mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2024-12-21 07:19:46 +08:00
Code Issues Packages Projects Releases Wiki Activity

Vectorized the packing of a col-major matrix used as the right hand side argument in a matrix-matrix product when AVX instructions are used. No vectorization takes place when SSE instructions are used, however this doesn't seem to impact performance.

Browse Source
This commit is contained in:
Benoit Steiner 2014-03-27 10:38:41 -07:00
parent b776458ccb
commit 3e1fe8e416
1 changed files with 17 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@ -1033,6 +1033,7 @@ EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, nr, ColMajor, Conjugate, Pan
conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj; conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
Index packet_cols = (cols/nr) * nr; Index packet_cols = (cols/nr) * nr;
Index count = 0; Index count = 0;
const Index peeled_k = (depth/PacketSize)*PacketSize;
for(Index j2=0; j2<packet_cols; j2+=nr) for(Index j2=0; j2<packet_cols; j2+=nr)
{ {
// skip what we have before // skip what we have before
@ -1045,7 +1046,22 @@ EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, nr, ColMajor, Conjugate, Pan
const Scalar* b5 = &rhs[(j2+5)*rhsStride]; const Scalar* b5 = &rhs[(j2+5)*rhsStride];
const Scalar* b6 = &rhs[(j2+6)*rhsStride]; const Scalar* b6 = &rhs[(j2+6)*rhsStride];
const Scalar* b7 = &rhs[(j2+7)*rhsStride]; const Scalar* b7 = &rhs[(j2+7)*rhsStride];
for(Index k=0; k<depth; k++) Index k=0;
if(nr == PacketSize)
{
for(; k<peeled_k; k+=PacketSize) {
Kernel<Packet> kernel;
for (int p = 0; p < PacketSize; ++p) {
kernel.packet[p] = ploadu<Packet>(&rhs[(j2+p)*rhsStride+k]);
}
ptranspose(kernel);
for (int p = 0; p < PacketSize; ++p) {
pstoreu(blockB+count, cj.pconj(kernel.packet[p]));
count+=PacketSize;
}
}
}
for(; k<depth; k++)
{ {
blockB[count+0] = cj(b0[k]); blockB[count+0] = cj(b0[k]);
blockB[count+1] = cj(b1[k]); blockB[count+1] = cj(b1[k]);