From: Connor Abbott <cwabbott0@gmail.com>
Date: Wed, 12 Jun 2019 16:58:13 +0000 (+0200)
Subject: ac/nir: Lower large indirect variables to scratch
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=74470baebbdacc8fd31c9912eb8c00c0cd102903;p=mesa.git

ac/nir: Lower large indirect variables to scratch

results from radeonsi NIR:

Totals from affected shaders:
SGPRS: 704 -> 464 (-34.09 %)
VGPRS: 2056 -> 672 (-67.32 %)
Spilled SGPRs: 24 -> 0 (-100.00 %)
Spilled VGPRs: 28406 -> 0 (-100.00 %)
Private memory VGPRs: 0 -> 3182 (0.00 %)
Scratch size: 1064 -> 3228 (203.38 %) dwords per thread
Code Size: 935260 -> 40180 (-95.70 %) bytes
LDS: 0 -> 0 (0.00 %) blocks
Max Waves: 28 -> 70 (150.00 %)
Wait states: 0 -> 0 (0.00 %)

results from radv:

Totals from affected shaders:
SGPRS: 80 -> 48 (-40.00 %)
VGPRS: 204 -> 108 (-47.06 %)
Spilled SGPRs: 0 -> 0 (0.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 256 (0.00 %) dwords per thread
Code Size: 15792 -> 9504 (-39.82 %) bytes
LDS: 0 -> 0 (0.00 %) blocks
Max Waves: 1 -> 2 (100.00 %)
Wait states: 0 -> 0 (0.00 %)

Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
---

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 53b93f7e1d3..5e25e838f8f 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -42,6 +42,8 @@ struct ac_nir_context {
 
 	LLVMValueRef *ssa_defs;
 
+	LLVMValueRef scratch;
+
 	struct hash_table *defs;
 	struct hash_table *phis;
 	struct hash_table *vars;
@@ -3573,6 +3575,36 @@ static void visit_intrinsic(struct ac_nir_context *ctx,
 	case nir_intrinsic_mbcnt_amd:
 		result = ac_build_mbcnt(&ctx->ac, get_src(ctx, instr->src[0]));
 		break;
+	case nir_intrinsic_load_scratch: {
+		LLVMValueRef offset = get_src(ctx, instr->src[0]);
+		LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch,
+						 offset);
+		LLVMTypeRef comp_type =
+			LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
+		LLVMTypeRef vec_type =
+			instr->dest.ssa.num_components == 1 ? comp_type :
+			LLVMVectorType(comp_type, instr->dest.ssa.num_components);
+		unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
+		ptr = LLVMBuildBitCast(ctx->ac.builder, ptr,
+				       LLVMPointerType(vec_type, addr_space), "");
+		result = LLVMBuildLoad(ctx->ac.builder, ptr, "");
+		break;
+	}
+	case nir_intrinsic_store_scratch: {
+		LLVMValueRef offset = get_src(ctx, instr->src[1]);
+		LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch,
+						 offset);
+		LLVMTypeRef comp_type =
+			LLVMIntTypeInContext(ctx->ac.context, instr->src[0].ssa->bit_size);
+		LLVMTypeRef vec_type =
+			instr->src[0].ssa->num_components == 1 ? comp_type :
+			LLVMVectorType(comp_type, instr->src[0].ssa->num_components);
+		unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
+		ptr = LLVMBuildBitCast(ctx->ac.builder, ptr,
+				       LLVMPointerType(vec_type, addr_space), "");
+		LLVMBuildStore(ctx->ac.builder, get_src(ctx, instr->src[0]), ptr);
+		break;
+	}
 	default:
 		fprintf(stderr, "Unknown intrinsic: ");
 		nir_print_instr(&instr->instr, stderr);
@@ -4473,6 +4505,18 @@ setup_locals(struct ac_nir_context *ctx,
 	}
 }
 
+static void
+setup_scratch(struct ac_nir_context *ctx,
+	      struct nir_shader *shader)
+{
+	if (shader->scratch_size == 0)
+		return;
+
+	ctx->scratch = ac_build_alloca_undef(&ctx->ac,
+					     LLVMArrayType(ctx->ac.i8, shader->scratch_size),
+					     "scratch");
+}
+
 static void
 setup_shared(struct ac_nir_context *ctx,
 	     struct nir_shader *nir)
@@ -4518,6 +4562,7 @@ void ac_nir_translate(struct ac_llvm_context *ac, struct ac_shader_abi *abi,
 	ctx.ssa_defs = calloc(func->impl->ssa_alloc, sizeof(LLVMValueRef));
 
 	setup_locals(&ctx, func);
+	setup_scratch(&ctx, nir);
 
 	if (gl_shader_stage_is_compute(nir->info.stage))
 		setup_shared(&ctx, nir);
@@ -4539,6 +4584,15 @@ void ac_nir_translate(struct ac_llvm_context *ac, struct ac_shader_abi *abi,
 void
 ac_lower_indirect_derefs(struct nir_shader *nir, enum chip_class chip_class)
 {
+	/* Lower large variables to scratch first so that we won't bloat the
+	 * shader by generating large if ladders for them. We later lower
+	 * scratch to alloca's, assuming LLVM won't generate VGPR indexing.
+	 */
+	NIR_PASS_V(nir, nir_lower_vars_to_scratch,
+		   nir_var_function_temp,
+		   256,
+		   glsl_get_natural_size_align_bytes);
+
 	/* While it would be nice not to have this flag, we are constrained
 	 * by the reality that LLVM 9.0 has buggy VGPR indexing on GFX9.
 	 */