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
|
#pragma once
#include "intgemm/intgemm_config.h"
#ifdef INTGEMM_COMPILER_SUPPORTS_AVX512VNNI
#include "avx512_gemm.h"
#include "types.h"
namespace intgemm {
namespace AVX512VNNI {
// Workaround extra vmovdqa64 https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94663
INTGEMM_AVX512VNNI static inline void VNNI8(__m512i &c, __m512i a, __m512i b) {
#if defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER)
asm ("vpdpbusds %2, %1, %0" : "+x"(c) : "x"(a), "mx"(b));
#else
c = _mm512_dpbusds_epi32(c, a, b);
#endif
}
struct Kernels8 : public AVX512BW::Kernels8 {
template <typename Callback>
INTGEMM_AVX512VNNI static void Multiply(const int8_t *A, const int8_t *B, Index A_rows, Index width, Index B_cols, Callback callback) {
assert(width % sizeof(Register) == 0);
assert(B_cols % 8 == 0);
assert(reinterpret_cast<uintptr_t>(A) % sizeof(Register) == 0);
assert(reinterpret_cast<uintptr_t>(B) % sizeof(Register) == 0);
auto callback_impl = callbacks::CallbackImpl<CPUType::AVX2, Callback>(callback);
const Index simd_width = width / sizeof(Register);
Register zeros = setzero_si<Register>();
// Go over 8 columns of B at a time.
#pragma omp for
for (Index B0_colidx = 0; B0_colidx < B_cols; B0_colidx += 8) {
const Register *B0_col = reinterpret_cast<const Register*>(B) + B0_colidx * simd_width;
// Process one row of A at a time. Doesn't seem to be faster to do multiple rows of A at once.
for (Index A_rowidx = 0; A_rowidx < A_rows; ++A_rowidx) {
// Iterate over shared (inner) dimension.
const Register *A_live = reinterpret_cast<const Register *>(A + A_rowidx * width);
const Register *A_end = A_live + simd_width;
const Register *B_live = B0_col;
// TODO: separate first step.
Register sum0 = zeros, sum1 = zeros, sum2 = zeros, sum3 = zeros, sum4 = zeros, sum5 = zeros, sum6 = zeros, sum7 = zeros;
for (; A_live != A_end; ++A_live, B_live += 8) {
Register a = *A_live;
// Retrieve the conveniently consecutive values of B.
Register b0 = *B_live;
Register b1 = *(B_live + 1);
Register b2 = *(B_live + 2);
Register b3 = *(B_live + 3);
Register b4 = *(B_live + 4);
Register b5 = *(B_live + 5);
Register b6 = *(B_live + 6);
Register b7 = *(B_live + 7);
// Get a mask where a is negative.
__mmask64 neg_mask = _mm512_test_epi8_mask(a, _mm512_set1_epi8(-128));
Register a_positive = _mm512_abs_epi8(a);
// Negate by subtracting from zero with a mask.
b0 = _mm512_mask_sub_epi8(b0, neg_mask, zeros, b0);
b1 = _mm512_mask_sub_epi8(b1, neg_mask, zeros, b1);
b2 = _mm512_mask_sub_epi8(b2, neg_mask, zeros, b2);
b3 = _mm512_mask_sub_epi8(b3, neg_mask, zeros, b3);
b4 = _mm512_mask_sub_epi8(b4, neg_mask, zeros, b4);
b5 = _mm512_mask_sub_epi8(b5, neg_mask, zeros, b5);
b6 = _mm512_mask_sub_epi8(b6, neg_mask, zeros, b6);
b7 = _mm512_mask_sub_epi8(b7, neg_mask, zeros, b7);
VNNI8(sum0, a_positive, b0);
VNNI8(sum1, a_positive, b1);
VNNI8(sum2, a_positive, b2);
VNNI8(sum3, a_positive, b3);
VNNI8(sum4, a_positive, b4);
VNNI8(sum5, a_positive, b5);
VNNI8(sum6, a_positive, b6);
VNNI8(sum7, a_positive, b7);
}
Register pack0123 = Pack0123(sum0, sum1, sum2, sum3);
Register pack4567 = Pack0123(sum4, sum5, sum6, sum7);
auto total = PermuteSummer(pack0123, pack4567);
callback_impl.Run(total, callbacks::OutputBufferInfo(A_rowidx, B0_colidx, A_rows, B_cols));
}
}
}
template <typename Callback>
INTGEMM_AVX512VNNI static void Multiply8Shift(const uint8_t *A, const int8_t *B, Index A_rows, Index width, Index B_cols, Callback callback) {
assert(width % sizeof(Register) == 0);
assert(B_cols % 8 == 0);
assert(reinterpret_cast<uintptr_t>(A) % sizeof(Register) == 0);
assert(reinterpret_cast<uintptr_t>(B) % sizeof(Register) == 0);
auto callback_impl = callbacks::CallbackImpl<CPUType::AVX2, Callback>(callback);
const Index simd_width = width / sizeof(Register);
Register zeros = setzero_si<Register>();
// Go over 8 columns of B at a time.
#pragma omp for
for (Index B0_colidx = 0; B0_colidx < B_cols; B0_colidx += 8) {
const Register *B0_col = reinterpret_cast<const Register*>(B) + B0_colidx * simd_width;
// Process one row of A at a time. Doesn't seem to be faster to do multiple rows of A at once.
for (Index A_rowidx = 0; A_rowidx < A_rows; ++A_rowidx) {
// Iterate over shared (inner) dimension.
const Register *A_live = reinterpret_cast<const Register *>(A + A_rowidx * width);
const Register *A_end = A_live + simd_width;
const Register *B_live = B0_col;
// TODO: separate first step.
Register sum0 = zeros, sum1 = zeros, sum2 = zeros, sum3 = zeros, sum4 = zeros, sum5 = zeros, sum6 = zeros, sum7 = zeros;
for (; A_live != A_end; ++A_live, B_live += 8) {
Register a = *A_live;
//MultiplyAdd
VNNI8(sum0, a, *B_live);
VNNI8(sum1, a, *(B_live + 1));
VNNI8(sum2, a, *(B_live + 2));
VNNI8(sum3, a, *(B_live + 3));
VNNI8(sum4, a, *(B_live + 4));
VNNI8(sum5, a, *(B_live + 5));
VNNI8(sum6, a, *(B_live + 6));
VNNI8(sum7, a, *(B_live + 7));
}
Register pack0123 = Pack0123(sum0, sum1, sum2, sum3);
Register pack4567 = Pack0123(sum4, sum5, sum6, sum7);
auto total = PermuteSummer(pack0123, pack4567);
callback_impl.Run(total, callbacks::OutputBufferInfo(A_rowidx, B0_colidx, A_rows, B_cols));
}
}
}
template <typename Callback>
INTGEMM_AVX512VNNI static void PrepareBias(const int8_t *B, Index width, Index B_cols, Callback callback) {
assert(width % sizeof(Register) == 0);
assert(B_cols % 8 == 0);
assert(reinterpret_cast<uintptr_t>(B) % sizeof(Register) == 0);
auto callback_impl = callbacks::CallbackImpl<CPUType::AVX2, Callback>(callback);
Index simd_width = width / sizeof(Register);
Register zeros = setzero_si<Register>();
const Register a = set1_epi8<Register>(1);
// Go over 8 columns of B at a time.
#pragma omp for
for (Index B0_colidx = 0; B0_colidx < B_cols; B0_colidx += 8) {
const Register *B0_col = reinterpret_cast<const Register*>(B) + B0_colidx * simd_width;
const Register *B_live = B0_col; //In order to make the code look as much as possible as the above function
const Register *B_end = B_live + simd_width*8;
// TODO: separate first step.
Register sum0 = zeros, sum1 = zeros, sum2 = zeros, sum3 = zeros, sum4 = zeros, sum5 = zeros, sum6 = zeros, sum7 = zeros;
for (; B_live != B_end; B_live += 8) {
// Retrieve the conveniently consecutive values of B.
VNNI8(sum0, a, *B_live);
VNNI8(sum1, a, *(B_live + 1));
VNNI8(sum2, a, *(B_live + 2));
VNNI8(sum3, a, *(B_live + 3));
VNNI8(sum4, a, *(B_live + 4));
VNNI8(sum5, a, *(B_live + 5));
VNNI8(sum6, a, *(B_live + 6));
VNNI8(sum7, a, *(B_live + 7));
}
Register pack0123 = Pack0123(sum0, sum1, sum2, sum3);
Register pack4567 = Pack0123(sum4, sum5, sum6, sum7);
auto total = PermuteSummer(pack0123, pack4567);
callback_impl.Run(total, callbacks::OutputBufferInfo(0, B0_colidx, 1, B_cols));
}
}
constexpr static const char *const kName = "8-bit AVX512VNNI";
static const CPUType kUses = CPUType::AVX512VNNI;
};
} // namespace AVX512VNNI
} // namespace intgemm
#endif
|
