双三次插值及优化Word文档下载推荐.docx
《双三次插值及优化Word文档下载推荐.docx》由会员分享，可在线阅读，更多相关《双三次插值及优化Word文档下载推荐.docx（45页珍藏版）》请在冰点文库上搜索。
3.进行矩阵运算，得到插值结果
iTemp1=Su0*P1+Su1*P5+Su2*P9+Su3*P13
iTemp2=Su0*P2+Su1*P6+Su2*P10+Su3*P14
iTemp3=Su0*P3+Su1*P7+Su2*P11+Su3*P15
iTemp4=Su0*P4+Su1*P8+Su2*P12+Su3*P16
iResult=Sv1*iTemp1+Sv2*iTemp2+Sv3*iTemp3+Sv4*iTemp4
4.在得到插值结果图后，我们发现图像中有“毛刺”，因此对插值结果做了个后处理，即：
设该点在原图中的像素值为pSrc，若abs（iResult-pSrc）大于某阈值，我们认为插值后的点可能污染原图，因此用原像素值pSrc代替。
3.算法优化
由于双三次插值计算一个点的坐标需要其周围16个点，更有多达20次的乘法及15次的加法，计算量可以说是非常大，势必要进行优化。
我们选择了Intel的SSE2优化技术，它只支持在P4及以上的机器。
测试当前CPU是否支持SSE2，可由CPUID指令得到，代码为：
BOOLg_bSSE2=FALSE;
__asm
{
moveax,1;
cpuid;
testedx,0x04000000;
jzNotSupport;
movg_bSSE2,1
NotSupport:
}
支持SSE2的CPU引入了8个128位的寄存器，这样一个寄存器中就可以存放4个点（RGB），有利于并行计算。
详细代码见Transform.cpp中函数Optimize_Bicubic。
优化中遇到的问题：
1.图像每个点由RGB通道组成，由于1个SSE2寄存器有16个字节，这样读入4个像素点后，要浪费4个字节，同时要花费时间将数据对齐，即由BRGB|RGBR|GBRG|BRGB对齐成0RGB|0RGB|0RGB|0RGB;
2.读16字节数据到寄存器时，由于图像地址不能保证是16字节对齐，因此需用更多时钟周期的MOVDQU指令（6个以上时钟周期）；
如能使地址16字节对齐，则可用MOVDQA指令（1个时钟周期）;
3.为了消除除法及浮点运算，对权值放大256倍，这样在计算插值核时，必须用2Bytes来表示1个系数，而图像数据都是1Byte，这样在对齐做乘法时，要浪费一半的SSE2寄存器的空间，导致运算时间变长；
而若降低插值核的精度，使其在1Byte表示围时，运算的精度又大为下降；
4.对各指令的周期以及若干行指令是否能够并行流水缺乏经验和认识。
附：
SSE2指令整理
算术（Arithmetic）指令：
ADDPD--PackedDouble-PrecisionFloating-PointAddSSE2
2个double对应相加
ADDPDxmm0,xmm1/m128
ADDPS--PackedSingle-PrecisionFloating-PointAddSSE
4个float对应相加
ADDPSxmm0,xmm1/m128
ADDSD--ScalarDouble-PrecisionFloating-PointAdd
1个double（低端）对应相加SSE2
ADDSDxmm0,xmm1/m64
ADDSS--ScalarSingle-PrecisionFloating-PointAddSSE
1个float（低端）对应相加
ADDSSxmm0,xmm1/m32
PADDB/PADDW/PADDD--PackedAdd
Opcode
Instruction
Description
0FFC/r
PADDBmm,mm/m64
Addpackedbyteintegersfrommm/m64andmm.
660FFC/r
PADDBxmm1,xmm2/m128
Addpackedbyteintegersfromxmm2/m128andxmm1.
0FFD/r
PADDWmm,mm/m64
Addpackedwordintegersfrommm/m64andmm.
660FFD/r
PADDWxmm1,xmm2/m128
Addpackedwordintegersfromxmm2/m128andxmm1.
0FFE/r
PADDDmm,mm/m64
Addpackeddoublewordintegersfrommm/m64andmm.
660FFE/r
PADDDxmm1,xmm2/m128
Addpackeddoublewordintegersfromxmm2/m128andxmm1.
PADDQ--PackedQuadwordAdd
0FD4/r
PADDQmm1,mm2/m64
Addquadwordintegermm2/m64tomm1
660FD4/r
PADDQxmm1,xmm2/m128
Addpackedquadwordintegersxmm2/m128toxmm1
PADDSB/PADDSW--PackedAddwithSaturation
0FEC/r
PADDSBmm,mm/m64
Addpackedsignedbyteintegersfrommm/m64andmmandsaturatetheresults.
660FEC/r
PADDSBxmm1,
xmm2/m128
Addpackedsignedbyteintegersfromxmm2/m128andxmm1saturatetheresults.
0FED/r
PADDSWmm,mm/m64
Addpackedsignedwordintegersfrommm/m64andmmandsaturatetheresults.
660FED/r
PADDSWxmm1,xmm2/m128
Addpackedsignedwordintegersfromxmm2/m128andxmm1andsaturatetheresults.
PADDUSB/PADDUSW--PackedAddUnsignedwithSaturation
0FDC/r
PADDUSBmm,mm/m64
Addpackedunsignedbyteintegersfrommm/m64andmmandsaturatetheresults.
660FDC/r
PADDUSBxmm1,xmm2/m128
Addpackedunsignedbyteintegersfromxmm2/m128andxmm1saturatetheresults.
0FDD/r
PADDUSWmm,mm/m64
Addpackedunsignedwordintegersfrommm/m64andmmandsaturatetheresults.
660FDD/r
PADDUSWxmm1,xmm2/m128
Addpackedunsignedwordintegersfromxmm2/m128toxmm1andsaturatetheresults.
PMADDWD--PackedMultiplyandAdd
0FF5/r
PMADDWDmm,mm/m64
Multiplythepackedwordsinmmbythepackedwordsinmm/m64.Addthe32-bitpairsofresultsandstoreinmmasdoubleword
660FF5/r
PMADDWDxmm1,xmm2/m128
Multiplythepackedwordintegersinxmm1bythepackedwordintegersinxmm2/m128,andaddtheadjacentdoublewordresults.
PSADBW--PackedSumofAbsoluteDifferences
0FF6/r
PSADBWmm1,mm2/m64
Absolutedifferenceofpackedunsignedbyteintegersfrommm2/m64andmm1;
differencesarethensummedtoproduceanunsignedwordintegerresult.
660FF6/r
PSADBWxmm1,xmm2/m128
Absolutedifferenceofpackedunsignedbyteintegersfromxmm2/m128andxmm1;
the8lowdifferencesand8highdifferencesarethensummedseparatelytoproducetwowordintegerresults.
PSUBB/PSUBW/PSUBD--PackedSubtract
0FF8/r
PSUBBmm,mm/m64
Subtractpackedbyteintegersinmm/m64frompackedbyteintegersinmm.
660FF8/r
PSUBBxmm1,xmm2/m128
Subtractpackedbyteintegersinxmm2/m128frompackedbyteintegersinxmm1.
0FF9/r
PSUBWmm,mm/m64
Subtractpackedwordintegersinmm/m64frompackedwordintegersinmm.
660FF9/r
PSUBWxmm1,xmm2/m128
Subtractpackedwordintegersinxmm2/m128frompackedwordintegersinxmm1.
0FFA/r
PSUBDmm,mm/m64
Subtractpackeddoublewordintegersinmm/m64frompackeddoublewordintegersinmm.
660FFA/r
PSUBDxmm1,xmm2/m128
Subtractpackeddoublewordintegersinxmm2/mem128frompackeddoublewordintegersinxmm1.
PSUBQ--PackedSubtractQuadword
0FFB/r
PSUBQmm1,mm2/m64
Subtractquadwordintegerinmm1frommm2/m64.
660FFB/r
PSUBQxmm1,xmm2/m128
Subtractpackedquadwordintegersinxmm1fromxmm2/m128.
PSUBSB/PSUBSW--PackedSubtractwithSaturation
0FE8/r
PSUBSBmm,mm/m64
Subtractsignedpackedbytesinmm/m64fromsignedpackedbytesinmmandsaturateresults.
660FE8/r
PSUBSBxmm1,xmm2/m128
Subtractpackedsignedbyteintegersinxmm2/m128frompackedsignedbyteintegersinxmm1andsaturateresults.
0FE9/r
PSUBSWmm,mm/m64
Subtractsignedpackedwordsinmm/m64fromsignedpackedwordsinmmandsaturateresults.
660FE9/r
PSUBSWxmm1,xmm2/m128
Subtractpackedsignedwordintegersinxmm2/m128frompackedsignedwordintegersinxmm1andsaturateresults.
PSUBUSB/PSUBUSW--PackedSubtractUnsignedwithSaturation
0FD8/r
PSUBUSBmm,mm/m64
Subtractunsignedpackedbytesinmm/m64fromunsignedpackedbytesinmmandsaturateresult.
660FD8/r
PSUBUSBxmm1,xmm2/m128
Subtractpackedunsignedbyteintegersinxmm2/m128frompackedunsignedbyteintegersinxmm1andsaturateresult.
0FD9/r
PSUBUSWmm,mm/m64
Subtractunsignedpackedwordsinmm/m64fromunsignedpackedwordsinmmandsaturateresult.
660FD9/r
PSUBUSWxmm1,xmm2/m128
Subtractpackedunsignedwordintegersinxmm2/m128frompackedunsignedwordintegersinxmm1andsaturateresult.
SUBPD--PackedDouble-PrecisionFloating-PointSubtract
660F5C/r
SUBPDxmm1,xmm2/m128
Subtractpackeddouble-precisionfloating-pointvaluesinxmm2/m128fromxmm1.
SUBPS--PackedSingle-PrecisionFloating-PointSubtract
0F5C/r
SUBPSxmm1xmm2/m128
Subtractpackedsingle-precisionfloating-pointvaluesinxmm2/memfromxmm1.
SUBSD--ScalarDouble-PrecisionFloating-PointSubtract
F20F5C/r
SUBSDxmm1,xmm2/m64
Subtractsthelowdouble-precisionfloating-pointnumbersinxmm2/mem64fromxmm1.
SUBSS--ScalarSingle-FPSubtract
F30F5C/r
SUBSSxmm1,xmm2/m32
Subtractthelowersingle-precisionfloating-pointnumbersinxmm2/m32fromxmm1.
------------------------------------------------------------------------------------------------------
PMULHUW--PackedMultiplyHighUnsigned
0FE4/r
PMULHUWmm1,mm2/m64
Multiplythepackedunsignedwordintegersinmm1registerandmm2/m64,andstorethehigh16bitsoftheresultsinmm1.
660FE4/r
PMULHUWxmm1,xmm2/m128
Multiplythepackedunsignedwordintegersinxmm1andxmm2/m128,andstorethehigh16bitsoftheresultsinxmm1.
PMULHW--PackedMultiplyHighSigned
0FE5/r
PMULHWmm,mm/m64
Multiplythepackedsignedwordintegersinmm1registerandmm2/m64,andstorethehigh16bitsoftheresultsinmm1.
660FE5/r
PMULHWxmm1,xmm2/m128
Multiplythepackedsignedwordintegersinxmm1andxmm2/m128,andstorethehigh16bitsoftheresultsinxmm1.
PMULLW--PackedMultiplyLowSigned
0FD5/r
PMULLWmm,mm/m64
Multiplythepackedsignedwordintegersinmm1registerandmm2/m64,andstorethelow16bitsoftheresultsinmm1.
660FD5/r
PMULLWxmm1,xmm2/m128
Multiplythepackedsignedwordintegersinxmm1andxmm2/m128,andstorethelow16bitsoftheresultsinxmm1.
PMULUDQ--MultiplyDoublewordUnsigned
0FF4/r
PMULUDQmm1,mm2/m64
Multiplyunsigneddoublewordintegerinmm1byunsigneddoublewordintegerinmm2/m64,andstorethequadwordresultinmm1.
66OFF4/r
PMULUDQxmm1,xmm2/m128
Multiplypackedunsigneddoublewordintegersinxmm1bypackedunsigneddoublewordintegersinxmm2/m128,andstorethequadwordresultsin
升级会员 