斌@NVIDIA&USTC2014年3matrix2MatrixMultiply:CPUvoidMatrixMulOnHost(float*M,float*N,float*P,int{for(inti=0;i<width;for(intj=0;j<width;{floatsum=for(intk=0;k<width;++{floata=M[i*width+k];floatb=N[k*width+j];sum+=a*b;}P[i*width+j]=}}3CodeMatrixMultiply:CUDA4CodeMatrixMultiply:CUDA5CodeMatrixMultiply:CUDA6CodeMatrixMultiply:CUDA7Code矩陣長度限僅用一個G80GT200–最多512個線程8PAGE9ImagePAGE9ImageImageImage例矩陣TILE_WIDTH=Block尺寸例矩陣TILE_WIDTH=Block尺寸:

Pd0,1Pd1,1Pd2,1Pd3,1Pd0,2Pd1,2Pd2,2Pd3,2ImagefromCodeCodeglobalvoidfloat*Md,float*Nd,float*Pd,int{intRow=blockIdx.y*blockDim.y+intCol=blockIdx.x*blockDim.x+floatPvalue=for(intk=0;k<Width;Pvalue+=Md[Row*Width+k]*Nd[k*Width+Pd[Row*Width+Col]=}globalvoidfloat*Md,float*Nd,float*Pd,intintintRow=blockIdx.y*blockDim.y+intCol=blockIdx.x*blockDim.x+floatPvalue=for(intk=0;k<Width;Pvalue+=Md[Row*Width+k]*Nd[k*Width+Pd[Row*Width+Col]=}globalvoidfloat*Md,float*Nd,float*Pd,int{intRow=blockIdx.y*blockDim.y+intCol=blockIdx.xintCol=blockIdx.x*blockDim.x+for(intk=0;k<Width;Pvalue+=Md[Row*Width+k]*Nd[k*Width+Pd[Row*Width+Col]=}globalvoidfloat*Md,float*Nd,float*Pd,int{intRow=blockIdx.y*+intCol=blockIdx.x*+for(intk=for(intk=0;k<Width;Pvalue+=Md[Row*Width+k]*Nd[k*Width+Pd[Row*Width+Col]=}Code調(diào)用dim3dimGrid(Width/TILE_WIDTH,Height/dim3dimBlock(TILE_WIDTH,Md,Nd,Pd,global讀寫怎么辦 N Image

102102Image012TILE_WIDTH-TILE_WIDTHTILE_WIDTH把kernel拆分成多個階MdNd的子集累 00212by

2

TILE_WIDTH

012TILE_WIDTH-TILE_WIDTHTILE_WIDTH每個線 MdNd的一個元素存入sharedmemory

012by 2

2

TILE_WIDTH

globalvoidfloat*Md,float*Nd,float*Pd,intCodeCode{sharedfloatsharedfloat

bx=blockIdx.x;inttx=threadIdx.x;

by=blockIdx.y;ty=threadIdx.y;Row=*+Col=*+floatPvalue=for(intm=0;m<Width/TILE_WIDTH;++m){Mds[ty][tx]=Md[Row*Width+(m*TILE_WIDTH+tx)];Nds[ty][tx]=Nd[Col+(m*TILE_WIDTH+for(intk=0;k<TILE_WIDTH;Pvalue+=Mds[ty][k]*}Pd[Row*Width+Col]=} globalglobalvoidfloat*Md,float*Nd,float*Pd,int{for(intfor(intk=0;k<TILE_WIDTH;Pvalue+=Mds[ty][k]*}Pd[Row*Width+Col]=}Code__shared____shared__float__shared__floatbxbx=blockIdx.x;inttx=threadIdx.x;by=blockty=Row=*+Col=*+SharedmemoryMdSharedmemoryMdNdfor(intm=0;m<Width/TILE_WIDTH;++m){Mds[ty][tx]=Md[Row*Width+(m*TILE_WIDTH+tx)];Nds[ty][tx]=Nd[Col+(m*TILE_WIDTH+globalglobalvoidfloat*Md,float*Nd,float*Pd,int{for(for(intk=0;k<TILE_WIDTH;Pvalue+=Mds[ty][k]*}Pd[Row*Width+Col]=}CodesharedfloatintRow=byintRow=by*TILE_WIDTH+intCol=bx*TILE_WIDTH+floatPvalue=mintbx=blockIdx.x;intby=inttx=threadIdx.x;intty=for(intm=0;m<for(intm=0;m<Width/TILE_WIDTH;++m)Nds[ty][tx]=Nd[Col+(m*TILE_WIDTH+globalglobalvoidfloat*Md,float*Nd,float*Pd,intfor(for(intk=0;k<TILE_WIDTH;Pvalue+=Mds[ty][k]*}Pd[Row*Width+Col]=}Code{sharedfloat

bx=blockIdx.x;inttx=threadIdx.x;

by=blockIdx.y;ty=threadIdx.y;Mds[ty][tx]Mds[ty][tx]=Md[Row*Width+(m*TILE_WIDTH+Nds[ty][tx]=Nd[Col+(m*TILEWIDTH+intintRow=by*TILE_WIDTH+intCol=bx*TILE_WIDTH+floatPvalue=for(intm=0;m<Width/TILE_WIDTH;++m)globalglobalvoidfloat*Md,float*Nd,float*Pd,int}Pd[Row*Width+Col]=}Pd[Row*Width+Col]=}Code{sharedfloatsharedfloat

bx=blockIdx.x;inttx=threadIdx.x;

by=blockIdx.y;ty=threadIdx.y;Row=*+Col=*+floatPvalue=for(intm=0;m<Width/TILE_WIDTH;++m){Mds[ty][tx]=Md[Row*Width+(m*TILE_WIDTH+tx)];Nds[ty][tx]=Nd[Col+(m*TILE_WIDTH+等待block內(nèi)所有線程,即,for(intk=0;等待block內(nèi)所有線程,即,Pvalue+=Mds[ty][k]*{sharedfloatsharedfloat

bx=blockIdx.x;inttx=threadIdx.x;

by=blockIdx.y;ty=threadIdx.y;Row=*+Col=*+floatPvalue=for(intm=0;m<Width/TILE_WIDTH;++m){Mds[ty][tx]=Md[Row*Width+(m*TILE_WIDTH+tx)];Nds[ty][tx]=Nd[Col+(m*TILE_WIDTH+for(intfor(intk=0;k<TILE_WIDTH;Pvalue+=Mds[ty][k]*globalglobalvoidfloat*Md,float*Nd,float*Pd,int}Pd[Row*Width+Col]=}Pd[Row*Width+Col]=}Code{sharedfloatsharedfloat

bx=blockIdx.x;inttx=threadIdx.x;

by=blockIdx.y;ty=threadIdx.y;Row=*+Col=*+floatPvalue=for(intm=0;m<Width/TILE_WIDTH;++m){Mds[ty][tx]=Md[Row*Width+(m*TILE_WIDTH+tx)];Nds[ty][tx]=Nd[Col+(m*TILE_WIDTH+for(intk=0;k<TILE_WIDTH;Pvalue+=Mds[ty][k]*globalglobalvoidfloat*Md,float*Nd,float*Pd,int}}Code{sharedfloatsharedfloat

bx=blockIdx.x;inttx=threadIdx.x;

by=blockIdx.y;ty=threadIdx.y;Row=*+Col=*+floatPvalue=for(intm=0;m<Width/TILE_WIDTH;++m){Mds[ty][tx]=Md[Row*Width+(m*TILE_WIDTH+tx)];Nds[ty][tx]=Nd[Col+(m*TILE_WIDTH+for(intk=0;k<TILE_WIDTH;Pvalue+=Mds[ty][k]*globalglobalPd[Row*Width+Col]=Matrix如何選取TILE_WIDTH如果太大的話會怎樣如何選取TILE_WIDTH如果太大的話會怎樣超出一個塊允許的最大線程G80andGT200–Kerpler–如何選取TILE_WIDTH的數(shù)G80andGT200–G80:16KBSM8blocks2KB1KB給Nds，1KB給Mds(16*16*TILE_WIDTH=更大的TILE_WIDTH將導致更少的塊Sharedmemory瓦片化的好