GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-04-30 07:44:10 +08:00
Code Issues Packages Projects Releases Wiki Activity

Artificially increase l1-blocking size for AVX512. +10% speedup with current kernels.

Browse Source 
With a 6pX4 kernel (not committed yet), this provides a +20% speedup.
This commit is contained in:
Gael Guennebaud 2018-12-11 15:36:27 +01:00
parent 0a7e7af6fd
commit f159cf3d75
1 changed files with 11 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@ -101,6 +101,16 @@ void evaluateProductBlockingSizesHeuristic(Index& k, Index& m, Index& n, Index n
// at the register level. This small horizontal panel has to stay within L1 cache. // at the register level. This small horizontal panel has to stay within L1 cache.
std::ptrdiff_t l1, l2, l3; std::ptrdiff_t l1, l2, l3;
manage_caching_sizes(GetAction, &l1, &l2, &l3); manage_caching_sizes(GetAction, &l1, &l2, &l3);
#ifdef EIGEN_VECTORIZE_AVX512
// We need to find a rationale for that, but without this adjustment,
// performance with AVX512 is pretty bad, like -20% slower.
// One reason is that with increasing packet-size, the blocking size k
// has to become pretty small if we want that 1 lhs panel fit within L1.
// For instance, with the 3pX4 kernel and double, the size of the lhs+rhs panels are:
// k*(3*64 + 4*8) Bytes, with l1=32kBytes, and k%8=0, we have k=144.
// This is quite small for a good reuse of the accumulation registers.
l1 *= 4;
#endif
if (num_threads > 1) { if (num_threads > 1) {
typedef typename Traits::ResScalar ResScalar; typedef typename Traits::ResScalar ResScalar;
@ -372,7 +382,7 @@ public:
default_mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*LhsPacketSize, default_mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*LhsPacketSize,
#if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) \ #if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX) \
&& ((!EIGEN_COMP_MSVC) || (EIGEN_COMP_MSVC>=1914)) && ((!EIGEN_COMP_MSVC) || (EIGEN_COMP_MSVC>=1914))
// we assume 16 registers // we assume 16 registers or more
// See bug 992, if the scalar type is not vectorizable but that EIGEN_HAS_SINGLE_INSTRUCTION_MADD is defined, // See bug 992, if the scalar type is not vectorizable but that EIGEN_HAS_SINGLE_INSTRUCTION_MADD is defined,
// then using 3*LhsPacketSize triggers non-implemented paths in syrk. // then using 3*LhsPacketSize triggers non-implemented paths in syrk.
// Bug 1515: MSVC prior to v19.14 yields to register spilling. // Bug 1515: MSVC prior to v19.14 yields to register spilling.