Mesa (master): ac: add ac_build_round

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: 77903c8cfbaa609d132ea5e5fbca062e0b2f0808
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=77903c8cfbaa609d132ea5e5fbca062e0b2f0808

Author: Marek Olšák <marek.olsak at amd.com>
Date:   Fri Sep 21 21:30:09 2018 -0400

ac: add ac_build_round

---

 src/amd/common/ac_llvm_build.c                    | 19 +++++++++++++++++--
 src/amd/common/ac_llvm_build.h                    |  1 +
 src/amd/common/ac_nir_to_llvm.c                   |  2 +-
 src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c |  4 +---
 4 files changed, 20 insertions(+), 6 deletions(-)

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 81158940b8..d11c12fcd1 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -558,6 +558,22 @@ LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 	return ac_build_gather_values(ctx, chan, 4);
 }
 
+LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value)
+{
+	unsigned type_size = ac_get_type_size(LLVMTypeOf(value));
+	const char *name;
+
+	if (type_size == 2)
+		name = "llvm.rint.f16";
+	else if (type_size == 4)
+		name = "llvm.rint.f32";
+	else
+		name = "llvm.rint.f64";
+
+	return ac_build_intrinsic(ctx, name, LLVMTypeOf(value), &value, 1,
+				  AC_FUNC_ATTR_READNONE);
+}
+
 LLVMValueRef
 ac_build_fdiv(struct ac_llvm_context *ctx,
 	      LLVMValueRef num,
@@ -675,8 +691,7 @@ ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 	LLVMValueRef invma;
 
 	if (is_array && !is_lod) {
-		LLVMValueRef tmp = coords_arg[3];
-		tmp = ac_build_intrinsic(ctx, "llvm.rint.f32", ctx->f32, &tmp, 1, 0);
+		LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
 
 		/* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
 		 *
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index 73f8a6b70c..65235fb735 100644
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -175,6 +175,7 @@ ac_build_gather_values(struct ac_llvm_context *ctx,
 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
 				     LLVMValueRef value,
 				     unsigned num_channels);
+LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
 
 LLVMValueRef
 ac_build_fdiv(struct ac_llvm_context *ctx,
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 312383db36..ffc64a79d9 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -3311,7 +3311,7 @@ static LLVMValueRef apply_round_slice(struct ac_llvm_context *ctx,
 				      LLVMValueRef coord)
 {
 	coord = ac_to_float(ctx, coord);
-	coord = ac_build_intrinsic(ctx, "llvm.rint.f32", ctx->f32, &coord, 1, 0);
+	coord = ac_build_round(ctx, coord);
 	coord = ac_to_integer(ctx, coord);
 	return coord;
 }
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
index cabc448a08..8c44831bcc 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_mem.c
@@ -1446,9 +1446,7 @@ static void build_tex_intrinsic(const struct lp_build_tgsi_action *action,
 		   opcode != TGSI_OPCODE_TXF_LZ &&
 		   ctx->screen->info.chip_class <= VI) {
 		unsigned array_coord = target == TGSI_TEXTURE_1D_ARRAY ? 1 : 2;
-		args.coords[array_coord] =
-			ac_build_intrinsic(&ctx->ac, "llvm.rint.f32", ctx->f32,
-					   &args.coords[array_coord], 1, 0);
+		args.coords[array_coord] = ac_build_round(&ctx->ac, args.coords[array_coord]);
 	}
 
 	/* 1D textures are allocated and used as 2D on GFX9. */