1 /*===---- __clang_cuda_runtime_wrapper.h - CUDA runtime support -------------===
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:
10 * The above copyright notice and this permission notice shall be included in
11 * all copies or substantial portions of the Software.
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
21 *===-----------------------------------------------------------------------===
25 * WARNING: This header is intended to be directly -include'd by
26 * the compiler and is not supposed to be included by users.
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.
36 * Similarly to NVCC which -include's cuda_runtime.h, Clang -include's
37 * this file during every CUDA compilation.
40 #ifndef __CLANG_CUDA_RUNTIME_WRAPPER_H__
41 #define __CLANG_CUDA_RUNTIME_WRAPPER_H__
43 #if defined(__CUDA__) && defined(__clang__)
45 // Include some forward declares that must come before cmath.
46 #include <__clang_cuda_math_forward_declares.h>
48 // Include some standard headers to avoid CUDA headers including them
49 // while some required macros (like __THROW) are in a weird state.
54 // Preserve common macros that will be changed below by us or by CUDA
56 #pragma push_macro("__THROW")
57 #pragma push_macro("__CUDA_ARCH__")
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.
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!"
69 // Make largest subset of device functions available during host
70 // compilation -- SM_35 for the time being.
72 #define __CUDA_ARCH__ 350
75 #include "__clang_cuda_builtin_vars.h"
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__
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__
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"
98 #include "cuda_runtime.h"
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)
108 #include "crt/device_runtime.h"
109 #include "crt/host_runtime.h"
110 // device_runtime.h defines __cxa_* macros that will conflict with
112 // FIXME: redefine these as __device__ functions.
113 #undef __cxa_vec_ctor
114 #undef __cxa_vec_cctor
115 #undef __cxa_vec_dtor
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
124 // math_functions.hpp expects this host function be defined on MacOS, but it
125 // ends up not being there because of the games we play here. Just define it
126 // ourselves; it's simple enough.
128 inline __host__ double __signbitd(double x) {
129 return std::signbit(x);
133 // We need decls for functions in CUDA's libdevice with __device__
134 // attribute only. Alas they come either as __host__ __device__ or
135 // with no attributes at all. To work around that, define __CUDA_RTC__
136 // which produces HD variant and undef __host__ which gives us desided
137 // decls with __device__ attribute.
138 #pragma push_macro("__host__")
140 #define __CUDACC_RTC__
141 #include "device_functions_decls.h"
142 #undef __CUDACC_RTC__
144 // Temporarily poison __host__ macro to ensure it's not used by any of
145 // the headers we're about to include.
146 #define __host__ UNEXPECTED_HOST_ATTRIBUTE
148 // CUDA 8.0.41 relies on __USE_FAST_MATH__ and __CUDA_PREC_DIV's values.
149 // Previous versions used to check whether they are defined or not.
150 // CU_DEVICE_INVALID macro is only defined in 8.0.41, so we use it
151 // here to detect the switch.
153 #if defined(CU_DEVICE_INVALID)
154 #if !defined(__USE_FAST_MATH__)
155 #define __USE_FAST_MATH__ 0
158 #if !defined(__CUDA_PREC_DIV)
159 #define __CUDA_PREC_DIV 0
163 // device_functions.hpp and math_functions*.hpp use 'static
164 // __forceinline__' (with no __device__) for definitions of device
165 // functions. Temporarily redefine __forceinline__ to include
167 #pragma push_macro("__forceinline__")
168 #define __forceinline__ __device__ __inline__ __attribute__((always_inline))
169 #include "device_functions.hpp"
171 // math_function.hpp uses the __USE_FAST_MATH__ macro to determine whether we
172 // get the slow-but-accurate or fast-but-inaccurate versions of functions like
173 // sin and exp. This is controlled in clang by -fcuda-approx-transcendentals.
175 // device_functions.hpp uses __USE_FAST_MATH__ for a different purpose (fast vs.
176 // slow divides), so we need to scope our define carefully here.
177 #pragma push_macro("__USE_FAST_MATH__")
178 #if defined(__CLANG_CUDA_APPROX_TRANSCENDENTALS__)
179 #define __USE_FAST_MATH__ 1
181 #include "math_functions.hpp"
182 #pragma pop_macro("__USE_FAST_MATH__")
184 #include "math_functions_dbl_ptx3.hpp"
185 #pragma pop_macro("__forceinline__")
187 // Pull in host-only functions that are only available when neither
188 // __CUDACC__ nor __CUDABE__ are defined.
189 #undef __MATH_FUNCTIONS_HPP__
191 #include "math_functions.hpp"
192 // Alas, additional overloads for these functions are hard to get to.
193 // Considering that we only need these overloads for a few functions,
194 // we can provide them here.
195 static inline float rsqrt(float __a) { return rsqrtf(__a); }
196 static inline float rcbrt(float __a) { return rcbrtf(__a); }
197 static inline float sinpi(float __a) { return sinpif(__a); }
198 static inline float cospi(float __a) { return cospif(__a); }
199 static inline void sincospi(float __a, float *__b, float *__c) {
200 return sincospif(__a, __b, __c);
202 static inline float erfcinv(float __a) { return erfcinvf(__a); }
203 static inline float normcdfinv(float __a) { return normcdfinvf(__a); }
204 static inline float normcdf(float __a) { return normcdff(__a); }
205 static inline float erfcx(float __a) { return erfcxf(__a); }
207 // For some reason single-argument variant is not always declared by
208 // CUDA headers. Alas, device_functions.hpp included below needs it.
209 static inline __device__ void __brkpt(int __c) { __brkpt(); }
211 // Now include *.hpp with definitions of various GPU functions. Alas,
212 // a lot of thins get declared/defined with __host__ attribute which
213 // we don't want and we have to define it out. We also have to include
214 // {device,math}_functions.hpp again in order to extract the other
215 // branch of #if/else inside.
220 #undef __DEVICE_FUNCTIONS_HPP__
221 #include "device_atomic_functions.hpp"
222 #include "device_functions.hpp"
223 #include "sm_20_atomic_functions.hpp"
224 #include "sm_20_intrinsics.hpp"
225 #include "sm_32_atomic_functions.hpp"
227 // Don't include sm_30_intrinsics.h and sm_32_intrinsics.h. These define the
228 // __shfl and __ldg intrinsics using inline (volatile) asm, but we want to
229 // define them using builtins so that the optimizer can reason about and across
230 // these instructions. In particular, using intrinsics for ldg gets us the
231 // [addr+imm] addressing mode, which, although it doesn't actually exist in the
232 // hardware, seems to generate faster machine code because ptxas can more easily
233 // reason about our code.
235 #if CUDA_VERSION >= 8000
236 #include "sm_60_atomic_functions.hpp"
237 #include "sm_61_intrinsics.hpp"
240 #undef __MATH_FUNCTIONS_HPP__
242 // math_functions.hpp defines ::signbit as a __host__ __device__ function. This
243 // conflicts with libstdc++'s constexpr ::signbit, so we have to rename
244 // math_function.hpp's ::signbit. It's guarded by #undef signbit, but that's
245 // conditional on __GNUC__. :)
246 #pragma push_macro("signbit")
247 #pragma push_macro("__GNUC__")
249 #define signbit __ignored_cuda_signbit
250 #include "math_functions.hpp"
251 #pragma pop_macro("__GNUC__")
252 #pragma pop_macro("signbit")
254 #pragma pop_macro("__host__")
256 #include "texture_indirect_functions.h"
258 // Restore state of __CUDA_ARCH__ and __THROW we had on entry.
259 #pragma pop_macro("__CUDA_ARCH__")
260 #pragma pop_macro("__THROW")
262 // Set up compiler macros expected to be seen during compilation.
267 // Device-side CUDA system calls.
268 // http://docs.nvidia.com/cuda/ptx-writers-guide-to-interoperability/index.html#system-calls
269 // We need these declarations and wrappers for device-side
270 // malloc/free/printf calls to work without relying on
271 // -fcuda-disable-target-call-checks option.
272 __device__ int vprintf(const char *, const char *);
273 __device__ void free(void *) __attribute((nothrow));
274 __device__ void *malloc(size_t) __attribute((nothrow)) __attribute__((malloc));
275 __device__ void __assertfail(const char *__message, const char *__file,
276 unsigned __line, const char *__function,
277 size_t __charSize) __attribute__((noreturn));
279 // In order for standard assert() macro on linux to work we need to
280 // provide device-side __assert_fail()
281 __device__ static inline void __assert_fail(const char *__message,
282 const char *__file, unsigned __line,
283 const char *__function) {
284 __assertfail(__message, __file, __line, __function, sizeof(char));
287 // Clang will convert printf into vprintf, but we still need
288 // device-side declaration for it.
289 __device__ int printf(const char *, ...);
292 // We also need device-side std::malloc and std::free.
294 __device__ static inline void free(void *__ptr) { ::free(__ptr); }
295 __device__ static inline void *malloc(size_t __size) {
296 return ::malloc(__size);
300 // Out-of-line implementations from __clang_cuda_builtin_vars.h. These need to
301 // come after we've pulled in the definition of uint3 and dim3.
303 __device__ inline __cuda_builtin_threadIdx_t::operator uint3() const {
311 __device__ inline __cuda_builtin_blockIdx_t::operator uint3() const {
319 __device__ inline __cuda_builtin_blockDim_t::operator dim3() const {
320 return dim3(x, y, z);
323 __device__ inline __cuda_builtin_gridDim_t::operator dim3() const {
324 return dim3(x, y, z);
327 #include <__clang_cuda_cmath.h>
328 #include <__clang_cuda_intrinsics.h>
329 #include <__clang_cuda_complex_builtins.h>
331 // curand_mtgp32_kernel helpfully redeclares blockDim and threadIdx in host
332 // mode, giving them their "proper" types of dim3 and uint3. This is
333 // incompatible with the types we give in __clang_cuda_builtin_vars.h. As as
334 // hack, force-include the header (nvcc doesn't include it by default) but
335 // redefine dim3 and uint3 to our builtin types. (Thankfully dim3 and uint3 are
336 // only used here for the redeclarations of blockDim and threadIdx.)
337 #pragma push_macro("dim3")
338 #pragma push_macro("uint3")
339 #define dim3 __cuda_builtin_blockDim_t
340 #define uint3 __cuda_builtin_threadIdx_t
341 #include "curand_mtgp32_kernel.h"
342 #pragma pop_macro("dim3")
343 #pragma pop_macro("uint3")
344 #pragma pop_macro("__USE_FAST_MATH__")
347 #endif // __CLANG_CUDA_RUNTIME_WRAPPER_H__