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
|
#include "intgemm.h"
// This is just for AlignedVector, which helps managed 64-byte aligned memory.
// Feel free to manage memory yourself.
#include "aligned.h"
#include <cassert>
#include <stdlib.h>
#include <math.h>
int main() {
const int A_rows = 1;
// The shared dimension: A's columns and B's rows.
const int width = 64;
const int B_cols = 8;
// This is a simple vector class that allocates memory aligned to 64 bytes.
// You don't have to use it; just use aligned_alloc and friends directly.
using intgemm::AlignedVector;
AlignedVector<float> A(A_rows * width);
AlignedVector<float> B(width * B_cols);
// Fill with random values in range [-2, 2].
srand(1);
for (int i = 0; i < A_rows * width; ++i) {
A[i] = ((float)rand()/(float)RAND_MAX)*4.0f - 2.0f;
}
for (int i = 0; i < width * B_cols; ++i) {
B[i] = ((float)rand()/(float)RAND_MAX)*4.0f - 2.0f;
}
// Compute the top left corner of C as a sanity check.
float top_left_reference = 0.0;
for (int w = 0; w < width; ++w) {
top_left_reference += A[w] * B[w * B_cols];
}
// 16-bit multiplication.
{
// For 16-bit, Jacob Devlin recommends 1024 so as to not overflow in 32-bit accumulation.
float quant_mult = 1024.0;
AlignedVector<int16_t> A_prepared(A_rows * width);
AlignedVector<int16_t> B_prepared(width * B_cols);
// Quantize A.
intgemm::Generic_16bit::PrepareA(A.get(), A_prepared.get(), quant_mult, A_rows, width);
// Quantize and reshape B.
// Typically you will do this once when parameters are loaded, not every time.
intgemm::Generic_16bit::PrepareB(B.get(), B_prepared.get(), quant_mult, width, B_cols);
AlignedVector<float> C(A_rows * B_cols);
// Do the actual multiply.
intgemm::Generic_16bit::Multiply(A_prepared.get(), B_prepared.get(), C.get(), 1.0 / (quant_mult * quant_mult), A_rows, width, B_cols);
// Sanity check. C will be row major.
assert(fabs(C[0] - top_left_reference) < 0.05);
}
// 8-bit multiplication.
{
// For 8-bit a good quantization multiplier is 127 / largest absolute value..
float quant_mult = 127.0 / 2.0;
AlignedVector<int8_t> A_prepared(A_rows * width);
AlignedVector<int8_t> B_prepared(width * B_cols);
// Quantize A.
intgemm::Generic_8bit::PrepareA(A.get(), A_prepared.get(), quant_mult, A_rows, width);
// Quantize and reshape B.
// Typically you will do this once when parameters are loaded, not every time.
intgemm::Generic_8bit::PrepareB(B.get(), B_prepared.get(), quant_mult, width, B_cols);
AlignedVector<float> C(A_rows * B_cols);
// Do the actual multiply.
intgemm::Generic_8bit::Multiply(A_prepared.get(), B_prepared.get(), C.get(), 1.0 / (quant_mult * quant_mult), A_rows, width, B_cols);
// Sanity check. C will be row major.
assert(fabs(C[0] - top_left_reference) < 0.05);
}
}
|