__clang_cuda_runtime_wrapper.h revision 309124
1292920Sdim/*===---- __clang_cuda_runtime_wrapper.h - CUDA runtime support -------------=== 2292920Sdim * 3292920Sdim * Permission is hereby granted, free of charge, to any person obtaining a copy 4292920Sdim * of this software and associated documentation files (the "Software"), to deal 5292920Sdim * in the Software without restriction, including without limitation the rights 6292920Sdim * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 7292920Sdim * copies of the Software, and to permit persons to whom the Software is 8292920Sdim * furnished to do so, subject to the following conditions: 9292920Sdim * 10292920Sdim * The above copyright notice and this permission notice shall be included in 11292920Sdim * all copies or substantial portions of the Software. 12292920Sdim * 13292920Sdim * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 14292920Sdim * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 15292920Sdim * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 16292920Sdim * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 17292920Sdim * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 18292920Sdim * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 19292920Sdim * THE SOFTWARE. 20292920Sdim * 21292920Sdim *===-----------------------------------------------------------------------=== 22292920Sdim */ 23292920Sdim 24292920Sdim/* 25292920Sdim * WARNING: This header is intended to be directly -include'd by 26292920Sdim * the compiler and is not supposed to be included by users. 27292920Sdim * 28292920Sdim * CUDA headers are implemented in a way that currently makes it 29292920Sdim * impossible for user code to #include directly when compiling with 30292920Sdim * Clang. They present different view of CUDA-supplied functions 31292920Sdim * depending on where in NVCC's compilation pipeline the headers are 32292920Sdim * included. Neither of these modes provides function definitions with 33292920Sdim * correct attributes, so we use preprocessor to force the headers 34292920Sdim * into a form that Clang can use. 35292920Sdim * 36292920Sdim * Similarly to NVCC which -include's cuda_runtime.h, Clang -include's 37292920Sdim * this file during every CUDA compilation. 38292920Sdim */ 39292920Sdim 40292920Sdim#ifndef __CLANG_CUDA_RUNTIME_WRAPPER_H__ 41292920Sdim#define __CLANG_CUDA_RUNTIME_WRAPPER_H__ 42292920Sdim 43292920Sdim#if defined(__CUDA__) && defined(__clang__) 44292920Sdim 45309124Sdim// Include some forward declares that must come before cmath. 46309124Sdim#include <__clang_cuda_math_forward_declares.h> 47309124Sdim 48292920Sdim// Include some standard headers to avoid CUDA headers including them 49292920Sdim// while some required macros (like __THROW) are in a weird state. 50309124Sdim#include <cmath> 51309124Sdim#include <cstdlib> 52292920Sdim#include <stdlib.h> 53292920Sdim 54292920Sdim// Preserve common macros that will be changed below by us or by CUDA 55292920Sdim// headers. 56292920Sdim#pragma push_macro("__THROW") 57292920Sdim#pragma push_macro("__CUDA_ARCH__") 58292920Sdim 59292920Sdim// WARNING: Preprocessor hacks below are based on specific details of 60292920Sdim// CUDA-7.x headers and are not expected to work with any other 61292920Sdim// version of CUDA headers. 62292920Sdim#include "cuda.h" 63292920Sdim#if !defined(CUDA_VERSION) 64292920Sdim#error "cuda.h did not define CUDA_VERSION" 65292920Sdim#elif CUDA_VERSION < 7000 || CUDA_VERSION > 7050 66292920Sdim#error "Unsupported CUDA version!" 67292920Sdim#endif 68292920Sdim 69292920Sdim// Make largest subset of device functions available during host 70292920Sdim// compilation -- SM_35 for the time being. 71292920Sdim#ifndef __CUDA_ARCH__ 72292920Sdim#define __CUDA_ARCH__ 350 73292920Sdim#endif 74292920Sdim 75292920Sdim#include "cuda_builtin_vars.h" 76292920Sdim 77292920Sdim// No need for device_launch_parameters.h as cuda_builtin_vars.h above 78292920Sdim// has taken care of builtin variables declared in the file. 79292920Sdim#define __DEVICE_LAUNCH_PARAMETERS_H__ 80292920Sdim 81292920Sdim// {math,device}_functions.h only have declarations of the 82292920Sdim// functions. We don't need them as we're going to pull in their 83292920Sdim// definitions from .hpp files. 84292920Sdim#define __DEVICE_FUNCTIONS_H__ 85292920Sdim#define __MATH_FUNCTIONS_H__ 86309124Sdim#define __COMMON_FUNCTIONS_H__ 87292920Sdim 88292920Sdim#undef __CUDACC__ 89292920Sdim#define __CUDABE__ 90292920Sdim// Disables definitions of device-side runtime support stubs in 91292920Sdim// cuda_device_runtime_api.h 92309124Sdim#include "driver_types.h" 93292920Sdim#include "host_config.h" 94292920Sdim#include "host_defines.h" 95292920Sdim 96292920Sdim#undef __CUDABE__ 97292920Sdim#define __CUDACC__ 98292920Sdim#include "cuda_runtime.h" 99292920Sdim 100292920Sdim#undef __CUDACC__ 101292920Sdim#define __CUDABE__ 102292920Sdim 103292920Sdim// CUDA headers use __nvvm_memcpy and __nvvm_memset which Clang does 104292920Sdim// not have at the moment. Emulate them with a builtin memcpy/memset. 105309124Sdim#define __nvvm_memcpy(s, d, n, a) __builtin_memcpy(s, d, n) 106309124Sdim#define __nvvm_memset(d, c, n, a) __builtin_memset(d, c, n) 107292920Sdim 108309124Sdim#include "crt/device_runtime.h" 109292920Sdim#include "crt/host_runtime.h" 110292920Sdim// device_runtime.h defines __cxa_* macros that will conflict with 111292920Sdim// cxxabi.h. 112292920Sdim// FIXME: redefine these as __device__ functions. 113292920Sdim#undef __cxa_vec_ctor 114292920Sdim#undef __cxa_vec_cctor 115292920Sdim#undef __cxa_vec_dtor 116292920Sdim#undef __cxa_vec_new2 117292920Sdim#undef __cxa_vec_new3 118292920Sdim#undef __cxa_vec_delete2 119292920Sdim#undef __cxa_vec_delete 120292920Sdim#undef __cxa_vec_delete3 121292920Sdim#undef __cxa_pure_virtual 122292920Sdim 123292920Sdim// We need decls for functions in CUDA's libdevice with __device__ 124292920Sdim// attribute only. Alas they come either as __host__ __device__ or 125292920Sdim// with no attributes at all. To work around that, define __CUDA_RTC__ 126292920Sdim// which produces HD variant and undef __host__ which gives us desided 127292920Sdim// decls with __device__ attribute. 128292920Sdim#pragma push_macro("__host__") 129292920Sdim#define __host__ 130292920Sdim#define __CUDACC_RTC__ 131292920Sdim#include "device_functions_decls.h" 132292920Sdim#undef __CUDACC_RTC__ 133292920Sdim 134292920Sdim// Temporarily poison __host__ macro to ensure it's not used by any of 135292920Sdim// the headers we're about to include. 136292920Sdim#define __host__ UNEXPECTED_HOST_ATTRIBUTE 137292920Sdim 138292920Sdim// device_functions.hpp and math_functions*.hpp use 'static 139292920Sdim// __forceinline__' (with no __device__) for definitions of device 140292920Sdim// functions. Temporarily redefine __forceinline__ to include 141292920Sdim// __device__. 142292920Sdim#pragma push_macro("__forceinline__") 143292920Sdim#define __forceinline__ __device__ __inline__ __attribute__((always_inline)) 144292920Sdim#include "device_functions.hpp" 145309124Sdim 146309124Sdim// math_function.hpp uses the __USE_FAST_MATH__ macro to determine whether we 147309124Sdim// get the slow-but-accurate or fast-but-inaccurate versions of functions like 148309124Sdim// sin and exp. This is controlled in clang by -fcuda-approx-transcendentals. 149309124Sdim// 150309124Sdim// device_functions.hpp uses __USE_FAST_MATH__ for a different purpose (fast vs. 151309124Sdim// slow divides), so we need to scope our define carefully here. 152309124Sdim#pragma push_macro("__USE_FAST_MATH__") 153309124Sdim#if defined(__CLANG_CUDA_APPROX_TRANSCENDENTALS__) 154309124Sdim#define __USE_FAST_MATH__ 155309124Sdim#endif 156292920Sdim#include "math_functions.hpp" 157309124Sdim#pragma pop_macro("__USE_FAST_MATH__") 158309124Sdim 159292920Sdim#include "math_functions_dbl_ptx3.hpp" 160292920Sdim#pragma pop_macro("__forceinline__") 161292920Sdim 162292920Sdim// Pull in host-only functions that are only available when neither 163292920Sdim// __CUDACC__ nor __CUDABE__ are defined. 164292920Sdim#undef __MATH_FUNCTIONS_HPP__ 165292920Sdim#undef __CUDABE__ 166292920Sdim#include "math_functions.hpp" 167292920Sdim// Alas, additional overloads for these functions are hard to get to. 168292920Sdim// Considering that we only need these overloads for a few functions, 169292920Sdim// we can provide them here. 170309124Sdimstatic inline float rsqrt(float __a) { return rsqrtf(__a); } 171309124Sdimstatic inline float rcbrt(float __a) { return rcbrtf(__a); } 172309124Sdimstatic inline float sinpi(float __a) { return sinpif(__a); } 173309124Sdimstatic inline float cospi(float __a) { return cospif(__a); } 174309124Sdimstatic inline void sincospi(float __a, float *__b, float *__c) { 175309124Sdim return sincospif(__a, __b, __c); 176292920Sdim} 177309124Sdimstatic inline float erfcinv(float __a) { return erfcinvf(__a); } 178309124Sdimstatic inline float normcdfinv(float __a) { return normcdfinvf(__a); } 179309124Sdimstatic inline float normcdf(float __a) { return normcdff(__a); } 180309124Sdimstatic inline float erfcx(float __a) { return erfcxf(__a); } 181292920Sdim 182292920Sdim// For some reason single-argument variant is not always declared by 183292920Sdim// CUDA headers. Alas, device_functions.hpp included below needs it. 184309124Sdimstatic inline __device__ void __brkpt(int __c) { __brkpt(); } 185292920Sdim 186292920Sdim// Now include *.hpp with definitions of various GPU functions. Alas, 187292920Sdim// a lot of thins get declared/defined with __host__ attribute which 188292920Sdim// we don't want and we have to define it out. We also have to include 189292920Sdim// {device,math}_functions.hpp again in order to extract the other 190292920Sdim// branch of #if/else inside. 191292920Sdim 192292920Sdim#define __host__ 193292920Sdim#undef __CUDABE__ 194292920Sdim#define __CUDACC__ 195292920Sdim#undef __DEVICE_FUNCTIONS_HPP__ 196309124Sdim#include "device_atomic_functions.hpp" 197292920Sdim#include "device_functions.hpp" 198292920Sdim#include "sm_20_atomic_functions.hpp" 199309124Sdim#include "sm_20_intrinsics.hpp" 200292920Sdim#include "sm_32_atomic_functions.hpp" 201309124Sdim 202309124Sdim// Don't include sm_30_intrinsics.h and sm_32_intrinsics.h. These define the 203309124Sdim// __shfl and __ldg intrinsics using inline (volatile) asm, but we want to 204309124Sdim// define them using builtins so that the optimizer can reason about and across 205309124Sdim// these instructions. In particular, using intrinsics for ldg gets us the 206309124Sdim// [addr+imm] addressing mode, which, although it doesn't actually exist in the 207309124Sdim// hardware, seems to generate faster machine code because ptxas can more easily 208309124Sdim// reason about our code. 209309124Sdim 210292920Sdim#undef __MATH_FUNCTIONS_HPP__ 211309124Sdim 212309124Sdim// math_functions.hpp defines ::signbit as a __host__ __device__ function. This 213309124Sdim// conflicts with libstdc++'s constexpr ::signbit, so we have to rename 214309124Sdim// math_function.hpp's ::signbit. It's guarded by #undef signbit, but that's 215309124Sdim// conditional on __GNUC__. :) 216309124Sdim#pragma push_macro("signbit") 217309124Sdim#pragma push_macro("__GNUC__") 218309124Sdim#undef __GNUC__ 219309124Sdim#define signbit __ignored_cuda_signbit 220292920Sdim#include "math_functions.hpp" 221309124Sdim#pragma pop_macro("__GNUC__") 222309124Sdim#pragma pop_macro("signbit") 223309124Sdim 224292920Sdim#pragma pop_macro("__host__") 225292920Sdim 226292920Sdim#include "texture_indirect_functions.h" 227292920Sdim 228292920Sdim// Restore state of __CUDA_ARCH__ and __THROW we had on entry. 229292920Sdim#pragma pop_macro("__CUDA_ARCH__") 230292920Sdim#pragma pop_macro("__THROW") 231292920Sdim 232292920Sdim// Set up compiler macros expected to be seen during compilation. 233292920Sdim#undef __CUDABE__ 234292920Sdim#define __CUDACC__ 235292920Sdim 236309124Sdimextern "C" { 237309124Sdim// Device-side CUDA system calls. 238309124Sdim// http://docs.nvidia.com/cuda/ptx-writers-guide-to-interoperability/index.html#system-calls 239309124Sdim// We need these declarations and wrappers for device-side 240309124Sdim// malloc/free/printf calls to work without relying on 241309124Sdim// -fcuda-disable-target-call-checks option. 242309124Sdim__device__ int vprintf(const char *, const char *); 243309124Sdim__device__ void free(void *) __attribute((nothrow)); 244309124Sdim__device__ void *malloc(size_t) __attribute((nothrow)) __attribute__((malloc)); 245309124Sdim__device__ void __assertfail(const char *__message, const char *__file, 246309124Sdim unsigned __line, const char *__function, 247309124Sdim size_t __charSize) __attribute__((noreturn)); 248309124Sdim 249309124Sdim// In order for standard assert() macro on linux to work we need to 250309124Sdim// provide device-side __assert_fail() 251309124Sdim__device__ static inline void __assert_fail(const char *__message, 252309124Sdim const char *__file, unsigned __line, 253309124Sdim const char *__function) { 254309124Sdim __assertfail(__message, __file, __line, __function, sizeof(char)); 255292920Sdim} 256292920Sdim 257309124Sdim// Clang will convert printf into vprintf, but we still need 258309124Sdim// device-side declaration for it. 259309124Sdim__device__ int printf(const char *, ...); 260309124Sdim} // extern "C" 261309124Sdim 262309124Sdim// We also need device-side std::malloc and std::free. 263309124Sdimnamespace std { 264309124Sdim__device__ static inline void free(void *__ptr) { ::free(__ptr); } 265309124Sdim__device__ static inline void *malloc(size_t __size) { 266309124Sdim return ::malloc(__size); 267309124Sdim} 268309124Sdim} // namespace std 269309124Sdim 270309124Sdim// Out-of-line implementations from cuda_builtin_vars.h. These need to come 271309124Sdim// after we've pulled in the definition of uint3 and dim3. 272309124Sdim 273309124Sdim__device__ inline __cuda_builtin_threadIdx_t::operator uint3() const { 274309124Sdim uint3 ret; 275309124Sdim ret.x = x; 276309124Sdim ret.y = y; 277309124Sdim ret.z = z; 278309124Sdim return ret; 279309124Sdim} 280309124Sdim 281309124Sdim__device__ inline __cuda_builtin_blockIdx_t::operator uint3() const { 282309124Sdim uint3 ret; 283309124Sdim ret.x = x; 284309124Sdim ret.y = y; 285309124Sdim ret.z = z; 286309124Sdim return ret; 287309124Sdim} 288309124Sdim 289309124Sdim__device__ inline __cuda_builtin_blockDim_t::operator dim3() const { 290309124Sdim return dim3(x, y, z); 291309124Sdim} 292309124Sdim 293309124Sdim__device__ inline __cuda_builtin_gridDim_t::operator dim3() const { 294309124Sdim return dim3(x, y, z); 295309124Sdim} 296309124Sdim 297309124Sdim#include <__clang_cuda_cmath.h> 298309124Sdim#include <__clang_cuda_intrinsics.h> 299309124Sdim 300309124Sdim// curand_mtgp32_kernel helpfully redeclares blockDim and threadIdx in host 301309124Sdim// mode, giving them their "proper" types of dim3 and uint3. This is 302309124Sdim// incompatible with the types we give in cuda_builtin_vars.h. As as hack, 303309124Sdim// force-include the header (nvcc doesn't include it by default) but redefine 304309124Sdim// dim3 and uint3 to our builtin types. (Thankfully dim3 and uint3 are only 305309124Sdim// used here for the redeclarations of blockDim and threadIdx.) 306309124Sdim#pragma push_macro("dim3") 307309124Sdim#pragma push_macro("uint3") 308309124Sdim#define dim3 __cuda_builtin_blockDim_t 309309124Sdim#define uint3 __cuda_builtin_threadIdx_t 310309124Sdim#include "curand_mtgp32_kernel.h" 311309124Sdim#pragma pop_macro("dim3") 312309124Sdim#pragma pop_macro("uint3") 313309124Sdim#pragma pop_macro("__USE_FAST_MATH__") 314309124Sdim 315292920Sdim#endif // __CUDA__ 316292920Sdim#endif // __CLANG_CUDA_RUNTIME_WRAPPER_H__ 317