基于NEON指令的图像旋转加速【armv7】
生活随笔
收集整理的這篇文章主要介紹了
基于NEON指令的图像旋转加速【armv7】
小編覺得挺不錯的,現(xiàn)在分享給大家,幫大家做個參考.
目錄
前言
知識直通車
NEON轉(zhuǎn)置指令
右旋90
4x4矩陣右旋實例
灰度圖（單通道）右旋90
彩圖（RGB三通道）右旋90
左旋90
4x4矩陣左旋實例
灰度圖（單通道）左旋90
彩圖（RGB三通道）左旋90?
RK3288-linux(硬浮點)?耗時對比
前言
圖像旋轉(zhuǎn)就是簡單的坐標(biāo)變換，1080p的圖像右旋90度耗時時間，在RK3288-linux(硬浮點)板子上，opencv4.1也就17ms搞定，本沒啥好講的，但是凡事追求極致嘛，能加速的就給他加速一把，當(dāng)然這態(tài)度的問題只是其中一個原因；另外一個原因是最近在公司遇到一個對旋轉(zhuǎn)90的加速需求，在某些場景下，需要做圖像旋轉(zhuǎn)然后再送入神經(jīng)網(wǎng)絡(luò)做人臉檢測，而由于整個人臉識別整個工程的運(yùn)行導(dǎo)致板子的資源極大消耗，使得簡單的圖像旋轉(zhuǎn)的時間延長至三四十ms，這是很可怕的，人臉檢測也就二三十ms左右吧，所以這是寫這篇文章的直接導(dǎo)火索。
當(dāng)然，類似的利用NEON對矩陣旋轉(zhuǎn)加速的解決方案，網(wǎng)上已經(jīng)有人發(fā)表過了，比如最早的文章https://www.cnblogs.com/hrlnw/p/3723072.html，本篇博客會在此基礎(chǔ)上進(jìn)行擴(kuò)展，增加RGB彩圖的加速旋轉(zhuǎn)代碼并補(bǔ)充opencv與本篇算法在RK3288上的旋轉(zhuǎn)耗時數(shù)據(jù)對比。
?
知識直通車
矩陣右旋90的NEON加速：https://www.cnblogs.com/hrlnw/p/3723072.html
矩陣右旋180度和270(-90)度：https://www.cnblogs.com/hrlnw/p/3767853.html
opencv右旋90、180、270度：https://www.cnblogs.com/alexYuin/p/9493242.html
opencv利用transpose+flip實現(xiàn)圖像旋轉(zhuǎn)，雖然比仿射變換函數(shù)warpaffine要快不少，但是分步的做法總是沒有一步到位的解決方案來得快的，接下來就開始介紹一步到位法了。
?
NEON轉(zhuǎn)置指令
可以用neon進(jìn)行旋轉(zhuǎn)加速的關(guān)鍵要感謝arm出品的vtrn_u8/u16/u32..系列的轉(zhuǎn)置指令，雖然是2x2的轉(zhuǎn)置指令，但是對于矩陣旋轉(zhuǎn)完全夠用，取決于怎么用了。
轉(zhuǎn)置的具體情況如下示意，應(yīng)該很清楚就不多少了
uint8x8_t mat0,mat1; uint8x8x2_t result=vtrn_u8(mat0,mat1);??? ? ? ? ? ? ?
?
右旋90
4x4矩陣右旋實例
在介紹圖形旋轉(zhuǎn)代碼之前先用個小矩陣實例說明一下算法的可行性。先說右旋90度：
? ? ? 旋轉(zhuǎn)前矩陣? ? ? ? ? ? ? ? ? ? ? ? ? ? 旋轉(zhuǎn)后矩陣
? ? ? ? ? ? ??
算法步驟：
1）相鄰兩行每2x2做轉(zhuǎn)置(但是注意輸入順序是倒的，比如前兩行，第二行作為vtrn的第一個參數(shù)輸入，第一行作為vtrn的第二個參數(shù)輸入)，結(jié)果如下：
? ? ? ? ? ?原矩陣? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?相鄰行倒置? ? ? ? ? ? ? ? ? ? ? ? ? ?相鄰行轉(zhuǎn)置? ? ? ? ? ? ? ? ? ? ? ? ??
? ? ? ? ? ? ? ?? ? ? ? ? ? ??
2）每行兩個像素值看做一個整體，隔行轉(zhuǎn)置，依然倒著輸入，即2、0行一組輸入，3、1行一組輸入，這樣便可以得到最終的矩陣旋轉(zhuǎn)結(jié)果了，過程如下：
? ? ? ?上一步結(jié)果? ? ? ? ? ? ? ? ? ? ? ? ? ? ? 隔行倒置? ? ? ? ? ? ? ? ? ? ? ? ? ? ? 隔行轉(zhuǎn)置
? ? ? ? ? ? ? ??? ? ? ? ? ? ?
例子看明白了，那下面圖像旋轉(zhuǎn)也就很簡單了，soeasy啊~
圖像旋轉(zhuǎn)與矩陣旋轉(zhuǎn)的區(qū)別就是.....嗯.....圖像比較大，下就分塊嘍，那從指令讀取數(shù)據(jù)的角度以及代碼量及速度方面的考慮我們將圖像分成8x8的小塊矩陣一個一個做旋轉(zhuǎn)然后再放到最終圖像相應(yīng)的位置就可以了。
灰度圖（單通道）右旋90
static void rotate_pos90_gray(uint8_t* src, size_t sstep, uint8_t* dst, size_t dstep, int w,int h) {uint8x8x4_t mat[2];uint8x8x2_t temp8x8x2[4];uint16x4x4_t temp16x4x4[2];uint16x4x2_t temp16x4x2[4];uint32x2x4_t temp32x2x4[2];uint32x2x2_t temp32x2x2[4];int dw=w&7; int sw=w-dw;int dh=h&7; int sh=h-dh;for(int y=0;y<sh;y+=8){uint8_t* pIndexY = src+y*sstep;int tr_x = h-y-8;for(int x=0;x<sw;x+=8){uint8_t* ptr_y = dst+x*dstep;//load 8x8 matrixmat[0].val[0] = vld1_u8(pIndexY+ x);mat[0].val[1] = vld1_u8(pIndexY+ sstep+x);mat[0].val[2] = vld1_u8(pIndexY+2*sstep+x);mat[0].val[3] = vld1_u8(pIndexY+3*sstep+x);mat[1].val[0] = vld1_u8(pIndexY+4*sstep+x);mat[1].val[1] = vld1_u8(pIndexY+5*sstep+x);mat[1].val[2] = vld1_u8(pIndexY+6*sstep+x);mat[1].val[3] = vld1_u8(pIndexY+7*sstep+x);//transpose for two nearby rows temp8x8x2[0] = vtrn_u8(mat[0].val[1],mat[0].val[0]); temp8x8x2[1] = vtrn_u8(mat[0].val[3],mat[0].val[2]); temp8x8x2[2] = vtrn_u8(mat[1].val[1],mat[1].val[0]); temp8x8x2[3] = vtrn_u8(mat[1].val[3],mat[1].val[2]); //interpret two 8bit to one 16bit datatemp16x4x4[0].val[0] = vreinterpret_u16_u8(temp8x8x2[0].val[0]);temp16x4x4[0].val[1] = vreinterpret_u16_u8(temp8x8x2[0].val[1]);temp16x4x4[0].val[2] = vreinterpret_u16_u8(temp8x8x2[1].val[0]);temp16x4x4[0].val[3] = vreinterpret_u16_u8(temp8x8x2[1].val[1]);temp16x4x4[1].val[0] = vreinterpret_u16_u8(temp8x8x2[2].val[0]);temp16x4x4[1].val[1] = vreinterpret_u16_u8(temp8x8x2[2].val[1]);temp16x4x4[1].val[2] = vreinterpret_u16_u8(temp8x8x2[3].val[0]);temp16x4x4[1].val[3] = vreinterpret_u16_u8(temp8x8x2[3].val[1]);//transpose for two interleaved rows temp16x4x2[0] = vtrn_u16(temp16x4x4[0].val[2],temp16x4x4[0].val[0]); temp16x4x2[1] = vtrn_u16(temp16x4x4[0].val[3],temp16x4x4[0].val[1]); temp16x4x2[2] = vtrn_u16(temp16x4x4[1].val[2],temp16x4x4[1].val[0]); temp16x4x2[3] = vtrn_u16(temp16x4x4[1].val[3],temp16x4x4[1].val[1]); temp32x2x4[0].val[0] = vreinterpret_u32_u16(temp16x4x2[0].val[0]);temp32x2x4[0].val[1] = vreinterpret_u32_u16(temp16x4x2[0].val[1]);temp32x2x4[0].val[2] = vreinterpret_u32_u16(temp16x4x2[1].val[0]);temp32x2x4[0].val[3] = vreinterpret_u32_u16(temp16x4x2[1].val[1]);temp32x2x4[1].val[0] = vreinterpret_u32_u16(temp16x4x2[2].val[0]);temp32x2x4[1].val[1] = vreinterpret_u32_u16(temp16x4x2[2].val[1]);temp32x2x4[1].val[2] = vreinterpret_u32_u16(temp16x4x2[3].val[0]);temp32x2x4[1].val[3] = vreinterpret_u32_u16(temp16x4x2[3].val[1]);temp32x2x2[0] = vtrn_u32(temp32x2x4[1].val[0],temp32x2x4[0].val[0]); temp32x2x2[1] = vtrn_u32(temp32x2x4[1].val[1],temp32x2x4[0].val[1]); temp32x2x2[2] = vtrn_u32(temp32x2x4[1].val[2],temp32x2x4[0].val[2]); temp32x2x2[3] = vtrn_u32(temp32x2x4[1].val[3],temp32x2x4[0].val[3]); mat[0].val[0] = vreinterpret_u8_u32(temp32x2x2[0].val[0]); mat[0].val[1] = vreinterpret_u8_u32(temp32x2x2[2].val[0]); mat[0].val[2] = vreinterpret_u8_u32(temp32x2x2[1].val[0]); mat[0].val[3] = vreinterpret_u8_u32(temp32x2x2[3].val[0]); mat[1].val[0] = vreinterpret_u8_u32(temp32x2x2[0].val[1]); mat[1].val[1] = vreinterpret_u8_u32(temp32x2x2[2].val[1]); mat[1].val[2] = vreinterpret_u8_u32(temp32x2x2[1].val[1]); mat[1].val[3] = vreinterpret_u8_u32(temp32x2x2[3].val[1]); vst1_u8(ptr_y +tr_x, mat[0].val[0]);vst1_u8(ptr_y+ dstep+tr_x, mat[0].val[1]);vst1_u8(ptr_y+2*dstep+tr_x, mat[0].val[2]);vst1_u8(ptr_y+3*dstep+tr_x, mat[0].val[3]);vst1_u8(ptr_y+4*dstep+tr_x, mat[1].val[0]);vst1_u8(ptr_y+5*dstep+tr_x, mat[1].val[1]);vst1_u8(ptr_y+6*dstep+tr_x, mat[1].val[2]);vst1_u8(ptr_y+7*dstep+tr_x, mat[1].val[3]);}}for(int y=sh;y<h;y++){int tr_x = h-y-1;for(int x=0;x<w;x++){uint8_t* ptr_y = dst+x*dstep;ptr_y[tr_x] = src[y*sstep+x];}}for(int x=sw;x<w;x++){ uint8_t* ptr_y = dst+x*dstep;for(int y=0;y<sh;y++){int tr_x = h-y-1;ptr_y[tr_x] = src[y*sstep+x];}}return; }彩圖（RGB三通道）右旋90
static void rotate_pos90_rgb(uint8_t* src, size_t sstep, uint8_t* dst, size_t dstep, int w,int h, int cn) {uint8x8x4_t mat10,mat11,mat12;uint8x8x4_t mat20,mat21,mat22;uint8x8x2_t temp10,temp11,temp12;uint8x8x2_t temp20,temp21,temp22;uint8x8x2_t temp30,temp31,temp32;uint8x8x2_t temp40,temp41,temp42;uint16x4x4_t temp110,temp111,temp112;uint16x4x4_t temp120,temp121,temp122;uint16x4x2_t temp50,temp51,temp52;uint16x4x2_t temp60,temp61,temp62;uint16x4x2_t temp70,temp71,temp72;uint16x4x2_t temp80,temp81,temp82;uint32x2x4_t temp210,temp211,temp212;uint32x2x4_t temp220,temp221,temp222;uint32x2x2_t res10,res11,res12;uint32x2x2_t res20,res21,res22;uint32x2x2_t res30,res31,res32;uint32x2x2_t res40,res41,res42;int dw=w&7; int sw=w-dw;int dh=h&7; int sh=h-dh;int x,y;for(y=0;y<sh;y=y+8){uint8_t* pIndexY = src+y*sstep;int tr_x = (h-y-8)*cn;for(x=0;x<sw;x=x+8){int xIndex = x * cn;uint8_t* ptr_y = dst+x*dstep;uint8x8x3_t v = vld3_u8(pIndexY + xIndex);mat10.val[0] = v.val[0]; mat11.val[0] = v.val[1]; mat12.val[0] = v.val[2];v = vld3_u8((pIndexY + sstep) + xIndex);mat10.val[1] = v.val[0]; mat11.val[1] = v.val[1]; mat12.val[1] = v.val[2];v = vld3_u8((pIndexY + 2*sstep) + xIndex);mat10.val[2] = v.val[0]; mat11.val[2] = v.val[1]; mat12.val[2] = v.val[2];v = vld3_u8((pIndexY + 3*sstep) + xIndex);mat10.val[3] = v.val[0]; mat11.val[3] = v.val[1]; mat12.val[3] = v.val[2];v = vld3_u8((pIndexY + 4*sstep) + xIndex);mat20.val[0] = v.val[0]; mat21.val[0] = v.val[1]; mat22.val[0] = v.val[2];v = vld3_u8((pIndexY + 5*sstep) + xIndex);mat20.val[1] = v.val[0]; mat21.val[1] = v.val[1]; mat22.val[1] = v.val[2];v = vld3_u8((pIndexY + 6*sstep) + xIndex);mat20.val[2] = v.val[0]; mat21.val[2] = v.val[1]; mat22.val[2] = v.val[2];v = vld3_u8((pIndexY + 7*sstep) + xIndex);mat20.val[3] = v.val[0]; mat21.val[3] = v.val[1]; mat22.val[3] = v.val[2];temp10=vtrn_u8(mat10.val[1],mat10.val[0]); temp11=vtrn_u8(mat11.val[1],mat11.val[0]); temp12=vtrn_u8(mat12.val[1],mat12.val[0]); temp20=vtrn_u8(mat10.val[3],mat10.val[2]); temp21=vtrn_u8(mat11.val[3],mat11.val[2]); temp22=vtrn_u8(mat12.val[3],mat12.val[2]);temp30=vtrn_u8(mat20.val[1],mat20.val[0]); temp31=vtrn_u8(mat21.val[1],mat21.val[0]); temp32=vtrn_u8(mat22.val[1],mat22.val[0]);temp40=vtrn_u8(mat20.val[3],mat20.val[2]); temp41=vtrn_u8(mat21.val[3],mat21.val[2]); temp42=vtrn_u8(mat22.val[3],mat22.val[2]);temp110.val[0]=vreinterpret_u16_u8(temp10.val[0]); temp111.val[0]=vreinterpret_u16_u8(temp11.val[0]); temp112.val[0]=vreinterpret_u16_u8(temp12.val[0]);temp110.val[1]=vreinterpret_u16_u8(temp10.val[1]); temp111.val[1]=vreinterpret_u16_u8(temp11.val[1]); temp112.val[1]=vreinterpret_u16_u8(temp12.val[1]);temp110.val[2]=vreinterpret_u16_u8(temp20.val[0]); temp111.val[2]=vreinterpret_u16_u8(temp21.val[0]); temp112.val[2]=vreinterpret_u16_u8(temp22.val[0]);temp110.val[3]=vreinterpret_u16_u8(temp20.val[1]); temp111.val[3]=vreinterpret_u16_u8(temp21.val[1]); temp112.val[3]=vreinterpret_u16_u8(temp22.val[1]);temp120.val[0]=vreinterpret_u16_u8(temp30.val[0]); temp121.val[0]=vreinterpret_u16_u8(temp31.val[0]); temp122.val[0]=vreinterpret_u16_u8(temp32.val[0]);temp120.val[1]=vreinterpret_u16_u8(temp30.val[1]); temp121.val[1]=vreinterpret_u16_u8(temp31.val[1]); temp122.val[1]=vreinterpret_u16_u8(temp32.val[1]);temp120.val[2]=vreinterpret_u16_u8(temp40.val[0]); temp121.val[2]=vreinterpret_u16_u8(temp41.val[0]); temp122.val[2]=vreinterpret_u16_u8(temp42.val[0]);temp120.val[3]=vreinterpret_u16_u8(temp40.val[1]); temp121.val[3]=vreinterpret_u16_u8(temp41.val[1]); temp122.val[3]=vreinterpret_u16_u8(temp42.val[1]);temp50=vtrn_u16(temp110.val[2],temp110.val[0]); temp51=vtrn_u16(temp111.val[2],temp111.val[0]); temp52=vtrn_u16(temp112.val[2],temp112.val[0]);temp60=vtrn_u16(temp110.val[3],temp110.val[1]); temp61=vtrn_u16(temp111.val[3],temp111.val[1]); temp62=vtrn_u16(temp112.val[3],temp112.val[1]);temp70=vtrn_u16(temp120.val[2],temp120.val[0]); temp71=vtrn_u16(temp121.val[2],temp121.val[0]); temp72=vtrn_u16(temp122.val[2],temp122.val[0]);temp80=vtrn_u16(temp120.val[3],temp120.val[1]); temp81=vtrn_u16(temp121.val[3],temp121.val[1]); temp82=vtrn_u16(temp122.val[3],temp122.val[1]);temp210.val[0]=vreinterpret_u32_u16(temp50.val[0]); temp211.val[0]=vreinterpret_u32_u16(temp51.val[0]); temp212.val[0]=vreinterpret_u32_u16(temp52.val[0]);temp210.val[1]=vreinterpret_u32_u16(temp50.val[1]); temp211.val[1]=vreinterpret_u32_u16(temp51.val[1]); temp212.val[1]=vreinterpret_u32_u16(temp52.val[1]);temp210.val[2]=vreinterpret_u32_u16(temp60.val[0]); temp211.val[2]=vreinterpret_u32_u16(temp61.val[0]); temp212.val[2]=vreinterpret_u32_u16(temp62.val[0]);temp210.val[3]=vreinterpret_u32_u16(temp60.val[1]); temp211.val[3]=vreinterpret_u32_u16(temp61.val[1]); temp212.val[3]=vreinterpret_u32_u16(temp62.val[1]);temp220.val[0]=vreinterpret_u32_u16(temp70.val[0]); temp221.val[0]=vreinterpret_u32_u16(temp71.val[0]); temp222.val[0]=vreinterpret_u32_u16(temp72.val[0]);temp220.val[1]=vreinterpret_u32_u16(temp70.val[1]); temp221.val[1]=vreinterpret_u32_u16(temp71.val[1]); temp222.val[1]=vreinterpret_u32_u16(temp72.val[1]);temp220.val[2]=vreinterpret_u32_u16(temp80.val[0]); temp221.val[2]=vreinterpret_u32_u16(temp81.val[0]); temp222.val[2]=vreinterpret_u32_u16(temp82.val[0]);temp220.val[3]=vreinterpret_u32_u16(temp80.val[1]); temp221.val[3]=vreinterpret_u32_u16(temp81.val[1]); temp222.val[3]=vreinterpret_u32_u16(temp82.val[1]);res10=vtrn_u32(temp220.val[0],temp210.val[0]); res11=vtrn_u32(temp221.val[0],temp211.val[0]); res12=vtrn_u32(temp222.val[0],temp212.val[0]);res20=vtrn_u32(temp220.val[1],temp210.val[1]); res21=vtrn_u32(temp221.val[1],temp211.val[1]); res22=vtrn_u32(temp222.val[1],temp212.val[1]);res30=vtrn_u32(temp220.val[2],temp210.val[2]); res31=vtrn_u32(temp221.val[2],temp211.val[2]); res32=vtrn_u32(temp222.val[2],temp212.val[2]);res40=vtrn_u32(temp220.val[3],temp210.val[3]); res41=vtrn_u32(temp221.val[3],temp211.val[3]); res42=vtrn_u32(temp222.val[3],temp212.val[3]);mat10.val[0]=vreinterpret_u8_u32(res10.val[0]); mat11.val[0]=vreinterpret_u8_u32(res11.val[0]); mat12.val[0]=vreinterpret_u8_u32(res12.val[0]);mat10.val[1]=vreinterpret_u8_u32(res30.val[0]); mat11.val[1]=vreinterpret_u8_u32(res31.val[0]); mat12.val[1]=vreinterpret_u8_u32(res32.val[0]);mat10.val[2]=vreinterpret_u8_u32(res20.val[0]); mat11.val[2]=vreinterpret_u8_u32(res21.val[0]); mat12.val[2]=vreinterpret_u8_u32(res22.val[0]);mat10.val[3]=vreinterpret_u8_u32(res40.val[0]); mat11.val[3]=vreinterpret_u8_u32(res41.val[0]); mat12.val[3]=vreinterpret_u8_u32(res42.val[0]);mat20.val[0]=vreinterpret_u8_u32(res10.val[1]); mat21.val[0]=vreinterpret_u8_u32(res11.val[1]); mat22.val[0]=vreinterpret_u8_u32(res12.val[1]);mat20.val[1]=vreinterpret_u8_u32(res30.val[1]); mat21.val[1]=vreinterpret_u8_u32(res31.val[1]); mat22.val[1]=vreinterpret_u8_u32(res32.val[1]);mat20.val[2]=vreinterpret_u8_u32(res20.val[1]); mat21.val[2]=vreinterpret_u8_u32(res21.val[1]); mat22.val[2]=vreinterpret_u8_u32(res22.val[1]);mat20.val[3]=vreinterpret_u8_u32(res40.val[1]); mat21.val[3]=vreinterpret_u8_u32(res41.val[1]); mat22.val[3]=vreinterpret_u8_u32(res42.val[1]);v_store_interleave_rotate(ptr_y +tr_x,mat10.val[0],mat11.val[0],mat12.val[0]); v_store_interleave_rotate(ptr_y+ dstep+tr_x,mat10.val[1],mat11.val[1],mat12.val[1]);v_store_interleave_rotate(ptr_y+2*dstep+tr_x,mat10.val[2],mat11.val[2],mat12.val[2]);v_store_interleave_rotate(ptr_y+3*dstep+tr_x,mat10.val[3],mat11.val[3],mat12.val[3]);v_store_interleave_rotate(ptr_y+4*dstep+tr_x,mat20.val[0],mat21.val[0],mat22.val[0]);v_store_interleave_rotate(ptr_y+5*dstep+tr_x,mat20.val[1],mat21.val[1],mat22.val[1]);v_store_interleave_rotate(ptr_y+6*dstep+tr_x,mat20.val[2],mat21.val[2],mat22.val[2]);v_store_interleave_rotate(ptr_y+7*dstep+tr_x,mat20.val[3],mat21.val[3],mat22.val[3]);}}for(y=sh;y<h;y++){int tr_x = (h-y-1)*cn;for(x=0;x<w;x++){uint8_t* ptr_y = dst+x*dstep;for(int n=0;n<cn;n++){ptr_y[tr_x+n]=src[y*sstep+x+n];}}}for(x=sw;x<w;x++){ uint8_t* ptr_y = dst+x*dstep;for(y=0;y<sh;y++){int tr_x = (h-y-1)*cn;for(int n=0;n<cn;n++){ptr_y[tr_x+n]=src[y*sstep+x+n];}}}return; }?
左旋90
4x4矩陣左旋實例
左旋90度相對右旋而言比較簡單了，不需要倒置行數(shù)據(jù)，其他的類似：
? ? ? ?旋轉(zhuǎn)前矩陣? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? 旋轉(zhuǎn)后矩陣
? ? ? ? ? ? ? ? ?? ? ?? ??
算法步驟：
1）直接每兩行做2x2轉(zhuǎn)置(相鄰行轉(zhuǎn)置)：
? ? ? ? ? 原矩陣? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? 相鄰行轉(zhuǎn)置? ? ? ? ? ? ? ? ? ? ? ? ??
? ? ? ? ? ? ? ?? ? ? ??
2）類似地，每行兩個像素值看做一個整體，隔行轉(zhuǎn)置，但是注意這里需要垂直flip一下，不要擔(dān)心會因此消耗時間，只需要在存儲的時候上下倒著水平存儲就可以了，并不耗時，?這樣便可以得到最終的矩陣左旋結(jié)果了：
? ? ? ?上一步結(jié)果? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?隔行轉(zhuǎn)置? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? 垂直倒置存儲
? ? ? ? ? ? ? ? ? ? ? ??? ? ? ? ? ? ? ? ? ? ?
灰度圖（單通道）左旋90
static void rotate_neg90_gray(uint8_t* src, size_t sstep, uint8_t* dst, size_t dstep, int w, int h) {uint8x8x4_t mat[2]; //use 2 register array to load a 8x8 patchuint8x8x2_t temp8x8x2[4];uint16x4x2_t temp16x4x2[8];uint32x2x2_t temp32x2x2[8];int dw=w&7; int sw=w-dw;int dh=h&7; int sh=h-dh;for(int i=0;i<sh;i+=8){for(int j=0;j<sw;j+=8){//step0 load 8x8 bytes in 8 registersmat[0].val[0] = vld1_u8(src+ i *sstep+j);mat[0].val[1] = vld1_u8(src+(i+1)*sstep+j);mat[0].val[2] = vld1_u8(src+(i+2)*sstep+j);mat[0].val[3] = vld1_u8(src+(i+3)*sstep+j);mat[1].val[0] = vld1_u8(src+(i+4)*sstep+j);mat[1].val[1] = vld1_u8(src+(i+5)*sstep+j);mat[1].val[2] = vld1_u8(src+(i+6)*sstep+j);mat[1].val[3] = vld1_u8(src+(i+7)*sstep+j);//step1 trn nearby registerstemp8x8x2[0] = vtrn_u8(mat[0].val[0], mat[0].val[1]);temp8x8x2[1] = vtrn_u8(mat[0].val[2], mat[0].val[3]);temp8x8x2[2] = vtrn_u8(mat[1].val[0], mat[1].val[1]);temp8x8x2[3] = vtrn_u8(mat[1].val[2], mat[1].val[3]);//step2 trn 1,3 2,4 5,7 6,8temp16x4x2[0].val[0] = vreinterpret_u16_u8(temp8x8x2[0].val[0]);temp16x4x2[0].val[1] = vreinterpret_u16_u8(temp8x8x2[0].val[1]);temp16x4x2[1].val[0] = vreinterpret_u16_u8(temp8x8x2[1].val[0]);temp16x4x2[1].val[1] = vreinterpret_u16_u8(temp8x8x2[1].val[1]);temp16x4x2[2].val[0] = vreinterpret_u16_u8(temp8x8x2[2].val[0]);temp16x4x2[2].val[1] = vreinterpret_u16_u8(temp8x8x2[2].val[1]);temp16x4x2[3].val[0] = vreinterpret_u16_u8(temp8x8x2[3].val[0]);temp16x4x2[3].val[1] = vreinterpret_u16_u8(temp8x8x2[3].val[1]);temp16x4x2[4] = vtrn_u16(temp16x4x2[0].val[0],temp16x4x2[1].val[0]);temp16x4x2[5] = vtrn_u16(temp16x4x2[0].val[1],temp16x4x2[1].val[1]);temp16x4x2[6] = vtrn_u16(temp16x4x2[2].val[0],temp16x4x2[3].val[0]);temp16x4x2[7] = vtrn_u16(temp16x4x2[2].val[1],temp16x4x2[3].val[1]);//step3 trn 1,5 2,6 3,7 4,8temp32x2x2[0].val[0] = vreinterpret_u32_u16(temp16x4x2[4].val[0]);temp32x2x2[0].val[1] = vreinterpret_u32_u16(temp16x4x2[4].val[1]);temp32x2x2[1].val[0] = vreinterpret_u32_u16(temp16x4x2[5].val[0]);temp32x2x2[1].val[1] = vreinterpret_u32_u16(temp16x4x2[5].val[1]);temp32x2x2[2].val[0] = vreinterpret_u32_u16(temp16x4x2[6].val[0]);temp32x2x2[2].val[1] = vreinterpret_u32_u16(temp16x4x2[6].val[1]);temp32x2x2[3].val[0] = vreinterpret_u32_u16(temp16x4x2[7].val[0]);temp32x2x2[3].val[1] = vreinterpret_u32_u16(temp16x4x2[7].val[1]);temp32x2x2[4] = vtrn_u32(temp32x2x2[0].val[0],temp32x2x2[2].val[0]);temp32x2x2[5] = vtrn_u32(temp32x2x2[0].val[1],temp32x2x2[2].val[1]);temp32x2x2[6] = vtrn_u32(temp32x2x2[1].val[0],temp32x2x2[3].val[0]);temp32x2x2[7] = vtrn_u32(temp32x2x2[1].val[1],temp32x2x2[3].val[1]);//step4 store bytes in correct position,the order now is 1,2,3,4,5,6,7,8temp8x8x2[0].val[0] = vreinterpret_u8_u32(temp32x2x2[7].val[1]);temp8x8x2[0].val[1] = vreinterpret_u8_u32(temp32x2x2[5].val[1]);temp8x8x2[1].val[0] = vreinterpret_u8_u32(temp32x2x2[6].val[1]);temp8x8x2[1].val[1] = vreinterpret_u8_u32(temp32x2x2[4].val[1]);temp8x8x2[2].val[0] = vreinterpret_u8_u32(temp32x2x2[7].val[0]);temp8x8x2[2].val[1] = vreinterpret_u8_u32(temp32x2x2[5].val[0]);temp8x8x2[3].val[0] = vreinterpret_u8_u32(temp32x2x2[6].val[0]);temp8x8x2[3].val[1] = vreinterpret_u8_u32(temp32x2x2[4].val[0]);vst1_u8(dst+(w-j-8)*dstep+i, temp8x8x2[0].val[0]);vst1_u8(dst+(w-j-7)*dstep+i, temp8x8x2[0].val[1]);vst1_u8(dst+(w-j-6)*dstep+i, temp8x8x2[1].val[0]);vst1_u8(dst+(w-j-5)*dstep+i, temp8x8x2[1].val[1]);vst1_u8(dst+(w-j-4)*dstep+i, temp8x8x2[2].val[0]);vst1_u8(dst+(w-j-3)*dstep+i, temp8x8x2[2].val[1]);vst1_u8(dst+(w-j-2)*dstep+i, temp8x8x2[3].val[0]);vst1_u8(dst+(w-j-1)*dstep+i, temp8x8x2[3].val[1]);}}for(int y=sh;y<h;y++){int yIndex = y*sstep;for(int x=0;x<w;x++){dst[(w-x-1)*h+y]=src[yIndex+x];}}for(int x=sw;x<w;x++){ int xIndex = (w-x-1)*h;for(int y=0;y<sh;y++){dst[xIndex+y]=src[y*sstep+x];}}return; }?
彩圖（RGB三通道）左旋90?
static void rotate_neg90_rgb(uint8_t* src, size_t sstep, uint8_t* dst, size_t dstep, int w,int h, int cn) {uint8x8x4_t mat0[2],mat1[2],mat2[2]; //use 2 register array to load a 8x8 patchuint8x8x2_t temp8x8x2_0[4],temp8x8x2_1[4],temp8x8x2_2[4];uint16x4x2_t temp16x4x2_0[8],temp16x4x2_1[8],temp16x4x2_2[8];uint32x2x2_t temp32x2x2_0[8],temp32x2x2_1[8],temp32x2x2_2[8];int dw=w&7; int sw=w-dw;int dh=h&7; int sh=h-dh;for(int i=0;i<sh;i+=8){for(int j=0;j<sw;j+=8){//step0 load 8x8 bytes in 8 registersuint8x8x3_t v = vld3_u8(src+ i *sstep+j*cn);mat0[0].val[0] = v.val[0]; mat1[0].val[0] = v.val[1]; mat2[0].val[0] = v.val[2];v = vld3_u8(src+(i+1)*sstep+j*cn);mat0[0].val[1] = v.val[0]; mat1[0].val[1] = v.val[1]; mat2[0].val[1] = v.val[2];v = vld3_u8(src+(i+2)*sstep+j*cn);mat0[0].val[2] = v.val[0]; mat1[0].val[2] = v.val[1]; mat2[0].val[2] = v.val[2];v = vld3_u8(src+(i+3)*sstep+j*cn);mat0[0].val[3] = v.val[0]; mat1[0].val[3] = v.val[1]; mat2[0].val[3] = v.val[2];v = vld3_u8(src+(i+4)*sstep+j*cn);mat0[1].val[0] = v.val[0]; mat1[1].val[0] = v.val[1]; mat2[1].val[0] = v.val[2];v = vld3_u8(src+(i+5)*sstep+j*cn);mat0[1].val[1] = v.val[0]; mat1[1].val[1] = v.val[1]; mat2[1].val[1] = v.val[2];v = vld3_u8(src+(i+6)*sstep+j*cn);mat0[1].val[2] = v.val[0]; mat1[1].val[2] = v.val[1]; mat2[1].val[2] = v.val[2];v = vld3_u8(src+(i+7)*sstep+j*cn);mat0[1].val[3] = v.val[0]; mat1[1].val[3] = v.val[1]; mat2[1].val[3] = v.val[2];//step1 trn nearby registerstemp8x8x2_0[0] = vtrn_u8(mat0[0].val[0], mat0[0].val[1]); temp8x8x2_1[0] = vtrn_u8(mat1[0].val[0], mat1[0].val[1]); temp8x8x2_2[0] = vtrn_u8(mat2[0].val[0], mat2[0].val[1]);temp8x8x2_0[1] = vtrn_u8(mat0[0].val[2], mat0[0].val[3]); temp8x8x2_1[1] = vtrn_u8(mat1[0].val[2], mat1[0].val[3]); temp8x8x2_2[1] = vtrn_u8(mat2[0].val[2], mat2[0].val[3]);temp8x8x2_0[2] = vtrn_u8(mat0[1].val[0], mat0[1].val[1]); temp8x8x2_1[2] = vtrn_u8(mat1[1].val[0], mat1[1].val[1]); temp8x8x2_2[2] = vtrn_u8(mat2[1].val[0], mat2[1].val[1]);temp8x8x2_0[3] = vtrn_u8(mat0[1].val[2], mat0[1].val[3]); temp8x8x2_1[3] = vtrn_u8(mat1[1].val[2], mat1[1].val[3]); temp8x8x2_2[3] = vtrn_u8(mat2[1].val[2], mat2[1].val[3]);//step2 trn 1,3 2,4 5,7 6,8temp16x4x2_0[0].val[0] = vreinterpret_u16_u8(temp8x8x2_0[0].val[0]); temp16x4x2_1[0].val[0] = vreinterpret_u16_u8(temp8x8x2_1[0].val[0]); temp16x4x2_2[0].val[0] = vreinterpret_u16_u8(temp8x8x2_2[0].val[0]);temp16x4x2_0[0].val[1] = vreinterpret_u16_u8(temp8x8x2_0[0].val[1]); temp16x4x2_1[0].val[1] = vreinterpret_u16_u8(temp8x8x2_1[0].val[1]); temp16x4x2_2[0].val[1] = vreinterpret_u16_u8(temp8x8x2_2[0].val[1]);temp16x4x2_0[1].val[0] = vreinterpret_u16_u8(temp8x8x2_0[1].val[0]); temp16x4x2_1[1].val[0] = vreinterpret_u16_u8(temp8x8x2_1[1].val[0]); temp16x4x2_2[1].val[0] = vreinterpret_u16_u8(temp8x8x2_2[1].val[0]);temp16x4x2_0[1].val[1] = vreinterpret_u16_u8(temp8x8x2_0[1].val[1]); temp16x4x2_1[1].val[1] = vreinterpret_u16_u8(temp8x8x2_1[1].val[1]); temp16x4x2_2[1].val[1] = vreinterpret_u16_u8(temp8x8x2_2[1].val[1]);temp16x4x2_0[2].val[0] = vreinterpret_u16_u8(temp8x8x2_0[2].val[0]); temp16x4x2_1[2].val[0] = vreinterpret_u16_u8(temp8x8x2_1[2].val[0]); temp16x4x2_2[2].val[0] = vreinterpret_u16_u8(temp8x8x2_2[2].val[0]);temp16x4x2_0[2].val[1] = vreinterpret_u16_u8(temp8x8x2_0[2].val[1]); temp16x4x2_1[2].val[1] = vreinterpret_u16_u8(temp8x8x2_1[2].val[1]); temp16x4x2_2[2].val[1] = vreinterpret_u16_u8(temp8x8x2_2[2].val[1]);temp16x4x2_0[3].val[0] = vreinterpret_u16_u8(temp8x8x2_0[3].val[0]); temp16x4x2_1[3].val[0] = vreinterpret_u16_u8(temp8x8x2_1[3].val[0]); temp16x4x2_2[3].val[0] = vreinterpret_u16_u8(temp8x8x2_2[3].val[0]);temp16x4x2_0[3].val[1] = vreinterpret_u16_u8(temp8x8x2_0[3].val[1]); temp16x4x2_1[3].val[1] = vreinterpret_u16_u8(temp8x8x2_1[3].val[1]); temp16x4x2_2[3].val[1] = vreinterpret_u16_u8(temp8x8x2_2[3].val[1]);temp16x4x2_0[4] = vtrn_u16(temp16x4x2_0[0].val[0],temp16x4x2_0[1].val[0]); temp16x4x2_1[4] = vtrn_u16(temp16x4x2_1[0].val[0],temp16x4x2_1[1].val[0]); temp16x4x2_2[4] = vtrn_u16(temp16x4x2_2[0].val[0],temp16x4x2_2[1].val[0]);temp16x4x2_0[5] = vtrn_u16(temp16x4x2_0[0].val[1],temp16x4x2_0[1].val[1]); temp16x4x2_1[5] = vtrn_u16(temp16x4x2_1[0].val[1],temp16x4x2_1[1].val[1]); temp16x4x2_2[5] = vtrn_u16(temp16x4x2_2[0].val[1],temp16x4x2_2[1].val[1]);temp16x4x2_0[6] = vtrn_u16(temp16x4x2_0[2].val[0],temp16x4x2_0[3].val[0]); temp16x4x2_1[6] = vtrn_u16(temp16x4x2_1[2].val[0],temp16x4x2_1[3].val[0]); temp16x4x2_2[6] = vtrn_u16(temp16x4x2_2[2].val[0],temp16x4x2_2[3].val[0]);temp16x4x2_0[7] = vtrn_u16(temp16x4x2_0[2].val[1],temp16x4x2_0[3].val[1]); temp16x4x2_1[7] = vtrn_u16(temp16x4x2_1[2].val[1],temp16x4x2_1[3].val[1]); temp16x4x2_2[7] = vtrn_u16(temp16x4x2_2[2].val[1],temp16x4x2_2[3].val[1]);//step3 trn 1,5 2,6 3,7 4,8temp32x2x2_0[0].val[0] = vreinterpret_u32_u16(temp16x4x2_0[4].val[0]); temp32x2x2_1[0].val[0] = vreinterpret_u32_u16(temp16x4x2_1[4].val[0]); temp32x2x2_2[0].val[0] = vreinterpret_u32_u16(temp16x4x2_2[4].val[0]);temp32x2x2_0[0].val[1] = vreinterpret_u32_u16(temp16x4x2_0[4].val[1]); temp32x2x2_1[0].val[1] = vreinterpret_u32_u16(temp16x4x2_1[4].val[1]); temp32x2x2_2[0].val[1] = vreinterpret_u32_u16(temp16x4x2_2[4].val[1]);temp32x2x2_0[1].val[0] = vreinterpret_u32_u16(temp16x4x2_0[5].val[0]); temp32x2x2_1[1].val[0] = vreinterpret_u32_u16(temp16x4x2_1[5].val[0]); temp32x2x2_2[1].val[0] = vreinterpret_u32_u16(temp16x4x2_2[5].val[0]);temp32x2x2_0[1].val[1] = vreinterpret_u32_u16(temp16x4x2_0[5].val[1]); temp32x2x2_1[1].val[1] = vreinterpret_u32_u16(temp16x4x2_1[5].val[1]); temp32x2x2_2[1].val[1] = vreinterpret_u32_u16(temp16x4x2_2[5].val[1]);temp32x2x2_0[2].val[0] = vreinterpret_u32_u16(temp16x4x2_0[6].val[0]); temp32x2x2_1[2].val[0] = vreinterpret_u32_u16(temp16x4x2_1[6].val[0]); temp32x2x2_2[2].val[0] = vreinterpret_u32_u16(temp16x4x2_2[6].val[0]);temp32x2x2_0[2].val[1] = vreinterpret_u32_u16(temp16x4x2_0[6].val[1]); temp32x2x2_1[2].val[1] = vreinterpret_u32_u16(temp16x4x2_1[6].val[1]); temp32x2x2_2[2].val[1] = vreinterpret_u32_u16(temp16x4x2_2[6].val[1]);temp32x2x2_0[3].val[0] = vreinterpret_u32_u16(temp16x4x2_0[7].val[0]); temp32x2x2_1[3].val[0] = vreinterpret_u32_u16(temp16x4x2_1[7].val[0]); temp32x2x2_2[3].val[0] = vreinterpret_u32_u16(temp16x4x2_2[7].val[0]);temp32x2x2_0[3].val[1] = vreinterpret_u32_u16(temp16x4x2_0[7].val[1]); temp32x2x2_1[3].val[1] = vreinterpret_u32_u16(temp16x4x2_1[7].val[1]); temp32x2x2_2[3].val[1] = vreinterpret_u32_u16(temp16x4x2_2[7].val[1]);temp32x2x2_0[4] = vtrn_u32(temp32x2x2_0[0].val[0],temp32x2x2_0[2].val[0]); temp32x2x2_1[4] = vtrn_u32(temp32x2x2_1[0].val[0],temp32x2x2_1[2].val[0]); temp32x2x2_2[4] = vtrn_u32(temp32x2x2_2[0].val[0],temp32x2x2_2[2].val[0]);temp32x2x2_0[5] = vtrn_u32(temp32x2x2_0[0].val[1],temp32x2x2_0[2].val[1]); temp32x2x2_1[5] = vtrn_u32(temp32x2x2_1[0].val[1],temp32x2x2_1[2].val[1]); temp32x2x2_2[5] = vtrn_u32(temp32x2x2_2[0].val[1],temp32x2x2_2[2].val[1]);temp32x2x2_0[6] = vtrn_u32(temp32x2x2_0[1].val[0],temp32x2x2_0[3].val[0]); temp32x2x2_1[6] = vtrn_u32(temp32x2x2_1[1].val[0],temp32x2x2_1[3].val[0]); temp32x2x2_2[6] = vtrn_u32(temp32x2x2_2[1].val[0],temp32x2x2_2[3].val[0]);temp32x2x2_0[7] = vtrn_u32(temp32x2x2_0[1].val[1],temp32x2x2_0[3].val[1]); temp32x2x2_1[7] = vtrn_u32(temp32x2x2_1[1].val[1],temp32x2x2_1[3].val[1]); temp32x2x2_2[7] = vtrn_u32(temp32x2x2_2[1].val[1],temp32x2x2_2[3].val[1]);//step4 store bytes in correct position,the order now is 1,2,3,4,5,6,7,8temp8x8x2_0[0].val[0] = vreinterpret_u8_u32(temp32x2x2_0[7].val[1]); temp8x8x2_1[0].val[0] = vreinterpret_u8_u32(temp32x2x2_1[7].val[1]); temp8x8x2_2[0].val[0] = vreinterpret_u8_u32(temp32x2x2_2[7].val[1]);temp8x8x2_0[0].val[1] = vreinterpret_u8_u32(temp32x2x2_0[5].val[1]); temp8x8x2_1[0].val[1] = vreinterpret_u8_u32(temp32x2x2_1[5].val[1]); temp8x8x2_2[0].val[1] = vreinterpret_u8_u32(temp32x2x2_2[5].val[1]);temp8x8x2_0[1].val[0] = vreinterpret_u8_u32(temp32x2x2_0[6].val[1]); temp8x8x2_1[1].val[0] = vreinterpret_u8_u32(temp32x2x2_1[6].val[1]); temp8x8x2_2[1].val[0] = vreinterpret_u8_u32(temp32x2x2_2[6].val[1]);temp8x8x2_0[1].val[1] = vreinterpret_u8_u32(temp32x2x2_0[4].val[1]); temp8x8x2_1[1].val[1] = vreinterpret_u8_u32(temp32x2x2_1[4].val[1]); temp8x8x2_2[1].val[1] = vreinterpret_u8_u32(temp32x2x2_2[4].val[1]);temp8x8x2_0[2].val[0] = vreinterpret_u8_u32(temp32x2x2_0[7].val[0]); temp8x8x2_1[2].val[0] = vreinterpret_u8_u32(temp32x2x2_1[7].val[0]); temp8x8x2_2[2].val[0] = vreinterpret_u8_u32(temp32x2x2_2[7].val[0]);temp8x8x2_0[2].val[1] = vreinterpret_u8_u32(temp32x2x2_0[5].val[0]); temp8x8x2_1[2].val[1] = vreinterpret_u8_u32(temp32x2x2_1[5].val[0]); temp8x8x2_2[2].val[1] = vreinterpret_u8_u32(temp32x2x2_2[5].val[0]);temp8x8x2_0[3].val[0] = vreinterpret_u8_u32(temp32x2x2_0[6].val[0]); temp8x8x2_1[3].val[0] = vreinterpret_u8_u32(temp32x2x2_1[6].val[0]); temp8x8x2_2[3].val[0] = vreinterpret_u8_u32(temp32x2x2_2[6].val[0]);temp8x8x2_0[3].val[1] = vreinterpret_u8_u32(temp32x2x2_0[4].val[0]); temp8x8x2_1[3].val[1] = vreinterpret_u8_u32(temp32x2x2_1[4].val[0]); temp8x8x2_2[3].val[1] = vreinterpret_u8_u32(temp32x2x2_2[4].val[0]);v_store_interleave_rotate(dst+(w-j-8)*dstep+i*cn, temp8x8x2_0[0].val[0], temp8x8x2_1[0].val[0], temp8x8x2_2[0].val[0]);v_store_interleave_rotate(dst+(w-j-7)*dstep+i*cn, temp8x8x2_0[0].val[1], temp8x8x2_1[0].val[1], temp8x8x2_2[0].val[1]);v_store_interleave_rotate(dst+(w-j-6)*dstep+i*cn, temp8x8x2_0[1].val[0], temp8x8x2_1[1].val[0], temp8x8x2_2[1].val[0]);v_store_interleave_rotate(dst+(w-j-5)*dstep+i*cn, temp8x8x2_0[1].val[1], temp8x8x2_1[1].val[1], temp8x8x2_2[1].val[1]);v_store_interleave_rotate(dst+(w-j-4)*dstep+i*cn, temp8x8x2_0[2].val[0], temp8x8x2_1[2].val[0], temp8x8x2_2[2].val[0]);v_store_interleave_rotate(dst+(w-j-3)*dstep+i*cn, temp8x8x2_0[2].val[1], temp8x8x2_1[2].val[1], temp8x8x2_2[2].val[1]);v_store_interleave_rotate(dst+(w-j-2)*dstep+i*cn, temp8x8x2_0[3].val[0], temp8x8x2_1[3].val[0], temp8x8x2_2[3].val[0]);v_store_interleave_rotate(dst+(w-j-1)*dstep+i*cn, temp8x8x2_0[3].val[1], temp8x8x2_1[3].val[1], temp8x8x2_2[3].val[1]);}}for(int y=sh;y<h;y++){int yIndex = y*sstep;for(int x=0;x<w;x++){for(int n=0;n<cn;n++){dst[(w-x-1)*h+cn*y+n]=src[yIndex+cn*x+n];}}}for(int x=sw;x<w;x++){ int xIndex = (w-x-1)*h;for(int y=0;y<sh;y++){for(int n=0;n<cn;n++){dst[xIndex+cn*y+n]=src[y*sstep+cn*x+n];}}}return; }v_store_interleave_rotate函數(shù)代碼：
inline void v_store_interleave_rotate(unsigned char* ptr, const uint8x8_t& a, const uint8x8_t& b, const uint8x8_t& c) {uint8x8x3_t v; v.val[0] = a; v.val[1] = b; v.val[2] = c; vst3_u8(ptr, v); }?
RK3288-linux(硬浮點)?耗時對比（opencv4.1與NEON加速）
總結(jié)
以上是生活随笔為你收集整理的基于NEON指令的图像旋转加速【armv7】的全部內(nèi)容，希望文章能夠幫你解決所遇到的問題。
- 上一篇： SqueezeNet网络模型详解
- 下一篇： |（与或移位等）的工程运用