This repository has been archived by the owner on Jul 1, 2023. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 133
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Lazy tensor: automatically promote constants to inputs based on histo…
…ry (#476)
- Loading branch information
Showing
10 changed files
with
306 additions
and
11 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,158 @@ | ||
| // Copyright 2019 The TensorFlow Authors. All Rights Reserved. | ||
| // | ||
| // Licensed 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. | ||
| import CTensorFlow | ||
|
|
||
| extension TFETensorHandle: Equatable {} | ||
|
|
||
| public func ==(_ lhs: TFETensorHandle, _ rhs: TFETensorHandle) -> Bool { | ||
| return lhs._cTensorHandle == rhs._cTensorHandle | ||
| } | ||
|
|
||
| extension TFETensorHandle { | ||
| /// Returns true if the underlying tensors are equal. | ||
| func elementsEqual(_ other: TFETensorHandle) -> Bool { | ||
| let selfDtype = TFE_TensorHandleDataType(self._cTensorHandle) | ||
| let otherDtype = TFE_TensorHandleDataType(other._cTensorHandle) | ||
| precondition( | ||
| selfDtype == otherDtype && selfDtype != TF_VARIANT && selfDtype != TF_RESOURCE, | ||
| "Datatypes of tensor handles don't match.") | ||
| let op = TFE_Op("Equal", 1) | ||
| op.updateAttribute("T", TensorDataType(selfDtype)) | ||
| op.addInput(self) | ||
| op.addInput(other) | ||
| let result: Tensor<Bool> = op.execute(Int(1)) | ||
| return result.scalars.allSatisfy { $0 } | ||
| } | ||
| } | ||
|
|
||
| extension LazyTensorHandle { | ||
| func isEquivalent(to other: LazyTensorHandle) -> Bool { | ||
| switch (self.handle, other.handle) { | ||
| case let (.concrete(x, _), .concrete(y, _)): | ||
| return x == y | ||
| case let (.symbolic(x, xi, _), .symbolic(y, yi, _)): | ||
| return xi == yi && x.id == y.id | ||
| default: return false | ||
| } | ||
| } | ||
| } | ||
|
|
||
| extension LazyTensorOperation.Input { | ||
| /// Returns true if these inputs are equivalent when comparing lazy tensor traces. | ||
| func isEquivalent(to other: LazyTensorOperation.Input) -> Bool { | ||
| switch (self, other) { | ||
| case let (.single(l), .single(r)): | ||
| return l.isEquivalent(to: r) | ||
| case let (.list(l), .list(r)): | ||
| return l.elementsEqual(r, by: { $0.isEquivalent(to: $1) }) | ||
| default: | ||
| return false | ||
| } | ||
| } | ||
| } | ||
|
|
||
| extension LazyTensorOperation { | ||
| /// Returns true if these operations are equivalent when comparing lazy tensor traces. | ||
| func isEquivalent(to other: LazyTensorOperation) -> Bool { | ||
| return self.name == other.name && | ||
| self.outputCount == other.outputCount && | ||
| self.deviceName == other.deviceName && | ||
| self.inputs.elementsEqual(other.inputs, by: { $0.isEquivalent(to: $1) }) && | ||
| self.attributes == other.attributes | ||
| } | ||
| } | ||
|
|
||
| // TODO(TF-693): This is not thread safe! | ||
| struct LazyTensorTraceCache { | ||
| /// Cache from signature to traces that match signature. | ||
| static private var cache: [String: [LazyTensorTrace]] = [:] | ||
| static func clearCache() { cache.removeAll() } | ||
|
|
||
| /// Returns a `MaterializationTraceInfo` with possibly some constants promoted to inputs. | ||
| static func traceWithPromotedConstants( | ||
| _ traceInfo: MaterializationTraceInfo | ||
| ) -> MaterializationTraceInfo { | ||
| let trace = traceInfo.trace | ||
| guard var traces = cache[trace.signature] else { | ||
| cache[trace.signature] = [trace] | ||
| return traceInfo | ||
| } | ||
| for cachedTrace in traces { | ||
| if let promotedTrace = traceInfo.withPromotedConstants(cachedTrace: cachedTrace) { | ||
| debugLog("Promoted: \(promotedTrace)\n") | ||
| return promotedTrace | ||
| } | ||
| } | ||
| // No match found; cache and return the input `traceInfo` itself. | ||
| traces.append(trace) | ||
| return traceInfo | ||
| } | ||
| } | ||
|
|
||
| private extension MaterializationTraceInfo { | ||
| func withPromotedConstants(cachedTrace: LazyTensorTrace) -> MaterializationTraceInfo? { | ||
| let currentTrace = self.trace | ||
| if currentTrace.operations.count != cachedTrace.operations.count { return nil } | ||
| var promotableConstants: [(Int, TFETensorHandle)] = [] | ||
| for (i, current) in currentTrace.operations.enumerated() { | ||
| let cached = cachedTrace.operations[i] | ||
| if let (currentTensor, cachedTensor) = Self.promotableConstants(current, cached) { | ||
| if currentTensor.elementsEqual(cachedTensor) { continue } | ||
| promotableConstants.append((i, currentTensor)) | ||
| continue | ||
| } | ||
| // TODO: we might avoid running the following check based on results of promotableConstant | ||
| if current.isEquivalent(to: cached) { continue } | ||
| return nil | ||
| } | ||
|
|
||
| let newConcreteInputs: [TFETensorHandle] = promotableConstants.map { return $0.1 } | ||
| let newOperations = currentTrace.operations | ||
| let newInputs = promotableConstants.map { | ||
| (promotableConstant: (Int, TFETensorHandle)) -> LazyTensorOperation in | ||
| let constantOp = newOperations[promotableConstant.0] | ||
| constantOp.name = "Placeholder" | ||
| constantOp.attributes.removeValue(forKey: "value") | ||
| return constantOp | ||
| } | ||
| let newTrace = LazyTensorTrace( | ||
| inputs: currentTrace.inputs + newInputs, | ||
| operations: newOperations, | ||
| outputs: currentTrace.outputs) | ||
| return MaterializationTraceInfo( | ||
| lazyOperations: self.lazyOperations, | ||
| trace: newTrace, | ||
| concreteInputs: self.concreteInputs + newConcreteInputs) | ||
| } | ||
|
|
||
| /// If `current` and `cached` are compatible constants, returns the constant tensors. | ||
| static private func promotableConstants( | ||
| _ current: LazyTensorOperation, | ||
| _ cached: LazyTensorOperation | ||
| ) -> (TFETensorHandle, TFETensorHandle)? { | ||
| if current.name != "Const" || cached.name != "Const" { return nil } | ||
| let currentValue = current.attributes["value"]! | ||
| let cachedValue = cached.attributes["value"]! | ||
| guard case let .constTensor(currentTensor) = currentValue, | ||
| case let .constTensor(cachedTensor) = cachedValue | ||
| else { return nil } | ||
| let currentDtype = TFE_TensorHandleDataType(currentTensor._cTensorHandle) | ||
| let cachedDtype = TFE_TensorHandleDataType(cachedTensor._cTensorHandle) | ||
| if currentDtype == TF_VARIANT || currentDtype == TF_RESOURCE { return nil } | ||
| if cachedDtype == TF_VARIANT || cachedDtype == TF_RESOURCE { return nil } | ||
| return currentTensor.shape == cachedTensor.shape && currentDtype == cachedDtype | ||
| ? (currentTensor, cachedTensor) | ||
| : nil | ||
| } | ||
| } |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,131 @@ | ||
| // Copyright 2019 The TensorFlow Authors. All Rights Reserved. | ||
| // | ||
| // Licensed 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. | ||
|
|
||
| import XCTest | ||
|
|
||
| @testable import TensorFlow | ||
| import CTensorFlow | ||
|
|
||
| final class LazyTensorTraceCacheTests: LazyTensorTestCase { | ||
| override class func setUp() { | ||
| super.setUp() | ||
| LazyTensorContext.local.shouldPromoteConstants = true | ||
| } | ||
|
|
||
| override class func tearDown() { | ||
| super.tearDown() | ||
| LazyTensorTraceCache.clearCache() | ||
| } | ||
|
|
||
| func testConstPromotion() { | ||
| LazyTensorTraceCache.clearCache() | ||
| let a = Tensor<Float>(1.0) | ||
| let b = Tensor<Float>(2.0) | ||
| let c = Tensor<Float>(3.0) | ||
| let d = Tensor<Float>(4.0) | ||
| let w = a * b | ||
| let x = c * d | ||
| // Trigger materialization for `w` so that a trace with constants and mul is added to cache. | ||
| XCTAssertEqual( | ||
| lazyTrace(w).description, | ||
| """ | ||
| lazyTrace_3() -> (%2) { | ||
| %0 = Const[dtype: float, value: 1.0]() | ||
| %1 = Const[dtype: float, value: 2.0]() | ||
| %2 = Mul[T: float](%0, %1) | ||
| } | ||
| """) | ||
| XCTAssertEqual(w.scalars, [2.0]) | ||
|
|
||
| // The trace for `x` should have the inputs to Mul as arguments instead of constants. | ||
| XCTAssertEqual( | ||
| lazyTrace(x).description, | ||
| """ | ||
| lazyTrace_3(%0: float, %1: float) -> (%2) { | ||
| %2 = Mul[T: float](%0, %1) | ||
| } | ||
| """) | ||
| XCTAssertEqual(x.scalarized(), 12.0) | ||
|
|
||
| let e = Tensor<Float>(shape: [1,3], scalars: [1, 2, 3]) | ||
| let f = Tensor<Float>(5.0) | ||
| let y = e * f | ||
| // We won't promote constants in 'y' as shape of constants is different. | ||
| XCTAssertEqual( | ||
| lazyTrace(y).description, | ||
| """ | ||
| lazyTrace_3() -> (%2) { | ||
| %0 = Const[dtype: float, value: [[1.0, 2.0, 3.0]]]() | ||
| %1 = Const[dtype: float, value: 5.0]() | ||
| %2 = Mul[T: float](%0, %1) | ||
| } | ||
| """) | ||
| XCTAssertEqual(y.scalars, [5.0, 10.0, 15.0]) | ||
| } | ||
|
|
||
| func testDoNotPromoteEqualConstants() { | ||
| LazyTensorTraceCache.clearCache() | ||
| let a = Tensor<Float>(1.0) | ||
| let b = Tensor<Float>(2.0) | ||
| let c = Tensor<Float>(3.0) | ||
| let w = a * b | ||
| let x = a * c | ||
| XCTAssertEqual( | ||
| lazyTrace(w).description, | ||
| """ | ||
| lazyTrace_3() -> (%2) { | ||
| %0 = Const[dtype: float, value: 1.0]() | ||
| %1 = Const[dtype: float, value: 2.0]() | ||
| %2 = Mul[T: float](%0, %1) | ||
| } | ||
| """) | ||
| XCTAssertEqual(w.scalars, [2.0]) | ||
| // Const 1.0 is not promoted. | ||
| XCTAssertEqual( | ||
| lazyTrace(x).description, | ||
| """ | ||
| lazyTrace_3(%1: float) -> (%2) { | ||
| %0 = Const[dtype: float, value: 1.0]() | ||
| %2 = Mul[T: float](%0, %1) | ||
| } | ||
| """) | ||
| } | ||
|
|
||
| private func lazyTensorOperation<T: TensorFlowScalar>( | ||
| _ input: Tensor<T> | ||
| ) -> LazyTensorOperation? { | ||
| let tensor = input.handle.handle | ||
| guard let lazyTensor = tensor as? LazyTensorHandle else { | ||
| XCTFail("Trying to get lazy trace for a non-lazy tensor.") | ||
| return nil | ||
| } | ||
| guard case let .symbolic(lazyOp, _, _) = lazyTensor.handle else { | ||
| XCTFail("Cannot get lazy trace for a concrete tensor.") | ||
| return nil | ||
| } | ||
| return lazyOp | ||
| } | ||
|
|
||
| private func lazyTrace<T: TensorFlowScalar>( | ||
| _ input: Tensor<T> | ||
| ) -> LazyTensorTrace { | ||
| let lazyOperation = lazyTensorOperation(input)! | ||
| return LazyTensorTraceBuilder.materializationTraceInfo(lazyOperation).trace | ||
| } | ||
|
|
||
| static var allTests = [ | ||
| ("testConstPromotion", testConstPromotion), | ||
| ("testDoNotPromoteEqualConstants", testDoNotPromoteEqualConstants) | ||
| ] | ||
| } |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters