TensorFlow Eigen - AshokBhat/ml GitHub Wiki

Description

TensorFlow Eigen backend
Supports FP32.
No support for INT8 as of TF v2.3

FP32 support

Uses Tensor Contraction kernel in eigen_contraction_kernel.h
Used as a fallback on x86 when OneDNN is used.
Default on other platforms.

Convolution

Convolution operations calling Eigen contraction kernel directly (MatMulConvFunctor) or indirectly (SpatialConvolution)

Benchmarking

TF has Eigen CPU kernels benchmark
Command to run bazel run -c opt //tensorflow/core/kernels:eigen_benchmark_cpu_test --test_output=all --cache_test_results=no -- --benchmarks=all
File : https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/eigen_benchmark_cpu_test.cc

INT8 support

No fallback tensor contraction kernel
A comment in eigen_contraction_kernel.h says "Currently not all data types are supported by Eigen data packing and default gebp_kernel".

FAQ

Does it support FP32?
Does it support FP16?
Does it support INT8?
Does it support BF16?
What is the performance as compared to other backends like OneDNN?

See also