Used neg ptx7.0 builtin for unary minus

Author: JackAKirk

clang/include/clang/Basic/BuiltinsNVPTX.def

@@ -182,6 +182,11 @@ BUILTIN(__nvvm_fabs_ftz_f, "ff", "")
 BUILTIN(__nvvm_fabs_f, "ff", "")
 BUILTIN(__nvvm_fabs_d, "dd", "")
 
+// Neg
+
+TARGET_BUILTIN(__nvvm_neg_bf16, "ZUsZUs", "", AND(SM_80,PTX70))
+TARGET_BUILTIN(__nvvm_neg_bf16x2, "ZUiZUi", "", AND(SM_80,PTX70))
+
 // Round
 
 BUILTIN(__nvvm_round_ftz_f, "ff", "")

llvm/include/llvm/IR/IntrinsicsNVVM.td

@@ -740,6 +740,19 @@ let TargetPrefix = "nvvm" in {
   def int_nvvm_fabs_d : GCCBuiltin<"__nvvm_fabs_d">,
       DefaultAttrsIntrinsic<[llvm_double_ty], [llvm_double_ty], [IntrNoMem, IntrSpeculatable]>;
 
+//
+// Neg bf16, bf16x2
+//
+
+  foreach unary = ["neg"] in {
+    def int_nvvm_ # unary # _bf16 :
+      GCCBuiltin<!strconcat("__nvvm_", unary, "_bf16")>,
+      DefaultAttrsIntrinsic<[llvm_i16_ty], [llvm_i16_ty], [IntrNoMem]>;
+    def int_nvvm_ # unary # _bf16x2 :
+      GCCBuiltin<!strconcat("__nvvm_", unary, "_bf16x2")>,
+      DefaultAttrsIntrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem]>;
+  }
+
 //
 // Round
 //

llvm/lib/Target/NVPTX/NVPTXIntrinsics.td

@@ -719,6 +719,15 @@ def INT_NVVM_FABS_F : F_MATH_1<"abs.f32 \t$dst, $src0;", Float32Regs,
 def INT_NVVM_FABS_D : F_MATH_1<"abs.f64 \t$dst, $src0;", Float64Regs,
   Float64Regs, int_nvvm_fabs_d>;
 
+//
+// Neg bf16, bf16x2
+//
+
+def INT_NVVM_NEG_BF16 : F_MATH_1<"neg.bf16 \t$dst, $src0;", Int16Regs,
+  Int16Regs, int_nvvm_neg_bf16>;
+def INT_NVVM_NEG_BF16X2 : F_MATH_1<"neg.bf16x2 \t$dst, $src0;", Int32Regs,
+  Int32Regs, int_nvvm_neg_bf16x2>;
+
 //
 // Round
 //

sycl/include/sycl/ext/oneapi/experimental/bfloat16.hpp

@@ -42,7 +42,7 @@ class [[sycl_detail::uses_aspects(ext_intel_bf16_conversion)]] bfloat16 {
   static float to_float(const storage_t &a) {
 #if defined(__SYCL_DEVICE_ONLY__)
 #if defined(__NVPTX__)
-    unsigned int y = a;
+    uint32_t y = a;
     y = y << 16;
     float *res = reinterpret_cast<float *>(&y);
     return *res;
@@ -81,7 +81,16 @@ class [[sycl_detail::uses_aspects(ext_intel_bf16_conversion)]] bfloat16 {
 
   // Unary minus operator overloading
   friend bfloat16 operator-(bfloat16 &lhs) {
-    return bfloat16{-to_float(lhs.value)};
+#if defined(__SYCL_DEVICE_ONLY__)
+#if defined(__NVPTX__)
+    return from_bits(__nvvm_neg_bf16(lhs.value));
+#else
+    return bfloat16{-__spirv_ConvertBF16ToFINTEL(lhs.value)};
+#endif
+#else
+    throw exception{errc::feature_not_supported,
+                    "Bfloat16 unary minus is not supported on host device"};
+#endif
   }
 
 // Increment and decrement operators overloading