icemelon9 commented on a change in pull request #7404:
URL: https://github.com/apache/tvm/pull/7404#discussion_r570606734
##########
File path: src/relay/op/nn/nn.h
##########
@@ -88,6 +90,29 @@ bool DenseRel(const Array<Type>& types, int num_inputs,
const Attrs& attrs,
return true;
}
+template <typename AttrType>
Review comment:
should we keep this function in `nn.cc`?
##########
File path: python/tvm/relay/op/strategy/x86.py
##########
@@ -364,14 +364,13 @@ def conv1d_strategy_cpu(attrs, inputs, out_type, target):
def dense_strategy_cpu(attrs, inputs, out_type, target):
"""dense x86 strategy"""
strategy = _op.OpStrategy()
- m, _ = inputs[0].shape
same_type = inputs[0].dtype == inputs[1].dtype == out_type.dtype
dtype = inputs[0].dtype
u8s8s32 = dtype == "uint8" and inputs[1].dtype == "int8" and
out_type.dtype == "int32"
strategy.add_implementation(
- wrap_compute_dense(topi.x86.dense_nopack),
- wrap_topi_schedule(topi.x86.schedule_dense_nopack),
- name="dense_nopack.x86",
+ wrap_compute_dense(topi.x86.dense_pack),
+ wrap_topi_schedule(topi.x86.schedule_dense_pack),
+ name="dense_pack.x86",
Review comment:
You could add both implementation but make `dense_pack` with higher
plevel.
----------------------------------------------------------------
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]
