contrib/llvm-project/clang/lib/Headers/__clang_cuda_builtin_vars.h

   1 /*===---- cuda_builtin_vars.h - CUDA built-in variables ---------------------===
   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 #ifndef __CUDA_BUILTIN_VARS_H
  11 #define __CUDA_BUILTIN_VARS_H
  12
  13 // Forward declares from vector_types.h.
  14 struct uint3;
  15 struct dim3;
  16
  17 // The file implements built-in CUDA variables using __declspec(property).
  18 // https://msdn.microsoft.com/en-us/library/yhfk0thd.aspx
  19 // All read accesses of built-in variable fields get converted into calls to a
  20 // getter function which in turn calls the appropriate builtin to fetch the
  21 // value.
  22 //
  23 // Example:
  24 //    int x = threadIdx.x;
  25 // IR output:
  26 //  %0 = call i32 @llvm.nvvm.read.ptx.sreg.tid.x() #3
  27 // PTX output:
  28 //  mov.u32     %r2, %tid.x;
  29
  30 #define __CUDA_DEVICE_BUILTIN(FIELD, INTRINSIC)                                \
  31   __declspec(property(get = __fetch_builtin_##FIELD)) unsigned int FIELD;      \
  32   static inline __attribute__((always_inline))                                 \
  33       __attribute__((device)) unsigned int __fetch_builtin_##FIELD(void) {     \
  34     return INTRINSIC;                                                          \
  35   }
  36
  37 #if __cplusplus >= 201103L
  38 #define __DELETE =delete
  39 #else
  40 #define __DELETE
  41 #endif
  42
  43 // Make sure nobody can create instances of the special variable types.  nvcc
  44 // also disallows taking address of special variables, so we disable address-of
  45 // operator as well.
  46 #define __CUDA_DISALLOW_BUILTINVAR_ACCESS(TypeName)                            \
  47   __attribute__((device)) TypeName() __DELETE;                                 \
  48   __attribute__((device)) TypeName(const TypeName &) __DELETE;                 \
  49   __attribute__((device)) void operator=(const TypeName &) const __DELETE;     \
  50   __attribute__((device)) TypeName *operator&() const __DELETE
  51
  52 struct __cuda_builtin_threadIdx_t {
  53   __CUDA_DEVICE_BUILTIN(x,__nvvm_read_ptx_sreg_tid_x());
  54   __CUDA_DEVICE_BUILTIN(y,__nvvm_read_ptx_sreg_tid_y());
  55   __CUDA_DEVICE_BUILTIN(z,__nvvm_read_ptx_sreg_tid_z());
  56   // threadIdx should be convertible to uint3 (in fact in nvcc, it *is* a
  57   // uint3).  This function is defined after we pull in vector_types.h.
  58   __attribute__((device)) operator uint3() const;
  59 private:
  60   __CUDA_DISALLOW_BUILTINVAR_ACCESS(__cuda_builtin_threadIdx_t);
  61 };
  62
  63 struct __cuda_builtin_blockIdx_t {
  64   __CUDA_DEVICE_BUILTIN(x,__nvvm_read_ptx_sreg_ctaid_x());
  65   __CUDA_DEVICE_BUILTIN(y,__nvvm_read_ptx_sreg_ctaid_y());
  66   __CUDA_DEVICE_BUILTIN(z,__nvvm_read_ptx_sreg_ctaid_z());
  67   // blockIdx should be convertible to uint3 (in fact in nvcc, it *is* a
  68   // uint3).  This function is defined after we pull in vector_types.h.
  69   __attribute__((device)) operator uint3() const;
  70 private:
  71   __CUDA_DISALLOW_BUILTINVAR_ACCESS(__cuda_builtin_blockIdx_t);
  72 };
  73
  74 struct __cuda_builtin_blockDim_t {
  75   __CUDA_DEVICE_BUILTIN(x,__nvvm_read_ptx_sreg_ntid_x());
  76   __CUDA_DEVICE_BUILTIN(y,__nvvm_read_ptx_sreg_ntid_y());
  77   __CUDA_DEVICE_BUILTIN(z,__nvvm_read_ptx_sreg_ntid_z());
  78   // blockDim should be convertible to dim3 (in fact in nvcc, it *is* a
  79   // dim3).  This function is defined after we pull in vector_types.h.
  80   __attribute__((device)) operator dim3() const;
  81 private:
  82   __CUDA_DISALLOW_BUILTINVAR_ACCESS(__cuda_builtin_blockDim_t);
  83 };
  84
  85 struct __cuda_builtin_gridDim_t {
  86   __CUDA_DEVICE_BUILTIN(x,__nvvm_read_ptx_sreg_nctaid_x());
  87   __CUDA_DEVICE_BUILTIN(y,__nvvm_read_ptx_sreg_nctaid_y());
  88   __CUDA_DEVICE_BUILTIN(z,__nvvm_read_ptx_sreg_nctaid_z());
  89   // gridDim should be convertible to dim3 (in fact in nvcc, it *is* a
  90   // dim3).  This function is defined after we pull in vector_types.h.
  91   __attribute__((device)) operator dim3() const;
  92 private:
  93   __CUDA_DISALLOW_BUILTINVAR_ACCESS(__cuda_builtin_gridDim_t);
  94 };
  95
  96 #define __CUDA_BUILTIN_VAR                                                     \
  97   extern const __attribute__((device)) __attribute__((weak))
  98 __CUDA_BUILTIN_VAR __cuda_builtin_threadIdx_t threadIdx;
  99 __CUDA_BUILTIN_VAR __cuda_builtin_blockIdx_t blockIdx;
 100 __CUDA_BUILTIN_VAR __cuda_builtin_blockDim_t blockDim;
 101 __CUDA_BUILTIN_VAR __cuda_builtin_gridDim_t gridDim;
 102
 103 // warpSize should translate to read of %WARP_SZ but there's currently no
 104 // builtin to do so. According to PTX v4.2 docs 'to date, all target
 105 // architectures have a WARP_SZ value of 32'.
 106 __attribute__((device)) const int warpSize = 32;
 107
 108 #undef __CUDA_DEVICE_BUILTIN
 109 #undef __CUDA_BUILTIN_VAR
 110 #undef __CUDA_DISALLOW_BUILTINVAR_ACCESS
 111
 112 #endif /* __CUDA_BUILTIN_VARS_H */