Extending Theano with a GPU Op


This covers the gpuarray back-end for the GPU.

This tutorial covers how to extend Theano with an op that offers a GPU implementation. It assumes you are familiar with how to write new Theano ops. If that is not the case you should probably follow the Creating a new Op: Python implementation and Extending Theano with a C Op sections before continuing on.

Writing a new GPU op can be done in Python for some simple tasks, but will usually done in C to access the complete API and avoid paying the overhead of a Python function call.

Dealing With the Context

One of the major differences with GPU ops is that they require a context (a.k.a. device) to execute. Most of the time you can infer the context to run on from your inputs. There is a way for the user to transfer things between contexts and to tag certain variables for transfer. It might also be the case that your inputs are not all from the same context and you would have to choose which one to run on.

In order to support all of those options and have a consistent interface, theano.gpuarray.basic_ops.infer_context_name() was written. An example usage is below:

def make_node(self, a, b, c):
    ctx = infer_context_name(a, b, c)
    a = as_gpuarray_variable(a, ctx)
    b = as_gpuarray_variable(b, ctx)
    c = as_gpuarray_variable(c, ctx)
    return Apply(self, [a, b, c], [a.type()])

In this example the Op takes three inputs, all on the GPU. In case one or more of your inputs is not supposed to be on the GPU, you should not pass it to infer_context_name() or call as_gpuarray_variable() on it.

Also note that theano.gpuarray.basic_ops.as_gpuarray_variable() takes context_name as a mandatory parameter. This is because it’s not enough to know you want the value to be on the GPU, you also want to know which GPU to put it on. In almost all cases, you can pass in the return value of infer_context_name() there.

If you also need the context during runtime (for example to allocate the output), you can use the context of one of your inputs to know which one to use. Here is another example:

def perform(self, node, inputs, output_storage):
    A, B = inputs
    C, = output_storage
    C[0] = pygpu.empty([A.shape[0], B.shape[1]], dtype=A.dtype, A.context)
    pygpu.blas.gemm(1, A, B, 0, C, overwrite_c=True)

Finally if you require the context before perform, such as during make_thunk() to initialize kernels and such, you can access the context of your inputs through the type of the variables:

def make_thunk(self, node, storage_map, compute_map, no_recycling):
    ctx = node.inputs[0].type.context

Note that GpuArrayType objects also have a context_name attribute which is the symbolic equivalent of context. It can’t be used for calls to pygpu or libgpuarray, but it should be used for theano operations and variables.

The last place where you might need the context is in the C initialization code. For that you will have to use the params. The params type should be theano.gpuarray.type.gpu_context_type and the params object should be a context object from one of your input variables:

def get_params(self, node):
    return node.inputs[0].type.context

If you don’t have any input variables on the GPU you can follow the the example of GpuFromHost or GpuEye. This is not a case that you should encounter often, so it will not be covered further.

Defining New Kernels

If your op needs to do some transformation on the data, chances are that you will need to write a new kernel. The best way to do this is to leverage GpuKernelBase (or CGpuKernelBase if you want to use the COp functionality).

For plain GpuKernelBase, you have to define a method called gpu_kernels which returns a list of Kernel objects. You can define as many kernels as you want for a single op. An example would look like this:

def gpu_kernels(self, node, name):
    code = """
KERNEL void k(GLOBAL_MEM ga_double *a, ga_size n, ga_size m) {
    ga_size nb = n < m ? n : m;
    for (ga_size i = LID_0; i < nb; i += LDIM_0) {
        a[i*m + i] = 1;
    return [Kernel(
            code=code, name="k",
            params=[gpuarray.GpuArray, gpuarray.SIZE, gpuarray.SIZE],

If you want to use COp, then you should use CGpuKernelBase instead. It adds a new section to the parsed files whose tag is kernels. Inside that section you can define some kernels with #kernel name:params:flags.

Here name is the name of the kernel function in the following code, params is a comma-separeted list of numpy typecode names. There are three exceptions for size_t which should be noted as size, ssize_t which should be noted as ssize and a pointer which should be noted as *.

flags is a |-separated list of C kernel flag values (can be empty). The same kernel definition as above would look like this with CGpuKernelBase:

#section kernels

#kernel k : *, size, size : GA_USE_DOUBLE

KERNEL void k(GLOBAL_MEM ga_double *a, ga_size n, ga_size m) {
    ga_size nb = n < m ? n : m;
    for (ga_size i = LID_0; i < nb; i += LDIM_0) {
    a[i*m + i] = 1;

The second method is to handle the kernel compilation and cache on your own. This is not recommended because there are lots of details to pay attention to that can cripple your performance if not done right, which GpuKernelBase handles for you. But if you really want to go this way, then you can look up the C API for kernels in libgpuarray.

In any case you will need to call your compiled kernel with some data, in most cases in your c_code() method. This is done by using the provided wrapper function. An example calling the above kernel would be:

size_t dims[2];
size_t n = 256;

// ...

err = k_scall(1, &n, 0, input->ga.data, dims[0], dims[1]);

// ...

If you want explicit control over the scheduling, you can use the _call wrapper instead which works like this:

size_t ls, gs;

// ...

gs = 1;
ls = 256;
err = k_call(1, &gs, &ls, 0, input->ga.data, dims[0], dims[1]);

The name of the wrapper function depends on the name you passed to Kernel() when you declared it (or the name in your #kernel statement). It defaults to name + ‘_call’ or ‘_scall’.

For other operations in the C code you should refer to the libgpuarray documentation.

Dealing with float16

To support limited-precision storage in a kernel you have to be careful to load values properly, declare working memory in float32 and write results properly. To help with that some functions have been declared in theano.gpuarray.fp16_help.

To load the inputs you should wrap the read with the function returned by load_w(). Similarly writes should be wrapped in the function returned by write_w(). Finally working data should have the type returned by work_dtype().

Here is a +1 kernel that is not ready to deal with float16 input:

type_x = dtype_to_ctype(x.dtype)
type_y = dtype_to_ctype(y.dtype)
KERNEL void k(const ga_size n, %(type_x)s *x, %(type_y)s *y) {
    ga_size i = GID_0 * LDIM_0 + LID_0;
    %(type_x) z = x[i];
    z += 1;
    y[i] = z;
""" % dict(dtype_x=dtype_x, dtype_y=dtype_y)

Here is the same kernel, but now ready to handle float16:

type_x = dtype_to_ctype(x.dtype)
type_y = dtype_to_ctype(y.dtype)
work_x = dtype_to_ctype(work_dtype(x.dtype))
load_x = load_w(x.dtype)
write_y = write_w(y.dtype)
KERNEL void k(const ga_size n, %(type_x)s *x, %(type_y)s *y) {
    ga_size i = GID_0 * LDIM_0 + LID_0;
    %(work_x) z = %(load_x)(x[i]);
    z += 1;
    y[i] = %(write_y)(z);
""" % dict(dtype_x=dtype_x, dtype_y=dtype_y, work_x=work_x, load_x=load_x,

Once you have converted your kernels for float16 support you need to tag your op with _f16_ok = True so that the linker will accept to generate C code on float16 types. This is done by inserting it as a class property like this:

class SomeOp(Op):
    _f16_ok = True

If this attribute is not present or is False, the linker will print a message saying that it’s refusing to use C code for float16 for the op.

A Complete Example

This is a complete example using both approches for a implementation of the Eye operation.


Python File

class GpuEye(GpuKernelBase, Op):
    Eye for GPU.


    __props__ = ("dtype", "context_name")
    _f16_ok = True

    def __init__(self, dtype=None, context_name=None):
        if dtype is None:
            dtype = config.floatX
        self.dtype = dtype
        self.context_name = context_name

    def get_params(self, node):
        return get_context(self.context_name)

    def make_node(self, n, m, k):
        n = tensor.as_tensor_variable(n)
        m = tensor.as_tensor_variable(m)
        k = tensor.as_tensor_variable(k)
        assert n.ndim == 0
        assert m.ndim == 0
        assert k.ndim == 0
        otype = GpuArrayType(
            broadcastable=(False, False),

        return Apply(self, [n, m, k], [otype()])

    def infer_shape(self, node, in_shapes):
        out_shape = [node.inputs[0], node.inputs[1]]
        return [out_shape]

    def grad(self, inp, grads):
        return [grad_undefined(self, i, inp[i]) for i in range(3)]

    def gpu_kernels(self, node, name):
        code = """#include "cluda.h"

KERNEL void eye(GLOBAL_MEM %(ctype)s *a, ga_size a_off,
                ga_size n, ga_size m, ga_ssize k) {
    a = (GLOBAL_MEM %(ctype)s *)(((GLOBAL_MEM char *)a) + a_off);
    ga_ssize coff = max(k, (ga_ssize) 0);
    ga_ssize roff = -min(k, (ga_ssize) 0);
    ga_size nb = (ga_size) min(n - roff, m - coff);
    for (ga_size i = LID_0; i < nb; i += LDIM_0) {
        a[(i + roff)*m + i + coff] = %(write_a)s(1);
}""" % dict(
        return [
                objvar="k_eye_" + name,

    def c_code(self, node, name, inp, out, sub):
        if len(inp) == 2:
            n, m = inp
            k = 0
        elif len(inp) == 3:
            n, m, k = inp

        (z,) = out
        fail = sub["fail"]
        ctx = sub["params"]
        typecode = pygpu.gpuarray.dtype_to_typecode(self.dtype)
        kname = self.gpu_kernels(node, name)[0].objvar
        s = (
        size_t dims[2] = {0, 0};
        size_t ls, gs;
        ssize_t k;
        size_t col_off;
        size_t row_off;
        int err;

        dims[0] = ((dtype_%(n)s*)PyArray_DATA(%(n)s))[0];
        dims[1] = ((dtype_%(m)s*)PyArray_DATA(%(m)s))[0];
        k = ((dtype_%(k)s*)PyArray_DATA(%(k)s))[0];


        %(z)s = pygpu_zeros(2, dims,
                            %(ctx)s, Py_None);
        if (%(z)s == NULL) {

        ls = 1;
        gs = 256;
        col_off = (size_t) (k > 0?k:0);
        row_off = (size_t) (k < 0?-k:0);
        if (row_off < dims[0] && col_off < dims[1]) {
            err = eye_call(1, &gs, &ls, 0, %(z)s->ga.data, %(z)s->ga.offset,
                           dims[0], dims[1], k);
            if (err != GA_NO_ERROR) {
                             "gpuarray error: kEye: %%s. n%%lu, m=%%lu.",
                             GpuKernel_error(&%(kname)s, err),
                             (unsigned long)dims[0], (unsigned long)dims[1]);

            % locals()

        return s

    def c_code_cache_version(self):
        return (10,)


Python File


Wrapping Exisiting Libraries


For things in PyCUDA (or things wrapped with PyCUDA), we usually need to create a PyCUDA context. This can be done with the following code:

with gpuarray_cuda_context:
    pycuda_context = pycuda.driver.Context.attach()

If you don’t need to create a context, because the library doesn’t require it, you can also just use the pygpu context and a with statement like above for all your code which will make the context the current context on the cuda stack.

GpuArray objects are compatible with PyCUDA and will expose the necessary interface so that they can be used in most things. One notable exception is PyCUDA kernels which require native objects. If you need to convert a pygpu GpuArray to a PyCUDA GPUArray, this code should do the trick:

assert pygpu_array.flags['IS_C_CONTIGUOUS']
pycuda_array = pycuda.gpuarray.GPUArray(pygpu_array.shape,
                                        gpudata=(pygpu_array.gpudata +

As long as the computations happen on the NULL stream there are no special considerations to watch for with regards to synchronization. Otherwise, you will have to make sure that you synchronize the pygpu objects by calling the .sync() method before scheduling any work and synchronize with the work that happens in the library after all the work is scheduled.