欧美free性护士vide0shd,老熟女,一区二区三区,久久久久夜夜夜精品国产,久久久久久综合网天天,欧美成人护士h版

首頁綜合正文

評論

柚子快報激活碼778899分享：縮放算法優(yōu)化步驟詳解

Snap優(yōu)選網(wǎng)綜合2025-05-05440

柚子快報激活碼778899分享：縮放算法優(yōu)化步驟詳解

http://yzkb.51969.com/

添加鏈接描述

背景

假設數(shù)據(jù)存放在在unsigned char* m_pData 里面，寬和高分別是：m_nDataWidth m_nDataHeight 給定縮放比例：fXZoom fYZoom，返回縮放后的unsigned char* dataZoom 這里采用最簡單的縮放算法即：根據(jù)比例計算原圖和縮放后圖坐標的對應關(guān)系：縮放后圖坐標*縮放比例 = 原圖坐標

原始代碼未優(yōu)化

#pragma once

class zoomBlock

{

public:

zoomBlock() {};

~zoomBlock();

void zoomDataSSE128(unsigned char* dataZoom, float fXZoom, float fYZoom);

void zoomData(unsigned char* dataZoom, float fXZoom, float fYZoom);

void test(float fXZoom =0.5, float fYZoom=0.5);

void init(int DataWidth, int DataHeight);

private:

void computeSrcValues(int* srcValues, size_t size, float zoom, int dataSize);

private:

unsigned char* m_pData = nullptr;

float m_fXZoom = 1 ;//x軸縮放比例 m_nXZoom=1時不縮放

float m_fYZoom = 1 ;//y軸縮放比例

int m_nDataWidth = 0;

int m_nDataHeight = 0;

};

#include "zoomBlock.h"

#include

#define SAFE_DELETE_ARRAY(p) { if( (p) != NULL ) delete[] (p); (p) = NULL; }

zoomBlock::~zoomBlock()

{

SAFE_DELETE_ARRAY(m_pData);

}

void zoomBlock::init(int DataWidth, int DataHeight)

{

m_nDataWidth = DataWidth;

m_nDataHeight = DataHeight;

m_pData = new unsigned char[m_nDataWidth* m_nDataHeight];

for (int i = 0; i < m_nDataWidth * m_nDataHeight; ++i)

{

m_pData[i] = static_cast(i); // Replace this with your data initialization logic

}

void zoomBlock::zoomData(unsigned char* dataZoom, float fXZoom, float fYZoom)

{

int nZoomDataWidth = fXZoom * m_nDataWidth;

int nZoomDataHeight = fYZoom * m_nDataHeight;

for (size_t row = 0; row < nZoomDataHeight; row++)

{

for (size_t column = 0; column < nZoomDataWidth; column ++)

{

//1

int srcx = std::min(int(row / fYZoom), m_nDataHeight - 1);

int srcy = std::min(int(column / fXZoom), m_nDataWidth - 1);

//2

int srcPos = srcx * m_nDataHeight + srcy;

int desPos = row * nZoomDataHeight + column;

dataZoom[desPos] = m_pData[srcPos];

}

void zoomBlock::test(float fXZoom, float fYZoom)

{

init(8,8);

std::cout << "Values in m_pData:" << std::endl;

for (int i = 0; i < m_nDataWidth * m_nDataHeight; ++i)

{

std::cout << std::setw(4) << static_cast(m_pData[i]) << " ";

if ((i + 1) % m_nDataWidth == 0)

{ // Adjust the value based on your data

std::cout << std::endl;

}

unsigned char* dataZoom = new unsigned char[fXZoom * m_nDataWidth * fYZoom * m_nDataHeight];

zoomData(dataZoom, fXZoom, fYZoom);

// Print or inspect the values in m_dataZoom

int nZoomDataWidth = fXZoom * m_nDataWidth;

int nZoomDataHeight = fYZoom * m_nDataHeight;

std::cout << "Values in m_dataZoom:" << std::endl;

for (int i = 0; i < nZoomDataHeight * nZoomDataWidth; ++i)

{

std::cout << std::setw(4)<< static_cast(dataZoom[i]) << " ";

if ((i + 1) % nZoomDataWidth == 0) { // Adjust the value based on your data

std::cout << std::endl;

}

SAFE_DELETE_ARRAY(dataZoom);

}

測試代碼

int main()

{

zoomBlock zoomBlocktest;

zoomBlocktest.test(1.5,1.5);

return 0;

}

其中函數(shù) ·void zoomBlock::zoomData(unsigned char* dataZoom, float fXZoom, float fYZoom)· 沒有使用任何加速優(yōu)化，現(xiàn)在來分析它。

sse128

我們知道sse128可以一次性處理4個int類型，所以我們把最后一層for循環(huán)改成，4個坐標的算法，不滿4個的單獨計算

void zoomBlock::zoomDataSSE128(unsigned char* dataZoom, float fXZoom, float fYZoom)

{

int nZoomDataWidth = fXZoom * m_nDataWidth;

int nZoomDataHeight = fYZoom * m_nDataHeight;

for (size_t row = 0; row < nZoomDataHeight; row++)

{

int remian = nZoomDataWidth % 4;

for (size_t column = 0; column < nZoomDataWidth - remian; column += 4)

{

//第一個坐標

int srcx = std::min(int(row / fYZoom), m_nDataHeight - 1);

int srcy = std::min(int(column / fXZoom), m_nDataWidth - 1);

int srcPos = srcx * m_nDataHeight + srcy;

int desPos = row * nZoomDataHeight + column;

dataZoom[desPos] = m_pData[srcPos];

//第二個坐標

int srcx1 = std::min(int((row+1) / fYZoom), m_nDataHeight - 1);

int srcy1 = std::min(int((column+1) / fXZoom), m_nDataWidth - 1);

int srcPos1 = srcx1 * m_nDataHeight + srcy1;

int desPos1 = (row+1) * nZoomDataHeight + column+1;

dataZoom[desPos1] = m_pData[srcPos1];

//第3個坐標

// 。。。

//第4個坐標

// 。。。

}

// Process the remaining elements (if any) without SSE

for (size_t column = nZoomDataWidth - remian; column < nZoomDataWidth; column++)

{

int srcx = std::min(int(row / fYZoom), m_nDataHeight - 1);

int srcy = std::min(int(column / fXZoom), m_nDataWidth - 1);

int srcPos = srcx * m_nDataHeight + srcy;

int desPos = row * nZoomDataHeight + column;

dataZoom[desPos] = m_pData[srcPos];

}

上面一次處理四個坐標的代碼要改成sse的代碼

在最里層的循環(huán)里面，每次都要計算 row / fYZoom 和 column / fXZoom，這個實際上可以挪出for循環(huán)，計算一次存到數(shù)組里

數(shù)據(jù)坐標desPos和srcPos ，必須放在最內(nèi)存的循環(huán)里

所以我們用calculateSrcIndex函數(shù)單獨處理 row / fYZoom 和 column / fXZoom，希望達到如下效果：

void calculateSrcIndex(int* srcValues, int size, float zoom,int max)

{

for (int i = 0; i < size; i++)

{

srcValues[i] = std::min(int(i/zoom),max);

}

改成sse：

void calculateSrcIndex(int* srcValues, int size, float zoom,int max)

{

__m128i mmIndex, mmSrcValue, mmMax;

mmMax = _mm_set1_epi32(max);

float zoomReciprocal = 1.0f / zoom;

int remian = size % 4;

for (size_t i = 0; i < size - remian; i += 4)

{

mmIndex = _mm_set_epi32(i + 3, i + 2, i + 1, i);

mmSrcValue = _mm_cvtps_epi32(_mm_mul_ps(_mm_cvtepi32_ps(mmIndex), _mm_set1_ps(zoomReciprocal)));

// Ensure srcValues are within the valid range [0, max]

mmSrcValue = _mm_min_epi32(mmSrcValue, mmMax);

// Store the result to the srcValues array

_mm_storeu_si128(reinterpret_cast<__m128i*>(&srcValues[i]), mmSrcValue);

}

// Process the remaining elements (if any) without SSE

for (size_t i = size - remian; i < size; i++)

{

srcValues[i] = std::min(int(i / zoom), max);

}

解釋：這里主要處理int型數(shù)據(jù)，為了使用sse加速，要使用__m128i類型來存儲4個int

加載int到__m128i：

__m128i _mm_set1_epi32(int i); 這個指令是使用1個i，來設置__m128i，將__m128i看做4個32位的部分，則每個部分都被賦為i； __m128i _mm_set_epi32(int i3, int i2,int i1, int i0); 說明：使用4個int(32bits)變量來設置__m128i變量；返回值：如果返回值__m128i，分為r0,r1,r2,r3返回值規(guī)則如下：

r0 := i0 r1 := i1 r2 := i2 r3 := i3

__m128i _mm_cvtps_epi32 (__m128 a) Converts packed 32-bit integers in a to packed single-precision (32-bit) floating-point elements.

加載float到__m128

__m128 _mm_set1_ps(float w) 對應于_mm_load1_ps的功能，不需要字節(jié)對齊，需要多條指令。（r0 = r1 = r2 = r3 = w）__m128 _mm_cvtepi32_ps (__m128i a) Converts packed 32-bit integers in a to packed single-precision (32-bit) floating-point elements.

float乘法

__m128 dst = _mm_mul_ps (__m128 a, __m128 b) 將a, b中的32位浮點數(shù)相乘，結(jié)果打包給dst

取最小值

__m128i _mm_min_epi32 (__m128i a, __m128i b) Compare packed signed 32-bit integers in a and b, and store packed minimum values in dst. Operation FOR j := 0 to 3 i := j*32 dst[i+31:i] := MIN(a[i+31:i], b[i+31:i]) ENDFOR

所以代碼修改為

int* srcX = new int[nZoomDataHeight];

int* srcY = new int[nZoomDataWidth];

calculateSrcIndex(srcX, nZoomDataHeight, fXZoom , m_nDataHeight - 1);

calculateSrcIndex(srcY, nZoomDataWidth, fYZoom, m_nDataWidth - 1);

for (size_t row = 0; row < nZoomDataHeight; row++)

{

int remian = nZoomDataWidth % 4;

for (size_t column = 0; column < nZoomDataWidth - remian; column += 4)

{

//第一個坐標

int srcPos = srcX[row] * m_nDataHeight + srcY[column];

int desPos = row * nZoomDataHeight + column;

dataZoom[desPos] = m_pData[srcPos];

...

}

然后把坐標的計算轉(zhuǎn)為sse

void zoomBlock::zoomDataSSE128(unsigned char* dataZoom, float fXZoom, float fYZoom)

{

int nZoomDataWidth = fXZoom * m_nDataWidth;

int nZoomDataHeight = fYZoom * m_nDataHeight;

int* srcX = new int[nZoomDataWidth];

int* srcY = new int[nZoomDataHeight];

calculateSrcIndex(srcX, nZoomDataWidth, fXZoom, m_nDataWidth - 1);

calculateSrcIndex(srcY, nZoomDataHeight, fYZoom, m_nDataHeight - 1);

for (size_t y = 0; y < nZoomDataHeight; y++)

{

int remian = nZoomDataWidth % 4;

for (size_t x = 0; x < nZoomDataWidth - remian; x += 4)

{

__m128i mmsrcX = _mm_set_epi32(srcX[x + 3], srcX[x + 2], srcX[x+1], srcX[x]);

__m128i srcPosIndices = _mm_add_epi32(

_mm_set1_epi32(srcY[y] * m_nDataWidth),

mmsrcX);

__m128i desPosIndices = _mm_add_epi32(

_mm_set1_epi32(y * nZoomDataWidth),

_mm_set_epi32(x + 3, x + 2, x + 1, x)

);

dataZoom[desPosIndices.m128i_i32[0]] = m_pData[srcPosIndices.m128i_i32[0]];

dataZoom[desPosIndices.m128i_i32[1]] = m_pData[srcPosIndices.m128i_i32[1]];

dataZoom[desPosIndices.m128i_i32[2]] = m_pData[srcPosIndices.m128i_i32[2]];

dataZoom[desPosIndices.m128i_i32[3]] = m_pData[srcPosIndices.m128i_i32[3]];

/*cout << "srcPosIndices: " << srcPosIndices.m128i_i32[0] << " , desPosIndices : " << desPosIndices.m128i_i32[0] << endl;

cout << "srcPosIndices: " << srcPosIndices.m128i_i32[1] << " , desPosIndices : " << desPosIndices.m128i_i32[1] << endl;

cout << "srcPosIndices: " << srcPosIndices.m128i_i32[2] << " , desPosIndices : " << desPosIndices.m128i_i32[2] << endl;

cout << "srcPosIndices: " << srcPosIndices.m128i_i32[3] << " , desPosIndices : " << desPosIndices.m128i_i32[3] << endl;*/

}

// Process the remaining elements (if any) without SSE

for (size_t x = nZoomDataWidth - remian; x < nZoomDataWidth; x++)

{

int srcy = std::min(int(y / fYZoom), m_nDataHeight - 1);

int srcx = std::min(int(x / fXZoom), m_nDataWidth - 1);

int srcPos = srcy * m_nDataHeight + srcx;

int desPos = y * nZoomDataHeight + x;

dataZoom[desPos] = m_pData[srcPos];

}

delete[] srcX;

delete[] srcY;

}

完整的代碼

#pragma once

class zoomBlock

{

public:

zoomBlock() {};

~zoomBlock();

void zoomDataSSE128(unsigned char* dataZoom, float fXZoom, float fYZoom);

void zoomData(unsigned char* dataZoom, float fXZoom, float fYZoom);

void test(float fXZoom =0.5, float fYZoom=0.5);

void init(int DataWidth, int DataHeight);

private:

inline void calculateSrcIndex(int* srcValues, int size, float zoom, int max);

private:

unsigned char* m_pData = nullptr;

float m_fXZoom = 1 ;//x軸縮放比例 m_nXZoom=1時不縮放

float m_fYZoom = 1 ;//y軸縮放比例

int m_nDataWidth = 0;

int m_nDataHeight = 0;

};

#include "zoomBlock.h"

#include

using namespace std;

#define SAFE_DELETE_ARRAY(p) { if( (p) != NULL ) delete[] (p); (p) = NULL; }

zoomBlock::~zoomBlock()

{

SAFE_DELETE_ARRAY(m_pData);

}

void zoomBlock::init(int DataWidth, int DataHeight)

{

m_nDataWidth = DataWidth;

m_nDataHeight = DataHeight;

m_pData = new unsigned char[m_nDataWidth* m_nDataHeight];

for (int i = 0; i < m_nDataWidth * m_nDataHeight; ++i)

{

m_pData[i] = static_cast(i); // Replace this with your data initialization logic

}

void zoomBlock::zoomData(unsigned char* dataZoom, float fXZoom, float fYZoom)

{

int nZoomDataWidth = fXZoom * m_nDataWidth;

int nZoomDataHeight = fYZoom * m_nDataHeight;

for (size_t y = 0; y < nZoomDataHeight; y++)

{

for (size_t x = 0; x < nZoomDataWidth; x ++)

{

//1

int srcy = std::min(int(y / fYZoom), m_nDataHeight - 1);

int srcx = std::min(int(x / fXZoom), m_nDataWidth - 1);

//2

int srcPos = srcy * m_nDataWidth + srcx;

int desPos = y * nZoomDataWidth + x;

dataZoom[desPos] = m_pData[srcPos];

}

inline void zoomBlock::calculateSrcIndex(int* srcValues, int size, float zoom,int max)

{

__m128i mmIndex, mmSrcValue, mmMax;

mmMax = _mm_set1_epi32(max);

float zoomReciprocal = 1.0f / zoom;

int remian = size % 4;

for (size_t i = 0; i < size - remian; i += 4)

{

mmIndex = _mm_set_epi32(i + 3, i + 2, i + 1, i);

mmSrcValue = _mm_cvttps_epi32(_mm_mul_ps(_mm_cvtepi32_ps(mmIndex), _mm_set1_ps(zoomReciprocal)));

// Ensure srcValues are within the valid range [0, max]

mmSrcValue = _mm_min_epi32(mmSrcValue, mmMax);

// Store the result to the srcValues array

_mm_storeu_si128(reinterpret_cast<__m128i*>(&srcValues[i]), mmSrcValue);

}

// Process the remaining elements (if any) without SSE

for (size_t i = size - remian; i < size; i++)

{

srcValues[i] = std::min(int(i / zoom), max);

}

void zoomBlock::zoomDataSSE128(unsigned char* dataZoom, float fXZoom, float fYZoom)

{

int nZoomDataWidth = fXZoom * m_nDataWidth;

int nZoomDataHeight = fYZoom * m_nDataHeight;

int* srcX = new int[nZoomDataWidth];

int* srcY = new int[nZoomDataHeight];

calculateSrcIndex(srcX, nZoomDataWidth, fXZoom, m_nDataWidth - 1);

calculateSrcIndex(srcY, nZoomDataHeight, fYZoom, m_nDataHeight - 1);

for (size_t y = 0; y < nZoomDataHeight; y++)

{

int remian = nZoomDataWidth % 4;

for (size_t x = 0; x < nZoomDataWidth - remian; x += 4)

{

/*int srcPos = srcx * m_nDataHeight + srcy;

int desPos = row * nZoomDataHeight + column;*/

//dataZoom[desPos] = m_pData[srcPos];

//__m128i mmsrcY = _mm_loadu_si128((__m128i*)(srcY));