diff options
| author | Louis Verhaard <louis.verhaard@arm.com> | 2020-05-07 08:12:58 +0200 |
|---|---|---|
| committer | Tim Hall <tim.hall@arm.com> | 2020-06-18 17:53:52 +0100 |
| commit | 3c07c97e0202c1cf01eba06c24b37a8f15ff7a7c (patch) | |
| tree | 5856b7727a99b3c0baa00f5486f0c3b53e8e38e6 /ethosu/vela/tensor_allocation.py | |
| parent | 86d49935c3736c7aaa419abda07fa20c37c991a8 (diff) | |
| download | ethos-u-vela-3c07c97e0202c1cf01eba06c24b37a8f15ff7a7c.tar.gz | |
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>
Diffstat (limited to 'ethosu/vela/tensor_allocation.py')
| -rw-r--r-- | ethosu/vela/tensor_allocation.py | 16 |
1 files changed, 12 insertions, 4 deletions
diff --git a/ethosu/vela/tensor_allocation.py b/ethosu/vela/tensor_allocation.py index cd2b570f..e3952df3 100644 --- a/ethosu/vela/tensor_allocation.py +++ b/ethosu/vela/tensor_allocation.py @@ -27,18 +27,26 @@ from .nn_graph import TensorAllocator 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 |