You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2021/03/17 03:39:01 UTC
[GitHub] [tvm] altanh opened a new pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
altanh opened a new pull request #7677:
URL: https://github.com/apache/tvm/pull/7677
Since there is negligible code sharing between the first-order and higher-order AD in Relay, I've factored out the first-order AD pass to a separate file. Additionally, I've made it a proper IRModule pass, and fixed a few spots where adding tuples wasn't being lifted correctly. I tried my best to add some diagnostics.
To complement AD, I've also added a pass called `ConcretizeLike` that transforms the `*_like` operators generated during AD to the explicit-shape equivalent (e.g. `zeros_like(x, y) -> zeros(x, y.shape)`), which removes some forward dependencies and could enable more opportunity for operator fusion. Note that this is the same thing as `LikeZapp` from @t-vi's [blogpost](https://lernapparat.de/transformers-pytorch-tvm/) with perhaps a bit more error checking.
cc @t-vi @MarisaKirisame @tqchen @jroesch
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] tqchen merged pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass
Posted by GitBox <gi...@apache.org>.
tqchen merged pull request #7677:
URL: https://github.com/apache/tvm/pull/7677
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] altanh commented on a change in pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
altanh commented on a change in pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#discussion_r596285634
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
+
+ data_pat_ = IsWildcard();
+ like_pat_ = IsWildcard();
+ unary_like_pat_ = (IsOp("zeros_like") || IsOp("ones_like"))({like_pat_});
+ binary_like_pat_ = (IsOp("reshape_like") || IsOp("collapse_sum_like") ||
+ IsOp("broadcast_to_like"))({data_pat_, like_pat_});
+ }
+
+ Expr Callback(const Expr& pre, const Expr& post,
+ const Map<DFPattern, Array<Expr>>& node_map) const {
+ // we will rewrite iff the like argument has fully concrete shape
+ const CallNode* call_node = post.as<CallNode>();
+ ICHECK(call_node);
+ const OpNode* op_node = call_node->op.as<OpNode>();
+ ICHECK(op_node);
+ const Op op_ref = GetRef<Op>(op_node);
+ ICHECK(concrete_map_.count(op_ref) > 0);
+
+ Expr like = node_map[like_pat_][0];
+
+ if (!like->checked_type_.defined()) {
+ // TODO(@altanh): maybe because of the input being rewritten?
+ return post;
+ }
+
+ // skip trying to support this for now (ironic, as I was the one who added the feature)
+ if (const auto* attrs = call_node->attrs.as<ReshapeLikeAttrs>()) {
+ if (attrs->lhs_begin != 0 || attrs->rhs_begin != 0 || attrs->lhs_end.defined() ||
+ attrs->rhs_end.defined()) {
+ return post;
+ }
+ }
Review comment:
yeah I agree, thanks
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] altanh commented on a change in pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
altanh commented on a change in pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#discussion_r596285172
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
+
+ data_pat_ = IsWildcard();
+ like_pat_ = IsWildcard();
+ unary_like_pat_ = (IsOp("zeros_like") || IsOp("ones_like"))({like_pat_});
+ binary_like_pat_ = (IsOp("reshape_like") || IsOp("collapse_sum_like") ||
+ IsOp("broadcast_to_like"))({data_pat_, like_pat_});
Review comment:
good points, thanks
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] tqchen commented on pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass
Posted by GitBox <gi...@apache.org>.
tqchen commented on pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#issuecomment-801955929
Thanks @altanh @comaniac @MarisaKirisame
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] MarisaKirisame commented on a change in pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
MarisaKirisame commented on a change in pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#discussion_r595693152
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
Review comment:
```suggestion
auto attrs = make_object<InitOpAttrs>();
attrs->shape = shape;
auto cshape =
MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
```
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
+
+ data_pat_ = IsWildcard();
+ like_pat_ = IsWildcard();
+ unary_like_pat_ = (IsOp("zeros_like") || IsOp("ones_like"))({like_pat_});
+ binary_like_pat_ = (IsOp("reshape_like") || IsOp("collapse_sum_like") ||
+ IsOp("broadcast_to_like"))({data_pat_, like_pat_});
+ }
+
+ Expr Callback(const Expr& pre, const Expr& post,
+ const Map<DFPattern, Array<Expr>>& node_map) const {
+ // we will rewrite iff the like argument has fully concrete shape
+ const CallNode* call_node = post.as<CallNode>();
+ ICHECK(call_node);
+ const OpNode* op_node = call_node->op.as<OpNode>();
+ ICHECK(op_node);
+ const Op op_ref = GetRef<Op>(op_node);
+ ICHECK(concrete_map_.count(op_ref) > 0);
+
+ Expr like = node_map[like_pat_][0];
+
+ if (!like->checked_type_.defined()) {
+ // TODO(@altanh): maybe because of the input being rewritten?
+ return post;
+ }
+
+ // skip trying to support this for now (ironic, as I was the one who added the feature)
+ if (const auto* attrs = call_node->attrs.as<ReshapeLikeAttrs>()) {
+ if (attrs->lhs_begin != 0 || attrs->rhs_begin != 0 || attrs->lhs_end.defined() ||
+ attrs->rhs_end.defined()) {
+ return post;
+ }
+ }
+
+ CHECK(like->checked_type_.defined())
+ << "ConcretizeLike requires checked types to be populated, please run type inference";
+ const TensorTypeNode* like_ty = like->checked_type().as<TensorTypeNode>();
+ ICHECK(like_ty) << "got non-Tensor argument type " << PrettyPrint(like->checked_type());
+
+ Array<Integer> cshape;
+ for (const auto& dim : like_ty->shape) {
+ if (const auto* imm = dim.as<IntImmNode>()) {
+ cshape.push_back(Integer(GetRef<IntImm>(imm)));
+ continue;
+ }
+ return post;
+ }
+
+ if (call_node->args.size() == 2) {
+ return concrete_map_.at(op_ref)(node_map[data_pat_][0], cshape, like_ty->dtype);
+ }
+ return concrete_map_.at(op_ref)(Expr(), cshape, like_ty->dtype);
Review comment:
why empty Expr()?
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
+
+ data_pat_ = IsWildcard();
+ like_pat_ = IsWildcard();
+ unary_like_pat_ = (IsOp("zeros_like") || IsOp("ones_like"))({like_pat_});
+ binary_like_pat_ = (IsOp("reshape_like") || IsOp("collapse_sum_like") ||
+ IsOp("broadcast_to_like"))({data_pat_, like_pat_});
+ }
+
+ Expr Callback(const Expr& pre, const Expr& post,
+ const Map<DFPattern, Array<Expr>>& node_map) const {
+ // we will rewrite iff the like argument has fully concrete shape
+ const CallNode* call_node = post.as<CallNode>();
+ ICHECK(call_node);
+ const OpNode* op_node = call_node->op.as<OpNode>();
+ ICHECK(op_node);
+ const Op op_ref = GetRef<Op>(op_node);
+ ICHECK(concrete_map_.count(op_ref) > 0);
+
+ Expr like = node_map[like_pat_][0];
+
+ if (!like->checked_type_.defined()) {
+ // TODO(@altanh): maybe because of the input being rewritten?
+ return post;
+ }
+
+ // skip trying to support this for now (ironic, as I was the one who added the feature)
+ if (const auto* attrs = call_node->attrs.as<ReshapeLikeAttrs>()) {
+ if (attrs->lhs_begin != 0 || attrs->rhs_begin != 0 || attrs->lhs_end.defined() ||
+ attrs->rhs_end.defined()) {
+ return post;
+ }
+ }
+
+ CHECK(like->checked_type_.defined())
+ << "ConcretizeLike requires checked types to be populated, please run type inference";
+ const TensorTypeNode* like_ty = like->checked_type().as<TensorTypeNode>();
+ ICHECK(like_ty) << "got non-Tensor argument type " << PrettyPrint(like->checked_type());
+
+ Array<Integer> cshape;
+ for (const auto& dim : like_ty->shape) {
+ if (const auto* imm = dim.as<IntImmNode>()) {
+ cshape.push_back(Integer(GetRef<IntImm>(imm)));
+ continue;
+ }
+ return post;
Review comment:
```suggestion
if (const auto* imm = dim.as<IntImmNode>()) {
cshape.push_back(Integer(GetRef<IntImm>(imm)));
} else {
return post;
}
```
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] tqchen commented on pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass
Posted by GitBox <gi...@apache.org>.
tqchen commented on pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#issuecomment-801913378
@MarisaKirisame please take another look and https://tvm.apache.org/docs/contribute/code_review.html#approve-and-request-changes-explicitly
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] altanh commented on a change in pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
altanh commented on a change in pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#discussion_r596288595
##########
File path: tests/python/relay/test_pass_concretize_like.py
##########
@@ -0,0 +1,108 @@
+# 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.
+"""Tests for the ConcretizeLike pass."""
+import tvm
+import tvm.relay.testing
+from tvm import relay
+from tvm.relay.testing import run_infer_type
+
+
+def test_reshape_like():
+ data = relay.var("data", shape=(2, 3, 4), dtype="float32")
+ shape_like = relay.var("shape_like", shape=(6, 2, 2), dtype="float32")
+ f = relay.Function([data, shape_like], relay.reshape_like(data, shape_like))
+ f_expected = relay.Function([data, shape_like], relay.reshape(data, (6, 2, 2)))
Review comment:
this is a good point that I was wondering about while I wrote the tests. However this might be complicated, as removing an parameter means we'll need to remove the argument at each callsite (and then maybe re-apply the optimization until fixed point). I imagine we won't want to remove the parameter from a global function as then the user will have to ensure they don't pass the removed argument.
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] tqchen commented on pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
tqchen commented on pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#issuecomment-801089545
cc @comaniac
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] comaniac commented on a change in pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
comaniac commented on a change in pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#discussion_r596290872
##########
File path: tests/python/relay/test_pass_concretize_like.py
##########
@@ -0,0 +1,108 @@
+# 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.
+"""Tests for the ConcretizeLike pass."""
+import tvm
+import tvm.relay.testing
+from tvm import relay
+from tvm.relay.testing import run_infer_type
+
+
+def test_reshape_like():
+ data = relay.var("data", shape=(2, 3, 4), dtype="float32")
+ shape_like = relay.var("shape_like", shape=(6, 2, 2), dtype="float32")
+ f = relay.Function([data, shape_like], relay.reshape_like(data, shape_like))
+ f_expected = relay.Function([data, shape_like], relay.reshape(data, (6, 2, 2)))
Review comment:
It should be fine to have another internal pass being called after concretizing like ops to check and mutate the function signatures.
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] comaniac commented on pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
comaniac commented on pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#issuecomment-801322861
> @comaniac thanks for the helpful feedback, it seems there's some design decisions that would be worth combining with existing approaches. I'll spend some time this week to incorporate the feedback, but in the meantime, I think I'll remove the ConcretizeLike pass from this PR and send a follow up (since the AD refactor can stand alone). How does that sound?
Sounds good to me and I'm good with the first-order AD changes in general.
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] altanh commented on a change in pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
altanh commented on a change in pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#discussion_r596444926
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
Review comment:
just to follow up on the idea of making these static, I'm mainly concerned about how the static lifetime interacts with the Operator registry, do you know how that works?
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] altanh commented on a change in pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
altanh commented on a change in pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#discussion_r596283755
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
Review comment:
1. thanks for the catch, this was left over from before
2. true, I'm not sure what the code style is for defining static variables like this but I'll try something
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] comaniac commented on a change in pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
comaniac commented on a change in pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#discussion_r596210282
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
Review comment:
1. Could you ellaborate what this loop does? It seems to me that it unions all the patterns that match ops in `concrete_map_`, but I didn't find `op_pat_` being used else where.
2. The construction of `concrete_map_` could be static, so we should be able to move it out of the constructor.
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
+
+ data_pat_ = IsWildcard();
+ like_pat_ = IsWildcard();
+ unary_like_pat_ = (IsOp("zeros_like") || IsOp("ones_like"))({like_pat_});
+ binary_like_pat_ = (IsOp("reshape_like") || IsOp("collapse_sum_like") ||
+ IsOp("broadcast_to_like"))({data_pat_, like_pat_});
+ }
+
+ Expr Callback(const Expr& pre, const Expr& post,
+ const Map<DFPattern, Array<Expr>>& node_map) const {
+ // we will rewrite iff the like argument has fully concrete shape
+ const CallNode* call_node = post.as<CallNode>();
+ ICHECK(call_node);
+ const OpNode* op_node = call_node->op.as<OpNode>();
+ ICHECK(op_node);
+ const Op op_ref = GetRef<Op>(op_node);
+ ICHECK(concrete_map_.count(op_ref) > 0);
+
+ Expr like = node_map[like_pat_][0];
+
+ if (!like->checked_type_.defined()) {
+ // TODO(@altanh): maybe because of the input being rewritten?
+ return post;
+ }
+
+ // skip trying to support this for now (ironic, as I was the one who added the feature)
+ if (const auto* attrs = call_node->attrs.as<ReshapeLikeAttrs>()) {
+ if (attrs->lhs_begin != 0 || attrs->rhs_begin != 0 || attrs->lhs_end.defined() ||
+ attrs->rhs_end.defined()) {
+ return post;
+ }
+ }
+
+ CHECK(like->checked_type_.defined())
+ << "ConcretizeLike requires checked types to be populated, please run type inference";
Review comment:
We already have L88-90 so this check seems useless.
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
+
+ data_pat_ = IsWildcard();
+ like_pat_ = IsWildcard();
+ unary_like_pat_ = (IsOp("zeros_like") || IsOp("ones_like"))({like_pat_});
+ binary_like_pat_ = (IsOp("reshape_like") || IsOp("collapse_sum_like") ||
+ IsOp("broadcast_to_like"))({data_pat_, like_pat_});
+ }
+
+ Expr Callback(const Expr& pre, const Expr& post,
+ const Map<DFPattern, Array<Expr>>& node_map) const {
+ // we will rewrite iff the like argument has fully concrete shape
+ const CallNode* call_node = post.as<CallNode>();
+ ICHECK(call_node);
+ const OpNode* op_node = call_node->op.as<OpNode>();
+ ICHECK(op_node);
+ const Op op_ref = GetRef<Op>(op_node);
+ ICHECK(concrete_map_.count(op_ref) > 0);
+
+ Expr like = node_map[like_pat_][0];
+
+ if (!like->checked_type_.defined()) {
+ // TODO(@altanh): maybe because of the input being rewritten?
+ return post;
+ }
+
+ // skip trying to support this for now (ironic, as I was the one who added the feature)
+ if (const auto* attrs = call_node->attrs.as<ReshapeLikeAttrs>()) {
+ if (attrs->lhs_begin != 0 || attrs->rhs_begin != 0 || attrs->lhs_end.defined() ||
+ attrs->rhs_end.defined()) {
+ return post;
+ }
+ }
+
+ CHECK(like->checked_type_.defined())
+ << "ConcretizeLike requires checked types to be populated, please run type inference";
+ const TensorTypeNode* like_ty = like->checked_type().as<TensorTypeNode>();
+ ICHECK(like_ty) << "got non-Tensor argument type " << PrettyPrint(like->checked_type());
+
+ Array<Integer> cshape;
+ for (const auto& dim : like_ty->shape) {
+ if (const auto* imm = dim.as<IntImmNode>()) {
+ cshape.push_back(Integer(GetRef<IntImm>(imm)));
+ continue;
+ }
+ return post;
+ }
+
+ if (call_node->args.size() == 2) {
+ return concrete_map_.at(op_ref)(node_map[data_pat_][0], cshape, like_ty->dtype);
+ }
+ return concrete_map_.at(op_ref)(Expr(), cshape, like_ty->dtype);
Review comment:
It might be better to refer to SimplifyExpr pass to separate unary and binary ops. Then maybe we could have a base struct to put the sharable logic.
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
+
+ data_pat_ = IsWildcard();
+ like_pat_ = IsWildcard();
+ unary_like_pat_ = (IsOp("zeros_like") || IsOp("ones_like"))({like_pat_});
+ binary_like_pat_ = (IsOp("reshape_like") || IsOp("collapse_sum_like") ||
+ IsOp("broadcast_to_like"))({data_pat_, like_pat_});
+ }
+
+ Expr Callback(const Expr& pre, const Expr& post,
+ const Map<DFPattern, Array<Expr>>& node_map) const {
+ // we will rewrite iff the like argument has fully concrete shape
+ const CallNode* call_node = post.as<CallNode>();
+ ICHECK(call_node);
+ const OpNode* op_node = call_node->op.as<OpNode>();
+ ICHECK(op_node);
+ const Op op_ref = GetRef<Op>(op_node);
+ ICHECK(concrete_map_.count(op_ref) > 0);
+
+ Expr like = node_map[like_pat_][0];
+
+ if (!like->checked_type_.defined()) {
+ // TODO(@altanh): maybe because of the input being rewritten?
+ return post;
+ }
+
+ // skip trying to support this for now (ironic, as I was the one who added the feature)
+ if (const auto* attrs = call_node->attrs.as<ReshapeLikeAttrs>()) {
+ if (attrs->lhs_begin != 0 || attrs->rhs_begin != 0 || attrs->lhs_end.defined() ||
+ attrs->rhs_end.defined()) {
+ return post;
+ }
+ }
Review comment:
This is too ad-hoc. It means we may not concretize *like ops in certain situations. Instead of hacking the unified callabck function, we should maintain this logic in the op specific function. I can think of two ways to achieve this:
1. Put the logic in the beginning of `concrete_map_` functions and return the same op if not applicable.
2. Construct another checker map that include checker functions of each op, and invoke the corresponding checker function here.
##########
File path: tests/python/relay/test_pass_concretize_like.py
##########
@@ -0,0 +1,108 @@
+# 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.
+"""Tests for the ConcretizeLike pass."""
+import tvm
+import tvm.relay.testing
+from tvm import relay
+from tvm.relay.testing import run_infer_type
+
+
+def test_reshape_like():
+ data = relay.var("data", shape=(2, 3, 4), dtype="float32")
+ shape_like = relay.var("shape_like", shape=(6, 2, 2), dtype="float32")
+ f = relay.Function([data, shape_like], relay.reshape_like(data, shape_like))
+ f_expected = relay.Function([data, shape_like], relay.reshape(data, (6, 2, 2)))
Review comment:
This exposes a question of the current implementation: after the ConcretizeLike pass we expect unused arguments. If we don't need shape_like tensor anymore, we should remove it from the function.
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
+
+ data_pat_ = IsWildcard();
+ like_pat_ = IsWildcard();
+ unary_like_pat_ = (IsOp("zeros_like") || IsOp("ones_like"))({like_pat_});
+ binary_like_pat_ = (IsOp("reshape_like") || IsOp("collapse_sum_like") ||
+ IsOp("broadcast_to_like"))({data_pat_, like_pat_});
Review comment:
1. These patterns can also be defined statically.
2. Now we have two places to specify the supported *like ops. I feel we should define a list of "unary like" and "binary like" ops once and used them both in `concrete_map_` and here. As a result, we could have the logic similar to L61-67 to construct the pattern.
##########
File path: tests/python/relay/test_pass_concretize_like.py
##########
@@ -0,0 +1,108 @@
+# 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.
+"""Tests for the ConcretizeLike pass."""
+import tvm
+import tvm.relay.testing
+from tvm import relay
+from tvm.relay.testing import run_infer_type
+
+
+def test_reshape_like():
+ data = relay.var("data", shape=(2, 3, 4), dtype="float32")
+ shape_like = relay.var("shape_like", shape=(6, 2, 2), dtype="float32")
+ f = relay.Function([data, shape_like], relay.reshape_like(data, shape_like))
+ f_expected = relay.Function([data, shape_like], relay.reshape(data, (6, 2, 2)))
+ f_expected = run_infer_type(f_expected)
+
+ mod = tvm.IRModule.from_expr(f)
+ mod_concrete = relay.transform.ConcretizeLike()(mod)
+ assert tvm.ir.structural_equal(mod_concrete["main"], f_expected)
+
+
+def test_zeros_like():
+ dtype = "int32"
+ shape_like = relay.var("shape_like", shape=(3, 4, 5), dtype=dtype)
+ f = relay.Function([shape_like], relay.zeros_like(shape_like))
+ f_expected = relay.Function([shape_like], relay.zeros((3, 4, 5), dtype))
+ f_expected = run_infer_type(f_expected)
+
+ mod = tvm.IRModule.from_expr(f)
+ mod_concrete = relay.transform.ConcretizeLike()(mod)
+ assert tvm.ir.structural_equal(mod_concrete["main"], f_expected)
+
+
+def test_ones_like():
+ dtype = "int32"
+ shape_like = relay.var("shape_like", shape=(3, 4, 5), dtype=dtype)
+ f = relay.Function([shape_like], relay.ones_like(shape_like))
+ f_expected = relay.Function([shape_like], relay.ones((3, 4, 5), dtype))
+ f_expected = run_infer_type(f_expected)
+
+ mod = tvm.IRModule.from_expr(f)
+ mod_concrete = relay.transform.ConcretizeLike()(mod)
+ assert tvm.ir.structural_equal(mod_concrete["main"], f_expected)
+
+
+def test_collapse_sum_like():
+ data = relay.var("data", shape=(3, 3, 3), dtype="float32")
+ shape_like = relay.var("shape_like", shape=(3,), dtype="float32")
+ f = relay.Function([data, shape_like], relay.collapse_sum_like(data, shape_like))
+ f_expected = relay.Function([data, shape_like], relay.collapse_sum_to(data, (3,)))
+ f_expected = run_infer_type(f_expected)
+
+ mod = tvm.IRModule.from_expr(f)
+ mod_concrete = relay.transform.ConcretizeLike()(mod)
+ assert tvm.ir.structural_equal(mod_concrete["main"], f_expected)
+
+
+def test_broadcast_to_like():
+ data = relay.var("data", shape=(3,), dtype="float32")
+ shape_like = relay.var("shape_like", shape=(3, 3, 3), dtype="float32")
+ f = relay.Function([data, shape_like], relay.broadcast_to_like(data, shape_like))
+ f_expected = relay.Function([data, shape_like], relay.broadcast_to(data, (3, 3, 3)))
+ f_expected = run_infer_type(f_expected)
+
+ mod = tvm.IRModule.from_expr(f)
+ mod_concrete = relay.transform.ConcretizeLike()(mod)
+ assert tvm.ir.structural_equal(mod_concrete["main"], f_expected)
+
+
+def test_multiple():
+ x = relay.var("x", shape=(2, 3), dtype="float32")
+ y = relay.var("x", shape=(3,), dtype="float32")
+ l = x + y
+
+ dl = relay.ones_like(l)
+ dx = relay.zeros_like(x)
+ dy = relay.zeros_like(y)
+ dx = dx + relay.collapse_sum_like(dl, dx)
+ dy = dy + relay.collapse_sum_like(dl, dy)
+ ret = relay.Tuple([dx, dy])
+ f = relay.Function([x, y], ret)
+
+ dl_c = relay.ones((2, 3), "float32")
+ dx_c = relay.zeros((2, 3), "float32")
+ dy_c = relay.zeros((3,), "float32")
+ dx_c = dx_c + relay.collapse_sum_to(dl_c, (2, 3))
+ dy_c = dy_c + relay.collapse_sum_to(dl_c, (3,))
+ ret_c = relay.Tuple([dx_c, dy_c])
+ f_expected = relay.Function([x, y], ret_c)
+ f_expected = run_infer_type(f_expected)
+
+ mod = tvm.IRModule.from_expr(f)
+ mod_concrete = relay.transform.ConcretizeLike()(mod)
+ assert tvm.ir.structural_equal(mod_concrete["main"], f_expected)
Review comment:
Add `if __name__ == "__main__":`
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] altanh commented on pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
altanh commented on pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#issuecomment-801300223
@comaniac thanks for the helpful feedback, it seems there's some design decisions that would be worth combining with existing approaches. I'll spend some time this week to incorporate the feedback, but in the meantime, I think I'll remove the ConcretizeLike pass from this PR and send a follow up (since the AD refactor can stand alone). How does that sound?
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] altanh commented on a change in pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
altanh commented on a change in pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#discussion_r595700369
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
+
+ data_pat_ = IsWildcard();
+ like_pat_ = IsWildcard();
+ unary_like_pat_ = (IsOp("zeros_like") || IsOp("ones_like"))({like_pat_});
+ binary_like_pat_ = (IsOp("reshape_like") || IsOp("collapse_sum_like") ||
+ IsOp("broadcast_to_like"))({data_pat_, like_pat_});
+ }
+
+ Expr Callback(const Expr& pre, const Expr& post,
+ const Map<DFPattern, Array<Expr>>& node_map) const {
+ // we will rewrite iff the like argument has fully concrete shape
+ const CallNode* call_node = post.as<CallNode>();
+ ICHECK(call_node);
+ const OpNode* op_node = call_node->op.as<OpNode>();
+ ICHECK(op_node);
+ const Op op_ref = GetRef<Op>(op_node);
+ ICHECK(concrete_map_.count(op_ref) > 0);
+
+ Expr like = node_map[like_pat_][0];
+
+ if (!like->checked_type_.defined()) {
+ // TODO(@altanh): maybe because of the input being rewritten?
+ return post;
+ }
+
+ // skip trying to support this for now (ironic, as I was the one who added the feature)
+ if (const auto* attrs = call_node->attrs.as<ReshapeLikeAttrs>()) {
+ if (attrs->lhs_begin != 0 || attrs->rhs_begin != 0 || attrs->lhs_end.defined() ||
+ attrs->rhs_end.defined()) {
+ return post;
+ }
+ }
+
+ CHECK(like->checked_type_.defined())
+ << "ConcretizeLike requires checked types to be populated, please run type inference";
+ const TensorTypeNode* like_ty = like->checked_type().as<TensorTypeNode>();
+ ICHECK(like_ty) << "got non-Tensor argument type " << PrettyPrint(like->checked_type());
+
+ Array<Integer> cshape;
+ for (const auto& dim : like_ty->shape) {
+ if (const auto* imm = dim.as<IntImmNode>()) {
+ cshape.push_back(Integer(GetRef<IntImm>(imm)));
+ continue;
+ }
+ return post;
+ }
+
+ if (call_node->args.size() == 2) {
+ return concrete_map_.at(op_ref)(node_map[data_pat_][0], cshape, like_ty->dtype);
+ }
+ return concrete_map_.at(op_ref)(Expr(), cshape, like_ty->dtype);
Review comment:
Yeah maybe this is too much of a hack, I'm just using it as a placeholder since the unary matches won't have a corresponding data Expr node. I'll rework this tmrw
----------------------------------------------------------------
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:
users@infra.apache.org
[GitHub] [tvm] comaniac commented on a change in pull request #7677: [Relay][Training][Pass] Factor out first-order AD to a module pass, and add ConcretizeLike pass
Posted by GitBox <gi...@apache.org>.
comaniac commented on a change in pull request #7677:
URL: https://github.com/apache/tvm/pull/7677#discussion_r596446640
##########
File path: src/relay/transforms/concretize_like.cc
##########
@@ -0,0 +1,160 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file concretize_like.cc
+ * \brief Converts `*_like` operators to their explicit shape equivalent (e.g. `zeros_like(x, y)` to
+ * `zeros(x, y.shape)`), when the target shape is concrete. This removes unnecessary dependencies
+ * and can enable more opportunities for operator fusion.
+ */
+#include <tvm/relay/dataflow_matcher.h>
+#include <tvm/relay/transform.h>
+
+#include "pattern_utils.h"
+
+namespace tvm {
+namespace relay {
+
+class ConcretizeLikeRewrite {
+ public:
+ ConcretizeLikeRewrite() {
+ concrete_map_[Op::Get("reshape_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeReshape(data, shape);
+ };
+ concrete_map_[Op::Get("zeros_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeZeros(shape, dtype);
+ };
+ concrete_map_[Op::Get("ones_like")] = [](Expr data, Array<Integer> shape, DataType dtype) {
+ return MakeOnes(shape, dtype);
+ };
+ concrete_map_[Op::Get("collapse_sum_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ ICHECK_LE(shape.size(), std::numeric_limits<int64_t>::max());
+ static const Op& op = Op::Get("collapse_sum_to");
+ auto attrs = make_object<InitOpAttrs>();
+ auto cshape =
+ MakeConstantTensor(DataType::Int(32), {static_cast<int64_t>(shape.size())}, shape);
+ attrs->shape = shape;
+ return Call(op, {data, cshape}, Attrs(attrs));
+ };
+ concrete_map_[Op::Get("broadcast_to_like")] = [](Expr data, Array<Integer> shape,
+ DataType dtype) {
+ return MakeBroadCastTo(data, shape);
+ };
+
+ for (const auto& pr : concrete_map_) {
+ if (!op_pat_.defined()) {
+ op_pat_ = IsExpr(pr.first);
+ } else {
+ op_pat_ = op_pat_ || IsExpr(pr.first);
+ }
+ }
Review comment:
I don't think it would be an issue. If you search `static const Op& op` in the code base, you'll find lots of use cases.
----------------------------------------------------------------
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:
users@infra.apache.org