halide-test

Test v14/v15 performance regression

Note this same simple codebase runs optimized on Jetson Nano (ARM64), Ubuntu Linux on Intel, and Windows with Visual Studio. Without writing any platform-specific code.

Setup

Requires an Ubuntu Linux server running on a modern Intel CPU.

git clone https://github.com/catid/halide-test
cd halide-test

sudo apt install cmake clang llvm

Set up Halide:

mkdir distros
cd distros

wget https://github.com/halide/Halide/releases/download/v14.0.0/Halide-14.0.0-x86-64-linux-6b9ed2afd1d6d0badf04986602c943e287d44e46.tar.gz
tar vzxf Halide-14.0.0-x86-64-linux-6b9ed2afd1d6d0badf04986602c943e287d44e46.tar.gz

wget https://github.com/halide/Halide/releases/download/v15.0.1/Halide-15.0.1-x86-64-linux-4c63f1befa1063184c5982b11b6a2cc17d4e5815.tar.gz
tar vzxf Halide-15.0.1-x86-64-linux-4c63f1befa1063184c5982b11b6a2cc17d4e5815.tar.gz

# Or the release for your platform from: https://github.com/halide/Halide/releases/

cd ..

Running

Run the benchmark with v14:

mkdir build14
cd build14
cmake -DHALIDE_VERSION=14.0.0 ..
make -j
./benchmark

Note the speed. Should be the same each time within 10 microseconds.

Run the benchmark with v15:

mkdir build15
cd build15
cmake -DHALIDE_VERSION=15.0.1 ..
make -j
./benchmark

Results

I'd expect them to produce the same speed output, but v15 runs 50% slower than v14 without running an auto-scheduler.

With auto-scheduler enabled, they're about the same.