IntelPython / dpctl / 7620610598 / 1

86%

master: 86%

push

github

Implementation of `matmul`, `tensordot`, and `vecdot` per array API (#1490)

* Adds ThreeOffsets_CombinedIndexer

This enables strided data processing by gemm kernels

* Remove unused `elementwise_functions.cpp`

* Implements `matmul`, `vecdot`, and `tensordot`

These three functions are implemented through a common `py_dot` binding,
which is also part of a new tensor submodule `_tensor_linalg_impl`

* Tweaks to `matmul` and gemm kernels

Fixes a missing indexer in gemm functor with threading along `nm` dimensions

Fixes `matmul` broadcasting, which was broadcasting in some unnecessary cases

* Remove double-counting of batch offset in gemm batch tree reduction

* Fixes missing dependency in vecdot

When the first argument would not be cast and the second argument would be, the copy dependency was not appended to the list of dependencies, creating race conditions

* Run test_matmul_simple2 in Windows before full test suite

Part of triaging crashes on Windows

* Test removing test_matmul_simple leaving only test_matmul_simple2

* Fix incorrect comments throughtout gemm kernels

Comments incorrectly stated that the third argument to `scale_gemm_k_parameters` is modified by reference

* Drastically reduced parameters used for gemm kernels which thread over k

Experimental change to see if this stabilizes CI

* Test removal of k-threading gemm kernel which writes to multiple outputs atomically

* Refactors `gemm_tree_impl`

Now uses two smaller functions, `gemm_tree_k_impl` and
`gemm_tree_nm_impl` for greater readability

* Reverse order of numeric types passed to test_matmul_simple2

May improve stability on CPU

* Refactors `gemm_contig_tree_impl`

`gemm_contig_tree_impl` now calls new functions `gemm_contig_tree_k_impl` and
`gemm_contig_tree_nm_impl`

* Refactoring `gemm_batch_tree` functions

Adds new functions for calling `nm` threading and `k` threading
kernels to improve readability

* Test... (continued)

2523 of 2833 branches covered (0.0%)

Branch coverage included in aggregate %.

7673 of 8496 relevant lines covered (90.31%)

4575.01 hits per line