Daming Feng

AMD Houston

Results 13 comments of


                                            Daming Feng

Numerical errors found in convolution backward data kernel using detetctron2 data

@bartekxk Thanks for the update. For 1, how much the slowdown it could cause? For 2, is there a default value of the splitK that can handle most of the...

[Don't merge] Refactor grouped convolution

> > May I ask when CK will deprecate the support of 2d convolution instances? This has a significant impact on MIOpen side because there are several solvers in MIOpen...

Grouped convolution forward missed the instances of datatype fp32 and int8 for layout (NHWGC, GKYXC, NHWGK)

> @iq136boy Could you specify which DeviceOp exactly? @zjing14, the one using in the client example: DeviceGroupedConvFwdMultipleD

[CK] Move computations of problem-invariant kernel arguments from Invoker to the Solver

@atamazov Thanks for the comments. This issue will be addressed.

2D and 3D NGCHW bacward weight convolution solvers based on CK instances

@averinevg thanks taking time for the code review. I will address them and post the update.

2D and 3D NGCHW bacward weight convolution solvers based on CK instances

close and will reopen if needed.

2D and 3D NGCHW forward convolution solvers based on CK instances

Close and will reopen if needed.

Impl adaptive avgpool

error log: [3301.txt](https://github.com/user-attachments/files/17300533/3301.txt)

Impl adaptive avgpool

CI passed. But the PR still need to solve conflict before it is getting merged.

Impl Kthvalue operation

The conflicts should be solved before merging of this PR.

1
2
›