ekalda commented on a change in pull request #9626:
URL: https://github.com/apache/tvm/pull/9626#discussion_r761199497
##########
File path: python/tvm/relay/backend/contrib/ethosu/legalize.py
##########
@@ -1156,6 +1156,7 @@ def transform_module(
mod = LegalizeReshape()(mod)
mod = LegalizeStridedSlice()(mod)
mod = LegalizeNoOps()(mod)
+ print(mod)
Review comment:
Nit: remove the print
##########
File path: python/tvm/relay/backend/contrib/ethosu/legalize.py
##########
@@ -1040,7 +1040,7 @@ def callback(
n = int(filter_height * filter_width)
eps = 1 / (256 * (n + 1)) if n % 2 == 0 else 0
- scalar_tensor = relay.const(np.ones([1, 1, 1, 1], dtype="uint8"),
dtype="uint8")
+ scalar_tensor = relay.const(np.ones([1, 1, 1, 1], dtype="int16"),
dtype="int16")
Review comment:
For enlightenment, why `int16` (as opposed to e.g. `int32`)?
##########
File path: python/tvm/relay/backend/contrib/ethosu/tir_to_cs_translator.py
##########
@@ -277,27 +276,24 @@ def classify_io(buffer):
raise ValueError(f"Unused IO : {buffer} in tir module.")
scratch_size = 0
- constant_tensor = None
+ constant_hex_data = []
+ total_constant_len = 0
buffer_addresses = dict()
for _buffer, info in buffer_info.items():
if info.values is not None:
- assert np.dtype(info.dtype) == np.uint8
assert info.btype == BufferType.constant
assert len(info.shape) == 1
- if constant_tensor is None:
- buffer_addresses[_buffer] = (0, info.btype)
- assert info.values.dtype == np.uint8
- size_in_bytes = info.values.size
- # Every memory address the NPU access have to be 16 byte
aligned
- size_in_bytes = util.round_up(size_in_bytes, 16)
- constant_tensor = np.resize(info.values, size_in_bytes)
- else:
- buffer_addresses[_buffer] = (constant_tensor.size, info.btype)
- assert info.values.dtype == np.uint8
- size_in_bytes = info.values.size
- # Every memory address the NPU access have to be 16 byte
aligned
- size_in_bytes = util.round_up(size_in_bytes, 16)
- constant_tensor = np.append(constant_tensor,
np.resize(info.values, size_in_bytes))
+ buffer_addresses[_buffer] = (
+ (total_constant_len, info.btype) if constant_hex_data else (0,
info.btype)
+ )
+ dtype_bytes = np.iinfo(np.dtype(info.dtype)).bits // 8
+ size_in_bytes = dtype_bytes * np.prod(list(info.shape))
+ # Every memory address the NPU access have to be 16 byte aligned
+ size_in_bytes = util.round_up(size_in_bytes, 16)
+ constant_tensor = np.resize(info.values, size_in_bytes //
dtype_bytes)
+ constant_tensor = constant_tensor.tobytes().hex()
+ constant_hex_data.append(constant_tensor)
+ total_constant_len += len(constant_tensor) // 2
Review comment:
For enlightenment, why division by 2?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
