kernel_x86.h


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

/* Copyright 2019 Google LLC. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#ifndef TENSORFLOW_LITE_EXPERIMENTAL_RUY_KERNEL_X86_H_
#define TENSORFLOW_LITE_EXPERIMENTAL_RUY_KERNEL_X86_H_

#include <cstdint>

#include "common.h"
#include "internal_matrix.h"
#include "kernel_common.h"
#include "matrix.h"
#include "opt_set.h"
#include "path.h"
#include "platform.h"
#include "spec.h"
#include "tune.h"

namespace ruy {

#if RUY_PLATFORM(X86)
// TODO(b/147376783): SSE 4.2 and AVX-VNNI support is incomplete / placeholder.
// Optimization is not finished. In particular the dimensions of the kernel
// blocks can be changed as desired.
//
void Kernel8bitSse42(const KernelParams8bit<8, 8>& params);

template <typename DstScalar>
struct Kernel<Path::kSse42, std::int8_t, std::int8_t, DstScalar,
              BasicSpec<std::int32_t, DstScalar>> {
  Tuning tuning = Tuning::kAuto;
  using LhsLayout = FixedKernelLayout<Order::kColMajor, 4, 8>;
  using RhsLayout = FixedKernelLayout<Order::kColMajor, 4, 8>;
  explicit Kernel(Tuning tuning_) : tuning(tuning_) {}
  void Run(const PackedMatrix<std::int8_t>& lhs,
           const PackedMatrix<std::int8_t>& rhs,
           const BasicSpec<std::int32_t, DstScalar>& spec, int start_row,
           int start_col, int end_row, int end_col,
           Matrix<DstScalar>* dst) const {
    KernelParams8bit<LhsLayout::kCols, RhsLayout::kCols> params;
    MakeKernelParams8bit(lhs, rhs, spec, start_row, start_col, end_row, end_col,
                         dst, &params);
    Kernel8bitSse42(params);
  }
};

void KernelFloatSse42(const KernelParamsFloat<8, 8>& params);

template <>
struct Kernel<Path::kSse42, float, float, float, BasicSpec<float, float>> {
  Tuning tuning = Tuning::kAuto;
  using LhsLayout = FixedKernelLayout<Order::kRowMajor, 1, 8>;
  using RhsLayout = FixedKernelLayout<Order::kRowMajor, 1, 8>;
  explicit Kernel(Tuning tuning_) : tuning(tuning_) {}
  void Run(const PackedMatrix<float>& lhs, const PackedMatrix<float>& rhs,
           const BasicSpec<float, float>& spec, int start_row, int start_col,
           int end_row, int end_col, Matrix<float>* dst) const {
    KernelParamsFloat<LhsLayout::kCols, RhsLayout::kCols> params;
    MakeKernelParamsFloat(lhs, rhs, spec, start_row, start_col, end_row,
                          end_col, dst, &params);
    KernelFloatSse42(params);
  }
};

void Kernel8bitAvx512(const KernelParams8bit<16, 16>& params);
void Kernel8bitAvx512SingleCol(const KernelParams8bit<16, 16>& params);

template <typename DstScalar>
struct Kernel<Path::kAvx512, std::int8_t, std::int8_t, DstScalar,
              BasicSpec<std::int32_t, DstScalar>> {
  Tuning tuning = Tuning::kAuto;
  using LhsLayout = FixedKernelLayout<Order::kColMajor, 4, 16>;
  using RhsLayout = FixedKernelLayout<Order::kColMajor, 4, 16>;
  explicit Kernel(Tuning tuning_) : tuning(tuning_) {}
  void Run(const PackedMatrix<std::int8_t>& lhs,
           const PackedMatrix<std::int8_t>& rhs,
           const BasicSpec<std::int32_t, DstScalar>& spec, int start_row,
           int start_col, int end_row, int end_col,
           Matrix<DstScalar>* dst) const {
    KernelParams8bit<LhsLayout::kCols, RhsLayout::kCols> params;
    MakeKernelParams8bit(lhs, rhs, spec, start_row, start_col, end_row, end_col,
                         dst, &params);
    if (dst->layout.cols == 1) {
      Kernel8bitAvx512SingleCol(params);
    } else {
      Kernel8bitAvx512(params);
    }
  }
};

void KernelFloatAvx512(const KernelParamsFloat<16, 16>& params);
void KernelFloatAvx512SingleCol(const KernelParamsFloat<16, 16>& param);

template <>
struct Kernel<Path::kAvx512, float, float, float, BasicSpec<float, float>> {
  Tuning tuning = Tuning::kAuto;
  using LhsLayout = FixedKernelLayout<Order::kRowMajor, 1, 16>;
  using RhsLayout = FixedKernelLayout<Order::kRowMajor, 1, 16>;
  explicit Kernel(Tuning tuning_) : tuning(tuning_) {}
  void Run(const PackedMatrix<float>& lhs, const PackedMatrix<float>& rhs,
           const BasicSpec<float, float>& spec, int start_row, int start_col,
           int end_row, int end_col, Matrix<float>* dst) const {
    KernelParamsFloat<LhsLayout::kCols, RhsLayout::kCols> params;
    MakeKernelParamsFloat(lhs, rhs, spec, start_row, start_col, end_row,
                          end_col, dst, &params);
    if (dst->layout.cols == 1) {
      KernelFloatAvx512SingleCol(params);
    } else {
      KernelFloatAvx512(params);
    }
  }
};

void Kernel8bitAvx2(const KernelParams8bit<8, 8>& params);
void Kernel8bitAvx2SingleCol(const KernelParams8bit<8, 8>& params);

template <typename DstScalar>
struct Kernel<Path::kAvx2, std::int8_t, std::int8_t, DstScalar,
              BasicSpec<std::int32_t, DstScalar>> {
  Tuning tuning = Tuning::kAuto;
  using LhsLayout = FixedKernelLayout<Order::kColMajor, 4, 8>;
  using RhsLayout = FixedKernelLayout<Order::kColMajor, 4, 8>;
  explicit Kernel(Tuning tuning_) : tuning(tuning_) {}
  void Run(const PackedMatrix<std::int8_t>& lhs,
           const PackedMatrix<std::int8_t>& rhs,
           const BasicSpec<std::int32_t, DstScalar>& spec, int start_row,
           int start_col, int end_row, int end_col,
           Matrix<DstScalar>* dst) const {
    KernelParams8bit<LhsLayout::kCols, RhsLayout::kCols> params;
    MakeKernelParams8bit(lhs, rhs, spec, start_row, start_col, end_row, end_col,
                         dst, &params);
    if (dst->layout.cols == 1) {
      Kernel8bitAvx2SingleCol(params);
    } else {
      Kernel8bitAvx2(params);
    }
  }
};

void KernelFloatAvx2(const KernelParamsFloat<8, 8>& params);
void KernelFloatAvx2SingleCol(const KernelParamsFloat<8, 8>& params);

template <>
struct Kernel<Path::kAvx2, float, float, float, BasicSpec<float, float>> {
  Tuning tuning = Tuning::kAuto;
  using LhsLayout = FixedKernelLayout<Order::kRowMajor, 1, 8>;
  using RhsLayout = FixedKernelLayout<Order::kRowMajor, 1, 8>;
  explicit Kernel(Tuning tuning_) : tuning(tuning_) {}
  void Run(const PackedMatrix<float>& lhs, const PackedMatrix<float>& rhs,
           const BasicSpec<float, float>& spec, int start_row, int start_col,
           int end_row, int end_col, Matrix<float>* dst) const {
    KernelParamsFloat<LhsLayout::kCols, RhsLayout::kCols> params;
    MakeKernelParamsFloat(lhs, rhs, spec, start_row, start_col, end_row,
                          end_col, dst, &params);
    if (dst->layout.cols == 1) {
      KernelFloatAvx2SingleCol(params);
    } else {
      KernelFloatAvx2(params);
    }
  }
};

// TODO(b/147376783): SSE 4.2 and AVX-VNNI support is incomplete / placeholder.
// Optimization is not finished. In particular the dimensions of the kernel
// blocks can be changed as desired.
//
void Kernel8bitAvxVnni(const KernelParams8bit<16, 16>& params);

template <typename DstScalar>
struct Kernel<Path::kAvxVnni, std::int8_t, std::int8_t, DstScalar,
              BasicSpec<std::int32_t, DstScalar>> {
  Tuning tuning = Tuning::kAuto;
  using LhsLayout = FixedKernelLayout<Order::kColMajor, 4, 16>;
  using RhsLayout = FixedKernelLayout<Order::kColMajor, 4, 16>;
  explicit Kernel(Tuning tuning_) : tuning(tuning_) {}
  void Run(const PackedMatrix<std::int8_t>& lhs,
           const PackedMatrix<std::int8_t>& rhs,
           const BasicSpec<std::int32_t, DstScalar>& spec, int start_row,
           int start_col, int end_row, int end_col,
           Matrix<DstScalar>* dst) const {
    KernelParams8bit<LhsLayout::kCols, RhsLayout::kCols> params;
    MakeKernelParams8bit(lhs, rhs, spec, start_row, start_col, end_row, end_col,
                         dst, &params);
    Kernel8bitAvxVnni(params);
  }
};

void KernelFloatAvxVnni(const KernelParamsFloat<16, 16>& params);

template <>
struct Kernel<Path::kAvxVnni, float, float, float, BasicSpec<float, float>> {
  Tuning tuning = Tuning::kAuto;
  using LhsLayout = FixedKernelLayout<Order::kRowMajor, 1, 16>;
  using RhsLayout = FixedKernelLayout<Order::kRowMajor, 1, 16>;
  explicit Kernel(Tuning tuning_) : tuning(tuning_) {}
  void Run(const PackedMatrix<float>& lhs, const PackedMatrix<float>& rhs,
           const BasicSpec<float, float>& spec, int start_row, int start_col,
           int end_row, int end_col, Matrix<float>* dst) const {
    KernelParamsFloat<LhsLayout::kCols, RhsLayout::kCols> params;
    MakeKernelParamsFloat(lhs, rhs, spec, start_row, start_col, end_row,
                          end_col, dst, &params);
    KernelFloatAvxVnni(params);
  }
};

#endif  // RUY_PLATFORM(X86)

}  // namespace ruy

#endif  // TENSORFLOW_LITE_EXPERIMENTAL_RUY_KERNEL_X86_H_