[CUDA] Implemented device-side support for functions in <cmath>.

[tra]

Summary

CUDA expects math functions in std:: namespace to work on device side.
In order to make it work with clang without allowing device-side code
generation for functions w/o appropriate target attributes, this patch
provides device-side implementations for <cmath> functions. Most of
them call global-scope math functions provided by CUDA headers. In few
cases we use clang builtins.

Tested out-of tree by compiling and running thrust's unit_tests.
https://github.com/thrust/thrust/tree/master/testing

[jlebar]

Missing (?) functions:

• div, ldiv, lldiv, imaxdiv
• imaxabs

If you left these out intentionally (I forget if nvidia supports div_t), that's probably fine, but maybe add a comment?

wrt the "::" comments, some are nits because I think we end up calling the right thing, assuming that nobody defines the wrong thing in namespace std. But some aren't nits because afaict they lead to infinite recursion. We should probably just give fully-qualified names for all functions we call.

lib/Headers/__clang_cuda_cmath.h

33

Why is std::abs a special case that needs to be called out here?

38

We should probably respect the standard and not define functions that aren't available in C++11 if we're not in c++11 mode.

93

Looking through bugzilla, it appears that this builtin may try to invoke library functions, which may or may not exist in our case (e.g. extern "C" fabs vs ::fabs). It's probably worth checking this one in particular to make sure it works.

99

Nit, ::

100

Need :: to avoid infinite recursion, right?

105

Where's finitef coming from? Same for the other functions used here.

105

Nit: "::" in front of all the __ functions.

146

Nit: ::ldexpf?

147

:: not optional on this one, right?

[tra]

Added missing ::

[tra]

Fixed few issues revealed by -Wdouble-promotion

[tra]

lib/Headers/__clang_cuda_cmath.h

33

It's a bit of an oddball because it operated on integer parameters and came from <cstdlib>.
Which is indeed irrelevant here.
Deleted.

40

Both libc++ and libstdc++ provide these symbols in pre-c++11 mode, so should we.

93

Checked -- as far as I can tell it generates valid PTX to test its arguments.

99

Fixed.

105

finitef and other non-builtin * functions come from CUDA headers. Added missing ::.

lib/Headers/CMakeLists.txt

set(files
   avx512vlbwintrin.h
   avx512vlintrin.h
   avx512dqintrin.h
   avx512vldqintrin.h
   pkuintrin.h
   avxintrin.h
   bmi2intrin.h
   bmiintrin.h
+  __clang_cuda_cmath.h
   __clang_cuda_runtime_wrapper.h
   cpuid.h
   cuda_builtin_vars.h
   emmintrin.h
   f16cintrin.h
   float.h
   fma4intrin.h
   fmaintrin.h

lib/Headers/__clang_cuda_cmath.h

+/*===---- __clang_cuda_cmath.h - Device-side CUDA cmath 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_CMATH_H__
+#define __CLANG_CUDA_CMATH_H__
+#ifndef __CUDA__
+#error "This file is for CUDA compilation only."
+#endif
+
+// CUDA allows using math functions form std:: on device side.  This
+// file provides __device__ overloads for math functions that map to
+// appropriate math functions provided by CUDA headers or to compiler
+// builtins if CUDA does not provide a suitable function.
+// We also provide device-side std::abs() for integer types.
+
+#define __DEVICE__ static __device__ __inline__ __attribute__((always_inline))
+
+namespace std {
+__DEVICE__ long long abs(long long n) { return ::llabs(n); }
+__DEVICE__ long abs(long n) { return ::labs(n); }
+__DEVICE__ int abs(int n) { return ::abs(n); }
+__DEVICE__ float abs(float x) { return ::fabsf(x); }
+__DEVICE__ double abs(double x) { return ::fabs(x); }
+__DEVICE__ float acos(float x) { return ::acosf(x); }
+__DEVICE__ double acos(double x) { return ::acos(x); }
+__DEVICE__ float acosh(float x) { return ::acoshf(x); }
+__DEVICE__ double acosh(double x) { return ::acosh(x); }
+__DEVICE__ float asin(float x) { return ::asinf(x); }
+__DEVICE__ double asin(double x) { return ::asin(x); }
+__DEVICE__ float asinh(float x) { return ::asinhf(x); }
+__DEVICE__ double asinh(double x) { return ::asinh(x); }
+__DEVICE__ float atan(float x) { return ::atanf(x); }
+__DEVICE__ double atan(double x) { return ::atan(x); }
+__DEVICE__ float atan2(float x, float y) { return ::atan2f(x, y); }
+__DEVICE__ double atan2(double x, double y) { return ::atan2(x, y); }
+__DEVICE__ float atanh(float x) { return ::atanhf(x); }
+__DEVICE__ double atanh(double x) { return ::atanh(x); }
+__DEVICE__ float cbrt(float x) { return ::cbrtf(x); }
+__DEVICE__ double cbrt(double x) { return ::cbrt(x); }
+__DEVICE__ float ceil(float x) { return ::ceilf(x); }
+__DEVICE__ double ceil(double x) { return ::ceil(x); }
+__DEVICE__ float copysign(float x, float y) { return ::copysignf(x, y); }
+__DEVICE__ double copysign(double x, double y) { return ::copysign(x, y); }
+__DEVICE__ float cos(float x) { return ::cosf(x); }
+__DEVICE__ double cos(double x) { return ::cos(x); }
+__DEVICE__ float cosh(float x) { return ::coshf(x); }
+__DEVICE__ double cosh(double x) { return ::cosh(x); }
+__DEVICE__ float erf(float x) { return ::erff(x); }
+__DEVICE__ double erf(double x) { return ::erf(x); }
+__DEVICE__ float erfc(float x) { return ::erfcf(x); }
+__DEVICE__ double erfc(double x) { return ::erfc(x); }
+__DEVICE__ float exp(float x) { return ::expf(x); }
+__DEVICE__ double exp(double x) { return ::exp(x); }
+__DEVICE__ float exp2(float x) { return ::exp2f(x); }
+__DEVICE__ double exp2(double x) { return ::exp2(x); }
+__DEVICE__ float expm1(float x) { return ::expm1f(x); }
+__DEVICE__ double expm1(double x) { return ::expm1(x); }
+__DEVICE__ float fabs(float x) { return ::fabsf(x); }
+__DEVICE__ double fabs(double x) { return ::fabs(x); }
+__DEVICE__ float fdim(float x, float y) { return ::fdimf(x, y); }
+__DEVICE__ double fdim(double x, double y) { return ::fdim(x, y); }
+__DEVICE__ float floor(float x) { return ::floorf(x); }
+__DEVICE__ double floor(double x) { return ::floor(x); }
+__DEVICE__ float fma(float x, float y, float z) { return ::fmaf(x, y, z); }
+__DEVICE__ double fma(double x, double y, double z) { return ::fma(x, y, z); }
+__DEVICE__ float fmax(float x, float y) { return ::fmaxf(x, y); }
+__DEVICE__ double fmax(double x, double y) { return ::fmax(x, y); }
+__DEVICE__ float fmin(float x, float y) { return ::fminf(x, y); }
+__DEVICE__ double fmin(double x, double y) { return ::fmin(x, y); }
+__DEVICE__ float fmod(float x, float y) { return ::fmodf(x, y); }
+__DEVICE__ double fmod(double x, double y) { return ::fmod(x, y); }
+__DEVICE__ int fpclassify(float x) {
+  return __builtin_fpclassify(FP_NAN, FP_INFINITE, FP_NORMAL, FP_SUBNORMAL,
+                              FP_ZERO, x);
+}
+__DEVICE__ int fpclassify(double x) {
+  return __builtin_fpclassify(FP_NAN, FP_INFINITE, FP_NORMAL, FP_SUBNORMAL,
+                              FP_ZERO, x);
+}
+__DEVICE__ float frexp(float arg, int *exp) { return frexpf(arg, exp); }
+__DEVICE__ double frexp(double arg, int *exp) { return frexp(arg, exp); }
+__DEVICE__ float hypot(float x, float y) { return ::hypotf(x, y); }
+__DEVICE__ double hypot(double x, double y) { return ::hypot(x, y); }
+__DEVICE__ int ilogb(float arg) { return ::ilogbf(arg); }
+__DEVICE__ int ilogb(double arg) { return ::ilogb(arg); }
+__DEVICE__ bool isfinite(float x) { return __finitef(x); }
+__DEVICE__ bool isfinite(double x) { return __finite(x); }
+__DEVICE__ bool isgreater(float x, float y) {
+  return __builtin_isgreater(x, y);
+}
+__DEVICE__ bool isgreater(double x, double y) {
+  return __builtin_isgreater(x, y);
+}
+__DEVICE__ bool isgreaterequal(float x, float y) {
+  return __builtin_isgreaterequal(x, y);
+}
+__DEVICE__ bool isgreaterequal(double x, double y) {
+  return __builtin_isgreaterequal(x, y);
+}
+__DEVICE__ bool isinf(float x) { return __isinff(x); }
+__DEVICE__ bool isinf(double x) { return __isinf(x); }
+__DEVICE__ bool isless(float x, float y) { return __builtin_isless(x, y); }
+__DEVICE__ bool isless(double x, double y) { return __builtin_isless(x, y); }
+__DEVICE__ bool islessequal(float x, float y) {
+  return __builtin_islessequal(x, y);
+}
+__DEVICE__ bool islessequal(double x, double y) {
+  return __builtin_islessequal(x, y);
+}
+__DEVICE__ bool islessgreater(float x, float y) {
+  return __builtin_islessgreater(x, y);
+}
+__DEVICE__ bool islessgreater(double x, double y) {
+  return __builtin_islessgreater(x, y);
+}
+__DEVICE__ bool isnan(float x) { return __isnanf(x); }
+__DEVICE__ bool isnan(double x) { return __isnan(x); }
+__DEVICE__ bool isnormal(float x) { return __builtin_isnormal(x); }
+__DEVICE__ bool isnormal(double x) { return __builtin_isnormal(x); }
+__DEVICE__ bool isunordered(float x, float y) {
+  return __builtin_isunordered(x, y);
+}
+__DEVICE__ bool isunordered(double x, double y) {
+  return __builtin_isunordered(x, y);
+}
+__DEVICE__ long labs(long n) { return ::labs(n); }
+__DEVICE__ float ldexp(float arg, int exp) { return ldexpf(arg, exp); }
+__DEVICE__ double ldexp(double arg, int exp) { return ldexp(arg, exp); }
+__DEVICE__ float lgamma(float x) { return ::lgammaf(x); }
+__DEVICE__ double lgamma(double x) { return ::lgamma(x); }
+__DEVICE__ long long llabs(long long n) { return ::llabs(n); }
+__DEVICE__ long long llrint(float x) { return ::llrintf(x); }
+__DEVICE__ long long llrint(double x) { return ::llrint(x); }
+__DEVICE__ float log(float x) { return ::logf(x); }
+__DEVICE__ double log(double x) { return ::log(x); }
+__DEVICE__ float log10(float x) { return ::log10f(x); }
+__DEVICE__ double log10(double x) { return ::log10(x); }
+__DEVICE__ float log1p(float x) { return ::log1pf(x); }
+__DEVICE__ double log1p(double x) { return ::log1p(x); }
+__DEVICE__ float log2(float x) { return ::log2f(x); }
+__DEVICE__ double log2(double x) { return ::log2(x); }
+__DEVICE__ float logb(float x) { return ::logbf(x); }
+__DEVICE__ double logb(double x) { return ::logb(x); }
+__DEVICE__ long lrint(float x) { return ::lrintf(x); }
+__DEVICE__ long lrint(double x) { return ::lrint(x); }
+__DEVICE__ long lround(float x) { return ::lroundf(x); }
+__DEVICE__ long lround(double x) { return ::lround(x); }
+__DEVICE__ float modf(float x, float *iptr) { return ::modff(x, iptr); }
+__DEVICE__ double modf(double x, double *iptr) { return ::modf(x, iptr); }
+__DEVICE__ double nan(const char *x) { return ::nan(x); }
+__DEVICE__ float nanf(const char *x) { return ::nanf(x); }
+__DEVICE__ float nearbyint(float x) { return ::nearbyintf(x); }
+__DEVICE__ double nearbyint(double x) { return ::nearbyint(x); }
+__DEVICE__ float nextafter(float from, float to) {
+  return ::nextafterf(from, to);
+}
+__DEVICE__ double nextafter(double from, double to) {
+  return ::nextafter(from, to);
+}
+__DEVICE__ float nexttoward(float from, float to) {
+  return ::nexttowardf(from, to);
+}
+__DEVICE__ double nexttoward(double from, double to) {
+  return ::nexttoward(from, to);
+}
+__DEVICE__ float pow(float base, float exp) { return ::powf(base, exp); }
+__DEVICE__ float pow(float base, int iexp) { return ::powif(base, iexp); }
+__DEVICE__ double pow(double base, double exp) { return ::pow(base, exp); }
+__DEVICE__ double pow(double base, int iexp) { return ::powi(base, iexp); }
+__DEVICE__ float remainder(float x, float y) { return ::remainderf(x, y); }
+__DEVICE__ double remainder(double x, double y) { return ::remainder(x, y); }
+__DEVICE__ float remquo(float x, float y, int *quo) {
+  return ::remquof(x, y, quo);
+}
+__DEVICE__ double remquo(double x, double y, int *quo) {
+  return ::remquo(x, y, quo);
+}
+__DEVICE__ float rint(float x) { return ::rintf(x); }
+__DEVICE__ double rint(double x) { return ::rint(x); }
+__DEVICE__ float round(float x) { return ::roundf(x); }
+__DEVICE__ double round(double x) { return ::round(x); }
+__DEVICE__ float scalbln(float x, long exp) { return ::scalblnf(x, exp); }
+__DEVICE__ double scalbln(double x, long exp) { return ::scalbln(x, exp); }
+__DEVICE__ float scalbn(float x, int exp) { return ::scalbnf(x, exp); }
+__DEVICE__ double scalbn(double x, int exp) { return ::scalbn(x, exp); }
+__DEVICE__ bool signbit(float x) { return __signbitf(x); }
+__DEVICE__ bool signbit(double x) { return __signbit(x); }
+__DEVICE__ float sin(float x) { return ::sinf(x); }
+__DEVICE__ double sin(double x) { return ::sin(x); }
+__DEVICE__ float sinh(float x) { return ::sinhf(x); }
+__DEVICE__ double sinh(double x) { return ::sinh(x); }
+__DEVICE__ float sqrt(float x) { return ::sqrtf(x); }
+__DEVICE__ double sqrt(double x) { return ::sqrt(x); }
+__DEVICE__ float tan(float x) { return ::tanf(x); }
+__DEVICE__ double tan(double x) { return ::tan(x); }
+__DEVICE__ float tanh(float x) { return ::tanhf(x); }
+__DEVICE__ double tanh(double x) { return ::tanh(x); }
+__DEVICE__ float tgamma(float x) { return ::tgammaf(x); }
+__DEVICE__ double tgamma(double x) { return ::tgamma(x); }
+__DEVICE__ float trunc(float x) { return ::truncf(x); }
+__DEVICE__ double trunc(double x) { return ::trunc(x); }
+
+} // namespace std
+
+#endif

lib/Headers/__clang_cuda_runtime_wrapper.h

 #define __CLANG_CUDA_RUNTIME_WRAPPER_H__
 
 #if defined(__CUDA__) && defined(__clang__)
 
 // Include some standard headers to avoid CUDA headers including them
 // while some required macros (like __THROW) are in a weird state.
 #include <stdlib.h>
 #include <cmath>
+#include <cstdlib>
 
 // Preserve common macros that will be changed below by us or by CUDA
 // headers.
 #pragma push_macro("__THROW")
 #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

 }
 
 // The nvptx vprintf syscall.  This doesn't actually need to be declared, but we
 // declare it so that if someone else declares it with a different signature,
 // we'll throw an error.
 extern "C" __device__ int vprintf(const char*, const char*);
 #endif
 
+#include <__clang_cuda_cmath.h>
+
 #endif // __CUDA__
 #endif // __CLANG_CUDA_RUNTIME_WRAPPER_H__