aboutsummaryrefslogtreecommitdiff
path: root/src/armnn/optimizations
AgeCommit message (Collapse)Author
2024-02-13MLCE-1165 Model failing to load when pad is folded into Conv2dTracy Narine
* Skipping the optimization which folds pad and conv2d together for a specific case: 1x1 filter and padding size >= filter size Signed-off-by: Tracy Narine <tracy.narine@arm.com> Change-Id: I46944e9f736df1ff60469b2d2852e1bba01ab8cd
2023-09-08IVGCVSW-7525 Add broadcast_to to TFLite ParserIdriss Chaouch
* Changing the optimizer * Changing EndToEnd Tests Signed-off-by: Idriss Chaouch <idriss.chaouch@arm.com> Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com> Change-Id: Ib581794280322a39cfc5ea3c4e6a6398cf723d5e
2023-08-31IVGCVSW-7525 Add broadcast_to operatorIdriss Chaouch
Signed-off-by: Idriss Chaouch <idriss.chaouch@arm.com> Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com> Change-Id: I94ec5f9120b2d736fdf98d00ec5137a4efd739b8
2023-07-03IVGCVSW-7828 Add an Optional TensorInfo to InputSlotMike Kelly
* Updated calls to use the new function From: GetInputSlot(n).GetConnection()->GetTensorInfo(); To: GetInputSlot(n).GetTensorInfo(); * Added UnitTests Signed-off-by: Mike Kelly <mike.kelly@arm.com> Change-Id: I43184cc05e4472011b9347aaa820eb8deb1cd4a0
2023-03-14IVGCVSW-3808 Add ElementwiseBinaryLayerMike Kelly
!android-nn-driver:9329 * Added ElementwiseBinaryLayer that can represent all ElementwiseBinary operations including Add, Div, Sub, Maximum, Mul and Minimum. * Updated Delegate to use ElementwiseBinaryLayer instead of the Add, Div, Sub, Maximum, Mul and Minimum layers. * Updated Deserializer to use ElementwiseBinaryLayer instead of the Add, Div, Sub, Maximum, Mul and Minimum layers. * Updated OnnxParser to use ElementwiseBinaryLayer instead of the Add layer. * Updated TfLiteParser to use ElementwiseBinaryLayer instead of the Add, Div, Sub, Maximum, Mul and Minimum layers. * Updated CL and Neon tests to use ElementwiseBinaryLayer. * Updated CL and Neon Backend Specific Optimizations to accept ElementBinaryLayers as well as Add, Div, Mul, Sub, Maximum and Minimum layers. Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com> Signed-off-by: Mike Kelly <mike.kelly@arm.com> Change-Id: I7cbb96b60eb01f0e2b57b0541016d48a08b86c75
2023-01-20IVGCVSW-7013 Removing the check for constant layer in FoldPadTeresa Charlin
Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com> Change-Id: Iaa3cff46217117aefdb92f281e9da2b0315f3af9
2022-12-12IVGCVSW-7209 Remove deprecated code due to be removed in 23.02Mike Kelly
* Removed weights and bias from Convolution, DepthwiseConv & FullyConnected layers * Removed the weight and bias ConstTensorHandles from the QueueDescriptors * Updated Workloads to take tensors from WorkloadInfo rather than the QueueDescriptors * Removed unused RedirectMembersToConstantInputs optimization and tests. Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com> Signed-off-by: Mike Kelly <mike.kelly@arm.com> Change-Id: I9ffcdc4a1c0dff725539dd69fc435b700bd98a56
2022-11-16IVGCVSW-7214 Disable BF16-Turbo-Mode and remove conversion layersRyan OShea
- Remove Bf16ToFp32 Conversion Layer - Remove Fp32ToBf16 Conversion Layer - Remove B16 Conversion tests * Throw exception if m_ReduceFp32ToBf16 optimzer option is set to true * Provide comments to enable fast math in order to use bf16 * Update docs to inform users to enable fast math for bf16 Execute Network Changes * Require bf16_turbo_mode to also have fast_math_enabled set to true - Remove setting m_ReduceFp32ToBf16 optimizer option Signed-off-by: Ryan OShea <ryan.oshea3@arm.com> Change-Id: Ibaa6da9d29c96a1ce32ff5196b0847fde9f04a1c
2022-10-19MLCE-545 INT8 TFLite model execution abnormalKeith Davis
* Add functionality to print output tensors to file in tempdir * UnitTests Signed-off-by: Keith Davis <keith.davis@arm.com> Change-Id: Idfb4c186544187db1fecdfca11c662540f645439
2022-10-04MLCE-545 INT8 TFLite model execution abnormalKeith Davis
* Fix for Debug mode in ExNet does not work with ConstTensorsAsInputs * Remove unnecessary assertion with ambiguous message in LoadedNetwork Signed-off-by: Keith Davis <keith.davis@arm.com> Change-Id: I9cd5d1f811dbbc89072d1190c510bf1b22e3069c
2022-09-28IVGCVSW-7209 Delay one release the removal of weights and biasTeresa Charlin
* This affects only to the layers (not workloads) Conv, DWConv and FC Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com> Change-Id: I66a91ed1a78bc0464e00423c7fc7c28c91d199ce
2022-08-29IVGCVSW-6954 'Arm NN SL Improvements'Sadik Armagan
* Move the Conv2D and DepthwiseConv2D validation to Optimization level when the weights and tensors are as constant inputs * Take into account offset and scales values when doing INT8 to FP32 dequantization Signed-off-by: Sadik Armagan <sadik.armagan@arm.com> Change-Id: I1f81f15640395ac041923b10dbe9151159715117
2022-08-05GitHub #667: Neon fold padding into average pool 2D quantization bug fix.Cathal Corbett
* Originated from a GitHub issue: https://github.com/ARM-software/armnn/issues/667 * Initially, Arm NN supports the pool 2D operation because there is no padding on the pool2d. Neon failure occurs when padding is followed by average pool 2D due to folding optimization. * Here we prevent the folding optimization from happening for the above special case and add it in as a backend specific optimization. Signed-off-by: Cathal Corbett <cathal.corbett@arm.com> Change-Id: Ia0fd90c3a6b4b9d29c81106f154617d2e893e26b
2022-07-27IVGCVSW-6978: RedirectMembersToConstantInputs does not work with ↵Francis Murtagh
Fp32NetworkToBf16Converter * Fuse FP32ToBF16Layers with Constant Layer so Conv2d/FullyConnected can have their weights redirected. * If BF16 Unsupported in Conv2d || FullyConnected revert fused Constant Layer to FP32 Change-Id: If523c708a822659d64597d9ae39cca1c2f84b76f Signed-off-by: Francis Murtagh <francis.murtagh@arm.com>
2022-05-18IVGCVSW-6147 ConstTensorsAsInput: Optimizer - FusePermuteIntoConstLayerCathal Corbett
* No trailing permute layer after a constant layer * Unit test for optimization Signed-off-by: Cathal Corbett <cathal.corbett@arm.com> Change-Id: I0d098f5af41d2c55df7cef1ccfb848093320ddc1
2022-05-18IVGCVSW-6455 Support Const + Dequantize layer and optimize it.Teresa Charlin
* Support Float16 as input to Dequantize layer * Add Optimization to substitute Const+Dequantize layers with Const layer Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com> Change-Id: I58bb7e3871ca480c7b6fca93c4efb2de84e09e64 Signed-off-by: David <david.monahan@arm.com>
2022-05-16IVGCVSW-6124 ConstTensorsAsInput: Conv2d - FrontEndKeith Davis
* Update Front-end and Tools. * Updated Serializer, Deserializer and unit tests to reflect this. * Updated TfLiteDelegate, TfLiteParser and OnnxParser. * Updated Ref. * Fixed resulting Neon / CL tests * Unified optimizers for conv2d ops * Optimizer Fix - Fp32ToBf16 * Partial implementation for ACL backends to fix VTS failures !android-nn-driver:7477 Signed-off-by: Keith Davis <keith.davis@arm.com> Change-Id: I5fb18877f7ee32643e15a9818945356274bb401b
2022-05-12IVGCVSW-6940 ConstTensorsAsInput: DepthwiseConvolution2d - Complete ACLCathal Corbett
* Added backend specific optimization & test for CpuAcc and GpuAcc: PermuteDepthwiseConv2dWeights Signed-off-by: Cathal Corbett <cathal.corbett@arm.com> Change-Id: I600476b2e9c557a39818a574c1091c9d650b21b1
2022-05-05IVGCVSW-6127 ConstTensorsAsInput: DepthwiseConvolution2dCathal Corbett
!android-nn-driver:7418 * Update Front-end and Tools. * Updated Serializer, Deserializer and unit tests to reflect this. * Updated TfLiteDelegate, TfLiteParser and OnnxParser. * Change NNDriver to new API. * Updated Ref. * Neon and Cl backend partially completed (Backend.cpp files). * Added dynamic or constant input EndToEnd tests. * Added ConstantTensorAsInputMemeberVariableRedirect Optimization. Signed-off-by: Cathal Corbett <cathal.corbett@arm.com> Change-Id: Ib18b6c10a093042e165e25237dc04a4c67ba82da
2022-02-07IVGCVSW-6635 Move MemCopyTestImpl from acl to armnnTestUtils.Colm Donelan
* Move MemCopyTestImpl.hpp from src/backends/aclCommon/test/ to include/armnnTestutils. * Refactor MemCopyTests in aclCommon, cl and Neon. * Introduce RefMemCopyTests to exercise this utility in x86 builds. Signed-off-by: Colm Donelan <colm.donelan@arm.com> Change-Id: I8824f013d3656658ed0a2904bb79384e3af68641
2021-12-15IVGCVSW-6626 Promote backend headers in backendCommon to armnn/backendsColm Donelan
Move the following header files from backendsCommon to armnn/backends. * MemCopyWorkload.hpp * TensorHandle.hpp * Workload.hpp * WorkloadData.hpp * WorkloadFactory.hpp Replace them with forwarding headers and a pragma deprecation message. Resolve the deprecation messages in Arm NN code. Signed-off-by: Colm Donelan <colm.donelan@arm.com> Change-Id: I47f116b30f86e478c9057795bc518c391a8ae514
2021-11-08IVGCVSW-6420: Constant flag in tensor info is not set correctlyCathal Corbett
!android-nn-driver:6532 !armnn-internal-tests:372451 * Made fix to 2 out of 3 ConstTensor() constructors in Tensor.hpp to throw InvalidArgumentException when TensorInfo isConstant parameter is false. * Added new ConstTensor() constructor in Tensor.cpp to accept vector<>.data() using template<typename MemoryType>. * Fixed runtime->GetOutputTensorInfo()/GetInputTensorInfo() methods and called submethods to return TensorInfo& rather than TensorInfo. * Fixed all failing unit tests for CpuRef/CpuAcc/GpuAcc to ensure any ConstTensor created has it's TensorInfo isConstant set to true. * Added unit tests in TensorTest.cpp to ensure ConstTensor constructors throw InvalidArgumentException when TensorInfo isConstat parameter is false. * Added unit test to ensure an empty ConstTensor constructor will set TensorInfo isConatant to true. * Indentation fixes. * Fix to arm_tensor.i to add isConstant parameter to TensorInfo constructor. Added methods IsConstant() and SetConstant(). * Fix to const_tensor.py to throw ValueError when TensorInfo isConstant is set to false when constructing a ConstTensor. * Fixed PyArmnn unit tests to set TensorInfo isConstant to True when ConstTensor is used. * Added unit tests in test_const_tensor.py to ensure ConstTensor constructors throw ValueError when TensorInfo isConstat parameter is false. Signed-off-by: Cathal Corbett <cathal.corbett@arm.com> Change-Id: I44e440dd0422c366d31bbdbc77ad2b4db0bde148
2021-09-15IVGCVSW-6331 Fixup include directive in RedirectMembersToConstantInputsFrancis Murtagh
Signed-off-by: Francis Murtagh <francis.murtagh@arm.com> Change-Id: I2503e6995c83316094257318743a756da600bb6c
2021-06-30IVGCVSW-6161 ConstTensorsAsInput: Optimizer - Redirect ConstTensor layer membersMatthew Sloyan
* Optimization that searches for layers with ConstantLayers as inputs. * The layer member variables are then redirected to these ConstantLayers. Signed-off-by: Matthew Sloyan <matthew.sloyan@arm.com> Change-Id: I24a2bf0e8575b808343e0bbe3897b344e94796ad
2021-06-16IVGCVSW-5826 Change weights layout for depthwise to [1,H,W,I*M]Jan Eilers
* This change is necessary because tflite uses a [1,H,W,I*M] format and uses the I*M dimension for per axis quantization. Our previous layout [M,I,H,W] can't handle the correlating quantization scales. * Updates Onnx-, TfLiteParser and TfliteDelegate * Updates the CpuRef, CpuAcc and GpuAcc backends * Adjusts unit tests * Adds test to ensure models with old layout can still be read and executed * Adds conversion function to previous layout [1,H,W,I*M] --> [M,I,H,W] which can be used by backend developers !android-nn-driver:5553 Signed-off-by: Jan Eilers <jan.eilers@arm.com> Change-Id: Ifef23368b8c3702cf315a5838d214f7dc13c0152
2021-05-25IVGCVSW-3649 Add Prelu with different alpha dimension test to TfLiteParserNarumol Prangnawarat
Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com> Change-Id: I982ecd66ea3ed4d88934cd8254832eecb4a7adb4
2021-05-24IVGCVSW-6069 Fold PAD into Depthwise ConvolutionTeresa Charlin
Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com> Change-Id: Ib01629256309cfe17f341909d5b9bbbb09361422
2021-05-06IVGCVSW-5815 Generalise ConstCpuTensorHandleJames Conroy
* Generalises ConstCpuTensorHandle and inherited classes by removing 'Cpu' from aliases. * New renamed classes: ConstTensorHandle, TensorHandle, ScopedTensorHandle, PassthroughTensorHandle, ConstPassthroughTensorHandle. Signed-off-by: James Conroy <james.conroy@arm.com> Change-Id: I1824e0e134202735fb77051f20a7252f161dfe16
2021-04-29IVGCVSW-5890 Prevent modification to const layers with multiple connectionsColm Donelan
* In AddBroadcastReshapeLayerImpl check if a constant layer has other connections before modifying its output tensor shape. * In ElementWiseBaseLayer replace an ARMNN_ASSERT with a proper error message. Signed-off-by: Colm Donelan <Colm.Donelan@arm.com> Change-Id: Id3f3796c260eede61f076660505257a8b65d93fc
2021-04-21Fold PAD into Pooling2d if possibleDiego Lopez Recas
Some models would add a PAD layer before a pooling when they can't express their padding configuration as SAME or VALID. ArmNN can potentially merge the two merge the two because pooling layers are described with explicit padding. The merge is possible if the extra padding is neutral in the combined pooling operation. A merged operation can only fuse paddings in the dimensions that accept explicit padding in a pooling operation, i.e. the spatial dimensions. Signed-off-by: Diego Lopez Recas <diego.lopez.recas@gmail.com> Signed-off-by: Colm Donelan <Colm.Donelan@arm.com> Change-Id: Icd54718dcd9e797c923456b7fa6e0213e288e668
2021-04-07Fix graph copy memory spikeFinn Williams
* Change layer storage of ConstTensors to std::shared_ptr<ConstCpuTensorHandle> * Change clone to share ConstTensor rather than copy * Remove uses of non-const GetTensor() call * Reduce scope of non-optimized network in ExeNet, so memory can be released after use Signed-off-by: Finn Williams <Finn.Williams@arm.com> Change-Id: Ibb2c7309d12411d21405bd6024c76bcdf5404545
2021-03-23Revert "Fold PAD into Pooling2d"Jim Flynn
This reverts commit 51ce7d487c761358de105f82ff90553570aedac0. Reason for revert: https://jira.arm.com/browse/IVGCVSW-5798 LargeGraph_TENSOR_FLOAT32 CTS tests failures Change-Id: Ib031a47f605340b2202ecf074ce96a8b54c51075
2021-03-22Fold PAD into Pooling2dDiego Lopez Recas
Some models would add a PAD layer before a pooling when they can't express their padding configuration as SAME or VALID. Arm NN can merge the two because pooling layers are described with explicit padding. Signed-off-by: Diego Lopez Recas <diego.lopez.recas@gmail.com> Change-Id: Id048186db6a005e0257bfbc1406c3b0dab2cdd58
2021-03-16IVGCVSW-5754 Change the behaviour of the AddBroadcastReshapeLayer ↵Finn Williams
Optimisation when the input is a const tensor Signed-off-by: Finn Williams <Finn.Williams@arm.com> Change-Id: I8b1357bdefc45880d064d7e448af364ac8644c0d
2021-01-25IVGCVSW-5619 Add OptimizerOptions and NetworkProperties to ArmNN DelegateNarumol Prangnawarat
* Add OptimizerOptions, NetworkProperties, DebugCallbackFunction to DelegateOptions * Enable OptimizerOptions when the network is being optimized * Enable NetworkProperties when loading network * Enable DebugCallbackFunction * Add error message when loading network * Log warning instead of error when operator is not supported but could fallback to another backend * Improve uint16_t CompareData * Unit tests Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com> Change-Id: I353035afb442774bfeb1c62570a90755c2ceaf38
2020-11-17IVGCVSW-5530 'Cannot run SSD Mobilenet f16/uint8 on CpuRef via ExecuteNetwork'Sadik Armagan
* Added FP16 DataType support to DetectionPostProcess * For DetectionPostProcess layer output is always Float32 regardless of input type Signed-off-by: Sadik Armagan <sadik.armagan@arm.com> Change-Id: I21f63dd08f0863e9a98e105b3009bab3da1ab0c3
2020-11-08IVGCVSW-5315 Create FuseBatchNorm classMike Kelly
Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com> Signed-off-by: Mike Kelly <mike.kelly@arm.com> Change-Id: Id0625c58dbeea79874bf986b70d136ed9390bf83
2020-10-29IVGCVSW-5314 Create OptimizeForExclusiveConnectionTeresa Charlin
* FuseBatchNorm class has been added to facilitate testing * Only Convolution2D FP32 being fused Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com> Change-Id: I049c4770946ddca21b08516d4c9f4d0d22bf9b45
2020-09-15IVGCVSW-5305 AddBroadcastReshapeLayer as optimizerNarumol Prangnawarat
* Remove AddBroadcastReshapeLayer from TfLiteParser * Add AddBroadcastReshapeLayer as optimizer * AddBroadcastReshapeLayer optimizer unit tests * Load-scope dynamic tensor broadcasting unit tests Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com> Change-Id: I3549e85b71b41cbd4d96c0f1ece7887acbca76d1
2020-06-03remove BOM from filesLaurent Carlier
Change-Id: Ia4b4bb3be0ed6e933c77d58f8e9879b1370e9537 Signed-off-by: Laurent Carlier <laurent.carlier@arm.com>
2020-05-19IVGCVSW-4669 Set destination tensorInfo in MoveAllConnections()Finn Williams
Signed-off-by: Finn Williams <Finn.Williams@arm.com> Change-Id: I3563209dcb3db1b40cf2db3855adc631b5e323be
2020-04-10IVGCVSW-4483 Remove boost::polymorphic_downcastJan Eilers
* exchange boost::polymorphic_downcast with armnn::PolymorphicDowncast * remove unnecessary includes of boost::polymorphic_downcast Signed-off-by: Jan Eilers <jan.eilers@arm.com> Change-Id: Ie603fb82860fe05fee547dc78073230cc62b2e1f
2020-04-06IVGCVSW-4485 Remove Boost assertNarumol Prangnawarat
* Change boost assert to armnn assert * Change include file to armnn assert * Fix ARMNN_ASSERT_MSG issue with multiple conditions * Change BOOST_ASSERT to BOOST_TEST where appropriate * Remove unused include statements Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com> Change-Id: I5d0fa3a37b7c1c921216de68f0073aa34702c9ff
2020-03-26IVGCVSW-4597 Modify BF16 optimizer to Convert only inputs and weights ofNarumol Prangnawarat
Convolution2d and FullyConnected layers * Add InsertConvertFp32ToBf16LayersBefore * Add ConvertWeight to ConvertFp32NetworkToBf16Impl for Conv2d and FullyConnected * Allow different input and output when input is BF16 and output is FP32 Conv2d and FullyConnected layers * Unit tests Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com> Change-Id: Ic8f92ff28edcae08a72a3114a28f50c4619f919b
2020-03-20IVGCVSW-4520 Implement BFloat16 OptimizerNarumol Prangnawarat
* Add ReduceFp32ToBf16 to OptimizerOptions * Add ConvertFp32NetworkToBf16 * Add utility functions to insert conversion layers * Add constant conversion BF16 <-> FP32 * Unit tests Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com> Change-Id: Iaff77e20c721400b052cb37eb9ef6fe16d7abaff
2020-03-10IVGCVSW-4482 Remove boost::ignore_unusedJan Eilers
!referencetests:229377 Signed-off-by: Jan Eilers <jan.eilers@arm.com> Change-Id: Ia9b360b4a057fe7bbce5b268092627c09a0dba82
2020-03-03IVGCVSW-4375 Add support for Transpose to optimizationsMike Kelly
* Changed some existing Permutation specific optimizations to also support Transpose * Added MoveTransposeUp optimization * Added TransposeAsReshape optimization * Added tests for Transpose optimizations * Added missing layer tests for Transpose Signed-off-by: Mike Kelly <mike.kelly@arm.com> Change-Id: I20d099b284861402ae94aaa5dbf34907327a485f
2019-12-30IVGCVSW-4246 Clean build optimizations with -WextraDerek Lamberti
Change-Id: I2e0884c66855071eb3aa72b86de06c6ed6389d50 Signed-off-by: Derek Lamberti <derek.lamberti@arm.com>
2019-11-29MLCE-143 Fixed driver crash during CTS testsMike Kelly
* Only apply the Optimization when the base ReshapeLayer is connected to the child ReshapeLayer and no other Layer. Signed-off-by: Mike Kelly <mike.kelly@arm.com> Change-Id: Iccd676d657f9e7c829813f1bec9c82db8745d069
2019-11-29IVGCVSW-4209 Create a public API for the ArmNN UtilsMatteo Martincigh
* Moved the relevant armnnUtils headers to the new location: include/armnnUtils * Update the header usage throughout the source code !android-nn-driver:2387 Signed-off-by: Matteo Martincigh <matteo.martincigh@arm.com> Change-Id: I2ba15cebcacafad2b5a1a7b9c3312ffc585e09d6
generated by cgit v1.2.1 (git 2.23.0) at 2024-07-12 17:19:16 +0000