bgawrych commented on a change in pull request #19587:
URL: https://github.com/apache/incubator-mxnet/pull/19587#discussion_r536238061
##########
File path: src/operator/subgraph/mkldnn/mkldnn_conv.cc
##########
@@ -137,17 +147,22 @@ void SgMKLDNNConvOperator::Forward(const OpContext &ctx,
auto in_beta = mkldnn_param.with_bn ? (idx++) : 0;
auto in_mean = mkldnn_param.with_bn ? (idx++) : 0;
auto in_var = mkldnn_param.with_bn ? (idx++) : 0;
- auto in_sum = mkldnn_param.with_sum ? (idx++) : 0;
+ auto in_sum = mkldnn_param.with_sum ? (mkldnn_param.dedup_sum? in_data :
idx++) : -1;
float data_min =
mkldnn_param.quantized ? inputs[idx++].data().dptr<float>()[0] : 0.0;
float data_max =
mkldnn_param.quantized ? inputs[idx++].data().dptr<float>()[0] : 0.0;
- float sum_min = (mkldnn_param.with_sum && mkldnn_param.quantized)
- ? inputs[idx++].data().dptr<float>()[0]
- : 0.0;
- float sum_max = (mkldnn_param.with_sum && mkldnn_param.quantized)
- ? inputs[idx++].data().dptr<float>()[0]
- : 0.0;
+ float sum_min = 0.0f;
Review comment:
Subgraph API provide deduplication of inputs to subgraph (e.g. residual
block in mobilenet)
Test for this is in tests/python/mkl/test_subgraph.py::test_dedup
----------------------------------------------------------------
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]
