diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
index 94ec569ad9f..a1e3ce6d241 100644
--- a/src/amd/common/ac_llvm_build.c
+++ b/src/amd/common/ac_llvm_build.c
@@ -2665,6 +2665,13 @@ LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
 			   LLVMValueRef s1, LLVMValueRef s2)
 {
+	/* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
+	if (ctx->chip_class >= GFX10) {
+		return ac_build_intrinsic(ctx, "llvm.fma.f32", ctx->f32,
+					  (LLVMValueRef []) {s0, s1, s2}, 3,
+					  AC_FUNC_ATTR_READNONE);
+	}
+
 	return LLVMBuildFAdd(ctx->builder,
 			     LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
 }
diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 5ed21b1650b..7294b293ca2 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -811,7 +811,8 @@ static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
 		}
 		break;
 	case nir_op_ffma:
-		result = emit_intrin_3f_param(&ctx->ac, "llvm.fmuladd",
+		/* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
+		result = emit_intrin_3f_param(&ctx->ac, ctx->ac.chip_class >= GFX10 ? "llvm.fma" : "llvm.fmuladd",
 		                              ac_to_float_type(&ctx->ac, def_type), src[0], src[1], src[2]);
 		break;
 	case nir_op_ldexp: