File size: 14,763 Bytes
dffe123
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.

// This file is modified from the ficus (https://github.com/vpisarev/ficus/blob/master/runtime/ficus/impl/gemm.impl.h).
// Here is the original license:
/*
    This file is a part of ficus language project.
    See ficus/LICENSE for the licensing terms
*/

#include "../../precomp.hpp"
#include "fast_gemm.hpp"

#define CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
#include "fast_gemm_kernels.simd.hpp"
#include "layers/cpu_kernels/fast_gemm_kernels.simd_declarations.hpp"
#undef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
#include "fast_gemm_kernels.default.hpp"

namespace cv { namespace dnn {

size_t fastGemmPackBSize(size_t N, size_t K, const FastGemmOpt &opt) {
#if CV_TRY_NEON
    if (opt.use_neon) {
        return static_cast<size_t>(opt_NEON::fastGemmPackBSize(N, K));
    } else
#endif
#if CV_TRY_AVX2
    if (opt.use_avx2) {
        return static_cast<size_t>(opt_AVX2::fastGemmPackBSize(N, K));
    } else
#endif
#if CV_TRY_AVX
    if (opt.use_avx) {
        return static_cast<size_t>(opt_AVX::fastGemmPackBSize(N, K));
    } else
#endif
#if CV_TRY_LASX
    if (opt.use_lasx) {
        return static_cast<size_t>(opt_LASX::fastGemmPackBSize(N, K));
    } else
#endif
    {
        return static_cast<size_t>(cpu_baseline::fastGemmPackBSize(N, K));
    }
}

void fastGemmPackB(const Mat &B, std::vector<float> &packed_B, bool trans, FastGemmOpt &opt) {
    CV_CheckTypeEQ(B.type(), CV_32F, "fastGemmPackB: only float32 is supported for now");

    auto B_shape = shape(B);
    int batch = total(B_shape, 0, B_shape.size() - 2),
        K = B_shape[B_shape.size() - 2], N = B_shape.back(), ldb0 = N, ldb1 = 1;
    if (trans) {
        std::swap(K, N);
        std::swap(ldb0, ldb1);
    }

    const auto *b = B.ptr<const char>();
    int esz = B.elemSize();

#if CV_TRY_NEON
    if (opt.use_neon) {
        int size_packed_B = opt_NEON::fastGemmPackBSize(N, K);
        packed_B.resize(size_packed_B * batch);
        auto *packed_b = (char*)packed_B.data();
        for (int i = 0; i < batch; i++) {
            opt_NEON::fastGemmPackBKernel(b, packed_b, N, K, ldb0, ldb1, esz);
            b += N * K * esz;
            packed_b += size_packed_B * esz;
        }
    } else
#endif
#if CV_TRY_AVX2
    if (opt.use_avx2) {
        int size_packed_B = opt_AVX2::fastGemmPackBSize(N, K);
        packed_B.resize(size_packed_B * batch);
        auto *packed_b = (char*)packed_B.data();
        for (int i = 0; i < batch; i++) {
            opt_AVX2::fastGemmPackBKernel(b, packed_b, N, K, ldb0, ldb1, esz);
            b += N * K * esz;
            packed_b += size_packed_B * esz;
        }
    } else
#endif
#if CV_TRY_AVX
    if (opt.use_avx) {
        int size_packed_B = opt_AVX::fastGemmPackBSize(N, K);
        packed_B.resize(size_packed_B * batch);
        auto *packed_b = (char*)packed_B.data();
        for (int i = 0; i < batch; i++) {
            opt_AVX::fastGemmPackBKernel(b, packed_b, N, K, ldb0, ldb1, esz);
            b += N * K * esz;
            packed_b += size_packed_B * esz;
        }
    } else
#endif
#if CV_TRY_LASX
    if (opt.use_lasx) {
        int size_packed_B = opt_LASX::fastGemmPackBSize(N, K);
        packed_B.resize(size_packed_B * batch);
        auto *packed_b = (char*)packed_B.data();
        for (int i = 0; i < batch; i++) {
            opt_LASX::fastGemmPackBKernel(b, packed_b, N, K, ldb0, ldb1, esz);
            b += N * K * esz;
            packed_b += size_packed_B * esz;
        }
    } else
#endif
    {
        int size_packed_B = cpu_baseline::fastGemmPackBSize(N, K);
        packed_B.resize(size_packed_B * batch);
        auto *packed_b = (char*)packed_B.data();
        for (int i = 0; i < batch; i++) {
            cpu_baseline::fastGemmPackBKernel(b, packed_b, N, K, ldb0, ldb1, esz);
            b += N * K * esz;
            packed_b += size_packed_B * esz;
        }
    }
}

void fastGemmPackB(bool trans, size_t N, size_t K, const float *B, size_t ldb, float *packed_B, const FastGemmOpt &opt) {
    size_t ldb0 = ldb, ldb1 = 1;
    if (trans) {
        std::swap(K, N);
        std::swap(ldb0, ldb1);
    }

    const auto &b = (const char *)B;
    auto *packed_b = (char *)packed_B;

#if CV_TRY_NEON
    if (opt.use_neon) {
        opt_NEON::fastGemmPackBKernel(b, packed_b, N, K, ldb0, ldb1, sizeof(float));
    } else
#endif
#if CV_TRY_AVX2
    if (opt.use_avx2) {
        opt_AVX2::fastGemmPackBKernel(b, packed_b, N, K, ldb0, ldb1, sizeof(float));
    } else
#endif
#if CV_TRY_AVX
    if (opt.use_avx) {
        opt_AVX::fastGemmPackBKernel(b, packed_b, N, K, ldb0, ldb1, sizeof(float));
    } else
#endif
#if CV_TRY_LASX
    if (opt.use_lasx) {
        opt_LASX::fastGemmPackBKernel(b, packed_b, N, K, ldb0, ldb1, sizeof(float));
    } else
#endif
    {
        cpu_baseline::fastGemmPackBKernel(b, packed_b, N, K, ldb0, ldb1, sizeof(float));
    }
}

static void fast_gemm_thin(float alpha, float beta, int M, int N, int K,
                           const char *a_, int lda0, int lda1,
                           const char *b_, int ldb,
                           char *c_, int ldc, bool multi_thread) {
    const float* a = (const float*)a_;

    auto fn = [&](const Range &r) {
        for(int start = r.start ; start < r.end; start++ ) {
            float* c_i = (float*)c_ + start * ldc;
            if (beta == 0.f)
                for(int j = 0; j < N; j++ ) c_i[j] = 0.f;
            else if (beta != 1.f)
                for(int j = 0; j < N; j++ ) c_i[j] *= beta;
            for(int k = 0; k < K; k++ ) {
                const float* b_k = (const float*)b_ + k * ldb;
                float aval = alpha * a[start * lda0 + k * lda1];
                for(int j = 0; j < N; j++ )
                    c_i[j] += aval * b_k[j];
            }
        }
    };

    if (multi_thread) {
        int total = M; // outer loops
        int cost_per_thread = static_cast<int>(K * N); // inner loops
        double nstripes = (size_t)total * cost_per_thread * (1 / 1024.0);
        parallel_for_(Range(0, total), fn, nstripes);
    } else {
        fn(Range(0, M));
    }
}

void fastGemm(bool trans_a, int M, int N, int K,
              float alpha, const float *A, int lda,
              const float *packed_B, float beta,
              float *C, int ldc, FastGemmOpt &opt) {
    const char *a = (const char *)A;
    const char *packed_b = (const char *)packed_B;
    char *c = (char *)C;

    int lda0 = lda, lda1 = 1;
    if (trans_a) {
        std::swap(lda0, lda1);
    }

#if CV_TRY_NEON
    if (opt.use_neon) {
        opt_NEON::fastGemmKernel(M, N, K, alpha, a, lda0, lda1, packed_b, beta, c, ldc, sizeof(float), opt.multi_thread);
    } else
#endif
#if CV_TRY_AVX2
    if (opt.use_avx2) {
        opt_AVX2::fastGemmKernel(M, N, K, alpha, a, lda0, lda1, packed_b, beta, c, ldc, sizeof(float), opt.multi_thread);
    } else
#endif
#if CV_TRY_AVX
    if (opt.use_avx) {
        opt_AVX::fastGemmKernel(M, N, K, alpha, a, lda0, lda1, packed_b, beta, c, ldc, sizeof(float), opt.multi_thread);
    } else
#endif
#if CV_TRY_LASX
    if (opt.use_lasx) {
        opt_LASX::fastGemmKernel(M, N, K, alpha, a, lda0, lda1, packed_b, beta, c, ldc, sizeof(float), opt.multi_thread);
    } else
#endif
    {
        cpu_baseline::fastGemmKernel(M, N, K, alpha, a, lda0, lda1, packed_b, beta, c, ldc, sizeof(float), opt.multi_thread);
    }
}

void fastGemm(bool trans_a, bool trans_b, int ma, int na, int mb, int nb,
              float alpha, const float *A, int lda0, int lda1, const float *B, int ldb0, int ldb1,
              float beta, float *C, int ldc, FastGemmOpt &opt) {
    const char *a = (const char *)A;
    const char *b = (const char *)B;
    char *c = (char *)C;

    int M = trans_a ? na : ma;
    int N = trans_b ? mb : nb;
    int K = trans_a ? ma : na;

    if (trans_a) {
        std::swap(lda0, lda1);
    }
    if (trans_b) {
        std::swap(ldb0, ldb1);
    }

    if (!trans_b && ldb1 == 1 && (M <= 4 || (uint64_t)M * N * K <= 10000)) {
        return fast_gemm_thin(alpha, beta, M, N, K, a, lda0, lda1, b, ldb0, c, ldc, opt.multi_thread);
    }

#if CV_TRY_NEON
    if (opt.use_neon) {
        opt_NEON::fastGemmKernel(M, N, K, alpha, a, lda0, lda1,
                                 b, ldb0, ldb1, beta,
                                 c, ldc, sizeof(float), opt.multi_thread);
    } else
#endif
#if CV_TRY_AVX2
    if (opt.use_avx2) {
        opt_AVX2::fastGemmKernel(M, N, K, alpha, a, lda0, lda1,
                                 b, ldb0, ldb1, beta,
                                 c, ldc, sizeof(float), opt.multi_thread);
    } else
#endif
#if CV_TRY_AVX
    if (opt.use_avx) {
        opt_AVX::fastGemmKernel(M, N, K, alpha, a, lda0, lda1,
                                 b, ldb0, ldb1, beta,
                                 c, ldc, sizeof(float), opt.multi_thread);
    } else
#endif
#if CV_TRY_LASX
    if (opt.use_lasx) {
        opt_LASX::fastGemmKernel(M, N, K, alpha, a, lda0, lda1,
                                 b, ldb0, ldb1, beta,
                                 c, ldc, sizeof(float), opt.multi_thread);
    } else
#endif
    {
        cpu_baseline::fastGemmKernel(M, N, K, alpha, a, lda0, lda1,
                                     b, ldb0, ldb1, beta,
                                     c, ldc, sizeof(float), opt.multi_thread);
    }
}

void fastGemm(bool trans_a, bool trans_b,
              float alpha, const Mat &A, const Mat &B,
              float beta, Mat &C, FastGemmOpt &opt) {
    CV_CheckTypeEQ(A.type(), CV_32F, "DNN/fastGemm: only support float32 for now");
    CV_CheckTypeEQ(A.type(), B.type(), "DNN/fastGemm: A and B should have the same type");
    CV_CheckTypeEQ(B.type(), C.type(), "DNN/fastGemm: B and C should have the same type");

    const auto shape_a = shape(A);
    CV_CheckEQ(shape_a.size(), static_cast<size_t>(2), "DNN/fastGemm: A must be 2-dimensional");
    const auto shape_b = shape(B);
    CV_CheckEQ(shape_b.size(), static_cast<size_t>(2), "DNN/fastGemm: B must be 2-dimensional");
    const auto shape_c = shape(C);
    CV_CheckEQ(shape_c.size(), static_cast<size_t>(2), "DNN/fastGemm: C must be 2-dimensional");

    int ma = shape_a[0], na = shape_a[1];
    int mb = shape_b[0], nb = shape_b[1];

    int lda0 = na, lda1 = 1, ldb0 = nb, ldb1 = 1, ldc = shape_c[1];

    const float *a = A.ptr<const float>();
    const float *b = B.ptr<const float>();
    float *c = C.ptr<float>();

    fastGemm(trans_a, trans_b, ma, na, mb, nb,
             alpha, a, lda0, lda1, b, ldb0, ldb1,
             beta, c, ldc, opt);
}

void fastGemmBatch(size_t batch, const size_t *A_offsets, const size_t *B_offsets, const size_t *C_offsets,
                   int M, int N, int K, float alpha, const float *A, int lda0, int lda1,
                   const float *B, int ldb0, int ldb1, float beta, float *C, int ldc, FastGemmOpt &opt) {
    const char *a = (const char *)A;
    const char *b = (const char *)B;
    char *c = (char *)C;

#if CV_TRY_NEON
    if (opt.use_neon) {
        opt_NEON::fastGemmBatchKernel(batch, A_offsets, B_offsets, C_offsets, M, N, K, alpha, a, lda0, lda1, b, ldb0, ldb1, beta, c, ldc, sizeof(float));
    } else
#endif
#if CV_TRY_AVX2
    if (opt.use_avx2) {
        opt_AVX2::fastGemmBatchKernel(batch, A_offsets, B_offsets, C_offsets, M, N, K, alpha, a, lda0, lda1, b, ldb0, ldb1, beta, c, ldc, sizeof(float));
    } else
#endif
#if CV_TRY_AVX
    if (opt.use_avx) {
        opt_AVX::fastGemmBatchKernel(batch, A_offsets, B_offsets, C_offsets, M, N, K, alpha, a, lda0, lda1, b, ldb0, ldb1, beta, c, ldc, sizeof(float));
    } else
#endif
#if CV_TRY_LASX
    if (opt.use_lasx) {
        opt_LASX::fastGemmBatchKernel(batch, A_offsets, B_offsets, C_offsets, M, N, K, alpha, a, lda0, lda1, b, ldb0, ldb1, beta, c, ldc, sizeof(float));
    } else
#endif
    {
        cpu_baseline::fastGemmBatchKernel(batch, A_offsets, B_offsets, C_offsets, M, N, K, alpha, a, lda0, lda1, b, ldb0, ldb1, beta, c, ldc, sizeof(float));
    }
}

void fastGemmBatch(size_t batch, const size_t *A_offsets, const size_t *packed_B_offsets, const size_t *C_offsets,
                   int M, int N, int K, float alpha, const float *A, int lda0, int lda1,
                   const float *packed_B, float beta, float *C, int ldc, FastGemmOpt &opt) {
    const char *a = (const char *)A;
    const char *b = (const char *)packed_B;
    char *c = (char *)C;

#if CV_TRY_NEON
    if (opt.use_neon) {
        opt_NEON::fastGemmBatchKernel(batch, A_offsets, packed_B_offsets, C_offsets, M, N, K, alpha, a, lda0, lda1, b, beta, c, ldc, sizeof(float));
    } else
#endif
#if CV_TRY_AVX2
    if (opt.use_avx2) {
        opt_AVX2::fastGemmBatchKernel(batch, A_offsets, packed_B_offsets, C_offsets, M, N, K, alpha, a, lda0, lda1, b, beta, c, ldc, sizeof(float));
    } else
#endif
#if CV_TRY_AVX
    if (opt.use_avx) {
        opt_AVX::fastGemmBatchKernel(batch, A_offsets, packed_B_offsets, C_offsets, M, N, K, alpha, a, lda0, lda1, b, beta, c, ldc, sizeof(float));
    } else
#endif
#if CV_TRY_LASX
    if (opt.use_lasx) {
        opt_LASX::fastGemmBatchKernel(batch, A_offsets, packed_B_offsets, C_offsets, M, N, K, alpha, a, lda0, lda1, b, beta, c, ldc, sizeof(float));
    } else
#endif
    {
        cpu_baseline::fastGemmBatchKernel(batch, A_offsets, packed_B_offsets, C_offsets, M, N, K, alpha, a, lda0, lda1, b, beta, c, ldc, sizeof(float));
    }
}

void fastGemmBatch(bool trans_a, bool trans_b,
                   float alpha, const Mat &A, const Mat &B,
                   float beta, Mat &C, FastGemmOpt &opt) {
    CV_CheckTypeEQ(A.type(), B.type(), "DNN/fastGemmBatch: A and B should have the same type");
    CV_CheckTypeEQ(B.type(), C.type(), "DNN/fastGemmBatch: B and C should have the same type");
    CV_CheckTypeEQ(A.type(), CV_32F, "DNN/fastGemmBatch: only support float32 for now");

    const auto shape_a = shape(A);
    const auto shape_b = shape(B);
    const auto shape_c = shape(C);
    CV_CheckGE(shape_a.size(), static_cast<size_t>(2), "DNN/fastGemmBatch: A must be n-dimensional (n >= 2)");
    CV_CheckGE(shape_b.size(), static_cast<size_t>(2), "DNN/fastGemmBatch: B must be n-dimensional (n >= 2)");

    const float *a = A.ptr<const float>();
    const float *b = B.ptr<const float>();
    float *c = C.ptr<float>();

    MatMulHelper helper;
    helper.compute(trans_a, trans_b, shape_a, shape_b, shape_c);

    fastGemmBatch(helper.batch, helper.A_offsets.data(), helper.B_offsets.data(), helper.C_offsets.data(),
                  helper.M, helper.N, helper.K, alpha, a, helper.lda0, helper.lda1, b, helper.ldb0,
                  helper.ldb1, beta, c, helper.ldc, opt);
}

}} // cv::dnn