Fix ConvCkIgemmFwdV6r1DlopsNchw solver to reflect that its not dynamic #2325
Conversation
|
@junliume Please review the urgency label. |
|
@cderb Can you please check if this solver is marked Dynamic in Tuna |
Marked as non-dynamic. |
|
@carlushuang mentioned that this solver might indeed be dynamic |
|
This solver will launch 2 kernels. 1st kernel calculate all the descriptors, and store into a workspace. second kernel will have 4 input, A/B/C pointer, and pointer to workspace. So if view it as a whole, this is actually dynamic kernel |
|
@carlushuang Thanks for the clarification, the compilation of the first kernel for each and every new conv problem means that the compiler will launch and we would have to wait a few seconds before the kernel can launch. I suggest we still mark it as static. |
|
@JehandadKhan sure, this make sense. simliar to what compiler will do. |
AFAICS from the source code, the most of convolution parameters are passed to the kernels at run time (not at compile time), therefore the kernels are dynamic. The problem is that tuning config may be different for each problem. But different tuning config means different compilation options, and that is why binary cache miss happens. In other words, it is the tuning which makes the Solver non-dynamic. This is known potential drawback of tuning. I am writing about this so that we all have a common understanding of this problem. Please do not hesitate to ask if something is unclear or if you not agree. |
* 3D group forward convolution solver (#2286) * [HotFix] Fix build issue after #2286 (#2328) * Fix ConvCkIgemmFwdV6r1DlopsNchw solver to reflect that it's not dynamic (#2325) * Remove target IDs from kdb entries (#2309) * Dropout: make seed and states_num kernel arguments (#2277) * [MI250] Adding missing kernel objects (#2329) * Post-merge fixups: Replace environment variable check with problem config check and reduce lambda capture for Invoker obj (#2305) * [HotFix][CI] fix HIP tidy issue from #2277 (#2335) * [HotFix] Update requirements.txt MLIR ignore PATH for Python conda LLD (#2324) * [NFC] Replace long integers with int64_t and size_t for better compatibility with Windows (#2323) * Windows compatibility: replace long integers with int64_t and size_t, replace uint with unsigned int, replace long with long long for numbers, proper casting * Fix formatting * Fix 3d group forward convolution * Resolve review comments * Fix formatting --------- Co-authored-by: Daming Feng <dmfeng8898@gmail.com> Co-authored-by: Evgenii Averin <86725875+averinevg@users.noreply.github.com> Co-authored-by: JD <jahandad@gmail.com> Co-authored-by: Tal Ben-Nun <tbennun@users.noreply.github.com> Co-authored-by: amberhassaan <mah@katanagraph.com> Co-authored-by: Jun Liu <Liu.Jun@amd.com>
ConvCkIgemmFwdV6r1DlopsNchwis not dynamic, however the class definition marks it as dynamic. This PR updates the definition so thatConvCkIgemmFwdV6r1DlopsNchwis no longer marked as Dynamic.