Android-GPU-Compute-Talk

WEC talk - GPU Compute on Android (using OpenGLES Compute)!

Prerequisites

• Excellent C programming skills
• An Android phone, running Android 11 or newer and
• Termux app installed on that phone

Project Structure

No need to panic looking at this huge project.

This is all you need to look at:

.
├── Resources -------(Books/Slides)
|
├── array-add-cpu ---(Adding 2 arrays on CPU)
|
├── array-add-gpu ---(Adding 2 arrays on GPU)
|
├── array-add-gpu-v2 (Adding 2 arrays on GPU-v2)
|
├── mat-cpu ---------(4x4 Matrix mult. on CPU)
|
├── mat-gpu ---------(4x4 Matrix mult. on GPU)
|
├── src -------------(Main Library and C code)
|
├── shaders ---------(ESSL shader source)
|
├── README.md -------(Read this FIRST, ofc)
|
├── LICENSE ---------(Read this, ofc)
|
└── Ignore all others!

Performance Stats

On my phone (Mediatek Helio-G85 SOC and 4GB RAM), I ran tests with 3 arrays (f32), each having 65535 x 384 elements (f32s) in all cases. The array is around 96MiB each.

CPU variant:

Compute time: ~140ms
Total time: ~1.2s
RAM consumed:

Output:

GPU variant:

Compute time: ~39ms
Total time: ~1.3s
RAM consumed:

Output:

GPU variant-v2: (Uses Hardware Buffers)

Compute time: ~37ms
Total time: ~1.1s
RAM consumed:

Output:

When the Array has 262140 x 384 elements, we see the following numbers:

CPU variant:

Compute time: ~600ms
Total time: ~4.8s
RAM consumed: ~1.2GB

GPU variant: Out of Memory

GPU variant-v2: (Uses Hardware Buffers)

Compute time: ~135ms
Total time: ~4.2s
RAM consumed: ~1.2GB

When doing 4x4 (f32) Matrix multiplication with 16384 x 384 matrices, we see the following:

CPU variant:

Compute time: ~700ms
Total time: ~20s
Output:

GPU variant: (Uses Hardware Buffers)

Compute time: ~250ms
Total time: ~4.6s
Output:

To see real benefits, the operation needs to be math-heavy and you need to queue up a ton of work!

These values are device dependent. Its just for illustration!

Build prerequisites

I recommend you to download the pre-compiled binaries which can be found in the releases section.

However, if you want to build/compile this project on your machine, youll need:

• Linux or MacOS
• GNU make
• Android NDK (LTS version)
• clang, llvm-ar, lld, rustc (2024 edition) and cargo

Before building, you must update the NDK_HOME variable in the Makefile. After that, run the following to build everything:

make

All the executables can be found in target/aarch64-linux-android/release folder. You will also need the shaders folder for the programs to work properly.

NOTE: The executables run ONLY on Android-Aarch64 devices!