trmm() refactoring #39
Conversation
mkatliar
force-pushed
the
trmm
branch
3 times, most recently
from
682b334 to
9b58ce3
Compare
| auto au = ~a; | ||
|
|
||
| if constexpr (SO == columnMajor) | ||
| { | ||
| #pragma unroll | ||
| for (size_t k = 0; k < N; ++k) | ||
| for (size_t k = 0; k < N; ++k) if (k < n) |
There was a problem hiding this comment.
This is intentional. If is my way to write an unrolled loop from 0 to n.
| for (size_t i = 0; i < A.rows(); ++i) | ||
| for (size_t j = i + 1; j < A.columns(); ++j) | ||
| blaze::reset(A(i, j)); | ||
| reference::trmm(alpha, B, A, UpLo::Lower, false, C); | ||
|
|
||
| // TODO: should be strictly equal? |
There was a problem hiding this comment.
You can always use integer calculations, if you want strict equality (floating point arithmetic on integers is accurate)
I did not want to go that way because I wanted to keep the API close to BLAS. Also, encoding the "upper-lower triangular" and "unit diagonal" properties in the matrix type would complicate the type system, and I don't see a big advantage of it. This is what Blaze does, and if you look at Blaze you see how much it complicates the implementation. |
No more
trmmLeftUpper()andtrmmRightLowe(), justtrmm(). There are overloads for left and right multiplication differing by the argument order, and there is theuploargument specifying whether the upper or lower-triangular part should be used.