7]. */
+		index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
+		index = LLVMBuildAdd(builder, index, LLVMConstInt(ctx->i32, 1, 0), "");
+		list = LLVMBuildPointerCast(builder, list,
+					    const_array(ctx->v4i32, 0), "");
+		break;
 	case DESC_FMASK:
 		/* The FMASK is at [8:15]. */
 		index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 2, 0), "");
 		index = LLVMBuildAdd(builder, index, LLVMConstInt(ctx->i32, 1, 0), "");
 		break;
 	case DESC_SAMPLER:
 		/* The sampler state is at [12:15]. */
 		index = LLVMBuildMul(builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
 		index = LLVMBuildAdd(builder, index, LLVMConstInt(ctx->i32, 3, 0), "");
 		list = LLVMBuildPointerCast(builder, list,
@@ -4228,56 +4226,57 @@ static void tex_fetch_ptrs(
 		const struct tgsi_full_src_register *reg = &emit_data->inst->Src[sampler_src];
 
 		index = get_bounded_indirect_index(ctx,
 						   &reg->Indirect,
 						   reg->Register.Index,
 						   SI_NUM_SAMPLERS);
 	} else {
 		index = LLVMConstInt(ctx->i32, sampler_index, 0);
 	}
 
-	*res_ptr = load_sampler_desc(ctx, index, DESC_IMAGE);
+	if (target == TGSI_TEXTURE_BUFFER)
+		*res_ptr = load_sampler_desc(ctx, index, DESC_BUFFER);
+	else
+		*res_ptr = load_sampler_desc(ctx, index, DESC_IMAGE);
+
+	if (samp_ptr)
+		*samp_ptr = NULL;
+	if (fmask_ptr)
+		*fmask_ptr = NULL;
 
 	if (target == TGSI_TEXTURE_2D_MSAA ||
 	    target == TGSI_TEXTURE_2D_ARRAY_MSAA) {
-		if (samp_ptr)
-			*samp_ptr = NULL;
 		if (fmask_ptr)
 			*fmask_ptr = load_sampler_desc(ctx, index, DESC_FMASK);
-	} else {
+	} else if (target != TGSI_TEXTURE_BUFFER) {
 		if (samp_ptr) {
 			*samp_ptr = load_sampler_desc(ctx, index, DESC_SAMPLER);
 			*samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
 		}
-		if (fmask_ptr)
-			*fmask_ptr = NULL;
 	}
 }
 
 static void txq_fetch_args(
 	struct lp_build_tgsi_context *bld_base,
 	struct lp_build_emit_data *emit_data)
 {
 	struct si_shader_context *ctx = si_shader_context(bld_base);
-	struct gallivm_state *gallivm = bld_base->base.gallivm;
-	LLVMBuilderRef builder = gallivm->builder;
 	const struct tgsi_full_instruction *inst = emit_data->inst;
 	unsigned target = inst->Texture.Texture;
 	LLVMValueRef res_ptr;
 	LLVMValueRef address;
 
 	tex_fetch_ptrs(bld_base, emit_data, &res_ptr, NULL, NULL);
 
 	if (target == TGSI_TEXTURE_BUFFER) {
 		/* Read the size from the buffer descriptor directly. */
-		LLVMValueRef res = LLVMBuildBitCast(builder, res_ptr, ctx->v8i32, "");
-		emit_data->args[0] = get_buffer_size(bld_base, res);
+		emit_data->args[0] = get_buffer_size(bld_base, res_ptr);
 		return;
 	}
 
 	/* Textures - set the mip level. */
 	address = lp_build_emit_fetch(bld_base, inst, 0, TGSI_CHAN_X);
 
 	set_tex_fetch_args(ctx, emit_data, TGSI_OPCODE_TXQ, target, res_ptr,
 			   NULL, &address, 1, 0xf);
 }
 
@@ -4331,30 +4330,23 @@ static void tex_fetch_args(
 	unsigned count = 0;
 	unsigned chan;
 	unsigned num_deriv_channels = 0;
 	bool has_offset = inst->Texture.NumOffsets > 0;
 	LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL;
 	unsigned dmask = 0xf;
 
 	tex_fetch_ptrs(bld_base, emit_data, &res_ptr, &samp_ptr, &fmask_ptr);
 
 	if (target == TGSI_TEXTURE_BUFFER) {
-		LLVMTypeRef v2i128 = LLVMVectorType(ctx->i128, 2);
-
-		/* Bitcast and truncate v8i32 to v16i8. */
-		LLVMValueRef res = res_ptr;
-		res = LLVMBuildBitCast(gallivm->builder, res, v2i128, "");
-		res = LLVMBuildExtractElement(gallivm->builder, res, bld_base->uint_bld.one, "");
-		res = LLVMBuildBitCast(gallivm->builder, res, ctx->v16i8, "");
-
 		emit_data->dst_type = ctx->v4f32;
-		emit_data->args[0] = res;
+		emit_data->args[0] = LLVMBuildBitCast(gallivm->builder, res_ptr,
+						      ctx->v16i8, "");
 		emit_data->args[1] = bld_base->uint_bld.zero;
 		emit_data->args[2] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_X);
 		emit_data->arg_count = 3;
 		return;
 	}
 
 	/* Fetch and project texture coordinates */
 	coords[3] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
 	for (chan = 0; chan < 3; chan++ ) {
 		coords[chan] = lp_build_emit_fetch(bld_base,
-- 
2.7.4



More information about the mesa-dev mailing list