Alex Eremin
Alex Eremin
I guess it can be closed, since #2984 has been merged.
Seems related to #2250
We talked one more time (with @atamazov and @DrizztDoUrden) and decided to remove the status from `ConvSolution` entirely and either return correctly initialized structure or throw an exception. But it's...
All of the CK, MLIR and GEMM based kernels do `return {};` if CK\MLIR\GEMM disabled and must do `return {miopenStatusNotInitialized}`. Current design is error-prone and must be fixed.
@ppanchad-amd no. It is still there.
1us for hipMemset of `4 * 4 * 1115 * 1115 * sizeof(float)` ~ 76MB seems strange. gfx1030 has 476GB/s global memory rate and around 1600GB/s L3 rate. Now lets...
There are also a `hipMemset2D` and `hipMemset3D` functions which deal with strides (I hope in an optimal way) Besides this particular PR - it seems we don't often use async...
@pxl-th can you check that the latest MIOpen develop works fine for you? You have to remove performance database from the previous runs or test it on a clean system....
> Recommend to add to the list of investigation: > > why skipping a test is taking so long? It _does not_ make sense to spend 3 seconds just to...
It seems a part of this PR included into https://github.com/ROCm/MIOpen/pull/3146 May I ask you to collaborate with @littlecutebird and fix all the common parts of both PRs?