switch mma instruction shape to 1684 from current 1688 for 3xTF32 L2/…

…cosine kernel (#1057)

-- switch mma instruction shape to 1684 from current 1688 as it is always faster for all the inputs tried from DISTANCE_BENCH for L2 and cosine distances.
-- the speedup in best case is 1.37x, and at minimum it is 1.05x faster.

Authors:
  - Mahesh Doijade (https://github.com/mdoijade)

Approvers:
  - Tamas Bela Feher (https://github.com/tfeher)

URL: #1057

cpp/include/raft/distance/detail/pairwise_distance_gemm.h

@@ -66,7 +66,7 @@ struct PairwiseDistanceGemm {
   /// Warp-level tile size (concept: GemmShape)
   // This code section describes the size of MMA op
   using InstructionShape =
-    cutlass::gemm::GemmShape<16, 8, 8>;  // <- MMA Op tile M = 16, N = 8, K = 8
+    cutlass::gemm::GemmShape<16, 8, 4>;  // <- MMA Op tile M = 16, N = 8, K = 4
 
   /// Operation performed by GEMM
   using Operator = cutlass::arch::OpMultiplyAddFastF32;