This is a pre-production deployment of Warehouse, however changes made here WILL affect the production instance of PyPI.
Latest Version Dependencies status unknown Test status unknown Test coverage unknown
Project Description

pytocl

A python library to seamlessly convert python functions to functions making use of OpenCL

Setup

python setup.py install

Dependencies

  • Python 3
  • Python libraries
  • numpy
  • pyopencl
  • OpenCL compatible hardware (version 1 is enough)
  • An OpenCL runtime (eg. AMDAPPSDK for AMD CPUs/GPUs)

Usage

The complete example can be found in examples/usage.py. For more examples check tests/cltest.py and examples/benchmarks.py.

1. Creating a function to be converted

Create a function (or convert an existing one) to be parallelized. In this example we calculate output = a + b for vectors, so we will be using one dimension. You can get the current index / global work id in the first dimension with get_global_id(0).

from pytocl import *

def parallel_add(a, b, output):
    i = get_global_id(0)
    output[i] = a[i] + b[i]

2. Converting the function

First we need to create the information for each argument of our function excluding the dimension parameters. For this CLArgDesc is used which holds information about the type of the argument as well as the array_size (0 for scalars) and whether the argument is used as an output (but not necessarily copied back to the host).

# Our vectors will have 16 elements
array_size = 16

dim_shape = (array_size,)

# Create the descriptors for the arguments of the function, excluding the dimension
arg_desc_a = CLArgDesc(CLArgType.float32_array, array_size=array_size) # a
arg_desc_b = CLArgDesc(CLArgType.float32_array, array_size=array_size) # b
arg_desc_output = CLArgDesc(CLArgType.float32_array, array_size=array_size) # output

Next we need to create the descriptor for the function itself which has the gloabl id / dimension shape information and more information about its arguments (ie. whether they’re copied from host to device before execution and whether theyre copied from device to host after execution)

"""
Create the function descriptor with the global id / dimension shape information.

Arguments can be added by chaining .arg() calls (the argument order has to match the original
function's argument order (ie. arg_desc_a -> a, arg_desc_b -> b, arg_desc_output -> output).
is_readonly has to be set to False for arguments that are assigned to in the function.

copy_in() or copy_out() can be called to copy the last added argument from host to
device before execution or from device to host after execution.
"""

func_desc = (CLFuncDesc(parallel_add, global_size)
            .arg(arg_desc_a).copy_in()
            .arg(arg_desc_b).copy_in()
            .arg(arg_desc_output, is_readonly=False).copy_out())

Now we can compile the function which gives us a normal python function we can call

# Compile the actual function, you can also pass an argument to compile with a CL context.
# By default it uses cl.create_some_context()
parallel_add_cl = CLFunc(func_desc).compile()

3. Calling the function

You can call the function like a normal python function now, all the copying will be done for you. The passed scalars can be normal python types but all arrays have to be numpy ndarrays.

import numpy as np

[...]

# Create the host buffers / vectors, the dtype needs to match the arg type of the arg desc
a = np.array([ 1 ] * array_size, dtype=np.float32)
b = np.array([ 2 ] * array_size, dtype=np.float32)
output = np.array([ 0 ] * array_size, dtype=np.float32)

# Now we can execute the compiled function, we need to provide buffers for all output copies.
# For input copies we could also pass None to not copy them
parallel_add_cl(a, b, output)

# output should now be a + b

print("A:", a)
print("B:", b)
print("Output:", output)

You can also pass None for arguments only used as copy-inputs meaning no data will be copied and the current content will be retained.

Limitations / Todo list

  • Only simple python functions are supported.
  • All mathematical and logical expressions
  • All literals except for strings
  • If statements and IfElse constructs
  • While loops
  • For loops currently only support range()
  • Function calls get converted to use the same name in the kernel, but the called functions themselves aren’t converted if they aren’t available yet (eg. currently you can from math import exp and use exp(x))
  • Type inference for local variables is currently limited. Defaults to float, variables named i, j, k or starting with i_ become int, variables starting with b_ become bool
  • return value is not supported, outputs have to be passed as an argument
  • Array slices are not supported
  • List comprehensions and other python-specific constructs are not supported
  • The source code of the function has to be available for conversion (which is often the case)
  • CUDA support?
  • Clean up the converter code

Benchmarks

in examples/benchmarks.py

Test hardware was an AMD Phenom II 1090T and an AMD 6970 The original functions and OpenCL with CPU are orders of magnitude slower (not shown here, you can uncomment the line in benchmarks.py though). Numpy versions are compared to the clified GPU versions.

Matrix Multiply 100 times for matrices of same size

Matrix size Runtime Numpy Runtime OpenCL GPU Relative Numpy Relative OpenCL GPU
(128, 128) 0.02s 0.08s 100.00% 372.00%
(256, 256) 0.08s 0.16s 100.00% 200.18%
(512, 512) 0.40s 0.60s 100.00% 148.81%
(1024, 1024) 2.64s 3.64s 100.00% 137.71%
(2048, 2048) 17.76s 28.07s 100.00% 158.03%

Neural Network sigmoid layer forward pass 100 times for input and weight matrices of same size

Matrix size Runtime Numpy Runtime OpenCL GPU Relative Numpy Relative OpenCL GPU
(128, 128) 0.05s 0.05s 100.00% 114.13%
(256, 256) 0.19s 0.13s 100.00% 70.13%
(512, 512) 2.62s 0.54s 100.00% 20.62%
(1024, 1024) 11.52s 3.65s 100.00% 31.66%
(2048, 2048) 60.01s 27.40s 100.00% 45.66%

2 layer MLP forward pass with 128 batch size and input vectors / weight matrices of same size 100 times

Matrix size Runtime Numpy Runtime OpenCL GPU Relative Numpy Relative OpenCL GPU
(128, 128) 0.09s 0.07s 100.00% 77.65%
(256, 256) 0.43s 0.11s 100.00% 26.72%
(512, 512) 1.35s 0.32s 100.00% 23.80%
(1024, 1024) 3.08s 1.08s 100.00% 35.09%
(2048, 2048) 7.71s 3.56s 100.00% 46.17%

Contributors

  • Toraxxx (Developer)
Release History

Release History

0.1.1

This version

History Node

TODO: Figure out how to actually get changelog content.

Changelog content for this version goes here.

Donec et mollis dolor. Praesent et diam eget libero egestas mattis sit amet vitae augue. Nam tincidunt congue enim, ut porta lorem lacinia consectetur. Donec ut libero sed arcu vehicula ultricies a non tortor. Lorem ipsum dolor sit amet, consectetur adipiscing elit.

Show More

Download Files

Download Files

TODO: Brief introduction on what you do with files - including link to relevant help section.

File Name & Checksum SHA256 Checksum Help Version File Type Upload Date
pytocl-0.1.1.zip (8.5 kB) Copy SHA256 Checksum SHA256 Source Jul 15, 2016

Supported By

WebFaction WebFaction Technical Writing Elastic Elastic Search Pingdom Pingdom Monitoring Dyn Dyn DNS HPE HPE Development Sentry Sentry Error Logging CloudAMQP CloudAMQP RabbitMQ Heroku Heroku PaaS Kabu Creative Kabu Creative UX & Design Fastly Fastly CDN DigiCert DigiCert EV Certificate Rackspace Rackspace Cloud Servers DreamHost DreamHost Log Hosting