[code-reflection] RFR: Integrate Java Triton example with Intel Triton Backend
hanklo6
duke at openjdk.org
Wed Nov 6 15:49:45 UTC 2024
On Tue, 5 Nov 2024 22:33:02 GMT, Paul Sandoz <psandoz at openjdk.org> wrote:
>> Babylon Java Triton example translates Java source code with Java Triton API into code model by code reflection.
>>
>> In this PR, we traverse the given code model and output Triton MLIR dialect in the generic form, and then inject generated MLIR dialect into the Intel Triton backend. We then utilize Intel Triton backend to compile the Triton MLIR dialect into a SPIR-V module. Use `Jextract` to create Java binding of Intel Level Zero runtime and launch the given kernel function with it on Intel GPUs.
>>
>> ## Usage
>> Navigate to the `cr-example/triton` directory and execute `mvn clean test`. This will generate multiple MLIR files in the `result` directory ready to be processed by the Triton backend.
>>
>> Next, modify the `compiler.py` file within the `intel-xpu-triton-backend` project by applying the patch `git apply add-mlir-insertion.patch`. Then run the Triton backend by running `python3 translate.py`.
>>
>> The Triton backend will generate SPIR-V files, which will be located under `~/.triton/cache/{hash_value}/{kernel_name}/{kernel_name}.spv`.
>>
>> To create a binding for Level Zero, execute the below commands:
>>
>> $JEXTRACT_DIR/bin/jextract --output src/gen/java -I /usr/include -t oneapi.levelzero level-zero/include/ze_api.h
>> $JAVA_HOME/bin/javac -cp target/classes -d target/classes src/gen/java/oneapi/levelzero/*.java
>> $JAVA_HOME/bin/jar cf levelzero.jar -C target/classes/ .
>>
>> The will generate `levelzero.jar` in the current directory.
>>
>> After getting JAR files for Level Zero and `JSON-java`, proceed to compile and run the launcher `LevelZero.java` with the following commands:
>>
>> babylon/build/linux-x86_64-server-release/jdk/bin/javac -cp .:levelzero.jar:json-java.jar LevelZero.java
>> babylon/build/linux-x86_64-server-release/jdk/bin/java -ea -cp .:levelzero.jar:json-java.jar LevelZero
>>
>>
>> Ensure the hash values in`~/.triton/cache` match those used in the `LevelZero.java`.
>>
>> ## Dependencies
>> - [intel-xpu-backend-for-triton](https://github.com/intel/intel-xpu-backend-for-triton)
>> - [intel-extension-for-pytorch](https://github.com/intel/intel-extension-for-pytorch)
>> - [Intel oneAPI base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit-download.html)
>> - [Jextract](https://github.com/openjdk/jextract)
>> - [Level Zero loader](https://github.com/oneapi-src/level-zero)
>> - [compute-runtime](https://github.com/intel/compute-runtime/releases)
>> - [JSON-java](https://github.com/stleary/JSON-java)
>
> cr-examples/triton/src/test/java/oracle/code/triton/TestMatrix.java line 264:
>
>> 262: LessThan);
>> 263: var b = load(b_ptrs, broadcast(offs_k_m_3, b_ptrs.type()));
>> 264: // We accumulate along the K dimension.
>
> Were these necessary changes because you encountered a bug?
Because I changed the return type of the `compare` function to `boolean`, It can't match the return type of the `expand` function.
-------------
PR Review Comment: https://git.openjdk.org/babylon/pull/241#discussion_r1831274958
More information about the babylon-dev
mailing list