jayavanth / chai Goto Github PK

This project is copyright 2012 by Chris Jang ([email protected]).

All source code is licensed under _The Artistic License 2.0_ of _The Perl
Foundation_ except for the sample code (under /sample). Some samples are
taken from PeakStream (TM) whitepapers and presentations. These are indicated
in source code comments.

===============================================================================
What is this?

Chai is a clean-room implementation of the PeakStream (TM) managed platform
for GPGPU.

PeakStream was an array programming language embedded as a domain specific
language in C++. It had a virtual machine and JIT which could target and
schedule across heterogeneous devices (CPUs, GPUs and accelerators). It was
still in a developmental state when the company was acquired in 2007 and the
product line discontinued.

Chai's JIT back-end generates OpenCL (which did not exist during PeakStream's
time). This means it works with the OpenCL SDKs from AMD, Intel, and NVIDIA.
Chai supports compute devices from all three vendors.



===============================================================================
Why recreate it?

PeakStream had a very practical approach to the GPGPU problem. It was a
solution which minimized software development costs. It struck a good balance
between old and new. GPGPU could be a managed platform embedded in C/C++
applications. Native and managed code could co-exist naturally and work
together.

We don't write applications software in low level languages. We use high level
languages. Why should applications that use GPUs be any different?



===============================================================================
How is Chai related to PeakStream?

Other than the inspirational idea, there is no other relationship. I never
used the PeakStream platform product in real life. All I have seen are
marketing whitepapers, presentations, and a few research papers that compare
competing GPGPU platforms circa 2007.

To be honest, this lack of exposure to PeakStream's technology is not out of
respect for clean-room development. Had I been able to find a copy of
PeakStream, I would have looked at it. It is just that Google did a thorough
job of discontinuing PeakStream's products after the acquisition.

The engineering and design of Chai is almost certainly very different from
PeakStream as a result of this independent history.



===============================================================================
Where is this going?

Write once and run anywhere. Code that adapts and optimizes across different
kinds of computers. This is the heterogeneous computing problem.



===============================================================================
Historical comments below:

July 12th 2011 - pre alpha, development code checkpoint
This code is not in a working state. The interpreter works. I believe memory
management works. That includes (primitive but hopefully correct) management
of host arrays and compute device memory buffers and images. Everything JIT
comes next: 1) tracing JIT for glue; 2) auto-tuned high arithmetic intensity
library kernels.

Feb 12 2012 - alpha release, working but needs more work
Code is in a working state with useful end-to-end functionality. Happy path
functionality appears to be reliable. Autotuned GEMM and GEMV work with
dynamically generated kernels from the JIT. Single threaded data parallel
vectorization works for GPU compute devices. Multi-threaded gather/scatter
vectorized scheduling works but is unstable (depends on compute device). Trace
continuation (typically loops) with multiple readouts is working. Autotuning
cold and warm start times are excessive.

Mar 25 2012 - alpha 2 release, refactored JIT and first class integer arrays
JIT code is reorganized into several subdirectories. The main new feature is
unsigned and signed integer array type support. Mixed integer and floating
point calculation is supported (includes autotuned GEMM and GEMV). Generated
kernels make better use of private registers in generated code. Gathering now
works in the kirch.cpp sample. The md5.cpp sample performs vectorized MD5 hash
code calculation on the GPU.

May 25 2012 - alpha 3 release, gathering and random number generation
Gather operations with constant translation stencils (many image processing
filters) are optimized to use images when possible. This takes advantage of the
high speed L1 texture cache. Random number generation on the GPU is supported
with the Random123 counter based PRNG. This includes uniform and normal
distributions (using the Box-Muller transform).

June 24 2012 - alpha 4 release, inline OpenCL kernels in managed code
Mixed language programming with OpenCL kernels and PeakStream DSL. Very large
.cpp files for API, enqueue trace, and memory manager broken up into more
manageably sized pieces. Fixed one unhappy path: array variable width can be
arbitrary, does not need to be a multiple of the underlying vector length.

July 14 2012 - alpha 5 release, extensive refactoring, no functionality changes
Source code restructuring and cleanup motivated by future embedded platform
support (configurable build in Buildroot style? i.e. Buildchai). Interpreter
and JIT are separate paths. Significant false sharing of common code removed -
recognizes the interpreter is mostly for debugging as full language API through
the JIT is supported.