MLBEDSW-4034: New Scheduler Size or Performance Optimisation

 - Merged dev/scheduler at 83639f90e8c828f70de6e29142355a940224959b

Signed-off-by: Tim Hall <tim.hall@arm.com>
Change-Id: I0050529d4b42da93768c7264296434dd877fb5b4

1 files changed, 32 insertions, 7 deletions

diff --git a/ethosu/vela/tensor_allocation.py b/ethosu/vela/tensor_allocation.py
index 724c7c0d..d3e2a037 100644
--- a/ethosu/vela/tensor_allocation.py
+++ b/ethosu/vela/tensor_allocation.py
@@ -106,6 +106,8 @@ def verify_allocation(live_ranges: LiveRangeGraph, alignment: int):
 
 
 def mark_sram_used_for_cascaded_passes(sg, lrs):
+    if len(sg.cascaded_passes) < 1:
+        return
     end_pos = max(ps.time for ps in sg.cascaded_passes) + 2
     mem_usage = np.zeros(end_pos, dtype=np.int64)
 
@@ -169,18 +171,14 @@ def memory_usage_histogram(lrs: List[LiveRange]):
     return histogram
 
 
-def allocate_tensors(
-    nng,
+def allocate(
     sg,
     arch,
     mem_area,
     mem_type_set,
     tensor_allocator=TensorAllocator.Greedy,
-    verbose_allocation=False,
     lr_graph=None,
     cpu_tensor_alignment=Tensor.AllocationQuantum,
-    max_size=None,
-    dry_test=False,
 ):
     # Allocates addresses to tensors, returns False if tensors could not be fit within max_size
     ignore_subgraph_input_output_tensors = False
@@ -192,7 +190,7 @@ def allocate_tensors(
         lr_graph=lr_graph,
         cpu_tensor_alignment=cpu_tensor_alignment,
     )
-
+    total_sz = 0
     if lrs.ranges:
         tens_alloc = tensor_allocator
         if tens_alloc == TensorAllocator.Greedy:
@@ -204,6 +202,34 @@ def allocate_tensors(
             total_sz = hillclimb_allocate_live_ranges(lrs, cpu_tensor_alignment)
         else:
             assert 0
+    return lrs, total_sz
+
+
+def allocate_tensors(
+    nng,
+    sg,
+    arch,
+    mem_area,
+    mem_type_set,
+    tensor_allocator=TensorAllocator.Greedy,
+    verbose_allocation=False,
+    lr_graph=None,
+    cpu_tensor_alignment=Tensor.AllocationQuantum,
+    max_size=None,
+    dry_test=False,
+):
+    # Allocates addresses to tensors, returns False if tensors could not be fit within max_size
+    lrs, total_sz = allocate(
+        sg,
+        arch,
+        mem_area,
+        mem_type_set,
+        tensor_allocator=tensor_allocator,
+        lr_graph=lr_graph,
+        cpu_tensor_alignment=cpu_tensor_alignment,
+    )
+
+    if lrs.ranges:
         alloc_ok = max_size is None or total_sz <= max_size
         if dry_test or not alloc_ok:
             # Dry test or allocation failed; undo allocation
@@ -233,5 +259,4 @@ def allocate_tensors(
 
     if sg == nng.get_root_subgraph():
         nng.memory_used = sg.memory_used
-
     return True