From b55b5c7280a0481f01fe5ec764d55c443a8b6496 Mon Sep 17 00:00:00 2001
From: Sameer Agarwal <sameeragarwal@google.com>
Date: Fri, 1 Feb 2019 15:23:53 -0800
Subject: [PATCH] Speed up row-major matrix-vector product on ARM

The row-major matrix-vector multiplication code uses a threshold to
check if processing 8 rows at a time would thrash the cache.

This change introduces two modifications to this logic.

1. A smaller threshold for ARM and ARM64 devices.

The value of this threshold was determined empirically using a Pixel2
phone, by benchmarking a large number of matrix-vector products in the
range [1..4096]x[1..4096] and measuring performance separately on
small and little cores with frequency pinning.

On big (out-of-order) cores, this change has little to no impact. But
on the small (in-order) cores, the matrix-vector products are up to
700% faster. Especially on large matrices.

The motivation for this change was some internal code at Google which
was using hand-written NEON for implementing similar functionality,
processing the matrix one row at a time, which exhibited substantially
better performance than Eigen.

With the current change, Eigen handily beats that code.

2. Make the logic for choosing number of simultaneous rows apply
unifiormly to 8, 4 and 2 rows instead of just 8 rows.

Since the default threshold for non-ARM devices is essentially
unchanged (32000 -> 32 * 1024), this change has no impact on non-ARM
performance. This was verified by running the same set of benchmarks
on a Xeon desktop.
---
 Eigen/src/Core/products/GeneralMatrixVector.h | 25 +++++++++++++++----
 1 file changed, 20 insertions(+), 5 deletions(-)

diff --git a/Eigen/src/Core/products/GeneralMatrixVector.h b/Eigen/src/Core/products/GeneralMatrixVector.h
index 767feb99d..e7dc25478 100644
--- a/Eigen/src/Core/products/GeneralMatrixVector.h
+++ b/Eigen/src/Core/products/GeneralMatrixVector.h
@@ -255,11 +255,20 @@ EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE void general_matrix_vector_product<Index,Lhs
   conj_helper<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> cj;
   conj_helper<LhsPacket,RhsPacket,ConjugateLhs,ConjugateRhs> pcj;
 
-  // TODO: fine tune the following heuristic. The rationale is that if the matrix is very large,
-  //       processing 8 rows at once might be counter productive wrt cache.
-  const Index n8 = lhs.stride()*sizeof(LhsScalar)>32000 ? 0 : rows-7;
-  const Index n4 = rows-3;
-  const Index n2 = rows-1;
+  // TODO: fine tune the following heuristic. The rationale is that if the
+  // matrix is very large, processing multiple rows at once might be counter
+  // productive wrt cache.
+#if EIGEN_ARCH_ARM_OR_ARM64
+  // This threshold was empirically determined using a Pixel2.
+  // The little cores are a lot more sensitive to this number
+  // than the big cores.
+  const Index cache_threshold = 1024;
+#else
+  const Index cache_threshold = 1024 * 256;
+#endif
+
+  const Index row_bytes = lhs.stride() * sizeof(LhsScalar);
+  const Index n8 = (8 * row_bytes > cache_threshold) ? 0 : (rows - 7);
 
   // TODO: for padded aligned inputs, we could enable aligned reads
   enum { LhsAlignment = Unaligned };
@@ -320,6 +329,9 @@ EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE void general_matrix_vector_product<Index,Lhs
     res[(i+6)*resIncr] += alpha*cc6;
     res[(i+7)*resIncr] += alpha*cc7;
   }
+
+  if (i == rows) return;
+  const Index n4 = (4 * row_bytes > cache_threshold) ? 0 : (rows - 3);
   for(; i<n4; i+=4)
   {
     ResPacket c0 = pset1<ResPacket>(ResScalar(0)),
@@ -355,6 +367,9 @@ EIGEN_DEVICE_FUNC EIGEN_DONT_INLINE void general_matrix_vector_product<Index,Lhs
     res[(i+2)*resIncr] += alpha*cc2;
     res[(i+3)*resIncr] += alpha*cc3;
   }
+
+  if (i == rows) return;
+  const Index n2 = (2 * row_bytes > cache_threshold) ? 0 : (rows - 1);
   for(; i<n2; i+=2)
   {
     ResPacket c0 = pset1<ResPacket>(ResScalar(0)),