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
|
#pragma once
#include <string>
#include <iostream>
#include <stdio.h>
#include "bytes_per_spmv.h"
#include "../timer.h"
#include <cusp/detail/device/spmv/coo_flat_k.h>
const char * BENCHMARK_OUTPUT_FILE_NAME = "benchmark_output.log";
template <typename HostMatrix, typename TestMatrix, typename TestKernel>
float check_spmv(HostMatrix& host_matrix, TestMatrix& test_matrix, TestKernel test_kernel)
{
typedef typename TestMatrix::index_type IndexType; // ASSUME same as HostMatrix::index_type
typedef typename TestMatrix::value_type ValueType; // ASSUME same as HostMatrix::value_type
typedef typename TestMatrix::memory_space MemorySpace;
const IndexType M = host_matrix.num_rows;
const IndexType N = host_matrix.num_cols;
// create host input (x) and output (y) vectors
cusp::array1d<ValueType,cusp::host_memory> host_x(N);
cusp::array1d<ValueType,cusp::host_memory> host_y(M);
//for(IndexType i = 0; i < N; i++) host_x[i] = (rand() % 21) - 10;
for(IndexType i = 0; i < N; i++) host_x[i] = (int(i % 21) - 10);
for(IndexType i = 0; i < M; i++) host_y[i] = 0;
// create test input (x) and output (y) vectors
cusp::array1d<ValueType, MemorySpace> test_x(host_x.begin(), host_x.end());
cusp::array1d<ValueType, MemorySpace> test_y(host_y.begin(), host_y.end());
// compute SpMV on host and device
cusp::multiply(host_matrix, host_x, host_y);
test_kernel(test_matrix, thrust::raw_pointer_cast(&test_x[0]), thrust::raw_pointer_cast(&test_y[0]));
// compare results
cusp::array1d<ValueType,cusp::host_memory> test_y_copy(test_y.begin(), test_y.end());
double error = l2_error(M, thrust::raw_pointer_cast(&test_y_copy[0]), thrust::raw_pointer_cast(&host_y[0]));
// if (error > 0.0001)
// {
// for(int i = 0; i < std::min<int>(N,256); i++)
// printf("host_x[%5d] = %10.8f\n", i, (float) host_x[i]);
//
// int limit = 256;
// for(int i = 0; i < M; i++)
// {
// if(std::abs(host_y[i] - test_y[i]) > 0.0)
// {
// printf("host_y[%5d] = %10.8f test_y[%5d] = %10.8f\n", i, (float) host_y[i], i, (float) test_y[i]);
//
// if(--limit <= 0)
// break;
// }
// }
// }
return error;
}
template <typename TestMatrix, typename TestKernel>
float time_spmv(TestMatrix& test_matrix, TestKernel test_spmv, double seconds = 3.0, size_t min_iterations = 100, size_t max_iterations = 500)
{
typedef typename TestMatrix::index_type IndexType; // ASSUME same as HostMatrix::index_type
typedef typename TestMatrix::value_type ValueType; // ASSUME same as HostMatrix::value_type
typedef typename TestMatrix::memory_space MemorySpace;
const IndexType M = test_matrix.num_rows;
const IndexType N = test_matrix.num_cols;
// create test input (x) and output (y) vectors
cusp::array1d<ValueType, MemorySpace> test_x(N);
cusp::array1d<ValueType, MemorySpace> test_y(M);
// warmup
timer time_one_iteration;
test_spmv(test_matrix, thrust::raw_pointer_cast(&test_x[0]), thrust::raw_pointer_cast(&test_y[0]));
cudaThreadSynchronize();
double estimated_time = time_one_iteration.seconds_elapsed();
// determine # of seconds dynamically
size_t num_iterations;
if (estimated_time == 0)
num_iterations = max_iterations;
else
num_iterations = std::min(max_iterations, std::max(min_iterations, (size_t) (seconds / estimated_time)) );
// time several SpMV iterations
timer t;
for(size_t i = 0; i < num_iterations; i++)
test_spmv(test_matrix, thrust::raw_pointer_cast(&test_x[0]), thrust::raw_pointer_cast(&test_y[0]));
cudaThreadSynchronize();
float sec_per_iteration = t.seconds_elapsed() / num_iterations;
return sec_per_iteration;
}
template <typename HostMatrix, typename TestMatrixOnHost, typename TestMatrixOnDevice, typename TestKernel>
void test_spmv(std::string kernel_name,
HostMatrix& host_matrix,
TestMatrixOnHost& test_matrix_on_host,
TestMatrixOnDevice& test_matrix_on_device,
TestKernel test_spmv)
{
float error = check_spmv(host_matrix, test_matrix_on_device, test_spmv);
float time = time_spmv( test_matrix_on_device, test_spmv);
float gbyte = bytes_per_spmv(test_matrix_on_host);
float GFLOPs = (time == 0) ? 0 : (2 * host_matrix.num_entries / time) / 1e9;
float GBYTEs = (time == 0) ? 0 : (gbyte / time) / 1e9;
printf("\t%-20s: %8.4f ms ( %5.2f GFLOP/s %5.1f GB/s) [L2 error %f]\n", kernel_name.c_str(), 1e3 * time, GFLOPs, GBYTEs, error);
//record results to file
FILE * fid = fopen(BENCHMARK_OUTPUT_FILE_NAME, "a");
fprintf(fid, "kernel=%s gflops=%f gbytes=%f msec=%f\n", kernel_name.c_str(), GFLOPs, GBYTEs, 1e3 * time);
fclose(fid);
}
/////////////////////////////////////////////////////
// These methods test specific formats and kernels //
/////////////////////////////////////////////////////
template <typename HostMatrix>
void test_coo(HostMatrix& host_matrix)
{
typedef typename HostMatrix::index_type IndexType;
typedef typename HostMatrix::value_type ValueType;
// convert HostMatrix to TestMatrix on host
cusp::coo_matrix<IndexType, ValueType, cusp::host_memory> test_matrix_on_host(host_matrix);
// transfer TestMatrix to device
typedef typename cusp::coo_matrix<IndexType, ValueType, cusp::device_memory> DeviceMatrix;
DeviceMatrix test_matrix_on_device(test_matrix_on_host);
test_spmv("coo_flat", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_coo_flat <DeviceMatrix,ValueType>);
test_spmv("coo_flat_tex", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_coo_flat_tex<DeviceMatrix,ValueType>);
// test_spmv("coo_flat_k", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_coo_flat_k <DeviceMatrix,ValueType>);
// test_spmv("coo_flat_k_tex", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_coo_flat_k_tex<DeviceMatrix,ValueType>);
}
template <typename HostMatrix>
void test_csr(HostMatrix& host_matrix)
{
typedef typename HostMatrix::index_type IndexType;
typedef typename HostMatrix::value_type ValueType;
// convert HostMatrix to TestMatrix on host
cusp::csr_matrix<IndexType, ValueType, cusp::host_memory> test_matrix_on_host(host_matrix);
// transfer csr_matrix to device
typedef typename cusp::csr_matrix<IndexType, ValueType, cusp::device_memory> DeviceMatrix;
DeviceMatrix test_matrix_on_device(test_matrix_on_host);
test_spmv("csr_scalar", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_csr_scalar <DeviceMatrix,ValueType>);
test_spmv("csr_scalar_tex", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_csr_scalar_tex<DeviceMatrix,ValueType>);
test_spmv("csr_vector", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_csr_vector <DeviceMatrix,ValueType>);
test_spmv("csr_vector_tex", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_csr_vector_tex<DeviceMatrix,ValueType>);
}
template <typename HostMatrix>
void test_dia(HostMatrix& host_matrix)
{
typedef typename HostMatrix::index_type IndexType;
typedef typename HostMatrix::value_type ValueType;
// convert HostMatrix to TestMatrix on host
cusp::dia_matrix<IndexType, ValueType, cusp::host_memory> test_matrix_on_host;
try
{
test_matrix_on_host = host_matrix;
}
catch (cusp::format_conversion_exception)
{
std::cout << "\tRefusing to convert to DIA format" << std::endl;
return;
}
// transfer TestMatrix to device
typedef typename cusp::dia_matrix<IndexType, ValueType, cusp::device_memory> DeviceMatrix;
DeviceMatrix test_matrix_on_device(test_matrix_on_host);
test_spmv("dia", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_dia <DeviceMatrix,ValueType>);
test_spmv("dia_tex", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_dia_tex<DeviceMatrix,ValueType>);
}
template <typename HostMatrix>
void test_ell(HostMatrix& host_matrix)
{
typedef typename HostMatrix::index_type IndexType;
typedef typename HostMatrix::value_type ValueType;
// convert HostMatrix to TestMatrix on host
cusp::ell_matrix<IndexType, ValueType, cusp::host_memory> test_matrix_on_host;
try
{
test_matrix_on_host = host_matrix;
}
catch (cusp::format_conversion_exception)
{
std::cout << "\tRefusing to convert to ELL format" << std::endl;
return;
}
// transfer TestMatrix to device
typedef typename cusp::ell_matrix<IndexType, ValueType, cusp::device_memory> DeviceMatrix;
DeviceMatrix test_matrix_on_device(test_matrix_on_host);
test_spmv("ell", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_ell <DeviceMatrix,ValueType>);
test_spmv("ell_tex", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_ell_tex<DeviceMatrix,ValueType>);
}
template <typename HostMatrix>
void test_hyb(HostMatrix& host_matrix)
{
typedef typename HostMatrix::index_type IndexType;
typedef typename HostMatrix::value_type ValueType;
// convert HostMatrix to TestMatrix on host
cusp::hyb_matrix<IndexType, ValueType, cusp::host_memory> test_matrix_on_host(host_matrix);
// transfer TestMatrix to device
typedef typename cusp::hyb_matrix<IndexType, ValueType, cusp::device_memory> DeviceMatrix;
DeviceMatrix test_matrix_on_device(test_matrix_on_host);
test_spmv("hyb", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_hyb <DeviceMatrix,ValueType>);
test_spmv("hyb_tex", host_matrix, test_matrix_on_host, test_matrix_on_device, cusp::detail::device::spmv_hyb_tex<DeviceMatrix,ValueType>);
}
|
