Go back to old Heuristic but ask for more bestKElementCountPerSubgroup

Signed-off-by: Nirvedh Meshram <nirvedh@gmail.com>

compiler/src/iree/compiler/Codegen/Dialect/GPU/TargetUtils/ConfigUtils.cpp

@@ -135,6 +135,7 @@ static std::optional<GPUMMASchedule> getMmaScheduleFromProblemAndTarget(
   GPUMMAHeuristicSeeds seeds;
   assert(problem.aType == problem.bType &&
          "expected the same aType and bType.");
+  int64_t inBitWidth = problem.aType.getIntOrFloatBitWidth();
 
   // Note that the following heuristic seeds are just placeholder values.
   // We need to clean it up and make it adjusting to different targets.
@@ -147,14 +148,23 @@ static std::optional<GPUMMASchedule> getMmaScheduleFromProblemAndTarget(
     // and a larger bestKTileCountPerSubgroup.
     seeds = {/*bestSubgroupCountPerWorkgroup=*/4,
              /*bestMNTileCountPerSubgroup=*/4,
-             /*bestKTileCountPerSubgroup=*/8};
+             /*bestKTileCountPerSubgroup=*/8,
+             /*bestKElementCountPerSubgroup*/ kCacheLineSizeBits * 4 /
+                 inBitWidth};
   } else {
     seeds = {/*bestSubgroupCountPerWorkgroup=*/4,
-             /*bestMNTileCountPerSubgroup=*/8,
-             /*bestKTileCountPerSubgroup=*/4};
-  }
-
-  int64_t maxSharedMemoryBytes = target.getWgp().getMaxWorkgroupMemoryBytes();
+             /*bestMNTileCountPerSubgroup=*/16,
+             /*bestKTileCountPerSubgroup=*/4,
+             /*bestKElementCountPerSubgroup*/ kCacheLineSizeBits * 2 /
+                 inBitWidth};
+  }
+
+  // We target slightly below the full available shared Memory to leave room for
+  // `GPUReduceBankConflictsPass` that will pad shared memory without keeping
+  // track of usage. We can drop this after fixing
+  // https://github.com/iree-org/iree/issues/19675
+  int64_t maxSharedMemoryBytes =
+      target.getWgp().getMaxWorkgroupMemoryBytes() - 64 * inBitWidth;
 
   // First try to find a schedule with an exactly matching intrinsic.
   std::optional<GPUMMASchedule> schedule = deduceMMASchedule(