Arm backend: Support constant negative pad values (#24108)

Lower negative constant padding to slices before applying non-negative
TOSA padding. This supports cropping on either side, mixed cropping and
padding, and cropping across multiple dimensions.

Preserve internal zero pairs and mixed-sign padding in the test module.
Add regression cases across floating-point and quantized TOSA and VGF
pipelines.

Validation: 148 padding tests pass, with three expected hardware-runner
xfails. Lintrunner passes for both changed files.

This commit was authored with Codex.

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils
@Sebastian-Larsson @robell @rascani

Signed-off-by: Erik Lundell <erik.lundell@arm.com>

Author: Erik-Lundell

backends/arm/_passes/rewrite_pad.py

@@ -25,6 +25,26 @@ class RewritePadPass(ArmOpTargetedPass):
     }
 
     def _rewrite_constant_pad(self, input_tensor, pad, value, meta):
+        input_rank = len(input_tensor.data.shape)
+        pad = list(pad)
+        for i in range(0, len(pad), 2):
+            left, right = pad[i : i + 2]
+            if left < 0 or right < 0:
+                # Negative padding crops the input; TOSA PAD only adds elements.
+                input_tensor = super().call_operator(
+                    exir_ops.edge.aten.slice_copy.Tensor,
+                    (
+                        input_tensor,
+                        input_rank - 1 - i // 2,
+                        max(-left, 0),
+                        right if right < 0 else None,
+                    ),
+                    {},
+                    meta,
+                    True,
+                )
+                pad[i : i + 2] = [max(left, 0), max(right, 0)]
+
         output_dtype = meta["val"].dtype
         if output_dtype in (torch.int8, torch.int16):
             input_qparams = meta.data.get("input_qparams", {})
@@ -43,7 +63,6 @@ def _rewrite_constant_pad(self, input_tensor, pad, value, meta):
         input_pad = []
         for pair in reversed(pad_pairs):
             input_pad.extend(pair)
-        input_rank = len(input_tensor.data.shape)
         # Place spatial dimensions last and pad non-spatial dimensions with 0 padding
         shape = [0] * ((input_rank * 2 - len(pad))) + input_pad
 

backends/arm/test/ops/test_constant_pad_nd.py

@@ -35,6 +35,23 @@
 input_t1 = Tuple[torch.Tensor]  # Input x
 
 test_data_suite = {
+    "crop_end": lambda: (torch.rand(1, 4, 16), (0, -1), 0, "constant"),
+    "crop_start": lambda: (torch.rand(1, 4, 16), (-2, 0), 0, "constant"),
+    "crop_both": lambda: (torch.rand(1, 4, 16), (-1, -2), 0, "constant"),
+    "crop_penultimate_dim": lambda: (
+        torch.rand(1, 16, 4),
+        (0, 0, 0, -1),
+        0,
+        "constant",
+    ),
+    "crop_and_pad": lambda: (torch.rand(1, 4, 16), (-2, 3), 2, "constant"),
+    "crop_shift": lambda: (torch.rand(1, 4, 16), (-1, 1), -0.5, "constant"),
+    "crop_multidim": lambda: (
+        torch.rand(1, 3, 8, 10),
+        (-1, 2, 3, -2),
+        1,
+        "constant",
+    ),
     "4dim_last1dim": lambda: (
         torch.rand(1, 1, 16, 16),
         (1, 1, 0, 0, 0, 0, 0, 0),