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 6.4 ms
那么使用 Horizontal普通buffer,Vertical image buffer 组合方式的话，是不是时间最少？只是Intermediate image仍使用image对象，Horizontal kernel中的写操作需要改变。
结果： Horizontal 的最大local_work_size只能是32, Horizontal 增至8ms, Vertical 6.4ms
const sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST; #define r(xc,y) read_imagef( source, sampler, (int2) (xc, y) ).x #define w16(x,y,sum) write_imagef( dest, (int2) (x, y), sum.s0 );write_imagef( dest, (int2) (x+1, y), sum.s1 ); write_imagef( dest, (int2) (x+2, y), sum.s2 );write_imagef( dest, (int2) (x+3, y), sum.s3 ); write_imagef( dest, (int2) (x+4, y), sum.s4 );write_imagef( dest, (int2) (x+5, y), sum.s5 ); write_imagef( dest, (int2) (x+6, y), sum.s6 );write_imagef( dest, (int2) (x+7, y), sum.s7 ); write_imagef( dest, (int2) (x+8, y), sum.s8 );write_imagef( dest, (int2) (x+9, y), sum.s9 ); write_imagef( dest, (int2) (x+10, y), sum.sa );write_imagef( dest, (int2) (x+11, y), sum.sb ); write_imagef( dest, (int2) (x+12, y), sum.sc );write_imagef( dest, (int2) (x+13, y), sum.sd ); write_imagef( dest, (int2) (x+14, y), sum.se );write_imagef( dest, (int2) (x+15, y), sum.sf ); __kernel __attribute__((work_group_size_hint(32,1,1))) void ImageGaussianFilterHorizontal(__global const uchar* restrict source, // Source image __write_only image2d_t dest, // Intermediate dest image const int imgWidth , // Image width const int imgHeight) { const int y = get_global_id(0); if(y>=(imgHeight)) return; const uchar m_nRightShiftNum = 8; const uchar Rounding = (1 << (m_nRightShiftNum - 1)); const uchar m_nFilter[11] = {1,4,8,16,32,134,32,16,8,4,1}; const int s = 11; const int nStart = 5; const int nWidth = imgWidth; __global const uchar* pInLine = source + y*nWidth; int j; for(j = 0; j < nStart; j ++) { ushort sum = 0; for (int m = 0; m<s / 2; m++) { int k1 = (j + m - nStart); k1 = k1<0 ? -k1 : k1; int k2 = (j + nStart - m ); sum += (pInLine[k1] + pInLine[k2])*m_nFilter[m]; } sum += pInLine[j] * m_nFilter[s / 2]; //sum = (sum + Rounding) >> 8; write_imagef( dest, (int2) (j, y), convert_float(sum)/(255.0*256) ); } ushort16 line0 = convert_ushort16(vload16(0,pInLine+j-nStart)); for ( ; (j+16)<= (nWidth - nStart); j+=16) { ushort16 line1 = convert_ushort16(vload16(0,pInLine+j-nStart+16)); 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[0]; temp0.s0123456789abcdef = temp0.s123456789abcdeff; temp0.sf = line1.s0; temp1.s0123456789abcdef = temp1.s00123456789abcde; temp1.s0 = line0.s9; sum += ( temp0 + temp1 ) * m_nFilter[1]; temp0.s0123456789abcdef = temp0.s123456789abcdeff; temp0.sf = line1.s1; temp1.s0123456789abcdef = temp1.s00123456789abcde; temp1.s0 = line0.s8; sum += ( temp0 + temp1 ) * m_nFilter[2]; temp0.s0123456789abcdef = temp0.s123456789abcdeff; temp0.sf = line1.s2; temp1.s0123456789abcdef = temp1.s00123456789abcde; temp1.s0 = line0.s7; sum += ( temp0 + temp1 ) * m_nFilter[3]; temp0.s0123456789abcdef = temp0.s123456789abcdeff; temp0.sf = line1.s3; temp1.s0123456789abcdef = temp1.s00123456789abcde; temp1.s0 = line0.s6; sum += ( temp0 + temp1 ) * m_nFilter[4]; temp0.s0123456789abcdef = temp0.s123456789abcdeff; temp0.sf = line1.s4; sum += ( temp0 ) * m_nFilter[5]; line0 = line1; float16 sum2 = (convert_float16(sum))/(255.0*256); w16(j,y,sum2 ); } for( ; j < nWidth; j ++) { ushort sum = 0; for (int m = 0; m<s / 2; m++) { int k1 = (j + m - nStart); int k2 = (j + nStart - m ); k2 = k2 >= nWidth ? 2 * nWidth - 2 - k2 : k2; sum += (pInLine[k1] + pInLine[k2])*m_nFilter[m]; } sum += pInLine[j] * m_nFilter[s / 2]; //sum = (sum + Rounding) >> m_nRightShiftNum; write_imagef( dest, (int2) (j, y), convert_float(sum)/(255.0*256) ); } }
2.使用各种办法,最终也只能降到13.7ms,Horizontal 7.5, Vertical 6ms,最终代码如下.
更新：H和V都 去掉__attribute__ 属性,local_work_size都设置NULL,让opencl自己选择，H 的最大local_work_size又变回了64，总时间13ms.因为在LG G4,adreno 418上运行却需要40ms,在adreno 418上的local_work_size最大可以是1024，却被强制设成了32.
a.使用mad指令做sum乘加，结果有误差，时间也略增.fma 是无限精度,mad 是快速方法，结果是近似值。
b.使用 pInTemp 读fisrt 16 bytes,避免重复读取，有0.x ms的优势
c.边界使用了mirror repeat
const sampler_t sampler = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP_TO_EDGE | CLK_FILTER_NEAREST; #define r(xc,y) read_imagef( source, sampler, (int2) (xc, y) ).x #define w16(x,y,sum) write_imagef( dest, (int2) (x, y), sum.s0 );write_imagef( dest, (int2) (x+1, y), sum.s1 ); write_imagef( dest, (int2) (x+2, y), sum.s2 );write_imagef( dest, (int2) (x+3, y), sum.s3 ); write_imagef( dest, (int2) (x+4, y), sum.s4 );write_imagef( dest, (int2) (x+5, y), sum.s5 ); write_imagef( dest, (int2) (x+6, y), sum.s6 );write_imagef( dest, (int2) (x+7, y), sum.s7 ); write_imagef( dest, (int2) (x+8, y), sum.s8 );write_imagef( dest, (int2) (x+9, y), sum.s9 ); write_imagef( dest, (int2) (x+10, y), sum.sa );write_imagef( dest, (int2) (x+11, y), sum.sb ); write_imagef( dest, (int2) (x+12, y), sum.sc );write_imagef( dest, (int2) (x+13, y), sum.sd ); write_imagef( dest, (int2) (x+14, y), sum.se );write_imagef( dest, (int2) (x+15, y), sum.sf ); //line0 start from j-5,line1 from j-5+16 #define GaussianShift16 { temp0 = line0; temp1.s0123 = line0.sabcd; temp1.s45 = line0.sef; temp1.s67 = line1.s01; temp1.s89abcdef = line1.s23456789; sum = ( temp0 + temp1 ) * m_nFilter[0]; temp0.s0123456789abcdef = temp0.s123456789abcdeff; temp0.sf = line1.s0; temp1.s0123456789abcdef = temp1.s00123456789abcde; temp1.s0 = line0.s9; sum += ( temp0 + temp1 ) * m_nFilter[1]; temp0.s0123456789abcdef = temp0.s123456789abcdeff; temp0.sf = line1.s1; temp1.s0123456789abcdef = temp1.s00123456789abcde; temp1.s0 = line0.s8; sum += ( temp0 + temp1 ) * m_nFilter[2]; temp0.s0123456789abcdef = temp0.s123456789abcdeff; temp0.sf = line1.s2; temp1.s0123456789abcdef = temp1.s00123456789abcde; temp1.s0 = line0.s7; sum += ( temp0 + temp1 ) * m_nFilter[3]; temp0.s0123456789abcdef = temp0.s123456789abcdeff; temp0.sf = line1.s3; temp1.s0123456789abcdef = temp1.s00123456789abcde; temp1.s0 = line0.s6; sum += ( temp0 + temp1 ) * m_nFilter[4]; temp0.s0123456789abcdef = temp0.s123456789abcdeff; temp0.sf = line1.s4; sum += ( temp0 ) * m_nFilter[5];} __kernel __attribute__((work_group_size_hint(32,1,1))) void ImageGaussianFilterHorizontal(__global const uchar* restrict source, // Source image __write_only image2d_t dest, // Intermediate dest image const int imgWidth , // Image width const int imgHeight) { const int y = get_global_id(0); if(y>=(imgHeight)) return; const uchar m_nFilter[11] = {1,4,8,16,32,134,32,16,8,4,1}; const int s = 11; const int nStart = 5; __global const uchar* pInLine = source + y*imgWidth; int j; uchar pInTemp[16]; *( (uint4*)(pInTemp) ) = *((__global uint4*)(pInLine)) ;//first 16 bytes for(j = 0; j < nStart; j ++) { ushort sum = 0; for (int m = 0; m<s / 2; m++) { int k1 = (j + m - nStart); k1 = k1<0 ? -k1 : k1; int k2 = (j + nStart - m ); sum += (pInTemp[k1] + pInTemp[k2])*m_nFilter[m]; } sum += pInTemp[j] * m_nFilter[s / 2]; write_imagef( dest, (int2) (j, y), convert_float(sum)/(255.0*256) ); } ushort16 temp0; ushort16 temp1; ushort16 sum; ushort16 line0,line1; line0 = convert_ushort16(*((uchar16*)pInTemp)); for ( ; j< (imgWidth-16); j+=16) { line1 = convert_ushort16(vload16(0,pInLine+j-nStart+16));//convert_ushort16( as_uchar16(*((__global uint4*)(pInLine+j-nStart+16))) ) ; GaussianShift16 line0 = line1; float16 sum2 = (convert_float16(sum))/(255.0*256); w16(j,y,sum2 ); } { //last 16 pixel,some pixels may caculate again j = imgWidth-16; line0 = convert_ushort16(vload16(0,pInLine+j-nStart)); //mirror repeat read line1.s0123 = convert_ushort4( vload4(0,pInLine+imgWidth-nStart) ); line1.s4567 = (ushort4)( pInLine[imgWidth-1],line1.s3,line1.s21 ) ; line1.s89 = (ushort2)(line1.s0,line0.sf); GaussianShift16 float16 sum2 = (convert_float16(sum))/(255.0*256); w16(j,y,sum2 ); } } __kernel __attribute__((work_group_size_hint(64,1,1))) void ImageGaussianFilterVertical(__read_only image2d_t source, // Source image __write_only image2d_t dest, const int imgWidth , const int imgHeight) { const int x = get_global_id(0); if(x>=(imgWidth)) return; const float m_nFilter[11] = {1/256.0,4/256.0,8/256.0,16/256.0,32/256.0,134/256.0,32/256.0,16/256.0,8/256.0,4/256.0,1/256.0}; #define rv16(x,y) (float16)( r(x,y),r(x,y+1),r(x,y+2),r(x,y+3),r(x,y+4),r(x,y+5),r(x,y+6),r(x,y+7), r(x,y+8),r(x,y+9),r(x,y+10),r(x,y+11),r(x,y+12),r(x,y+13),r(x,y+14),r(x,y+15)) #define wv16(x,y,sum) write_imagef( dest, (int2) (x,y), sum.s0 );write_imagef( dest, (int2) (x,y+1), sum.s1 ); write_imagef( dest, (int2) (x,y+2), sum.s2 );write_imagef( dest, (int2) (x,y+3), sum.s3 ); write_imagef( dest, (int2) (x,y+4), sum.s4 );write_imagef( dest, (int2) (x,y+5), sum.s5 ); write_imagef( dest, (int2) (x,y+6), sum.s6 );write_imagef( dest, (int2) (x,y+7), sum.s7 ); write_imagef( dest, (int2) (x,y+8), sum.s8 );write_imagef( dest, (int2) (x,y+9), sum.s9 ); write_imagef( dest, (int2) (x,y+10), sum.sa );write_imagef( dest, (int2) (x,y+11), sum.sb ); write_imagef( dest, (int2) (x,y+12), sum.sc );write_imagef( dest, (int2) (x,y+13), sum.sd ); write_imagef( dest, (int2) (x,y+14), sum.se );write_imagef( dest, (int2) (x,y+15), sum.sf ); float16 temp0; float16 temp1; float16 sum; float16 line0,line1; line0 = rv16(x,-5); line0.s0123 = line0.sa987;//mirror repeat line0.s4 = line0.s6; int j; for(j=0;j<imgHeight-16;j+=16){ line1 = rv16(x,j-5+16); GaussianShift16 line0 = line1; wv16(x,j,sum ); } //last 16 pixel,some pixels may caculate again if imgHeight not 16 bytes align j = imgHeight-16; line0 = rv16(x,j-5); //mirror repeat read const int y = imgHeight-5; line1.s0123 = (float4)( r(x,y),r(x,y+1),r(x,y+2),r(x,y+3) ); line1.s4567 = (float4)( r(x,y+4),line1.s3,line1.s21 ); line1.s89 = (float2)(line1.s0,line0.sf); GaussianShift16 wv16(x,j,sum ); }
总结：1.local_work_size 对时间的影响比较大,有时使用NULL默认的就可以，有时需要一个个去试。
使用vector 类型,local memory,kernel代码结构 都会对 local_work_size 最大值有影响
2.profile中的wait time可能是读写memory还有其它的等待时间，rum time是ALU计算执行的时间。
3.避免对global memory的重复读写,预先缓存下来再用
4.image buffer的读写比普通buffer快，也没有按行按列读写的效率差异.尽量使用image buffer
5.read/write_imageui 并不比 read/write_imagef 快，一般就使用float
6.write 比read 要慢很多，内存未对齐也会慢些
7.使用vector 读写,计算 都会更快.image buffer虽然是单点读，组合成vector计算也更快.
8.half类型存在精度问题，会引入误差，在这里也不比float快
9.如果不确定local_work_size,就设置成NULL,让opencl自己选择。
不同的GPU 上local_work_size最大值不一样，比如这个kernel 在Adreno 330上最大64,在adreno 418上最大1024.