Matmul_benchmark/matmul.cpp at master · JosephChenHub/Matmul_benchmark · GitHub

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
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
#include <iostream>
#include <xmmintrin.h> // SSE
#include <immintrin.h> // AVX

#include "matmul.hpp"

static inline float _mm_reduce_add_ps(__m128 x128) {
    const __m128 x64 = _mm_add_ps(x128, _mm_movehl_ps(x128, x128));
    const __m128 x32 = _mm_add_ss(x64, _mm_shuffle_ps(x64, x64, 0x55));
    return _mm_cvtss_f32(x32);
}
static inline float _mm256_reduce_add_ps(__m256 x) {
    /* ( x3+x7, x2+x6, x1+x5, x0+x4 ) */
    const __m128 x128 = _mm_add_ps(_mm256_extractf128_ps(x, 1), _mm256_castps256_ps128(x));
    /* ( -, -, x1+x3+x5+x7, x0+x2+x4+x6 ) */
    const __m128 x64 = _mm_add_ps(x128, _mm_movehl_ps(x128, x128));
    /* ( -, -, -, x0+x1+x2+x3+x4+x5+x6+x7 ) */
    const __m128 x32 = _mm_add_ss(x64, _mm_shuffle_ps(x64, x64, 0x55));
    /* Conversion to float is a no-op on x86-64 */
    return _mm_cvtss_f32(x32);
}


template <typename Dtype>
void matmul_naive(const Matrix<Dtype>& A, const Matrix<Dtype>& B, Matrix<Dtype>& C) {
    assert (A.rows() == C.rows() && A.cols() == B.rows() && B.cols() == C.cols());
    for(int i = 0; i < C.rows(); ++i) {
        for(int j = 0; j < C.cols(); ++j) {
            Dtype tmp(0);
            for(int k = 0; k < A.cols(); ++k) {
                tmp += A[i][k] * B[k][j];
            }
            C[i][j] = tmp;
        }
    }
}


template <typename Dtype>
void matmul_openmp(const Matrix<Dtype>& A, const Matrix<Dtype>& B, Matrix<Dtype>& C) {
    assert (A.rows() == C.rows() && A.cols() == B.rows() && B.cols() == C.cols());
    int i, j, k;
    #pragma omp parallel for private(i, j)
    for(i = 0; i < C.rows(); ++i) {
        for(j = 0; j < C.cols(); ++j) {
            Dtype tmp(0);
            for(k = 0; k < A.cols(); ++k) {
                tmp += A[i][k] * B[k][j];
            }
            C[i][j] = tmp;
        }
    }
}


template <typename Dtype>
void matmul_trans(const Matrix<Dtype>& A, const Matrix<Dtype>& B, Matrix<Dtype>& C) {
    assert (A.rows() == C.rows() && A.cols() == B.cols() && B.rows() == C.cols());
    int i, j;
    for(i = 0; i < C.rows(); ++i) {
        for(j = 0; j < C.cols(); ++j) {
            Dtype tmp(0);
            for(int k = 0; k < A.cols(); ++k) {
                tmp += A[i][k] * B[j][k];
            }
            C[i][j] = tmp;
        }
    }
}

#define BLOCK_SIZE 32

template <typename Dtype>
void matmul_trans_block(const Matrix<Dtype>& A, const Matrix<Dtype>& B, Matrix<Dtype>& C) {
    assert (A.rows() == C.rows() && A.cols() == B.cols() && B.rows() == C.cols());
    int i, j, k;
    int ib, jb, kb;
    if (A.rows() < BLOCK_SIZE) { matmul_trans(A, B, C); return; }

    for(i = 0; i < A.rows(); i += BLOCK_SIZE) {
        for(j = 0; j < B.rows(); j += BLOCK_SIZE) {
            for(k = 0; k < A.cols(); k += BLOCK_SIZE) {
            // block matmul
	    for(ib = i; ib < i + BLOCK_SIZE ; ++ib) {
               for(jb = j; jb < j + BLOCK_SIZE ; ++jb) {
                  for(kb = k; kb < k + BLOCK_SIZE ; ++kb) {
                     C[ib][jb] += A[ib][kb] * B[jb][kb];
                  }
	       }
            }
            }
        }
    }

}


template <typename Dtype>
void matmul_cuda_naive(const Matrix<Dtype>& A, const Matrix<Dtype>& B, Matrix<Dtype>& C) {
    matmul_naive_kernel<Dtype>(
            reinterpret_cast<Dtype*>(A.gpu_data()),
            reinterpret_cast<Dtype*>(B.gpu_data()),
            reinterpret_cast<Dtype*>(C.gpu_data()),
            A.rows(), A.cols(), B.cols());
}

template <typename Dtype>
void matmul_cuda_tile(const Matrix<Dtype>& A, const Matrix<Dtype>& B, Matrix<Dtype>& C) {
    matmul_tile_kernel<Dtype>(
            reinterpret_cast<Dtype*>(A.gpu_data()),
            reinterpret_cast<Dtype*>(B.gpu_data()),
            reinterpret_cast<Dtype*>(C.gpu_data()),
            A.rows(), A.cols(), B.cols());
}
template <typename Dtype>
void matmul_cuda_unroll(const Matrix<Dtype>& A, const Matrix<Dtype>& B, Matrix<Dtype>& C) {
    matmul_unroll_kernel<Dtype>(
            reinterpret_cast<Dtype*>(A.gpu_data()),
            reinterpret_cast<Dtype*>(B.gpu_data()),
            reinterpret_cast<Dtype*>(C.gpu_data()),
            A.rows(), A.cols(), B.cols());
}

template <typename Dtype>
void matmul_cuda_comopt(const Matrix<Dtype>& A, const Matrix<Dtype>& B, Matrix<Dtype>& C) {
    matmul_comopt_kernel<Dtype>(
            reinterpret_cast<Dtype*>(A.gpu_data()),
            reinterpret_cast<Dtype*>(B.gpu_data()),
            reinterpret_cast<Dtype*>(C.gpu_data()),
            A.rows(), A.cols(), B.cols());
}

template <>
void matmul_cublas(const Matrix<float>& A, const Matrix<float>& B, Matrix<float>& C) {
    cublas_sgemm(reinterpret_cast<float*>(A.gpu_data()),
            reinterpret_cast<float*>(B.gpu_data()),
            reinterpret_cast<float*>(C.gpu_data()),
            A.rows(), A.cols(), B.cols());
}

//template <typename Dtype>
//void matmul_cuda_vectorize(const Matrix<Dtype>& A, const Matrix<Dtype>& B, Matrix<Dtype>& C) {
//    matmul_vectorize_kernel<Dtype>(
//            reinterpret_cast<Dtype*>(A.gpu_data()),
//            reinterpret_cast<Dtype*>(B.gpu_data()),
//            reinterpret_cast<Dtype*>(C.gpu_data()),
//            A.rows(), A.cols(), B.cols());
//}

template <>
void matmul_omp_sse(const Matrix<float>& A, const Matrix<float>& Bt, Matrix<float>& C) {
    // data must be aligned with 16 bytes
    int i, j, k;
    #pragma omp parallel for private(i, j)
    for(i = 0; i < C.rows(); ++i) {
	for(j = 0; j < C.cols(); ++j) {
	    __m128 partial_sum = _mm_setzero_ps();

            for(k = 0; k < A.cols() / 4; ++k) {
                auto a = _mm_load_ps(A.data() + i * A.cols() + 4 * k);
                auto b = _mm_load_ps(Bt.data() + j * Bt.cols() + 4 * k);
                auto c = _mm_mul_ps(a, b);
	        partial_sum = _mm_add_ps(partial_sum, c);
            }
	    C[i][j]  = _mm_reduce_add_ps(partial_sum);
	}
    }
}

template <>
void matmul_omp_avx(const Matrix<float>& A, const Matrix<float>& Bt, Matrix<float>& C) {
    // data must be aligned with 32 bytes
    int i, j, k;
    #pragma omp parallel for private(i, j)
    for(i = 0; i < C.rows(); ++i) {
	for(j = 0; j < C.cols(); ++j) {
	    auto partial_sum = _mm256_setzero_ps();
            for(k = 0; k < A.cols() / 8; ++k) {
                auto a = _mm256_load_ps(A.data() + i * A.cols() + 8 * k);
                auto b = _mm256_load_ps(Bt.data() + j * Bt.cols() + 8 * k);
                auto c = _mm256_mul_ps(a, b);
	        partial_sum = _mm256_add_ps(partial_sum, c);
            }
	    C[i][j] = _mm256_reduce_add_ps(partial_sum);
	}
    }
}

template <>
void matmul_omp_avx512(const Matrix<float>& A, const Matrix<float>& Bt, Matrix<float>& C) {
    // data must be aligned with 32 bytes
    int i, j, k;
    #pragma omp parallel for private(i, j)
    for(i = 0; i < C.rows(); ++i) {
	for(j = 0; j < C.cols(); ++j) {
	    __m512 partial_sum = _mm512_set1_ps(0.0);
            for(k = 0; k < A.cols() / 16; ++k) {
                auto a = _mm512_loadu_ps(A.data() + i * A.cols() + 16 * k);
                auto b = _mm512_loadu_ps(Bt.data() + j * Bt.cols() + 16 * k);
		partial_sum = _mm512_fmadd_ps(a, b, partial_sum);
            }
	    C[i][j] = _mm512_reduce_add_ps(partial_sum);
	}
    }
}

template <typename Dtype>
inline void _strassen_2x2(const Dtype* a, const int stride_a,
		const Dtype* b, const int stride_b,
		Dtype* c, const int stride_c) {
    Dtype p1 = (b[1] - b[stride_b + 1]) * a[0];
    Dtype p2 = (a[0] + a[1]) * b[stride_b + 1];
    Dtype p3 = (a[stride_a] + a[stride_a+1]) * b[0];
    Dtype p4 = (b[stride_b] - b[0]) * a[stride_a+1];
    Dtype p5 = (a[0] + a[stride_a+1]) * (b[0] + b[stride_b+1]);
    Dtype p6 = (a[1] - a[stride_a+1]) * (b[stride_b] + b[stride_b+1]);
    Dtype p7 = (a[0] - a[stride_a]) * (b[0] + b[1]);

    c[0] = p5 + p4 - p2 + p6;
    c[1] = p1 + p2;
    c[stride_c] = p3 + p4;
    c[stride_c+1] = p1 + p5 - p3 - p7;
}


template <typename Dtype>
void matmul_strassen(const Matrix<Dtype>& A,
		const Matrix<Dtype>& B,
		Matrix<Dtype>& C) {
     assert (A.rows() == A.cols());
     const int n = A.rows();
     assert ((n & (n-1)) == 0);
     if (n == 2) {
	_strassen_2x2(A.data(), A.stride(), B.data(), B.stride(), C.data(), C.stride());
	//n=1, C.data()[0] = A.data()[0] * B.data()[0];
	return;
     }
     const int half_n = n >> 1;
     // split
     Matrix<Dtype> A11(A, 0, 0, half_n, half_n);  // sub-matrix
     Matrix<Dtype> A12(A, half_n, 0, half_n, half_n);
     Matrix<Dtype> A21(A, 0, half_n, half_n, half_n);
     Matrix<Dtype> A22(A, half_n, half_n, half_n, half_n);

     Matrix<Dtype> B11(B, 0, 0, half_n, half_n);
     Matrix<Dtype> B12(B, half_n, 0, half_n, half_n);
     Matrix<Dtype> B21(B, 0, half_n, half_n, half_n);
     Matrix<Dtype> B22(B, half_n, half_n, half_n, half_n);


     Matrix<Dtype> P1 = B12 - B22;
     Matrix<Dtype> P2 = A11 + A12;
     Matrix<Dtype> P3 = A21 + A22;
     Matrix<Dtype> P4 = B21 - B11;
     Matrix<Dtype> P51 = A11 + A22;
     Matrix<Dtype> P52 = B11 + B22;
     Matrix<Dtype> P61 = A12 - A22;
     Matrix<Dtype> P62 = B21 + B22;
     Matrix<Dtype> P71 = A11 - A21;
     Matrix<Dtype> P72 = B11 + B12;


     matmul_strassen(A11, P1, P1);
     matmul_strassen(P2, B22, P2);
     matmul_strassen(P3, B11, P3);
     matmul_strassen(A22, P4, P4);
     matmul_strassen(P51, P52, P52);
     matmul_strassen(P61, P62, P62);
     matmul_strassen(P71, P72, P72);


     // merge
     int i, j;
     #pragma omp parallel for private(i, j)
     for(i = 0; i < half_n; ++i) {
        for(j = 0; j < half_n; ++j) {
           C[i][j] = P52[i][j] + P4[i][j] - P2[i][j] + P62[i][j];
        }
     }
     #pragma omp parallel for private(i, j)
     for(i = 0; i < half_n; ++i) {
        for(j = 0; j < half_n; ++j) {
           C[i][j+half_n] = P1[i][j] + P2[i][j];
        }
     }
     #pragma omp parallel for private(i, j)
     for(i = 0; i < half_n; ++i) {
        for(j = 0; j < half_n; ++j) {
           C[i+half_n][j] = P3[i][j] + P4[i][j];
        }
     }
     #pragma omp parallel for private(i, j)
     for(i = 0; i < half_n; ++i) {
        for(j = 0; j < half_n; ++j) {
           C[i+half_n][j+half_n] = P1[i][j] + P52[i][j] - P3[i][j] - P72[i][j];
        }
     }
}


template void matmul_naive(const Matrix<float>&, const Matrix<float>&, Matrix<float>& );
template void matmul_naive(const Matrix<double>&, const Matrix<double>&, Matrix<double>& );
template void matmul_naive(const Matrix<int>&, const Matrix<int>&, Matrix<int>& );

template void matmul_openmp(const Matrix<float>&, const Matrix<float>&, Matrix<float>& );
template void matmul_openmp(const Matrix<double>&, const Matrix<double>&, Matrix<double>& );
template void matmul_openmp(const Matrix<int>&, const Matrix<int>&, Matrix<int>& );

template void matmul_trans(const Matrix<float>&, const Matrix<float>&, Matrix<float>& );
template void matmul_trans(const Matrix<double>&, const Matrix<double>&, Matrix<double>& );
template void matmul_trans(const Matrix<int>&, const Matrix<int>&, Matrix<int>& );

template void matmul_trans_block(const Matrix<float>&, const Matrix<float>&, Matrix<float>& );
template void matmul_trans_block(const Matrix<double>&, const Matrix<double>&, Matrix<double>& );
template void matmul_trans_block(const Matrix<int>&, const Matrix<int>&, Matrix<int>& );

template void matmul_strassen(const Matrix<float>&, const Matrix<float>&, Matrix<float>& );
template void matmul_strassen(const Matrix<double>&, const Matrix<double>&, Matrix<double>& );
template void matmul_strassen(const Matrix<int>&, const Matrix<int>&, Matrix<int>& );

template void matmul_cuda_naive(const Matrix<float>&, const Matrix<float>&, Matrix<float>& );
template void matmul_cuda_naive(const Matrix<double>&, const Matrix<double>&, Matrix<double>& );
template void matmul_cuda_naive(const Matrix<int>&, const Matrix<int>&, Matrix<int>& );

template void matmul_cuda_tile(const Matrix<float>&, const Matrix<float>&, Matrix<float>& );
template void matmul_cuda_tile(const Matrix<double>&, const Matrix<double>&, Matrix<double>& );
template void matmul_cuda_tile(const Matrix<int>&, const Matrix<int>&, Matrix<int>& );

template void matmul_cuda_unroll(const Matrix<float>&, const Matrix<float>&, Matrix<float>& );
template void matmul_cuda_unroll(const Matrix<double>&, const Matrix<double>&, Matrix<double>& );
template void matmul_cuda_unroll(const Matrix<int>&, const Matrix<int>&, Matrix<int>& );

template void matmul_cuda_comopt(const Matrix<float>&, const Matrix<float>&, Matrix<float>& );
template void matmul_cuda_comopt(const Matrix<double>&, const Matrix<double>&, Matrix<double>& );
template void matmul_cuda_comopt(const Matrix<int>&, const Matrix<int>&, Matrix<int>& );

//template void matmul_cuda_vectorize(const Matrix<float>&, const Matrix<float>&, Matrix<float>& );
//template void matmul_cuda_vectorize(const Matrix<double>&, const Matrix<double>&, Matrix<double>& );
//template void matmul_cuda_vectorize(const Matrix<int>&, const Matrix<int>&, Matrix<int>& );