[Mesa-dev] [PATCH 8/9] ac,radeonsi: use ac_build_fmad

[Marek Olšák]

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

---
 src/amd/common/ac_llvm_build.c           |  3 +--
 src/amd/common/ac_nir_to_llvm.c          |  7 ++-----
 src/gallium/drivers/radeonsi/si_shader.c | 17 +++++------------
 3 files changed, 8 insertions(+), 19 deletions(-)

diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 6d5bfb1a1be..c741a1ab62d 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -755,22 +755,21 @@ ac_prepare_cube_coords(struct ac_llvm_context *ctx,
 
 	/* Shift the texture coordinate. This must be applied after the
 	 * derivative calculation.
 	 */
 	for (int i = 0; i < 2; ++i)
 		coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
 
 	if (is_array) {
 		/* for cube arrays coord.z = coord.w(array_index) * 8 + face */
 		/* coords_arg.w component - array_index for cube arrays */
-		LLVMValueRef tmp = LLVMBuildFMul(ctx->builder, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), "");
-		coords[2] = LLVMBuildFAdd(ctx->builder, tmp, coords[2], "");
+		coords[2] = ac_build_fmad(ctx, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), coords[2]);
 	}
 
 	memcpy(coords_arg, coords, sizeof(coords));
 }
 
 
 LLVMValueRef
 ac_build_fs_interp(struct ac_llvm_context *ctx,
 		   LLVMValueRef llvm_chan,
 		   LLVMValueRef attr_number,
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 1584fef7ab7..537ac33c044 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -2846,25 +2846,22 @@ static LLVMValueRef visit_interp(struct ac_nir_context *ctx,
 								      ddxy_out, ix_ll, "");
 			LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->ac.builder,
 								      ddxy_out, iy_ll, "");
 			LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->ac.builder,
 									 interp_param, ix_ll, "");
 			LLVMValueRef temp1, temp2;
 
 			interp_el = LLVMBuildBitCast(ctx->ac.builder, interp_el,
 						     ctx->ac.f32, "");
 
-			temp1 = LLVMBuildFMul(ctx->ac.builder, ddx_el, src_c0, "");
-			temp1 = LLVMBuildFAdd(ctx->ac.builder, temp1, interp_el, "");
-
-			temp2 = LLVMBuildFMul(ctx->ac.builder, ddy_el, src_c1, "");
-			temp2 = LLVMBuildFAdd(ctx->ac.builder, temp2, temp1, "");
+			temp1 = ac_build_fmad(&ctx->ac, ddx_el, src_c0, interp_el);
+			temp2 = ac_build_fmad(&ctx->ac, ddy_el, src_c1, temp1);
 
 			ij_out[i] = LLVMBuildBitCast(ctx->ac.builder,
 						     temp2, ctx->ac.i32, "");
 		}
 		interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2);
 
 	}
 
 	for (chan = 0; chan < 4; chan++) {
 		LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, chan, false);
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 24ee45f578a..66fe5fad218 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -2652,24 +2652,22 @@ static void si_llvm_emit_clipvertex(struct si_shader_context *ctx,
 		args->out[3] = LLVMConstReal(ctx->f32, 0.0f);
 
 		/* Compute dot products of position and user clip plane vectors */
 		for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
 			for (const_chan = 0; const_chan < TGSI_NUM_CHANNELS; const_chan++) {
 				LLVMValueRef addr =
 					LLVMConstInt(ctx->i32, ((reg_index * 4 + chan) * 4 +
 								const_chan) * 4, 0);
 				base_elt = buffer_load_const(ctx, const_resource,
 							     addr);
-				args->out[chan] =
-					LLVMBuildFAdd(ctx->ac.builder, args->out[chan],
-						      LLVMBuildFMul(ctx->ac.builder, base_elt,
-								    out_elts[const_chan], ""), "");
+				args->out[chan] = ac_build_fmad(&ctx->ac, base_elt,
+								out_elts[const_chan], args->out[chan]);
 			}
 		}
 
 		args->enabled_channels = 0xf;
 		args->valid_mask = 0;
 		args->done = 0;
 		args->target = V_008DFC_SQ_EXP_POS + 2 + reg_index;
 		args->compr = 0;
 	}
 }
@@ -4107,31 +4105,26 @@ static void build_interp_intrinsic(const struct lp_build_tgsi_action *action,
 		 */
 		for (i = 0; i < 2; i++) {
 			LLVMValueRef ix_ll = LLVMConstInt(ctx->i32, i, 0);
 			LLVMValueRef iy_ll = LLVMConstInt(ctx->i32, i + 2, 0);
 			LLVMValueRef ddx_el = LLVMBuildExtractElement(ctx->ac.builder,
 								      ddxy_out, ix_ll, "");
 			LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->ac.builder,
 								      ddxy_out, iy_ll, "");
 			LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->ac.builder,
 									 interp_param, ix_ll, "");
-			LLVMValueRef temp1, temp2;
+			LLVMValueRef temp;
 
 			interp_el = ac_to_float(&ctx->ac, interp_el);
 
-			temp1 = LLVMBuildFMul(ctx->ac.builder, ddx_el, offset_x, "");
-
-			temp1 = LLVMBuildFAdd(ctx->ac.builder, temp1, interp_el, "");
-
-			temp2 = LLVMBuildFMul(ctx->ac.builder, ddy_el, offset_y, "");
-
-			ij_out[i] = LLVMBuildFAdd(ctx->ac.builder, temp2, temp1, "");
+			temp = ac_build_fmad(&ctx->ac, ddx_el, offset_x, interp_el);
+			ij_out[i] = ac_build_fmad(&ctx->ac, ddy_el, offset_y, temp);
 		}
 		interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2);
 	}
 
 	if (interp_param)
 		interp_param = ac_to_float(&ctx->ac, interp_param);
 
 	for (chan = 0; chan < 4; chan++) {
 		LLVMValueRef gather = LLVMGetUndef(LLVMVectorType(ctx->f32, input_array_size));
 		unsigned schan = tgsi_util_get_full_src_register_swizzle(&inst->Src[0], chan);
-- 
2.17.1