[Mlir-commits] [mlir] [mlir] Clamp UnPackOp tiling sizes from operand tile (PR #112429)

[llvmlistbot at llvm.org]

================
@@ -265,7 +265,7 @@ module {
         %c4 = arith.constant 4 : index
         %c64 = arith.constant 64 : index
         %c0 = arith.constant 0 : index
-        %1 = scf.forall (%arg3, %arg4) in (2, 2) shared_outs(%arg5 = %arg2) -> (tensor<64x32xf32>) {
+        %1 = scf.forall (%arg3, %arg4) = (0, 0) to (64, 32) step (32, 32) shared_outs(%arg5 = %arg2) -> (tensor<64x32xf32>) {
----------------
Max191 wrote:

This test also tests the case, because offsets are dynamic, but it is not entirely clear in this case because the unpack is perfectly aligned (i.e., the dynamic result sizes could potentially be optimized away). I will add another test with an imperfectly aligned unpack op.

https://github.com/llvm/llvm-project/pull/112429