MLBEDSW-3212 Remove CLI opt ifm-ofm-overlap

Removed the CLI opt ifm-ofm-overlap

Signed-off-by: Patrik Gustavsson <patrik.gustavsson@arm.com>
Change-Id: I23faa0d10c3e71972c543e22e8155086fce73556

7 files changed, 6 insertions, 55 deletions

diff --git a/OPTIONS.md b/OPTIONS.md
index 07cf78e..a7b513c 100644
--- a/OPTIONS.md
+++ b/OPTIONS.md
@@ -72,18 +72,6 @@ system's SRAM this optimisation is required.
 vela network.tflite --cascading False
 ```
 
-### IFM/OFM Overlap
-
-Controls the overlapping of IFM and OFM buffers. This means that IFM and OFM
-buffers may overlap if possible. This allows for lower memory usage.  
-**Type: Boolean**  
-**Default: True**  
-
-```bash
-vela network.tflite --ifm-ofm-overlap False
-```
-
-
 ### Force Block Config
 
 Force a specific block configuration in the format HxWxC, where H, W, and C are
diff --git a/ethosu/vela/compiler_driver.py b/ethosu/vela/compiler_driver.py
index 1d7521b..9263305 100644
--- a/ethosu/vela/compiler_driver.py
+++ b/ethosu/vela/compiler_driver.py
@@ -194,7 +194,6 @@ def compiler_driver(nng, arch, options, scheduler_options):
             arch,
             permanent_storage,
             set((MemType.Permanent_NPU,)),
-            use_ifm_ofm_overlap=scheduler_options.use_ifm_ofm_overlap,
             tensor_allocator=TensorAllocator.LinearAlloc,
             verbose_allocation=options.verbose_allocation,
             show_minimum_possible_allocation=options.show_minimum_possible_allocation,
@@ -232,7 +231,6 @@ def compiler_driver(nng, arch, options, scheduler_options):
                 arch,
                 mem_area,
                 mem_type_set,
-                use_ifm_ofm_overlap=scheduler_options.use_ifm_ofm_overlap,
                 tensor_allocator=options.tensor_allocator,
                 verbose_allocation=options.verbose_allocation,
                 show_minimum_possible_allocation=options.show_minimum_possible_allocation,
@@ -259,7 +257,6 @@ def compiler_driver(nng, arch, options, scheduler_options):
                     mem_type_set,
                     max_size=arch.sram_size,
                     dry_test=dry_test,
-                    use_ifm_ofm_overlap=scheduler_options.use_ifm_ofm_overlap,
                     tensor_allocator=options.tensor_allocator,
                     verbose_allocation=options.verbose_allocation,
                     show_minimum_possible_allocation=options.show_minimum_possible_allocation,
@@ -305,7 +302,6 @@ def compiler_driver(nng, arch, options, scheduler_options):
         arch,
         permanent_storage,
         set((MemType.Permanent_CPU,)),
-        use_ifm_ofm_overlap=scheduler_options.use_ifm_ofm_overlap,
         tensor_allocator=TensorAllocator.LinearAlloc,
         verbose_allocation=options.verbose_allocation,
         show_minimum_possible_allocation=options.show_minimum_possible_allocation,
diff --git a/ethosu/vela/high_level_command_stream_generator.py b/ethosu/vela/high_level_command_stream_generator.py
index 3e3cda1..01fab0e 100644
--- a/ethosu/vela/high_level_command_stream_generator.py
+++ b/ethosu/vela/high_level_command_stream_generator.py
@@ -394,7 +394,3 @@ def calc_allowed_ofm_ifm_overlap_for_pass_list(strat, passes, block_configs):
 
     min_overlap = max(min_overlap, 0)
     return min_overlap
-
-
-def calc_allowed_ofm_ifm_overlap_for_cascaded_pass(cps):
-    return calc_allowed_ofm_ifm_overlap_for_pass_list(cps.strategy, cps.passes, [ps.block_config for ps in cps.passes])
diff --git a/ethosu/vela/live_range.py b/ethosu/vela/live_range.py
index b884035..a29cafe 100644
--- a/ethosu/vela/live_range.py
+++ b/ethosu/vela/live_range.py
@@ -16,7 +16,6 @@
 # Description:
 # Build a live range graph for tensors in one or more subgraphs. Used for tensor allocation as well as in the scheduler.
 # Can work with either a pass packed subgraph or a scheduled subgraph.
-from .high_level_command_stream_generator import calc_allowed_ofm_ifm_overlap_for_cascaded_pass
 from .nn_graph import PassPlacement
 from .operation import Op
 from .tensor import MemType
@@ -101,7 +100,6 @@ class LiveRange:
 class LiveRangeGraph:
     def __init__(self):
         self.ranges = {}  # tens -> range
-        self.allowed_overlaps = {}  # (tens,tens) -> overlap_int
         self.ignore_tensors = set()
         self.processed_subgraphs = set()
         self.current_time = 0
@@ -198,7 +196,7 @@ def merge_elementwise_op_ranges(ps, lr_graph, target_mem_area, target_mem_type_s
 def extract_live_ranges_from_passes(
     sg,
     target_mem_area,
-    target_mem_type=set((MemType.Scratch, MemType.Scratch_fast)),
+    target_mem_type_set=set((MemType.Scratch, MemType.Scratch_fast)),
     ignore_subgraph_input_output_tensors=False,
 ):
     lr_graph = LiveRangeGraph()
@@ -209,7 +207,7 @@ def extract_live_ranges_from_passes(
 
     # Try to merge live ranges of operations in the NPU subgraphs
     if sg.placement == PassPlacement.Npu:
-        merge_op_ranges(sg, lr_graph, target_mem_area, target_mem_type)
+        merge_op_ranges(sg, lr_graph, target_mem_area, target_mem_type_set)
 
     for idx, ps in enumerate(sg.passes):
         ps.time = 2 * idx
@@ -217,14 +215,14 @@ def extract_live_ranges_from_passes(
         time_for_pass = ps.time
 
         for tens in ps.inputs + ps.intermediates + ps.outputs:
-            if tensor_should_be_ignored(lr_graph, tens, target_mem_area, target_mem_type):
+            if tensor_should_be_ignored(lr_graph, tens, target_mem_area, target_mem_type_set):
                 continue
             rng = lr_graph.get_or_create_range(tens)
             rng.mark_usage(time_for_pass)
 
     end_time = len(sg.passes) * 2
     for tens in sg.output_tensors:
-        if tensor_should_be_ignored(lr_graph, tens, target_mem_area, target_mem_type):
+        if tensor_should_be_ignored(lr_graph, tens, target_mem_area, target_mem_type_set):
             continue
         rng = lr_graph.get_or_create_range(tens)
         rng.mark_usage(end_time)
@@ -236,7 +234,6 @@ def extract_live_ranges_from_cascaded_passes(
     sg,
     target_mem_area,
     target_mem_type_set,
-    use_ifm_ofm_overlap=True,
     ignore_subgraph_input_output_tensors=False,
     lr_graph=None,
     allocation_alignment=Tensor.AllocationQuantum,
@@ -279,7 +276,7 @@ def extract_live_ranges_from_cascaded_passes(
             # Use default allocation alignment of 16 for Npu tensors
             npu_sg = cps_primary_op.attrs["subgraph"]
             lr_graph = extract_live_ranges_from_cascaded_passes(
-                npu_sg, target_mem_area, target_mem_type_set, use_ifm_ofm_overlap, False, lr_graph,
+                npu_sg, target_mem_area, target_mem_type_set, False, lr_graph,
             )
             # Set the new time after handling the Npu subgraph
             time_for_pass = lr_graph.current_time
@@ -291,20 +288,6 @@ def extract_live_ranges_from_cascaded_passes(
             rng = lr_graph.get_or_create_range(tens, allocation_alignment)
             rng.mark_usage(time_for_pass)
 
-        if use_ifm_ofm_overlap:
-            # fill allowed overlap for ifm and ofm tensor
-            ifm_tensor = cps.passes[0].ifm_tensor
-            ofm_tensor = cps.passes[-1].ofm_tensor
-            if (
-                ifm_tensor is not None
-                and ofm_tensor is not None
-                and not tensor_should_be_ignored(lr_graph, ifm_tensor, target_mem_area, target_mem_type_set)
-                and not tensor_should_be_ignored(lr_graph, ofm_tensor, target_mem_area, target_mem_type_set)
-            ):
-                lr_graph.allowed_overlaps[(ifm_tensor, ofm_tensor)] = calc_allowed_ofm_ifm_overlap_for_cascaded_pass(
-                    cps
-                )
-
         lr_graph.current_time += 2
 
     end_time = 0
diff --git a/ethosu/vela/scheduler.py b/ethosu/vela/scheduler.py
index 31e6383..59c2b58 100644
--- a/ethosu/vela/scheduler.py
+++ b/ethosu/vela/scheduler.py
@@ -59,7 +59,6 @@ class SchedulerOptions:
     def __init__(
         self,
         use_cascading=True,
-        use_ifm_ofm_overlap=True,
         verbose_schedule=False,
         verbose_pareto_frontier_schedules=False,
         use_ifm_streaming=True,
@@ -67,7 +66,6 @@ class SchedulerOptions:
         use_nhcwb16_between_cascaded_passes=True,
     ):
         self.use_cascading = use_cascading
-        self.use_ifm_ofm_overlap = use_ifm_ofm_overlap
         self.verbose_schedule = verbose_schedule
         self.verbose_pareto_frontier_schedules = verbose_pareto_frontier_schedules
         self.use_ifm_streaming = use_ifm_streaming
@@ -236,7 +234,7 @@ class DynamicProgrammingScheduler:
         if self.arch.feature_map_storage_mem_area != MemArea.Sram:
             self.use_ifm_ofm_overlap = False  # force off IFM/OFM overlap if IFMs and OFMs are not in the SRAM
         else:
-            self.use_ifm_ofm_overlap = options.use_ifm_ofm_overlap
+            self.use_ifm_ofm_overlap = True
 
         self.verbose_schedule = options.verbose_schedule
         self.verbose_pareto_frontier_schedules = options.verbose_pareto_frontier_schedules
diff --git a/ethosu/vela/tensor_allocation.py b/ethosu/vela/tensor_allocation.py
index 9f14ec4..8329a61 100644
--- a/ethosu/vela/tensor_allocation.py
+++ b/ethosu/vela/tensor_allocation.py
@@ -122,7 +122,6 @@ def allocate_tensors(
     arch,
     mem_area,
     mem_type_set,
-    use_ifm_ofm_overlap=True,
     tensor_allocator=TensorAllocator.Greedy,
     verbose_allocation=False,
     show_minimum_possible_allocation=False,
@@ -137,7 +136,6 @@ def allocate_tensors(
         sg,
         mem_area,
         mem_type_set,
-        use_ifm_ofm_overlap=use_ifm_ofm_overlap,
         ignore_subgraph_input_output_tensors=ignore_subgraph_input_output_tensors,
         lr_graph=lr_graph,
         allocation_alignment=allocation_alignment,
diff --git a/ethosu/vela/vela.py b/ethosu/vela/vela.py
index b9e224c..6d54187 100644
--- a/ethosu/vela/vela.py
+++ b/ethosu/vela/vela.py
@@ -158,13 +158,6 @@ def main(args=None):
         choices=[True, False],
         help="Controls the packing of multiple passes into a cascade (default: %(default)s)",
     )
-    parser.add_argument(
-        "--ifm-ofm-overlap",
-        type=ast.literal_eval,
-        default=True,
-        choices=[True, False],
-        help="Controls the overlapping of IFM and OFM buffers (default: %(default)s)",
-    )
     parser.add_argument("--force-block-config", type=str, default="", help="Force a specific block configuration HxWxC")
     parser.add_argument("--timing", action="store_true", help="Time the compiler doing operations")
     parser.add_argument(
@@ -313,7 +306,6 @@ def main(args=None):
 
     scheduler_options = scheduler.SchedulerOptions(
         use_cascading=args.cascading,
-        use_ifm_ofm_overlap=args.ifm_ofm_overlap,
         verbose_schedule=args.verbose_schedule,
         verbose_pareto_frontier_schedules=args.verbose_pareto_frontier_schedules,
         use_ifm_streaming=args.ifm_streaming,