increase k=64

NVIDIA · Jan 31, 2025 · e68f85c · e68f85c
1 parent 8b399d2
commit e68f85c
Showing 1 changed file with 5 additions and 1 deletion.
diff --git a/csrc/scheduler/matmul_utils.cpp b/csrc/scheduler/matmul_utils.cpp
@@ -227,7 +227,11 @@ bool fillDefaultHopperHeuristic(
 
   // TODO: it might be advantageous in some cases to issue multiple wgmma
   // instructions per warp group
-  warp_tile = instruction_tile;
+  constexpr int64_t k_ratio = 4;
+  warp_tile = {
+        instruction_tile.m,
+        instruction_tile.n,
+        instruction_tile.k * k_ratio};
 
   // The MmaOp output is a 32-bit float which requires one register per value