[CUDA] Added partial support for CUDA-9.1

[tra]

Summary

Clang can use CUDA-9.1 now, though new builtins (__hmma_m32n8k16*) are not implemented yet.

The major change is that headers in CUDA-9.1 went through substantial
changes that started in CUDA-9.0 which required substantial changes
in the cuda compatibility headers provided by clang.

There are two major issues:

• CUDA SDK no longer provides declarations for libdevice functions.
• A lot of device-side functions have become nvcc's builtins and CUDA headers no longer contain their implementations.

This patch changes the way CUDA headers are handled if we compile
with CUDA 9.x. Both 9.0 and 9.1 are affected.

• Clang provides its own declarations of libdevice functions.
• For CUDA-9.x clang now provides implementation of device-side 'standard library' functions using libdevice.

This patch should not affect compilation with CUDA-8. There may be
some observable differences for CUDA-9.0, though they are not expected
to affect functionality.

Tested: CUDA test-suite tests for all supported combinations of:

CUDA: 7.0,7.5,8.0,9.0,9.1
GPU: sm_20, sm_35, sm_60, sm_70

[jlebar]

Gosh I think the chances we get through this without a copy-paste error are way too high. I'm not sure what to do about it, though. I don't have the focus to check this as carefully as it needs.

Perhaps we need to test this in the test-suite (eventually)? We've seen that even when our own implementation is correct, sometimes nvptx breaks it, and so we need to test e2e *anyway*...

clang/lib/Driver/ToolChains/Cuda.cpp

125

Perhaps update to CUDA 9+

clang/lib/Headers/__clang_cuda_device_functions.h

32

I don't think we should need __forceinline__? I'm sure they had that in their headers, but clang doesn't rely on its inliner for correctness, unlike some *other* compilers. :)

I'm also not sure about static. I think we mark all CUDA code as internal, so it shouldn't be necessary? If they're not static, they should be marked as inline, though.

51

Sanity check: Ignoring the __a argument here is correct?

clang/lib/Headers/__clang_cuda_runtime_wrapper.h

143

a particular (or "our particular", if you like)

144

Sentence doesn't read right if you skip the parens.

But also: Our header provides *everything* from cuda/device_functions_decls.h, not just the things that were once in that header and, in newer CUDA versions, are no longer there, right? I'm actually even more confused now that I read the big header, because here we say the header "defines or declares" these functions, but that suggests that for every function, it decides whether to define or to declare it, but that's not the case...

155

Should we be pushing/popping the macro? That is, do we want this macro exposed to user code?

[tra]

clang/lib/Headers/__clang_cuda_device_functions.h

32

The idea is to make those wrappers always inlined, so they are effectively replaced with a call to __nv_* libdevice function. Whether the callee itself get inlined is controlled by the attributes in the libdevice bitcode.

__attribute__((always_inline)) which __forceinline__ expands to is handled differently from regular inline. We have Transforms/IPO/AlwaysInliner.cpp which handles them with -O1. Higher opt levels will presumably inline these functions, too. Without always_inline these wrappers will remain uninlined at -O1. Replacing it with just inline will work OK (only affects -O1 in a minor way), but __forceinline__ is a bit closer to what I want from these wrappers.

As for the static, unused non-static functions will be emitted by clang -- as far as clang is concerned they are externally visible. I've checked and we do end up emitting them in PTX, even though they don't have .visible directive (I think that's what you referred to when you mentioned 'mark as internal'). I don't think we want that.

I'll keep this combination for now. We can adjust it later, if necessary.

51

Yes. It's a leftover from the old versions and has been deprecated in CUDA-8:

__DEPRECATED__("Please use __brkpt() instead.") void brkpt(int c = 0)

CUDA-9.1 does not have it at all.

clang/lib/Headers/__clang_cuda_runtime_wrapper.h

144

I should rephrase that then.

__clang_cuda_device_functions.h does two jobs. It always includes __clang_cuda_libdevice_declares.h which provides declarations for the libdevice functions. Those are gone in CUDA-9.1.

It also provides definitions for the standard library and __* device-side functions that call libdevice.

I can explicitly include _clang_cuda_libdevice_declares.h here and update the comment to make more sense.

155

We do push __THROW at the beginning of this header and pop it closer to the end, so this redefinition is not visible outside.

[jlebar]

clang/lib/Headers/__clang_cuda_device_functions.h

32

Can you help me understand what makes you want this particular inlining behavior (i.e., always inline, even at -O0) for these wrappers?

Does e.g. libc++ do the same thing for the functions it has that are thin wrappers around e.g. libc functions?

As for the static, unused non-static functions will be emitted by clang -- as far as clang is concerned they are externally visible. I've checked and we do end up emitting them in PTX, even though they don't have .visible directive (I think that's what you referred to when you mentioned 'mark as internal'). I don't think we want that.

Agree. This is kind of unfortunate, though -- it's not How You Write Headers. Could we add a comment?

clang/lib/Headers/__clang_cuda_runtime_wrapper.h

144

I think that would help clarify things, for sure.

[tra]

clang/lib/Headers/__clang_cuda_device_functions.h

32

libc++ does seem to use _LIBCPP_ALWAYS_INLINE for a lot of wrapper-like functions:

E.g. libcxx/include/__locale:

535:    _LIBCPP_ALWAYS_INLINE
536-    const char_type* tolower(char_type* __low, const char_type* __high) const
537-    {
538-        return do_tolower(__low, __high);
539-    }

[jlebar]

clang/lib/Headers/__clang_cuda_device_functions.h

32

sgtm then!

[tra]

Addressed Justin's comments.

[tra]

Rebased to HEAD.

clang/include/clang/Basic/Cuda.h

 namespace clang {
 
 enum class CudaVersion {
   UNKNOWN,
   CUDA_70,
   CUDA_75,
   CUDA_80,
   CUDA_90,
-  LATEST = CUDA_90,
+  CUDA_91,
+  LATEST = CUDA_91,
 };
 const char *CudaVersionToString(CudaVersion V);
 
 // No string -> CudaVersion conversion function because there's no canonical
 // spelling of the various CUDA versions.
 
 enum class CudaArch {
   UNKNOWN,
   SM_20,
   SM_21,
   SM_30,
   SM_32,
   SM_35,
   SM_37,
   SM_50,
   SM_52,
   SM_53,
   SM_60,
   SM_61,
   SM_62,
   SM_70,
+  SM_72,
 };
 const char *CudaArchToString(CudaArch A);
 
 // The input should have the form "sm_20".
 CudaArch StringToCudaArch(llvm::StringRef S);
 
 enum class CudaVirtualArch {
   UNKNOWN,
   COMPUTE_20,
   COMPUTE_30,
   COMPUTE_32,
   COMPUTE_35,
   COMPUTE_37,
   COMPUTE_50,
   COMPUTE_52,
   COMPUTE_53,
   COMPUTE_60,
   COMPUTE_61,
   COMPUTE_62,
   COMPUTE_70,
+  COMPUTE_72,
 };
 const char *CudaVirtualArchToString(CudaVirtualArch A);
 
 // The input should have the form "compute_20".
 CudaVirtualArch StringToCudaVirtualArch(llvm::StringRef S);
 
 /// Get the compute_xx corresponding to an sm_yy.
 CudaVirtualArch VirtualArchForCudaArch(CudaArch A);

clang/lib/Basic/Cuda.cpp

const char *CudaVersionToString(CudaVersion V) {
   case CudaVersion::CUDA_70:
     return "7.0";
   case CudaVersion::CUDA_75:
     return "7.5";
   case CudaVersion::CUDA_80:
     return "8.0";
   case CudaVersion::CUDA_90:
     return "9.0";
+  case CudaVersion::CUDA_91:
+    return "9.1";
   }
   llvm_unreachable("invalid enum");
 }
 
 const char *CudaArchToString(CudaArch A) {
   switch (A) {
   case CudaArch::UNKNOWN:
     return "unknown";
const char *CudaArchToString(CudaArch A) {
   case CudaArch::SM_60:
     return "sm_60";
   case CudaArch::SM_61:
     return "sm_61";
   case CudaArch::SM_62:
     return "sm_62";
   case CudaArch::SM_70:
     return "sm_70";
+  case CudaArch::SM_72:
+    return "sm_72";
   }
   llvm_unreachable("invalid enum");
 }
 
 CudaArch StringToCudaArch(llvm::StringRef S) {
   return llvm::StringSwitch<CudaArch>(S)
       .Case("sm_20", CudaArch::SM_20)
       .Case("sm_21", CudaArch::SM_21)
       .Case("sm_30", CudaArch::SM_30)
       .Case("sm_32", CudaArch::SM_32)
       .Case("sm_35", CudaArch::SM_35)
       .Case("sm_37", CudaArch::SM_37)
       .Case("sm_50", CudaArch::SM_50)
       .Case("sm_52", CudaArch::SM_52)
       .Case("sm_53", CudaArch::SM_53)
       .Case("sm_60", CudaArch::SM_60)
       .Case("sm_61", CudaArch::SM_61)
       .Case("sm_62", CudaArch::SM_62)
       .Case("sm_70", CudaArch::SM_70)
+      .Case("sm_72", CudaArch::SM_72)
       .Default(CudaArch::UNKNOWN);
 }
 
 const char *CudaVirtualArchToString(CudaVirtualArch A) {
   switch (A) {
   case CudaVirtualArch::UNKNOWN:
     return "unknown";
   case CudaVirtualArch::COMPUTE_20:
const char *CudaVirtualArchToString(CudaVirtualArch A) {
   case CudaVirtualArch::COMPUTE_60:
     return "compute_60";
   case CudaVirtualArch::COMPUTE_61:
     return "compute_61";
   case CudaVirtualArch::COMPUTE_62:
     return "compute_62";
   case CudaVirtualArch::COMPUTE_70:
     return "compute_70";
+  case CudaVirtualArch::COMPUTE_72:
+    return "compute_72";
   }
   llvm_unreachable("invalid enum");
 }
 
 CudaVirtualArch StringToCudaVirtualArch(llvm::StringRef S) {
   return llvm::StringSwitch<CudaVirtualArch>(S)
       .Case("compute_20", CudaVirtualArch::COMPUTE_20)
       .Case("compute_30", CudaVirtualArch::COMPUTE_30)
       .Case("compute_32", CudaVirtualArch::COMPUTE_32)
       .Case("compute_35", CudaVirtualArch::COMPUTE_35)
       .Case("compute_37", CudaVirtualArch::COMPUTE_37)
       .Case("compute_50", CudaVirtualArch::COMPUTE_50)
       .Case("compute_52", CudaVirtualArch::COMPUTE_52)
       .Case("compute_53", CudaVirtualArch::COMPUTE_53)
       .Case("compute_60", CudaVirtualArch::COMPUTE_60)
       .Case("compute_61", CudaVirtualArch::COMPUTE_61)
       .Case("compute_62", CudaVirtualArch::COMPUTE_62)
       .Case("compute_70", CudaVirtualArch::COMPUTE_70)
+      .Case("compute_72", CudaVirtualArch::COMPUTE_72)
       .Default(CudaVirtualArch::UNKNOWN);
 }
 
 CudaVirtualArch VirtualArchForCudaArch(CudaArch A) {
   switch (A) {
   case CudaArch::UNKNOWN:
     return CudaVirtualArch::UNKNOWN;
   case CudaArch::SM_20:
CudaVirtualArch VirtualArchForCudaArch(CudaArch A) {
   case CudaArch::SM_60:
     return CudaVirtualArch::COMPUTE_60;
   case CudaArch::SM_61:
     return CudaVirtualArch::COMPUTE_61;
   case CudaArch::SM_62:
     return CudaVirtualArch::COMPUTE_62;
   case CudaArch::SM_70:
     return CudaVirtualArch::COMPUTE_70;
+  case CudaArch::SM_72:
+    return CudaVirtualArch::COMPUTE_72;
   }
   llvm_unreachable("invalid enum");
 }
 
 CudaVersion MinVersionForCudaArch(CudaArch A) {
   switch (A) {
   case CudaArch::UNKNOWN:
     return CudaVersion::UNKNOWN;
   case CudaArch::SM_20:
   case CudaArch::SM_21:
   case CudaArch::SM_30:
   case CudaArch::SM_32:
   case CudaArch::SM_35:
   case CudaArch::SM_37:
   case CudaArch::SM_50:
   case CudaArch::SM_52:
   case CudaArch::SM_53:
     return CudaVersion::CUDA_70;
   case CudaArch::SM_60:
   case CudaArch::SM_61:
   case CudaArch::SM_62:
     return CudaVersion::CUDA_80;
   case CudaArch::SM_70:
     return CudaVersion::CUDA_90;
+  case CudaArch::SM_72:
+    return CudaVersion::CUDA_91;
   }
   llvm_unreachable("invalid enum");
 }
 
 CudaVersion MaxVersionForCudaArch(CudaArch A) {
   switch (A) {
   case CudaArch::UNKNOWN:
     return CudaVersion::UNKNOWN;

clang/lib/Basic/Targets/NVPTX.cpp

std::string CUDAArchCode = [this] {
       case CudaArch::SM_60:
         return "600";
       case CudaArch::SM_61:
         return "610";
       case CudaArch::SM_62:
         return "620";
       case CudaArch::SM_70:
         return "700";
+      case CudaArch::SM_72:
+        return "720";
       }
       llvm_unreachable("unhandled CudaArch");
     }();
     Builder.defineMacro("__CUDA_ARCH__", CUDAArchCode);
   }
 }
 
 ArrayRef<Builtin::Info> NVPTXTargetInfo::getTargetBuiltins() const {
   return llvm::makeArrayRef(BuiltinInfo, clang::NVPTX::LastTSBuiltin -
                                              Builtin::FirstTSBuiltin);
 }

clang/lib/Driver/ToolChains/Cuda.cpp

if (Major == 7 && Minor == 0) {
     return CudaVersion::CUDA_70;
   }
   if (Major == 7 && Minor == 5)
     return CudaVersion::CUDA_75;
   if (Major == 8 && Minor == 0)
     return CudaVersion::CUDA_80;
   if (Major == 9 && Minor == 0)
     return CudaVersion::CUDA_90;
+  if (Major == 9 && Minor == 1)
+    return CudaVersion::CUDA_91;
   return CudaVersion::UNKNOWN;
 }
 
 CudaInstallationDetector::CudaInstallationDetector(
     const Driver &D, const llvm::Triple &HostTriple,
     const llvm::opt::ArgList &Args)
     : D(D) {
   SmallVector<std::string, 4> CudaPathCandidates;
for (const auto &CudaPath : CudaPathCandidates) {
     if (!VersionFile) {
       // CUDA 7.0 doesn't have a version.txt, so guess that's our version if
       // version.txt isn't present.
       Version = CudaVersion::CUDA_70;
     } else {
       Version = ParseCudaVersionFile((*VersionFile)->getBuffer());
     }
 
-    if (Version == CudaVersion::CUDA_90) {
+    if (Version >= CudaVersion::CUDA_90) {
       // CUDA-9 uses single libdevice file for all GPU variants.
       std::string FilePath = LibDevicePath + "/libdevice.10.bc";
       if (FS.exists(FilePath)) {
-        for (const char *GpuArch :
+        for (const char *GpuArchName :
              {"sm_20", "sm_30", "sm_32", "sm_35", "sm_50", "sm_52", "sm_53",
-              "sm_60", "sm_61", "sm_62", "sm_70"})
-          LibDeviceMap[GpuArch] = FilePath;
+                   "sm_60", "sm_61", "sm_62", "sm_70", "sm_72"}) {
+          const CudaArch GpuArch = StringToCudaArch(GpuArchName);
+          if (Version >= MinVersionForCudaArch(GpuArch) &&
+              Version <= MaxVersionForCudaArch(GpuArch))
+            LibDeviceMap[GpuArchName] = FilePath;
+        }
       }
     } else {
       std::error_code EC;
       for (llvm::sys::fs::directory_iterator LI(LibDevicePath, EC), LE;
            !EC && LI != LE; LI = LI.increment(EC)) {
         StringRef FilePath = LI->path();
         StringRef FileName = llvm::sys::path::filename(FilePath);
         // Process all bitcode filenames that look like

clang/lib/Headers/CMakeLists.txt

set(files
   avx512vnniintrin.h
   avx512vlvnniintrin.h
   avxintrin.h
   bmi2intrin.h
   bmiintrin.h
   __clang_cuda_builtin_vars.h
   __clang_cuda_cmath.h
   __clang_cuda_complex_builtins.h
+  __clang_cuda_device_functions.h
   __clang_cuda_intrinsics.h
+  __clang_cuda_libdevice_declares.h
   __clang_cuda_math_forward_declares.h
   __clang_cuda_runtime_wrapper.h
   cetintrin.h
   clzerointrin.h
   cpuid.h
   clflushoptintrin.h
   clwbintrin.h
   emmintrin.h

clang/lib/Headers/__clang_cuda_device_functions.h

+/*===---- __clang_cuda_device_functions.h - CUDA runtime support -----------===
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ *
+ *===-----------------------------------------------------------------------===
+ */
+
+#ifndef __CLANG_CUDA_DEVICE_FUNCTIONS_H__
+#define __CLANG_CUDA_DEVICE_FUNCTIONS_H__
+
+#include <__clang_cuda_libdevice_declares.h>
+
+#pragma push_macro("__DEVICE__")
+#pragma push_macro("__FAST_OR_SLOW")
+
+#define __DEVICE__ static __device__ __forceinline__
+// There are number of functions that became compiler builtins in CUDA-9 and are
+// no longer implemented in device_functions.hpp.
+// Standard math library functions are also NVCC builtins now.
+#if defined(__CLANG_CUDA_APPROX_TRANSCENDENTALS__)
+#define __FAST_OR_SLOW(fast, slow) fast
+#else
+#define __FAST_OR_SLOW(fast, slow) slow
+#endif
+
+#if CUDA_VERSION >= 9000
+__DEVICE__ int __all(int __a) { return __nvvm_vote_all(__a); }
+__DEVICE__ int __any(int __a) { return __nvvm_vote_any(__a); }
+__DEVICE__ unsigned int __ballot(int __a) { return __nvvm_vote_ballot(__a); }
+__DEVICE__ unsigned int __brev(unsigned int __a) { return __nv_brev(__a); }
+__DEVICE__ unsigned long long __brevll(unsigned long long __a) {
+  return __nv_brevll(__a);
+}
+__DEVICE__ void __brkpt() { asm volatile("brkpt;"); }
+__DEVICE__ void __brkpt(int __a) { __brkpt(); }
+__DEVICE__ unsigned int __byte_perm(unsigned int __a, unsigned int __b,
+                                    unsigned int __c) {
+  return __nv_byte_perm(__a, __b, __c);
+}
+__DEVICE__ int __clz(int __a) { return __nv_clz(__a); }
+__DEVICE__ int __clzll(long long __a) { return __nv_clzll(__a); }
+__DEVICE__ float __cosf(float __a) { return __nv_fast_cosf(__a); }
+__DEVICE__ double __dAtomicAdd(double *__p, double __v) {
+  return __nvvm_atom_add_gen_d(__p, __v);
+}
+__DEVICE__ double __dAtomicAdd_block(double *__p, double __v) {
+  return __nvvm_atom_cta_add_gen_d(__p, __v);
+}
+__DEVICE__ double __dAtomicAdd_system(double *__p, double __v) {
+  return __nvvm_atom_sys_add_gen_d(__p, __v);
+}
+__DEVICE__ double __dadd_rd(double __a, double __b) {
+  return __nv_dadd_rd(__a, __b);
+}
+__DEVICE__ double __dadd_rn(double __a, double __b) {
+  return __nv_dadd_rn(__a, __b);
+}
+__DEVICE__ double __dadd_ru(double __a, double __b) {
+  return __nv_dadd_ru(__a, __b);
+}
+__DEVICE__ double __dadd_rz(double __a, double __b) {
+  return __nv_dadd_rz(__a, __b);
+}
+__DEVICE__ double __ddiv_rd(double __a, double __b) {
+  return __nv_ddiv_rd(__a, __b);
+}
+__DEVICE__ double __ddiv_rn(double __a, double __b) {
+  return __nv_ddiv_rn(__a, __b);
+}
+__DEVICE__ double __ddiv_ru(double __a, double __b) {
+  return __nv_ddiv_ru(__a, __b);
+}
+__DEVICE__ double __ddiv_rz(double __a, double __b) {
+  return __nv_ddiv_rz(__a, __b);
+}
+__DEVICE__ double __dmul_rd(double __a, double __b) {
+  return __nv_dmul_rd(__a, __b);
+}
+__DEVICE__ double __dmul_rn(double __a, double __b) {
+  return __nv_dmul_rn(__a, __b);
+}
+__DEVICE__ double __dmul_ru(double __a, double __b) {
+  return __nv_dmul_ru(__a, __b);
+}
+__DEVICE__ double __dmul_rz(double __a, double __b) {
+  return __nv_dmul_rz(__a, __b);
+}
+__DEVICE__ float __double2float_rd(double __a) {
+  return __nv_double2float_rd(__a);
+}
+__DEVICE__ float __double2float_rn(double __a) {
+  return __nv_double2float_rn(__a);
+}
+__DEVICE__ float __double2float_ru(double __a) {
+  return __nv_double2float_ru(__a);
+}
+__DEVICE__ float __double2float_rz(double __a) {
+  return __nv_double2float_rz(__a);
+}
+__DEVICE__ int __double2hiint(double __a) { return __nv_double2hiint(__a); }
+__DEVICE__ int __double2int_rd(double __a) { return __nv_double2int_rd(__a); }
+__DEVICE__ int __double2int_rn(double __a) { return __nv_double2int_rn(__a); }
+__DEVICE__ int __double2int_ru(double __a) { return __nv_double2int_ru(__a); }
+__DEVICE__ int __double2int_rz(double __a) { return __nv_double2int_rz(__a); }
+__DEVICE__ long long __double2ll_rd(double __a) {
+  return __nv_double2ll_rd(__a);
+}
+__DEVICE__ long long __double2ll_rn(double __a) {
+  return __nv_double2ll_rn(__a);
+}
+__DEVICE__ long long __double2ll_ru(double __a) {
+  return __nv_double2ll_ru(__a);
+}
+__DEVICE__ long long __double2ll_rz(double __a) {
+  return __nv_double2ll_rz(__a);
+}
+__DEVICE__ int __double2loint(double __a) { return __nv_double2loint(__a); }
+__DEVICE__ unsigned int __double2uint_rd(double __a) {
+  return __nv_double2uint_rd(__a);
+}
+__DEVICE__ unsigned int __double2uint_rn(double __a) {
+  return __nv_double2uint_rn(__a);
+}
+__DEVICE__ unsigned int __double2uint_ru(double __a) {
+  return __nv_double2uint_ru(__a);
+}
+__DEVICE__ unsigned int __double2uint_rz(double __a) {
+  return __nv_double2uint_rz(__a);
+}
+__DEVICE__ unsigned long long __double2ull_rd(double __a) {
+  return __nv_double2ull_rd(__a);
+}
+__DEVICE__ unsigned long long __double2ull_rn(double __a) {
+  return __nv_double2ull_rn(__a);
+}
+__DEVICE__ unsigned long long __double2ull_ru(double __a) {
+  return __nv_double2ull_ru(__a);
+}
+__DEVICE__ unsigned long long __double2ull_rz(double __a) {
+  return __nv_double2ull_rz(__a);
+}
+__DEVICE__ long long __double_as_longlong(double __a) {
+  return __nv_double_as_longlong(__a);
+}
+__DEVICE__ double __drcp_rd(double __a) { return __nv_drcp_rd(__a); }
+__DEVICE__ double __drcp_rn(double __a) { return __nv_drcp_rn(__a); }
+__DEVICE__ double __drcp_ru(double __a) { return __nv_drcp_ru(__a); }
+__DEVICE__ double __drcp_rz(double __a) { return __nv_drcp_rz(__a); }
+__DEVICE__ double __dsqrt_rd(double __a) { return __nv_dsqrt_rd(__a); }
+__DEVICE__ double __dsqrt_rn(double __a) { return __nv_dsqrt_rn(__a); }
+__DEVICE__ double __dsqrt_ru(double __a) { return __nv_dsqrt_ru(__a); }
+__DEVICE__ double __dsqrt_rz(double __a) { return __nv_dsqrt_rz(__a); }
+__DEVICE__ double __dsub_rd(double __a, double __b) {
+  return __nv_dsub_rd(__a, __b);
+}
+__DEVICE__ double __dsub_rn(double __a, double __b) {
+  return __nv_dsub_rn(__a, __b);
+}
+__DEVICE__ double __dsub_ru(double __a, double __b) {
+  return __nv_dsub_ru(__a, __b);
+}
+__DEVICE__ double __dsub_rz(double __a, double __b) {
+  return __nv_dsub_rz(__a, __b);
+}
+__DEVICE__ float __exp10f(float __a) { return __nv_fast_exp10f(__a); }
+__DEVICE__ float __expf(float __a) { return __nv_fast_expf(__a); }
+__DEVICE__ float __fAtomicAdd(float *__p, float __v) {
+  return __nvvm_atom_add_gen_f(__p, __v);
+}
+__DEVICE__ float __fAtomicAdd_block(float *__p, float __v) {
+  return __nvvm_atom_cta_add_gen_f(__p, __v);
+}
+__DEVICE__ float __fAtomicAdd_system(float *__p, float __v) {
+  return __nvvm_atom_sys_add_gen_f(__p, __v);
+}
+__DEVICE__ float __fAtomicExch(float *__p, float __v) {
+  return __nv_int_as_float(
+      __nvvm_atom_xchg_gen_i((int *)__p, __nv_float_as_int(__v)));
+}
+__DEVICE__ float __fAtomicExch_block(float *__p, float __v) {
+  return __nv_int_as_float(
+      __nvvm_atom_cta_xchg_gen_i((int *)__p, __nv_float_as_int(__v)));
+}
+__DEVICE__ float __fAtomicExch_system(float *__p, float __v) {
+  return __nv_int_as_float(
+      __nvvm_atom_sys_xchg_gen_i((int *)__p, __nv_float_as_int(__v)));
+}
+__DEVICE__ float __fadd_rd(float __a, float __b) {
+  return __nv_fadd_rd(__a, __b);
+}
+__DEVICE__ float __fadd_rn(float __a, float __b) {
+  return __nv_fadd_rn(__a, __b);
+}
+__DEVICE__ float __fadd_ru(float __a, float __b) {
+  return __nv_fadd_ru(__a, __b);
+}
+__DEVICE__ float __fadd_rz(float __a, float __b) {
+  return __nv_fadd_rz(__a, __b);
+}
+__DEVICE__ float __fdiv_rd(float __a, float __b) {
+  return __nv_fdiv_rd(__a, __b);
+}
+__DEVICE__ float __fdiv_rn(float __a, float __b) {
+  return __nv_fdiv_rn(__a, __b);
+}
+__DEVICE__ float __fdiv_ru(float __a, float __b) {
+  return __nv_fdiv_ru(__a, __b);
+}
+__DEVICE__ float __fdiv_rz(float __a, float __b) {
+  return __nv_fdiv_rz(__a, __b);
+}
+__DEVICE__ float __fdividef(float __a, float __b) {
+  return __nv_fast_fdividef(__a, __b);
+}
+__DEVICE__ int __ffs(int __a) { return __nv_ffs(__a); }
+__DEVICE__ int __ffsll(long long __a) { return __nv_ffsll(__a); }
+__DEVICE__ int __finite(double __a) { return __nv_isfinited(__a); }
+__DEVICE__ int __finitef(float __a) { return __nv_finitef(__a); }
+__DEVICE__ int __float2int_rd(float __a) { return __nv_float2int_rd(__a); }
+__DEVICE__ int __float2int_rn(float __a) { return __nv_float2int_rn(__a); }
+__DEVICE__ int __float2int_ru(float __a) { return __nv_float2int_ru(__a); }
+__DEVICE__ int __float2int_rz(float __a) { return __nv_float2int_rz(__a); }
+__DEVICE__ long long __float2ll_rd(float __a) { return __nv_float2ll_rd(__a); }
+__DEVICE__ long long __float2ll_rn(float __a) { return __nv_float2ll_rn(__a); }
+__DEVICE__ long long __float2ll_ru(float __a) { return __nv_float2ll_ru(__a); }
+__DEVICE__ long long __float2ll_rz(float __a) { return __nv_float2ll_rz(__a); }
+__DEVICE__ unsigned int __float2uint_rd(float __a) {
+  return __nv_float2uint_rd(__a);
+}
+__DEVICE__ unsigned int __float2uint_rn(float __a) {
+  return __nv_float2uint_rn(__a);
+}
+__DEVICE__ unsigned int __float2uint_ru(float __a) {
+  return __nv_float2uint_ru(__a);
+}
+__DEVICE__ unsigned int __float2uint_rz(float __a) {
+  return __nv_float2uint_rz(__a);
+}
+__DEVICE__ unsigned long long __float2ull_rd(float __a) {
+  return __nv_float2ull_rd(__a);
+}
+__DEVICE__ unsigned long long __float2ull_rn(float __a) {
+  return __nv_float2ull_rn(__a);
+}
+__DEVICE__ unsigned long long __float2ull_ru(float __a) {
+  return __nv_float2ull_ru(__a);
+}
+__DEVICE__ unsigned long long __float2ull_rz(float __a) {
+  return __nv_float2ull_rz(__a);
+}
+__DEVICE__ int __float_as_int(float __a) { return __nv_float_as_int(__a); }
+__DEVICE__ unsigned int __float_as_uint(float __a) {
+  return __nv_float_as_uint(__a);
+}
+__DEVICE__ double __fma_rd(double __a, double __b, double __c) {
+  return __nv_fma_rd(__a, __b, __c);
+}
+__DEVICE__ double __fma_rn(double __a, double __b, double __c) {
+  return __nv_fma_rn(__a, __b, __c);
+}
+__DEVICE__ double __fma_ru(double __a, double __b, double __c) {
+  return __nv_fma_ru(__a, __b, __c);
+}
+__DEVICE__ double __fma_rz(double __a, double __b, double __c) {
+  return __nv_fma_rz(__a, __b, __c);
+}
+__DEVICE__ float __fmaf_ieee_rd(float __a, float __b, float __c) {
+  return __nv_fmaf_ieee_rd(__a, __b, __c);
+}
+__DEVICE__ float __fmaf_ieee_rn(float __a, float __b, float __c) {
+  return __nv_fmaf_ieee_rn(__a, __b, __c);
+}
+__DEVICE__ float __fmaf_ieee_ru(float __a, float __b, float __c) {
+  return __nv_fmaf_ieee_ru(__a, __b, __c);
+}
+__DEVICE__ float __fmaf_ieee_rz(float __a, float __b, float __c) {
+  return __nv_fmaf_ieee_rz(__a, __b, __c);
+}
+__DEVICE__ float __fmaf_rd(float __a, float __b, float __c) {
+  return __nv_fmaf_rd(__a, __b, __c);
+}
+__DEVICE__ float __fmaf_rn(float __a, float __b, float __c) {
+  return __nv_fmaf_rn(__a, __b, __c);
+}
+__DEVICE__ float __fmaf_ru(float __a, float __b, float __c) {
+  return __nv_fmaf_ru(__a, __b, __c);
+}
+__DEVICE__ float __fmaf_rz(float __a, float __b, float __c) {
+  return __nv_fmaf_rz(__a, __b, __c);
+}
+__DEVICE__ float __fmul_rd(float __a, float __b) {
+  return __nv_fmul_rd(__a, __b);
+}
+__DEVICE__ float __fmul_rn(float __a, float __b) {
+  return __nv_fmul_rn(__a, __b);
+}
+__DEVICE__ float __fmul_ru(float __a, float __b) {
+  return __nv_fmul_ru(__a, __b);
+}
+__DEVICE__ float __fmul_rz(float __a, float __b) {
+  return __nv_fmul_rz(__a, __b);
+}
+__DEVICE__ float __frcp_rd(float __a) { return __nv_frcp_rd(__a); }
+__DEVICE__ float __frcp_rn(float __a) { return __nv_frcp_rn(__a); }
+__DEVICE__ float __frcp_ru(float __a) { return __nv_frcp_ru(__a); }
+__DEVICE__ float __frcp_rz(float __a) { return __nv_frcp_rz(__a); }
+__DEVICE__ float __frsqrt_rn(float __a) { return __nv_frsqrt_rn(__a); }
+__DEVICE__ float __fsqrt_rd(float __a) { return __nv_fsqrt_rd(__a); }
+__DEVICE__ float __fsqrt_rn(float __a) { return __nv_fsqrt_rn(__a); }
+__DEVICE__ float __fsqrt_ru(float __a) { return __nv_fsqrt_ru(__a); }
+__DEVICE__ float __fsqrt_rz(float __a) { return __nv_fsqrt_rz(__a); }
+__DEVICE__ float __fsub_rd(float __a, float __b) {
+  return __nv_fsub_rd(__a, __b);
+}
+__DEVICE__ float __fsub_rn(float __a, float __b) {
+  return __nv_fsub_rn(__a, __b);
+}
+__DEVICE__ float __fsub_ru(float __a, float __b) {
+  return __nv_fsub_ru(__a, __b);
+}
+__DEVICE__ float __fsub_rz(float __a, float __b) {
+  return __nv_fsub_rz(__a, __b);
+}
+__DEVICE__ int __hadd(int __a, int __b) { return __nv_hadd(__a, __b); }
+__DEVICE__ double __hiloint2double(int __a, int __b) {
+  return __nv_hiloint2double(__a, __b);
+}
+__DEVICE__ int __iAtomicAdd(int *__p, int __v) {
+  return __nvvm_atom_add_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicAdd_block(int *__p, int __v) {
+  __nvvm_atom_cta_add_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicAdd_system(int *__p, int __v) {
+  __nvvm_atom_sys_add_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicAnd(int *__p, int __v) {
+  return __nvvm_atom_and_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicAnd_block(int *__p, int __v) {
+  return __nvvm_atom_cta_and_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicAnd_system(int *__p, int __v) {
+  return __nvvm_atom_sys_and_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicCAS(int *__p, int __cmp, int __v) {
+  return __nvvm_atom_cas_gen_i(__p, __cmp, __v);
+}
+__DEVICE__ int __iAtomicCAS_block(int *__p, int __cmp, int __v) {
+  return __nvvm_atom_cta_cas_gen_i(__p, __cmp, __v);
+}
+__DEVICE__ int __iAtomicCAS_system(int *__p, int __cmp, int __v) {
+  return __nvvm_atom_sys_cas_gen_i(__p, __cmp, __v);
+}
+__DEVICE__ int __iAtomicExch(int *__p, int __v) {
+  return __nvvm_atom_xchg_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicExch_block(int *__p, int __v) {
+  return __nvvm_atom_cta_xchg_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicExch_system(int *__p, int __v) {
+  return __nvvm_atom_sys_xchg_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicMax(int *__p, int __v) {
+  return __nvvm_atom_max_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicMax_block(int *__p, int __v) {
+  return __nvvm_atom_cta_max_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicMax_system(int *__p, int __v) {
+  return __nvvm_atom_sys_max_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicMin(int *__p, int __v) {
+  return __nvvm_atom_min_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicMin_block(int *__p, int __v) {
+  return __nvvm_atom_cta_min_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicMin_system(int *__p, int __v) {
+  return __nvvm_atom_sys_min_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicOr(int *__p, int __v) {
+  return __nvvm_atom_or_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicOr_block(int *__p, int __v) {
+  return __nvvm_atom_cta_or_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicOr_system(int *__p, int __v) {
+  return __nvvm_atom_sys_or_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicXor(int *__p, int __v) {
+  return __nvvm_atom_xor_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicXor_block(int *__p, int __v) {
+  return __nvvm_atom_cta_xor_gen_i(__p, __v);
+}
+__DEVICE__ int __iAtomicXor_system(int *__p, int __v) {
+  return __nvvm_atom_sys_xor_gen_i(__p, __v);
+}
+__DEVICE__ long long __illAtomicMax(long long *__p, long long __v) {
+  return __nvvm_atom_max_gen_ll(__p, __v);
+}
+__DEVICE__ long long __illAtomicMax_block(long long *__p, long long __v) {
+  return __nvvm_atom_cta_max_gen_ll(__p, __v);
+}
+__DEVICE__ long long __illAtomicMax_system(long long *__p, long long __v) {
+  return __nvvm_atom_sys_max_gen_ll(__p, __v);
+}
+__DEVICE__ long long __illAtomicMin(long long *__p, long long __v) {
+  return __nvvm_atom_min_gen_ll(__p, __v);
+}
+__DEVICE__ long long __illAtomicMin_block(long long *__p, long long __v) {
+  return __nvvm_atom_cta_min_gen_ll(__p, __v);
+}
+__DEVICE__ long long __illAtomicMin_system(long long *__p, long long __v) {
+  return __nvvm_atom_sys_min_gen_ll(__p, __v);
+}
+__DEVICE__ double __int2double_rn(int __a) { return __nv_int2double_rn(__a); }
+__DEVICE__ float __int2float_rd(int __a) { return __nv_int2float_rd(__a); }
+__DEVICE__ float __int2float_rn(int __a) { return __nv_int2float_rn(__a); }
+__DEVICE__ float __int2float_ru(int __a) { return __nv_int2float_ru(__a); }
+__DEVICE__ float __int2float_rz(int __a) { return __nv_int2float_rz(__a); }
+__DEVICE__ float __int_as_float(int __a) { return __nv_int_as_float(__a); }
+__DEVICE__ int __isfinited(double __a) { return __nv_isfinited(__a); }
+__DEVICE__ int __isinf(double __a) { return __nv_isinfd(__a); }
+__DEVICE__ int __isinff(float __a) { return __nv_isinff(__a); }
+__DEVICE__ int __isnan(double __a) { return __nv_isnand(__a); }
+__DEVICE__ int __isnanf(float __a) { return __nv_isnanf(__a); }
+__DEVICE__ double __ll2double_rd(long long __a) {
+  return __nv_ll2double_rd(__a);
+}
+__DEVICE__ double __ll2double_rn(long long __a) {
+  return __nv_ll2double_rn(__a);
+}
+__DEVICE__ double __ll2double_ru(long long __a) {
+  return __nv_ll2double_ru(__a);
+}
+__DEVICE__ double __ll2double_rz(long long __a) {
+  return __nv_ll2double_rz(__a);
+}
+__DEVICE__ float __ll2float_rd(long long __a) { return __nv_ll2float_rd(__a); }
+__DEVICE__ float __ll2float_rn(long long __a) { return __nv_ll2float_rn(__a); }
+__DEVICE__ float __ll2float_ru(long long __a) { return __nv_ll2float_ru(__a); }
+__DEVICE__ float __ll2float_rz(long long __a) { return __nv_ll2float_rz(__a); }
+__DEVICE__ long long __llAtomicAnd(long long *__p, long long __v) {
+  return __nvvm_atom_and_gen_ll(__p, __v);
+}
+__DEVICE__ long long __llAtomicAnd_block(long long *__p, long long __v) {
+  return __nvvm_atom_cta_and_gen_ll(__p, __v);
+}
+__DEVICE__ long long __llAtomicAnd_system(long long *__p, long long __v) {
+  return __nvvm_atom_sys_and_gen_ll(__p, __v);
+}
+__DEVICE__ long long __llAtomicOr(long long *__p, long long __v) {
+  return __nvvm_atom_or_gen_ll(__p, __v);
+}
+__DEVICE__ long long __llAtomicOr_block(long long *__p, long long __v) {
+  return __nvvm_atom_cta_or_gen_ll(__p, __v);
+}
+__DEVICE__ long long __llAtomicOr_system(long long *__p, long long __v) {
+  return __nvvm_atom_sys_or_gen_ll(__p, __v);
+}
+__DEVICE__ long long __llAtomicXor(long long *__p, long long __v) {
+  return __nvvm_atom_xor_gen_ll(__p, __v);
+}
+__DEVICE__ long long __llAtomicXor_block(long long *__p, long long __v) {
+  return __nvvm_atom_cta_xor_gen_ll(__p, __v);
+}
+__DEVICE__ long long __llAtomicXor_system(long long *__p, long long __v) {
+  return __nvvm_atom_sys_xor_gen_ll(__p, __v);
+}
+__DEVICE__ float __log10f(float __a) { return __nv_fast_log10f(__a); }
+__DEVICE__ float __log2f(float __a) { return __nv_fast_log2f(__a); }
+__DEVICE__ float __logf(float __a) { return __nv_fast_logf(__a); }
+__DEVICE__ double __longlong_as_double(long long __a) {
+  return __nv_longlong_as_double(__a);
+}
+__DEVICE__ int __mul24(int __a, int __b) { return __nv_mul24(__a, __b); }
+__DEVICE__ long long __mul64hi(long long __a, long long __b) {
+  return __nv_mul64hi(__a, __b);
+}
+__DEVICE__ int __mulhi(int __a, int __b) { return __nv_mulhi(__a, __b); }
+__DEVICE__ unsigned int __pm0(void) { return __nvvm_read_ptx_sreg_pm0(); }
+__DEVICE__ unsigned int __pm1(void) { return __nvvm_read_ptx_sreg_pm1(); }
+__DEVICE__ unsigned int __pm2(void) { return __nvvm_read_ptx_sreg_pm2(); }
+__DEVICE__ unsigned int __pm3(void) { return __nvvm_read_ptx_sreg_pm3(); }
+__DEVICE__ int __popc(int __a) { return __nv_popc(__a); }
+__DEVICE__ int __popcll(long long __a) { return __nv_popcll(__a); }
+__DEVICE__ float __powf(float __a, float __b) {
+  return __nv_fast_powf(__a, __b);
+}
+
+// Parameter must have a known integer value.
+#define __prof_trigger(__a) asm __volatile__("pmevent \t%0;" ::"i"(__a))
+__DEVICE__ int __rhadd(int __a, int __b) { return __nv_rhadd(__a, __b); }
+__DEVICE__ unsigned int __sad(int __a, int __b, unsigned int __c) {
+  return __nv_sad(__a, __b, __c);
+}
+__DEVICE__ float __saturatef(float __a) { return __nv_saturatef(__a); }
+__DEVICE__ int __signbitd(double __a) { return __nv_signbitd(__a); }
+__DEVICE__ int __signbitf(float __a) { return __nv_signbitf(__a); }
+__DEVICE__ void __sincosf(float __a, float *__sptr, float *__cptr) {
+  return __nv_fast_sincosf(__a, __sptr, __cptr);
+}
+__DEVICE__ float __sinf(float __a) { return __nv_fast_sinf(__a); }
+__DEVICE__ int __syncthreads_and(int __a) { return __nvvm_bar0_and(__a); }
+__DEVICE__ int __syncthreads_count(int __a) { return __nvvm_bar0_popc(__a); }
+__DEVICE__ int __syncthreads_or(int __a) { return __nvvm_bar0_or(__a); }
+__DEVICE__ float __tanf(float __a) { return __nv_fast_tanf(__a); }
+__DEVICE__ void __threadfence(void) { __nvvm_membar_gl(); }
+__DEVICE__ void __threadfence_block(void) { __nvvm_membar_cta(); };
+__DEVICE__ void __trap(void) { asm volatile("trap;"); }
+__DEVICE__ unsigned int __uAtomicAdd(unsigned int *__p, unsigned int __v) {
+  return __nvvm_atom_add_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicAdd_block(unsigned int *__p,
+                                           unsigned int __v) {
+  return __nvvm_atom_cta_add_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicAdd_system(unsigned int *__p,
+                                            unsigned int __v) {
+  return __nvvm_atom_sys_add_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicAnd(unsigned int *__p, unsigned int __v) {
+  return __nvvm_atom_and_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicAnd_block(unsigned int *__p,
+                                           unsigned int __v) {
+  return __nvvm_atom_cta_and_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicAnd_system(unsigned int *__p,
+                                            unsigned int __v) {
+  return __nvvm_atom_sys_and_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicCAS(unsigned int *__p, unsigned int __cmp,
+                                     unsigned int __v) {
+  return __nvvm_atom_cas_gen_i((int *)__p, __cmp, __v);
+}
+__DEVICE__ unsigned int
+__uAtomicCAS_block(unsigned int *__p, unsigned int __cmp, unsigned int __v) {
+  return __nvvm_atom_cta_cas_gen_i((int *)__p, __cmp, __v);
+}
+__DEVICE__ unsigned int
+__uAtomicCAS_system(unsigned int *__p, unsigned int __cmp, unsigned int __v) {
+  return __nvvm_atom_sys_cas_gen_i((int *)__p, __cmp, __v);
+}
+__DEVICE__ unsigned int __uAtomicDec(unsigned int *__p, unsigned int __v) {
+  return __nvvm_atom_dec_gen_ui(__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicDec_block(unsigned int *__p,
+                                           unsigned int __v) {
+  return __nvvm_atom_cta_dec_gen_ui(__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicDec_system(unsigned int *__p,
+                                            unsigned int __v) {
+  return __nvvm_atom_sys_dec_gen_ui(__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicExch(unsigned int *__p, unsigned int __v) {
+  return __nvvm_atom_xchg_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicExch_block(unsigned int *__p,
+                                            unsigned int __v) {
+  return __nvvm_atom_cta_xchg_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicExch_system(unsigned int *__p,
+                                             unsigned int __v) {
+  return __nvvm_atom_sys_xchg_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicInc(unsigned int *__p, unsigned int __v) {
+  return __nvvm_atom_inc_gen_ui(__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicInc_block(unsigned int *__p,
+                                           unsigned int __v) {
+  return __nvvm_atom_cta_inc_gen_ui(__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicInc_system(unsigned int *__p,
+                                            unsigned int __v) {
+  return __nvvm_atom_sys_inc_gen_ui(__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicMax(unsigned int *__p, unsigned int __v) {
+  return __nvvm_atom_max_gen_ui(__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicMax_block(unsigned int *__p,
+                                           unsigned int __v) {
+  return __nvvm_atom_cta_max_gen_ui(__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicMax_system(unsigned int *__p,
+                                            unsigned int __v) {
+  return __nvvm_atom_sys_max_gen_ui(__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicMin(unsigned int *__p, unsigned int __v) {
+  return __nvvm_atom_min_gen_ui(__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicMin_block(unsigned int *__p,
+                                           unsigned int __v) {
+  return __nvvm_atom_cta_min_gen_ui(__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicMin_system(unsigned int *__p,
+                                            unsigned int __v) {
+  return __nvvm_atom_sys_min_gen_ui(__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicOr(unsigned int *__p, unsigned int __v) {
+  return __nvvm_atom_or_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicOr_block(unsigned int *__p, unsigned int __v) {
+  return __nvvm_atom_cta_or_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicOr_system(unsigned int *__p,
+                                           unsigned int __v) {
+  return __nvvm_atom_sys_or_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicXor(unsigned int *__p, unsigned int __v) {
+  return __nvvm_atom_xor_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicXor_block(unsigned int *__p,
+                                           unsigned int __v) {
+  return __nvvm_atom_cta_xor_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uAtomicXor_system(unsigned int *__p,
+                                            unsigned int __v) {
+  return __nvvm_atom_sys_xor_gen_i((int *)__p, __v);
+}
+__DEVICE__ unsigned int __uhadd(unsigned int __a, unsigned int __b) {
+  return __nv_uhadd(__a, __b);
+}
+__DEVICE__ double __uint2double_rn(unsigned int __a) {
+  return __nv_uint2double_rn(__a);
+}
+__DEVICE__ float __uint2float_rd(unsigned int __a) {
+  return __nv_uint2float_rd(__a);
+}
+__DEVICE__ float __uint2float_rn(unsigned int __a) {
+  return __nv_uint2float_rn(__a);
+}
+__DEVICE__ float __uint2float_ru(unsigned int __a) {
+  return __nv_uint2float_ru(__a);
+}
+__DEVICE__ float __uint2float_rz(unsigned int __a) {
+  return __nv_uint2float_rz(__a);
+}
+__DEVICE__ float __uint_as_float(unsigned int __a) {
+  return __nv_uint_as_float(__a);
+} //
+__DEVICE__ double __ull2double_rd(unsigned long long __a) {
+  return __nv_ull2double_rd(__a);
+}
+__DEVICE__ double __ull2double_rn(unsigned long long __a) {
+  return __nv_ull2double_rn(__a);
+}
+__DEVICE__ double __ull2double_ru(unsigned long long __a) {
+  return __nv_ull2double_ru(__a);
+}
+__DEVICE__ double __ull2double_rz(unsigned long long __a) {
+  return __nv_ull2double_rz(__a);
+}
+__DEVICE__ float __ull2float_rd(unsigned long long __a) {
+  return __nv_ull2float_rd(__a);
+}
+__DEVICE__ float __ull2float_rn(unsigned long long __a) {
+  return __nv_ull2float_rn(__a);
+}
+__DEVICE__ float __ull2float_ru(unsigned long long __a) {
+  return __nv_ull2float_ru(__a);
+}
+__DEVICE__ float __ull2float_rz(unsigned long long __a) {
+  return __nv_ull2float_rz(__a);
+}
+__DEVICE__ unsigned long long __ullAtomicAdd_block(unsigned long long *__p,
+                                                   unsigned long long __v) {
+  return __nvvm_atom_cta_add_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicAdd_system(unsigned long long *__p,
+                                                    unsigned long long __v) {
+  return __nvvm_atom_sys_add_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicAnd(unsigned long long *__p,
+                                             unsigned long long __v) {
+  return __nvvm_atom_and_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicAnd_block(unsigned long long *__p,
+                                                   unsigned long long __v) {
+  return __nvvm_atom_cta_and_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicAnd_system(unsigned long long *__p,
+                                                    unsigned long long __v) {
+  return __nvvm_atom_sys_and_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicCAS_block(unsigned long long *__p,
+                                                   unsigned long long __cmp,
+                                                   unsigned long long __v) {
+  return __nvvm_atom_cta_cas_gen_ll((long long *)__p, __cmp, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicCAS_system(unsigned long long *__p,
+                                                    unsigned long long __cmp,
+                                                    unsigned long long __v) {
+  return __nvvm_atom_sys_cas_gen_ll((long long *)__p, __cmp, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicExch_block(unsigned long long *__p,
+                                                    unsigned long long __v) {
+  return __nvvm_atom_cta_xchg_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicExch_system(unsigned long long *__p,
+                                                     unsigned long long __v) {
+  return __nvvm_atom_sys_xchg_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicMax(unsigned long long *__p,
+                                             unsigned long long __v) {
+  return __nvvm_atom_max_gen_ull(__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicMax_block(unsigned long long *__p,
+                                                   unsigned long long __v) {
+  return __nvvm_atom_cta_max_gen_ull(__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicMax_system(unsigned long long *__p,
+                                                    unsigned long long __v) {
+  return __nvvm_atom_sys_max_gen_ull(__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicMin(unsigned long long *__p,
+                                             unsigned long long __v) {
+  return __nvvm_atom_min_gen_ull(__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicMin_block(unsigned long long *__p,
+                                                   unsigned long long __v) {
+  return __nvvm_atom_cta_min_gen_ull(__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicMin_system(unsigned long long *__p,
+                                                    unsigned long long __v) {
+  return __nvvm_atom_sys_min_gen_ull(__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicOr(unsigned long long *__p,
+                                            unsigned long long __v) {
+  return __nvvm_atom_or_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicOr_block(unsigned long long *__p,
+                                                  unsigned long long __v) {
+  return __nvvm_atom_cta_or_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicOr_system(unsigned long long *__p,
+                                                   unsigned long long __v) {
+  return __nvvm_atom_sys_or_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicXor(unsigned long long *__p,
+                                             unsigned long long __v) {
+  return __nvvm_atom_xor_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicXor_block(unsigned long long *__p,
+                                                   unsigned long long __v) {
+  return __nvvm_atom_cta_xor_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned long long __ullAtomicXor_system(unsigned long long *__p,
+                                                    unsigned long long __v) {
+  return __nvvm_atom_sys_xor_gen_ll((long long *)__p, __v);
+}
+__DEVICE__ unsigned int __umul24(unsigned int __a, unsigned int __b) {
+  return __nv_umul24(__a, __b);
+}
+__DEVICE__ unsigned long long __umul64hi(unsigned long long __a,
+                                         unsigned long long __b) {
+  return __nv_umul64hi(__a, __b);
+}
+__DEVICE__ unsigned int __umulhi(unsigned int __a, unsigned int __b) {
+  return __nv_umulhi(__a, __b);
+}
+__DEVICE__ unsigned int __urhadd(unsigned int __a, unsigned int __b) {
+  return __nv_urhadd(__a, __b);
+}
+__DEVICE__ unsigned int __usad(unsigned int __a, unsigned int __b,
+                               unsigned int __c) {
+  return __nv_usad(__a, __b, __c);
+}
+__DEVICE__ unsigned int __vabs2(unsigned int __a) { return __nv_vabs2(__a); }
+__DEVICE__ unsigned int __vabs4(unsigned int __a) { return __nv_vabs4(__a); }
+__DEVICE__ unsigned int __vabsdiffs2(unsigned int __a, unsigned int __b) {
+  return __nv_vabsdiffs2(__a, __b);
+}
+__DEVICE__ unsigned int __vabsdiffs4(unsigned int __a, unsigned int __b) {
+  return __nv_vabsdiffs4(__a, __b);
+}
+__DEVICE__ unsigned int __vabsdiffu2(unsigned int __a, unsigned int __b) {
+  return __nv_vabsdiffu2(__a, __b);
+}
+__DEVICE__ unsigned int __vabsdiffu4(unsigned int __a, unsigned int __b) {
+  return __nv_vabsdiffu4(__a, __b);
+}
+__DEVICE__ unsigned int __vabsss2(unsigned int __a) {
+  return __nv_vabsss2(__a);
+}
+__DEVICE__ unsigned int __vabsss4(unsigned int __a) {
+  return __nv_vabsss4(__a);
+}
+__DEVICE__ unsigned int __vadd2(unsigned int __a, unsigned int __b) {
+  return __nv_vadd2(__a, __b);
+}
+__DEVICE__ unsigned int __vadd4(unsigned int __a, unsigned int __b) {
+  return __nv_vadd4(__a, __b);
+}
+__DEVICE__ unsigned int __vaddss2(unsigned int __a, unsigned int __b) {
+  return __nv_vaddss2(__a, __b);
+}
+__DEVICE__ unsigned int __vaddss4(unsigned int __a, unsigned int __b) {
+  return __nv_vaddss4(__a, __b);
+}
+__DEVICE__ unsigned int __vaddus2(unsigned int __a, unsigned int __b) {
+  return __nv_vaddus2(__a, __b);
+}
+__DEVICE__ unsigned int __vaddus4(unsigned int __a, unsigned int __b) {
+  return __nv_vaddus4(__a, __b);
+}
+__DEVICE__ unsigned int __vavgs2(unsigned int __a, unsigned int __b) {
+  return __nv_vavgs2(__a, __b);
+}
+__DEVICE__ unsigned int __vavgs4(unsigned int __a, unsigned int __b) {
+  return __nv_vavgs4(__a, __b);
+}
+__DEVICE__ unsigned int __vavgu2(unsigned int __a, unsigned int __b) {
+  return __nv_vavgu2(__a, __b);
+}
+__DEVICE__ unsigned int __vavgu4(unsigned int __a, unsigned int __b) {
+  return __nv_vavgu4(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpeq2(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpeq2(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpeq4(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpeq4(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpges2(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpges2(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpges4(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpges4(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpgeu2(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpgeu2(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpgeu4(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpgeu4(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpgts2(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpgts2(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpgts4(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpgts4(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpgtu2(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpgtu2(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpgtu4(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpgtu4(__a, __b);
+}
+__DEVICE__ unsigned int __vcmples2(unsigned int __a, unsigned int __b) {
+  return __nv_vcmples2(__a, __b);
+}
+__DEVICE__ unsigned int __vcmples4(unsigned int __a, unsigned int __b) {
+  return __nv_vcmples4(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpleu2(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpleu2(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpleu4(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpleu4(__a, __b);
+}
+__DEVICE__ unsigned int __vcmplts2(unsigned int __a, unsigned int __b) {
+  return __nv_vcmplts2(__a, __b);
+}
+__DEVICE__ unsigned int __vcmplts4(unsigned int __a, unsigned int __b) {
+  return __nv_vcmplts4(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpltu2(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpltu2(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpltu4(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpltu4(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpne2(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpne2(__a, __b);
+}
+__DEVICE__ unsigned int __vcmpne4(unsigned int __a, unsigned int __b) {
+  return __nv_vcmpne4(__a, __b);
+}
+__DEVICE__ unsigned int __vhaddu2(unsigned int __a, unsigned int __b) {
+  return __nv_vhaddu2(__a, __b);
+}
+__DEVICE__ unsigned int __vhaddu4(unsigned int __a, unsigned int __b) {
+  return __nv_vhaddu4(__a, __b);
+}
+__DEVICE__ unsigned int __vmaxs2(unsigned int __a, unsigned int __b) {
+  return __nv_vmaxs2(__a, __b);
+}
+__DEVICE__ unsigned int __vmaxs4(unsigned int __a, unsigned int __b) {
+  return __nv_vmaxs4(__a, __b);
+}
+__DEVICE__ unsigned int __vmaxu2(unsigned int __a, unsigned int __b) {
+  return __nv_vmaxu2(__a, __b);
+}
+__DEVICE__ unsigned int __vmaxu4(unsigned int __a, unsigned int __b) {
+  return __nv_vmaxu4(__a, __b);
+}
+__DEVICE__ unsigned int __vmins2(unsigned int __a, unsigned int __b) {
+  return __nv_vmins2(__a, __b);
+}
+__DEVICE__ unsigned int __vmins4(unsigned int __a, unsigned int __b) {
+  return __nv_vmins4(__a, __b);
+}
+__DEVICE__ unsigned int __vminu2(unsigned int __a, unsigned int __b) {
+  return __nv_vminu2(__a, __b);
+}
+__DEVICE__ unsigned int __vminu4(unsigned int __a, unsigned int __b) {
+  return __nv_vminu4(__a, __b);
+}
+__DEVICE__ unsigned int __vneg2(unsigned int __a) { return __nv_vneg2(__a); }
+__DEVICE__ unsigned int __vneg4(unsigned int __a) { return __nv_vneg4(__a); }
+__DEVICE__ unsigned int __vnegss2(unsigned int __a) {
+  return __nv_vnegss2(__a);
+}
+__DEVICE__ unsigned int __vnegss4(unsigned int __a) {
+  return __nv_vnegss4(__a);
+}
+__DEVICE__ unsigned int __vsads2(unsigned int __a, unsigned int __b) {
+  return __nv_vsads2(__a, __b);
+}
+__DEVICE__ unsigned int __vsads4(unsigned int __a, unsigned int __b) {
+  return __nv_vsads4(__a, __b);
+}
+__DEVICE__ unsigned int __vsadu2(unsigned int __a, unsigned int __b) {
+  return __nv_vsadu2(__a, __b);
+}
+__DEVICE__ unsigned int __vsadu4(unsigned int __a, unsigned int __b) {
+  return __nv_vsadu4(__a, __b);
+}
+__DEVICE__ unsigned int __vseteq2(unsigned int __a, unsigned int __b) {
+  return __nv_vseteq2(__a, __b);
+}
+__DEVICE__ unsigned int __vseteq4(unsigned int __a, unsigned int __b) {
+  return __nv_vseteq4(__a, __b);
+}
+__DEVICE__ unsigned int __vsetges2(unsigned int __a, unsigned int __b) {
+  return __nv_vsetges2(__a, __b);
+}
+__DEVICE__ unsigned int __vsetges4(unsigned int __a, unsigned int __b) {
+  return __nv_vsetges4(__a, __b);
+}
+__DEVICE__ unsigned int __vsetgeu2(unsigned int __a, unsigned int __b) {
+  return __nv_vsetgeu2(__a, __b);
+}
+__DEVICE__ unsigned int __vsetgeu4(unsigned int __a, unsigned int __b) {
+  return __nv_vsetgeu4(__a, __b);
+}
+__DEVICE__ unsigned int __vsetgts2(unsigned int __a, unsigned int __b) {
+  return __nv_vsetgts2(__a, __b);
+}
+__DEVICE__ unsigned int __vsetgts4(unsigned int __a, unsigned int __b) {
+  return __nv_vsetgts4(__a, __b);
+}
+__DEVICE__ unsigned int __vsetgtu2(unsigned int __a, unsigned int __b) {
+  return __nv_vsetgtu2(__a, __b);
+}
+__DEVICE__ unsigned int __vsetgtu4(unsigned int __a, unsigned int __b) {
+  return __nv_vsetgtu4(__a, __b);
+}
+__DEVICE__ unsigned int __vsetles2(unsigned int __a, unsigned int __b) {
+  return __nv_vsetles2(__a, __b);
+}
+__DEVICE__ unsigned int __vsetles4(unsigned int __a, unsigned int __b) {
+  return __nv_vsetles4(__a, __b);
+}
+__DEVICE__ unsigned int __vsetleu2(unsigned int __a, unsigned int __b) {
+  return __nv_vsetleu2(__a, __b);
+}
+__DEVICE__ unsigned int __vsetleu4(unsigned int __a, unsigned int __b) {
+  return __nv_vsetleu4(__a, __b);
+}
+__DEVICE__ unsigned int __vsetlts2(unsigned int __a, unsigned int __b) {
+  return __nv_vsetlts2(__a, __b);
+}
+__DEVICE__ unsigned int __vsetlts4(unsigned int __a, unsigned int __b) {
+  return __nv_vsetlts4(__a, __b);
+}
+__DEVICE__ unsigned int __vsetltu2(unsigned int __a, unsigned int __b) {
+  return __nv_vsetltu2(__a, __b);
+}
+__DEVICE__ unsigned int __vsetltu4(unsigned int __a, unsigned int __b) {
+  return __nv_vsetltu4(__a, __b);
+}
+__DEVICE__ unsigned int __vsetne2(unsigned int __a, unsigned int __b) {
+  return __nv_vsetne2(__a, __b);
+}
+__DEVICE__ unsigned int __vsetne4(unsigned int __a, unsigned int __b) {
+  return __nv_vsetne4(__a, __b);
+}
+__DEVICE__ unsigned int __vsub2(unsigned int __a, unsigned int __b) {
+  return __nv_vsub2(__a, __b);
+}
+__DEVICE__ unsigned int __vsub4(unsigned int __a, unsigned int __b) {
+  return __nv_vsub4(__a, __b);
+}
+__DEVICE__ unsigned int __vsubss2(unsigned int __a, unsigned int __b) {
+  return __nv_vsubss2(__a, __b);
+}
+__DEVICE__ unsigned int __vsubss4(unsigned int __a, unsigned int __b) {
+  return __nv_vsubss4(__a, __b);
+}
+__DEVICE__ unsigned int __vsubus2(unsigned int __a, unsigned int __b) {
+  return __nv_vsubus2(__a, __b);
+}
+__DEVICE__ unsigned int __vsubus4(unsigned int __a, unsigned int __b) {
+  return __nv_vsubus4(__a, __b);
+}
+__DEVICE__ int abs(int __a) { return __nv_abs(__a); }
+__DEVICE__ double acos(double __a) { return __nv_acos(__a); }
+__DEVICE__ float acosf(float __a) { return __nv_acosf(__a); }
+__DEVICE__ double acosh(double __a) { return __nv_acosh(__a); }
+__DEVICE__ float acoshf(float __a) { return __nv_acoshf(__a); }
+__DEVICE__ double asin(double __a) { return __nv_asin(__a); }
+__DEVICE__ float asinf(float __a) { return __nv_asinf(__a); }
+__DEVICE__ double asinh(double __a) { return __nv_asinh(__a); }
+__DEVICE__ float asinhf(float __a) { return __nv_asinhf(__a); }
+__DEVICE__ double atan(double __a) { return __nv_atan(__a); }
+__DEVICE__ double atan2(double __a, double __b) { return __nv_atan2(__a, __b); }
+__DEVICE__ float atan2f(float __a, float __b) { return __nv_atan2f(__a, __b); }
+__DEVICE__ float atanf(float __a) { return __nv_atanf(__a); }
+__DEVICE__ double atanh(double __a) { return __nv_atanh(__a); }
+__DEVICE__ float atanhf(float __a) { return __nv_atanhf(__a); }
+__DEVICE__ double cbrt(double __a) { return __nv_cbrt(__a); }
+__DEVICE__ float cbrtf(float __a) { return __nv_cbrtf(__a); }
+__DEVICE__ double ceil(double __a) { return __nv_ceil(__a); }
+__DEVICE__ float ceilf(float __a) { return __nv_ceilf(__a); }
+__DEVICE__ int clock() { return __nvvm_read_ptx_sreg_clock(); }
+__DEVICE__ long long clock64() { return __nvvm_read_ptx_sreg_clock64(); }
+__DEVICE__ double copysign(double __a, double __b) {
+  return __nv_copysign(__a, __b);
+}
+__DEVICE__ float copysignf(float __a, float __b) {
+  return __nv_copysignf(__a, __b);
+}
+__DEVICE__ double cos(double __a) { return __nv_cos(__a); }
+__DEVICE__ float cosf(float __a) {
+  return __FAST_OR_SLOW(__nv_fast_cosf, __nv_cosf)(__a);
+}
+__DEVICE__ double cosh(double __a) { return __nv_cosh(__a); }
+__DEVICE__ float coshf(float __a) { return __nv_coshf(__a); }
+__DEVICE__ double cospi(double __a) { return __nv_cospi(__a); }
+__DEVICE__ float cospif(float __a) { return __nv_cospif(__a); }
+__DEVICE__ double cyl_bessel_i0(double __a) { return __nv_cyl_bessel_i0(__a); }
+__DEVICE__ float cyl_bessel_i0f(float __a) { return __nv_cyl_bessel_i0f(__a); }
+__DEVICE__ double cyl_bessel_i1(double __a) { return __nv_cyl_bessel_i1(__a); }
+__DEVICE__ float cyl_bessel_i1f(float __a) { return __nv_cyl_bessel_i1f(__a); }
+__DEVICE__ double erf(double __a) { return __nv_erf(__a); }
+__DEVICE__ double erfc(double __a) { return __nv_erfc(__a); }
+__DEVICE__ float erfcf(float __a) { return __nv_erfcf(__a); }
+__DEVICE__ double erfcinv(double __a) { return __nv_erfcinv(__a); }
+__DEVICE__ float erfcinvf(float __a) { return __nv_erfcinvf(__a); }
+__DEVICE__ double erfcx(double __a) { return __nv_erfcx(__a); }
+__DEVICE__ float erfcxf(float __a) { return __nv_erfcxf(__a); }
+__DEVICE__ float erff(float __a) { return __nv_erff(__a); }
+__DEVICE__ double erfinv(double __a) { return __nv_erfinv(__a); }
+__DEVICE__ float erfinvf(float __a) { return __nv_erfinvf(__a); }
+__DEVICE__ double exp(double __a) { return __nv_exp(__a); }
+__DEVICE__ double exp10(double __a) { return __nv_exp10(__a); }
+__DEVICE__ float exp10f(float __a) { return __nv_exp10f(__a); }
+__DEVICE__ double exp2(double __a) { return __nv_exp2(__a); }
+__DEVICE__ float exp2f(float __a) { return __nv_exp2f(__a); }
+__DEVICE__ float expf(float __a) { return __nv_expf(__a); }
+__DEVICE__ double expm1(double __a) { return __nv_expm1(__a); }
+__DEVICE__ float expm1f(float __a) { return __nv_expm1f(__a); }
+__DEVICE__ double fabs(double __a) { return __nv_fabs(__a); }
+__DEVICE__ float fabsf(float __a) { return __nv_fabsf(__a); }
+__DEVICE__ double fdim(double __a, double __b) { return __nv_fdim(__a, __b); }
+__DEVICE__ float fdimf(float __a, float __b) { return __nv_fdimf(__a, __b); }
+__DEVICE__ double fdivide(double __a, double __b) { return __a / __b; }
+__DEVICE__ float fdividef(float __a, float __b) {
+#if __FAST_MATH__ && !__CUDA_PREC_DIV
+  return __nv_fast_fdividef(__a, __b);
+#else
+  return __a / __b;
+#endif
+}
+__DEVICE__ double floor(double __f) { return __nv_floor(__f); }
+__DEVICE__ float floorf(float __f) { return __nv_floorf(__f); }
+__DEVICE__ double fma(double __a, double __b, double __c) {
+  return __nv_fma(__a, __b, __c);
+}
+__DEVICE__ float fmaf(float __a, float __b, float __c) {
+  return __nv_fmaf(__a, __b, __c);
+}
+__DEVICE__ double fmax(double __a, double __b) { return __nv_fmax(__a, __b); }
+__DEVICE__ float fmaxf(float __a, float __b) { return __nv_fmaxf(__a, __b); }
+__DEVICE__ double fmin(double __a, double __b) { return __nv_fmin(__a, __b); }
+__DEVICE__ float fminf(float __a, float __b) { return __nv_fminf(__a, __b); }
+__DEVICE__ double fmod(double __a, double __b) { return __nv_fmod(__a, __b); }
+__DEVICE__ float fmodf(float __a, float __b) { return __nv_fmodf(__a, __b); }
+__DEVICE__ double frexp(double __a, int *__b) { return __nv_frexp(__a, __b); }
+__DEVICE__ float frexpf(float __a, int *__b) { return __nv_frexpf(__a, __b); }
+__DEVICE__ double hypot(double __a, double __b) { return __nv_hypot(__a, __b); }
+__DEVICE__ float hypotf(float __a, float __b) { return __nv_hypotf(__a, __b); }
+__DEVICE__ int ilogb(double __a) { return __nv_ilogb(__a); }
+__DEVICE__ int ilogbf(float __a) { return __nv_ilogbf(__a); }
+__DEVICE__ double j0(double __a) { return __nv_j0(__a); }
+__DEVICE__ float j0f(float __a) { return __nv_j0f(__a); }
+__DEVICE__ double j1(double __a) { return __nv_j1(__a); }
+__DEVICE__ float j1f(float __a) { return __nv_j1f(__a); }
+__DEVICE__ double jn(int __n, double __a) { return __nv_jn(__n, __a); }
+__DEVICE__ float jnf(int __n, float __a) { return __nv_jnf(__n, __a); }
+__DEVICE__ double ldexp(double __a, int __b) { return __nv_ldexp(__a, __b); }
+__DEVICE__ float ldexpf(float __a, int __b) { return __nv_ldexpf(__a, __b); }
+__DEVICE__ double lgamma(double __a) { return __nv_lgamma(__a); }
+__DEVICE__ float lgammaf(float __a) { return __nv_lgammaf(__a); }
+__DEVICE__ long long llmax(long long __a, long long __b) {
+  return __nv_llmax(__a, __b);
+}
+__DEVICE__ long long llmin(long long __a, long long __b) {
+  return __nv_llmin(__a, __b);
+}
+__DEVICE__ long long llrint(double __a) { return __nv_llrint(__a); }
+__DEVICE__ long long llrintf(float __a) { return __nv_llrintf(__a); }
+__DEVICE__ long long llround(double __a) { return __nv_llround(__a); }
+__DEVICE__ long long llroundf(float __a) { return __nv_llroundf(__a); }
+__DEVICE__ double log(double __a) { return __nv_log(__a); }
+__DEVICE__ double log10(double __a) { return __nv_log10(__a); }
+__DEVICE__ float log10f(float __a) { return __nv_log10f(__a); }
+__DEVICE__ double log1p(double __a) { return __nv_log1p(__a); }
+__DEVICE__ float log1pf(float __a) { return __nv_log1pf(__a); }
+__DEVICE__ double log2(double __a) { return __nv_log2(__a); }
+__DEVICE__ float log2f(float __a) {
+  return __FAST_OR_SLOW(__nv_fast_log2f, __nv_log2f)(__a);
+}
+__DEVICE__ double logb(double __a) { return __nv_logb(__a); }
+__DEVICE__ float logbf(float __a) { return __nv_logbf(__a); }
+__DEVICE__ float logf(float __a) {
+  return __FAST_OR_SLOW(__nv_fast_logf, __nv_logf)(__a);
+}
+#if defined(__LP64__)
+__DEVICE__ long lrint(double __a) { return llrint(__a); }
+__DEVICE__ long lrintf(float __a) { return __float2ll_rn(__a); }
+__DEVICE__ long lround(double __a) { return llround(__a); }
+__DEVICE__ long lroundf(float __a) { return llroundf(__a); }
+#else
+__DEVICE__ long lrint(double __a) { return (long)rint(__a); }
+__DEVICE__ long lrintf(float __a) { return __float2int_rn(__a); }
+__DEVICE__ long lround(double __a) { return round(__a); }
+__DEVICE__ long lroundf(float __a) { return roundf(__a); }
+#endif
+__DEVICE__ int max(int __a, int __b) { return __nv_max(__a, __b); }
+__DEVICE__ void *memcpy(void *__a, const void *__b, size_t __c) {
+  return __builtin_memcpy(__a, __b, __c);
+}
+__DEVICE__ void *memset(void *__a, int __b, size_t __c) {
+  return __builtin_memset(__a, __b, __c);
+}
+__DEVICE__ int min(int __a, int __b) { return __nv_min(__a, __b); }
+__DEVICE__ double modf(double __a, double *__b) { return __nv_modf(__a, __b); }
+__DEVICE__ float modff(float __a, float *__b) { return __nv_modff(__a, __b); }
+__DEVICE__ double nearbyint(double __a) { return __nv_nearbyint(__a); }
+__DEVICE__ float nearbyintf(float __a) { return __nv_nearbyintf(__a); }
+__DEVICE__ double nextafter(double __a, double __b) {
+  return __nv_nextafter(__a, __b);
+}
+__DEVICE__ float nextafterf(float __a, float __b) {
+  return __nv_nextafterf(__a, __b);
+}
+__DEVICE__ double norm(int __dim, const double *__t) {
+  return __nv_norm(__dim, __t);
+}
+__DEVICE__ double norm3d(double __a, double __b, double __c) {
+  return __nv_norm3d(__a, __b, __c);
+}
+__DEVICE__ float norm3df(float __a, float __b, float __c) {
+  return __nv_norm3df(__a, __b, __c);
+}
+__DEVICE__ double norm4d(double __a, double __b, double __c, double __d) {
+  return __nv_norm4d(__a, __b, __c, __d);
+}
+__DEVICE__ float norm4df(float __a, float __b, float __c, float __d) {
+  return __nv_norm4df(__a, __b, __c, __d);
+}
+__DEVICE__ double normcdf(double __a) { return __nv_normcdf(__a); }
+__DEVICE__ float normcdff(float __a) { return __nv_normcdff(__a); }
+__DEVICE__ double normcdfinv(double __a) { return __nv_normcdfinv(__a); }
+__DEVICE__ float normcdfinvf(float __a) { return __nv_normcdfinvf(__a); }
+__DEVICE__ float normf(int __dim, const float *__t) {
+  return __nv_normf(__dim, __t);
+}
+__DEVICE__ double pow(double __a, double __b) { return __nv_pow(__a, __b); }
+__DEVICE__ float powf(float __a, float __b) { return __nv_powf(__a, __b); }
+__DEVICE__ double powi(double __a, int __b) { return __nv_powi(__a, __b); }
+__DEVICE__ float powif(float __a, int __b) { return __nv_powif(__a, __b); }
+__DEVICE__ double rcbrt(double __a) { return __nv_rcbrt(__a); }
+__DEVICE__ float rcbrtf(float __a) { return __nv_rcbrtf(__a); }
+__DEVICE__ double remainder(double __a, double __b) {
+  return __nv_remainder(__a, __b);
+}
+__DEVICE__ float remainderf(float __a, float __b) {
+  return __nv_remainderf(__a, __b);
+}
+__DEVICE__ double remquo(double __a, double __b, int *__c) {
+  return __nv_remquo(__a, __b, __c);
+}
+__DEVICE__ float remquof(float __a, float __b, int *__c) {
+  return __nv_remquof(__a, __b, __c);
+}
+__DEVICE__ double rhypot(double __a, double __b) {
+  return __nv_rhypot(__a, __b);
+}
+__DEVICE__ float rhypotf(float __a, float __b) {
+  return __nv_rhypotf(__a, __b);
+}
+__DEVICE__ double rint(double __a) { return __nv_rint(__a); }
+__DEVICE__ float rintf(float __a) { return __nv_rintf(__a); }
+__DEVICE__ double rnorm(int __a, const double *__b) {
+  return __nv_rnorm(__a, __b);
+}
+__DEVICE__ double rnorm3d(double __a, double __b, double __c) {
+  return __nv_rnorm3d(__a, __b, __c);
+}
+__DEVICE__ float rnorm3df(float __a, float __b, float __c) {
+  return __nv_rnorm3df(__a, __b, __c);
+}
+__DEVICE__ double rnorm4d(double __a, double __b, double __c, double __d) {
+  return __nv_rnorm4d(__a, __b, __c, __d);
+}
+__DEVICE__ float rnorm4df(float __a, float __b, float __c, float __d) {
+  return __nv_rnorm4df(__a, __b, __c, __d);
+}
+__DEVICE__ float rnormf(int __dim, const float *__t) {
+  return __nv_rnormf(__dim, __t);
+}
+__DEVICE__ double round(double __a) { return __nv_round(__a); }
+__DEVICE__ float roundf(float __a) { return __nv_roundf(__a); }
+__DEVICE__ double rsqrt(double __a) { return __nv_rsqrt(__a); }
+__DEVICE__ float rsqrtf(float __a) { return __nv_rsqrtf(__a); }
+__DEVICE__ double scalbn(double __a, int __b) { return __nv_scalbn(__a, __b); }
+__DEVICE__ float scalbnf(float __a, int __b) { return __nv_scalbnf(__a, __b); }
+__DEVICE__ double scalbln(double __a, long __b) {
+  if (__b > INT_MAX)
+    return __a > 0 ? HUGE_VAL : -HUGE_VAL;
+  if (__b < INT_MIN)
+    return __a > 0 ? 0.0 : -0.0;
+  return scalbn(__a, (int)__b);
+}
+__DEVICE__ float scalblnf(float __a, long __b) {
+  if (__b > INT_MAX)
+    return __a > 0 ? HUGE_VALF : -HUGE_VALF;
+  if (__b < INT_MIN)
+    return __a > 0 ? 0.f : -0.f;
+  return scalbnf(__a, (int)__b);
+}
+__DEVICE__ double sin(double __a) { return __nv_sin(__a); }
+__DEVICE__ void sincosf(float __a, float *__sptr, float *__cptr) {
+  return __FAST_OR_SLOW(__nv_fast_sincosf, __nv_sincosf)(__a, __sptr, __cptr);
+}
+__DEVICE__ void sincospi(double __a, double *__sptr, double *__cptr) {
+  return __nv_sincospi(__a, __sptr, __cptr);
+}
+__DEVICE__ void sincospif(float __a, float *__sptr, float *__cptr) {
+  return __nv_sincospif(__a, __sptr, __cptr);
+}
+__DEVICE__ float sinf(float __a) {
+  return __FAST_OR_SLOW(__nv_fast_sinf, __nv_sinf)(__a);
+}
+__DEVICE__ double sinh(double __a) { return __nv_sinh(__a); }
+__DEVICE__ float sinhf(float __a) { return __nv_sinhf(__a); }
+__DEVICE__ double sinpi(double __a) { return __nv_sinpi(__a); }
+__DEVICE__ float sinpif(float __a) { return __nv_sinpif(__a); }
+__DEVICE__ double sqrt(double __a) { return __nv_sqrt(__a); }
+__DEVICE__ float sqrtf(float __a) { return __nv_sqrtf(__a); }
+__DEVICE__ double tan(double __a) { return __nv_tan(__a); }
+__DEVICE__ float tanf(float __a) { return __nv_tanf(__a); }
+__DEVICE__ double tanh(double __a) { return __nv_tanh(__a); }
+__DEVICE__ float tanhf(float __a) { return __nv_tanhf(__a); }
+__DEVICE__ double tgamma(double __a) { return __nv_tgamma(__a); }
+__DEVICE__ float tgammaf(float __a) { return __nv_tgammaf(__a); }
+__DEVICE__ double trunc(double __a) { return __nv_trunc(__a); }
+__DEVICE__ float truncf(float __a) { return __nv_truncf(__a); }
+__DEVICE__ unsigned long long ullmax(unsigned long long __a,
+                                     unsigned long long __b) {
+  return __nv_ullmax(__a, __b);
+}
+__DEVICE__ unsigned long long ullmin(unsigned long long __a,
+                                     unsigned long long __b) {
+  return __nv_ullmin(__a, __b);
+}
+__DEVICE__ unsigned int umax(unsigned int __a, unsigned int __b) {
+  return __nv_umax(__a, __b);
+}
+__DEVICE__ unsigned int umin(unsigned int __a, unsigned int __b) {
+  return __nv_umin(__a, __b);
+}
+__DEVICE__ double y0(double __a) { return __nv_y0(__a); }
+__DEVICE__ float y0f(float __a) { return __nv_y0f(__a); }
+__DEVICE__ double y1(double __a) { return __nv_y1(__a); }
+__DEVICE__ float y1f(float __a) { return __nv_y1f(__a); }
+__DEVICE__ double yn(int __a, double __b) { return __nv_yn(__a, __b); }
+__DEVICE__ float ynf(int __a, float __b) { return __nv_ynf(__a, __b); }
+
+#endif // CUDA_VERSION >= 9000
+#pragma pop_macro("__DEVICE__")
+#pragma pop_macro("__FAST_OR_SLOW")
+#endif // __CLANG_CUDA_DEVICE_FUNCTIONS_H__

clang/lib/Headers/__clang_cuda_libdevice_declares.h

+/*===-- __clang_cuda_libdevice_declares.h - decls for libdevice functions --===
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ *
+ *===-----------------------------------------------------------------------===
+ */
+
+#ifndef __CLANG_CUDA_LIBDEVICE_DECLARES_H__
+#define __CLANG_CUDA_LIBDEVICE_DECLARES_H__
+
+extern "C" {
+
+__device__ int __nv_abs(int __a);
+__device__ double __nv_acos(double __a);
+__device__ float __nv_acosf(float __a);
+__device__ double __nv_acosh(double __a);
+__device__ float __nv_acoshf(float __a);
+__device__ double __nv_asin(double __a);
+__device__ float __nv_asinf(float __a);
+__device__ double __nv_asinh(double __a);
+__device__ float __nv_asinhf(float __a);
+__device__ double __nv_atan2(double __a, double __b);
+__device__ float __nv_atan2f(float __a, float __b);
+__device__ double __nv_atan(double __a);
+__device__ float __nv_atanf(float __a);
+__device__ double __nv_atanh(double __a);
+__device__ float __nv_atanhf(float __a);
+__device__ int __nv_brev(int __a);
+__device__ long long __nv_brevll(long long __a);
+__device__ int __nv_byte_perm(int __a, int __b, int __c);
+__device__ double __nv_cbrt(double __a);
+__device__ float __nv_cbrtf(float __a);
+__device__ double __nv_ceil(double __a);
+__device__ float __nv_ceilf(float __a);
+__device__ int __nv_clz(int __a);
+__device__ int __nv_clzll(long long __a);
+__device__ double __nv_copysign(double __a, double __b);
+__device__ float __nv_copysignf(float __a, float __b);
+__device__ double __nv_cos(double __a);
+__device__ float __nv_cosf(float __a);
+__device__ double __nv_cosh(double __a);
+__device__ float __nv_coshf(float __a);
+__device__ double __nv_cospi(double __a);
+__device__ float __nv_cospif(float __a);
+__device__ double __nv_cyl_bessel_i0(double __a);
+__device__ float __nv_cyl_bessel_i0f(float __a);
+__device__ double __nv_cyl_bessel_i1(double __a);
+__device__ float __nv_cyl_bessel_i1f(float __a);
+__device__ double __nv_dadd_rd(double __a, double __b);
+__device__ double __nv_dadd_rn(double __a, double __b);
+__device__ double __nv_dadd_ru(double __a, double __b);
+__device__ double __nv_dadd_rz(double __a, double __b);
+__device__ double __nv_ddiv_rd(double __a, double __b);
+__device__ double __nv_ddiv_rn(double __a, double __b);
+__device__ double __nv_ddiv_ru(double __a, double __b);
+__device__ double __nv_ddiv_rz(double __a, double __b);
+__device__ double __nv_dmul_rd(double __a, double __b);
+__device__ double __nv_dmul_rn(double __a, double __b);
+__device__ double __nv_dmul_ru(double __a, double __b);
+__device__ double __nv_dmul_rz(double __a, double __b);
+__device__ float __nv_double2float_rd(double __a);
+__device__ float __nv_double2float_rn(double __a);
+__device__ float __nv_double2float_ru(double __a);
+__device__ float __nv_double2float_rz(double __a);
+__device__ int __nv_double2hiint(double __a);
+__device__ int __nv_double2int_rd(double __a);
+__device__ int __nv_double2int_rn(double __a);
+__device__ int __nv_double2int_ru(double __a);
+__device__ int __nv_double2int_rz(double __a);
+__device__ long long __nv_double2ll_rd(double __a);
+__device__ long long __nv_double2ll_rn(double __a);
+__device__ long long __nv_double2ll_ru(double __a);
+__device__ long long __nv_double2ll_rz(double __a);
+__device__ int __nv_double2loint(double __a);
+__device__ unsigned int __nv_double2uint_rd(double __a);
+__device__ unsigned int __nv_double2uint_rn(double __a);
+__device__ unsigned int __nv_double2uint_ru(double __a);
+__device__ unsigned int __nv_double2uint_rz(double __a);
+__device__ unsigned long long __nv_double2ull_rd(double __a);
+__device__ unsigned long long __nv_double2ull_rn(double __a);
+__device__ unsigned long long __nv_double2ull_ru(double __a);
+__device__ unsigned long long __nv_double2ull_rz(double __a);
+__device__ unsigned long long __nv_double_as_longlong(double __a);
+__device__ double __nv_drcp_rd(double __a);
+__device__ double __nv_drcp_rn(double __a);
+__device__ double __nv_drcp_ru(double __a);
+__device__ double __nv_drcp_rz(double __a);
+__device__ double __nv_dsqrt_rd(double __a);
+__device__ double __nv_dsqrt_rn(double __a);
+__device__ double __nv_dsqrt_ru(double __a);
+__device__ double __nv_dsqrt_rz(double __a);
+__device__ double __nv_dsub_rd(double __a, double __b);
+__device__ double __nv_dsub_rn(double __a, double __b);
+__device__ double __nv_dsub_ru(double __a, double __b);
+__device__ double __nv_dsub_rz(double __a, double __b);
+__device__ double __nv_erfc(double __a);
+__device__ float __nv_erfcf(float __a);
+__device__ double __nv_erfcinv(double __a);
+__device__ float __nv_erfcinvf(float __a);
+__device__ double __nv_erfcx(double __a);
+__device__ float __nv_erfcxf(float __a);
+__device__ double __nv_erf(double __a);
+__device__ float __nv_erff(float __a);
+__device__ double __nv_erfinv(double __a);
+__device__ float __nv_erfinvf(float __a);
+__device__ double __nv_exp10(double __a);
+__device__ float __nv_exp10f(float __a);
+__device__ double __nv_exp2(double __a);
+__device__ float __nv_exp2f(float __a);
+__device__ double __nv_exp(double __a);
+__device__ float __nv_expf(float __a);
+__device__ double __nv_expm1(double __a);
+__device__ float __nv_expm1f(float __a);
+__device__ double __nv_fabs(double __a);
+__device__ float __nv_fabsf(float __a);
+__device__ float __nv_fadd_rd(float __a, float __b);
+__device__ float __nv_fadd_rn(float __a, float __b);
+__device__ float __nv_fadd_ru(float __a, float __b);
+__device__ float __nv_fadd_rz(float __a, float __b);
+__device__ float __nv_fast_cosf(float __a);
+__device__ float __nv_fast_exp10f(float __a);
+__device__ float __nv_fast_expf(float __a);
+__device__ float __nv_fast_fdividef(float __a, float __b);
+__device__ float __nv_fast_log10f(float __a);
+__device__ float __nv_fast_log2f(float __a);
+__device__ float __nv_fast_logf(float __a);
+__device__ float __nv_fast_powf(float __a, float __b);
+__device__ void __nv_fast_sincosf(float __a, float *__sptr, float *__cptr);
+__device__ float __nv_fast_sinf(float __a);
+__device__ float __nv_fast_tanf(float __a);
+__device__ double __nv_fdim(double __a, double __b);
+__device__ float __nv_fdimf(float __a, float __b);
+__device__ float __nv_fdiv_rd(float __a, float __b);
+__device__ float __nv_fdiv_rn(float __a, float __b);
+__device__ float __nv_fdiv_ru(float __a, float __b);
+__device__ float __nv_fdiv_rz(float __a, float __b);
+__device__ int __nv_ffs(int __a);
+__device__ int __nv_ffsll(long long __a);
+__device__ int __nv_finitef(float __a);
+__device__ unsigned short __nv_float2half_rn(float __a);
+__device__ int __nv_float2int_rd(float __a);
+__device__ int __nv_float2int_rn(float __a);
+__device__ int __nv_float2int_ru(float __a);
+__device__ int __nv_float2int_rz(float __a);
+__device__ long long __nv_float2ll_rd(float __a);
+__device__ long long __nv_float2ll_rn(float __a);
+__device__ long long __nv_float2ll_ru(float __a);
+__device__ long long __nv_float2ll_rz(float __a);
+__device__ unsigned int __nv_float2uint_rd(float __a);
+__device__ unsigned int __nv_float2uint_rn(float __a);
+__device__ unsigned int __nv_float2uint_ru(float __a);
+__device__ unsigned int __nv_float2uint_rz(float __a);
+__device__ unsigned long long __nv_float2ull_rd(float __a);
+__device__ unsigned long long __nv_float2ull_rn(float __a);
+__device__ unsigned long long __nv_float2ull_ru(float __a);
+__device__ unsigned long long __nv_float2ull_rz(float __a);
+__device__ int __nv_float_as_int(float __a);
+__device__ unsigned int __nv_float_as_uint(float __a);
+__device__ double __nv_floor(double __a);
+__device__ float __nv_floorf(float __a);
+__device__ double __nv_fma(double __a, double __b, double __c);
+__device__ float __nv_fmaf(float __a, float __b, float __c);
+__device__ float __nv_fmaf_ieee_rd(float __a, float __b, float __c);
+__device__ float __nv_fmaf_ieee_rn(float __a, float __b, float __c);
+__device__ float __nv_fmaf_ieee_ru(float __a, float __b, float __c);
+__device__ float __nv_fmaf_ieee_rz(float __a, float __b, float __c);
+__device__ float __nv_fmaf_rd(float __a, float __b, float __c);
+__device__ float __nv_fmaf_rn(float __a, float __b, float __c);
+__device__ float __nv_fmaf_ru(float __a, float __b, float __c);
+__device__ float __nv_fmaf_rz(float __a, float __b, float __c);
+__device__ double __nv_fma_rd(double __a, double __b, double __c);
+__device__ double __nv_fma_rn(double __a, double __b, double __c);
+__device__ double __nv_fma_ru(double __a, double __b, double __c);
+__device__ double __nv_fma_rz(double __a, double __b, double __c);
+__device__ double __nv_fmax(double __a, double __b);
+__device__ float __nv_fmaxf(float __a, float __b);
+__device__ double __nv_fmin(double __a, double __b);
+__device__ float __nv_fminf(float __a, float __b);
+__device__ double __nv_fmod(double __a, double __b);
+__device__ float __nv_fmodf(float __a, float __b);
+__device__ float __nv_fmul_rd(float __a, float __b);
+__device__ float __nv_fmul_rn(float __a, float __b);
+__device__ float __nv_fmul_ru(float __a, float __b);
+__device__ float __nv_fmul_rz(float __a, float __b);
+__device__ float __nv_frcp_rd(float __a);
+__device__ float __nv_frcp_rn(float __a);
+__device__ float __nv_frcp_ru(float __a);
+__device__ float __nv_frcp_rz(float __a);
+__device__ double __nv_frexp(double __a, int *__b);
+__device__ float __nv_frexpf(float __a, int *__b);
+__device__ float __nv_frsqrt_rn(float __a);
+__device__ float __nv_fsqrt_rd(float __a);
+__device__ float __nv_fsqrt_rn(float __a);
+__device__ float __nv_fsqrt_ru(float __a);
+__device__ float __nv_fsqrt_rz(float __a);
+__device__ float __nv_fsub_rd(float __a, float __b);
+__device__ float __nv_fsub_rn(float __a, float __b);
+__device__ float __nv_fsub_ru(float __a, float __b);
+__device__ float __nv_fsub_rz(float __a, float __b);
+__device__ int __nv_hadd(int __a, int __b);
+__device__ float __nv_half2float(unsigned short __h);
+__device__ double __nv_hiloint2double(int __a, int __b);
+__device__ double __nv_hypot(double __a, double __b);
+__device__ float __nv_hypotf(float __a, float __b);
+__device__ int __nv_ilogb(double __a);
+__device__ int __nv_ilogbf(float __a);
+__device__ double __nv_int2double_rn(int __a);
+__device__ float __nv_int2float_rd(int __a);
+__device__ float __nv_int2float_rn(int __a);
+__device__ float __nv_int2float_ru(int __a);
+__device__ float __nv_int2float_rz(int __a);
+__device__ float __nv_int_as_float(int __a);
+__device__ int __nv_isfinited(double __a);
+__device__ int __nv_isinfd(double __a);
+__device__ int __nv_isinff(float __a);
+__device__ int __nv_isnand(double __a);
+__device__ int __nv_isnanf(float __a);
+__device__ double __nv_j0(double __a);
+__device__ float __nv_j0f(float __a);
+__device__ double __nv_j1(double __a);
+__device__ float __nv_j1f(float __a);
+__device__ float __nv_jnf(int __a, float __b);
+__device__ double __nv_jn(int __a, double __b);
+__device__ double __nv_ldexp(double __a, int __b);
+__device__ float __nv_ldexpf(float __a, int __b);
+__device__ double __nv_lgamma(double __a);
+__device__ float __nv_lgammaf(float __a);
+__device__ double __nv_ll2double_rd(long long __a);
+__device__ double __nv_ll2double_rn(long long __a);
+__device__ double __nv_ll2double_ru(long long __a);
+__device__ double __nv_ll2double_rz(long long __a);
+__device__ float __nv_ll2float_rd(long long __a);
+__device__ float __nv_ll2float_rn(long long __a);
+__device__ float __nv_ll2float_ru(long long __a);
+__device__ float __nv_ll2float_rz(long long __a);
+__device__ long long __nv_llabs(long long __a);
+__device__ long long __nv_llmax(long long __a, long long __b);
+__device__ long long __nv_llmin(long long __a, long long __b);
+__device__ long long __nv_llrint(double __a);
+__device__ long long __nv_llrintf(float __a);
+__device__ long long __nv_llround(double __a);
+__device__ long long __nv_llroundf(float __a);
+__device__ double __nv_log10(double __a);
+__device__ float __nv_log10f(float __a);
+__device__ double __nv_log1p(double __a);
+__device__ float __nv_log1pf(float __a);
+__device__ double __nv_log2(double __a);
+__device__ float __nv_log2f(float __a);
+__device__ double __nv_logb(double __a);
+__device__ float __nv_logbf(float __a);
+__device__ double __nv_log(double __a);
+__device__ float __nv_logf(float __a);
+__device__ double __nv_longlong_as_double(long long __a);
+__device__ int __nv_max(int __a, int __b);
+__device__ int __nv_min(int __a, int __b);
+__device__ double __nv_modf(double __a, double *__b);
+__device__ float __nv_modff(float __a, float *__b);
+__device__ int __nv_mul24(int __a, int __b);
+__device__ long long __nv_mul64hi(long long __a, long long __b);
+__device__ int __nv_mulhi(int __a, int __b);
+__device__ double __nv_nan(const signed char *__a);
+__device__ float __nv_nanf(const signed char *__a);
+__device__ double __nv_nearbyint(double __a);
+__device__ float __nv_nearbyintf(float __a);
+__device__ double __nv_nextafter(double __a, double __b);
+__device__ float __nv_nextafterf(float __a, float __b);
+__device__ double __nv_norm3d(double __a, double __b, double __c);
+__device__ float __nv_norm3df(float __a, float __b, float __c);
+__device__ double __nv_norm4d(double __a, double __b, double __c, double __d);
+__device__ float __nv_norm4df(float __a, float __b, float __c, float __d);
+__device__ double __nv_normcdf(double __a);
+__device__ float __nv_normcdff(float __a);
+__device__ double __nv_normcdfinv(double __a);
+__device__ float __nv_normcdfinvf(float __a);
+__device__ float __nv_normf(int __a, const float *__b);
+__device__ double __nv_norm(int __a, const double *__b);
+__device__ int __nv_popc(int __a);
+__device__ int __nv_popcll(long long __a);
+__device__ double __nv_pow(double __a, double __b);
+__device__ float __nv_powf(float __a, float __b);
+__device__ double __nv_powi(double __a, int __b);
+__device__ float __nv_powif(float __a, int __b);
+__device__ double __nv_rcbrt(double __a);
+__device__ float __nv_rcbrtf(float __a);
+__device__ double __nv_rcp64h(double __a);
+__device__ double __nv_remainder(double __a, double __b);
+__device__ float __nv_remainderf(float __a, float __b);
+__device__ double __nv_remquo(double __a, double __b, int *__c);
+__device__ float __nv_remquof(float __a, float __b, int *__c);
+__device__ int __nv_rhadd(int __a, int __b);
+__device__ double __nv_rhypot(double __a, double __b);
+__device__ float __nv_rhypotf(float __a, float __b);
+__device__ double __nv_rint(double __a);
+__device__ float __nv_rintf(float __a);
+__device__ double __nv_rnorm3d(double __a, double __b, double __c);
+__device__ float __nv_rnorm3df(float __a, float __b, float __c);
+__device__ double __nv_rnorm4d(double __a, double __b, double __c, double __d);
+__device__ float __nv_rnorm4df(float __a, float __b, float __c, float __d);
+__device__ float __nv_rnormf(int __a, const float *__b);
+__device__ double __nv_rnorm(int __a, const double *__b);
+__device__ double __nv_round(double __a);
+__device__ float __nv_roundf(float __a);
+__device__ double __nv_rsqrt(double __a);
+__device__ float __nv_rsqrtf(float __a);
+__device__ int __nv_sad(int __a, int __b, int __c);
+__device__ float __nv_saturatef(float __a);
+__device__ double __nv_scalbn(double __a, int __b);
+__device__ float __nv_scalbnf(float __a, int __b);
+__device__ int __nv_signbitd(double __a);
+__device__ int __nv_signbitf(float __a);
+__device__ void __nv_sincos(double __a, double *__b, double *__c);
+__device__ void __nv_sincosf(float __a, float *__b, float *__c);
+__device__ void __nv_sincospi(double __a, double *__b, double *__c);
+__device__ void __nv_sincospif(float __a, float *__b, float *__c);
+__device__ double __nv_sin(double __a);
+__device__ float __nv_sinf(float __a);
+__device__ double __nv_sinh(double __a);
+__device__ float __nv_sinhf(float __a);
+__device__ double __nv_sinpi(double __a);
+__device__ float __nv_sinpif(float __a);
+__device__ double __nv_sqrt(double __a);
+__device__ float __nv_sqrtf(float __a);
+__device__ double __nv_tan(double __a);
+__device__ float __nv_tanf(float __a);
+__device__ double __nv_tanh(double __a);
+__device__ float __nv_tanhf(float __a);
+__device__ double __nv_tgamma(double __a);
+__device__ float __nv_tgammaf(float __a);
+__device__ double __nv_trunc(double __a);
+__device__ float __nv_truncf(float __a);
+__device__ int __nv_uhadd(unsigned int __a, unsigned int __b);
+__device__ double __nv_uint2double_rn(unsigned int __i);
+__device__ float __nv_uint2float_rd(unsigned int __a);
+__device__ float __nv_uint2float_rn(unsigned int __a);
+__device__ float __nv_uint2float_ru(unsigned int __a);
+__device__ float __nv_uint2float_rz(unsigned int __a);
+__device__ float __nv_uint_as_float(unsigned int __a);
+__device__ double __nv_ull2double_rd(unsigned long long __a);
+__device__ double __nv_ull2double_rn(unsigned long long __a);
+__device__ double __nv_ull2double_ru(unsigned long long __a);
+__device__ double __nv_ull2double_rz(unsigned long long __a);
+__device__ float __nv_ull2float_rd(unsigned long long __a);
+__device__ float __nv_ull2float_rn(unsigned long long __a);
+__device__ float __nv_ull2float_ru(unsigned long long __a);
+__device__ float __nv_ull2float_rz(unsigned long long __a);
+__device__ unsigned long long __nv_ullmax(unsigned long long __a,
+                                          unsigned long long __b);
+__device__ unsigned long long __nv_ullmin(unsigned long long __a,
+                                          unsigned long long __b);
+__device__ unsigned int __nv_umax(unsigned int __a, unsigned int __b);
+__device__ unsigned int __nv_umin(unsigned int __a, unsigned int __b);
+__device__ unsigned int __nv_umul24(unsigned int __a, unsigned int __b);
+__device__ unsigned long long __nv_umul64hi(unsigned long long __a,
+                                            unsigned long long __b);
+__device__ unsigned int __nv_umulhi(unsigned int __a, unsigned int __b);
+__device__ unsigned int __nv_urhadd(unsigned int __a, unsigned int __b);
+__device__ unsigned int __nv_usad(unsigned int __a, unsigned int __b,
+                                  unsigned int __c);
+__device__ int __nv_vabs2(int __a);
+__device__ int __nv_vabs4(int __a);
+__device__ int __nv_vabsdiffs2(int __a, int __b);
+__device__ int __nv_vabsdiffs4(int __a, int __b);
+__device__ int __nv_vabsdiffu2(int __a, int __b);
+__device__ int __nv_vabsdiffu4(int __a, int __b);
+__device__ int __nv_vabsss2(int __a);
+__device__ int __nv_vabsss4(int __a);
+__device__ int __nv_vadd2(int __a, int __b);
+__device__ int __nv_vadd4(int __a, int __b);
+__device__ int __nv_vaddss2(int __a, int __b);
+__device__ int __nv_vaddss4(int __a, int __b);
+__device__ int __nv_vaddus2(int __a, int __b);
+__device__ int __nv_vaddus4(int __a, int __b);
+__device__ int __nv_vavgs2(int __a, int __b);
+__device__ int __nv_vavgs4(int __a, int __b);
+__device__ int __nv_vavgu2(int __a, int __b);
+__device__ int __nv_vavgu4(int __a, int __b);
+__device__ int __nv_vcmpeq2(int __a, int __b);
+__device__ int __nv_vcmpeq4(int __a, int __b);
+__device__ int __nv_vcmpges2(int __a, int __b);
+__device__ int __nv_vcmpges4(int __a, int __b);
+__device__ int __nv_vcmpgeu2(int __a, int __b);
+__device__ int __nv_vcmpgeu4(int __a, int __b);
+__device__ int __nv_vcmpgts2(int __a, int __b);
+__device__ int __nv_vcmpgts4(int __a, int __b);
+__device__ int __nv_vcmpgtu2(int __a, int __b);
+__device__ int __nv_vcmpgtu4(int __a, int __b);
+__device__ int __nv_vcmples2(int __a, int __b);
+__device__ int __nv_vcmples4(int __a, int __b);
+__device__ int __nv_vcmpleu2(int __a, int __b);
+__device__ int __nv_vcmpleu4(int __a, int __b);
+__device__ int __nv_vcmplts2(int __a, int __b);
+__device__ int __nv_vcmplts4(int __a, int __b);
+__device__ int __nv_vcmpltu2(int __a, int __b);
+__device__ int __nv_vcmpltu4(int __a, int __b);
+__device__ int __nv_vcmpne2(int __a, int __b);
+__device__ int __nv_vcmpne4(int __a, int __b);
+__device__ int __nv_vhaddu2(int __a, int __b);
+__device__ int __nv_vhaddu4(int __a, int __b);
+__device__ int __nv_vmaxs2(int __a, int __b);
+__device__ int __nv_vmaxs4(int __a, int __b);
+__device__ int __nv_vmaxu2(int __a, int __b);
+__device__ int __nv_vmaxu4(int __a, int __b);
+__device__ int __nv_vmins2(int __a, int __b);
+__device__ int __nv_vmins4(int __a, int __b);
+__device__ int __nv_vminu2(int __a, int __b);
+__device__ int __nv_vminu4(int __a, int __b);
+__device__ int __nv_vneg2(int __a);
+__device__ int __nv_vneg4(int __a);
+__device__ int __nv_vnegss2(int __a);
+__device__ int __nv_vnegss4(int __a);
+__device__ int __nv_vsads2(int __a, int __b);
+__device__ int __nv_vsads4(int __a, int __b);
+__device__ int __nv_vsadu2(int __a, int __b);
+__device__ int __nv_vsadu4(int __a, int __b);
+__device__ int __nv_vseteq2(int __a, int __b);
+__device__ int __nv_vseteq4(int __a, int __b);
+__device__ int __nv_vsetges2(int __a, int __b);
+__device__ int __nv_vsetges4(int __a, int __b);
+__device__ int __nv_vsetgeu2(int __a, int __b);
+__device__ int __nv_vsetgeu4(int __a, int __b);
+__device__ int __nv_vsetgts2(int __a, int __b);
+__device__ int __nv_vsetgts4(int __a, int __b);
+__device__ int __nv_vsetgtu2(int __a, int __b);
+__device__ int __nv_vsetgtu4(int __a, int __b);
+__device__ int __nv_vsetles2(int __a, int __b);
+__device__ int __nv_vsetles4(int __a, int __b);
+__device__ int __nv_vsetleu2(int __a, int __b);
+__device__ int __nv_vsetleu4(int __a, int __b);
+__device__ int __nv_vsetlts2(int __a, int __b);
+__device__ int __nv_vsetlts4(int __a, int __b);
+__device__ int __nv_vsetltu2(int __a, int __b);
+__device__ int __nv_vsetltu4(int __a, int __b);
+__device__ int __nv_vsetne2(int __a, int __b);
+__device__ int __nv_vsetne4(int __a, int __b);
+__device__ int __nv_vsub2(int __a, int __b);
+__device__ int __nv_vsub4(int __a, int __b);
+__device__ int __nv_vsubss2(int __a, int __b);
+__device__ int __nv_vsubss4(int __a, int __b);
+__device__ int __nv_vsubus2(int __a, int __b);
+__device__ int __nv_vsubus4(int __a, int __b);
+__device__ double __nv_y0(double __a);
+__device__ float __nv_y0f(float __a);
+__device__ double __nv_y1(double __a);
+__device__ float __nv_y1f(float __a);
+__device__ float __nv_ynf(int __a, float __b);
+__device__ double __nv_yn(int __a, double __b);
+
+} // extern "C"
+#endif // __CLANG_CUDA_LIBDEVICE_DECLARES_H__

clang/lib/Headers/__clang_cuda_runtime_wrapper.h

 #pragma push_macro("__CUDA_ARCH__")
 
 // WARNING: Preprocessor hacks below are based on specific details of
 // CUDA-7.x headers and are not expected to work with any other
 // version of CUDA headers.
 #include "cuda.h"
 #if !defined(CUDA_VERSION)
 #error "cuda.h did not define CUDA_VERSION"
-#elif CUDA_VERSION < 7000 || CUDA_VERSION > 9000
+#elif CUDA_VERSION < 7000 || CUDA_VERSION > 9010
 #error "Unsupported CUDA version!"
 #endif
 
 // Make largest subset of device functions available during host
 // compilation -- SM_35 for the time being.
 #ifndef __CUDA_ARCH__
 #define __CUDA_ARCH__ 350
 #endif
 
 #include "__clang_cuda_builtin_vars.h"
 
 // No need for device_launch_parameters.h as __clang_cuda_builtin_vars.h above
 // has taken care of builtin variables declared in the file.
 #define __DEVICE_LAUNCH_PARAMETERS_H__
 
 // {math,device}_functions.h only have declarations of the
 // functions. We don't need them as we're going to pull in their
 // definitions from .hpp files.
 #define __DEVICE_FUNCTIONS_H__
 #define __MATH_FUNCTIONS_H__
 #define __COMMON_FUNCTIONS_H__
+// device_functions_decls is replaced by __clang_cuda_device_functions.h
+// included below.
+#define __DEVICE_FUNCTIONS_DECLS_H__
 
 #undef __CUDACC__
 #if CUDA_VERSION < 9000
 #define __CUDABE__
 #else
 #define __CUDA_LIBDEVICE__
 #endif
 // Disables definitions of device-side runtime support stubs in

 // ends up not being there because of the games we play here.  Just define it
 // ourselves; it's simple enough.
 #ifdef __APPLE__
 inline __host__ double __signbitd(double x) {
   return std::signbit(x);
 }
 #endif
 
-// We need decls for functions in CUDA's libdevice with __device__
-// attribute only. Alas they come either as __host__ __device__ or
-// with no attributes at all. To work around that, define __CUDA_RTC__
-// which produces HD variant and undef __host__ which gives us desided
-// decls with __device__ attribute.
-#pragma push_macro("__host__")
-#define __host__
-#define __CUDACC_RTC__
-#include "device_functions_decls.h"
-#undef __CUDACC_RTC__
-
-// Temporarily poison __host__ macro to ensure it's not used by any of
-// the headers we're about to include.
-#define __host__ UNEXPECTED_HOST_ATTRIBUTE
+// Declare or define device-side functions that particular CUDA version relies
+// on but does (no longer) declare in its headers. E.g. some of the device-side
+// functions that used to be implemented in a header in CUDA-8, became NVCC's
+// builtins and recent CUDA versions. Another example would be declarations for
+// the functions provided by libdevice bitcode. The declarations used to be in
+// device_functions_decls.h but are no longer provided starting with CUDA 9.1.
+#include <__clang_cuda_device_functions.h>
+
+// __THROW is redefined to be empty by device_functions_decls.h in CUDA. Clang's
+// counterpart does not do it, so we need to make it empty here to keep
+// following CUDA includes happy.
+#undef __THROW
+#define __THROW
 
 // CUDA 8.0.41 relies on __USE_FAST_MATH__ and __CUDA_PREC_DIV's values.
 // Previous versions used to check whether they are defined or not.
 // CU_DEVICE_INVALID macro is only defined in 8.0.41, so we use it
 // here to detect the switch.
 
 #if defined(CU_DEVICE_INVALID)
 #if !defined(__USE_FAST_MATH__)
 #define __USE_FAST_MATH__ 0
 #endif
 
 #if !defined(__CUDA_PREC_DIV)
 #define __CUDA_PREC_DIV 0
 #endif
 #endif
 
+// Temporarily poison __host__ macro to ensure it's not used by any of
+// the headers we're about to include.
+#pragma push_macro("__host__")
+#define __host__ UNEXPECTED_HOST_ATTRIBUTE
+
 // device_functions.hpp and math_functions*.hpp use 'static
 // __forceinline__' (with no __device__) for definitions of device
 // functions. Temporarily redefine __forceinline__ to include
 // __device__.
 #pragma push_macro("__forceinline__")
 #define __forceinline__ __device__ __inline__ __attribute__((always_inline))
-
-#pragma push_macro("__float2half_rn")
-#if CUDA_VERSION >= 9000
-// CUDA-9 has conflicting prototypes for __float2half_rn(float f) in
-// cuda_fp16.h[pp] and device_functions.hpp. We need to get the one in
-// device_functions.hpp out of the way.
-#define __float2half_rn  __float2half_rn_disabled
-#endif
-
+#if CUDA_VERSION < 9000
 #include "device_functions.hpp"
-#pragma pop_macro("__float2half_rn")
-
+#endif
 
 // math_function.hpp uses the __USE_FAST_MATH__ macro to determine whether we
 // get the slow-but-accurate or fast-but-inaccurate versions of functions like
 // sin and exp.  This is controlled in clang by -fcuda-approx-transcendentals.
 //
 // device_functions.hpp uses __USE_FAST_MATH__ for a different purpose (fast vs.
 // slow divides), so we need to scope our define carefully here.
 #pragma push_macro("__USE_FAST_MATH__")
 #if defined(__CLANG_CUDA_APPROX_TRANSCENDENTALS__)
 #define __USE_FAST_MATH__ 1
 #endif
+
+#if CUDA_VERSION >= 9000
+#include "crt/math_functions.hpp"
+#else
 #include "math_functions.hpp"
+#endif
+
 #pragma pop_macro("__USE_FAST_MATH__")
 
+#if CUDA_VERSION < 9000
 #include "math_functions_dbl_ptx3.hpp"
+#endif
 #pragma pop_macro("__forceinline__")
 
 // Pull in host-only functions that are only available when neither
 // __CUDACC__ nor __CUDABE__ are defined.
 #undef __MATH_FUNCTIONS_HPP__
 #undef __CUDABE__
+#if CUDA_VERSION < 9000
 #include "math_functions.hpp"
+#endif
 // Alas, additional overloads for these functions are hard to get to.
 // Considering that we only need these overloads for a few functions,
 // we can provide them here.
 static inline float rsqrt(float __a) { return rsqrtf(__a); }
 static inline float rcbrt(float __a) { return rcbrtf(__a); }
 static inline float sinpi(float __a) { return sinpif(__a); }
 static inline float cospi(float __a) { return cospif(__a); }
 static inline void sincospi(float __a, float *__b, float *__c) {
   return sincospif(__a, __b, __c);
 }
 static inline float erfcinv(float __a) { return erfcinvf(__a); }
 static inline float normcdfinv(float __a) { return normcdfinvf(__a); }
 static inline float normcdf(float __a) { return normcdff(__a); }
 static inline float erfcx(float __a) { return erfcxf(__a); }
 
+#if CUDA_VERSION < 9000
 // For some reason single-argument variant is not always declared by
 // CUDA headers. Alas, device_functions.hpp included below needs it.
 static inline __device__ void __brkpt(int __c) { __brkpt(); }
+#endif
 
 // Now include *.hpp with definitions of various GPU functions.  Alas,
 // a lot of thins get declared/defined with __host__ attribute which
 // we don't want and we have to define it out. We also have to include
 // {device,math}_functions.hpp again in order to extract the other
 // branch of #if/else inside.
-
 #define __host__
 #undef __CUDABE__
 #define __CUDACC__
+#if CUDA_VERSION >= 9000
+// Some atomic functions became compiler builtins in CUDA-9 , so we need their
+// declarations.
+#include "device_atomic_functions.h"
+#endif
 #undef __DEVICE_FUNCTIONS_HPP__
 #include "device_atomic_functions.hpp"
+#if CUDA_VERSION >= 9000
+#include "crt/device_functions.hpp"
+#include "crt/device_double_functions.hpp"
+#else
 #include "device_functions.hpp"
+#define __CUDABE__
+#include "device_double_functions.h"
+#undef __CUDABE__
+#endif
 #include "sm_20_atomic_functions.hpp"
 #include "sm_20_intrinsics.hpp"
 #include "sm_32_atomic_functions.hpp"
 
 // Don't include sm_30_intrinsics.h and sm_32_intrinsics.h.  These define the
 // __shfl and __ldg intrinsics using inline (volatile) asm, but we want to
 // define them using builtins so that the optimizer can reason about and across
 // these instructions.  In particular, using intrinsics for ldg gets us the
 // [addr+imm] addressing mode, which, although it doesn't actually exist in the
 // hardware, seems to generate faster machine code because ptxas can more easily
 // reason about our code.
 
 #if CUDA_VERSION >= 8000
+#pragma push_macro("__CUDA_ARCH__")
+#undef __CUDA_ARCH__
 #include "sm_60_atomic_functions.hpp"
 #include "sm_61_intrinsics.hpp"
+#pragma pop_macro("__CUDA_ARCH__")
 #endif
 
 #undef __MATH_FUNCTIONS_HPP__
 
 // math_functions.hpp defines ::signbit as a __host__ __device__ function.  This
 // conflicts with libstdc++'s constexpr ::signbit, so we have to rename
 // math_function.hpp's ::signbit.  It's guarded by #undef signbit, but that's
 // conditional on __GNUC__.  :)

 #if CUDA_VERSION >= 9000
 #undef _GLIBCXX_MATH_H
 // We also need to undo another guard that checks for libc++ 3.8+
 #ifdef _LIBCPP_VERSION
 #define _LIBCPP_VERSION 3700
 #endif
 #endif
 
+#if CUDA_VERSION >= 9000
+#include "crt/math_functions.hpp"
+#else
 #include "math_functions.hpp"
+#endif
 #pragma pop_macro("_GLIBCXX_MATH_H")
 #pragma pop_macro("_LIBCPP_VERSION")
 #pragma pop_macro("__GNUC__")
 #pragma pop_macro("signbit")
 
 #pragma pop_macro("__host__")
 
 #include "texture_indirect_functions.h"