MLBEDSW-3623: Diff on semantic_segmentation

The root cause of this diff is precision errors caused by rounding
several times when performing a resize bilinear upscaling to more than
twice the initial size. This is solved by rewriting the algorithm to
perform nearest neighbour upscaling to the correct size and then
applying one larger average pool instead of several 2x2 pools. Avgpool
with padding is limited to kernel size 8x8, which constraints the
largest possible bilinear upscaling to 8 times the input size.

Signed-off-by: Rickard Bolin <rickard.bolin@arm.com>
Change-Id: I846232f309ba26aab6c385e593cbe25b646c6668

2 files changed, 42 insertions, 46 deletions

diff --git a/ethosu/vela/tflite_graph_optimiser.py b/ethosu/vela/tflite_graph_optimiser.py
index 8cfc373..4098798 100644
--- a/ethosu/vela/tflite_graph_optimiser.py
+++ b/ethosu/vela/tflite_graph_optimiser.py
@@ -299,13 +299,13 @@ def convert_resizebilinear_1x1_to_add(op):
     return op
 
 
-# Convert ResizeBilinear to a number of 2x2 pool ops
-def convert_resizebilinear_to_2x2_pool(op):
-    count = 0
+# Convert ResizeBilinear to a number of 2x2 nearest neighbor upscaling and one avgpool op with kernel size dependent
+# on the upscaling factor. Avgpool kernel limit of 8x8 when padding is applied limits upscaling to 8x8.
+def convert_resizebilinear_to_nearest_neighbor_upscaling_and_pool(op):
     pre_op = op
     outputs = op.outputs
-
-    op.attrs.update({"strides": (1, 1, 1, 1), "ksize": (1, 2, 2, 1)})
+    dtype = op.ifm.dtype
+    op.attrs.update({"strides": (1, 1, 1, 1), "ksize": (1, 1, 1, 1)})
     if op.attrs["align_corners"]:
         shape_modifier = 1
         op.attrs["padding"] = Padding.VALID
@@ -316,41 +316,41 @@ def convert_resizebilinear_to_2x2_pool(op):
 
     upscaled_shape = np.array(op.ifm_shapes[0].get_hw_as_list())
     out_shape = np.array(op.ofm_shapes[0].get_hw_as_list())
-    if (upscaled_shape == upscaled_shape * 2 - shape_modifier).all():
-        return op
 
-    while (upscaled_shape < out_shape).all():
-        if count == 0:
-            scaled_op = pre_op
-        else:
-            scaled_op = op.clone("_{}".format(count))
+    # Calculate how many times 2x2 upscaling needs to be performed
+    upscale_factor = round(out_shape[1] / upscaled_shape[1])
+    n = int(np.log2(upscale_factor))
+
+    # Perform 2x2 upscaling n-1 times
+    scaled_op = pre_op
+    for count in range(n - 1):
+        if count > 0:
+            scaled_op = op.clone(f"_{count}")
             scaled_op.inputs[0] = pre_op.outputs[0]
 
+        # Nearest neighbor 2x2 upscaling
         upscaled_shape = upscaled_shape * 2 - shape_modifier
+        shape = op.ofm_shapes[0].as_list()
+        shape[1:3] = upscaled_shape
+        out_tens = Tensor(shape, dtype, f"{op.outputs[0].name}_{count}")
+        out_tens.quantization = op.outputs[0].quantization.clone()
+        scaled_op.set_output_tensor(out_tens)
+        pre_op = scaled_op
 
-        if (upscaled_shape == out_shape).all():
-            scaled_op.outputs = outputs
-            scaled_op.outputs[0].ops = [scaled_op]
-        else:
-            shape = op.ofm_shapes[0].as_list()
-            shape[1:3] = upscaled_shape
-            out_tens = Tensor(shape, DataType.int16, "{}_{}".format(op.outputs[0].name, count))
-            out_tens.quantization = op.outputs[0].quantization.clone()
-            out_tens.quantization.quant_min = np.iinfo(np.int16).min
-            out_tens.quantization.quant_max = np.iinfo(np.int16).max
-            scaled_op.set_output_tensor(out_tens)
-            pre_op = scaled_op
-            count += 1
-
-        # Setup the scale value
-        if scaled_op.inputs[0].dtype.bits == 8 and scaled_op.outputs[0].dtype.bits == 16:
-            scaled_op.rescale = 128
-        elif scaled_op.inputs[0].dtype.bits == 16 and scaled_op.outputs[0].dtype.bits == 8:
-            scaled_op.rescale = 1 / 128
-        else:
-            scaled_op.rescale = None
         scaled_op.set_ifm_ofm_shapes()
 
+    # Last 2x2 upscaling also applies avgpool with kernel size dependent on the upscaling factor and adds
+    # padding to the right and bottom.
+    if n > 1:
+        scaled_op = op.clone(f"_{n-1}")
+        scaled_op.inputs[0] = pre_op.outputs[0]
+    scaled_op.attrs["padding"] = Padding.EXPLICIT
+    scaled_op.attrs["explicit_padding"] = [0, 0, upscale_factor - 1, upscale_factor - 1]
+    scaled_op.attrs.update({"ksize": (1, upscale_factor, upscale_factor, 1)})
+    scaled_op.outputs = outputs
+    scaled_op.outputs[0].ops = [scaled_op]
+    scaled_op.set_ifm_ofm_shapes()
+
     return op
 
 
@@ -363,7 +363,7 @@ def fixup_resizebilinear(op, arch, nng):
         elif op.ifm_shapes[0].height == 1 and op.ifm_shapes[0].width == 1:
             convert_resizebilinear_1x1_to_add(op)
         else:
-            convert_resizebilinear_to_2x2_pool(op)
+            convert_resizebilinear_to_nearest_neighbor_upscaling_and_pool(op)
 
     return op
 
diff --git a/ethosu/vela/tflite_supported_operators.py b/ethosu/vela/tflite_supported_operators.py
index 60bc6fd..193a23f 100644
--- a/ethosu/vela/tflite_supported_operators.py
+++ b/ethosu/vela/tflite_supported_operators.py
@@ -511,8 +511,8 @@ class TFLiteSupportedOperators:
         """The width and height of the IFM and OFM must match one of the following criteria:
         IFM W and H must both be 1
         IFM must match OFM
-        OFM W and H must be 2x IFM -1, if align_corners is True
-        OFM W and H must be 2x IFM, if align_corners is False"""
+        OFM W and H must be equal and 2/4/8x IFM -1, if align_corners is True
+        OFM W and H must be equal and 2/4/8x IFM, if align_corners is False"""
         # Easier to start with False condition as very few cases result in a supported resize
         valid = False
         ifm_shape = op.ifm.shape
@@ -523,16 +523,12 @@ class TFLiteSupportedOperators:
             if ((ifm_shape[1] == 1) and (ifm_shape[2] == 1)) or (ifm_shape == ofm_shape):
                 valid = True
             else:
-                upscaled_shape = np.array(ifm_shape[1:3])
-                out_shape = np.array(ofm_shape[1:3])
-                while (upscaled_shape < out_shape).all():
-                    upscaled_shape *= 2
-                    if align_corners:
-                        upscaled_shape -= 1
-                    # Valid if OFM is 2x IFM (-1 for align corners)
-                    if np.array_equal(out_shape, upscaled_shape):
-                        valid = True
-                        break
+                # Valid if OFM is 2/4/8x IFM (-1 for align corners)
+                w_upscale_factor = (ofm_shape[1] + 1) / ifm_shape[1] if align_corners else ofm_shape[1] / ifm_shape[1]
+                h_upscale_factor = (ofm_shape[2] + 1) / ifm_shape[2] if align_corners else ofm_shape[2] / ifm_shape[2]
+
+                valid = w_upscale_factor == h_upscale_factor and w_upscale_factor in [2, 4, 8]
+
         return valid, f"Op has ifm_shape={ifm_shape}, ofm_shape={ofm_shape} and align_corners={align_corners}"
 
     @staticmethod