ac: initial Wave32 support in LLVM build helpers

Reviewed-by: Samuel Pitoiset <[email protected]>

4 files changed, 21 insertions, 13 deletions

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 041b6cd797e..1551df07959 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -58,7 +58,8 @@ struct ac_llvm_flow {
  */
 void
 ac_llvm_context_init(struct ac_llvm_context *ctx,
-		     enum chip_class chip_class, enum radeon_family family)
+		     enum chip_class chip_class, enum radeon_family family,
+		     unsigned wave_size)
 {
 	LLVMValueRef args[1];
 
@@ -66,6 +67,7 @@ ac_llvm_context_init(struct ac_llvm_context *ctx,
 
 	ctx->chip_class = chip_class;
 	ctx->family = family;
+	ctx->wave_size = wave_size;
 	ctx->module = NULL;
 	ctx->builder = NULL;
 
@@ -2225,10 +2227,14 @@ ac_get_thread_id(struct ac_llvm_context *ctx)
 					 "llvm.amdgcn.mbcnt.lo", ctx->i32,
 					 tid_args, 2, AC_FUNC_ATTR_READNONE);
 
-	tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
-				 ctx->i32, tid_args,
-				 2, AC_FUNC_ATTR_READNONE);
-	set_range_metadata(ctx, tid, 0, 64);
+	if (ctx->wave_size == 32) {
+		tid = tid_args[1];
+	} else {
+		tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
+					 ctx->i32, tid_args,
+					 2, AC_FUNC_ATTR_READNONE);
+	}
+	set_range_metadata(ctx, tid, 0, ctx->wave_size);
 	return tid;
 }
 
@@ -4260,7 +4266,7 @@ ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op
 		get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
 	result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
 				  LLVMTypeOf(identity), "");
-	result = ac_build_scan(ctx, op, result, identity, 64, true);
+	result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, true);
 
 	return ac_build_wwm(ctx, result);
 }
@@ -4284,7 +4290,7 @@ ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op
 		get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
 	result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
 				  LLVMTypeOf(identity), "");
-	result = ac_build_scan(ctx, op, result, identity, 64, false);
+	result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, false);
 
 	return ac_build_wwm(ctx, result);
 }
@@ -4360,12 +4366,12 @@ ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
 	if (ws->maxwaves <= 1)
 		return;
 
-	const LLVMValueRef i32_63 = LLVMConstInt(ctx->i32, 63, false);
+	const LLVMValueRef last_lane = LLVMConstInt(ctx->i32, ctx->wave_size - 1, false);
 	LLVMBuilderRef builder = ctx->builder;
 	LLVMValueRef tid = ac_get_thread_id(ctx);
 	LLVMValueRef tmp;
 
-	tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, i32_63, "");
+	tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, last_lane, "");
 	ac_build_ifcc(ctx, tmp, 1000);
 	LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
 	ac_build_endif(ctx, 1000);
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index 7f6139e5cdd..588ef242c20 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -103,13 +103,15 @@ struct ac_llvm_context {
 
 	enum chip_class chip_class;
 	enum radeon_family family;
+	unsigned wave_size;
 
 	LLVMValueRef lds;
 };
 
 void
 ac_llvm_context_init(struct ac_llvm_context *ctx,
-		     enum chip_class chip_class, enum radeon_family family);
+		     enum chip_class chip_class, enum radeon_family family,
+		     unsigned wave_size);
 
 void
 ac_llvm_context_dispose(struct ac_llvm_context *ctx);
diff --git a/src/amd/vulkan/radv_nir_to_llvm.c b/src/amd/vulkan/radv_nir_to_llvm.c
index f5fb54f2d11..a18718ef081 100644
--- a/src/amd/vulkan/radv_nir_to_llvm.c
+++ b/src/amd/vulkan/radv_nir_to_llvm.c
@@ -4319,7 +4319,7 @@ LLVMModuleRef ac_translate_nir_to_llvm(struct ac_llvm_compiler *ac_llvm,
 	ctx.options = options;
 	ctx.shader_info = shader_info;
 
-	ac_llvm_context_init(&ctx.ac, options->chip_class, options->family);
+	ac_llvm_context_init(&ctx.ac, options->chip_class, options->family, 64);
 	ctx.context = ctx.ac.context;
 	ctx.ac.module = ac_create_module(ac_llvm->tm, ctx.context);
 
@@ -4838,7 +4838,7 @@ radv_compile_gs_copy_shader(struct ac_llvm_compiler *ac_llvm,
 	ctx.options = options;
 	ctx.shader_info = shader_info;
 
-	ac_llvm_context_init(&ctx.ac, options->chip_class, options->family);
+	ac_llvm_context_init(&ctx.ac, options->chip_class, options->family, 64);
 	ctx.context = ctx.ac.context;
 	ctx.ac.module = ac_create_module(ac_llvm->tm, ctx.context);
 
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
index 1690c3cb8a8..9a9f3d63cc6 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
@@ -962,7 +962,7 @@ void si_llvm_context_init(struct si_shader_context *ctx,
 	ctx->screen = sscreen;
 	ctx->compiler = compiler;
 
-	ac_llvm_context_init(&ctx->ac, sscreen->info.chip_class, sscreen->info.family);
+	ac_llvm_context_init(&ctx->ac, sscreen->info.chip_class, sscreen->info.family, 64);
 	ctx->ac.module = ac_create_module(compiler->tm, ctx->ac.context);
 
 	enum ac_float_mode float_mode =