From 3f8d844ba099017d93ae17cfb6dc15d0dd0040e5 Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Thu, 4 Aug 2016 22:16:06 -0500
Subject: [PATCH] A first cut of SVM support
---
doc/runtime_memory.rst | 43 +++--
pyopencl/__init__.py | 165 ++++++++++++++++++-
pyopencl/cffi_cl.py | 308 ++++++++++++++++++++++++++++++++---
setup.py | 6 +-
src/c_wrapper/kernel.cpp | 13 ++
src/c_wrapper/wrap_cl_core.h | 1 +
test/test_wrapper.py | 28 +++-
7 files changed, 522 insertions(+), 42 deletions(-)
diff --git a/doc/runtime_memory.rst b/doc/runtime_memory.rst
index 3f85a6ca..3816abaa 100644
--- a/doc/runtime_memory.rst
+++ b/doc/runtime_memory.rst
@@ -33,7 +33,7 @@ OpenCL Runtime: Memory
|comparable|
Memory Migration
-^^^^^^^^^^^^^^^^
+----------------
.. function:: enqueue_migrate_mem_objects(queue, mem_objects, flags=0, wait_for=None)
@@ -52,8 +52,8 @@ Memory Migration
Only available with the `cl_ext_migrate_memobject`
extension.
-Buffers
-^^^^^^^
+Buffer
+------
.. class:: Buffer(context, flags, size=0, hostbuf=None)
@@ -107,24 +107,39 @@ Buffers
.. _svm:
-Shared Virtual Memory
-^^^^^^^^^^^^^^^^^^^^^
+Shared Virtual Memory (SVM)
+---------------------------
-Shared virtual memory requires OpenCL 2.0.
+Shared virtual memory allows the host and the compute device to share
+address space, so that pointers on the host and on the device may have
+the same meaning. In addition, it allows the same memory to be accessed
+by both the host and the device. *Coarse-grain* SVM requires that
+buffers be mapped before being accessed on the host, *fine-grain* SVM
+does away with that requirement.
-.. autoclass:: SVMAllocation
+SVM requires OpenCL 2.0.
.. autoclass:: SVM
.. autoclass:: SVMMap
+Allocating SVM
+^^^^^^^^^^^^^^
+
+.. autofunction:: svm_empty
+.. autofunction:: svm_empty_like
.. autofunction:: csvm_empty
-.. autofunction:: fsvm_empty
.. autofunction:: csvm_empty_like
+.. autofunction:: fsvm_empty
.. autofunction:: fsvm_empty_like
-Image Formats
-^^^^^^^^^^^^^
+SVM Allocation Holder
+^^^^^^^^^^^^^^^^^^^^^
+
+.. autoclass:: SVMAllocation
+
+Image
+-----
.. class:: ImageFormat([channel_order, channel_type])
@@ -169,8 +184,6 @@ Image Formats
See :class:`mem_flags` for possible values of *flags*
and :class:`mem_object_type` for possible values of *image_type*.
-Images
-^^^^^^
.. class:: Image(context, flags, format, shape=None, pitches=None, hostbuf=None, is_array=False, buffer=None):
See :class:`mem_flags` for values of *flags*.
@@ -255,12 +268,12 @@ Images
.. versionadded:: 2011.2
Transfers
-^^^^^^^^^
+---------
.. autofunction:: enqueue_copy(queue, dest, src, **kwargs)
Mapping Memory into Host Address Space
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+--------------------------------------
.. autoclass:: MemoryMap
@@ -309,7 +322,7 @@ Mapping Memory into Host Address Space
Added *strides* argument.
Samplers
-^^^^^^^^
+--------
.. class:: Sampler(context, normalized_coords, addressing_mode, filter_mode)
diff --git a/pyopencl/__init__.py b/pyopencl/__init__.py
index 9100c844..166601f2 100644
--- a/pyopencl/__init__.py
+++ b/pyopencl/__init__.py
@@ -112,6 +112,10 @@ from pyopencl.cffi_cl import ( # noqa
MemoryObject,
MemoryMap,
Buffer,
+ SVMAllocation,
+ SVM,
+ SVMMap,
+
CompilerWarning,
_Program,
Kernel,
@@ -695,7 +699,8 @@ def enqueue_copy(queue, dest, src, **kwargs):
Two types of 'buffer' occur in the arguments to this function,
:class:`Buffer` and 'host-side buffers'. The latter are
defined by Python and commonly called `buffer objects
- <https://docs.python.org/3.4/c-api/buffer.html>`_.
+ <https://docs.python.org/3.4/c-api/buffer.html>`_. :mod:`numpy`
+ arrays are a very common example.
Make sure to always be clear on whether a :class:`Buffer` or a
Python buffer object is needed.
@@ -788,6 +793,15 @@ def enqueue_copy(queue, dest, src, **kwargs):
:arg region: :class:`tuple` of :class:`int` of length
three or shorter. (mandatory)
+ .. ------------------------------------------------------------------------
+ .. rubric :: Transfer :class:`SVM`/host ↔ :class:`SVM`/host
+ .. ------------------------------------------------------------------------
+
+ :arg byte_count: (optional) If not specified, defaults to the
+ size of the source in versions 2012.x and earlier,
+ and to the minimum of the size of the source and target
+ from 2013.1 on.
+
|std-enqueue-blurb|
.. versionadded:: 2011.1
@@ -983,4 +997,153 @@ def enqueue_fill_buffer(queue, mem, pattern, offset, size, wait_for=None):
# }}}
+# {{{ numpy-like svm allocation
+
+def svm_empty(ctx, flags, shape, dtype, order="C", alignment=None):
+ """Allocate an empty :class:`numpy.ndarray` of the given *shape*, *dtype*
+ and *order*. (See :func:`numpy.empty` for the meaning of these arguments.)
+ The array will be allocated in shared virtual memory belonging
+ to *ctx*.
+
+ :arg ctx: a :class:`Context`
+ :arg flags: a combination of flags from :class:`svm_mem_flags`.
+ :arg alignment: the number of bytes to which the beginning of the memory
+ is aligned. Defaults to the :attr:`numpy.dtype.itemsize` of *dtype*.
+
+ :returns: a :class:`numpy.ndarray` whose :attr:`numpy.ndarray.base` attribute
+ is a :class:`SVMAllocation`.
+
+ To pass the resulting array to an OpenCL kernel or :func:`enqueue_copy`, you
+ will likely want to wrap the returned array in an :class:`SVM` tag.
+
+ .. versionadded:: 2016.2
+ """
+
+ dtype = np.dtype(dtype)
+
+ try:
+ s = 1
+ for dim in shape:
+ s *= dim
+ except TypeError:
+ import sys
+ if sys.version_info >= (3,):
+ admissible_types = (int, np.integer)
+ else:
+ admissible_types = (np.integer,) + six.integer_types
+
+ if not isinstance(shape, admissible_types):
+ raise TypeError("shape must either be iterable or "
+ "castable to an integer")
+ s = shape
+ shape = (shape,)
+
+ itemsize = dtype.itemsize
+ nbytes = s * itemsize
+
+ from pyopencl.compyte.array import c_contiguous_strides, f_contiguous_strides
+
+ if order in "fF":
+ strides = f_contiguous_strides(itemsize, shape)
+ elif order in "cC":
+ strides = c_contiguous_strides(itemsize, shape)
+ else:
+ raise ValueError("order not recognized: %s" % order)
+
+ descr = dtype.descr
+ if len(descr) == 1:
+ typestr = descr[0][1]
+ else:
+ typestr = "V%d" % itemsize
+
+ interface = {
+ "version": 3,
+ "shape": shape,
+ "typestr": typestr,
+ "descr": descr,
+ "strides": strides,
+ }
+
+ if alignment is None:
+ alignment = itemsize
+
+ svm_alloc = SVMAllocation(ctx, nbytes, alignment, flags, _interface=interface)
+ return SVM(np.asarray(svm_alloc))
+
+
+def svm_empty_like(ctx, flags, ary, alignment=None):
+ """Allocate an empty :class:`numpy.ndarray` like the existing
+ :class:`numpy.ndarray` *ary*. The array will be allocated in shared
+ virtual memory belonging to *ctx*.
+
+ :arg ctx: a :class:`Context`
+ :arg flags: a combination of flags from :class:`svm_mem_flags`.
+ :arg alignment: the number of bytes to which the beginning of the memory
+ is aligned. Defaults to the :attr:`numpy.dtype.itemsize` of *dtype*.
+
+ :returns: a :class:`numpy.ndarray` whose :attr:`numpy.ndarray.base` attribute
+ is a :class:`SVMAllocation`.
+
+ To pass the resulting array to an OpenCL kernel or :func:`enqueue_copy`, you
+ will likely want to wrap the returned array in an :class:`SVM` tag.
+
+ .. versionadded:: 2016.2
+ """
+ if ary.flags.c_contiguous:
+ order = "C"
+ elif ary.flags.f_contiguous:
+ order = "F"
+ else:
+ raise ValueError("array is neither C- nor Fortran-contiguous")
+
+ return svm_empty(ctx, ary.shape, ary.dtype, order,
+ flags=flags, alignment=alignment)
+
+
+def csvm_empty(ctx, shape, dtype, order="C", alignment=None):
+ """
+ Like :func:`svm_empty`, but with *flags* set for a coarse-grain read-write
+ buffer.
+
+ .. versionadded:: 2016.2
+ """
+ return svm_empty(ctx, svm_mem_flags.READ_WRITE, shape, dtype, order, alignment)
+
+
+def csvm_empty_like(ctx, ary, alignment=None):
+ """
+ Like :func:`svm_empty_like`, but with *flags* set for a coarse-grain
+ read-write buffer.
+
+ .. versionadded:: 2016.2
+ """
+ return svm_empty_like(ctx, svm_mem_flags.READ_WRITE, ary)
+
+
+def fsvm_empty(ctx, shape, dtype, order="C", alignment=None):
+ """
+ Like :func:`svm_empty`, but with *flags* set for a fine-grain read-write
+ buffer.
+
+ .. versionadded:: 2016.2
+ """
+ return svm_empty(ctx,
+ svm_mem_flags.READ_WRITE | svm_mem_flags.SVM_FINE_GRAIN_BUFFER,
+ shape, dtype, order, alignment)
+
+
+def fsvm_empty_like(ctx, ary, alignment=None):
+ """
+ Like :func:`svm_empty_like`, but with *flags* set for a fine-grain
+ read-write buffer.
+
+ .. versionadded:: 2016.2
+ """
+ return svm_empty_like(
+ ctx,
+ svm_mem_flags.READ_WRITE | svm_mem_flags.SVM_FINE_GRAIN_BUFFER,
+ ary)
+
+# }}}
+
# vim: foldmethod=marker
diff --git a/pyopencl/cffi_cl.py b/pyopencl/cffi_cl.py
index b4c104f0..31df8050 100644
--- a/pyopencl/cffi_cl.py
+++ b/pyopencl/cffi_cl.py
@@ -804,7 +804,7 @@ class CommandQueue(_Common):
# }}}
-# {{{ memory objects
+# {{{ _norm_shape_dtype and cffi_array
def _norm_shape_dtype(shape, dtype, order="C", strides=None, name=""):
dtype = np.dtype(dtype)
@@ -839,17 +839,10 @@ class cffi_array(np.ndarray): # noqa
def base(self):
return self.__base
+# }}}
-class LocalMemory(_CLKernelArg):
- __slots__ = ('_size',)
-
- def __init__(self, size):
- self._size = size
-
- @property
- def size(self):
- return self._size
+# {{{ MemoryObjectHolder base class
class MemoryObjectHolder(_Common, _CLKernelArg):
def get_host_array(self, shape, dtype, order="C"):
@@ -867,6 +860,10 @@ class MemoryObjectHolder(_Common, _CLKernelArg):
"MemoryObjectHolder.get_host_array")
return ary
+# }}}
+
+
+# {{{ MemoryObject
class MemoryObject(MemoryObjectHolder):
def __init__(self, hostbuf=None):
@@ -894,8 +891,18 @@ class MemoryObject(MemoryObjectHolder):
def release(self):
_handle_error(_lib.memory_object__release(self.ptr))
+# }}}
+
+
+# {{{ MemoryMap
class MemoryMap(_Common):
+ """"
+ .. automethod:: release
+
+ This class may also be used as a context manager in a ``with`` statement.
+ """
+
@classmethod
def _create(cls, ptr, shape, typestr, strides):
self = _Common._create.__func__(cls, ptr)
@@ -908,6 +915,12 @@ class MemoryMap(_Common):
}
return self
+ def __enter__(self):
+ return self
+
+ def __exit__(self, exc_type, exc_val, exc_tb):
+ self.release()
+
def release(self, queue=None, wait_for=None):
c_wait_for, num_wait_for = _clobj_list(wait_for)
_event = _ffi.new('clobj_t*')
@@ -916,6 +929,8 @@ class MemoryMap(_Common):
c_wait_for, num_wait_for, _event))
return Event._create(_event[0])
+# }}}
+
# {{{ _c_buffer_from_obj
@@ -960,8 +975,6 @@ elif sys.version_info >= (2, 7, 4):
_ssize_t = ctypes.c_size_t
def _c_buffer_from_obj(obj, writable=False, retain=False):
- # {{{ fall back to the old CPython buffer protocol API
-
# {{{ try the numpy array interface first
# avoid slow ctypes-based buffer interface wrapper
@@ -976,6 +989,8 @@ elif sys.version_info >= (2, 7, 4):
# }}}
+ # {{{ fall back to the old CPython buffer protocol API
+
from pyopencl._buffers import Py_buffer, PyBUF_ANY_CONTIGUOUS, PyBUF_WRITABLE
flags = PyBUF_ANY_CONTIGUOUS
@@ -1071,6 +1086,233 @@ class Buffer(MemoryObject):
# }}}
+
+# {{{ SVMAllocation
+
+class SVMAllocation(object):
+ """An object whose lifetime is tied to an allocation of shared virtual memory.
+
+ .. note::
+
+ Most likely, you will not want to use this directly, but rather
+ :func:`svm_empty` and related functions which allow access to this
+ functionality using a friendlier, more Pythonic interface.
+
+ .. versionadded:: 2016.2
+
+ .. automethod:: __init__(self, ctx, size, alignment, flags=None)
+ .. automethod:: release
+ .. automethod:: enqueue_release
+ """
+ def __init__(self, ctx, size, alignment, flags, _interface=None):
+ """
+ :arg ctx: a :class:`Context`
+ :arg flags: some of :class:`svm_mem_flags`.
+ """
+
+ self.ptr = None
+
+ ptr = _ffi.new('void**')
+ _handle_error(_lib.svm_alloc(
+ ctx.ptr, flags, size, alignment,
+ ptr))
+
+ self.ctx = ctx
+ self.ptr = ptr[0]
+ self.is_fine_grain = flags & svm_mem_flags.SVM_FINE_GRAIN_BUFFER
+
+ if _interface is not None:
+ _interface["data"] = (
+ int(_ffi.cast("intptr_t", self.ptr)),
+ flags & mem_flags.WRITE_ONLY != 0
+ or flags & mem_flags.READ_WRITE != 0)
+ self.__array_interface__ = _interface
+
+ def __del__(self):
+ if self.ptr is not None:
+ self.release()
+
+ def release(self):
+ _handle_error(_lib.svm_free(self.ctx.ptr, self.ptr))
+ self.ptr = None
+
+ def enqueue_release(self, queue, wait_for=None):
+ """
+ :arg flags: a combination of :class:`pyopencl.map_flags`
+ :returns: a :class:`pyopencl.Event`
+
+ |std-enqueue-blurb|
+ """
+ ptr_event = _ffi.new('clobj_t*')
+ c_wait_for, num_wait_for = _clobj_list(wait_for)
+ _handle_error(_lib.enqueue_svm_free(
+ ptr_event, queue.ptr, 1, self.ptr,
+ c_wait_for, num_wait_for))
+
+ self.ctx = None
+ self.ptr = None
+
+ return Event._create(ptr_event[0])
+
+# }}}
+
+
+# {{{ SVM
+
+#TODO:
+# doc example
+# finish copy
+# test
+# fill
+# test
+# migrate
+
+class SVM(_CLKernelArg):
+ """Tags an object exhibiting the Python buffer interface (such as a
+ :class:`numpy.ndarray`) as referring to shared virtual memory.
+
+ Depending on the features of the OpenCL implementation, the following
+ types of objects may be passed to/wrapped in this type:
+
+ * coarse-grain shared memory as returned by (e.g.) :func:`csvm_empty`
+ for any implementation of OpenCL 2.0.
+
+ * fine-grain shared memory as returned by (e.g.) :func:`fsvm_empty`,
+ if the implementation supports fine-grained shared virtual memory.
+
+ * any :class:`numpy.ndarray` (or other Python object with a buffer
+ interface) if the implementation supports fine-grained *system* shared
+ virtual memory.
+
+ Objects of this type may be passed to kernel calls and :func:`enqueue_copy`.
+ Coarse-grain shared-memory *must* be mapped into host address space using
+ :meth:`map` before being accessed through the :mod:`numpy` interface.
+
+ .. note::
+
+ This object merely serves as a 'tag' that changes the meaning
+ of functions to which it is passed. It has no special management
+ relationship to the memory it tags. For example, it is permissible
+ to grab a :mod:`numpy.array` out of :attr:`SVM.memory` of one
+ :class:`SVM` instance and use the array to construct another.
+ Neither of the tags needs to be kept alive.
+
+ .. versionadded:: 2016.2
+
+ .. attribute:: mem
+
+ The wrapped object.
+
+ .. automethod:: __init__
+ .. automethod:: map
+ .. automethod:: as_buffer
+ """
+
+ def __init__(self, mem):
+ self.mem = mem
+
+ def map(self, queue, is_blocking=True, flags=None, wait_for=None):
+ """
+ :arg is_blocking: If *False*, subsequent code must wait on
+ :attr:`SVMMap.event` in the returned object before accessing the
+ mapped memory.
+ :arg flags: a combination of :class:`pyopencl.map_flags`, defaults to
+ read-write.
+ :returns: an :class:`SVMMap` instance
+
+ |std-enqueue-blurb|
+ """
+ if flags is None:
+ flags = map_flags.READ | map_flags.WRITE
+
+ c_buf, size, _ = _c_buffer_from_obj(self.mem, writable=bool(
+ flags & (map_flags.WRITE | map_flags.INVALIDATE_REGION)))
+
+ ptr_event = _ffi.new('clobj_t*')
+ c_wait_for, num_wait_for = _clobj_list(wait_for)
+ _handle_error(_lib.enqueue_svm_map(
+ ptr_event, queue.ptr, is_blocking, flags,
+ c_buf, size,
+ c_wait_for, num_wait_for))
+
+ evt = Event._create(ptr_event[0]), SVMMap(self.mem)
+ return SVMMap(self, queue, evt)
+
+ def _enqueue_unmap(self, queue, wait_for=None):
+ c_buf, _, _ = _c_buffer_from_obj(self.mem)
+
+ ptr_event = _ffi.new('clobj_t*')
+ c_wait_for, num_wait_for = _clobj_list(wait_for)
+ _handle_error(_lib.enqueue_svm_unmap(
+ ptr_event, queue.ptr,
+ c_buf,
+ c_wait_for, num_wait_for))
+
+ return Event._create(ptr_event[0]), SVMMap(self.mem)
+
+ def as_buffer(self, ctx, flags=None):
+ """
+ :arg ctx: a :class:`Context`
+ :arg flags: a combination of :class:`pyopencl.map_flags`, defaults to
+ read-write.
+ :returns: a :class:`Buffer` corresponding to *self*.
+
+ The memory referred to by this object must not be freed before
+ the returned :class:`Buffer` is released.
+ """
+
+ if flags is None:
+ flags = mem_flags.READ_WRITE
+
+ return Buffer(ctx, flags, size=self.mem.nbytes, hostbuf=self.mem)
+
+# }}}
+
+
+# {{{ SVMMap
+
+class SVMMap(_CLKernelArg):
+ """
+ .. attribute:: event
+
+ .. versionadded:: 2016.2
+
+ .. automethod:: release
+
+ This class may also be used as a context manager in a ``with`` statement.
+ :meth:`release` will be called upon exit from the ``with`` region.
+ The value returned to the ``as`` part of the context manager is the
+ mapped Python object (e.g. a :mod:`numpy` array).
+ """
+ def __init__(self, svm, queue, event):
+ self.svm = svm
+ self.queue = queue
+ self.event = event
+
+ def __del__(self):
+ if self.svm is not None:
+ self.release()
+
+ def __enter__(self):
+ return self.svm.mem
+
+ def __exit__(self, exc_type, exc_val, exc_tb):
+ self.release()
+
+ def release(self, queue=None, wait_for=None):
+ """
+ :arg queue: a :class:`pyopencl.CommandQueue`. Defaults to the one
+ with which the map was created, if not specified.
+ :returns: a :class:`pyopencl.Event`
+
+ |std-enqueue-blurb|
+ """
+
+ evt = self.svm._enqueue_unmap(self.queue)
+ self.svm = None
+
+ return evt
+
# }}}
@@ -1246,6 +1488,17 @@ class _Program(_Common):
# }}}
+class LocalMemory(_CLKernelArg):
+ __slots__ = ('_size',)
+
+ def __init__(self, size):
+ self._size = size
+
+ @property
+ def size(self):
+ return self._size
+
+
# {{{ Kernel
# {{{ arg packing helpers
@@ -1407,8 +1660,8 @@ class Kernel(_Common):
status = _lib.kernel__set_arg_null(self.ptr, {arg_idx})
if status != _ffi.NULL:
_handle_error(status)
- elif isinstance({arg_var}, _CLKernelArg):
- self.set_arg({arg_idx}, {arg_var})
+ else:
+ self._set_arg_clkernelarg({arg_idx}, {arg_var})
"""
.format(arg_idx=arg_idx, arg_var=arg_var))
@@ -1588,19 +1841,30 @@ class Kernel(_Common):
capture_kernel_call(self, filename, queue, global_size, local_size,
*args, **kwargs)
+ def _set_arg_clkernelarg(self, arg_index, arg):
+ if isinstance(arg, MemoryObjectHolder):
+ _handle_error(_lib.kernel__set_arg_mem(self.ptr, arg_index, arg.ptr))
+ elif isinstance(arg, SVM):
+ c_buf, _, _ = _c_buffer_from_obj(arg.mem)
+ _handle_error(_lib.kernel__set_arg_svm_pointer(
+ self.ptr, arg_index, c_buf))
+ elif isinstance(arg, Sampler):
+ _handle_error(_lib.kernel__set_arg_sampler(self.ptr, arg_index,
+ arg.ptr))
+ elif isinstance(arg, LocalMemory):
+ _handle_error(_lib.kernel__set_arg_buf(self.ptr, arg_index,
+ _ffi.NULL, arg.size))
+ else:
+ raise RuntimeError("unexpected _CLKernelArg subclass"
+ "dimensions", status_code.INVALID_VALUE,
+ "clSetKernelArg")
+
def set_arg(self, arg_index, arg):
# If you change this, also change the kernel call generation logic.
if arg is None:
_handle_error(_lib.kernel__set_arg_null(self.ptr, arg_index))
elif isinstance(arg, _CLKernelArg):
- if isinstance(arg, MemoryObjectHolder):
- _handle_error(_lib.kernel__set_arg_mem(self.ptr, arg_index, arg.ptr))
- elif isinstance(arg, Sampler):
- _handle_error(_lib.kernel__set_arg_sampler(self.ptr, arg_index,
- arg.ptr))
- elif isinstance(arg, LocalMemory):
- _handle_error(_lib.kernel__set_arg_buf(self.ptr, arg_index,
- _ffi.NULL, arg.size))
+ self._set_arg_clkernelarg(self, arg_index, arg)
elif _CPY2 and isinstance(arg, np.generic):
# https://github.com/numpy/numpy/issues/5381
c_buf, size, _ = _c_buffer_from_obj(np.getbuffer(arg))
diff --git a/setup.py b/setup.py
index ba00bb94..1d735904 100644
--- a/setup.py
+++ b/setup.py
@@ -72,11 +72,11 @@ def get_config_schema():
return ConfigSchema([
Switch("CL_TRACE", False, "Enable OpenCL API tracing"),
Switch("CL_ENABLE_GL", False, "Enable OpenCL<->OpenGL interoperability"),
- Switch("CL_USE_SHIPPED_EXT", True,
+ Switch(
+ "CL_USE_SHIPPED_EXT", True,
"Use the pyopencl version of CL/cl_ext.h which includes" +
" a broader range of vendor-specific OpenCL extension attributes" +
- " than the standard Khronos (or vendor specific) CL/cl_ext.h."
- ),
+ " than the standard Khronos (or vendor specific) CL/cl_ext.h."),
Option("CL_PRETEND_VERSION", None,
"Dotted CL version (e.g. 1.2) which you'd like to use."),
diff --git a/src/c_wrapper/kernel.cpp b/src/c_wrapper/kernel.cpp
index e787ff8e..817e1061 100644
--- a/src/c_wrapper/kernel.cpp
+++ b/src/c_wrapper/kernel.cpp
@@ -144,6 +144,19 @@ kernel__set_arg_buf(clobj_t _knl, cl_uint arg_index,
});
}
+error*
+kernel__set_arg_svm_pointer(clobj_t _knl, cl_uint arg_index, void *value)
+{
+#if PYOPENCL_CL_VERSION >= 0x2000
+ auto knl = static_cast<kernel*>(_knl);
+ return c_handle_error([&] {
+ pyopencl_call_guarded(clSetKernelArgSVMPointer, knl, arg_index, value);
+ });
+#else
+ PYOPENCL_UNSUPPORTED_BEFORE(clSetKernelArgSVMPointer, "CL 2.0")
+#endif
+}
+
error*
kernel__get_work_group_info(clobj_t _knl, cl_kernel_work_group_info param,
clobj_t _dev, generic_info *out)
diff --git a/src/c_wrapper/wrap_cl_core.h b/src/c_wrapper/wrap_cl_core.h
index 8ef6f2e8..7a4992f4 100644
--- a/src/c_wrapper/wrap_cl_core.h
+++ b/src/c_wrapper/wrap_cl_core.h
@@ -211,6 +211,7 @@ error *kernel__set_arg_sampler(clobj_t kernel, cl_uint arg_index,
clobj_t sampler);
error *kernel__set_arg_buf(clobj_t kernel, cl_uint arg_index,
const void *buffer, size_t size);
+error *kernel__set_arg_svm_pointer(clobj_t kernel, cl_uint arg_index, void *value);
error *kernel__get_work_group_info(clobj_t kernel,
cl_kernel_work_group_info param,
clobj_t device, generic_info *out);
diff --git a/test/test_wrapper.py b/test/test_wrapper.py
index 903da758..43624afe 100644
--- a/test/test_wrapper.py
+++ b/test/test_wrapper.py
@@ -914,7 +914,7 @@ def test_spirv(ctx_factory):
if (ctx._get_cl_version() < (2, 1) or
cl.get_cl_header_version() < (2, 1)):
from pytest import skip
- skip("SPIR-V program creation only available in OpenCL 2.1")
+ skip("SPIR-V program creation only available in OpenCL 2.1 and higher")
n = 50000
@@ -932,6 +932,32 @@ def test_spirv(ctx_factory):
assert la.norm((dest_dev - (a_dev+b_dev)).get()) < 1e-7
+def test_coarse_grain_svm(ctx_factory):
+ ctx = ctx_factory()
+ # queue = cl.CommandQueue(ctx)
+
+ if (ctx._get_cl_version() < (2, 0) or
+ cl.get_cl_header_version() < (2, 0)):
+ from pytest import skip
+ skip("SVM only available in OpenCL 2.0 and higher")
+
+ svm_ary = cl.csvm_empty(ctx, (100, 100), np.float32, alignment=64)
+ assert isinstance(svm_ary.base, cl.SVMAllocation)
+
+
+def test_fine_grain_svm(ctx_factory):
+ ctx = ctx_factory()
+ # queue = cl.CommandQueue(ctx)
+
+ if (ctx._get_cl_version() < (2, 0) or
+ cl.get_cl_header_version() < (2, 0)):
+ from pytest import skip
+ skip("SVM only available in OpenCL 2.0 and higher")
+
+ svm_ary = cl.fsvm_empty(ctx, (100, 100), np.float32, alignment=64)
+ assert isinstance(svm_ary.base, cl.SVMAllocation)
+
+
if __name__ == "__main__":
# make sure that import failures get reported, instead of skipping the tests.
import pyopencl # noqa
--
GitLab