Tutorial

Getting started

Before you can use PyCuda, you have to import and initialize it:

import pycuda.driver as cuda
import pycuda.autoinit
from pycuda.compiler import SourceModule

Note that you do not have to use pycuda.autoinit– initialization, context creation, and cleanup can also be performed manually, if desired.

Transferring Data

The next step in most programs is to transfer data onto the device. In PyCuda, you will mostly transfer data from numpy arrays on the host. (But indeed, everything that satisfies the Python buffer interface will work, even a str.) Let’s make a 4x4 array of random numbers:

import numpy
a = numpy.random.randn(4,4)

But wait–a consists of double precision numbers, but most nVidia devices only support single precision:

a = a.astype(numpy.float32)

Finally, we need somewhere to transfer data to, so we need to allocate memory on the device:

a_gpu = cuda.mem_alloc(a.nbytes)

As a last step, we need to transfer the data to the GPU:

cuda.memcpy_htod(a_gpu, a)

Executing a Kernel

For this tutorial, we’ll stick to something simple: We will write code to double each entry in a_gpu. To this end, we write the corresponding CUDA C code, and feed it into the constructor of a pycuda.compiler.SourceModule:

mod = SourceModule("""
  __global__ void doublify(float *a)
  {
    int idx = threadIdx.x + threadIdx.y*4;
    a[idx] *= 2;
  }
  """)

If there aren’t any errors, the code is now compiled and loaded onto the device. We find a reference to our pycuda.driver.Function and call it, specifying a_gpu as the argument, and a block size of 4x4:

func = mod.get_function("doublify")
func(a_gpu, block=(4,4,1))

Finally, we fetch the data back from the GPU and display it, together with the original a:

a_doubled = numpy.empty_like(a)
cuda.memcpy_dtoh(a_doubled, a_gpu)
print a_doubled
print a

This will print something like this:

[[ 0.51360393  1.40589952  2.25009012  3.02563429]
 [-0.75841576 -1.18757617  2.72269917  3.12156057]
 [ 0.28826082 -2.92448163  1.21624792  2.86353827]
 [ 1.57651746  0.63500965  2.21570683 -0.44537592]]
[[ 0.25680196  0.70294976  1.12504506  1.51281714]
 [-0.37920788 -0.59378809  1.36134958  1.56078029]
 [ 0.14413041 -1.46224082  0.60812396  1.43176913]
 [ 0.78825873  0.31750482  1.10785341 -0.22268796]]

It worked! That completes our walkthrough. Thankfully, PyCuda takes over from here and does all the cleanup for you, so you’re done. Stick around for some bonus material in the next section, though.

(You can find the code for this demo as examples/demo.py in the PyCuda source distribution.)

Shortcuts for Explicit Memory Copies

The pycuda.driver.In, pycuda.driver.Out, and pycuda.driver.InOut argument handlers can simplify some of the memory transfers. For example, instead of creating a_gpu, if replacing a is fine, the following code can be used:

func(cuda.InOut(a), block=(4, 4, 1))

Prepared Invocations

Function invocation using the built-in pycuda.driver.Function.__call__() method incurs overhead for type identification (see Device Interface). To achieve the same effect as above without this overhead, the function is bound to argument types (as designated by Python’s standard library struct module), and then called. This also avoids having to assign explicit argument sizes using the numpy.number classes:

func.prepare("P", block=(4,4,1))
func.prepared_call((1, 1), a_gpu)

Bonus: Abstracting Away the Complications

Using a pycuda.gpuarray.GPUArray, the same effect can be achieved with much less writing:

import pycuda.gpuarray as gpuarray
import pycuda.driver as cuda
import pycuda.autoinit
import numpy

a_gpu = gpuarray.to_gpu(numpy.random.randn(4,4).astype(numpy.float32))
a_doubled = (2*a_gpu).get()
print a_doubled
print a_gpu

Advanced Topics

Structures

(contributed by Nicholas Tung, find the code in examples/demo_struct.py)

Suppose we have the following structure, for doubling a number of variable length arrays:

mod = SourceModule("""
    struct DoubleOperation {
        int datalen, __padding; // so 64-bit ptrs can be aligned
        float *ptr;
    };

    __global__ void double_array(DoubleOperation *a) {
        a = &a[blockIdx.x];
        for (int idx = threadIdx.x; idx < a->datalen; idx += blockDim.x) {
            a->ptr[idx] *= 2;
        }
    }
    """)

Each block in the grid (see CUDA documentation) will double one of the arrays. The for loop allows for more data elements than threads to be doubled, though is not efficient if one can guarantee that there will be a sufficient number of threads. Next, a wrapper class for the structure is created, and two arrays are instantiated:

class DoubleOpStruct:
    mem_size = 8 + numpy.intp(0).nbytes
    def __init__(self, array, struct_arr_ptr):
        self.data = cuda.to_device(array)
        self.shape, self.dtype = array.shape, array.dtype
        cuda.memcpy_htod(int(struct_arr_ptr), numpy.int32(array.size))
        cuda.memcpy_htod(int(struct_arr_ptr) + 8, numpy.intp(int(self.data)))
    def __str__(self):
        return str(cuda.from_device(self.data, self.shape, self.dtype))

struct_arr = cuda.mem_alloc(2 * DoubleOpStruct.mem_size)
do2_ptr = int(struct_arr) + DoubleOpStruct.mem_size

array1 = DoubleOpStruct(numpy.array([1, 2, 3], dtype=numpy.float32), struct_arr)
array2 = DoubleOpStruct(numpy.array([0, 4], dtype=numpy.float32), do2_ptr)
print("original arrays", array1, array2)

This code uses the pycuda.driver.to_device() and pycuda.driver.from_device() functions to allocate and copy values, and demonstrates how offsets to an allocated block of memory can be used. Finally, the code can be executed; the following demonstrates doubling both arrays, then only the second:

func = mod.get_function("double_array")
func(struct_arr, block = (32, 1, 1), grid=(2, 1))
print("doubled arrays", array1, array2)

func(numpy.intp(do2_ptr), block = (32, 1, 1), grid=(1, 1))
print("doubled second only", array1, array2, "\n")

Where to go from here

Once you feel sufficiently familiar with the basics, feel free to dig into the Device Interface. For more examples, check the in the examples/ subdirectory of the distribution. This folder also contains several benchmarks to see the difference between GPU and CPU based calculations. As a reference for how stuff is done, PyCuda’s test suite in the test/ subdirectory of the distribution may also be of help.