Add Grid stride pairwise dist and fused L2 NN kernels #232

Merged

rapids-bot[bot] merged 11 commits intorapidsai:branch-21.06from mdoijade:gridStridedDist

Jun 2, 2021

+359-328

Commits on May 12, 2021

Refactor fusedL2NN to use pairwiseDistance class. invert block y/x dir usage in all contraction based kernels so that n is along x dir and m is along y dir blocks
mdoijade
committed

Commits on May 17, 2021

-- add grid stride support to pairwise distance based cosine, l2, l1 kernels. --add launch config generator function to launch optimal grid size kernel for these pairwise dist kernels
mdoijade
committed

Commits on May 20, 2021

--Add additional Mblk + Nblk shmem for storing norms, and reuse xNorm for subsequent gridStrideX variations. this overall improves perf of fusedL2NN to 1.85x over previous version. --Also remove ch…
mdoijade
committed

Commits on May 24, 2021

Use cudaOccupancyMaxActiveBlocksPerSM instead of hard-coded launch bound in launchConfigGenerator. --Use constexpr in shmemSize.
mdoijade
committed

Commits on May 27, 2021

Merge branch 'branch-21.06' into gridStridedDist
mdoijade
committed

Commits on Jun 1, 2021