mirror of
https://github.com/tesseract-ocr/tesseract.git
synced 2024-12-03 00:49:01 +08:00
104 lines
3.7 KiB
C++
104 lines
3.7 KiB
C++
///////////////////////////////////////////////////////////////////////
|
|
// File: dotproductavx.cpp
|
|
// Description: Architecture-specific dot-product function.
|
|
// Author: Ray Smith
|
|
// Created: Wed Jul 22 10:48:05 PDT 2015
|
|
//
|
|
// (C) Copyright 2015, Google Inc.
|
|
// 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.
|
|
///////////////////////////////////////////////////////////////////////
|
|
|
|
#if !defined(__AVX__)
|
|
// Implementation for non-avx archs.
|
|
|
|
#include "dotproductavx.h"
|
|
#include <stdio.h>
|
|
#include <stdlib.h>
|
|
|
|
namespace tesseract {
|
|
double DotProductAVX(const double* u, const double* v, int n) {
|
|
fprintf(stderr, "DotProductAVX can't be used on Android\n");
|
|
abort();
|
|
}
|
|
} // namespace tesseract
|
|
|
|
#else // !defined(__AVX__)
|
|
// Implementation for avx capable archs.
|
|
#include <immintrin.h>
|
|
#include <stdint.h>
|
|
#include "dotproductavx.h"
|
|
#include "host.h"
|
|
|
|
namespace tesseract {
|
|
|
|
// Computes and returns the dot product of the n-vectors u and v.
|
|
// Uses Intel AVX intrinsics to access the SIMD instruction set.
|
|
double DotProductAVX(const double* u, const double* v, int n) {
|
|
int max_offset = n - 4;
|
|
int offset = 0;
|
|
// Accumulate a set of 4 sums in sum, by loading pairs of 4 values from u and
|
|
// v, and multiplying them together in parallel.
|
|
__m256d sum = _mm256_setzero_pd();
|
|
if (offset <= max_offset) {
|
|
offset = 4;
|
|
// Aligned load is reputedly faster but requires 32 byte aligned input.
|
|
if ((reinterpret_cast<const uintptr_t>(u) & 31) == 0 &&
|
|
(reinterpret_cast<const uintptr_t>(v) & 31) == 0) {
|
|
// Use aligned load.
|
|
__m256d floats1 = _mm256_load_pd(u);
|
|
__m256d floats2 = _mm256_load_pd(v);
|
|
// Multiply.
|
|
sum = _mm256_mul_pd(floats1, floats2);
|
|
while (offset <= max_offset) {
|
|
floats1 = _mm256_load_pd(u + offset);
|
|
floats2 = _mm256_load_pd(v + offset);
|
|
offset += 4;
|
|
__m256d product = _mm256_mul_pd(floats1, floats2);
|
|
sum = _mm256_add_pd(sum, product);
|
|
}
|
|
} else {
|
|
// Use unaligned load.
|
|
__m256d floats1 = _mm256_loadu_pd(u);
|
|
__m256d floats2 = _mm256_loadu_pd(v);
|
|
// Multiply.
|
|
sum = _mm256_mul_pd(floats1, floats2);
|
|
while (offset <= max_offset) {
|
|
floats1 = _mm256_loadu_pd(u + offset);
|
|
floats2 = _mm256_loadu_pd(v + offset);
|
|
offset += 4;
|
|
__m256d product = _mm256_mul_pd(floats1, floats2);
|
|
sum = _mm256_add_pd(sum, product);
|
|
}
|
|
}
|
|
}
|
|
// Add the 4 product sums together horizontally. Not so easy as with sse, as
|
|
// there is no add across the upper/lower 128 bit boundary, so permute to
|
|
// move the upper 128 bits to lower in another register.
|
|
__m256d sum2 = _mm256_permute2f128_pd(sum, sum, 1);
|
|
sum = _mm256_hadd_pd(sum, sum2);
|
|
sum = _mm256_hadd_pd(sum, sum);
|
|
double result;
|
|
// _mm256_extract_f64 doesn't exist, but resist the temptation to use an sse
|
|
// instruction, as that introduces a 70 cycle delay. All this casting is to
|
|
// fool the instrinsics into thinking we are extracting the bottom int64.
|
|
*(reinterpret_cast<inT64*>(&result)) =
|
|
_mm256_extract_epi64(_mm256_castpd_si256(sum), 0);
|
|
while (offset < n) {
|
|
result += u[offset] * v[offset];
|
|
++offset;
|
|
}
|
|
return result;
|
|
}
|
|
|
|
} // namespace tesseract.
|
|
|
|
#endif // ANDROID_BUILD
|