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.
Prior to this change, a product with a LHS having 8 rows was faster with AVX-only than with AVX+FMA.
With AVX+FMA I measured a speed up of about x1.25 in such cases.
This changeset also includes:
* add HouseholderSequence::conjugateIf
* define int as the StorageIndex type for all dense solvers
* dedicated unit tests, including assertion checking
* _check_solve_assertion(): this method can be implemented in derived solver classes to implement custom checks
* CompleteOrthogonalDecompositions: add applyZOnTheLeftInPlace, fix scalar type in applyZAdjointOnTheLeftInPlace(), add missing assertions
* Cholesky: add missing assertions
* FullPivHouseholderQR: Corrected Scalar type in _solve_impl()
* BDCSVD: Unambiguous return type for ternary operator
* SVDBase: Corrected Scalar type in _solve_impl()
The isometric transform, like the affine transform, has an implicit last
row of [0, 0, 0, 1]. This was not being properly initialized, as verified
by a new test function.
This makes both the small and huge argument cases faster because:
- for small inputs this removes the last pselect
- for large inputs only the reduction part follows a scalar path,
the rest use the same SIMD path as the small-argument case.
The reinterpret_casts used in ptranspose(PacketBlock<Packet8cf,4>&)
ptranspose(PacketBlock<Packet8cf,8>&) don't appear to be working
correctly. They're used to convert the kernel parameters to
PacketBlock<Packet8d,T>& so that the complex number versions of
ptranspose can be written using the existing double implementations.
Unfortunately, they don't seem to work and are responsible for 9 unit
test failures in the AVX512 build of tensorflow master. This commit
fixes the issue by manually initialising PacketBlock<Packet8d,T>
variables with the contents of the kernel parameter before calling
the double version of ptranspose, and then copying the resulting
values back into the kernel parameter before returning.
- no FMA: 1ULP up to 3pi, 2ULP up to sin(25966) and cos(18838), fallback to std::sin/cos for larger inputs
- FMA: 1ULP up to sin(117435.992) and cos(71476.0625), fallback to std::sin/cos for larger inputs
Commit c53eececb0
introduced AVX512 support for complex numbers but required
avx512dq to build. Commit 1d683ae2f5
fixed some but not, it would seem all,
of the hard avx512dq dependencies. Build failures are still evident on
Eigen and TensorFlow when compiling with just avx512f and no avx512dq
using gcc 7.3. Looking at the code there does indeed seem to be a problem.
Commit c53eececb0
calls avx512dq intrinsics directly, e.g, _mm512_extractf32x8_ps
and _mm512_and_ps. This commit fixes the issue by replacing the direct
intrinsic calls with the various wrapper functions that are safe to use on
avx512f only builds.
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 4ad359237a
.
Since packing is changed, Eigen users which would go for very
low-level API usage, like TensorFlow, will have to be adapted to work
fine with the changes.
This is a preparation to a change on gebp_traits, where a new template
argument will be introduced to dictate the packet size, so it won't be
bound to the current/max packet size only anymore.
By having packet types defined early on gebp_traits, one has now to
act on packet types, not scalars anymore, for the enum values defined
on that class. One approach for reaching the vectorizable/size
properties one needs there could be getting the packet's scalar again
with unpacket_traits<>, then the size/Vectorizable enum entries from
packet_traits<>. It turns out guards like "#ifndef
EIGEN_VECTORIZE_AVX512" at AVX/PacketMath.h will hide smaller packet
variations of packet_traits<> for some types (and it makes sense to
keep that). In other words, one can't go back to the scalar and create
a new PacketType, as this will always lead to the maximum packet type
for the architecture.
The less costly/invasive solution for that, thus, is to add the
vectorizable info on every unpacket_traits struct as well.
It is based on the SSE version which is much more accurate, though very slightly slower.
This changeset also includes the following required changes:
- add packet-float to packet-int type traits
- add packet float<->int reinterpret casts
- add faster pselect for AVX based on blendv
1. Eigen/src/Core/arch/GPU/Half.h
Updating the HIPCC implementation half so that it can declared as a __shared__ variable
2. Eigen/src/Core/util/Macros.h, Eigen/src/Core/util/Memory.h
introducing a EIGEN_USE_STD(func) macro that calls
- std::func be default
- ::func when eigen is being compiled with HIPCC
This change was requested in the previous HIP PR
(https://bitbucket.org/eigen/eigen/pull-requests/518/pr-with-hip-specific-fixes-for-the-eigen/diff)
3. unsupported/Eigen/CXX11/src/Tensor/TensorDeviceThreadPool.h
Removing EIGEN_DEVICE_FUNC attribute from pure virtual methods as it is not supported by HIPCC
4. unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
Disabling the template specializations of InnerMostDimReducer as they run into HIPCC link errors
Commit aa110e681b
optimised the multiplication of small dyanmically
sized matrices by restricting the packet size to a maximum of 4, increasing
the chances that SIMD instructions are used in the computation. However, it
introduced a mismatch between the packet size and the requestedAlignment. This
mismatch can lead to crashes when the destination is not aligned. This patch
fixes the issue by ensuring that the AssignmentTraits are correctly computed
when using a restricted packet size.
* * *
Bind LinearPacketType to MaxPacketSize
This commit applies any packet size limit specified when instantiating
copy_using_evaluator_traits to the LinearPacketType, providing that the
size of the destination is not known at compile time.
* * *
Add unit test for restricted packet assignment
A new unit test is added to check that multiplication of small dynamically
sized matrices works correctly when the packet size is restricted to 4 and
the destination is unaligned.
* Support compiling without IO streams
Add the preprocessor definition EIGEN_NO_IO which, if defined,
disables all use of the IO streams part of the standard library.
INFO: From Compiling tensorflow/core/kernels/maxpooling_op_gpu.cu.cc:
/b/f/w/run/external/eigen_archive/Eigen/src/Core/arch/GPU/Half.h(197): error: calling a __host__ function("std::equal_to<float> ::operator () const") from a __global__ function("tensorflow::_NV_ANON_NAMESPACE::MaxPoolGradBackwardNoMaskNHWC< ::Eigen::half> ") is not allowed
/b/f/w/run/external/eigen_archive/Eigen/src/Core/arch/GPU/Half.h(197): error: identifier "std::equal_to<float> ::operator () const" is undefined in device code"
/b/f/w/run/external/eigen_archive/Eigen/src/Core/arch/GPU/Half.h(197): error: calling a __host__ function("std::equal_to<float> ::operator () const") from a __global__ function("tensorflow::_NV_ANON_NAMESPACE::MaxPoolGradBackwardNoMaskNCHW< ::Eigen::half> ") is not allowed
/b/f/w/run/external/eigen_archive/Eigen/src/Core/arch/GPU/Half.h(197): error: identifier "std::equal_to<float> ::operator () const" is undefined in device code
4 errors detected in the compilation of "/tmp/tmpxft_00000011_00000000-6_maxpooling_op_gpu.cu.cpp1.ii".
ERROR: /tmpfs/tensor_flow/tensorflow/core/kernels/BUILD:3753:1: output 'tensorflow/core/kernels/_objs/pooling_ops_gpu/maxpooling_op_gpu.cu.pic.o' was not created
ERROR: /tmpfs/tensor_flow/tensorflow/core/kernels/BUILD:3753:1: Couldn't build file tensorflow/core/kernels/_objs/pooling_ops_gpu/maxpooling_op_gpu.cu.pic.o: not all outputs were created or valid
The Packet16f, Packet8f and Packet8d types are too large to use with dynamically
sized matrices typically processed by the SliceVectorizedTraversal specialization of
the dense_assignment_loop. Using these types is likely to lead to little or no
vectorization. Significant slowdown in the multiplication of these small matrices can
be observed when building with AVX and AVX512 enabled.
This patch introduces a new dense_assignment_kernel that is used when
computing small products whose operands have dynamic dimensions. It ensures that the
PacketSize used is no larger than 4, thereby increasing the chance that vectorized
instructions will be used when computing the product.
I tested all 969 possible combinations of M, K, and N that are handled by the
dense_assignment_loop on x86 builds. Although a few combinations are slowed down
by this patch they are far outnumbered by the cases that are sped up, as the
following results demonstrate.
Disabling Packed8d on AVX512 builds:
Total Cases: 969
Better: 511
Worse: 85
Same: 373
Max Improvement: 169.00% (4 8 6)
Max Degradation: 36.50% (8 5 3)
Median Improvement: 35.46%
Median Degradation: 17.41%
Total FLOPs Improvement: 19.42%
Disabling Packet16f and Packed8f on AVX512 builds:
Total Cases: 969
Better: 658
Worse: 5
Same: 306
Max Improvement: 214.05% (8 6 5)
Max Degradation: 22.26% (16 2 1)
Median Improvement: 60.05%
Median Degradation: 13.32%
Total FLOPs Improvement: 59.58%
Disabling Packed8f on AVX builds:
Total Cases: 969
Better: 663
Worse: 96
Same: 210
Max Improvement: 155.29% (4 10 5)
Max Degradation: 35.12% (8 3 2)
Median Improvement: 34.28%
Median Degradation: 15.05%
Total FLOPs Improvement: 26.02%
