HIP: Heterogenous-computing Interface for Portability
hip_cooperative_groups_helper.h
Go to the documentation of this file.
1 /*
2 Copyright (c) 2015 - 2023 Advanced Micro Devices, Inc. All rights reserved.
3 
4 Permission is hereby granted, free of charge, to any person obtaining a copy
5 of this software and associated documentation files (the "Software"), to deal
6 in the Software without restriction, including without limitation the rights
7 to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
8 copies of the Software, and to permit persons to whom the Software is
9 furnished to do so, subject to the following conditions:
10 
11 The above copyright notice and this permission notice shall be included in
12 all copies or substantial portions of the Software.
13 
14 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
17 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
19 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
20 THE SOFTWARE.
21 */
22 
31 #ifndef HIP_INCLUDE_HIP_AMD_DETAIL_HIP_COOPERATIVE_GROUPS_HELPER_H
32 #define HIP_INCLUDE_HIP_AMD_DETAIL_HIP_COOPERATIVE_GROUPS_HELPER_H
33 
34 #if __cplusplus
35 #if !defined(__HIPCC_RTC__)
36 #include <hip/amd_detail/amd_hip_runtime.h> // threadId, blockId
37 #include <hip/amd_detail/amd_device_functions.h>
38 #endif
39 #if !defined(__align__)
40 #define __align__(x) __attribute__((aligned(x)))
41 #endif
42 
43 #if !defined(__CG_QUALIFIER__)
44 #define __CG_QUALIFIER__ __device__ __forceinline__
45 #endif
46 
47 #if !defined(__CG_STATIC_QUALIFIER__)
48 #define __CG_STATIC_QUALIFIER__ __device__ static __forceinline__
49 #endif
50 
51 #if !defined(_CG_STATIC_CONST_DECL_)
52 #define _CG_STATIC_CONST_DECL_ static constexpr
53 #endif
54 
55 #if __AMDGCN_WAVEFRONT_SIZE == 32
56 using lane_mask = unsigned int;
57 #else
58 using lane_mask = unsigned long long int;
59 #endif
60 
61 namespace cooperative_groups {
62 
63 /* Global scope */
64 template <unsigned int size>
65 using is_power_of_2 = std::integral_constant<bool, (size & (size - 1)) == 0>;
66 
67 template <unsigned int size>
68 using is_valid_wavefront = std::integral_constant<bool, (size <= __AMDGCN_WAVEFRONT_SIZE)>;
69 
70 template <unsigned int size>
71 using is_valid_tile_size =
72  std::integral_constant<bool, is_power_of_2<size>::value && is_valid_wavefront<size>::value>;
73 
74 template <typename T>
75 using is_valid_type =
76  std::integral_constant<bool, std::is_integral<T>::value || std::is_floating_point<T>::value>;
77 
78 namespace internal {
79 
85 typedef enum {
86  cg_invalid,
87  cg_multi_grid,
88  cg_grid,
89  cg_workgroup,
90  cg_tiled_group,
91  cg_coalesced_group
92 } group_type;
111 namespace multi_grid {
112 
113 __CG_STATIC_QUALIFIER__ uint32_t num_grids() {
114  return static_cast<uint32_t>(__ockl_multi_grid_num_grids()); }
115 
116 __CG_STATIC_QUALIFIER__ uint32_t grid_rank() {
117  return static_cast<uint32_t>(__ockl_multi_grid_grid_rank()); }
118 
119 __CG_STATIC_QUALIFIER__ uint32_t size() { return static_cast<uint32_t>(__ockl_multi_grid_size()); }
120 
121 __CG_STATIC_QUALIFIER__ uint32_t thread_rank() {
122  return static_cast<uint32_t>(__ockl_multi_grid_thread_rank()); }
123 
124 __CG_STATIC_QUALIFIER__ bool is_valid() { return static_cast<bool>(__ockl_multi_grid_is_valid()); }
125 
126 __CG_STATIC_QUALIFIER__ void sync() { __ockl_multi_grid_sync(); }
127 
128 } // namespace multi_grid
129 
134 namespace grid {
135 
136 __CG_STATIC_QUALIFIER__ uint32_t size() {
137  return static_cast<uint32_t>((blockDim.z * gridDim.z) * (blockDim.y * gridDim.y) *
138  (blockDim.x * gridDim.x));
139 }
140 
141 __CG_STATIC_QUALIFIER__ uint32_t thread_rank() {
142  // Compute global id of the workgroup to which the current thread belongs to
143  uint32_t blkIdx = static_cast<uint32_t>((blockIdx.z * gridDim.y * gridDim.x) +
144  (blockIdx.y * gridDim.x) + (blockIdx.x));
145 
146  // Compute total number of threads being passed to reach current workgroup
147  // within grid
148  uint32_t num_threads_till_current_workgroup =
149  static_cast<uint32_t>(blkIdx * (blockDim.x * blockDim.y * blockDim.z));
150 
151  // Compute thread local rank within current workgroup
152  uint32_t local_thread_rank = static_cast<uint32_t>((threadIdx.z * blockDim.y * blockDim.x) +
153  (threadIdx.y * blockDim.x) + (threadIdx.x));
154 
155  return (num_threads_till_current_workgroup + local_thread_rank);
156 }
157 
158 __CG_STATIC_QUALIFIER__ bool is_valid() { return static_cast<bool>(__ockl_grid_is_valid()); }
159 
160 __CG_STATIC_QUALIFIER__ void sync() { __ockl_grid_sync(); }
161 
162 } // namespace grid
163 
169 namespace workgroup {
170 
171 __CG_STATIC_QUALIFIER__ dim3 group_index() {
172  return (dim3(static_cast<uint32_t>(blockIdx.x), static_cast<uint32_t>(blockIdx.y),
173  static_cast<uint32_t>(blockIdx.z)));
174 }
175 
176 __CG_STATIC_QUALIFIER__ dim3 thread_index() {
177  return (dim3(static_cast<uint32_t>(threadIdx.x), static_cast<uint32_t>(threadIdx.y),
178  static_cast<uint32_t>(threadIdx.z)));
179 }
180 
181 __CG_STATIC_QUALIFIER__ uint32_t size() {
182  return (static_cast<uint32_t>(blockDim.x * blockDim.y * blockDim.z));
183 }
184 
185 __CG_STATIC_QUALIFIER__ uint32_t thread_rank() {
186  return (static_cast<uint32_t>((threadIdx.z * blockDim.y * blockDim.x) +
187  (threadIdx.y * blockDim.x) + (threadIdx.x)));
188 }
189 
190 __CG_STATIC_QUALIFIER__ bool is_valid() {
191  return true;
192 }
193 
194 __CG_STATIC_QUALIFIER__ void sync() { __syncthreads(); }
195 
196 __CG_STATIC_QUALIFIER__ dim3 block_dim() {
197  return (dim3(static_cast<uint32_t>(blockDim.x), static_cast<uint32_t>(blockDim.y),
198  static_cast<uint32_t>(blockDim.z)));
199 }
200 
201 } // namespace workgroup
202 
203 namespace tiled_group {
204 
205 // enforce ordering for memory intructions
206 __CG_STATIC_QUALIFIER__ void sync() { __builtin_amdgcn_fence(__ATOMIC_ACQ_REL, "agent"); }
207 
208 } // namespace tiled_group
209 
210 namespace coalesced_group {
211 
212 // enforce ordering for memory intructions
213 __CG_STATIC_QUALIFIER__ void sync() { __builtin_amdgcn_fence(__ATOMIC_ACQ_REL, "agent"); }
214 
215 // Masked bit count
216 //
217 // For each thread, this function returns the number of active threads which
218 // have i-th bit of x set and come before the current thread.
219 __CG_STATIC_QUALIFIER__ unsigned int masked_bit_count(lane_mask x, unsigned int add = 0) {
220  unsigned int counter=0;
221  #if __AMDGCN_WAVEFRONT_SIZE == 32
222  counter = __builtin_amdgcn_mbcnt_lo(x, add);
223  #else
224  counter = __builtin_amdgcn_mbcnt_lo(static_cast<lane_mask>(x), add);
225  counter = __builtin_amdgcn_mbcnt_hi(static_cast<lane_mask>(x >> 32), counter);
226  #endif
227 
228  return counter;
229 }
230 
231 } // namespace coalesced_group
232 
233 
234 } // namespace internal
235 
236 } // namespace cooperative_groups
241 #endif // __cplusplus
242 #endif // HIP_INCLUDE_HIP_AMD_DETAIL_HIP_COOPERATIVE_GROUPS_HELPER_H