IntelPython / dpctl / 7979394689 / 1

86%

master: 86%

push

github

Changes `py_dot` dispatching for boolean data (#1553)

* Changes py_dot dispatching for boolean data

py_dot for boolean inputs now gives boolean results, which means boolean input arrays will no longer be copied and cast to uint8, improving performance

* Added logical_or to need_workaround

* Refactors reductions.hpp

Adds functions for submitting reductions which handle the choice of using sycl::reduce_over_group or custom_reduce_over_group internally

* Added meta-template arg to submit_(no_)atomic_reduction functions

The last template parameter is a templated class that takes 5
template parameters. This class, instantiated with types, this class
serves as a KernelName for the submitted functor.

The invocation sites were modified to provide such a class as
reduction_*._krn.

The custom_reduction_*_krn class was removed, in favor of using
custom_reduction_wrapper. The generated kernel name, in case
custom reduction functor is called, is custom_reduction_wrapper<KN>,
where KN would be the kernel name for Functor using built-in
sycl::reduce_over_group function.

* Used submit_no_atomic_reduction wrapper throughout gemm.hpp

* Refactors dot_product.hpp to permit using custom_reduce_over_group

* Pass indexers as const, or constexpr as appropriate

Functor constructors take const references for indexers, and
store them with const qualifiers.

* Fix assertions in dot_product.hpp and reductions.hpp

Assertions were asserting for reduction_groups rather than final_reduction_groups. Now final_reduction_groups has been removed.

Also removes unnecessary scope creation during middle portion of tree reductions

* Make indexers const and constexpr, functor constructors take const &

Made indexer instances `const`, or `constexpr` as appropriate.
Functors store indexers as const members, and constructors take
const reference.

Modularized repeated code to compute work-group size into an inline
function in detail n... (continued)

2553 of 2843 branches covered (89.8%)

Branch coverage included in aggregate %.

7764 of 8562 relevant lines covered (90.68%)

4556.18 hits per line