opencl gauss filter优化(一)
Platform: LG G3, Adreno 330 ,img size 3264x2448
|
C code |
neon |
GPU |
|
300 |
60 |
29 |
单位:ms
1.
目前按如下行列分解的方式最快29ms,Horizontal
kernel
globalWorksize[1] = {height+256-height%256};Vertical kernel
globalWorksize2[1] = {width+256-width%256};
localWorksize2[]
= {64}; localWorksize2 手动设为64时最快。
Porfile的结果为:Horizontal
kernel 的wait
time 有11ms,实际rum
time 18ms.
这个wait
time是什么呢？注释掉Horizontal
kernel中的
vstore16(convert_uchar16(sum>>(ushort)8),0,pOutLine+j)
; 则wait
time只有0.x
ms.并且
localWorksize
越小wait
time越长,为1时达到200ms,16时20ms.
难道是写内存等待时间，没有足够的ALU指令隐藏访存延时？写内存后进入下一个for循环，马上又读内存，所以没有ALU指令隐藏这个延时。然而Horizontal
kernel的profile结果实际run
time只有0.x
ms,所有时间基本都是在wait.(更正：注释掉vstore16后,sum的计算被优化掉了，0.x
ms是读内存的时间)
__kernel void ImageGaussianFilterHorizontal(__global const uchar* restrict source, // Source image
__global uchar* restrict dest, // Intermediate dest image
const int imgWidth , // Image width
const int imgHeight)
{
const int y = get_global_id();
if(y>=(imgHeight))
return;
const uchar m_nRightShiftNum = ;
const uchar Rounding = ( << (m_nRightShiftNum - ));
const uchar m_nFilter[] = {,,,,,,,,,,}; const int s = ;
const int nStart = ;
const int nWidth = imgWidth; __global const uchar* pInLine = source + y*nWidth;
__global uchar* pOutLine = dest + y*nWidth;
int j;
for(j = ; j < nStart; j ++)
{
ushort sum = ; for (int m = ; m<s / ; m++)
{
int k1 = (j + m - nStart);
k1 = k1< ? -k1 : k1; int k2 = (j + nStart - m );
sum += (pInLine[k1] + pInLine[k2])*m_nFilter[m];
}
sum += pInLine[j] * m_nFilter[s / ];
sum = (sum + Rounding) >> ;
pOutLine[j] = (uchar)clamp(sum,(ushort),(ushort));
} for ( ; (j+)<= (nWidth - nStart); j+=)
{
#define GAUSSIAN_LINE_NEON(m) \
sum += ( convert_ushort16(vload16(,pInLine+j-nStart+m))* m_nFilter[m] ); ushort16 sum = (convert_ushort16(vload16(,pInLine+j-nStart)) * m_nFilter[]);
GAUSSIAN_LINE_NEON();
GAUSSIAN_LINE_NEON();
GAUSSIAN_LINE_NEON();
GAUSSIAN_LINE_NEON();
GAUSSIAN_LINE_NEON();
GAUSSIAN_LINE_NEON();
GAUSSIAN_LINE_NEON();
GAUSSIAN_LINE_NEON();
GAUSSIAN_LINE_NEON();
GAUSSIAN_LINE_NEON(); sum += (ushort)Rounding;
vstore16(convert_uchar16(sum>>(ushort)),,pOutLine+j) ;
} for( ; j < nWidth; j ++)
{
ushort sum = ; for (int m = ; m<s / ; m++)
{
int k1 = (j + m - nStart); int k2 = (j + nStart - m );
k2 = k2 >= nWidth ? * nWidth - - k2 : k2;
sum += (pInLine[k1] + pInLine[k2])*m_nFilter[m];
}
sum += pInLine[j] * m_nFilter[s / ];
sum = (sum + Rounding) >> m_nRightShiftNum;
pOutLine[j] = (uchar)clamp(sum,(ushort),(ushort));
}
} __kernel void ImageGaussianFilterVertical( __global uchar* restrict source, // Intermediate image processed by ImageGaussianFilterHorizontal()
__global uchar* restrict dest, // Final destination image
const int imgWidth,
const int imgHeight
)
{
const int x = get_global_id();
if(x>=(imgWidth))
return;
const int x_offset = x; const int s = ;
const int nStart = s / ;
const int m_nRightShiftNum = ;
const int Rounding = ( << (m_nRightShiftNum - ));
const uchar m_nFilter[] = {,,,,,,,,,,}; int y;
// mem_fence(CLK_LOCAL_MEM_FENCE); ushort lines[];
lines[nStart] = (ushort)( source[x_offset] );
for(y=;y<=nStart;y++)
{
lines[nStart+y] = (ushort)( source[y*imgWidth+x_offset] );
lines[nStart-y] = lines[nStart+y];
} for(y=;y<(imgHeight-nStart-);)
{ ushort sum = lines[nStart] * m_nFilter[nStart];
#define GaussianTwoLines(m) \
sum += ( (lines[m] + lines[s--m])*m_nFilter[m] );
GaussianTwoLines()
GaussianTwoLines()
GaussianTwoLines()
GaussianTwoLines()
GaussianTwoLines() sum += (ushort)Rounding;
dest[y*imgWidth+x_offset] = (uchar)(sum>>(ushort)); y++;
for(int i = ; i<s-; i++) lines[i] = lines[i+]; lines[s-] = (ushort)( source[(y+nStart)*imgWidth+x_offset] ); } for(y=imgHeight-nStart-;y<(imgHeight-);)
{
ushort sum = lines[nStart] * m_nFilter[nStart];
GaussianTwoLines()
GaussianTwoLines()
GaussianTwoLines()
GaussianTwoLines()
GaussianTwoLines() sum += (ushort)Rounding;
dest[y*imgWidth+x_offset] = (uchar)(sum>>(ushort)); y++;
for(int i = ; i<s-; i++) {
lines[i] = lines[i+];
}
lines[s-] = lines[(imgHeight-y)*-] ; //
}
//last y=imgHeight-1
ushort sum = lines[nStart] * m_nFilter[nStart];
GaussianTwoLines()
GaussianTwoLines()
GaussianTwoLines()
GaussianTwoLines()
GaussianTwoLines() sum += (ushort)Rounding;
dest[y*imgWidth+x_offset] = (uchar)(sum>>(ushort));
}
kernel
2.Horizontal kernel改进，预先load 2x16个所需的pixel,计算时从中提取,这样每次循环只需读一次内存。需要26ms,wait time 8ms.
ushort16 line0 = convert_ushort16(vload16(,pInLine+j-nStart));
for ( ; (j+)<= (nWidth - nStart); j+=)
{
ushort16 line1 = convert_ushort16(vload16(,pInLine+j-nStart+)); ushort16 temp0;
ushort16 temp1;
temp0 = line0;
temp1.s0123 = line0.sabcd;
temp1.s45 = line0.sef;
temp1.s67 = line1.s01;
temp1.s89abcdef = line1.s23456789;
ushort16 sum = ( temp0 + temp1 ) * m_nFilter[];
temp0.s0123456789abcdef = temp0.s123456789abcdeff;
temp0.sf = line1.s0;
temp1.s0123456789abcdef = temp1.s00123456789abcde;
temp1.s0 = line0.s9;
sum += ( temp0 + temp1 ) * m_nFilter[];
temp0.s0123456789abcdef = temp0.s123456789abcdeff;
temp0.sf = line1.s1;
temp1.s0123456789abcdef = temp1.s00123456789abcde;
temp1.s0 = line0.s8;
sum += ( temp0 + temp1 ) * m_nFilter[];
temp0.s0123456789abcdef = temp0.s123456789abcdeff;
temp0.sf = line1.s2;
temp1.s0123456789abcdef = temp1.s00123456789abcde;
temp1.s0 = line0.s7;
sum += ( temp0 + temp1 ) * m_nFilter[];
temp0.s0123456789abcdef = temp0.s123456789abcdeff;
temp0.sf = line1.s3;
temp1.s0123456789abcdef = temp1.s00123456789abcde;
temp1.s0 = line0.s6;
sum += ( temp0 + temp1 ) * m_nFilter[];
temp0.s0123456789abcdef = temp0.s123456789abcdeff;
temp0.sf = line1.s4;
sum += ( temp0 ) * m_nFilter[]; sum += (ushort)Rounding;
line0 = line1;
vstore16(convert_uchar16(sum>>(ushort)),,pOutLine+j) ;
}
3.不计算，只读写内存测试。那么wait time 3.2 ms,run time 18.2 ms.说明Horizontal kernel 耗时的极限也需3.2ms. 但是只是注释掉vstore16，还保留了读和计算，反而wait time还只有0.x ms，这又是为何？是读几乎没有wait,3.2ms都是写的wait time? (更正：注释掉vstore16后,sum的计算被优化掉了，0.x ms是读内存的时间)
a.再次测试，只有读wait time 0.xms ,只有写wait time 3.2ms.写比读的周期长.
for ( ; (j+16)<= (nWidth - nStart); j+=16)
{
ushort16 line1 = convert_ushort16(vload16(0,pInLine+j-nStart+16));
vstore16(0,0,pOutLine+j) ;
}
b.另外发现使用*((__global uint4*)(pOutLine+j)) = as_uint4(result);比vstore16快，wait time 2.5ms.高通 80-N8592-1_L_OpenCL_Programming_Guide 中提到：
Vectorized load/store of a larger data type is more optimal than a small data type; e.g., a load of uint2* is more optimal than uchar8* .
For optimal SP to L2 bandwidth performance, align read access to a 32-bit address and write access to a 128-bit address.
c.原来写的内存没有对齐，使用*((__global uint4*)(pOutLine+j-5)) = as_uint4(result);wait time 1.9ms.
d.最后加上sum计算，采用的Horizontal kernel如下，localWorksize[] = {64};时时间最少，需要23ms,wait time 4.7ms , localWorksize = 128时,wait 6ms.
并且使用__attribute__((work_group_size_hint(64,1,1))) ,耗时22ms.
__kernel __attribute__((work_group_size_hint(,,)))
void ImageGaussianFilterHorizontal(__global const uchar* restrict source, // Source image
__global uchar* restrict dest, // Intermediate dest image
const int imgWidth , // Image width
const int imgHeight)
{
const int y = get_global_id();
if(y>=(imgHeight))
return;
const uchar m_nRightShiftNum = ;
const uchar Rounding = ( << (m_nRightShiftNum - ));
const uchar m_nFilter[] = {,,,,,,,,,,}; const int s = ;
const int nStart = ;
const int nWidth = imgWidth; __global const uchar* pInLine = source + y*nWidth;
__global uchar* pOutLine = dest + y*nWidth; int j;
uchar temp[];
for(j = ; j < nStart; j ++)
{
ushort sum = ; for (int m = ; m<s / ; m++)
{
int k1 = (j + m - nStart);
k1 = k1< ? -k1 : k1; int k2 = (j + nStart - m );
sum += (pInLine[k1] + pInLine[k2])*m_nFilter[m];
}
sum += pInLine[j] * m_nFilter[s / ];
sum = (sum + Rounding) >> ;
temp[j] = (uchar)clamp(sum,(ushort),(ushort));
} uchar16 result,pre_result;
pre_result.sbcde = (uchar4)(temp[],temp[],temp[],temp[]);
pre_result.sf = temp[]; ushort16 line0 = convert_ushort16(vload16(,pInLine+j-nStart));
for ( ; (j+)<= (nWidth - nStart); j+=)
{
//prefetch(pInLine+j-nStart,32); //无变化
ushort16 line1 = convert_ushort16(vload16(,pInLine+j-nStart+)); ushort16 temp0;
ushort16 temp1;
temp0 = line0;
temp1.s0123 = line0.sabcd;
temp1.s45 = line0.sef;
temp1.s67 = line1.s01;
temp1.s89abcdef = line1.s23456789;
ushort16 sum = ( temp0 + temp1 ) * m_nFilter[];
temp0.s0123456789abcdef = temp0.s123456789abcdeff;
temp0.sf = line1.s0;
temp1.s0123456789abcdef = temp1.s00123456789abcde;
temp1.s0 = line0.s9;
sum += ( temp0 + temp1 ) * m_nFilter[];
temp0.s0123456789abcdef = temp0.s123456789abcdeff;
temp0.sf = line1.s1;
temp1.s0123456789abcdef = temp1.s00123456789abcde;
temp1.s0 = line0.s8;
sum += ( temp0 + temp1 ) * m_nFilter[];
temp0.s0123456789abcdef = temp0.s123456789abcdeff;
temp0.sf = line1.s2;
temp1.s0123456789abcdef = temp1.s00123456789abcde;
temp1.s0 = line0.s7;
sum += ( temp0 + temp1 ) * m_nFilter[];
temp0.s0123456789abcdef = temp0.s123456789abcdeff;
temp0.sf = line1.s3;
temp1.s0123456789abcdef = temp1.s00123456789abcde;
temp1.s0 = line0.s6;
sum += ( temp0 + temp1 ) * m_nFilter[];
temp0.s0123456789abcdef = temp0.s123456789abcdeff;
temp0.sf = line1.s4;
sum += ( temp0 ) * m_nFilter[]; sum += (ushort)Rounding;
line0 = line1; result.s0123 = pre_result.sbcde;
result.s4 = pre_result.sf;
pre_result = convert_uchar16(sum>>(ushort)) ; result.s5 = pre_result.s0;
result.s67 = pre_result.s12;
result.s89abcdef = pre_result.s3456789a;
*( (__global uint4*)(pOutLine+j-) ) = (as_uint4)(result) ;
} *( (__global uint*)(pOutLine+j-) ) = (as_uint)(pre_result.sbcde);//last 5 bytes
pOutLine[j-] = pre_result.sf; for( ; j < nWidth; j ++)
{
ushort sum = ; for (int m = ; m<s / ; m++)
{
int k1 = (j + m - nStart); int k2 = (j + nStart - m );
k2 = k2 >= nWidth ? * nWidth - - k2 : k2;
sum += (pInLine[k1] + pInLine[k2])*m_nFilter[m];
}
sum += pInLine[j] * m_nFilter[s / ];
sum = (sum + Rounding) >> m_nRightShiftNum;
pOutLine[j] = (uchar)clamp(sum,(ushort),(ushort));
}
}
opencl gauss filter优化(一)的更多相关文章
- opencl gauss filter优化(三)
1.根据前两次的最终结果: 使用普通buffer,Horizontal 5ms, Vertical 17 ms 使用image buffer:Horizontal 9.4ms, Vertical 6. ...
- opencl gauss filter优化(二)
1.buffer使用image的方式:Horizontal 与 Vertical 算法一样, 共需30ms,wait time 19ms. const sampler_t sampler = CLK_ ...
- Anisotropic gauss filter
最近一直在做版面分析,其中文本行检测方面,许多文章涉及到了Anigauss也就是各向异性高斯滤波. 顾名思义,简单的理解就是参数不同的二维高斯滤波. 在文章Fast Anisotropic Gauss ...
- OpenCL Kernel设计优化
使用Intel® FPGA SDK for OpenCL™ 离线编译器,不需要调整kernel代码便可以将其最佳的适应于固定的硬件设备,而是离线编译器会根据kernel的要求自适应调整硬件的结构. 通 ...
- FILTER优化
explain plan for select a.* from fxqd_list_20131115_new_100 a where (acct_no, oper_no, seqno, trans_ ...
- 二维高斯滤波器（gauss filter）的实现
我们以一个二维矩阵表示二元高斯滤波器,显然此二维矩阵的具体形式仅于其形状(shape)有关: def gauss_filter(kernel_shape): 为实现二维高斯滤波器,需要首先定义二元高斯 ...
- 一次性能优化将filter转换
有一条SQL性能有问题,在运行计划中发现filter.遇到它要小心了,类似于nestloop.我曾经的blog对它有研究探索运行计划中filter的原理.用exists极易引起filter. 优化前: ...
- 安卓平台ARM Mali OpenCL例子-灰度转换（转）
手头一块RK3288的板子,在板子上测试了一张1080p的彩色图灰度转换的OpenCL例子.OpenCL没有任何优化.例子请移步这里. 该例子是编译成安卓平台下的可执行程序. 进入jni文件夹,进行如 ...
- OpenCV、OpenCL、OpenGL、OpenPCL
对于几个开源库的总结,作为标记,以前看过,现在开始重视起来!更详细资料请移步 开源中国社区! 涉及:OpenCV,OpenCL,OpenGL,OpenPCL 截止到目前: OpenGL的最新版本为4. ...
随机推荐
- logo上线
- 【算法题目】包含min函数的栈
题目来源:<剑指offer>面试题21 题目:定义栈的数据结构,请在该类型中实现一个能够得到栈的最小元素的min函数.在该栈中,调用min,push以及pop的时间复杂度都是O(1). 分 ...
- 《利用python进行数据分析》读书笔记--第十章 时间序列（三）
7.时间序列绘图 pandas时间序列的绘图功能在日期格式化方面比matplotlib原生的要好. #-*- coding:utf-8 -*- import numpy as np import pa ...
- web错误
“/Web”应用程序中的服务器错误. 服务器 'LD-PC' 上的 MSDTC 不可用. 说明: 执行当前 Web 请求期间,出现未经处理的异常.请检查堆栈跟踪信息,以了解有关该错误以及代码中导致错误 ...
- Jnotify文件监控的用法以及Jar文件导入的方法
简介Jnotiy, 支持动态监控(支持级联监控)文件夹和文件的jar包.在linux中,调用linux底层的jnotify服务.在windows中,需要添加附件的dll文件. 因为通用的Maven仓库 ...
- SqlSever基础 getdate函数 返回系统当前的年月日，时分秒 毫秒
镇场诗:---大梦谁觉,水月中建博客.百千磨难,才知世事无常.---今持佛语,技术无量愿学.愿尽所学,铸一良心博客.------------------------------------------ ...
- 高通、猎户机型Android典型bootloader分析
1.bootloader是什么? 简单地说,bootloader 就是在操作系统内核运行之前运行的一段小程序.通过这段小程序,我们可以初始化硬件设备.建立内存空间的映射图,从而将系统的软硬件环境带到一 ...
- 【leetcode❤python】 88. Merge Sorted Array
#-*- coding: UTF-8 -*-class Solution(object): def merge(self, nums1, m, nums2, n): "& ...
- 熟悉 Objective-C
## 1. OC 的起源 OC 的方法(本质上讲是消息)在运行时决定.使用函数调用的语言,由编译器决定.如果涉及多态,则用到虚函数表. ## 2. 少在头文件中引用其他文件 1. 两个头文件互相引用会 ...
- javascript encodeURI和encodeURIComponent的比较
总结:encodeURI对除三种字符()之外的字符进行编码 encodeURIComponent对除两种字符之外的字符进行编码,保留字符会被转义 在进行SaaS前端开发的时候,大家经常会用到两个Jav ...