tpu-ppl 1.6.19

A language and compiler for custom Deep Learning operations

ppl

Primitive Programming Language

Build instructions

Requirements

• Working C and C++ toolchains(compiler[clang], linker)
• cmake
• ninja

1. install git lfs

curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh | sudo bash

sudo apt-get install git-lfs

setting up lfs for the first time

git lfs install

2. clone PPL

2.1 If you had not pull code using the ssh method:

git config --global http.sslVerify false git clone git config --global http.sslVerify true

2.2 If you had pull code using the ssh method:

Change ssh to http

git remote set-url origin https://<your_name>@

2.3 set up for http

2.3.1 set http repo

cd ppl

2.3.2 Disable Locked Authentication

git config lfs.https://gerrit-ai.so#ph#go.vip/ppl.git/info/lfs.locksverify false

2.3.3 Local disk to save credentials (have security risk, but no need to enter a password every time you interact with the remote library)

note : If you have changed your password, just execute the following statement

git config --global credential.helper store

2.3.4 Skip ssl certificate verification

git config http.sslVerify false

2.3.5 When you change the password, you need to execute the statement below to update the locally saved password

git config --global credential.helper store

3. Get the LFS objects contained in the current commit

git lfs fetch git lfs checkout or: git lfs pull

4. If need update runtime

source envsetup.sh git config --global http.sslVerify false (suggested update outside docker) git submodule update --init git config --global http.sslVerify true ./update.sh

5. How to use git lfs

refer to: https://wiki.so#ph#go.com/pages/viewpage.action?pageId=127019078

6. Install LLVM, MLIR, Clang, and PPL

Using LLVM release files

download llvm+mlir-17.0.0-x86_64-linux-gnu-ubuntu-18.04-release.tar.gz to third_party

cd ppl/third_party
scp guest@172.22.12.22:/data/ppl_third_party/llvm+mlir-17.0.0-x86_64-linux-gnu-ubuntu-18.04-release.tar.gz .
(password:123456)
tar -xvhf llvm+mlir-17.0.0-x86_64-linux-gnu-ubuntu-18.04-release.tar.gz
ln -sf llvm+mlir-17.0.0-x86_64-linux-gnu-ubuntu-18.04-release llvm_release

7. build ppl

cd ppl
./build.sh
# ./build.sh DEBUG (debug mode)

Push Code

cd existing_repo
git pull -r
git push origin HEAD:refs/for/main

Run

ppl-compile options: --function : function need to be converted --I : include search path --print-debug-info : print loc information --print-ir : print ir --gen-ref : generate reference groups --chip : chip type --x : treat input as language --O0 : Optimisation level 0: none optimisation --O1 : Optimisation level 1: do ppl canonicalize --O2 : Optimisation level 2: do ppl pipeline and ppl canonicalize --O3 : Optimisation level 3: do all ppl optimisation --o : Output path for ir and c files --pm-enable-printing : Enable printing of IR before and after all passed

Test

You can use the script named ppl_compile.py to test the xx.pl file

For example: you can test xx.pl by executing the following command.

ppl_compile.py --src ./examples/matmul/mm2.pl --chip bm1684x --gen_test

Profiling

preprare PerfAI tool

download PerfAI tool from FTP:172.28.141.89 /perfAI_release/perfAI_release mv PerfAI_Release_xxx.tar.gz to ppl/third_party tar -xvhf PerfAI_Release_xxx.tar.gz ln -sf PerfAI_Release_xxx.tar.gz PerfAI cd PerfAI pip install -r requirements.txt

run

cd ppl/third_party/PerfAI/ source envsetup.sh ppl_compile.py --src examples/sdma/add_c_dual_loop.pl --chip bm1690 --profiling

profiling files will generate in test_add_c_dual_loop/profiling

CI

git clone ssh://[e-mail]#ph#go.vip:29418/jenkins_pipeline.git

jenkins script regression/ppl_ci_test.groovy

dockerfile envsetup/docker/ppl_dailybuild.dockerfile

third-party version

triton 5df904233c11a65bd131ead7268f84cca7804275 llvm 2538e550420f