X86 3x performance difference NNVM vs Relay

deimsdeutsch · September 15, 2020, 8:05am

I have read previous posts and i understand that relay does not use avx2. Is this resolved as i have compared performance with various different parameters and the results are always consistent.

tqchen · September 15, 2020, 2:12pm

i don;t think there will be any limitations in terms of avx2. As of now we always recommend relay over nnvm

deimsdeutsch · September 15, 2020, 3:57pm

Thanks for replying.

Build Configuration: TVM github build LLVM 6.0.1 Openblas Centos 8.2

target = tvm.target.create(“llvm -mcpu=core-avx2”)

This means i am missing something in the relay tuning script. Please suggest me the changes i can make in following script to improve performance.

github.com

apache/incubator-tvm/blob/master/tutorials/autotvm/tune_relay_x86.py

# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements.  See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership.  The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License.  You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied.  See the License for the
# specific language governing permissions and limitations
# under the License.
"""
.. _tune_relay_x86:

Auto-tuning a convolutional network for x86 CPU

This file has been truncated. show original

tqchen · September 15, 2020, 4:28pm

It would be helpful if you can dig deeper, e.g. what is the model you observed perf and look into the generated code. It is hard to tell as it is what could be wrong