vulkanese

Installation

pip install git+https://github.com/julianfl0w/vulkanese

It's Vulkan-Ease!

This repository

• Imposes a hierarchical structure on Vulkan
• Dramatically simplifies Vulkan usage
• Is pure python
• Runs SPIR-V compute shaders efficiently, across all modern GPUs
• Makes compute shader debugging easy
• Easily integrates with Numpy
• Easily chain up GPU operations using semaphores

It is comparable to

• Nvidia's CUDA
• Linux's Kompute
• Alibaba's MNN

The Topology

For simplicity, the Python Classes contain one another in the following topology

Sponsors

Mandlebrot Example

import vulkanese as ve
# device selection and instantiation
instance_inst = ve.instance.Instance(verbose=True)
# choose a device
print("naively choosing device 0")
device = instance_inst.getDevice(0)
screen = ve.screen.FullScreen()

ve.image.Mandlebrot(
    device=device,
    instance=instance_inst,
    parent=device,
    WIDTH=screen.display.width,
    HEIGHT=screen.display.height
    ).runDemo()

GPGPU Example: Pitch Detection

I've implemented a world class pitch detector in GPU, based on the Loiacono Transform. Here is a snapshot of that code, which shows how to use Vulkanese to manage compute shaders:

import vulkanese as ve

# generate a sine wave at A440, SR=48000
sr = 48000
A4 = 440
z = np.sin(np.arange(2**15)*2*np.pi*A4/sr)

multiple = 40
normalizedStep = 5.0/sr
# create a linear distribution of desired frequencies
fprime = np.arange(100/sr,3000/sr,normalizedStep)

# generate a Loiacono based on this SR
# (this one runs in CPU. reference only)
linst = Loiacono(
    fprime = fprime,
    multiple=multiple,
    dtftlen=2**15
)
linst.debugRun(z)

# begin GPU test
instance = ve.instance.Instance(verbose=False)
device = instance.getDevice(0)
linst_gpu = Loiacono_GPU(
    device = device,
    fprime = fprime,
    multiple = linst.multiple,
)
linst_gpu.gpuBuffers.x.set(z)
for i in range(10):
    linst_gpu.debugRun()
#linst_gpu.dumpMemory()
readstart = time.time()
linst_gpu.spectrum = linst_gpu.gpuBuffers.L.getAsNumpyArray()
print("Readtime " + str(time.time()- readstart))

fig, ((ax1, ax2)) = plt.subplots(1, 2)
ax1.plot(linst.fprime*sr, linst_gpu.spectrum)
ax1.set_title("GPU Result")
ax2.plot(linst.fprime*sr, linst.spectrum)
ax2.set_title("CPU Result")

plt.show()

We have sucessfully detected the 440Hz signal in this simple example:

Future Development

Vulkanese has potential to simplify the fields of

• High-power Computing
• Gaming
• Graphics Rendering

Possible avenues for development include

• A guide for Distributed High-power Computing (HPC) using Docker and Kubernetes (ACTIVE PROJECT)
• A more complete math library
• Machine Learning / AI (Tensorflow Backend)
• Mobile (Android) port for gaming
• Mobile (Android) port for embedded devices
• No-code GUI for pipeline development
• Raytracing support
• Virtual Reality