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
|
#pragma once
#include "intrinsics.h"
#include "vec_utils.h"
#include <cassert>
#include <exception>
namespace intgemm {
class JustUnquantizeC {
public:
JustUnquantizeC(float *C, float unquant_mult) : C_(C), unquant_mult_(unquant_mult) {}
class OnSSE2 {
public:
INTGEMM_SSE2 explicit OnSSE2(const JustUnquantizeC &from)
: C_(from.C_), unquant_mult_(set1_ps<__m128>(from.unquant_mult_)) {
assert(reinterpret_cast<uintptr_t>(C_) % sizeof(__m128i) == 0);
}
INTGEMM_SSE2 inline void operator()(Index rowIDX, Index cols, Index colIDX, MultiplyResult128 result) {
storeu_ps(C_ + rowIDX*cols + colIDX , unquantize(result.pack0123, unquant_mult_));
storeu_ps(C_ + rowIDX*cols + colIDX + 4, unquantize(result.pack4567, unquant_mult_));
}
private:
float *C_;
__m128 unquant_mult_;
};
class OnAVX2 {
public:
INTGEMM_AVX2 explicit OnAVX2(const JustUnquantizeC &from)
: C_(from.C_), unquant_mult_(set1_ps<__m256>(from.unquant_mult_)) {
assert(reinterpret_cast<uintptr_t>(C_) % sizeof(__m256i) == 0);
}
INTGEMM_AVX2 inline void operator()(Index rowIDX, Index cols, Index colIDX, __m256i result) {
storeu_ps(C_ + rowIDX*cols + colIDX, unquantize(result, unquant_mult_));
}
private:
float *C_;
__m256 unquant_mult_;
};
private:
float *C_;
float unquant_mult_;
};
class BiasAddUnquantizeC {
public:
BiasAddUnquantizeC(float *C, const float *bias, float unquant_mult) : C_(C), bias_(bias), unquant_mult_(unquant_mult) {}
class OnSSE2 {
public:
INTGEMM_SSE2 explicit OnSSE2(const BiasAddUnquantizeC &from)
: C_(from.C_), bias_(from.bias_), unquant_mult_(_mm_set1_ps(from.unquant_mult_)) {
assert(reinterpret_cast<uintptr_t>(C_) % sizeof(__m128i) == 0);
}
INTGEMM_SSE2 inline void operator()(Index rowIDX, Index cols, Index colIDX, MultiplyResult128 result) {
const __m128* biasSection = reinterpret_cast<const __m128*>(bias_ + colIDX);
const __m128* biasSection2 = reinterpret_cast<const __m128*>(bias_ + colIDX + 4);
storeu_ps(C_ + rowIDX*cols + colIDX, add_ps(unquantize(result.pack0123, unquant_mult_), *biasSection));
storeu_ps(C_ + rowIDX*cols + colIDX + 4, add_ps(unquantize(result.pack4567, unquant_mult_), *biasSection2));
}
private:
float *C_;
const float *bias_;
__m128 unquant_mult_;
};
class OnAVX2 {
public:
INTGEMM_AVX2 explicit OnAVX2(const BiasAddUnquantizeC &from)
: C_(from.C_), bias_(from.bias_), unquant_mult_(_mm256_set1_ps(from.unquant_mult_)) {
assert(reinterpret_cast<uintptr_t>(C_) % sizeof(__m256i) == 0);
}
INTGEMM_AVX2 inline void operator()(Index rowIDX, Index cols, Index colIDX, __m256i result) {
const __m256* biasSection = reinterpret_cast<const __m256*>(bias_ + colIDX);
storeu_ps(C_ + rowIDX*cols + colIDX, add_ps(unquantize(result, unquant_mult_), *biasSection));
}
private:
float *C_;
const float *bias_;
__m256 unquant_mult_;
};
private:
float *C_;
const float *bias_;
float unquant_mult_;
};
class Identity {
public:
explicit Identity(int32_t *C) : C_(C) {}
class OnSSE2 {
public:
INTGEMM_SSE2 explicit OnSSE2(const Identity &from)
: C_(from.C_) {
assert(reinterpret_cast<uintptr_t>(C_) % sizeof(__m128i) == 0);
}
INTGEMM_SSE2 inline void operator()(Index rowIDX, Index cols, Index colIDX, MultiplyResult128 result) {
_mm_storeu_si128(reinterpret_cast<__m128i*>(C_ + rowIDX*cols + colIDX), result.pack0123);
_mm_storeu_si128(reinterpret_cast<__m128i*>(C_ + rowIDX*cols + colIDX + 4), result.pack4567);
}
private:
int32_t *C_;
};
class OnAVX2 {
public:
INTGEMM_AVX2 explicit OnAVX2(const Identity &from)
: C_(from.C_) {
assert(reinterpret_cast<uintptr_t>(C_) % sizeof(__m256i) == 0);
}
INTGEMM_AVX2 inline void operator()(Index rowIDX, Index cols, Index colIDX, __m256i result) {
_mm256_storeu_si256(reinterpret_cast<__m256i*>(C_ + rowIDX*cols + colIDX), result);
}
private:
int32_t *C_;
};
private:
int32_t *C_;
};
class ReLU {
public:
explicit ReLU(float *C, float unquant_mult) : C_(C), unquant_mult_(unquant_mult) {}
class OnSSE2 {
public:
INTGEMM_SSE2 explicit OnSSE2(const ReLU& from)
: C_(from.C_), zeros_(setzero_ps<__m128>()), unquant_mult_(set1_ps<__m128>(from.unquant_mult_)) {
assert(reinterpret_cast<uintptr_t>(C_) % sizeof(__m128i) == 0);
}
INTGEMM_SSE2 inline void operator()(Index rowIDX, Index cols, Index colIDX, MultiplyResult128 result) {
auto unquantized0123 = unquantize(result.pack0123, unquant_mult_);
auto nonnegative0123 = max_ps(zeros_, unquantized0123);
storeu_ps(C_ + rowIDX*cols + colIDX, nonnegative0123);
auto unquantized4567 = unquantize(result.pack4567, unquant_mult_);
auto nonnegative4567 = max_ps(zeros_, unquantized4567);
storeu_ps(C_ + rowIDX*cols + colIDX + 4, nonnegative4567);
}
private:
float* C_;
__m128 unquant_mult_;
__m128 zeros_;
};
using OnSSSE2 = OnSSE2;
class OnAVX2 {
public:
INTGEMM_AVX2 explicit OnAVX2(const ReLU& from)
: C_(from.C_), zeros_(setzero_ps<__m256>()), unquant_mult_(set1_ps<__m256>(from.unquant_mult_)) {
assert(reinterpret_cast<uintptr_t>(C_) % sizeof(__m256i) == 0);
}
INTGEMM_AVX2 inline void operator()(Index rowIDX, Index cols, Index colIDX, __m256i result) {
auto nonnegative = max_ps(zeros_, unquantize(result, unquant_mult_));
storeu_ps(C_ + rowIDX*cols + colIDX, nonnegative);
}
private:
float* C_;
__m256 unquant_mult_;
__m256 zeros_;
};
#ifndef INTGEMM_NO_AVX512
class OnAVX512 {
public:
INTGEMM_AVX512BW explicit OnAVX512(const ReLU& from)
: C_(from.C_), zeros_(setzero_ps<__m512>()), unquant_mult_(set1_ps<__m512>(from.unquant_mult_)) {
assert(reinterpret_cast<uintptr_t>(C_) % sizeof(__m512i) == 0);
}
INTGEMM_AVX512BW inline void operator()(Index rowIDX, Index cols, Index colIDX, __m512i result) {
auto nonnegative = max_ps(zeros_, unquantize(result, unquant_mult_));
storeu_ps(C_ + rowIDX*cols + colIDX, nonnegative);
}
private:
float* C_;
__m512 unquant_mult_;
__m512 zeros_;
};
#endif
private:
float* C_;
float unquant_mult_;
};
}
|
