/vml

Primary LanguageC++BSD 2-Clause "Simplified" LicenseBSD-2-Clause

This repository includes machine learning kernels for Vector Engine and its wrapper for TensorFlow for SX-Aurora.

Prerequisites

  • ncc/nc++
  • llvm-ve
  • vednn

build

build veorun_tf for TensorFlow-VE

% ./scripts/build-release-veorun_tf.sh build

build with prebuilt vednn in extra/vednn directory

% mkdir build
% (cd build && cmake3 -DUSE_PREBUILT_VEDNN=ON .. && make)

build vednn and vetfkernel

% git clone <vednn> libs/vednn
% mkdir build
% (cd build && cmake3 .. && make)

test

% (cd build && ctest3)
% ./build/bench/bench --validation-only
% python test/python/avgpoolgrad.py

or

% ./test.sh [BUILD_DIR]

Then, run perf test

% python perf.py [-e <BUILD_DIR>/bench/bench] -d perfdb/10BE test

See here for details.

profiler

  1. cmake with -DUSE_PROFILE=ON
  2. run with VML_PROFILE=conv2d,conv2d_backprop_input

VML API Document

The document is generated into doc/api directory.

% doxygen

See in your browser. If you need http server, try

% python3 -m http.server