1 /*===---- __clang_cuda_runtime_wrapper.h - CUDA runtime support -------------===
2  *
3  * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4  * See https://llvm.org/LICENSE.txt for license information.
5  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6  *
7  *===-----------------------------------------------------------------------===
8  */
9 
10 /*
11  * WARNING: This header is intended to be directly -include'd by
12  * the compiler and is not supposed to be included by users.
13  *
14  * CUDA headers are implemented in a way that currently makes it
15  * impossible for user code to #include directly when compiling with
16  * Clang. They present different view of CUDA-supplied functions
17  * depending on where in NVCC's compilation pipeline the headers are
18  * included. Neither of these modes provides function definitions with
19  * correct attributes, so we use preprocessor to force the headers
20  * into a form that Clang can use.
21  *
22  * Similarly to NVCC which -include's cuda_runtime.h, Clang -include's
23  * this file during every CUDA compilation.
24  */
25 
26 #ifndef __CLANG_CUDA_RUNTIME_WRAPPER_H__
27 #define __CLANG_CUDA_RUNTIME_WRAPPER_H__
28 
29 #if defined(__CUDA__) && defined(__clang__)
30 
31 // Include some forward declares that must come before cmath.
32 #include <__clang_cuda_math_forward_declares.h>
33 
34 // Define __CUDACC__ early as libstdc++ standard headers with GNU extensions
35 // enabled depend on it to avoid using __float128, which is unsupported in
36 // CUDA.
37 #define __CUDACC__
38 
39 // Include some standard headers to avoid CUDA headers including them
40 // while some required macros (like __THROW) are in a weird state.
41 #include <cmath>
42 #include <cstdlib>
43 #include <stdlib.h>
44 #undef __CUDACC__
45 
46 // Preserve common macros that will be changed below by us or by CUDA
47 // headers.
48 #pragma push_macro("__THROW")
49 #pragma push_macro("__CUDA_ARCH__")
50 
51 // WARNING: Preprocessor hacks below are based on specific details of
52 // CUDA-7.x headers and are not expected to work with any other
53 // version of CUDA headers.
54 #include "cuda.h"
55 #if !defined(CUDA_VERSION)
56 #error "cuda.h did not define CUDA_VERSION"
57 #elif CUDA_VERSION < 7000
58 #error "Unsupported CUDA version!"
59 #endif
60 
61 #pragma push_macro("__CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__")
62 #if CUDA_VERSION >= 10000
63 #define __CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__
64 #endif
65 
66 // Make largest subset of device functions available during host
67 // compilation -- SM_35 for the time being.
68 #ifndef __CUDA_ARCH__
69 #define __CUDA_ARCH__ 350
70 #endif
71 
72 #include "__clang_cuda_builtin_vars.h"
73 
74 // No need for device_launch_parameters.h as __clang_cuda_builtin_vars.h above
75 // has taken care of builtin variables declared in the file.
76 #define __DEVICE_LAUNCH_PARAMETERS_H__
77 
78 // {math,device}_functions.h only have declarations of the
79 // functions. We don't need them as we're going to pull in their
80 // definitions from .hpp files.
81 #define __DEVICE_FUNCTIONS_H__
82 #define __MATH_FUNCTIONS_H__
83 #define __COMMON_FUNCTIONS_H__
84 // device_functions_decls is replaced by __clang_cuda_device_functions.h
85 // included below.
86 #define __DEVICE_FUNCTIONS_DECLS_H__
87 
88 #undef __CUDACC__
89 #if CUDA_VERSION < 9000
90 #define __CUDABE__
91 #else
92 #define __CUDACC__
93 #define __CUDA_LIBDEVICE__
94 #endif
95 // Disables definitions of device-side runtime support stubs in
96 // cuda_device_runtime_api.h
97 #include "host_defines.h"
98 #undef __CUDACC__
99 #include "driver_types.h"
100 #include "host_config.h"
101 
102 // Temporarily replace "nv_weak" with weak, so __attribute__((nv_weak)) in
103 // cuda_device_runtime_api.h ends up being __attribute__((weak)) which is the
104 // functional equivalent of what we need.
105 #pragma push_macro("nv_weak")
106 #define nv_weak weak
107 #undef __CUDABE__
108 #undef __CUDA_LIBDEVICE__
109 #define __CUDACC__
110 #include "cuda_runtime.h"
111 
112 #pragma pop_macro("nv_weak")
113 #undef __CUDACC__
114 #define __CUDABE__
115 
116 // CUDA headers use __nvvm_memcpy and __nvvm_memset which Clang does
117 // not have at the moment. Emulate them with a builtin memcpy/memset.
118 #define __nvvm_memcpy(s, d, n, a) __builtin_memcpy(s, d, n)
119 #define __nvvm_memset(d, c, n, a) __builtin_memset(d, c, n)
120 
121 #if CUDA_VERSION < 9000
122 #include "crt/device_runtime.h"
123 #endif
124 #include "crt/host_runtime.h"
125 // device_runtime.h defines __cxa_* macros that will conflict with
126 // cxxabi.h.
127 // FIXME: redefine these as __device__ functions.
128 #undef __cxa_vec_ctor
129 #undef __cxa_vec_cctor
130 #undef __cxa_vec_dtor
131 #undef __cxa_vec_new
132 #undef __cxa_vec_new2
133 #undef __cxa_vec_new3
134 #undef __cxa_vec_delete2
135 #undef __cxa_vec_delete
136 #undef __cxa_vec_delete3
137 #undef __cxa_pure_virtual
138 
139 // math_functions.hpp expects this host function be defined on MacOS, but it
140 // ends up not being there because of the games we play here.  Just define it
141 // ourselves; it's simple enough.
142 #ifdef __APPLE__
__signbitd(double x)143 inline __host__ double __signbitd(double x) {
144   return std::signbit(x);
145 }
146 #endif
147 
148 // CUDA 9.1 no longer provides declarations for libdevice functions, so we need
149 // to provide our own.
150 #include <__clang_cuda_libdevice_declares.h>
151 
152 // Wrappers for many device-side standard library functions, incl. math
153 // functions, became compiler builtins in CUDA-9 and have been removed from the
154 // CUDA headers. Clang now provides its own implementation of the wrappers.
155 #if CUDA_VERSION >= 9000
156 #include <__clang_cuda_device_functions.h>
157 #include <__clang_cuda_math.h>
158 #endif
159 
160 // __THROW is redefined to be empty by device_functions_decls.h in CUDA. Clang's
161 // counterpart does not do it, so we need to make it empty here to keep
162 // following CUDA includes happy.
163 #undef __THROW
164 #define __THROW
165 
166 // CUDA 8.0.41 relies on __USE_FAST_MATH__ and __CUDA_PREC_DIV's values.
167 // Previous versions used to check whether they are defined or not.
168 // CU_DEVICE_INVALID macro is only defined in 8.0.41, so we use it
169 // here to detect the switch.
170 
171 #if defined(CU_DEVICE_INVALID)
172 #if !defined(__USE_FAST_MATH__)
173 #define __USE_FAST_MATH__ 0
174 #endif
175 
176 #if !defined(__CUDA_PREC_DIV)
177 #define __CUDA_PREC_DIV 0
178 #endif
179 #endif
180 
181 // Temporarily poison __host__ macro to ensure it's not used by any of
182 // the headers we're about to include.
183 #pragma push_macro("__host__")
184 #define __host__ UNEXPECTED_HOST_ATTRIBUTE
185 
186 // device_functions.hpp and math_functions*.hpp use 'static
187 // __forceinline__' (with no __device__) for definitions of device
188 // functions. Temporarily redefine __forceinline__ to include
189 // __device__.
190 #pragma push_macro("__forceinline__")
191 #define __forceinline__ __device__ __inline__ __attribute__((always_inline))
192 #if CUDA_VERSION < 9000
193 #include "device_functions.hpp"
194 #endif
195 
196 // math_function.hpp uses the __USE_FAST_MATH__ macro to determine whether we
197 // get the slow-but-accurate or fast-but-inaccurate versions of functions like
198 // sin and exp.  This is controlled in clang by -fcuda-approx-transcendentals.
199 //
200 // device_functions.hpp uses __USE_FAST_MATH__ for a different purpose (fast vs.
201 // slow divides), so we need to scope our define carefully here.
202 #pragma push_macro("__USE_FAST_MATH__")
203 #if defined(__CLANG_CUDA_APPROX_TRANSCENDENTALS__)
204 #define __USE_FAST_MATH__ 1
205 #endif
206 
207 #if CUDA_VERSION >= 9000
208 // CUDA-9.2 needs host-side memcpy for some host functions in
209 // device_functions.hpp
210 #if CUDA_VERSION >= 9020
211 #include <string.h>
212 #endif
213 #include "crt/math_functions.hpp"
214 #else
215 #include "math_functions.hpp"
216 #endif
217 
218 #pragma pop_macro("__USE_FAST_MATH__")
219 
220 #if CUDA_VERSION < 9000
221 #include "math_functions_dbl_ptx3.hpp"
222 #endif
223 #pragma pop_macro("__forceinline__")
224 
225 // Pull in host-only functions that are only available when neither
226 // __CUDACC__ nor __CUDABE__ are defined.
227 #undef __MATH_FUNCTIONS_HPP__
228 #undef __CUDABE__
229 #if CUDA_VERSION < 9000
230 #include "math_functions.hpp"
231 #endif
232 // Alas, additional overloads for these functions are hard to get to.
233 // Considering that we only need these overloads for a few functions,
234 // we can provide them here.
rsqrt(float __a)235 static inline float rsqrt(float __a) { return rsqrtf(__a); }
rcbrt(float __a)236 static inline float rcbrt(float __a) { return rcbrtf(__a); }
sinpi(float __a)237 static inline float sinpi(float __a) { return sinpif(__a); }
cospi(float __a)238 static inline float cospi(float __a) { return cospif(__a); }
sincospi(float __a,float * __b,float * __c)239 static inline void sincospi(float __a, float *__b, float *__c) {
240   return sincospif(__a, __b, __c);
241 }
erfcinv(float __a)242 static inline float erfcinv(float __a) { return erfcinvf(__a); }
normcdfinv(float __a)243 static inline float normcdfinv(float __a) { return normcdfinvf(__a); }
normcdf(float __a)244 static inline float normcdf(float __a) { return normcdff(__a); }
erfcx(float __a)245 static inline float erfcx(float __a) { return erfcxf(__a); }
246 
247 #if CUDA_VERSION < 9000
248 // For some reason single-argument variant is not always declared by
249 // CUDA headers. Alas, device_functions.hpp included below needs it.
__brkpt(int __c)250 static inline __device__ void __brkpt(int __c) { __brkpt(); }
251 #endif
252 
253 // Now include *.hpp with definitions of various GPU functions.  Alas,
254 // a lot of thins get declared/defined with __host__ attribute which
255 // we don't want and we have to define it out. We also have to include
256 // {device,math}_functions.hpp again in order to extract the other
257 // branch of #if/else inside.
258 #define __host__
259 #undef __CUDABE__
260 #define __CUDACC__
261 #if CUDA_VERSION >= 9000
262 // Some atomic functions became compiler builtins in CUDA-9 , so we need their
263 // declarations.
264 #include "device_atomic_functions.h"
265 #endif
266 #undef __DEVICE_FUNCTIONS_HPP__
267 #include "device_atomic_functions.hpp"
268 #if CUDA_VERSION >= 9000
269 #include "crt/device_functions.hpp"
270 #include "crt/device_double_functions.hpp"
271 #else
272 #include "device_functions.hpp"
273 #define __CUDABE__
274 #include "device_double_functions.h"
275 #undef __CUDABE__
276 #endif
277 #include "sm_20_atomic_functions.hpp"
278 #include "sm_20_intrinsics.hpp"
279 #include "sm_32_atomic_functions.hpp"
280 
281 // Don't include sm_30_intrinsics.h and sm_32_intrinsics.h.  These define the
282 // __shfl and __ldg intrinsics using inline (volatile) asm, but we want to
283 // define them using builtins so that the optimizer can reason about and across
284 // these instructions.  In particular, using intrinsics for ldg gets us the
285 // [addr+imm] addressing mode, which, although it doesn't actually exist in the
286 // hardware, seems to generate faster machine code because ptxas can more easily
287 // reason about our code.
288 
289 #if CUDA_VERSION >= 8000
290 #pragma push_macro("__CUDA_ARCH__")
291 #undef __CUDA_ARCH__
292 #include "sm_60_atomic_functions.hpp"
293 #include "sm_61_intrinsics.hpp"
294 #pragma pop_macro("__CUDA_ARCH__")
295 #endif
296 
297 #undef __MATH_FUNCTIONS_HPP__
298 
299 // math_functions.hpp defines ::signbit as a __host__ __device__ function.  This
300 // conflicts with libstdc++'s constexpr ::signbit, so we have to rename
301 // math_function.hpp's ::signbit.  It's guarded by #undef signbit, but that's
302 // conditional on __GNUC__.  :)
303 #pragma push_macro("signbit")
304 #pragma push_macro("__GNUC__")
305 #undef __GNUC__
306 #define signbit __ignored_cuda_signbit
307 
308 // CUDA-9 omits device-side definitions of some math functions if it sees
309 // include guard from math.h wrapper from libstdc++. We have to undo the header
310 // guard temporarily to get the definitions we need.
311 #pragma push_macro("_GLIBCXX_MATH_H")
312 #pragma push_macro("_LIBCPP_VERSION")
313 #if CUDA_VERSION >= 9000
314 #undef _GLIBCXX_MATH_H
315 // We also need to undo another guard that checks for libc++ 3.8+
316 #ifdef _LIBCPP_VERSION
317 #define _LIBCPP_VERSION 3700
318 #endif
319 #endif
320 
321 #if CUDA_VERSION >= 9000
322 #include "crt/math_functions.hpp"
323 #else
324 #include "math_functions.hpp"
325 #endif
326 #pragma pop_macro("_GLIBCXX_MATH_H")
327 #pragma pop_macro("_LIBCPP_VERSION")
328 #pragma pop_macro("__GNUC__")
329 #pragma pop_macro("signbit")
330 
331 #pragma pop_macro("__host__")
332 
333 #include "texture_indirect_functions.h"
334 
335 // Restore state of __CUDA_ARCH__ and __THROW we had on entry.
336 #pragma pop_macro("__CUDA_ARCH__")
337 #pragma pop_macro("__THROW")
338 
339 // Set up compiler macros expected to be seen during compilation.
340 #undef __CUDABE__
341 #define __CUDACC__
342 
343 extern "C" {
344 // Device-side CUDA system calls.
345 // http://docs.nvidia.com/cuda/ptx-writers-guide-to-interoperability/index.html#system-calls
346 // We need these declarations and wrappers for device-side
347 // malloc/free/printf calls to work without relying on
348 // -fcuda-disable-target-call-checks option.
349 __device__ int vprintf(const char *, const char *);
350 __device__ void free(void *) __attribute((nothrow));
351 __device__ void *malloc(size_t) __attribute((nothrow)) __attribute__((malloc));
352 __device__ void __assertfail(const char *__message, const char *__file,
353                              unsigned __line, const char *__function,
354                              size_t __charSize) __attribute__((noreturn));
355 
356 // In order for standard assert() macro on linux to work we need to
357 // provide device-side __assert_fail()
__assert_fail(const char * __message,const char * __file,unsigned __line,const char * __function)358 __device__ static inline void __assert_fail(const char *__message,
359                                             const char *__file, unsigned __line,
360                                             const char *__function) {
361   __assertfail(__message, __file, __line, __function, sizeof(char));
362 }
363 
364 // Clang will convert printf into vprintf, but we still need
365 // device-side declaration for it.
366 __device__ int printf(const char *, ...);
367 } // extern "C"
368 
369 // We also need device-side std::malloc and std::free.
370 namespace std {
free(void * __ptr)371 __device__ static inline void free(void *__ptr) { ::free(__ptr); }
malloc(size_t __size)372 __device__ static inline void *malloc(size_t __size) {
373   return ::malloc(__size);
374 }
375 } // namespace std
376 
377 // Out-of-line implementations from __clang_cuda_builtin_vars.h.  These need to
378 // come after we've pulled in the definition of uint3 and dim3.
379 
dim3()380 __device__ inline __cuda_builtin_threadIdx_t::operator dim3() const {
381   return dim3(x, y, z);
382 }
383 
uint3()384 __device__ inline __cuda_builtin_threadIdx_t::operator uint3() const {
385   return {x, y, z};
386 }
387 
dim3()388 __device__ inline __cuda_builtin_blockIdx_t::operator dim3() const {
389   return dim3(x, y, z);
390 }
391 
uint3()392 __device__ inline __cuda_builtin_blockIdx_t::operator uint3() const {
393   return {x, y, z};
394 }
395 
dim3()396 __device__ inline __cuda_builtin_blockDim_t::operator dim3() const {
397   return dim3(x, y, z);
398 }
399 
uint3()400 __device__ inline __cuda_builtin_blockDim_t::operator uint3() const {
401   return {x, y, z};
402 }
403 
dim3()404 __device__ inline __cuda_builtin_gridDim_t::operator dim3() const {
405   return dim3(x, y, z);
406 }
407 
uint3()408 __device__ inline __cuda_builtin_gridDim_t::operator uint3() const {
409   return {x, y, z};
410 }
411 
412 #include <__clang_cuda_cmath.h>
413 #include <__clang_cuda_intrinsics.h>
414 #include <__clang_cuda_complex_builtins.h>
415 
416 // curand_mtgp32_kernel helpfully redeclares blockDim and threadIdx in host
417 // mode, giving them their "proper" types of dim3 and uint3.  This is
418 // incompatible with the types we give in __clang_cuda_builtin_vars.h.  As as
419 // hack, force-include the header (nvcc doesn't include it by default) but
420 // redefine dim3 and uint3 to our builtin types.  (Thankfully dim3 and uint3 are
421 // only used here for the redeclarations of blockDim and threadIdx.)
422 #pragma push_macro("dim3")
423 #pragma push_macro("uint3")
424 #define dim3 __cuda_builtin_blockDim_t
425 #define uint3 __cuda_builtin_threadIdx_t
426 #include "curand_mtgp32_kernel.h"
427 #pragma pop_macro("dim3")
428 #pragma pop_macro("uint3")
429 #pragma pop_macro("__USE_FAST_MATH__")
430 #pragma pop_macro("__CUDA_INCLUDE_COMPILER_INTERNAL_HEADERS__")
431 
432 // CUDA runtime uses this undocumented function to access kernel launch
433 // configuration. The declaration is in crt/device_functions.h but that file
434 // includes a lot of other stuff we don't want. Instead, we'll provide our own
435 // declaration for it here.
436 #if CUDA_VERSION >= 9020
437 extern "C" unsigned __cudaPushCallConfiguration(dim3 gridDim, dim3 blockDim,
438                                                 size_t sharedMem = 0,
439                                                 void *stream = 0);
440 #endif
441 
442 #endif // __CUDA__
443 #endif // __CLANG_CUDA_RUNTIME_WRAPPER_H__
444