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
|
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef CXX11_SRC_FIXEDPOINT_MATMATPRODUCT_H_
#define CXX11_SRC_FIXEDPOINT_MATMATPRODUCT_H_
namespace Eigen {
namespace internal {
// Accumulate the product of 2 QInt8 inputs on 32 bits to prevent
// overflows
template<> struct scalar_product_traits<QInt8, QInt8>
{
enum {
Defined = 1
};
typedef QInt32 ReturnType;
};
// Accumulate the product of 2 QInt16 inputs on 32 bits to prevent
// overflows
template <>
struct scalar_product_traits<QInt16, QInt16> {
enum { Defined = 1 };
typedef QInt32 ReturnType;
};
// Accumulate the product of QInt8 inputs with QUint8 inputs on 32 bits
// to prevent overflows
template<> struct scalar_product_traits<QInt8, QUInt8>
{
enum {
Defined = 1
};
typedef QInt32 ReturnType;
};
// Description of the product implementation. It's pretty simple now since
// nothing is vectorized yet.
// This definition tackle the case where both lhs and rhs are encoded using
// signed 8bit integers
#ifndef EIGEN_USE_OPTIMIZED_INT8_INT8_MAT_MAT_PRODUCT
template<bool _ConjLhs, bool _ConjRhs>
class gebp_traits<QInt8, QInt8, _ConjLhs, _ConjRhs>
{
public:
typedef QInt8 LhsScalar;
typedef QInt8 RhsScalar;
typedef QInt32 ResScalar;
enum {
// register block size along the M and N directions
// One for the current implementation
nr = 1,
mr = 1,
// Progress made at each iteration of the product loop
// also 1 for the current implementation
LhsProgress = 1,
RhsProgress = 1
};
};
// The signed 8bit Mat-Mat product itself.
template<typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs>
struct gebp_kernel<QInt8, QInt8, Index, DataMapper, mr, nr, ConjugateLhs, ConjugateRhs>
{
EIGEN_DONT_INLINE
void operator()(const DataMapper& res, const QInt8* blockA, const QInt8* blockB,
Index rows, Index depth, Index cols, QInt32 alpha,
Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0);
};
template<typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs>
EIGEN_DONT_INLINE
void gebp_kernel<QInt8, QInt8, Index, DataMapper, mr, nr, ConjugateLhs, ConjugateRhs>
::operator()(const DataMapper& res, const QInt8* blockA, const QInt8* blockB,
Index rows, Index depth, Index cols, QInt32 alpha,
Index strideA, Index strideB, Index offsetA, Index offsetB)
{
EIGEN_STATIC_ASSERT(!ConjugateLhs, YOU_MADE_A_PROGRAMMING_MISTAKE);
EIGEN_STATIC_ASSERT(!ConjugateRhs, YOU_MADE_A_PROGRAMMING_MISTAKE);
eigen_assert(alpha.value == 1);
eigen_assert(strideA == -1);
eigen_assert(strideB == -1);
eigen_assert(offsetA == 0);
eigen_assert(offsetB == 0);
eigen_assert(rows > 0);
eigen_assert(cols > 0);
eigen_assert(depth > 0);
eigen_assert(blockA);
eigen_assert(blockB);
for (Index j = 0; j < cols; ++j) {
Index startB = j * depth;
for (Index i = 0; i < rows; ++i) {
Index startA = i * depth;
for (Index k = 0; k < depth; ++k) {
res(i, j) += blockA[startA + k] * blockB[startB + k];
}
}
}
}
#endif
// This definition tackle the case where the lhs is encoded using signed 8bit
// integers and the rhs using unsigned 8bit integers.
#ifndef EIGEN_USE_OPTIMIZED_INT8_UINT8_MAT_MAT_PRODUCT
template<bool _ConjLhs, bool _ConjRhs>
class gebp_traits<QInt8, QUInt8, _ConjLhs, _ConjRhs>
{
public:
typedef QInt8 LhsScalar;
typedef QUInt8 RhsScalar;
typedef QInt32 ResScalar;
enum {
// register block size along the M and N directions
// One for the current implementation
nr = 1,
mr = 1,
// Progress made at each iteration of the product loop
// also 1 for the current implementation
LhsProgress = 1,
RhsProgress = 1
};
};
// Mat-Mat product of a signed 8bit lhs with an unsigned 8bit rhs
template<typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs>
struct gebp_kernel<QInt8, QUInt8, Index, DataMapper, mr, nr, ConjugateLhs, ConjugateRhs>
{
EIGEN_DONT_INLINE
void operator()(const DataMapper& res, const QInt8* blockA, const QUInt8* blockB,
Index rows, Index depth, Index cols, QInt32 alpha,
Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0);
};
template<typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs>
EIGEN_DONT_INLINE
void gebp_kernel<QInt8, QUInt8, Index, DataMapper, mr, nr, ConjugateLhs, ConjugateRhs>
::operator()(const DataMapper& res, const QInt8* blockA, const QUInt8* blockB,
Index rows, Index depth, Index cols, QInt32 alpha,
Index strideA, Index strideB, Index offsetA, Index offsetB)
{
EIGEN_STATIC_ASSERT(!ConjugateLhs, YOU_MADE_A_PROGRAMMING_MISTAKE);
EIGEN_STATIC_ASSERT(!ConjugateRhs, YOU_MADE_A_PROGRAMMING_MISTAKE);
eigen_assert(alpha.value == 1);
eigen_assert(strideA == -1);
eigen_assert(strideB == -1);
eigen_assert(offsetA == 0);
eigen_assert(offsetB == 0);
eigen_assert(rows > 0);
eigen_assert(cols > 0);
eigen_assert(depth > 0);
eigen_assert(blockA);
eigen_assert(blockB);
for (Index j = 0; j < cols; ++j) {
Index startB = j * depth;
for (Index i = 0; i < rows; ++i) {
Index startA = i * depth;
for (Index k = 0; k < depth; ++k) {
res(i, j) += blockA[startA + k] * blockB[startB + k];
}
}
}
}
#endif
// This definition tackle the case where the khs is encoded using unsigned 8bit
// integers and the rhs using signed 8bit integers.
#ifndef EIGEN_USE_OPTIMIZED_UINT8_INT8_MAT_MAT_PRODUCT
template<bool _ConjLhs, bool _ConjRhs>
class gebp_traits<QUInt8, QInt8, _ConjLhs, _ConjRhs>
{
public:
typedef QUInt8 LhsScalar;
typedef QInt8 RhsScalar;
typedef QInt32 ResScalar;
enum {
// register block size along the M and N directions
// One for the current implementation
nr = 1,
mr = 1,
// Progress made at each iteration of the product loop
// also 1 for the current implementation
LhsProgress = 1,
RhsProgress = 1
};
};
// Mat-Mat product of an unsigned 8bit lhs with a signed 8bit rhs
template<typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs>
struct gebp_kernel<QUInt8, QInt8, Index, DataMapper, mr, nr, ConjugateLhs, ConjugateRhs>
{
EIGEN_DONT_INLINE
void operator()(const DataMapper& res, const QUInt8* blockA, const QInt8* blockB,
Index rows, Index depth, Index cols, QInt32 alpha,
Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0);
};
template<typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs>
EIGEN_DONT_INLINE
void gebp_kernel<QUInt8, QInt8, Index, DataMapper, mr, nr, ConjugateLhs, ConjugateRhs>
::operator()(const DataMapper& res, const QUInt8* blockA, const QInt8* blockB,
Index rows, Index depth, Index cols, QInt32 alpha,
Index strideA, Index strideB, Index offsetA, Index offsetB)
{
EIGEN_STATIC_ASSERT(!ConjugateLhs, YOU_MADE_A_PROGRAMMING_MISTAKE);
EIGEN_STATIC_ASSERT(!ConjugateRhs, YOU_MADE_A_PROGRAMMING_MISTAKE);
eigen_assert(alpha.value == 1);
eigen_assert(strideA == -1);
eigen_assert(strideB == -1);
eigen_assert(offsetA == 0);
eigen_assert(offsetB == 0);
eigen_assert(rows > 0);
eigen_assert(cols > 0);
eigen_assert(depth > 0);
eigen_assert(blockA);
eigen_assert(blockB);
for (Index j = 0; j < cols; ++j) {
Index startB = j * depth;
for (Index i = 0; i < rows; ++i) {
Index startA = i * depth;
for (Index k = 0; k < depth; ++k) {
res(i, j) += blockA[startA + k] * blockB[startB + k];
}
}
}
}
#endif
#ifndef EIGEN_USE_OPTIMIZED_INT16_INT16_MAT_MAT_PRODUCT
template <bool _ConjLhs, bool _ConjRhs>
class gebp_traits<QInt16, QInt16, _ConjLhs, _ConjRhs> {
public:
typedef QInt16 LhsScalar;
typedef QInt16 RhsScalar;
typedef QInt32 ResScalar;
enum {
// register block size along the M and N directions
// One for the current implementation
nr = 1,
mr = 1,
// Progress made at each iteration of the product loop
// also 1 for the current implementation
LhsProgress = 1,
RhsProgress = 1
};
};
// The signed 16bit Mat-Mat product itself.
template <typename Index, typename DataMapper, int mr, int nr,
bool ConjugateLhs, bool ConjugateRhs>
struct gebp_kernel<QInt16, QInt16, Index, DataMapper, mr, nr, ConjugateLhs,
ConjugateRhs> {
EIGEN_DONT_INLINE
void operator()(const DataMapper& res, const QInt16* blockA,
const QInt16* blockB, Index rows, Index depth, Index cols,
QInt32 alpha, Index strideA = -1, Index strideB = -1,
Index offsetA = 0, Index offsetB = 0);
};
template <typename Index, typename DataMapper, int mr, int nr,
bool ConjugateLhs, bool ConjugateRhs>
EIGEN_DONT_INLINE void gebp_kernel<QInt16, QInt16, Index, DataMapper, mr, nr,
ConjugateLhs, ConjugateRhs>::
operator()(const DataMapper& res, const QInt16* blockA, const QInt16* blockB,
Index rows, Index depth, Index cols, QInt32 alpha, Index strideA,
Index strideB, Index offsetA, Index offsetB) {
EIGEN_STATIC_ASSERT(!ConjugateLhs, YOU_MADE_A_PROGRAMMING_MISTAKE);
EIGEN_STATIC_ASSERT(!ConjugateRhs, YOU_MADE_A_PROGRAMMING_MISTAKE);
eigen_assert(alpha.value == 1);
eigen_assert(strideA == -1);
eigen_assert(strideB == -1);
eigen_assert(offsetA == 0);
eigen_assert(offsetB == 0);
eigen_assert(rows > 0);
eigen_assert(cols > 0);
eigen_assert(depth > 0);
eigen_assert(blockA);
eigen_assert(blockB);
for (Index j = 0; j < cols; ++j) {
Index startB = j * depth;
for (Index i = 0; i < rows; ++i) {
Index startA = i * depth;
for (Index k = 0; k < depth; ++k) {
res(i, j) += blockA[startA + k] * blockB[startB + k];
}
}
}
}
#endif
} // namespace internal
} // namespace Eigen
#endif // CXX11_SRC_FIXEDPOINT_MATMATPRODUCT_H_
|
