fix: compare mrgsort_format2 by stable output prefix

test/npu_validation/scripts/generate_testcase.py

@@ -1623,6 +1623,16 @@ def generate_testcase(
         file_cnt = ptr_elem_counts.get(name, logical_elem_count)
         if file_cnt and req < int(file_cnt):
             compare_prefix_counts[name] = req
+    # TMRGSORT format2 testcase writes three contiguous regions:
+    # 2-way (256) + 3-way (384) + 4-way (up to 512).
+    # With 4-way exhausted mode, the stable worst-case valid prefix for 4-way
+    # is 128 elements, so compare 256 + 384 + 128 = 768 elements.
+    testcase_lc = testcase.lower()
+    if testcase_lc == "mrgsort_format2":
+        for p in output_ptrs:
+            name = p["name"]
+            file_cnt = int(ptr_elem_counts.get(name, logical_elem_count))
+            compare_prefix_counts[name] = min(file_cnt, 768)
     for p in output_ptrs:
         np_dtype = _np_dtype_for_cpp(p["cpp_type"])
         name = p["name"]

test/samples/Mrgsort/mrgsort_format2.py

@@ -125,15 +125,15 @@ def build():
                     srcs=[tb_s0, tb_s1, tb_s2],
                     dsts=[tb_dst3, tb_tmp3],
                     excuted=excuted,
-                    exhausted=True,
+                    exhausted=False,
                 )
 
                 # 4-way: src0 + src1 + src2 + src3 -> 1x512
                 pto.TMrgSortOp(
                     srcs=[tb_s0, tb_s1, tb_s2, tb_s3],
                     dsts=[tb_dst4, tb_tmp4],
                     excuted=excuted,
-                    exhausted=False,
+                    exhausted=True,
                 )
 
                 sv_out2 = pto.PartitionViewOp(part_view_1x256, tv_out, offsets=[c0, c0], sizes=[c1, c256]).result