Fixes

- Check operand type instead of dotop which returns accum type.
- Fix medium size typo
- Enable pingpong by default, still can disable it by using `TRITON_HIP_USE_BLOCK_PINGPONG=0`

third_party/amd/backend/compiler.py

@@ -243,7 +243,7 @@ def make_ttgir(mod, metadata, options):
         passes.ttgpuir.add_reduce_data_duplication(pm)
         if amd.has_matrix_core_feature(options.arch):
             amd.passes.ttgpuir.add_reorder_instructions(pm)
-            use_block_pingpong = os.getenv("TRITON_HIP_USE_BLOCK_PINGPONG", "0") == "1"
+            use_block_pingpong = os.getenv("TRITON_HIP_USE_BLOCK_PINGPONG", "1") == "1"
             if use_block_pingpong and options.num_stages == 2:
                 amd.passes.ttgpuir.add_block_pingpong(pm)
 

third_party/amd/lib/TritonAMDGPUTransforms/BlockPingpong.cpp

@@ -433,11 +433,11 @@ void Pingponger::getDotPingponged() {
   auto dotShape = dotType.getShape();
   auto AType = dotOps[0].getA().getType();
   auto AShape = AType.getShape();
-  auto elemWidth = dotType.getElementType().getIntOrFloatBitWidth();
+  auto elemWidth = AType.getElementType().getIntOrFloatBitWidth();
   int64_t tileSize = dotShape[0] * dotShape[1] * AShape[1] * elemWidth;
 
   const int64_t SmallTile = 16777216;  // e.g. 128x128x64x16bit
-  const int64_t MediumTile = 35554432; // SmallTile x 2
+  const int64_t MediumTile = 33554432; // SmallTile x 2
   const int64_t LargeTile = 67108864;  // e.g. 256x256x64x16bit
 
   auto encoding = cast<RankedTensorType>(AType).getEncoding();