MLBEDSW-1941: Bug fix shared weights
If same weight tensor was used with different block configs,
errors would occur.
Fixed by always cloning weight tensors, using a global weight
compression cache and modifying the linear allocator to
detect multiple usage of same weight compression.
Change-Id: I91ca59176e1c59c66e0ac7a4227f2b5f0b47053f
Signed-off-by: Louis Verhaard <louis.verhaard@arm.com>
diff --git a/ethosu/vela/tensor_allocation.py b/ethosu/vela/tensor_allocation.py
index cd2b570..e3952df 100644
--- a/ethosu/vela/tensor_allocation.py
+++ b/ethosu/vela/tensor_allocation.py
@@ -27,18 +27,26 @@
from .tensor import MemArea
-def linear_allocate_live_ranges(live_ranges, alloc_granularity=256):
+def linear_allocate_live_ranges(live_ranges, alloc_granularity=16):
+ # Allocates using increasing addresses. Duplicate constant tensors will be allocated to the same address
total_sz = 0
allocated_tensors = []
- # just assign increasing addresses
+ # just assign increasing addresses, except for duplicates
for tens, lr in live_ranges.ranges.items():
if tens in allocated_tensors:
continue
- lr.set_address(total_sz)
+ address = total_sz
+ if tens.weight_compression_config is not None:
+ for allocated_tens in allocated_tensors:
+ if allocated_tens.weight_compression_config == tens.weight_compression_config:
+ address = allocated_tens.address
+ break
+ lr.set_address(address)
allocated_tensors += lr.tensors
- total_sz += numeric_util.round_up(int(math.ceil(lr.size)), alloc_granularity)
+ if address == total_sz:
+ total_sz += numeric_util.round_up(int(math.ceil(lr.size)), alloc_granularity)
return total_sz