Add Grid stride pairwise dist and fused L2 NN kernels #232
Conversation
…r usage in all contraction based kernels so that n is along x dir and m is along y dir blocks
…kernels. --add launch config generator function to launch optimal grid size kernel for these pairwise dist kernels
…ed up over previous version. -- improve logic of the grid launch config generator for x-dir blocks
…ced val for pre-volta arch
… for subsequent gridStrideX variations. this overall improves perf of fusedL2NN to 1.85x over previous version. --Also remove checking keys only check values in fusedL2nn test case, as it may happen a row has multiple keys with same min val
|
@teju85 for help with reviewing and adding appropriate label |
teju85
added
improvement
…und in launchConfigGenerator. --Use constexpr in shmemSize.
|
some intermittent CI related issue can you rerun the tests @teju85 |
|
@teju85 can we merge this ? if not in branch-21.06 then in branch-21.08? |
|
@BradReesWork and/or @afender can we get approval for this from cugraph side? (I believe since you are not using pairwise distances logic anywhere in cugraph, the changes in this PR should not affect you) |
|
@divyegala you are still listed as a code owner and thus we'll need your approval too before merging! |
|
@teju85 @dantegd can we merge this now? the cuML side test PR is working fine as well the L1 dist intermittent issue seems resolved - rapidsai/cuml#3891 |
|
@gpucibot merge |
After the merge of #232, a few different tests failed in rapidsai/cuml#3891, given the timing I think it'd be best to target 232 (again) to 21.08 after triaging the issues. Authors: - Dante Gama Dessavre (https://github.com/dantegd) Approvers: - Corey J. Nolet (https://github.com/cjnolet) - Divye Gala (https://github.com/divyegala) - Brad Rees (https://github.com/BradReesWork) URL: #246
This PR addresses issues mentioned in #221 -- Adds grid stride based fusedL2NN kernel, this gives approx 1.85x speed up over previous version of this kernel. -- Adds support in pairwise dist base class to work for any input size by adding support for grid stride based work distribution. This was submitted to branch-21.06 through PR - #232 but later reverted due to intermittent failure by - #246 Authors: - Mahesh Doijade (https://github.com/mdoijade) Approvers: - Thejaswi. N. S (https://github.com/teju85) - Brad Rees (https://github.com/BradReesWork) URL: #250
This PR addresses issues mentioned in #221
-- Adds grid stride based fusedL2NN kernel, this gives approx 1.85x speed up over previous version of this kernel.
-- Adds support in pairwise dist base class to work for any input size by adding support for grid stride based work distribution.