commit d03dc504452dbb32db383121f3dad81f4280bb3c
author Fredrik Svedberg <fredrik.svedberg@arm.com> Thu Jun 30 10:44:12 2022 +0200
committer Fredrik Svedberg <fredrik.svedberg@arm.com> Wed Jul 13 15:29:38 2022 +0000
tree a491e706d5d18a099116b30b8d924667ff0f7245
parent a04f2f7322e7b83d93e875313d2e5b4d0dca94fb

MLBEDSW-6496 mlperf_deeplabv3_mnv2_ade20k_int8 fails at verify_output for u65

Added check to see if additional stripe data is needed from producer op
when cascading to make sure the stripes are not overwriting data still
being used. Also changed scheduler to make sure ResizeBilinear always
runs with even stripe height.

Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>
Change-Id: If7d723e6be29575c2b55c400eebbe8275a1aa328

ethosu/vela/high_level_command_stream_generator.py

diff --git a/ethosu/vela/high_level_command_stream_generator.py b/ethosu/vela/high_level_command_stream_generator.py
index 9506808..a52bdc3 100644
--- a/ethosu/vela/high_level_command_stream_generator.py
+++ b/ethosu/vela/high_level_command_stream_generator.py
@@ -185,13 +185,14 @@
                 if producer_op:
                     assert op_info.cascade != 0
                     assert op_info.cascade == schedule.cost_map[producer_op].cascade
-                    for prev_cmd in prev_cmd_gen:
-                        yield prev_cmd
-                        if prev_cmd.is_npu_pass_command() and prev_cmd.ps == producer_op.parent_ps:
-                            ifm_present.end_coord = prev_cmd.ofm_box.end_coord
-                            if ifm_required.is_subbox_of(ifm_present):
-                                # There is enough IFM data - exit loop
-                                break
+                    if not ifm_required.is_subbox_of(ifm_present):
+                        for prev_cmd in prev_cmd_gen:
+                            yield prev_cmd
+                            if prev_cmd.is_npu_pass_command() and prev_cmd.ps == producer_op.parent_ps:
+                                ifm_present.end_coord = prev_cmd.ofm_box.end_coord
+                                if ifm_required.is_subbox_of(ifm_present):
+                                    # There is enough IFM data - exit loop
+                                    break
 
                 # Information about the current stripe's location in the cascade
                 is_first_h_stripe = ofm_box_start.height == ofm_start.height

ethosu/vela/scheduler.py

diff --git a/ethosu/vela/scheduler.py b/ethosu/vela/scheduler.py
index 7e989a7..d01942b 100644
--- a/ethosu/vela/scheduler.py
+++ b/ethosu/vela/scheduler.py
@@ -44,6 +44,7 @@
 from .architecture_allocator import ArchitectureBlockConfig
 from .architecture_allocator import find_block_config
 from .architecture_allocator import get_ifm_area_required
+from .architecture_allocator import to_upscale
 from .architecture_features import ArchitectureFeatures
 from .architecture_features import Block
 from .cascade_builder import CascadeBuilder
@@ -906,7 +907,8 @@
             striped_schedule.cost_map[sched_op] = cost
 
             # Calculate the preceeding Op's stripe
-            stripe = sched_op.ifm.shape.with_height(stripe.height * sched_op.kernel.stride.y)
+            height = stripe.height + stripe.height % to_upscale(sched_op.resampling_mode)
+            stripe = sched_op.ifm.shape.with_height(height * sched_op.kernel.stride.y)
 
         return striped_schedule