vela: Add support for CPU only networks

 - Fix various problems when no operators run on Ethos-U55

Signed-off-by: Tim Hall <tim.hall@arm.com>
Change-Id: I44a1a914fabb7ca26c921a02753da8abeecd9c7b

2 files changed, 24 insertions, 20 deletions

diff --git a/ethosu/vela/compiler_driver.py b/ethosu/vela/compiler_driver.py
index b6a98a64..9c345dba 100644
--- a/ethosu/vela/compiler_driver.py
+++ b/ethosu/vela/compiler_driver.py
@@ -138,24 +138,23 @@ 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]
-    assert first_npu_sg.placement == PassPlacement.Npu
-    # Use the linear allocator for constant tensors
-    tensor_allocation.allocate_tensors(
-        nng,
-        first_npu_sg,
-        arch,
-        permanent_storage,
-        scheduler_options.use_ifm_ofm_overlap,
-        TensorAllocator.LinearAlloc,
-        options.verbose_allocation,
-        options.show_minimum_possible_allocation,
-        lr_graph_flash,
-    )
+    if len(nng.subgraphs) > 1:
+        # 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]
+        assert first_npu_sg.placement == PassPlacement.Npu
+        # Use the linear allocator for constant tensors
+        tensor_allocation.allocate_tensors(
+            nng,
+            first_npu_sg,
+            arch,
+            permanent_storage,
+            scheduler_options.use_ifm_ofm_overlap,
+            TensorAllocator.LinearAlloc,
+            options.verbose_allocation,
+            options.show_minimum_possible_allocation,
+            lr_graph_flash,
+        )
 
     # Allocate all non-constant tensors to the root, i.e. Cpu, subgraph. This step
     # will start at the root subgraph's input and traverse from top to bottom. When
diff --git a/ethosu/vela/tflite_writer.py b/ethosu/vela/tflite_writer.py
index 99df849b..675b6985 100644
--- a/ethosu/vela/tflite_writer.py
+++ b/ethosu/vela/tflite_writer.py
@@ -134,12 +134,17 @@ class TFLiteSerialiser:
         return builder.EndVector(len(v))
 
     def assign_buffers_to_tensors(self, tensors):
+        scratch_tensors = [tens for tens in tensors if tens.purpose == TensorPurpose.Scratch]
+        if len(scratch_tensors) > 0:
+            scratch_tensor_mem_area = scratch_tensors[0].mem_area
+        else:
+            scratch_tensor_mem_area = None  # all tensors are initialised to MemArea.Unknown
+
         buffer_map = {}
-        scratch_tensor = [tens for tens in tensors if tens.purpose == TensorPurpose.Scratch][0]
         buf_idx = 1
 
         for tens in tensors:
-            if tens.mem_area == scratch_tensor.mem_area:
+            if tens.mem_area == scratch_tensor_mem_area:
                 buffer_map[tens] = self.scratch_buf_id
             else:
                 buffer_map[tens] = buf_idx