MLBEDSW-2822 Account for NHCWB16 in scheduler est.
Added that NHCWB16 is accounted for in the sram estimates
in the scheduler, for intermediate buffers in ifm streaming.
Signed-off-by: Patrik Gustavsson <patrik.gustavsson@arm.com>
Change-Id: Icda5e05dd3663935f528f1a06d36d9e1de123cc8
diff --git a/ethosu/vela/tensor.py b/ethosu/vela/tensor.py
index c41a7eb..66bed59 100644
--- a/ethosu/vela/tensor.py
+++ b/ethosu/vela/tensor.py
@@ -461,6 +461,12 @@
if sub_purpose == TensorSubPurpose.DoubleBuffer:
raw_size = elems * self.element_size() * self.compression_scale_for_worst_weight_stream
else:
+ # Rolling buffers are used for intermediate data in ifm streaming
+ # These will all use the NHCWB16 format, and need to be aligned to 16 in the C-dimension
+ if alt_shape[-1] % 16 != 0:
+ nhcwb16_shape = alt_shape[0:-1] + [numeric_util.round_up(alt_shape[-1], 16)]
+ elems = shape_num_elements(nhcwb16_shape)
+
raw_size = elems * self.element_size() * self.storage_compression_scale
rounded_size = numeric_util.round_up(numeric_util.round_up_to_int(raw_size), self.alignment)
return rounded_size