root
/
MNN
mirror of https://github.com/alibaba/MNN.git
1
0
Fork 0
Code Issues Actions 1 Packages Projects Releases Wiki Activity
MNN/source/cv/ImageFloatBlitter.cpp

414 lines
18 KiB
C++
Raw Blame History

//
// ImageFloatBlitter.cpp
// MNN
//
// Created by MNN on 2018/12/24.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include "cv/ImageFloatBlitter.hpp"
extern "C" {
void MNNBlitC1ToFloatRGBA(const unsigned char* source, float* dest, const float* mean, const float* normal,
size_t count);
void MNNBlitC3ToFloatRGBA(const unsigned char* source, float* dest, const float* mean, const float* normal,
size_t count);
void MNNBlitC4ToFloatC4(const unsigned char* source, float* dest, const float* mean, const float* normal, size_t count);
}
#ifdef MNN_USE_SSE
#if defined(_MSC_VER)
#include <intrin.h>
#else
#include <x86intrin.h>
#endif
#ifndef _MM_TRANSPOSE4_PS
#define _MM_TRANSPOSE4_PS(row0, row1, row2, row3) \
do { \
__m128 tmp3, tmp2, tmp1, tmp0; \
tmp0 = _mm_unpacklo_ps((row0), (row1)); \
tmp2 = _mm_unpacklo_ps((row2), (row3)); \
tmp1 = _mm_unpackhi_ps((row0), (row1)); \
tmp3 = _mm_unpackhi_ps((row2), (row3)); \
(row0) = _mm_movelh_ps(tmp0, tmp2); \
(row1) = _mm_movehl_ps(tmp2, tmp0); \
(row2) = _mm_movelh_ps(tmp1, tmp3); \
(row3) = _mm_movehl_ps(tmp3, tmp1); \
} while (0)
#endif
#endif
#include "core/Macro.h"
#ifdef MNN_USE_NEON
#include <arm_neon.h>
#endif
namespace MNN {
namespace CV {
static void _blitC1ToFloatC1(const unsigned char* source, float* dest, const float* mean, const float* normal,
size_t count) {
#ifdef MNN_USE_NEON
unsigned long size = count >> 4;
float32x4_t cache = vdupq_n_f32(0);
float32x4_t _mean = vdupq_n_f32(-mean[0]);
float32x4_t _normal = vdupq_n_f32(normal[0]);
for (int i = 0; i < size; i++, source += 16) {
uint8x16_t v = vld1q_u8(source);
int16x8_t vl = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(v))); // 0..7
int16x8_t vh = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(v))); // 8..15
// unpack to 32 bits
float32x4_t vll = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vl))); // 0..3
cache = vaddq_f32(_mean, vll);
cache = vmulq_f32(cache, _normal);
vst1q_f32(dest, cache);
dest += 4;
float32x4_t vlh = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vl))); // 4..7
cache = vaddq_f32(_mean, vlh);
cache = vmulq_f32(cache, _normal);
vst1q_f32(dest, cache);
dest += 4;
float32x4_t vhl = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vh))); // 8..11
cache = vaddq_f32(_mean, vhl);
cache = vmulq_f32(cache, _normal);
vst1q_f32(dest, cache);
dest += 4;
float32x4_t vhh = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vh))); // 12..15
cache = vaddq_f32(_mean, vhh);
cache = vmulq_f32(cache, _normal);
vst1q_f32(dest, cache);
dest += 4;
}
int left = count & 15;
if (left == 0) {
return;
}
for (int i = 0; i < left; ++i, ++dest, ++source) {
*dest = normal[0] * (*source - mean[0]);
}
#else
int remain = 0;
#ifdef MNN_USE_SSE
int countC16 = count / 16;
remain = countC16 * 16;
const auto meanC4 = _mm_set1_ps(mean[0]);
const auto normalC4 = _mm_set1_ps(normal[0]);
const __m128i l1 = _mm_setr_epi8(4,5,6,7, 6,5,4,7, 10,9,8,11, 14,13,12,15);
const __m128i l2 = _mm_setr_epi8(8,9,10,11, 6,5,4,7, 10,9,8,11, 14,13,12,15);
const __m128i l3 = _mm_setr_epi8(12,13,14,15, 6,5,4,7, 10,9,8,11, 14,13,12,15);
for (int i=0; i<countC16; ++i) {
auto srcInt8 = _mm_loadu_si128((const __m128i*)(source + i * 16));
auto int3200 = _mm_cvtepu8_epi32(srcInt8);
auto int3201 = _mm_cvtepu8_epi32(_mm_shuffle_epi8(srcInt8, l1));
auto int3210 = _mm_cvtepu8_epi32(_mm_shuffle_epi8(srcInt8, l2));
auto int3211 = _mm_cvtepu8_epi32(_mm_shuffle_epi8(srcInt8, l3));
auto float00 = _mm_cvtepi32_ps(int3200);
auto float01 = _mm_cvtepi32_ps(int3201);
auto float10 = _mm_cvtepi32_ps(int3210);
auto float11 = _mm_cvtepi32_ps(int3211);
_mm_storeu_ps(dest + 16 * i + 4 * 0, _mm_mul_ps(_mm_sub_ps(float00, meanC4), normalC4));
_mm_storeu_ps(dest + 16 * i + 4 * 1, _mm_mul_ps(_mm_sub_ps(float01, meanC4), normalC4));
_mm_storeu_ps(dest + 16 * i + 4 * 2, _mm_mul_ps(_mm_sub_ps(float10, meanC4), normalC4));
_mm_storeu_ps(dest + 16 * i + 4 * 3, _mm_mul_ps(_mm_sub_ps(float11, meanC4), normalC4));
}
#endif
for (int i = remain; i < count; ++i) {
dest[i + 0] = normal[0] * (source[i + 0] - mean[0]);
}
#endif
}
static void _blitC3ToFloatC3(const unsigned char* source, float* dest, const float* mean, const float* normal,
size_t count) {
#ifdef MNN_USE_NEON
int size = (int)count / 16;
float32x4x3_t cachell = {vmovq_n_f32(0), vmovq_n_f32(0), vmovq_n_f32(0)};
float32x4x3_t cachelh = {vmovq_n_f32(0), vmovq_n_f32(0), vmovq_n_f32(0)};
float32x4x3_t cachehl = {vmovq_n_f32(0), vmovq_n_f32(0), vmovq_n_f32(0)};
float32x4x3_t cachehh = {vmovq_n_f32(0), vmovq_n_f32(0), vmovq_n_f32(0)};
float32x4x3_t _mean;
float32x4x3_t _normal;
for (int c = 0; c < 3; c++) {
_mean.val[c] = vmovq_n_f32(-mean[c]);
_normal.val[c] = vmovq_n_f32(normal[c]);
}
for (int i = 0; i < size; i++) {
uint8x16x3_t v = vld3q_u8(source + 16 * 3 * i);
int c = 0;
{
int16x8_t vl = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(v.val[c]))); // 0..7
// unpack to 32 bits
float32x4_t vll = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vl))); // 0..3
cachell.val[c] = vaddq_f32(_mean.val[c], vll);
cachell.val[c] = vmulq_f32(cachell.val[c], _normal.val[c]);
float32x4_t vlh = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vl))); // 4..7
cachelh.val[c] = vaddq_f32(_mean.val[c], vlh);
cachelh.val[c] = vmulq_f32(cachelh.val[c], _normal.val[c]);
int16x8_t vh = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(v.val[c]))); // 8..15
// unpack to 32 bits
float32x4_t vhl = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vh))); // 8..11
cachehl.val[c] = vaddq_f32(_mean.val[c], vhl);
cachehl.val[c] = vmulq_f32(cachehl.val[c], _normal.val[c]);
float32x4_t vhh = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vh))); // 12..15
cachehh.val[c] = vaddq_f32(_mean.val[c], vhh);
cachehh.val[c] = vmulq_f32(cachehh.val[c], _normal.val[c]);
}
c = 1;
{
int16x8_t vl = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(v.val[c]))); // 0..7
// unpack to 32 bits
float32x4_t vll = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vl))); // 0..3
cachell.val[c] = vaddq_f32(_mean.val[c], vll);
cachell.val[c] = vmulq_f32(cachell.val[c], _normal.val[c]);
float32x4_t vlh = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vl))); // 4..7
cachelh.val[c] = vaddq_f32(_mean.val[c], vlh);
cachelh.val[c] = vmulq_f32(cachelh.val[c], _normal.val[c]);
int16x8_t vh = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(v.val[c]))); // 8..15
// unpack to 32 bits
float32x4_t vhl = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vh))); // 8..11
cachehl.val[c] = vaddq_f32(_mean.val[c], vhl);
cachehl.val[c] = vmulq_f32(cachehl.val[c], _normal.val[c]);
float32x4_t vhh = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vh))); // 12..15
cachehh.val[c] = vaddq_f32(_mean.val[c], vhh);
cachehh.val[c] = vmulq_f32(cachehh.val[c], _normal.val[c]);
}
c = 2;
{
int16x8_t vl = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(v.val[c]))); // 0..7
// unpack to 32 bits
float32x4_t vll = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vl))); // 0..3
cachell.val[c] = vaddq_f32(_mean.val[c], vll);
cachell.val[c] = vmulq_f32(cachell.val[c], _normal.val[c]);
float32x4_t vlh = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vl))); // 4..7
cachelh.val[c] = vaddq_f32(_mean.val[c], vlh);
cachelh.val[c] = vmulq_f32(cachelh.val[c], _normal.val[c]);
int16x8_t vh = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(v.val[c]))); // 8..15
// unpack to 32 bits
float32x4_t vhl = vcvtq_f32_s32(vmovl_s16(vget_low_s16(vh))); // 8..11
cachehl.val[c] = vaddq_f32(_mean.val[c], vhl);
cachehl.val[c] = vmulq_f32(cachehl.val[c], _normal.val[c]);
float32x4_t vhh = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vh))); // 12..15
cachehh.val[c] = vaddq_f32(_mean.val[c], vhh);
cachehh.val[c] = vmulq_f32(cachehh.val[c], _normal.val[c]);
}
vst3q_f32(dest + 48 * i + 0 * 3, cachell);
vst3q_f32(dest + 48 * i + 4 * 3, cachelh);
vst3q_f32(dest + 48 * i + 8 * 3, cachehl);
vst3q_f32(dest + 48 * i + 12 * 3, cachehh);
}
int remain = size * 16;
for (int i = remain; i < count; i++) {
dest[3 * i + 0] = normal[0] * (source[3 * i + 0] - mean[0]);
dest[3 * i + 1] = normal[1] * (source[3 * i + 1] - mean[1]);
dest[3 * i + 2] = normal[2] * (source[3 * i + 2] - mean[2]);
}
#else
int remain = 0;
#ifdef MNN_USE_SSE
int countC4 = count / 4;
if (countC4 > 1) {
if ((count % 4) * 3 < 4) {
//Avoid load extra memory
countC4 -=1;
}
// RGBRGBRGBRGB -> RGBR , GBRG, BRGB
auto alpha0 = _mm_setr_ps(normal[0], normal[1], normal[2], normal[0]);
auto alpha1 = _mm_setr_ps(normal[1], normal[2], normal[0], normal[1]);
auto alpha2 = _mm_setr_ps(normal[2], normal[0], normal[1], normal[2]);
auto beta0 = _mm_setr_ps(mean[0], mean[1], mean[2], mean[0]);
auto beta1 = _mm_setr_ps(mean[1], mean[2], mean[0], mean[1]);
auto beta2 = _mm_setr_ps(mean[2], mean[0], mean[1], mean[2]);
remain = countC4 * 4;
const __m128i gM = _mm_setr_epi8(4, 5, 6, 7, 6,5,4,7, 10,9,8,11, 14,13,12,15);
const __m128i bM = _mm_setr_epi8(8, 9,10,11, 6,5,4,7, 10,9,8,11, 14,13,12,15);
for (int i = 0; i < countC4; ++i) {
auto sInt8 = _mm_loadu_si128((const __m128i*)(source + 12 * i));
auto s0 = _mm_cvtepu8_epi32(sInt8);
auto s1 = _mm_cvtepu8_epi32(_mm_shuffle_epi8(sInt8, gM));
auto s2 = _mm_cvtepu8_epi32(_mm_shuffle_epi8(sInt8, bM));
auto f0 = _mm_cvtepi32_ps(s0);
auto f1 = _mm_cvtepi32_ps(s1);
auto f2 = _mm_cvtepi32_ps(s2);
_mm_storeu_ps(dest + 12 * i + 4 * 0, _mm_mul_ps(_mm_sub_ps(f0, beta0), alpha0));
_mm_storeu_ps(dest + 12 * i + 4 * 1, _mm_mul_ps(_mm_sub_ps(f1, beta1), alpha1));
_mm_storeu_ps(dest + 12 * i + 4 * 2, _mm_mul_ps(_mm_sub_ps(f2, beta2), alpha2));
}
}
#endif
for (int i = remain; i < count; ++i) {
dest[3 * i + 0] = normal[0] * (source[3 * i + 0] - mean[0]);
dest[3 * i + 1] = normal[1] * (source[3 * i + 1] - mean[1]);
dest[3 * i + 2] = normal[2] * (source[3 * i + 2] - mean[2]);
}
#endif
}
void MNNBlitC4ToFloatC4(const unsigned char* source, float* dest, const float* mean, const float* normal,
size_t count) {
for (int i = 0; i < count; ++i) {
dest[4 * i + 0] = normal[0] * (source[4 * i + 0] - mean[0]);
dest[4 * i + 1] = normal[1] * (source[4 * i + 1] - mean[1]);
dest[4 * i + 2] = normal[2] * (source[4 * i + 2] - mean[2]);
dest[4 * i + 3] = normal[3] * (source[4 * i + 3] - mean[3]);
}
}
#ifndef MNN_USE_NEON
void MNNBlitC1ToFloatRGBA(const unsigned char* source, float* dest, const float* mean, const float* normal,
size_t count) {
// MNN_PRINT("normal = %f\n", normal[0]);
::memset(dest, 0, 4 * sizeof(float) * count);
int remain = 0;
#ifdef MNN_USE_SSE
int countC16 = count / 16;
remain = countC16 * 16;
if (countC16 > 0) {
const __m128i gM = _mm_setr_epi8(4, 5, 6, 7, 6,5,4,7, 10,9,8,11, 14,13,12,15);
const __m128i bM = _mm_setr_epi8(8, 9,10,11, 6,5,4,7, 10,9,8,11, 14,13,12,15);
const __m128i aM = _mm_setr_epi8(12,13,14,15, 6,5,4,7, 10,9,8,11, 14,13,12,15);
auto normalC4 = _mm_set1_ps(normal[0]);
auto meanC4 = _mm_set1_ps(mean[0]);
for (int i=0; i<countC16; ++i) {
auto sInt8 = _mm_loadu_si128((const __m128i*)(source + 16 * i));
auto s0 = _mm_cvtepu8_epi32(sInt8);
auto s1 = _mm_cvtepu8_epi32(_mm_shuffle_epi8(sInt8, gM));
auto s2 = _mm_cvtepu8_epi32(_mm_shuffle_epi8(sInt8, bM));
auto s3 = _mm_cvtepu8_epi32(_mm_shuffle_epi8(sInt8, aM));
auto float00 = _mm_cvtepi32_ps(s0);
auto float01 = _mm_cvtepi32_ps(s1);
auto float10 = _mm_cvtepi32_ps(s2);
auto float11 = _mm_cvtepi32_ps(s3);
auto f0 = _mm_mul_ps(_mm_sub_ps(float00, meanC4), normalC4);
auto f1 = _mm_mul_ps(_mm_sub_ps(float01, meanC4), normalC4);
auto f2 = _mm_mul_ps(_mm_sub_ps(float10, meanC4), normalC4);
auto f3 = _mm_mul_ps(_mm_sub_ps(float11, meanC4), normalC4);
auto r1 = _mm_set1_ps(0.0f);
auto r2 = _mm_set1_ps(0.0f);
auto r3 = _mm_set1_ps(0.0f);
auto curDst = dest + 4 * i * 16;
_MM_TRANSPOSE4_PS(f0, r1, r2, r3);
_mm_store_ps(curDst + 4 * 0, f0);
_mm_store_ps(curDst + 4 * 1, r1);
_mm_store_ps(curDst + 4 * 2, r2);
_mm_store_ps(curDst + 4 * 3, r3);
curDst += 16;
_MM_TRANSPOSE4_PS(f1, r1, r2, r3);
_mm_store_ps(curDst + 4 * 0, f1);
_mm_store_ps(curDst + 4 * 1, r1);
_mm_store_ps(curDst + 4 * 2, r2);
_mm_store_ps(curDst + 4 * 3, r3);
curDst += 16;
_MM_TRANSPOSE4_PS(f2, r1, r2, r3);
_mm_store_ps(curDst + 4 * 0, f2);
_mm_store_ps(curDst + 4 * 1, r1);
_mm_store_ps(curDst + 4 * 2, r2);
_mm_store_ps(curDst + 4 * 3, r3);
curDst += 16;
_MM_TRANSPOSE4_PS(f3, r1, r2, r3);
_mm_store_ps(curDst + 4 * 0, f3);
_mm_store_ps(curDst + 4 * 1, r1);
_mm_store_ps(curDst + 4 * 2, r2);
_mm_store_ps(curDst + 4 * 3, r3);
}
}
#endif
for (int i = remain; i < count; ++i) {
dest[4 * i + 0] = normal[0] * (source[i + 0] - mean[0]);
}
}
void MNNBlitC3ToFloatRGBA(const unsigned char* source, float* dest, const float* mean, const float* normal,
size_t count) {
int remain = 0;
#ifdef MNN_USE_SSE
int countC4 = count / 4;
if (countC4 > 1) {
if ((count % 4) * 3 < 4) {
//Avoid load extra memory
countC4 -=1;
}
// RGBRGBRGBRGB -> RGB0 , RGB0, RGB0
auto alpha0 = _mm_setr_ps(normal[0], normal[1], normal[2], 0.0f);
auto beta0 = _mm_setr_ps(mean[0], mean[1], mean[2], 0.0f);
remain = countC4 * 4;
const __m128i rM = _mm_setr_epi8(0, 1, 2, 0, 6,5,4,7, 10,9,8,11, 14,13,12,15);
const __m128i gM = _mm_setr_epi8(3, 4, 5, 3, 6,5,4,7, 10,9,8,11, 14,13,12,15);
const __m128i bM = _mm_setr_epi8(6, 7, 8, 6, 6,5,4,7, 10,9,8,11, 14,13,12,15);
const __m128i aM = _mm_setr_epi8(9, 10, 11, 9, 6,5,4,7, 10,9,8,11, 14,13,12,15);
for (int i = 0; i < countC4; ++i) {
auto sInt8 = _mm_loadu_si128((const __m128i*)(source + 12 * i));
auto s0 = _mm_cvtepu8_epi32(_mm_shuffle_epi8(sInt8, rM));
auto s1 = _mm_cvtepu8_epi32(_mm_shuffle_epi8(sInt8, gM));
auto s2 = _mm_cvtepu8_epi32(_mm_shuffle_epi8(sInt8, bM));
auto s3 = _mm_cvtepu8_epi32(_mm_shuffle_epi8(sInt8, aM));
auto f0 = _mm_cvtepi32_ps(s0);
auto f1 = _mm_cvtepi32_ps(s1);
auto f2 = _mm_cvtepi32_ps(s2);
auto f3 = _mm_cvtepi32_ps(s3);
_mm_storeu_ps(dest + 16 * i + 4 * 0, _mm_mul_ps(_mm_sub_ps(f0, beta0), alpha0));
_mm_storeu_ps(dest + 16 * i + 4 * 1, _mm_mul_ps(_mm_sub_ps(f1, beta0), alpha0));
_mm_storeu_ps(dest + 16 * i + 4 * 2, _mm_mul_ps(_mm_sub_ps(f2, beta0), alpha0));
_mm_storeu_ps(dest + 16 * i + 4 * 3, _mm_mul_ps(_mm_sub_ps(f3, beta0), alpha0));
}
}
#endif
for (int i = remain; i < count; ++i) {
dest[4 * i + 0] = normal[0] * (source[3 * i + 0] - mean[0]);
dest[4 * i + 1] = normal[1] * (source[3 * i + 1] - mean[1]);
dest[4 * i + 2] = normal[2] * (source[3 * i + 2] - mean[2]);
dest[4 * i + 3] = 0.0f;
}
}
#endif
ImageFloatBlitter::BLIT_FLOAT ImageFloatBlitter::choose(ImageFormat format, int dstBpp) {
if (4 == dstBpp) {
switch (format) {
case GRAY:
return MNNBlitC1ToFloatRGBA;
case RGBA:
case BGRA:
return MNNBlitC4ToFloatC4;
case RGB:
case BGR:
return MNNBlitC3ToFloatRGBA;
default:
break;
}
}
switch (format) {
case GRAY:
return _blitC1ToFloatC1;
case RGBA:
case BGRA:
return MNNBlitC4ToFloatC4;
case RGB:
case BGR:
return _blitC3ToFloatC3;
default:
break;
}
return nullptr;
}
} // namespace CV
} // namespace MNN
Reference in New Issue View Git Blame Copy Permalink