jackwish commented on a change in pull request #4307: [QNN][Legalize]
Specialize for Platforms w/o fast Int8 support
URL: https://github.com/apache/incubator-tvm/pull/4307#discussion_r345586389
##########
File path: python/tvm/relay/qnn/op/legalizations.py
##########
@@ -137,4 +165,124 @@ def _is_int8_hw_support(target):
new_attrs = {k : attrs[k] for k in attrs.keys()}
new_attrs['input_zero_point'] = input_zp
new_attrs['kernel_zero_point'] = kernel_zp
- return relay.qnn.op.conv2d(data, kernel, **new_attrs)
+ return relay_op(data, kernel, **new_attrs)
+
+# Helper function to change dtypes to be same. ARM dotprod instructions prefer
this setting.
+def helper_change_dtypes_to_be_same(attrs, inputs, types, relay_op):
+ """ Sometimes MxNet + MLDNN can lead to uint8 x int8 datatypes for the
conv inputs. However,
+ many devices like ARM prefer the datatypes to be same for the HW units.
This helper transforms
+ conv2d/dense such that both the dtypes are same.
+
+ Parameters
+ ----------
+ attrs : tvm.attrs.Attrs
+ Attributes of current convolution
+ inputs : list of tvm.relay.Expr
+ The args of the Relay expr to be legalized
+ types : list of types
+ List of input and output types
+
+ Returns
+ -------
+ result : tvm.relay.Expr
+ The legalized expr
+ """
+
+ def _shift(data, out_dtype):
+ """Shifts (add/subtracts) the qnn tensor with +/-128)"""
Review comment:
I guess this is *add or subtract with 128* :)
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
