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 > 7050
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 "cuda_builtin_vars.h"
76
77 // No need for device_launch_parameters.h as 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_new2
117 #undef __cxa_vec_new3
118 #undef __cxa_vec_delete2
119 #undef __cxa_vec_delete
120 #undef __cxa_vec_delete3
121 #undef __cxa_pure_virtual
122
123 // We need decls for functions in CUDA's libdevice with __device__
124 // attribute only. Alas they come either as __host__ __device__ or
125 // with no attributes at all. To work around that, define __CUDA_RTC__
126 // which produces HD variant and undef __host__ which gives us desided
127 // decls with __device__ attribute.
128 #pragma push_macro("__host__")
129 #define __host__
130 #define __CUDACC_RTC__
131 #include "device_functions_decls.h"
132 #undef __CUDACC_RTC__
133
134 // Temporarily poison __host__ macro to ensure it's not used by any of
135 // the headers we're about to include.
136 #define __host__ UNEXPECTED_HOST_ATTRIBUTE
137
138 // device_functions.hpp and math_functions*.hpp use 'static
139 // __forceinline__' (with no __device__) for definitions of device
140 // functions. Temporarily redefine __forceinline__ to include
141 // __device__.
142 #pragma push_macro("__forceinline__")
143 #define __forceinline__ __device__ __inline__ __attribute__((always_inline))
144 #include "device_functions.hpp"
145
146 // math_function.hpp uses the __USE_FAST_MATH__ macro to determine whether we
147 // get the slow-but-accurate or fast-but-inaccurate versions of functions like
148 // sin and exp. This is controlled in clang by -fcuda-approx-transcendentals.
149 //
150 // device_functions.hpp uses __USE_FAST_MATH__ for a different purpose (fast vs.
151 // slow divides), so we need to scope our define carefully here.
152 #pragma push_macro("__USE_FAST_MATH__")
153 #if defined(__CLANG_CUDA_APPROX_TRANSCENDENTALS__)
154 #define __USE_FAST_MATH__
155 #endif
156 #include "math_functions.hpp"
157 #pragma pop_macro("__USE_FAST_MATH__")
158
159 #include "math_functions_dbl_ptx3.hpp"
160 #pragma pop_macro("__forceinline__")
161
162 // Pull in host-only functions that are only available when neither
163 // __CUDACC__ nor __CUDABE__ are defined.
164 #undef __MATH_FUNCTIONS_HPP__
165 #undef __CUDABE__
166 #include "math_functions.hpp"
167 // Alas, additional overloads for these functions are hard to get to.
168 // Considering that we only need these overloads for a few functions,
169 // we can provide them here.
rsqrt(float __a)170 static inline float rsqrt(float __a) { return rsqrtf(__a); }
rcbrt(float __a)171 static inline float rcbrt(float __a) { return rcbrtf(__a); }
sinpi(float __a)172 static inline float sinpi(float __a) { return sinpif(__a); }
cospi(float __a)173 static inline float cospi(float __a) { return cospif(__a); }
sincospi(float __a,float * __b,float * __c)174 static inline void sincospi(float __a, float *__b, float *__c) {
175 return sincospif(__a, __b, __c);
176 }
erfcinv(float __a)177 static inline float erfcinv(float __a) { return erfcinvf(__a); }
normcdfinv(float __a)178 static inline float normcdfinv(float __a) { return normcdfinvf(__a); }
normcdf(float __a)179 static inline float normcdf(float __a) { return normcdff(__a); }
erfcx(float __a)180 static inline float erfcx(float __a) { return erfcxf(__a); }
181
182 // For some reason single-argument variant is not always declared by
183 // CUDA headers. Alas, device_functions.hpp included below needs it.
__brkpt(int __c)184 static inline __device__ void __brkpt(int __c) { __brkpt(); }
185
186 // Now include *.hpp with definitions of various GPU functions. Alas,
187 // a lot of thins get declared/defined with __host__ attribute which
188 // we don't want and we have to define it out. We also have to include
189 // {device,math}_functions.hpp again in order to extract the other
190 // branch of #if/else inside.
191
192 #define __host__
193 #undef __CUDABE__
194 #define __CUDACC__
195 #undef __DEVICE_FUNCTIONS_HPP__
196 #include "device_atomic_functions.hpp"
197 #include "device_functions.hpp"
198 #include "sm_20_atomic_functions.hpp"
199 #include "sm_20_intrinsics.hpp"
200 #include "sm_32_atomic_functions.hpp"
201
202 // Don't include sm_30_intrinsics.h and sm_32_intrinsics.h. These define the
203 // __shfl and __ldg intrinsics using inline (volatile) asm, but we want to
204 // define them using builtins so that the optimizer can reason about and across
205 // these instructions. In particular, using intrinsics for ldg gets us the
206 // [addr+imm] addressing mode, which, although it doesn't actually exist in the
207 // hardware, seems to generate faster machine code because ptxas can more easily
208 // reason about our code.
209
210 #undef __MATH_FUNCTIONS_HPP__
211
212 // math_functions.hpp defines ::signbit as a __host__ __device__ function. This
213 // conflicts with libstdc++'s constexpr ::signbit, so we have to rename
214 // math_function.hpp's ::signbit. It's guarded by #undef signbit, but that's
215 // conditional on __GNUC__. :)
216 #pragma push_macro("signbit")
217 #pragma push_macro("__GNUC__")
218 #undef __GNUC__
219 #define signbit __ignored_cuda_signbit
220 #include "math_functions.hpp"
221 #pragma pop_macro("__GNUC__")
222 #pragma pop_macro("signbit")
223
224 #pragma pop_macro("__host__")
225
226 #include "texture_indirect_functions.h"
227
228 // Restore state of __CUDA_ARCH__ and __THROW we had on entry.
229 #pragma pop_macro("__CUDA_ARCH__")
230 #pragma pop_macro("__THROW")
231
232 // Set up compiler macros expected to be seen during compilation.
233 #undef __CUDABE__
234 #define __CUDACC__
235
236 extern "C" {
237 // Device-side CUDA system calls.
238 // http://docs.nvidia.com/cuda/ptx-writers-guide-to-interoperability/index.html#system-calls
239 // We need these declarations and wrappers for device-side
240 // malloc/free/printf calls to work without relying on
241 // -fcuda-disable-target-call-checks option.
242 __device__ int vprintf(const char *, const char *);
243 __device__ void free(void *) __attribute((nothrow));
244 __device__ void *malloc(size_t) __attribute((nothrow)) __attribute__((malloc));
245 __device__ void __assertfail(const char *__message, const char *__file,
246 unsigned __line, const char *__function,
247 size_t __charSize) __attribute__((noreturn));
248
249 // In order for standard assert() macro on linux to work we need to
250 // provide device-side __assert_fail()
__assert_fail(const char * __message,const char * __file,unsigned __line,const char * __function)251 __device__ static inline void __assert_fail(const char *__message,
252 const char *__file, unsigned __line,
253 const char *__function) {
254 __assertfail(__message, __file, __line, __function, sizeof(char));
255 }
256
257 // Clang will convert printf into vprintf, but we still need
258 // device-side declaration for it.
259 __device__ int printf(const char *, ...);
260 } // extern "C"
261
262 // We also need device-side std::malloc and std::free.
263 namespace std {
free(void * __ptr)264 __device__ static inline void free(void *__ptr) { ::free(__ptr); }
malloc(size_t __size)265 __device__ static inline void *malloc(size_t __size) {
266 return ::malloc(__size);
267 }
268 } // namespace std
269
270 // Out-of-line implementations from cuda_builtin_vars.h. These need to come
271 // after we've pulled in the definition of uint3 and dim3.
272
uint3()273 __device__ inline __cuda_builtin_threadIdx_t::operator uint3() const {
274 uint3 ret;
275 ret.x = x;
276 ret.y = y;
277 ret.z = z;
278 return ret;
279 }
280
uint3()281 __device__ inline __cuda_builtin_blockIdx_t::operator uint3() const {
282 uint3 ret;
283 ret.x = x;
284 ret.y = y;
285 ret.z = z;
286 return ret;
287 }
288
dim3()289 __device__ inline __cuda_builtin_blockDim_t::operator dim3() const {
290 return dim3(x, y, z);
291 }
292
dim3()293 __device__ inline __cuda_builtin_gridDim_t::operator dim3() const {
294 return dim3(x, y, z);
295 }
296
297 #include <__clang_cuda_cmath.h>
298 #include <__clang_cuda_intrinsics.h>
299
300 // curand_mtgp32_kernel helpfully redeclares blockDim and threadIdx in host
301 // mode, giving them their "proper" types of dim3 and uint3. This is
302 // incompatible with the types we give in cuda_builtin_vars.h. As as hack,
303 // force-include the header (nvcc doesn't include it by default) but redefine
304 // dim3 and uint3 to our builtin types. (Thankfully dim3 and uint3 are only
305 // used here for the redeclarations of blockDim and threadIdx.)
306 #pragma push_macro("dim3")
307 #pragma push_macro("uint3")
308 #define dim3 __cuda_builtin_blockDim_t
309 #define uint3 __cuda_builtin_threadIdx_t
310 #include "curand_mtgp32_kernel.h"
311 #pragma pop_macro("dim3")
312 #pragma pop_macro("uint3")
313 #pragma pop_macro("__USE_FAST_MATH__")
314
315 #endif // __CUDA__
316 #endif // __CLANG_CUDA_RUNTIME_WRAPPER_H__
317