composable_kernel
composable_kernel copied to clipboard
ROCmSoftwarePlatform
[Don't merge] Refactor grouped convolution
- Remove 1d and 2d grouped convolution instances
- Move MakeDescriptor.... functions to transform headers
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 based on 2d ck instances and APIs. They are in the rocm6.2 mainline staging and (at lease some of them) are now being in tuning process or has done the tuning.
Hi @iq136boy , the reason is to reduce number of instances. 3D instances can handle 1d and 2d cases (for example, for 2d user can set Z=1, Di=1, Do=1, d_padding=0, d_stride=1, d_dilation=1). Of course, if MIOpen still uses 2d instances this change could be suspended until MIOpen will update instances to 3d.
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 based on 2d ck instances and APIs. They are in the rocm6.2 mainline staging and (at lease some of them) are now being in tuning process or has done the tuning.
Hi @iq136boy , the reason is to reduce number of instances. 3D instances can handle 1d and 2d cases (for example, for 2d user can set Z=1, Di=1, Do=1, d_padding=0, d_stride=1, d_dilation=1). Of course, if MIOpen still uses 2d instances this change could be suspended until MIOpen will update instances to 3d.
Yes. We do need this changes to be suspended until we updated the MIOpen solvers to using only 3d ck instances.
