zehanort/oclude

An OpenCL driver to test, run and profile standalone kernels on arbitrary devices

oclude's Issues

Make oclude importable

for easier usage from Python scripts (e.g. machine learning applications)

API for custom user input

Create an API for custom user input - a plugin interface.

E.g. support plugins (as shared objects) that the user can write (this needs proper documentation obviously) to parse input from a file/stdin to OpenCL objects/buffers.

See the oclgrind plugin support for inspiration.

remove inline-related functionalities

useless now that no optimizations are used

Rewrite instrumentation with pycparserext, not manually

It most certainly is an enhancement, but it is also a bug, given that some rodinia kernels can not even compile after the instrumentation...

Implement flags for cache manipulation

Specifically:

• --clear-cache: Remove the whole .cache directory
• --ignore-cache: Do not use (possibly) cached info regarding the provided kernel file
• --no-cache-warnings: If the size of the .cache directory exceeds a certain limit, say 100MiB, a warning is shown to the user (implement that, too). This flag suppresses this warning.

Support user-defined buffer lengths and argument numerical limits

possibly with the use of a yaml file

Examine most alarming errors from the tests

More specifically:

• cfd/Kernels.cl
• heartwall/kernel_gpu_opencl.cl
• lavaMD/kernel_gpu_opencl.cl
• particlefilter/particle_double.cl

Use SWIG to export C++ utilities to Python

Proper README/documentation

Must include the following, at least:

• system requirements ☑
• installation ☑
• usage ☑
• limitations & known issues ☑
• license ☑

Add option to run kernel multiple times

maybe discard cache?

Relocate cache directory

Currently, the .cache directory resides in the utils directory. This is probably a bad choice if oclude gets released, i.e. downloadable and installable from pip.
A better idea would be to move it somewhere predefined (e.g. to /home/user/.oclude/.cache e.g. in *nix).

Clear memory at the end of C++ (frees, deletes etc)

More elegant stderr output, more verbosity levels

Commands and kernel info should be printed only on debug level verbosity

compilation with no debug info

it also messes with instructions, i.e. some are not accounted for

Color-coded output

Examine NO OUTPUT errors

Bug with `timeout` flag

only found in the module API.

• sometimes default timeout value (i.e. not specifying a timeout value) leads to the same results every time.
• sometimes defining a timeout value different than 0 leads to oclude hanging.

OpenCL vector support

Custom user structs as OpenCL kernel arguments

cache

Debug (actually, redesign and rewrite) the file caching utility.

Is it possible to rewrite instrumentation-parser in Python?

Separate kernel and context profiling

Export an extra functionality, that of context profiling, but separately. Context profiling every time the --time-it flag is given is both redundant and extremely expensive.

Get address space of ALL memory access LLVM instructions

i.e. not only for load and store. Most importantly getelementptr. These are the instructions: https://llvm.org/docs/LangRef.html#memory-access-and-addressing-operations

API for output formatting

Create a simple interface to pretty-print output (as an addition to what is proposed in issue #4).

Support plugins (as shared objects) to which the OpenCL kernel buffers are exported in order to be processed by the user.

Do not forget proper documentation.