MLBEDSW-1998: Add support for new_axis_mask

-Added support for new_axis_mask
-Added support for more than 1 bit set in new/shrink_axis mask
-Added checks for strided slice in supported operator check
-Added assert if nothing has been put on NPU

Change-Id: I66e2d04784f14d7ad82371f5d649a455d576a818
Signed-off-by: Patrik Gustavsson <patrik.gustavsson@arm.com>

4 files changed, 49 insertions, 17 deletions

diff --git a/ethosu/vela/compiler_driver.py b/ethosu/vela/compiler_driver.py
index 7f8c4ca4..db669ac7 100644
--- a/ethosu/vela/compiler_driver.py
+++ b/ethosu/vela/compiler_driver.py
@@ -141,6 +141,8 @@ def compiler_driver(nng, arch, options, scheduler_options):
                 sg, permanent_storage, ignore_subgraph_input_output_tensors=True, lr_graph=lr_graph_flash
             )
 
+    assert len(nng.subgraphs) > 1, "Error: No operators can be hardware accelerated; cancelling compilation"
+
     # Allocate all Npu constant tensors to the first Npu subgraph since it is
     # processed first during serialization into tensors
     first_npu_sg = nng.subgraphs[1]
diff --git a/ethosu/vela/graph_optimiser.py b/ethosu/vela/graph_optimiser.py
index f0afcf8f..a4ed39ff 100644
--- a/ethosu/vela/graph_optimiser.py
+++ b/ethosu/vela/graph_optimiser.py
@@ -208,23 +208,49 @@ def fixup_unpack_output(tens, arch):
     if op.type in set(("Unpack", "StridedSlice")):
         # Unpack is also referred to as Unstack
         # Requires the rewrite_split function to be called on the op afterwards
+
+        reshape_input_shape = tens.shape
         if op.type == "StridedSlice":
+            new_axis_mask = op.attrs["new_axis_mask"]
             shrink_axis_mask = op.attrs["shrink_axis_mask"]
-            if shrink_axis_mask == 0:
+            ellipsis_mask =  op.attrs["ellipsis_mask"]
+
+            if (new_axis_mask != 0 and shrink_axis_mask != 0) or ellipsis_mask != 0:
+                # Not supported, will be put on CPU
+                return tens
+            if shrink_axis_mask == 0 and new_axis_mask == 0:
                 # Equal Rank StridedSlice, no need to insert reshape
                 return tens
-
-            # Only allow shrinking 1 axis for now
-            assert shrink_axis_mask & (shrink_axis_mask - 1) == 0
-            assert len(tens.shape) == (len(op.inputs[0].shape) - 1)
-
-            axis = int(math.log2(shrink_axis_mask))
-            op.attrs["shrink_axis_mask"] = 0
+            elif shrink_axis_mask != 0:
+                n = 0
+                axis = 0
+                while shrink_axis_mask:
+                    prev_mask = shrink_axis_mask
+                    n += 1
+                    shrink_axis_mask &= shrink_axis_mask - 1
+                    axis = int(math.log2(prev_mask - shrink_axis_mask))
+                    reshape_input_shape = reshape_input_shape[:axis] + [1] + reshape_input_shape[axis:]
+
+                assert len(tens.shape) == (len(op.inputs[0].shape) - n)
+                op.attrs["shrink_axis_mask"] = 0
+
+            elif new_axis_mask != 0:
+                n = 0
+                axis = 0
+                while new_axis_mask:
+                    prev_mask = new_axis_mask
+                    n += 1
+                    new_axis_mask &= new_axis_mask - 1
+                    axis = int(math.log2(prev_mask - new_axis_mask))
+                    reshape_input_shape = reshape_input_shape[:axis] + reshape_input_shape[(axis + 1):]
+                    new_axis_mask >>= 1
+
+                assert len(tens.shape) == (len(op.inputs[0].shape) + n)
+                op.attrs["new_axis_mask"] = 0
         else:
             axis = int(op.attrs["axis"])
             op.type = "UnpackReshaped"
-
-        desired_shape = tens.shape[:axis] + [1] + tens.shape[axis:]
+            reshape_input_shape = tens.shape[:axis] + [1] + tens.shape[axis:]
 
         # Construct 1 shape tensor to be used by all inserted reshape ops
         new_shape_name = op.name + "_reshape_shape"
@@ -239,7 +265,7 @@ def fixup_unpack_output(tens, arch):
             reshape_op = Operation("Reshape", reshape_name)
             reshape_op.outputs = [out_tens]
             reshape_in = out_tens.clone("_reshaped")
-            reshape_in.shape = reshape_in.storage_shape = reshape_in.bandwidth_shape = desired_shape
+            reshape_in.shape = reshape_in.storage_shape = reshape_in.bandwidth_shape = reshape_input_shape
             reshape_in.ops = [op]
             out_tens.ops = [reshape_op]
             reshape_op.inputs = [reshape_in, new_shape_tens]
diff --git a/ethosu/vela/operation.py b/ethosu/vela/operation.py
index f36e61c6..a25574a6 100644
--- a/ethosu/vela/operation.py
+++ b/ethosu/vela/operation.py
@@ -248,12 +248,10 @@ input and output tensors, as well as an attribute dictionary."""
             end_mask = self.attrs["end_mask"]
             new_axis_mask = self.attrs["new_axis_mask"]
             shrink_axis_mask = self.attrs["shrink_axis_mask"]
-            # TODO: Either extend this to support these different masks or check
-            # for this at an earlier stage and place the op on Cpu if needed
-            assert new_axis_mask == ellipsis_mask == 0
-            # shrink_axis_mask is not supported by the Operation class but the operation
+
+            # shrink_axis_mask/new_axis_mask/ellipsis_mask is not supported by the Operation class but the operation
             # may have the attribute modified and handled in the graph optimization phase.
-            assert shrink_axis_mask == 0
+            assert shrink_axis_mask == new_axis_mask == ellipsis_mask == 0
             assert len(input_tens.shape) == len(out_tens.shape)
 
             for idx in range(len(input_tens.shape)):
diff --git a/ethosu/vela/supported_operators.py b/ethosu/vela/supported_operators.py
index 1e11d788..1a25887f 100644
--- a/ethosu/vela/supported_operators.py
+++ b/ethosu/vela/supported_operators.py
@@ -240,8 +240,14 @@ class SupportedOperators:
         return True
 
     def check_memory_only_restrictions(self, op):
-        # check stride size
         if op.type == "StridedSlice":
+            # check stride size
             if len(op.inputs) > 3 and any(stride != 1 for stride in op.inputs[3].values):
                 return False
+            # check ellipsis_mask
+            if op.attrs["ellipsis_mask"] != 0:
+                return False
+            # check if both new_axis_mask and shrink_axis_mask have bit set
+            if op.attrs["new_axis_mask"] != 0 and op.attrs["shrink_axis_mask"] != 0:
+                return False
         return True