Parallel Algorithms¶

Element-wise expression evaluation (“map”)¶

Evaluating involved expressions on pyopencl.array.Array instances by using overloaded operators can be somewhat inefficient, because a new temporary is created for each intermediate result. The functionality in the module pyopencl.elementwise contains tools to help generate kernels that evaluate multi-stage expressions on one or several operands in a single pass.

class pyopencl.elementwise.ElementwiseKernel(context: Context, arguments: str | List[DtypedArgument], operation: str, name: str = 'elwise_kernel', options: Any = None, **kwargs: Any)[source]¶

A kernel that takes a number of scalar or vector arguments and performs an operation specified as a snippet of C on these arguments.

Parameters:

arguments – a string formatted as a C argument list.
operation – a snippet of C that carries out the desired ‘map’ operation. The current index is available as the variable i. operation may contain the statement PYOPENCL_ELWISE_CONTINUE, which will terminate processing for the current element.
name – the function name as which the kernel is compiled
options – passed unmodified to pyopencl.Program.build().
preamble – a piece of C source code that gets inserted outside of the function context in the elementwise operation’s kernel source code.

Warning

Using a return statement in operation will lead to incorrect results, as some elements may never get processed. Use PYOPENCL_ELWISE_CONTINUE instead.

Changed in version 2013.1: Added PYOPENCL_ELWISE_CONTINUE.

__call__(*args, **kwargs) → Event[source]¶

Invoke the generated scalar kernel.

The arguments may either be scalars or pyopencl.array.Array instances.

Returns a new pyopencl.Event. wait_for may either be None or a list of pyopencl.Event instances for whose completion this command waits before starting execution.

Here’s a usage example:

import numpy as np

import pyopencl as cl
import pyopencl.array
from pyopencl.elementwise import ElementwiseKernel


n = 10

rng = np.random.default_rng()
a_np = rng.random(n, dtype=np.float32)
b_np = rng.random(n, dtype=np.float32)

ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)

a_g = cl.array.to_device(queue, a_np)
b_g = cl.array.to_device(queue, b_np)

lin_comb = ElementwiseKernel(ctx,
    "float k1, float *a_g, float k2, float *b_g, float *res_g",
    "res_g[i] = k1 * a_g[i] + k2 * b_g[i]",
    "lin_comb")

res_g = cl.array.empty_like(a_g)
lin_comb(2, a_g, 3, b_g, res_g)

# Check on GPU with PyOpenCL Array:
print((res_g - (2 * a_g + 3 * b_g)).get())

# Check on CPU with Numpy:
res_np = res_g.get()
print(res_np - (2 * a_np + 3 * b_np))
print(np.linalg.norm(res_np - (2 * a_np + 3 * b_np)))

(You can find this example as examples/demo_elementwise.py in the PyOpenCL distribution.)

Sums and counts (“reduce”)¶

class pyopencl.reduction.ReductionKernel(ctx: Context, dtype_out: Any, neutral: str, reduce_expr: str, map_expr: str | None = None, arguments: str | List[DtypedArgument] | None = None, name: str = 'reduce_kernel', options: Any = None, preamble: str = '')[source]¶

A kernel that performs a generic reduction on arrays.

Generate a kernel that takes a number of scalar or vector arguments (at least one vector argument), performs the map_expr on each entry of the vector argument and then the reduce_expr on the outcome of that. neutral serves as an initial value. preamble offers the possibility to add preprocessor directives and other code (such as helper functions) to be added before the actual reduction kernel code.

Vectors in map_expr should be indexed by the variable i. reduce_expr uses the formal values “a” and “b” to indicate two operands of a binary reduction operation. If you do not specify a map_expr, in[i] is automatically assumed and treated as the only one input argument.

dtype_out specifies the numpy.dtype in which the reduction is performed and in which the result is returned. neutral is specified as float or integer formatted as string. reduce_expr and map_expr are specified as string formatted operations and arguments is specified as a string formatted as a C argument list. name specifies the name as which the kernel is compiled. options are passed unmodified to pyopencl.Program.build(). preamble specifies a string of code that is inserted before the actual kernels.

__init__(ctx: Context, dtype_out: Any, neutral: str, reduce_expr: str, map_expr: str | None = None, arguments: str | List[DtypedArgument] | None = None, name: str = 'reduce_kernel', options: Any = None, preamble: str = '') → None[source]¶

__call__(*args: Any, **kwargs: Any) → Event[source]¶

Invoke the generated kernel.

wait_for may either be None or a list of pyopencl.Event instances for whose completion this command waits before starting execution.

With out the resulting single-entry pyopencl.array.Array can be specified. Because offsets are supported one can store results anywhere (e.g. out=a[3]).

Note

The returned pyopencl.Event corresponds only to part of the execution of the reduction. It is not suitable for profiling.

Added in version 2011.1.

Changed in version 2014.2: Added out parameter.

Changed in version 2016.2: range_ and slice_ added.

Parameters:

range – A slice object. Specifies the range of indices on which the kernel will be executed. May not be given at the same time as slice.
slice – A slice object. Specifies the range of indices on which the kernel will be executed, relative to the first vector-like argument. May not be given at the same time as range.
return_event – a boolean flag used to return an event for the reduction.

Returns:

the resulting scalar as a single-entry pyopencl.array.Array if return_event is False, otherwise a tuple (scalar_array, event).

Here’s a usage example:

a = pyopencl.array.arange(queue, 400, dtype=numpy.float32)
b = pyopencl.array.arange(queue, 400, dtype=numpy.float32)

krnl = ReductionKernel(ctx, numpy.float32, neutral="0",
        reduce_expr="a+b", map_expr="x[i]*y[i]",
        arguments="__global float *x, __global float *y")

my_dot_prod = krnl(a, b).get()

Prefix Sums (“scan”)¶

A prefix sum is a running sum of an array, as provided by e.g. numpy.cumsum():

>>> import numpy as np
>>> a = [1,1,1,1,1,2,2,2,2,2]
>>> np.cumsum(a)
array([ 1,  2,  3,  4,  5,  7,  9, 11, 13, 15])

This is a very simple example of what a scan can do. It turns out that scans are significantly more versatile. They are a basic building block of many non-trivial parallel algorithms. Many of the operations enabled by scans seem difficult to parallelize because of loop-carried dependencies.

Usage Example¶

This example illustrates the implementation of a simplified version of pyopencl.algorithm.copy_if(), which copies integers from an array into the (variable-size) output if they are greater than 300:

knl = GenericScanKernel(
        ctx, np.int32,
        arguments="__global int *ary, __global int *out",
        input_expr="(ary[i] > 300) ? 1 : 0",
        scan_expr="a+b", neutral="0",
        output_statement="""
            if (prev_item != item) out[item-1] = ary[i];
            """)

out = a.copy()
knl(a, out)

a_host = a.get()
out_host = a_host[a_host > 300]

assert (out_host == out.get()[:len(out_host)]).all()

The value being scanned over is a number of flags indicating whether each array element is greater than 300. These flags are computed by input_expr. The prefix sum over this array gives a running count of array items greater than 300. The output_statement the compares prev_item (the previous item’s scan result, i.e. index) to item (the current item’s scan result, i.e. index). If they differ, i.e. if the predicate was satisfied at this position, then the item is stored in the output at the computed index.

This example does not make use of the following advanced features also available in PyOpenCL:

Segmented scans
Access to the previous item in input_expr (e.g. for comparisons) See the implementation of pyopencl.algorithm.unique() for an example.

Making Custom Scan Kernels¶

Added in version 2013.1.

class pyopencl.scan.GenericScanKernel(ctx: Context, dtype: Any, arguments: str | List[DtypedArgument], input_expr: str, scan_expr: str, neutral: str | None, output_statement: str, is_segment_start_expr: str | None = None, input_fetch_exprs: List[Tuple[str, str, int]] | None = None, index_dtype: Any = None, name_prefix: str = 'scan', options: Any = None, preamble: str = '', devices: Device | None = None)[source]¶

Generates and executes code that performs prefix sums (“scans”) on arbitrary types, with many possible tweaks.

Usage example:

from pyopencl.scan import GenericScanKernel
knl = GenericScanKernel(
        context, np.int32,
        arguments="__global int *ary",
        input_expr="ary[i]",
        scan_expr="a+b", neutral="0",
        output_statement="ary[i+1] = item;")

a = cl.array.arange(queue, 10000, dtype=np.int32)
knl(a, queue=queue)

__init__(ctx: Context, dtype: Any, arguments: str | List[DtypedArgument], input_expr: str, scan_expr: str, neutral: str | None, output_statement: str, is_segment_start_expr: str | None = None, input_fetch_exprs: List[Tuple[str, str, int]] | None = None, index_dtype: Any = None, name_prefix: str = 'scan', options: Any = None, preamble: str = '', devices: Device | None = None) → None[source]¶

Parameters:

ctx – a pyopencl.Context within which the code for this scan kernel will be generated.
dtype – the numpy.dtype with which the scan will be performed. May be a structured type if that type was registered through pyopencl.tools.get_or_register_dtype().
arguments – A string of comma-separated C argument declarations. If arguments is specified, then input_expr must also be specified. All types used here must be known to PyOpenCL. (see pyopencl.tools.get_or_register_dtype()).
scan_expr –
The associative, binary operation carrying out the scan, represented as a C string. Its two arguments are available as a and b when it is evaluated. b is guaranteed to be the ‘element being updated’, and a is the increment. Thus, if some data is supposed to just propagate along without being modified by the scan, it should live in b.

This expression may call functions given in the preamble.

Another value available to this expression is across_seg_boundary, a C bool indicating whether this scan update is crossing a segment boundary, as defined by is_segment_start_expr. The scan routine does not implement segmentation semantics on its own. It relies on scan_expr to do this. This value is available (but always false) even for a non-segmented scan.

Note

In early pre-releases of the segmented scan, segmentation semantics were implemented without relying on scan_expr.
input_expr –
A C expression, encoded as a string, resulting in the values to which the scan is applied. This may be used to apply a mapping to values stored in arguments before being scanned. The result of this expression must match dtype. The index intended to be mapped is available as i in this expression. This expression may also use the variables defined by input_fetch_expr.

This expression may also call functions given in the preamble.
output_statement –
a C statement that writes the output of the scan. It has access to the scan result as item, the preceding scan result item as prev_item, and the current index as i. prev_item in a segmented scan will be the neutral element at a segment boundary, not the immediately preceding item.

Using prev_item in output statement has a small run-time cost. prev_item enables the construction of an exclusive scan.

For non-segmented scans, output_statement may also reference last_item, which evaluates to the scan result of the last array entry.
is_segment_start_expr –
A C expression, encoded as a string, resulting in a C bool value that determines whether a new scan segments starts at index i. If given, makes the scan a segmented scan. Has access to the current index i, the result of input_expr as a, and in addition may use arguments and input_fetch_expr variables just like input_expr.

If it returns true, then previous sums will not spill over into the item with index i or subsequent items.
input_fetch_exprs –
a list of tuples (NAME, ARG_NAME, OFFSET). An entry here has the effect of doing the equivalent of the following before input_expr:
```
ARG_NAME_TYPE NAME = ARG_NAME[i+OFFSET];
```
OFFSET is allowed to be 0 or -1, and ARG_NAME_TYPE is the type of ARG_NAME.
preamble – A snippet of C that is inserted into the compiled kernel before the actual kernel function. May be used for, e.g. type definitions or include statements.

The first array in the argument list determines the size of the index space over which the scan is carried out, and thus the values over which the index i occurring in a number of code fragments in arguments above will vary.

All code fragments further have access to N, the number of elements being processed in the scan.

__call__(*args: Any, **kwargs: Any) → Event[source]¶

Returns a new pyopencl.Event. wait_for may either be None or a list of pyopencl.Event instances for whose completion this command waits before starting execution.

Note

The returned pyopencl.Event corresponds only to part of the execution of the scan. It is not suitable for profiling.

Parameters:

queue – queue on which to execute the scan. If not given, the queue of the first pyopencl.array.Array in args is used
allocator – an allocator for the temporary arrays and results. If not given, the allocator of the first pyopencl.array.Array in args is used.
size – specify the length of the scan to be carried out. If not given, this length is inferred from the first argument
wait_for – a list of events to wait for.

Debugging aids¶

class pyopencl.scan.GenericDebugScanKernel(ctx: Context, dtype: Any, arguments: str | List[DtypedArgument], input_expr: str, scan_expr: str, neutral: str | None, output_statement: str, is_segment_start_expr: str | None = None, input_fetch_exprs: List[Tuple[str, str, int]] | None = None, index_dtype: Any = None, name_prefix: str = 'scan', options: Any = None, preamble: str = '', devices: Device | None = None)[source]¶

Performs the same function and has the same interface as GenericScanKernel, but uses a dead-simple, sequential scan. Works best on CPU platforms, and helps isolate bugs in scans by removing the potential for issues originating in parallel execution.

__call__(*args: Any, **kwargs: Any) → Event[source]¶: See GenericScanKernel.__call__().

Simple / Legacy Interface¶

class pyopencl.scan.ExclusiveScanKernel(ctx, dtype, scan_expr, neutral, name_prefix='scan', options=[], preamble='', devices=None)[source]¶

Generates a kernel that can compute a prefix sum using any associative operation given as scan_expr. scan_expr uses the formal values “a” and “b” to indicate two operands of an associative binary operation. neutral is the neutral element of scan_expr, obeying scan_expr(a, neutral) == a.

dtype specifies the type of the arrays being operated on. name_prefix is used for kernel names to ensure recognizability in profiles and logs. options is a list of compiler options to use when building. preamble specifies a string of code that is inserted before the actual kernels. devices may be used to restrict the set of devices on which the kernel is meant to run. (defaults to all devices in the context ctx.

__call__(self, input_ary, output_ary=None, allocator=None, queue=None)[source]¶

class pyopencl.scan.InclusiveScanKernel(ctx, dtype, scan_expr, neutral=None, name_prefix='scan', options=[], preamble='', devices=None)[source]¶: Works like ExclusiveScanKernel.

Changed in version 2013.1: neutral is now always required.

For the array [1, 2, 3], inclusive scan results in [1, 3, 6], and exclusive scan results in [0, 1, 3].

Here’s a usage example:

knl = InclusiveScanKernel(context, np.int32, "a+b")

n = 2**20-2**18+5
rng = np.random.default_rng(seed=42)
host_data = rng.integers(0, 10, size=n, dtype=np.int32)
dev_data = cl_array.to_device(queue, host_data)

knl(dev_data)
assert (dev_data.get() == np.cumsum(host_data, axis=0)).all()

Predicated copies (“partition”, “unique”, …)¶

pyopencl.algorithm.copy_if(ary, predicate, extra_args=None, preamble='', queue=None, wait_for=None)[source]¶

Copy the elements of ary satisfying predicate to an output array.

Parameters:

predicate – a C expression evaluating to a bool, represented as a string. The value to test is available as ary[i], and if the expression evaluates to true, then this value ends up in the output.
extra_args – a list of tuples (name, value) specifying extra arguments to pass to the scan procedure. For version 2013.1, value must be a of a numpy sized scalar type. As of version 2013.2, value may also be a pyopencl.array.Array.
preamble – A snippet of C that is inserted into the compiled kernel before the actual kernel function. May be used for, e.g. type definitions or include statements.
wait_for – wait_for may either be None or a list of pyopencl.Event instances for whose completion this command waits before starting execution.

Returns:

a tuple (out, count, event) where out is the output array, count is an on-device scalar (fetch to host with count.get()) indicating how many elements satisfied predicate, and event is a pyopencl.Event for dependency management. out is allocated to the same length as ary, but only the first count entries carry meaning.

Added in version 2013.1.

pyopencl.algorithm.remove_if(ary, predicate, extra_args=None, preamble='', queue=None, wait_for=None)[source]¶

Copy the elements of ary not satisfying predicate to an output array.

Parameters:

predicate – a C expression evaluating to a bool, represented as a string. The value to test is available as ary[i], and if the expression evaluates to false, then this value ends up in the output.
extra_args – a list of tuples (name, value) specifying extra arguments to pass to the scan procedure. For version 2013.1, value must be a of a numpy sized scalar type. As of version 2013.2, value may also be a pyopencl.array.Array.
preamble – A snippet of C that is inserted into the compiled kernel before the actual kernel function. May be used for, e.g. type definitions or include statements.
wait_for – wait_for may either be None or a list of pyopencl.Event instances for whose completion this command waits before starting execution.

Returns:

a tuple (out, count, event) where out is the output array, count is an on-device scalar (fetch to host with count.get()) indicating how many elements did not satisfy predicate, and event is a pyopencl.Event for dependency management.

Added in version 2013.1.

pyopencl.algorithm.partition(ary, predicate, extra_args=None, preamble='', queue=None, wait_for=None)[source]¶

Copy the elements of ary into one of two arrays depending on whether they satisfy predicate.

Parameters:

predicate – a C expression evaluating to a bool, represented as a string. The value to test is available as ary[i].
extra_args – a list of tuples (name, value) specifying extra arguments to pass to the scan procedure. For version 2013.1, value must be a of a numpy sized scalar type. As of version 2013.2, value may also be a pyopencl.array.Array.
preamble – A snippet of C that is inserted into the compiled kernel before the actual kernel function. May be used for, e.g. type definitions or include statements.
wait_for – wait_for may either be None or a list of pyopencl.Event instances for whose completion this command waits before starting execution.

Returns:

a tuple (out_true, out_false, count, event) where count is an on-device scalar (fetch to host with count.get()) indicating how many elements satisfied the predicate, and event is a pyopencl.Event for dependency management.

Added in version 2013.1.

pyopencl.algorithm.unique(ary, is_equal_expr='a == b', extra_args=None, preamble='', queue=None, wait_for=None)[source]¶

Copy the elements of ary into the output if is_equal_expr, applied to the array element and its predecessor, yields false.

Works like the UNIX command uniq, with a potentially custom comparison. This operation is often used on sorted sequences.

Parameters:

is_equal_expr – a C expression evaluating to a bool, represented as a string. The elements being compared are available as a and b. If this expression yields false, the two are considered distinct.
extra_args – a list of tuples (name, value) specifying extra arguments to pass to the scan procedure. For version 2013.1, value must be a of a numpy sized scalar type. As of version 2013.2, value may also be a pyopencl.array.Array.
preamble – A snippet of C that is inserted into the compiled kernel before the actual kernel function. May be used for, e.g. type definitions or include statements.
wait_for – wait_for may either be None or a list of pyopencl.Event instances for whose completion this command waits before starting execution.

Returns:

a tuple (out, count, event) where out is the output array, count is an on-device scalar (fetch to host with count.get()) indicating how many elements satisfied the predicate, and event is a pyopencl.Event for dependency management.

Added in version 2013.1.

Sorting (radix sort)¶

class pyopencl.algorithm.RadixSort(context, arguments, key_expr, sort_arg_names, bits_at_a_time=2, index_dtype=<class 'numpy.int32'>, key_dtype=<class 'numpy.uint32'>, scan_kernel=<class 'pyopencl.scan.GenericScanKernel'>, options=None)[source]¶

Provides a general radix sort on the compute device.

Building many variable-size lists¶

class pyopencl.algorithm.ListOfListsBuilder(context, list_names_and_dtypes, generate_template, arg_decls, count_sharing=None, devices=None, name_prefix='plb_build_list', options=None, preamble='', debug=False, complex_kernel=False, eliminate_empty_output_lists=False)[source]¶

Generates and executes code to produce a large number of variable-size lists, simply.

Note

This functionality is provided as a preview. Its interface is subject to change until this notice is removed.

Added in version 2013.1.

Here’s a usage example:

from pyopencl.algorithm import ListOfListsBuilder
builder = ListOfListsBuilder(context, [("mylist", np.int32)], """
        void generate(LIST_ARG_DECL USER_ARG_DECL index_type i)
        {
            int count = i % 4;
            for (int j = 0; j < count; ++j)
            {
                APPEND_mylist(count);
            }
        }
        """, arg_decls=[])

result, event = builder(queue, 2000)

inf = result["mylist"]
assert inf.count == 3000
assert (inf.list.get()[-6:] == [1, 2, 2, 3, 3, 3]).all()

The function generate above is called once for each “input object”. Each input object can then generate zero or more list entries. The number of these input objects is given to __call__() as n_objects. List entries are generated by calls to APPEND_<list name>(value). Multiple lists may be generated at once.

__init__(context, list_names_and_dtypes, generate_template, arg_decls, count_sharing=None, devices=None, name_prefix='plb_build_list', options=None, preamble='', debug=False, complex_kernel=False, eliminate_empty_output_lists=False)[source]¶

Parameters:

context – A pyopencl.Context.
list_names_and_dtypes – a list of (name, dtype) tuples indicating the lists to be built.
generate_template – a snippet of C as described below
arg_decls – A string of comma-separated C argument declarations.
count_sharing – A mapping consisting of (child, mother) indicating that mother and child will always have the same number of indices, and the APPEND to mother will always happen before the APPEND to the child.
name_prefix – the name prefix to use for the compiled kernels
options – OpenCL compilation options for kernels using generate_template.
complex_kernel – If True, prevents vectorization on CPUs.
eliminate_empty_output_lists – A Python list of list names for which the empty output lists are eliminated.

generate_template may use the following C macros/identifiers:

index_type: expands to C identifier for the index type used for the calculation
USER_ARG_DECL: expands to the C declarator for arg_decls
USER_ARGS: a list of C argument values corresponding to user_arg_decl
LIST_ARG_DECL: expands to a C argument list representing the data for the output lists. These are escaped prefixed with "plg_" so as to not interfere with user-provided names.
LIST_ARGS: a list of C argument values corresponding to LIST_ARG_DECL
APPEND_name(entry): inserts entry into the list name. entry must be a valid C expression of the correct type.

All argument-list related macros have a trailing comma included if they are non-empty.

generate_template must supply a function:

void generate(USER_ARG_DECL LIST_ARG_DECL index_type i)
{
    APPEND_mylist(5);
}

Internally, the kernel_template is expanded (at least) twice. Once, for a ‘counting’ stage where the size of all the lists is determined, and a second time, for a ‘generation’ stage where the lists are actually filled. A generate function that has side effects beyond calling append is therefore ill-formed.

Changed in version 2018.1: Change eliminate_empty_output_lists argument type from bool to list.

__call__(queue, n_objects, *args, **kwargs)[source]¶

Parameters:

args – arguments corresponding to arg_decls in the constructor. Array-like arguments must be either 1D pyopencl.array.Array objects or pyopencl.MemoryObject objects, of which the latter can be obtained from a pyopencl.array.Array using the pyopencl.array.Array.data attribute.
allocator – optionally, the allocator to use to allocate new arrays.
omit_lists – an iterable of list names that should not be built with this invocation. The kernel code may not call APPEND_name for these omitted lists. If it does, undefined behavior will result. The returned lists dictionary will not contain an entry for names in omit_lists.
wait_for – wait_for may either be None or a list of pyopencl.Event instances for whose completion this command waits before starting execution.

Returns:

a tuple (lists, event), where lists is a mapping from (built) list names to objects which have attributes

count for the total number of entries in all lists combined
lists for the array containing all lists.
starts for the array of starting indices in lists. starts is built so that it has n+1 entries, so that the i’th entry is the start of the i’th list, and the i’th entry is the index one past the i’th list’s end, even for the last list.

This implies that all lists are contiguous.

If the list name is specified in eliminate_empty_output_lists constructor argument, lists has two additional attributes num_nonempty_lists and nonempty_indices

num_nonempty_lists for the number of nonempty lists.
nonempty_indices for the index of nonempty list in input objects.

In this case, starts has num_nonempty_lists + 1 entries. The i’s entry is the start of the i’th nonempty list, which is generated by the object with index nonempty_indices[i].

event is a pyopencl.Event for dependency management.

Changed in version 2016.2: Added omit_lists.

Bitonic Sort¶

class pyopencl.bitonic_sort.BitonicSort(context)[source]¶

Sort an array (or one axis of one) using a sorting network.

Will only work if the axis of the array to be sorted has a length that is a power of 2.

Added in version 2015.2.