MLBEDSW-2809: Redo the Tensor addressing

Added a static class TensorAddressMap that stores all Tensor addresses
based on their equivalence_id. Made the "address" field into a property
which getter and setter looks up/sets the tensor's address in
TensorAddressMap.

This makes the references to cpu_tensor/npu_tensor obsolete and they
have been removed.

Addition to scheduler: avoid SRAM spilling if an op has consumers in
other subgraphs.

Minor rework in LUTState; it will now assign a unique equivalence_id to
the SHRAM lut tensor to avoid issues with addressing. The equivalent
checks in LUTState now compares the values of the LUT instead of the the
equivalence_id.

Updated LUT unit tests accordingly.

Signed-off-by: Jacob Bohlin <jacob.bohlin@arm.com>
Change-Id: I41de5a8a4e5f07b77d6544d8d4034b754993e503

1 files changed, 14 insertions, 3 deletions

diff --git a/ethosu/vela/scheduler.py b/ethosu/vela/scheduler.py
index e9a93c19..47f8a47f 100644
--- a/ethosu/vela/scheduler.py
+++ b/ethosu/vela/scheduler.py
@@ -35,6 +35,7 @@ from .npu_performance import make_cycles_array
 from .npu_performance import make_macs_array
 from .npu_performance import make_metrics_arrays
 from .npu_performance import PassCycles
+from .numeric_util import full_shape
 from .operation import NpuBlockType
 from .shared_buffer_allocation import find_block_configs_suitable_for_pass_and_shared_buffer
 from .shared_buffer_allocation import shared_buffer_allocation_for_pass_and_block_config
@@ -43,7 +44,7 @@ from .tensor import MemType
 from .tensor import TensorFormat
 from .tensor import TensorPurpose
 from .tensor import TensorSubPurpose
-from .numeric_util import full_shape
+
 
 class ParetoMetric(enum.Enum):
     BwCycMem = 1
@@ -652,6 +653,9 @@ class DynamicProgrammingScheduler:
         for op in pred_candidate.ops:
             if op.type == "ConcatSliceWrite":
                 return True
+            if len(op.outputs) > 1 or len(op.outputs[0].consumer_list) > 1:
+                # The op has consumers in other subgraphs
+                return True
         return False
 
     def search_ifm_streaming_partial(self, ps, block_config):
@@ -976,8 +980,15 @@ class DynamicProgrammingScheduler:
                                 # be processed by CPU operations. No-op reshape consumers with empty lists
                                 # (those that have no consumers, or null-consumers used as list terminators)
                                 # must use normal NHWC output.
-                                incompatible_consumers = [ (not consumer.run_on_npu or consumer.type == "Reshape" or (consumer is last_op_in_subgraph))
-                                                           for consumer in op.outputs[0].consumer_list if consumer is not None ]
+                                incompatible_consumers = [
+                                    (
+                                        not consumer.run_on_npu
+                                        or consumer.type == "Reshape"
+                                        or (consumer is last_op_in_subgraph)
+                                    )
+                                    for consumer in op.outputs[0].consumer_list
+                                    if consumer is not None
+                                ]
                                 if (outshape == inshape) and incompatible_consumers and not any(incompatible_consumers):
                                     rewrites.append(op)
                                 else: