dpnp.divide() doesn't work properly with a scalar #1295
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
…s are not NaNs or +-Inf
antonwolfy
force-pushed
the
div_by_scalar
branch
from
9b4e2d3 to
28ae93d
Compare
…to div_by_scalar
antonwolfy
force-pushed
the
div_by_scalar
branch
from
28ae93d to
a0a91bd
Compare
|
|
Thanks for the comment. New commit with the fix is pushed. |
antonwolfy
force-pushed
the
div_by_scalar
branch
from
098a57c to
9c3010c
Compare
antonwolfy
added a commit
that referenced
this pull request
Feb 23, 2023
* Complete support of python 3.10 in external CI (#1269) * Set minimum required versions & fix debug building (#1270) * Set minimum required versions & fix debug building * Fix typo * Add support of NumPy 1.24 (#1276) * Set minimum required versions & fix debug building * Add support of numpy 1.24 * Get rid of 'has_aspect_host' property in tests (#1274) * Set minimum required versions & fix debug building * Get rid of 'has_aspect_host' property in tests * Update tests/test_sycl_queue.py Co-authored-by: Oleksandr Pavlyk <[email protected]> Co-authored-by: Oleksandr Pavlyk <[email protected]> * Add support of dpnp.less_equal() (#1275) * Set minimum required versions & fix debug building * Add support of dpnp.less_equal() * Test no broadcast together with input shapes * Add support of comparison operations (#1278) * Use eye() function from dpctl.tensor. (#1271) * Use eye() function from dpctl.tensor. * Add missed order in test for eye() function. * Updated copyright year. Added parameter like for eye() function. * Removed input argumet additional kwards for eye() function. * Get rid of unsupported types in array creation tests (#1283) * Add support of logical comparison operations (#1280) * Add device and sycl_queue keyword arguments to random calls (#1277) * Set minimum required versions & fix debug building * Add device and sycl_queue keyword arguments to random calls * Add device and sycl_queue to dpnp.random.seed() & use random values if seed is None * Update dpnp/random/dpnp_iface_random.py Co-authored-by: Oleksandr Pavlyk <[email protected]> --------- Co-authored-by: Oleksandr Pavlyk <[email protected]> * add __repr__ * add __str__ * reviewer's comments * Fixed gh-1272 (#1287) * Support high=None in dpnp.randint() (#1284) * linter changes applied * Add operation __index__ and __complex__ (#1285) * Add operation __index__ and __complex__ * Add tests * Update tests with proper call of dpctl.SyclQueue() (#1290) * Update minimum required versions of dependent components. (#1289) * dpnp.add() doesn't work properly with a scalar (#1288) * dpnp.add() doesn't work properly with a scalar * get rid of dpctl.SyclQueue() call in tests with unsupported device keyword * Add a fix for crash on CPU device * USM type in operations with a scalar * Porting fix for crash to logic kernel * Add dlpack support with tests and docstrings * Add a test for dlpack with dpt * Fix remarks, add _create_from_usm_ndarray func and move tests to test_sycl_queue * Use tril() and triu() function from dpctl.tensor (#1286) * Use tril() function from dpctl.tensor * Use triu() function from dpctl.tensor * Changed tests for tril() and triu() functions. * Skip tests for tril() and triu() functions with usm_type. * dpnp.subtract() doesn't work properly with a scalar (#1292) * dpnp.add() doesn't work properly with a scalar * dpnp.subtract() doesn't work properly with a scalar * USM type in operations with a scalar * Rollback excluded 'floor_divide' tests from skip scope * Explicit vector operations instead of saturation functions * Use std::int32_t and std::int64_t types * Tune tail's loop of kernel for the vector op * dpnp.divide() doesn't work properly with a scalar (#1295) * dpnp.add() doesn't work properly with a scalar * dpnp.subtract() doesn't work properly with a scalar * dpnp.divide() doesn't work properly with a scalar * dpnp.divide() doesn't work properly with a scalar * Use std::int32_t and std::int64_t types * Disable floating-point optimizations that assume arguments and results are not NaNs or +-Inf * Fix issue with divide on Iris Xe * Updae example3 building from debug build script (#1298) * Remove temporary solution accepting CFD with equal SYCL context instead of queue (#1303) * Intel LLVM is to use conda's gcc toolchain, sysroot and target libraries (#1306) * Tests are crashing if no default device (#1311) * Setting version to 0.11.1 (#1308) --------- Co-authored-by: Oleksandr Pavlyk <[email protected]> Co-authored-by: Natalia Polina <[email protected]> Co-authored-by: Vahid Tavanashad <[email protected]> Co-authored-by: Vladislav Perevezentsev <[email protected]>
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.
Since the change implemented in #1201 and until now, implicit
operation /of array with a scalar or explicit one throughdpnp.divideanddpnp.true_divideleads to raising of the not-implemented exception in dpnp. This is because dpnp was falling back on NumPy implementation in that use case, but has to call an internal kernel function from the backend.The issue was reported in #1273.
The PR proposes to copy a scalar data from host memory into USM allocated memory, which means to create zero-dimension dpnp array from the scalar. This way the division of array with a scalar will be handled in the same way as division of two arrays, which already works fine in dpnp.
Note that if some input argument is a scalar and another one is an array, memory for the scalar will be allocated on the same SYCL device and with the same USM type as the array has. It's required to proceed with further computing and to be compliant with compute follows data paradigm (all input data has to reside on the same device).
The default type of result array will depend on device capabilities, that is dpnp will return the result of type either
float32orcomplex64if nofp64support provided be the device.The compiler flag
fno-finite-math-onlyis included now to explicitly disable a floating-point optimizations that assume arguments and results are not NaNs or +/-Inf, since it is not the case indpnp.