Evan Harvey

with [kokkos@75feb42](https://github.com/kokkos/kokkos/tree/75feb4266108da5800e6f3b0c83cfdc4fedffe97): ``` $ git diff diff --git a/core/unit_test/TestReducers_d.hpp b/core/unit_test/TestReducers_d.hpp index 67f30e6..dc945f6 100644 --- a/core/unit_test/TestReducers_d.hpp +++ b/core/unit_test/TestReducers_d.hpp @@ -81,18 +81,33 @@ TEST(TEST_CATEGORY, reducers_half_t) { // TODO: File a bug report...

After further investigation, we narrowed down the reproducer to: ``` float t = 257; std::cerr

I see that gap in the bhalf integer value range now. I will update the test.

> actually its 8 bits so 256+1 is iffy. Just weird that it behaves differently in different cuda versions. We fall back to float for bhalf_t prior to cuda 11.

For bhalf, we start overflowing integer values at 2^8. after 2^8, we lose representation of odd numbers; after 2^9, we lose representation of odd and even numbers in position 1....

CI checks in #1405 show this too.

> Thanks, Evan! Is the idea that this API is preferred over just transposing the original CSR? No problem. This PR is just a code cleanup in csc2csr right now....

> > Thanks, Evan! Is the idea that this API is preferred over just transposing the original CSR? > > No problem. This PR is just a code cleanup in...

OK, thank you.

> The data structure will be identical, but we have to change `row_map` -> `column_map`, and swap the meanings of `numRows()`/`numCols()` (e.g. column_map will be length numCols+1 instead of numRows+1,...