This patch modifies the TensorContraction class to ensure that the kc_ field is
always a multiple of the packet_size, if the packet_size is > 8. Without this
change spatial convolutions in Tensorflow do not work properly as the code that
re-arranges the input matrices can assert if kc_ is not a multiple of the
packet_size. This leads to a unit test failure,
//tensorflow/python/kernel_tests:conv_ops_test, on AVX512 builds of tensorflow.
"... always use {NonBlocking}ThreadPool". It seems the non-blocking
implementation was me the default/only one, but a reference to the old
name was left unmodified. Fix that.
block
Builds configured without the -DEIGEN_TEST_CXX11=ON flag would fail
right away without this, as this test seems to rely on those language
features. The skip under compilation with MSVC was kept.
Without explicit conversion Tensorflow fails to compile, pset1 template deduction fails.
cannot convert '((const Eigen::internal::MeanReducer<Eigen::half>*)this)
->Eigen::internal::MeanReducer<Eigen::half>::packetCount_'
(type 'const DenseIndex {aka const long int}')
to type 'const type& {aka const Eigen::half&}'
return pdiv(vaccum, pset1<Packet>(packetCount_));
Honestly I’m not sure why it works in Eigen tests, because Eigen::half constructor is explicit, and why it stopped working in TF, I didn’t find any relevant changes since previous Eigen upgrade.
static_cast<T>(packetCount_) - breaks cxx11_tensor_reductions test for Eigen::half, also quite surprising.
When supplied, this allocator will be used in place of
internal::aligned_malloc. This permits e.g. use of a NUMA-node specific
allocator where the thread-pool is also restricted a single NUMA-node.
The type used in Eigen::DSizes needs to be at least 8 bytes long. Internally Tensor tries to convert this to an __int64 on Windows and this fails to build. On Linux, long and long long are both 8 byte integer types.
* * *
Changing from "long long" to "std::int64_t".
The major changes are
1. Moving CUDA/PacketMath.h to GPU/PacketMath.h
2. Moving CUDA/MathFunctions.h to GPU/MathFunction.h
3. Moving CUDA/CudaSpecialFunctions.h to GPU/GpuSpecialFunctions.h
The above three changes effectively enable the Eigen "Packet" layer for the HIP platform
4. Merging the "hip_basic" and "cuda_basic" unit tests into one ("gpu_basic")
5. Updating the "EIGEN_DEVICE_FUNC" marking in some places
The change has been tested on the HIP and CUDA platforms.
