cltorch-benchmarking

cltorch benchmarking, for evaluating where to focus optimization effort

This is cltorch-specific for now, though if someone wants to make it more general, I'm happy to change the name, eg to torch-benchmarking :-)

Current direction is to measure why char-rnn runs really slowly, on opencl, on certain devices. Examples of things to check:

• is it because of kernel launch time?
• is it because of passing in structs?
• is it because of passing in non-const structs?
• is it because of all those dimension loops?
• is it because the various reduceAll calls are causing sync points?
• is it because of excessive sync points generally?

Contents

• test_launch: measure kernel launch times, by adding 1 to a constant-sized array (about 100MB), and varying the number of kernel launches used
• test_apply1: varies vector size, float vs float4. varies operation used, ie + vs -, exp, etc
• test_apply1b: varying operation, as test_apply1, but adds an additional temporary variable out
• test_applystrided: (in progress) mix up the memory access a bit, and/or add an inner loop over dimensions (tbd)

To build

pre-requisites:

• EasyCL installed, using make -j 4 install, into ~/git/EasyCL/dist (ie install easycl, with a CMAKE_INSTALL_PREFIX of [your home directory]/git/EasyCL/dist
• cmake and ccmake installed
• gcc, g++ etc

method

git clone https://github.com/hughperkins/cltorch-benchmarking.git
cd cltorch-benchmarking
mkdir build
cd build
cmake ..
make -j 4