Feature/memory zero copy #152
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
oleksandr-pavlyk
force-pushed
the
feature/memory-zero-copy
branch
2 times, most recently
from
9eaf8a5 to
d6e6483
Compare
Merged
oleksandr-pavlyk
force-pushed
the
feature/memory-zero-copy
branch
from
c261e9a to
bd9f1f1
Compare
|
By the way, reconstruction of |
oleksandr-pavlyk
force-pushed
the
feature/memory-zero-copy
branch
2 times, most recently
from
75690ba to
bc861ab
Compare
|
Implement support for aligned memory allocation, exposing Changed constructor of Non-positive alignment values are ignored and |
oleksandr-pavlyk
force-pushed
the
feature/memory-zero-copy
branch
2 times, most recently
from
2c07fab to
0a8da06
Compare
1. Memory class exposes the interface
2. Memory variants constructors can consum objects exposing
the said interface, and take over memory zero copy.
3. Class implements
memInst.copy_to_host(pyobj=None)
If `pyobj` supports Python's buffer protocol,
content of USM memory in the instance is copied
to the host buffer.
Otherwise, bytearray is allocated, populated
and returned
memInst.copy_from_host(pyobj)
Copies buffer of `pyobj` into USM memory of the
instance.
Raises exception if pyobj is not a byte array
memInst.copy_from_device(sycl_usm_obj)
Copies USM memory of sycl_usm_obj exposing
__sycl_usm_array_interface__ into USM memory
of the instance
4. Class is pickleable
5. Class implements tobytes method that produces bytes object
populated by the content of USM memory.
Methods are currently not releasing GIL, but I think they should.
modularized test, + changes per black
Previously it would always produced shared memory on unpickling.
queue can no longer be specified via positional argument, only through a keyword to allow a user to specify alignment but not queue. SYCL spec says that aligned allocation may return null pointer when the requested alignment is not supported by the device. Non-positive alignments silently go unused (i.e. DPPLmalloc_* is instead of DPPL_aligned_alloc_*)
Added DPPLDevice_AreEq to check if two devices are pointer equal. Used in the test.
oleksandr-pavlyk
force-pushed
the
feature/memory-zero-copy
branch
from
0a8da06 to
cebc064
Compare
1. In that case we can avoid making change of reference objects very long. ``` In [1]: import dpctl, dpctl.memory as dpmem In [2]: m = dpmem.MemoryUSMShared(256) In [3]: m2 = dpmem.MemoryUSMShared(m) In [4]: m3 = dpmem.MemoryUSMShared(m2) In [5]: m3.reference_obj is m Out[5]: True In [6]: m2.reference_obj is m Out[6]: True In [7]: m2._pointer Out[7]: 94798596370432 In [8]: m3._pointer Out[8]: 94798596370432 In [9]: m._pointer Out[9]: 94798596370432 ```
1. Removed dpctl.memory 2. Exposed MemoryUSMShared, MemoryUSMDevice, MemoryUSMHost to dpctl 3. When dpctl is cimported MemoryUSMShared, MemoryUSMHost, MemoryUSMDevice and Memory classes are exposed.
Also simplified queue construction from context and device per PR feedback
…r user consumption
Now to access memory objects one does import dpctl import dpctl.memory
|
@diptorupd The refactoring is complete. Lapse in logic on L109 of |
+ formatting changes, + EXPECT_NO_FATAL_FAILURE on prefetch call.
diptorupd
approved these changes
Oct 27, 2020
|
Windows build failure in CI is not related to changes in this PR at all - test environment fails to activate. I was able to build the package on Windows manually, and test suite run was clean. |
oleksandr-pavlyk
force-pushed
the
feature/memory-zero-copy
branch
from
0a66066 to
948a67b
Compare
oleksandr-pavlyk
force-pushed
the
feature/memory-zero-copy
branch
from
431dc12 to
6d7375e
Compare
Closed
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Closes #43, closes #76, closes #45, closes #120.
Work on
Memoryclass and its derivatives.Classes
MemoryUSMShared,MemoryUSMDevice, andMemoryUSMHostconstructors allow to specify the number of bytes to allocate, and optionally thequeue, or give a Python object expected to implement__sycl_usm_array_interface__attribute to create the USM memory object pointing the same USM memory with zero copy.Classes also implement several new features.
copy_from_host(pyobj)to copy content of Python object buffer to USM memorycopy_to_hist(pyobj=None)to copy content of USM memory to pre-allocated Python object buffer. If object does not implement the buffer protocol,bytearrayof the appropriate size is allocated and populated. The method returns the populated object.copy_from_device(sycl_usm_obj)to copy content of USM memory behind Python object that is expected to implement__sycl_usm_array_interface__to the USM memory of the instance.Classes are now pickleable, with it being implemented via copying to host and picking that.
@Alexander-Makaryev