1 /*===---- __clang_cuda_runtime_wrapper.h - CUDA runtime support -------------=== 2 * 3 * Permission is hereby granted, free of charge, to any person obtaining a copy 4 * of this software and associated documentation files (the "Software"), to deal 5 * in the Software without restriction, including without limitation the rights 6 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 7 * copies of the Software, and to permit persons to whom the Software is 8 * furnished to do so, subject to the following conditions: 9 * 10 * The above copyright notice and this permission notice shall be included in 11 * all copies or substantial portions of the Software. 12 * 13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 15 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 16 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 17 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 18 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 19 * THE SOFTWARE. 20 * 21 *===-----------------------------------------------------------------------=== 22 */ 23 24 /* 25 * WARNING: This header is intended to be directly -include'd by 26 * the compiler and is not supposed to be included by users. 27 * 28 * CUDA headers are implemented in a way that currently makes it 29 * impossible for user code to #include directly when compiling with 30 * Clang. They present different view of CUDA-supplied functions 31 * depending on where in NVCC's compilation pipeline the headers are 32 * included. Neither of these modes provides function definitions with 33 * correct attributes, so we use preprocessor to force the headers 34 * into a form that Clang can use. 35 * 36 * Similarly to NVCC which -include's cuda_runtime.h, Clang -include's 37 * this file during every CUDA compilation. 38 */ 39 40 #ifndef __CLANG_CUDA_RUNTIME_WRAPPER_H__ 41 #define __CLANG_CUDA_RUNTIME_WRAPPER_H__ 42 43 #if defined(__CUDA__) && defined(__clang__) 44 45 // Include some forward declares that must come before cmath. 46 #include <__clang_cuda_math_forward_declares.h> 47 48 // Include some standard headers to avoid CUDA headers including them 49 // while some required macros (like __THROW) are in a weird state. 50 #include <cmath> 51 #include <cstdlib> 52 #include <stdlib.h> 53 54 // Preserve common macros that will be changed below by us or by CUDA 55 // headers. 56 #pragma push_macro("__THROW") 57 #pragma push_macro("__CUDA_ARCH__") 58 59 // WARNING: Preprocessor hacks below are based on specific details of 60 // CUDA-7.x headers and are not expected to work with any other 61 // version of CUDA headers. 62 #include "cuda.h" 63 #if !defined(CUDA_VERSION) 64 #error "cuda.h did not define CUDA_VERSION" 65 #elif CUDA_VERSION < 7000 || CUDA_VERSION > 8000 66 #error "Unsupported CUDA version!" 67 #endif 68 69 // Make largest subset of device functions available during host 70 // compilation -- SM_35 for the time being. 71 #ifndef __CUDA_ARCH__ 72 #define __CUDA_ARCH__ 350 73 #endif 74 75 #include "__clang_cuda_builtin_vars.h" 76 77 // No need for device_launch_parameters.h as __clang_cuda_builtin_vars.h above 78 // has taken care of builtin variables declared in the file. 79 #define __DEVICE_LAUNCH_PARAMETERS_H__ 80 81 // {math,device}_functions.h only have declarations of the 82 // functions. We don't need them as we're going to pull in their 83 // definitions from .hpp files. 84 #define __DEVICE_FUNCTIONS_H__ 85 #define __MATH_FUNCTIONS_H__ 86 #define __COMMON_FUNCTIONS_H__ 87 88 #undef __CUDACC__ 89 #define __CUDABE__ 90 // Disables definitions of device-side runtime support stubs in 91 // cuda_device_runtime_api.h 92 #include "driver_types.h" 93 #include "host_config.h" 94 #include "host_defines.h" 95 96 #undef __CUDABE__ 97 #define __CUDACC__ 98 #include "cuda_runtime.h" 99 100 #undef __CUDACC__ 101 #define __CUDABE__ 102 103 // CUDA headers use __nvvm_memcpy and __nvvm_memset which Clang does 104 // not have at the moment. Emulate them with a builtin memcpy/memset. 105 #define __nvvm_memcpy(s, d, n, a) __builtin_memcpy(s, d, n) 106 #define __nvvm_memset(d, c, n, a) __builtin_memset(d, c, n) 107 108 #include "crt/device_runtime.h" 109 #include "crt/host_runtime.h" 110 // device_runtime.h defines __cxa_* macros that will conflict with 111 // cxxabi.h. 112 // FIXME: redefine these as __device__ functions. 113 #undef __cxa_vec_ctor 114 #undef __cxa_vec_cctor 115 #undef __cxa_vec_dtor 116 #undef __cxa_vec_new 117 #undef __cxa_vec_new2 118 #undef __cxa_vec_new3 119 #undef __cxa_vec_delete2 120 #undef __cxa_vec_delete 121 #undef __cxa_vec_delete3 122 #undef __cxa_pure_virtual 123 124 // We need decls for functions in CUDA's libdevice with __device__ 125 // attribute only. Alas they come either as __host__ __device__ or 126 // with no attributes at all. To work around that, define __CUDA_RTC__ 127 // which produces HD variant and undef __host__ which gives us desided 128 // decls with __device__ attribute. 129 #pragma push_macro("__host__") 130 #define __host__ 131 #define __CUDACC_RTC__ 132 #include "device_functions_decls.h" 133 #undef __CUDACC_RTC__ 134 135 // Temporarily poison __host__ macro to ensure it's not used by any of 136 // the headers we're about to include. 137 #define __host__ UNEXPECTED_HOST_ATTRIBUTE 138 139 // CUDA 8.0.41 relies on __USE_FAST_MATH__ and __CUDA_PREC_DIV's values. 140 // Previous versions used to check whether they are defined or not. 141 // CU_DEVICE_INVALID macro is only defined in 8.0.41, so we use it 142 // here to detect the switch. 143 144 #if defined(CU_DEVICE_INVALID) 145 #if !defined(__USE_FAST_MATH__) 146 #define __USE_FAST_MATH__ 0 147 #endif 148 149 #if !defined(__CUDA_PREC_DIV) 150 #define __CUDA_PREC_DIV 0 151 #endif 152 #endif 153 154 // device_functions.hpp and math_functions*.hpp use 'static 155 // __forceinline__' (with no __device__) for definitions of device 156 // functions. Temporarily redefine __forceinline__ to include 157 // __device__. 158 #pragma push_macro("__forceinline__") 159 #define __forceinline__ __device__ __inline__ __attribute__((always_inline)) 160 #include "device_functions.hpp" 161 162 // math_function.hpp uses the __USE_FAST_MATH__ macro to determine whether we 163 // get the slow-but-accurate or fast-but-inaccurate versions of functions like 164 // sin and exp. This is controlled in clang by -fcuda-approx-transcendentals. 165 // 166 // device_functions.hpp uses __USE_FAST_MATH__ for a different purpose (fast vs. 167 // slow divides), so we need to scope our define carefully here. 168 #pragma push_macro("__USE_FAST_MATH__") 169 #if defined(__CLANG_CUDA_APPROX_TRANSCENDENTALS__) 170 #define __USE_FAST_MATH__ 1 171 #endif 172 #include "math_functions.hpp" 173 #pragma pop_macro("__USE_FAST_MATH__") 174 175 #include "math_functions_dbl_ptx3.hpp" 176 #pragma pop_macro("__forceinline__") 177 178 // Pull in host-only functions that are only available when neither 179 // __CUDACC__ nor __CUDABE__ are defined. 180 #undef __MATH_FUNCTIONS_HPP__ 181 #undef __CUDABE__ 182 #include "math_functions.hpp" 183 // Alas, additional overloads for these functions are hard to get to. 184 // Considering that we only need these overloads for a few functions, 185 // we can provide them here. 186 static inline float rsqrt(float __a) { return rsqrtf(__a); } 187 static inline float rcbrt(float __a) { return rcbrtf(__a); } 188 static inline float sinpi(float __a) { return sinpif(__a); } 189 static inline float cospi(float __a) { return cospif(__a); } 190 static inline void sincospi(float __a, float *__b, float *__c) { 191 return sincospif(__a, __b, __c); 192 } 193 static inline float erfcinv(float __a) { return erfcinvf(__a); } 194 static inline float normcdfinv(float __a) { return normcdfinvf(__a); } 195 static inline float normcdf(float __a) { return normcdff(__a); } 196 static inline float erfcx(float __a) { return erfcxf(__a); } 197 198 // For some reason single-argument variant is not always declared by 199 // CUDA headers. Alas, device_functions.hpp included below needs it. 200 static inline __device__ void __brkpt(int __c) { __brkpt(); } 201 202 // Now include *.hpp with definitions of various GPU functions. Alas, 203 // a lot of thins get declared/defined with __host__ attribute which 204 // we don't want and we have to define it out. We also have to include 205 // {device,math}_functions.hpp again in order to extract the other 206 // branch of #if/else inside. 207 208 #define __host__ 209 #undef __CUDABE__ 210 #define __CUDACC__ 211 #undef __DEVICE_FUNCTIONS_HPP__ 212 #include "device_atomic_functions.hpp" 213 #include "device_functions.hpp" 214 #include "sm_20_atomic_functions.hpp" 215 #include "sm_20_intrinsics.hpp" 216 #include "sm_32_atomic_functions.hpp" 217 218 // Don't include sm_30_intrinsics.h and sm_32_intrinsics.h. These define the 219 // __shfl and __ldg intrinsics using inline (volatile) asm, but we want to 220 // define them using builtins so that the optimizer can reason about and across 221 // these instructions. In particular, using intrinsics for ldg gets us the 222 // [addr+imm] addressing mode, which, although it doesn't actually exist in the 223 // hardware, seems to generate faster machine code because ptxas can more easily 224 // reason about our code. 225 226 #undef __MATH_FUNCTIONS_HPP__ 227 228 // math_functions.hpp defines ::signbit as a __host__ __device__ function. This 229 // conflicts with libstdc++'s constexpr ::signbit, so we have to rename 230 // math_function.hpp's ::signbit. It's guarded by #undef signbit, but that's 231 // conditional on __GNUC__. :) 232 #pragma push_macro("signbit") 233 #pragma push_macro("__GNUC__") 234 #undef __GNUC__ 235 #define signbit __ignored_cuda_signbit 236 #include "math_functions.hpp" 237 #pragma pop_macro("__GNUC__") 238 #pragma pop_macro("signbit") 239 240 #pragma pop_macro("__host__") 241 242 #include "texture_indirect_functions.h" 243 244 // Restore state of __CUDA_ARCH__ and __THROW we had on entry. 245 #pragma pop_macro("__CUDA_ARCH__") 246 #pragma pop_macro("__THROW") 247 248 // Set up compiler macros expected to be seen during compilation. 249 #undef __CUDABE__ 250 #define __CUDACC__ 251 252 extern "C" { 253 // Device-side CUDA system calls. 254 // http://docs.nvidia.com/cuda/ptx-writers-guide-to-interoperability/index.html#system-calls 255 // We need these declarations and wrappers for device-side 256 // malloc/free/printf calls to work without relying on 257 // -fcuda-disable-target-call-checks option. 258 __device__ int vprintf(const char *, const char *); 259 __device__ void free(void *) __attribute((nothrow)); 260 __device__ void *malloc(size_t) __attribute((nothrow)) __attribute__((malloc)); 261 __device__ void __assertfail(const char *__message, const char *__file, 262 unsigned __line, const char *__function, 263 size_t __charSize) __attribute__((noreturn)); 264 265 // In order for standard assert() macro on linux to work we need to 266 // provide device-side __assert_fail() 267 __device__ static inline void __assert_fail(const char *__message, 268 const char *__file, unsigned __line, 269 const char *__function) { 270 __assertfail(__message, __file, __line, __function, sizeof(char)); 271 } 272 273 // Clang will convert printf into vprintf, but we still need 274 // device-side declaration for it. 275 __device__ int printf(const char *, ...); 276 } // extern "C" 277 278 // We also need device-side std::malloc and std::free. 279 namespace std { 280 __device__ static inline void free(void *__ptr) { ::free(__ptr); } 281 __device__ static inline void *malloc(size_t __size) { 282 return ::malloc(__size); 283 } 284 } // namespace std 285 286 // Out-of-line implementations from __clang_cuda_builtin_vars.h. These need to 287 // come after we've pulled in the definition of uint3 and dim3. 288 289 __device__ inline __cuda_builtin_threadIdx_t::operator uint3() const { 290 uint3 ret; 291 ret.x = x; 292 ret.y = y; 293 ret.z = z; 294 return ret; 295 } 296 297 __device__ inline __cuda_builtin_blockIdx_t::operator uint3() const { 298 uint3 ret; 299 ret.x = x; 300 ret.y = y; 301 ret.z = z; 302 return ret; 303 } 304 305 __device__ inline __cuda_builtin_blockDim_t::operator dim3() const { 306 return dim3(x, y, z); 307 } 308 309 __device__ inline __cuda_builtin_gridDim_t::operator dim3() const { 310 return dim3(x, y, z); 311 } 312 313 #include <__clang_cuda_cmath.h> 314 #include <__clang_cuda_intrinsics.h> 315 #include <__clang_cuda_complex_builtins.h> 316 317 // curand_mtgp32_kernel helpfully redeclares blockDim and threadIdx in host 318 // mode, giving them their "proper" types of dim3 and uint3. This is 319 // incompatible with the types we give in __clang_cuda_builtin_vars.h. As as 320 // hack, force-include the header (nvcc doesn't include it by default) but 321 // redefine dim3 and uint3 to our builtin types. (Thankfully dim3 and uint3 are 322 // only used here for the redeclarations of blockDim and threadIdx.) 323 #pragma push_macro("dim3") 324 #pragma push_macro("uint3") 325 #define dim3 __cuda_builtin_blockDim_t 326 #define uint3 __cuda_builtin_threadIdx_t 327 #include "curand_mtgp32_kernel.h" 328 #pragma pop_macro("dim3") 329 #pragma pop_macro("uint3") 330 #pragma pop_macro("__USE_FAST_MATH__") 331 332 #endif // __CUDA__ 333 #endif // __CLANG_CUDA_RUNTIME_WRAPPER_H__ 334