csullivan commented on code in PR #11066: URL: https://github.com/apache/tvm/pull/11066#discussion_r859213141
########## python/tvm/utils/__init__.py: ########## @@ -0,0 +1,303 @@ +# 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. +"""Utilities operating at a graph/model or other "high" level""" +import csv +import subprocess +from typing import Dict, Union, Optional +import numpy as np + +from .. import auto_scheduler, relay, tir, device, nd, IRModule, build, topi, transform +from ..target import Target +from ..runtime import profiler_vm, profiling, Device, num_threads +from ..script import tir as T + + +def _create_args(mod, dev, func_name="main"): + args = [] + for arg in mod[func_name].params: + args.append( + nd.array( + np.zeros([x.value for x in arg.type_annotation.shape], arg.type_annotation.dtype), + device=dev, + ) + ) + return args + + +def _estimated_features(mod, params, target): + comp = relay.vm.VMCompiler() + mod, params = comp.optimize(mod, params=params, target=target) + return { + prim.attrs["hash"]: (name, auto_scheduler.feature.named_features_from_primfunc(prim)) + for name, prim in mod.functions.items() + if isinstance(prim, tir.PrimFunc) + } + + +def _vec_width_registers(target, vec_width, num_vector_registers): + if vec_width is None: + if target.device_name == "": # indicates x86 Review Comment: Yeah, having an empty device name is more just a special case for llvm. > target = tvm.target.Target(" > llvm -keys=hexagon,cpu \ > -link-params=0 -mattr=+hvxv68, \ > **+hvx-length128b**, +hvx-qfloat, -hvx-ieee-fp \ > -mcpu=hexagonv68 -mtriple=hexagon > ") will also have an empty device name similar to Andrew's comment. I think we need a different way to dispatch here. In general I'd like to see this information queried as attrs from the target which, when not provided, can be queried from the device via an attr [preprocessor](https://github.com/apache/tvm/blob/main/src/target/target_kind.cc#L295) used to update the attributes for a target kind -- which can [query](https://github.com/apache/tvm/blob/main/src/target/target_kind.cc#L114) the local or remote device api for the necessary information. -- 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]
