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
|
/******************************************************************************
* Copyright (c) Intel Corporation - All rights reserved. *
* This file is part of the LIBXSMM library. *
* *
* For information on the license, see the LICENSE file. *
* Further information: https://github.com/hfp/libxsmm/ *
* SPDX-License-Identifier: BSD-3-Clause *
******************************************************************************/
/* Hans Pabst (Intel Corp.)
******************************************************************************/
#include "magazine.h"
#if !defined(__EIGEN) && 0
# define __EIGEN
#endif
#if defined(__EIGEN)
# if !defined(EIGEN_DONT_PARALLELIZE)
# define EIGEN_DONT_PARALLELIZE
# endif
# if defined(EIGEN_USE_MKL_ALL)
# undef EIGEN_USE_MKL_ALL
# endif
# include <Eigen/Dense>
# if defined(__EIGEN_TIMER)
# include <bench/BenchTimer.h>
# endif
#endif
#include <memory>
#include <cstdlib>
#if defined(__EIGEN)
template<bool pad> struct stride_helper {
stride_helper(int pad_a, int pad_b, int pad_c): a(pad_a), b(pad_b), c(pad_c) {}
/* dynamic strides may slow-down also if lda == m, etc. */
Eigen::OuterStride<Eigen::Dynamic> a, b, c;
};
template<> struct stride_helper<false> {
stride_helper(...) {}
Eigen::OuterStride<0> a, b, c;
};
#endif
int main(int argc, char* argv[])
{
#if defined(__EIGEN)
typedef TYPE T;
typedef Eigen::Matrix<T,Eigen::Dynamic,Eigen::Dynamic,Eigen::ColMajor> matrix_type;
/* batch-size is used to stream matrix-operands from memory */
const int batchsize = (1 < argc ? atoi(argv[1]) : 0/*auto*/);
/* default: M, N, and K are 13, 5, and 7 respectively */
const int m = (2 < argc ? atoi(argv[2]) : 13);
const int n = (3 < argc ? atoi(argv[3]) : 5);
const int k = (4 < argc ? atoi(argv[4]) : 7);
/* leading dimensions are used to each pad (row-major!) */
const int lda = (5 < argc ? (m < atoi(argv[5]) ? atoi(argv[5]) : m) : static_cast<int>(((sizeof(T) * m + PAD - 1) & ~(PAD - 1)) / sizeof(T)));
const int ldb = (6 < argc ? (k < atoi(argv[6]) ? atoi(argv[6]) : k) : static_cast<int>(((sizeof(T) * k + PAD - 1) & ~(PAD - 1)) / sizeof(T)));
const int ldc = (7 < argc ? (m < atoi(argv[7]) ? atoi(argv[7]) : m) : static_cast<int>(((sizeof(T) * m + PAD - 1) & ~(PAD - 1)) / sizeof(T)));
/* Eigen specifies leading dimensions per "outer stride" */
stride_helper<(sizeof(T)<PAD)> stride(lda, ldb, ldc);
#if 0
const char transa = 'n', transb = 'n';
#endif
const T alpha = 1, beta = 1;
/* calculate matrix sizes incl. padded elements */
const size_t na = ((sizeof(T) * lda * k + PAD - 1) & ~(PAD - 1)) / sizeof(T);
const size_t nb = ((sizeof(T) * ldb * n + PAD - 1) & ~(PAD - 1)) / sizeof(T);
const size_t nc = ((sizeof(T) * ldc * n + PAD - 1) & ~(PAD - 1)) / sizeof(T);
/* calculate default batch-size to hit work-set size of approx. 2 GB */
const int size = (0 >= batchsize ? static_cast<int>((2ULL << 30/*2 GB*/) / (sizeof(T) * (na + nb + nc))) : batchsize);
#if defined(SHUFFLE)
const size_t shuffle = libxsmm_shuffle((unsigned int)size);
#endif
size_t sa = sizeof(T) * na * size + PAD - 1;
size_t sb = sizeof(T) * nb * size + PAD - 1;
size_t sc = sizeof(T) * nc * size + PAD - 1;
/* allocate A, B, and C matrix buffers */
void *const va = malloc(sa), *const vb = malloc(sb), *const vc = malloc(sc), *wa = va, *wb = vb, *wc = vc;
/* align memory according to PAD */
#if defined(PAD) && (1 < (PAD))
T *const pa = static_cast<T*>(std::align(PAD, sa - PAD + 1, wa, sa));
T *const pb = static_cast<T*>(std::align(PAD, sb - PAD + 1, wb, sb));
T *const pc = static_cast<T*>(std::align(PAD, sc - PAD + 1, wc, sc));
#else
T *const pa = static_cast<T*>(wa);
T *const pb = static_cast<T*>(wb);
T *const pc = static_cast<T*>(wc);
#endif
const double scale = 1.0 / size;
double duration = 0;
#if defined(__EIGEN_TIMER)
Eigen::BenchTimer timer;
#endif
/* initialize data according to touch-first policy */
#if defined(_OPENMP)
# pragma omp parallel for
#endif
for (int i = 0; i < size; ++i) {
#if defined(SHUFFLE)
const int j = (i * shuffle) % size;
#else
const int j = i;
#endif
init(25 + i, pa + j * na, m, k, lda, scale);
init(75 + i, pb + j * nb, k, n, ldb, scale);
if (0 != beta) { /* no need to initialize for beta=0 */
init(42 + i, pc + j * nc, m, n, ldc, scale);
}
}
#if defined(_OPENMP)
# pragma omp parallel
#endif
{
#if defined(_OPENMP)
# pragma omp single
#endif
#if defined(__EIGEN_TIMER)
timer.start();
#else
duration = seconds();
#endif
#if defined(_OPENMP)
# pragma omp for
#endif
for (int i = 0; i < size; ++i) {
#if defined(SHUFFLE)
const int j = (i * shuffle) % size;
#else
const int j = i;
#endif
/* using "matrix_type" instead of "auto" induces an unnecessary copy */
const auto a = matrix_type::Map/*Aligned*/(pa + STREAM_A(j * na), m, k, stride.a);
const auto b = matrix_type::Map/*Aligned*/(pb + STREAM_B(j * nb), k, n, stride.b);
auto c = matrix_type::Map/*Aligned*/(pc + STREAM_C(j * nc), m, n, stride.c);
/**
* Expression templates attempt to delay evaluation until the sequence point
* is reached, or an "expression object" goes out of scope and hence must
* materialize the effect. Ideally, a complex expression is mapped to the
* best possible implementation, e.g., c = alpha * a * b + beta * c may be
* mapped to GEMM or definitely omits alpha*a in case of alpha=1, or similar
* for special cases for beta=0 and beta=1. However, to not rely on an ideal
* transformation a *manually specialized* expression is written for, e.g.,
* alpha=1 and beta=1 (c += a * b) or tweaked manually ("noalias").
* NOTE: changing alpha or beta from above may not have an effect
* depending on what is selected below (expression).
*/
#if 0 /* alpha=1 anyway */
c.noalias() = alpha * a * b + beta * c;
#elif 0
(void)alpha; /* unused */
c.noalias() = a * b + beta * c;
#elif 0 /* beta=0 */
(void)alpha; /* unused */
(void)beta; /* unused */
c.noalias() = a * b;
#else /* beta=1 */
(void)alpha; /* unused */
(void)beta; /* unused */
c.noalias() += a * b;
#endif
}
}
#if defined(__EIGEN_TIMER)
timer.stop();
duration = timer.total();
#else
duration = seconds() - duration;
#endif
if (0 < duration) {
const double gflops = 2.0 * m * n * k * 1E-9;
printf("%.1f GFLOPS/s\n", gflops / duration * size);
printf("%.1f ms\n", 1000.0 * duration);
}
{ /* calculate checksum */
double check = 0;
for (int i = 0; i < size; ++i) {
const double cn = norm(pc + STREAM_C(i * nc), m, n, ldc);
if (check < cn) check = cn;
}
printf("\n%f (check)\n", check);
}
free(va);
free(vb);
free(vc);
return EXIT_SUCCESS;
#else
(void)argc; /* unused */
(void)argv; /* unused */
return EXIT_FAILURE;
#endif
}
|
