MLBEDSW-4602: Fix Deepspeech scale & bias reuse issue.

 - Deepspeech reuses identical weights and biases throughout
   the network. Since biases are now interleaved with weights
   there is a scaling issue when the ifm scales differ between
   operations using the same weight and scale tensor.

 - This commit uses interleaved weights/scales on their first use
   but separates scales to source memory on subsequent use (if
   the ifm scale is different).

Signed-off-by: Tim Hall <tim.hall@arm.com>
Change-Id: I7aae163438160a919cae04e235966e75355a6148

1 files changed, 4 insertions, 6 deletions

diff --git a/ethosu/vela/live_range.py b/ethosu/vela/live_range.py
index d75a167d..b687a9e7 100644
--- a/ethosu/vela/live_range.py
+++ b/ethosu/vela/live_range.py
@@ -344,16 +344,14 @@ def create_linear_live_range_graph(sg, target_mem_area, target_mem_type_set, lr_
                 lr_graph, tens, target_mem_area, target_mem_type_set
             ):
                 continue
-
             rng = lr_graph.get_or_create_range(tens)
             rng.mark_usage(sg_time)
 
     for sched_op, op_info in sg.schedule.cost_map.items():
-        if op_info.npu_weights_tensor and not (
-            tensor_should_be_ignored(lr_graph, op_info.npu_weights_tensor, target_mem_area, target_mem_type_set)
-        ):
-            rng = lr_graph.get_or_create_range(op_info.npu_weights_tensor)
-            rng.mark_usage(sg_time)
+        for tensor in [op_info.npu_weights_tensor, op_info.npu_scales_tensor]:
+            if tensor and not (tensor_should_be_ignored(lr_graph, tensor, target_mem_area, target_mem_type_set)):
+                rng = lr_graph.get_or_create_range(tensor)
+                rng.mark_usage(sg_time)
 
     lr_graph.current_time += 1
     return lr_graph