mirror of
https://gitlab.com/libeigen/eigen.git
synced 2025-01-24 14:45:14 +08:00
1024a70e82
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
The patch works by altering the gebp lhs packing routines to also
consider ½ and ¼ packet lenght rows when packing, besides the original
whole package and row-by-row attempts. Finally, gebp itself will try
to fit a fraction of a packet at a time if:
i) ½ and/or ¼ packets are available for the current context (e.g. AVX2
and SSE-sized SIMD register for x86)
ii) The matrix's height is favorable to it (it may be it's too small
in that dimension to take full advantage of the current/maximum
packet width or it may be the case that last rows may take
advantage of smaller packets before gebp goes row-by-row)
This helps mitigate huge slowdowns one had on AVX512 builds when
compared to AVX2 ones, for some dimensions. Gains top at an extra 1x
in throughput. This patch is a complement to changeset
|
||
|---|---|---|
| .. | ||
| src | ||
| Cholesky | ||
| CholmodSupport | ||
| CMakeLists.txt | ||
| Core | ||
| Dense | ||
| Eigen | ||
| Eigenvalues | ||
| Geometry | ||
| Householder | ||
| IterativeLinearSolvers | ||
| Jacobi | ||
| KLUSupport | ||
| LU | ||
| MetisSupport | ||
| OrderingMethods | ||
| PardisoSupport | ||
| PaStiXSupport | ||
| QR | ||
| QtAlignedMalloc | ||
| Sparse | ||
| SparseCholesky | ||
| SparseCore | ||
| SparseLU | ||
| SparseQR | ||
| SPQRSupport | ||
| StdDeque | ||
| StdList | ||
| StdVector | ||
| SuperLUSupport | ||
| SVD | ||
| UmfPackSupport | ||