1 /*===---- cuda_builtin_vars.h - CUDA built-in variables ---------------------=== 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 #ifndef __CUDA_BUILTIN_VARS_H 25 #define __CUDA_BUILTIN_VARS_H 26 27 // Forward declares from vector_types.h. 28 struct uint3; 29 struct dim3; 30 31 // The file implements built-in CUDA variables using __declspec(property). 32 // https://msdn.microsoft.com/en-us/library/yhfk0thd.aspx 33 // All read accesses of built-in variable fields get converted into calls to a 34 // getter function which in turn calls the appropriate builtin to fetch the 35 // value. 36 // 37 // Example: 38 // int x = threadIdx.x; 39 // IR output: 40 // %0 = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() #3 41 // PTX output: 42 // mov.u32 %r2, %tid.x; 43 44 #define __CUDA_DEVICE_BUILTIN(FIELD, INTRINSIC) \ 45 __declspec(property(get = __fetch_builtin_##FIELD)) unsigned int FIELD; \ 46 static inline __attribute__((always_inline)) \ 47 __attribute__((device)) unsigned int __fetch_builtin_##FIELD(void) { \ 48 return INTRINSIC; \ 49 } 50 51 #if __cplusplus >= 201103L 52 #define __DELETE =delete 53 #else 54 #define __DELETE 55 #endif 56 57 // Make sure nobody can create instances of the special varible types. nvcc 58 // also disallows taking address of special variables, so we disable address-of 59 // operator as well. 60 #define __CUDA_DISALLOW_BUILTINVAR_ACCESS(TypeName) \ 61 __attribute__((device)) TypeName() __DELETE; \ 62 __attribute__((device)) TypeName(const TypeName &) __DELETE; \ 63 __attribute__((device)) void operator=(const TypeName &) const __DELETE; \ 64 __attribute__((device)) TypeName *operator&() const __DELETE 65 66 struct __cuda_builtin_threadIdx_t { 67 __CUDA_DEVICE_BUILTIN(x,__nvvm_read_ptx_sreg_tid_x()); 68 __CUDA_DEVICE_BUILTIN(y,__nvvm_read_ptx_sreg_tid_y()); 69 __CUDA_DEVICE_BUILTIN(z,__nvvm_read_ptx_sreg_tid_z()); 70 // threadIdx should be convertible to uint3 (in fact in nvcc, it *is* a 71 // uint3). This function is defined after we pull in vector_types.h. 72 __attribute__((device)) operator uint3() const; 73 private: 74 __CUDA_DISALLOW_BUILTINVAR_ACCESS(__cuda_builtin_threadIdx_t); 75 }; 76 77 struct __cuda_builtin_blockIdx_t { 78 __CUDA_DEVICE_BUILTIN(x,__nvvm_read_ptx_sreg_ctaid_x()); 79 __CUDA_DEVICE_BUILTIN(y,__nvvm_read_ptx_sreg_ctaid_y()); 80 __CUDA_DEVICE_BUILTIN(z,__nvvm_read_ptx_sreg_ctaid_z()); 81 // blockIdx should be convertible to uint3 (in fact in nvcc, it *is* a 82 // uint3). This function is defined after we pull in vector_types.h. 83 __attribute__((device)) operator uint3() const; 84 private: 85 __CUDA_DISALLOW_BUILTINVAR_ACCESS(__cuda_builtin_blockIdx_t); 86 }; 87 88 struct __cuda_builtin_blockDim_t { 89 __CUDA_DEVICE_BUILTIN(x,__nvvm_read_ptx_sreg_ntid_x()); 90 __CUDA_DEVICE_BUILTIN(y,__nvvm_read_ptx_sreg_ntid_y()); 91 __CUDA_DEVICE_BUILTIN(z,__nvvm_read_ptx_sreg_ntid_z()); 92 // blockDim should be convertible to dim3 (in fact in nvcc, it *is* a 93 // dim3). This function is defined after we pull in vector_types.h. 94 __attribute__((device)) operator dim3() const; 95 private: 96 __CUDA_DISALLOW_BUILTINVAR_ACCESS(__cuda_builtin_blockDim_t); 97 }; 98 99 struct __cuda_builtin_gridDim_t { 100 __CUDA_DEVICE_BUILTIN(x,__nvvm_read_ptx_sreg_nctaid_x()); 101 __CUDA_DEVICE_BUILTIN(y,__nvvm_read_ptx_sreg_nctaid_y()); 102 __CUDA_DEVICE_BUILTIN(z,__nvvm_read_ptx_sreg_nctaid_z()); 103 // gridDim should be convertible to dim3 (in fact in nvcc, it *is* a 104 // dim3). This function is defined after we pull in vector_types.h. 105 __attribute__((device)) operator dim3() const; 106 private: 107 __CUDA_DISALLOW_BUILTINVAR_ACCESS(__cuda_builtin_gridDim_t); 108 }; 109 110 #define __CUDA_BUILTIN_VAR \ 111 extern const __attribute__((device)) __attribute__((weak)) 112 __CUDA_BUILTIN_VAR __cuda_builtin_threadIdx_t threadIdx; 113 __CUDA_BUILTIN_VAR __cuda_builtin_blockIdx_t blockIdx; 114 __CUDA_BUILTIN_VAR __cuda_builtin_blockDim_t blockDim; 115 __CUDA_BUILTIN_VAR __cuda_builtin_gridDim_t gridDim; 116 117 // warpSize should translate to read of %WARP_SZ but there's currently no 118 // builtin to do so. According to PTX v4.2 docs 'to date, all target 119 // architectures have a WARP_SZ value of 32'. 120 __attribute__((device)) const int warpSize = 32; 121 122 #undef __CUDA_DEVICE_BUILTIN 123 #undef __CUDA_BUILTIN_VAR 124 #undef __CUDA_DISALLOW_BUILTINVAR_ACCESS 125 126 #endif /* __CUDA_BUILTIN_VARS_H */ 127