1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
|
//
// MIT License
// Copyright (c) 2018 Jonathan R. Madsen
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED
// "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT
// LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR
// PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
// HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
// ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
// WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
//
// ---------------------------------------------------------------
//
// PTL CUDA implementation
//
//
//============================================================================//
#include "sum.hh"
#define PRINT_HERE(extra) \
printf("> %s@'%s':%i %s\n", __FUNCTION__, __FILE__, __LINE__, extra)
//============================================================================//
// gridDim: This variable contains the dimensions of the grid.
// blockIdx: This variable contains the block index within the grid.
// blockDim: This variable and contains the dimensions of the block.
// threadIdx: This variable contains the thread index within the block.
//============================================================================//
//
// efficient reduction
// https://developer.download.nvidia.com/assets/cuda/files/reduction.pdf
//
//============================================================================//
template <unsigned int blockSize, typename _Tp>
__device__ void
warpReduce(volatile _Tp* _data, unsigned int tid)
{
if(blockSize >= 64)
_data[tid] += _data[tid + 32];
if(blockSize >= 32)
_data[tid] += _data[tid + 16];
if(blockSize >= 16)
_data[tid] += _data[tid + 8];
if(blockSize >= 8)
_data[tid] += _data[tid + 4];
if(blockSize >= 4)
_data[tid] += _data[tid + 2];
if(blockSize >= 2)
_data[tid] += _data[tid + 1];
}
//----------------------------------------------------------------------------//
template <unsigned int blockSize, typename _Tp>
__global__ void
reduce(_Tp* _idata, _Tp* _odata, unsigned int n)
{
extern __shared__ _Tp _data[];
unsigned int tid = threadIdx.x;
unsigned int i = (2 * blockSize) * blockIdx.x + tid;
unsigned int gridSize = 2 * blockSize * gridDim.x;
_data[tid] = 0;
while(i < n)
{
_data[tid] += _idata[i] + _idata[i + blockSize];
i += gridSize;
}
__syncthreads();
if(blockSize >= 512)
{
if(tid < 256)
{
_data[tid] += _data[tid + 256];
}
__syncthreads();
}
if(blockSize >= 256)
{
if(tid < 128)
{
_data[tid] += _data[tid + 128];
}
__syncthreads();
}
if(blockSize >= 128)
{
if(tid < 64)
{
_data[tid] += _data[tid + 64];
}
__syncthreads();
}
if(tid < 32)
warpReduce<blockSize, _Tp>(_data, tid);
if(tid == 0)
_odata[blockIdx.x] = _data[0];
}
//----------------------------------------------------------------------------//
template <typename _Tp>
void
compute_reduction(int threads, _Tp* _idata, _Tp* _odata, int dimGrid, int dimBlock,
int smemSize, cudaStream_t stream)
{
cudaStreamSynchronize(stream);
CUDA_CHECK_LAST_ERROR();
switch(threads)
{
case 512:
reduce<512, _Tp>
<<<dimGrid, dimBlock, smemSize, stream>>>(_idata, _odata, threads);
break;
case 256:
reduce<256, _Tp>
<<<dimGrid, dimBlock, smemSize, stream>>>(_idata, _odata, threads);
break;
case 128:
reduce<128, _Tp>
<<<dimGrid, dimBlock, smemSize, stream>>>(_idata, _odata, threads);
break;
case 64:
reduce<64, _Tp>
<<<dimGrid, dimBlock, smemSize, stream>>>(_idata, _odata, threads);
break;
case 32:
reduce<32, _Tp>
<<<dimGrid, dimBlock, smemSize, stream>>>(_idata, _odata, threads);
break;
case 16:
reduce<16, _Tp>
<<<dimGrid, dimBlock, smemSize, stream>>>(_idata, _odata, threads);
break;
case 8:
reduce<8, _Tp>
<<<dimGrid, dimBlock, smemSize, stream>>>(_idata, _odata, threads);
break;
case 4:
reduce<4, _Tp>
<<<dimGrid, dimBlock, smemSize, stream>>>(_idata, _odata, threads);
break;
case 2:
reduce<2, _Tp>
<<<dimGrid, dimBlock, smemSize, stream>>>(_idata, _odata, threads);
break;
case 1:
reduce<1, _Tp>
<<<dimGrid, dimBlock, smemSize, stream>>>(_idata, _odata, threads);
break;
}
CUDA_CHECK_LAST_ERROR();
cudaStreamSynchronize(stream);
CUDA_CHECK_LAST_ERROR();
}
//============================================================================//
template <typename _Tp>
void
call_compute_reduction(int64_t& _i, uint64_t& _offset, int nthreads, _Tp* _idata,
_Tp* _odata, int dimGrid, int dimBlock, int smemSize,
cudaStream_t stream)
{
// assumes nthreads < cuda_max_threads_per_block()
compute_reduction(nthreads, _idata + _offset, _odata + _offset, dimGrid, dimBlock,
smemSize, stream);
_i -= nthreads;
_offset += nthreads;
}
//============================================================================//
float
compute_sum_host(aligned_ptr<float>& data, cudaStream_t stream, bool with_thrust,
float* buffer)
{
float _sum;
if(with_thrust)
{
cudaStreamSynchronize(stream);
CUDA_CHECK_LAST_ERROR();
_sum = thrust::reduce(thrust::system::cuda::par.on(stream), data.ptr,
data.ptr + data.size, 0.0f, thrust::plus<float>());
CUDA_CHECK_LAST_ERROR();
cudaStreamSynchronize(stream);
CUDA_CHECK_LAST_ERROR();
}
else
{
// PRINT_HERE("");
// PRINT_HERE(std::string(std::string("size : ") +
// std::to_string(data.size)).c_str());
// PRINT_HERE(std::string(std::string("padding : ") +
// std::to_string(data.padding)).c_str());
// PRINT_HERE(std::string(std::string("storage : ") +
// std::to_string(data.storage_size)).c_str());
if(data.size < 1 || data.storage_size < 1)
return 0.0f;
int64_t remain = data.size;
uint64_t offset = 0;
int smemSize = cuda_shared_memory_per_block();
int dimGrid = cuda_multi_processor_count();
int dimBlock = cuda_max_threads_per_block();
float* _idata = data.ptr;
float* _odata = buffer;
async_gpu_memset<float>(_odata, data.storage_size, stream);
CUDA_CHECK_LAST_ERROR();
while(remain > 0)
{
for(const auto& itr : { 512, 256, 128, 64, 32, 16, 8, 4, 2, 1 })
{
if(remain >= itr)
{
call_compute_reduction(remain, offset, itr, _idata, _odata, dimGrid,
dimBlock, smemSize, stream);
break;
}
}
}
cudaMemcpyAsync(&_sum, _odata, 1 * sizeof(float), cudaMemcpyDeviceToHost, stream);
CUDA_CHECK_LAST_ERROR();
cudaDeviceSynchronize();
CUDA_CHECK_LAST_ERROR();
}
return _sum;
}
//============================================================================//
|
