Cumo

Cumo (pronounced "koomo") is a CUDA-aware, GPU-optimized numerical library that offers a significant performance boost over Ruby Numo, while (mostly) maintaining drop-in compatibility.

Requirements

• Ruby 2.5 or later
• NVIDIA GPU Compute Capability 3.5 (Kepler) or later
• CUDA 9.0 or later

Preparation

Install CUDA and set your environment variables as follows:

export CUDA_PATH="/usr/local/cuda"
export CPATH="$CUDA_PATH/include:$CPATH"
export LD_LIBRARY_PATH="$CUDA_PATH/lib64:$CUDA_PATH/lib:$LD_LIBRARY_PATH"
export PATH="$CUDA_PATH/bin:$PATH"
export LIBRARY_PATH="$CUDA_PATH/lib64:$CUDA_PATH/lib:$LIBRARY_PATH"

To use cuDNN features, install cuDNN and set your environment variables as follows:

export CUDNN_ROOT_DIR=/path/to/cudnn
export CPATH=$CUDNN_ROOT_DIR/include:$CPATH
export LD_LIBRARY_PATH=$CUDNN_ROOT_DIR/lib64:$LD_LIBRARY_PATH
export LIBRARY_PATH=$CUDNN_ROOT_DIR/lib64:$LIBRARY_PATH

FYI: I use cudnnenv to install cudnn under my home directory like export CUDNN_ROOT_DIR=/home/sonots/.cudnn/active/cuda.

Installation

Add the following line to your Gemfile:

gem 'cumo'

And then execute:

$ bundle

Or install it yourself as:

$ gem install cumo

How To Use

Quick start

An example:

[1] pry(main)> require "cumo/narray"
=> true
[2] pry(main)> a = Cumo::DFloat.new(3,5).seq
=> Cumo::DFloat#shape=[3,5]
[[0, 1, 2, 3, 4],
 [5, 6, 7, 8, 9],
 [10, 11, 12, 13, 14]]
[3] pry(main)> a.shape
=> [3, 5]
[4] pry(main)> a.ndim
=> 2
[5] pry(main)> a.class
=> Cumo::DFloat
[6] pry(main)> a.size
=> 15

Switching from Numo to Cumo

The following find-and-replace should just work:

find . -type f | xargs sed -i -e 's/Numo/Cumo/g' -e 's/numo/cumo/g'

If you want to dynamically switch between Numo and Cumo, something like the following will work:

if gpu
  require 'cumo/narray'
  xm = Cumo
else
  require 'numo/narray'
  xm = Numo
end

a = xm::DFloat.new(3,5).seq

Incompatibility With Numo

The following methods behave incompatibly with Numo by default for performance reasons:

• extract
• []
• count_true
• count_false

Numo returns a Ruby numeric object for 0-dimensional NArray, while Cumo returns the 0-dimensional NArray instead of a Ruby numeric object. Cumo differs in this way to avoid synchronization and minimize CPU ⇄ GPU data transfer.

Set the CUMO_COMPATIBLE_MODE environment variable to ON to force Numo NArray compatibility (for worse performance).

You may enable or disable compatible_mode as:

require 'cumo'
Cumo.enable_compatible_mode # enable
Cumo.compatible_mode_enabled? #=> true
Cumo.disable_compatible_mode # disable
Cumo.compatible_mode_enabled? #=> false

You can also use the following methods which behave like Numo's NArray methods. The behavior of these methods does not depend on compatible_mode.

• extract_cpu
• aref_cpu(*idx)
• count_true_cpu
• count_false_cpu

Select a GPU device ID

Set the CUDA_VISIBLE_DEVICES=id environment variable, or

require 'cumo'
Cumo::CUDA::Runtime.cudaSetDevice(id)

where id is an integer.

Disable GPU Memory Pool

GPU memory pool is enabled by default. To disable it, set CUMO_MEMORY_POOL=OFF, or:

require 'cumo'
Cumo::CUDA::MemoryPool.disable

Documentation

See https://github.com/ruby-numo/numo-narray#documentation, replacing Numo with Cumo.

Contributions

This project is under active development. See issues for future works.

Development

Install ruby dependencies:

bundle install --path vendor/bundle

Compile:

bundle exec rake compile

Run tests:

bundle exec rake test

Generate docs:

bundle exec rake docs

Advanced Development Tips

ccache

ccache would be useful to speedup compilation time. Install ccache and configure with:

export PATH="$HOME/opt/ccache/bin:$PATH"
ln -sf "$HOME/opt/ccache/bin/ccache" "$HOME/opt/ccache/bin/gcc"
ln -sf "$HOME/opt/ccache/bin/ccache" "$HOME/opt/ccache/bin/g++"
ln -sf "$HOME/opt/ccache/bin/ccache" "$HOME/opt/ccache/bin/nvcc"

Build in parallel

Set MAKEFLAGS to specify make command options. You can build in parallel as:

bundle exec env MAKEFLAG=-j8 rake compile

Specify nvcc --generate-code options

bundle exec env CUMO_NVCC_GENERATE_CODE=arch=compute_60,code=sm_60 rake compile

This is useful even on development because it makes it possible to skip JIT compilation of PTX to cubin during runtime.

Run tests with gdb

Compile with debugging enabled:

bundle exec DEBUG=1 rake compile

Run tests with gdb:

bundle exec gdb -x run.gdb --args ruby test/narray_test.rb

You may put a breakpoint by calling cumo_debug_breakpoint() at C source codes.

Run tests only a specific line

--location option is available as:

bundle exec ruby test/narray_test.rb --location 121

Compile and run tests only a specific type

DTYPE environment variable is available as:

bundle exec DTYPE=dfloat rake compile

bundle exec DTYPE=dfloat ruby test/narray_test.rb

Run program always synchronizing CPU and GPU

bundle exec CUDA_LAUNCH_BLOCKING=1

Show GPU synchronization warnings

Cumo shows warnings if CPU and GPU synchronization occurs if:

export CUMO_SHOW_WARNING=ON

By default, Cumo shows warnings that occurred at the same place only once. To show all, multiple warnings, set:

export CUMO_SHOW_WARNING=ON
export CUMO_SHOW_WARNING_ONCE=OFF

Contributing

Bug reports and pull requests are welcome on GitHub at https://github.com/sonots/cumo.

License

• LICENSE.txt
• 3rd_party/LICENSE.txt

Related Materials

• Fast Numerical Computing and Deep Learning in Ruby with Cumo - Presentation Slide at RubyKaigi 2018