[CI] Add larger shape e2e bfp16 Ukernel test (#1067)

-- This commit adds larger shape e2e bfp16 Ukernel test to CI for
   both `pack-peel` as well as `pack-peel-4-level-tiling` targeting
   4x8 array on Strix.
-- It wasn't working earlier (3 months ago) but now it does - hence
    enabling.

Signed-off-by: Abhishek Varma <[email protected]>

build_tools/ci/cpu_comparison/run.py

@@ -1751,7 +1751,6 @@ def __init__(self):
                 )
             )
 
-        # TODO: Failure is expected for the 128x128 case we don't yet understand why.
         self.register(
             Matmul(
                 64,
@@ -1816,6 +1815,24 @@ def __init__(self):
                 use_chess=True,
             )
         )
+        self.register(
+            Matmul(
+                512,
+                512,
+                512,
+                "bf16",
+                "f32",
+                name_suffix="4rows_8cols_npu4_pack_peel_4_level_tiling_ukernel",
+                use_ukernel=True,
+                tile_pipeline="pack-peel-4-level-tiling",
+                run_on_target=["npu4"],
+                aie_compilation_flags=[
+                    "--iree-amdaie-num-rows=4",
+                    "--iree-amdaie-num-cols=8",
+                ],
+                use_chess=True,
+            )
+        )
 
         # Matmul test on 2(rows)x2(cols) cores
         self.register(