Attachment #408 for bug #717

This bugzilla service is closed. All entries have been migrated to https://gitlab.com/libeigen/eigen

View | Details | Raw Unified | Return to bug 717
Collapse All | Expand All




    if(PanelMode) count += (stride-offset-depth);
  }
}

// this version is optimized for row major matrices
template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode>
struct gemm_pack_rhs<Scalar, Index, nr, RowMajor, Conjugate, PanelMode>
{
  typedef typename packet_traits<Scalar>::type Packet;
  enum { PacketSize = packet_traits<Scalar>::size };
  EIGEN_DONT_INLINE void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride=0, Index offset=0);
};

template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode>
EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, nr, RowMajor, Conjugate, PanelMode>
  ::operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride, Index offset)
{

  Index packet_cols = (cols/nr) * nr;
  Index count = 0;
  for(Index j2=0; j2<packet_cols; j2+=nr)
  {
    // skip what we have before
    if(PanelMode) count += nr * offset;
    for(Index k=0; k<depth; k++)
    {
      if (nr == PacketSize) {
        Packet A = ploadu<Packet>(&rhs[k*rhsStride + j2]);
        pstoreu(blockB+count, cj.pconj(A));
        count += PacketSize;
      } else {
        const Scalar* b0 = &rhs[k*rhsStride + j2];
                  blockB[count+0] = cj(b0[0]);
                  blockB[count+1] = cj(b0[1]);
        if(nr==4) blockB[count+2] = cj(b0[2]);
        if(nr==4) blockB[count+3] = cj(b0[3]);
        count += nr;
      }
    }
    // skip what we have after
    if(PanelMode) count += nr * (stride-offset-depth);
  }
  // copy the remaining columns one at a time (nr==1)
  for(Index j2=packet_cols; j2<cols; ++j2)
  {
    if(PanelMode) count += offset;

Return to bug 717

Lines 1252-1267 EIGEN_DONT_INLINE void gemm_pack_rhs<Sca Link Here

(-)a/Eigen/src/Core/products/GeneralBlockPanelKernel.h (-6 / +13 lines)
1252	if(PanelMode) count += (stride-offset-depth);	1252	if(PanelMode) count += (stride-offset-depth);
1253	}	1253	}
1254	}	1254	}
1255		1255
1256	// this version is optimized for row major matrices	1256	// this version is optimized for row major matrices
1257	template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode>	1257	template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode>
1258	struct gemm_pack_rhs<Scalar, Index, nr, RowMajor, Conjugate, PanelMode>	1258	struct gemm_pack_rhs<Scalar, Index, nr, RowMajor, Conjugate, PanelMode>
1259	{	1259	{
		1260	typedef typename packet_traits<Scalar>::type Packet;
1260	enum { PacketSize = packet_traits<Scalar>::size };	1261	enum { PacketSize = packet_traits<Scalar>::size };
1261	EIGEN_DONT_INLINE void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride=0, Index offset=0);	1262	EIGEN_DONT_INLINE void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride=0, Index offset=0);
1262	};	1263	};
1263		1264
1264	template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode>	1265	template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode>
1265	EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, nr, RowMajor, Conjugate, PanelMode>	1266	EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, nr, RowMajor, Conjugate, PanelMode>
1266	::operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride, Index offset)	1267	::operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride, Index offset)
1267	{	1268	{
Lines 1271-1292 EIGEN_DONT_INLINE void gemm_pack_rhs<Sca Link Here
1271	Index packet_cols = (cols/nr) * nr;	1272	Index packet_cols = (cols/nr) * nr;
1272	Index count = 0;	1273	Index count = 0;
1273	for(Index j2=0; j2<packet_cols; j2+=nr)	1274	for(Index j2=0; j2<packet_cols; j2+=nr)
1274	{	1275	{
1275	// skip what we have before	1276	// skip what we have before
1276	if(PanelMode) count += nr * offset;	1277	if(PanelMode) count += nr * offset;
1277	for(Index k=0; k<depth; k++)	1278	for(Index k=0; k<depth; k++)
1278	{	1279	{
1279	const Scalar* b0 = &rhs[k*rhsStride + j2];	1280	if (nr == PacketSize) {
1280	blockB[count+0] = cj(b0[0]);	1281	Packet A = ploadu<Packet>(&rhs[k*rhsStride + j2]);
1281	blockB[count+1] = cj(b0[1]);	1282	pstoreu(blockB+count, cj.pconj(A));
1282	if(nr==4) blockB[count+2] = cj(b0[2]);	1283	count += PacketSize;
1283	if(nr==4) blockB[count+3] = cj(b0[3]);	1284	} else {
1284	count += nr;	1285	const Scalar* b0 = &rhs[k*rhsStride + j2];
		1286	blockB[count+0] = cj(b0[0]);
		1287	blockB[count+1] = cj(b0[1]);
		1288	if(nr==4) blockB[count+2] = cj(b0[2]);
		1289	if(nr==4) blockB[count+3] = cj(b0[3]);
		1290	count += nr;
		1291	}
1285	}	1292	}
1286	// skip what we have after	1293	// skip what we have after
1287	if(PanelMode) count += nr * (stride-offset-depth);	1294	if(PanelMode) count += nr * (stride-offset-depth);
1288	}	1295	}
1289	// copy the remaining columns one at a time (nr==1)	1296	// copy the remaining columns one at a time (nr==1)
1290	for(Index j2=packet_cols; j2<cols; ++j2)	1297	for(Index j2=packet_cols; j2<cols; ++j2)
1291	{	1298	{
1292	if(PanelMode) count += offset;	1299	if(PanelMode) count += offset;