Enable 4x4 herd for pack-peel (not elementwise fusion)

erwei-xilinx · Oct 23, 2024 · cb6a36d · cb6a36d
1 parent bea71d6
commit cb6a36d
Showing 1 changed file with 4 additions and 2 deletions.
diff --git a/compiler/plugins/target/AMD-AIE/iree-amd-aie/Transforms/Passes.cpp b/compiler/plugins/target/AMD-AIE/iree-amd-aie/Transforms/Passes.cpp
@@ -809,10 +809,12 @@ void addMLIRAIRLoweringPasses(OpPassManager &passManager, AMDAIEDevice device,
     // with given factors, and subsequently unrolled in
     // AIRUnrollOuterPerfectlyNestedLoopsPass, to enforce SHIM DMA BD count
     // within the hardware limit.
-    if (useTilePipeline == TilePassPipeline::PackPeelPipeline) {
+    if (useTilePipeline == TilePassPipeline::PackPeelPipeline &&
+        matmulElementwiseFusion) {
       const static llvm::SmallVector<unsigned> tile_sizes = {2, 2};
       options.clTileSizes = tile_sizes;
-    } else if (useTilePipeline == TilePassPipeline::PadPackPipeline) {
+    } else if (useTilePipeline == TilePassPipeline::PadPackPipeline ||
+               useTilePipeline == TilePassPipeline::PackPeelPipeline) {
       const static llvm::SmallVector<unsigned> tile_sizes = {4, 4};
       options.clTileSizes = tile_sizes;
     }