Adding lowlevel APIs for optimized RHS packet load in TensorFlow

 SpatialConvolution

Low-level APIs are added in order to optimized packet load in gemm_pack_rhs
in TensorFlow SpatialConvolution. The optimization is for scenario when a
packet is split across 2 adjacent columns. In this case we read it as two
'partial' packets and then merge these into 1. Currently this only works for
Packet16f (AVX512) and Packet8f (AVX2). We plan to add this for other
packet types (such as Packet8d) also.

This optimization shows significant speedup in SpatialConvolution with
certain parameters. Some examples are below.

Benchmark parameters are specified as:
Batch size, Input dim, Depth, Num of filters, Filter dim

Speedup numbers are specified for number of threads 1, 2, 4, 8, 16.

AVX512:

Parameters                  | Speedup (Num of threads: 1, 2, 4, 8, 16)
----------------------------|------------------------------------------
128,   24x24,  3, 64,   5x5 |2.18X, 2.13X, 1.73X, 1.64X, 1.66X
128,   24x24,  1, 64,   8x8 |2.00X, 1.98X, 1.93X, 1.91X, 1.91X
 32,   24x24,  3, 64,   5x5 |2.26X, 2.14X, 2.17X, 2.22X, 2.33X
128,   24x24,  3, 64,   3x3 |1.51X, 1.45X, 1.45X, 1.67X, 1.57X
 32,   14x14, 24, 64,   5x5 |1.21X, 1.19X, 1.16X, 1.70X, 1.17X
128, 128x128,  3, 96, 11x11 |2.17X, 2.18X, 2.19X, 2.20X, 2.18X

AVX2:

Parameters                  | Speedup (Num of threads: 1, 2, 4, 8, 16)
----------------------------|------------------------------------------
128,   24x24,  3, 64,   5x5 | 1.66X, 1.65X, 1.61X, 1.56X, 1.49X
 32,   24x24,  3, 64,   5x5 | 1.71X, 1.63X, 1.77X, 1.58X, 1.68X
128,   24x24,  1, 64,   5x5 | 1.44X, 1.40X, 1.38X, 1.37X, 1.33X
128,   24x24,  3, 64,   3x3 | 1.68X, 1.63X, 1.58X, 1.56X, 1.62X
128, 128x128,  3, 96, 11x11 | 1.36X, 1.36X, 1.37X, 1.37X, 1.37X

In the higher level benchmark cifar10, we observe a runtime improvement
of around 6% for AVX512 on Intel Skylake server (8 cores).

On lower level PackRhs micro-benchmarks specified in TensorFlow
tensorflow/core/kernels/eigen_spatial_convolutions_test.cc, we observe
the following runtime numbers:

AVX512:

Parameters                                                     | Runtime without patch (ns) | Runtime with patch (ns) | Speedup
---------------------------------------------------------------|----------------------------|-------------------------|---------
BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56)  |  41350                     | 15073                   | 2.74X
BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56)  |   7277                     |  7341                   | 0.99X
BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56)  |   8675                     |  8681                   | 1.00X
BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56)  |  24155                     | 16079                   | 1.50X
BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56)  |  25052                     | 17152                   | 1.46X
BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) |  18269                     | 18345                   | 1.00X
BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) |  19468                     | 19872                   | 0.98X
BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432)   | 156060                     | 42432                   | 3.68X
BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432)   | 132701                     | 36944                   | 3.59X

AVX2:

Parameters                                                     | Runtime without patch (ns) | Runtime with patch (ns) | Speedup
---------------------------------------------------------------|----------------------------|-------------------------|---------
BM_RHS_NAME(PackRhs, 128, 24, 24, 3, 64, 5, 5, 1, 1, 256, 56)  | 26233                      | 12393                   | 2.12X
BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 1, 1, 256, 56)  |  6091                      |  6062                   | 1.00X
BM_RHS_NAME(PackRhs, 32, 64, 64, 32, 64, 5, 5, 2, 2, 256, 56)  |  7427                      |  7408                   | 1.00X
BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 1, 1, 256, 56)  | 23453                      | 20826                   | 1.13X
BM_RHS_NAME(PackRhs, 32, 64, 64, 30, 64, 5, 5, 2, 2, 256, 56)  | 23167                      | 22091                   | 1.09X
BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 1, 1, 256, 56) | 23422                      | 23682                   | 0.99X
BM_RHS_NAME(PackRhs, 32, 256, 256, 4, 16, 8, 8, 2, 4, 256, 56) | 23165                      | 23663                   | 0.98X
BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 1, 1, 36, 432)   | 72689                      | 44969                   | 1.62X
BM_RHS_NAME(PackRhs, 32, 64, 64, 4, 16, 3, 3, 2, 2, 36, 432)   | 61732                      | 39779                   | 1.55X

All benchmarks on Intel Skylake server with 8 cores.

1 files changed, 24 insertions, 0 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
index 4ca6b3d8c..05c684ff9 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
@@ -95,6 +95,18 @@ struct TensorEvaluator
     return internal::ploadt<PacketReturnType, LoadMode>(m_data + index);
   }
 
+  // Return a packet starting at `index` where `umask` specifies which elements
+  // have to be loaded. Type/size of mask depends on PacketReturnType, e.g. for
+  // Packet16f, `umask` is of type uint16_t and if a bit is 1, corresponding
+  // float element will be loaded, otherwise 0 will be loaded.
+  // Function has been templatized to enable Sfinae.
+  template <typename PacketReturnTypeT = PacketReturnType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  typename internal::enable_if<internal::unpacket_traits<PacketReturnTypeT>::masked_load_available, PacketReturnTypeT>::type
+  partialPacket(Index index, typename internal::unpacket_traits<PacketReturnTypeT>::mask_t umask) const
+  {
+    return internal::ploadu<PacketReturnTypeT>(m_data + index, umask);
+  }
+
   template <int StoreMode> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
   void writePacket(Index index, const PacketReturnType& x)
   {
@@ -244,6 +256,18 @@ struct TensorEvaluator<const Derived, Device>
     return internal::ploadt_ro<PacketReturnType, LoadMode>(m_data + index);
   }
 
+  // Return a packet starting at `index` where `umask` specifies which elements
+  // have to be loaded. Type/size of mask depends on PacketReturnType, e.g. for
+  // Packet16f, `umask` is of type uint16_t and if a bit is 1, corresponding
+  // float element will be loaded, otherwise 0 will be loaded.
+  // Function has been templatized to enable Sfinae.
+  template <typename PacketReturnTypeT = PacketReturnType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  typename internal::enable_if<internal::unpacket_traits<PacketReturnTypeT>::masked_load_available, PacketReturnTypeT>::type
+  partialPacket(Index index, typename internal::unpacket_traits<PacketReturnTypeT>::mask_t umask) const
+  {
+    return internal::ploadu<PacketReturnTypeT>(m_data + index, umask);
+  }
+
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(const array<DenseIndex, NumCoords>& coords) const {
     eigen_assert(m_data);
     const Index index = (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? m_dims.IndexOfColMajor(coords)