mirror/tesseract
1
0
Fork 0
mirror of https://github.com/tesseract-ocr/tesseract.git synced 2024-11-24 02:59:07 +08:00
Code Issues Actions 10 Packages Projects Releases Wiki Activity

Facilitate vectorization for generic build (#4223)

Browse Source 
* Update intsimdmatrix.cpp

Unroll loop into chunks of four operations to facilitate compilers recognizing vectorization opportunity.

* Apply suggestions from code review

Co-authored-by: Stefan Weil <sw@weilnetz.de>
This commit is contained in:
Mahesh Madhav 2024-04-16 10:28:55 -07:00 committed by GitHub
parent d5e000bc58
commit bae520ea00
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 26 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
src/arch/intsimdmatrix.cpp
View File

@ -80,7 +80,32 @@ void IntSimdMatrix::MatrixDotVector(const GENERIC_2D_ARRAY<int8_t> &w,
int num_out = w.dim1(); int num_out = w.dim1();
int num_in = w.dim2() - 1; int num_in = w.dim2() - 1;
// Base implementation. // Base implementation.
for (int i = 0; i < num_out; ++i) { int i;
// Break up into chunks of four to facilitate vectorization
for (i = 0; i < (num_out / 4) * 4; i += 4) {
const int8_t *wi0 = w[i + 0];
const int8_t *wi1 = w[i + 1];
const int8_t *wi2 = w[i + 2];
const int8_t *wi3 = w[i + 3];
int total0 = 0;
int total1 = 0;
int total2 = 0;
int total3 = 0;
for (int j = 0; j < num_in; ++j) {
total0 += wi0[j] * u[j];
total1 += wi1[j] * u[j];
total2 += wi2[j] * u[j];
total3 += wi3[j] * u[j];
}
// Add in the bias and correct for integer values.
v[i + 0] = (total0 + wi0[num_in] * INT8_MAX) * scales[i + 0];
v[i + 1] = (total1 + wi1[num_in] * INT8_MAX) * scales[i + 1];
v[i + 2] = (total2 + wi2[num_in] * INT8_MAX) * scales[i + 2];
v[i + 3] = (total3 + wi3[num_in] * INT8_MAX) * scales[i + 3];
}
// Capture the remainder mod four
for (; i < num_out; ++i) {
const int8_t *wi = w[i]; const int8_t *wi = w[i];
int total = 0; int total = 0;
for (int j = 0; j < num_in; ++j) { for (int j = 0; j < num_in; ++j) {