박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 09:26:26 +0000 (18:26 +0900)]
[res] TensorFlow Sqrt_000 (#6617)
This will introduce TensorFlow basic test for Sqrt node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 14 Aug 2019 09:22:37 +0000 (18:22 +0900)]
[loco] Clean up GraphTestcase (#6613)
Let's remove deprecated code in GraphTestcase.h
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
[mir_onnx] Reimplement Gemm converter through FullyConnected (#6604)
`GemmOp` is almost the same thing as `FullyConnectedOp`. It was only used in the ONNX importer.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 09:22:00 +0000 (18:22 +0900)]
[tf2tflite] Enabled test by FixedReshape support (#6609)
This commit adds several enabled tests in tf2tflite by newly supported
FixedReshape
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
[neurun] Fix fining scheduler and tryBackend (#6616)
CPU_DELAY must be just for parallel executor.
Also tryBackend: PR 5894 commented a line and this
reduced performance by 10%
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
[mir] Accout for data format in shape inference (#6584)
* Add several convenient functions to work with data format.
* Use the functions to implement data format aware shape inference.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 08:32:54 +0000 (17:32 +0900)]
[moco-tf] Enable TFSqrt Canonicalizer (#6614)
This will enable TFSqrt canonicalizer
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 08:29:58 +0000 (17:29 +0900)]
[moco-tf] Knobs for Reshape elimination transform (#6611)
This commit introduces knobs for following Reshape elimination
transforms
- ResolveDuplicateReshape
- ResolveRedundantReshape
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 08:29:12 +0000 (17:29 +0900)]
[moco-tf] TypeInference (#6593)
* [moco-tf] TypeInference
This will introduce TF dialect TypeInference and rule class
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* explicit use name space
* update commentf for TypeInference
* add todo
* update class name and comment
* comment for not supported yet node
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Wed, 14 Aug 2019 08:23:20 +0000 (17:23 +0900)]
[moco-tf] Resolve Alloc-Dealloc mismatch (#6599)
Let's use std::unique_ptr<float []> instead of std::unique_ptr<float>!
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 08:15:35 +0000 (17:15 +0900)]
[moco-tf] TFSqrt Canonicalizer (#6592)
* [moco-tf] TFSqrt Canonicalizer
This will introduce TFSqrt node Canonicalizer
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* remove unused header
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 08:00:24 +0000 (17:00 +0900)]
[loco] Forward Type inference (#6607)
This commit introduces type inference for forward node
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 07:40:44 +0000 (16:40 +0900)]
[moco-tf] GraphBuilder for TFSqrt (#6581)
* [moco-tf] GraphBuilder for TFSqrt
This will introduce GraphBuilder for TFSqrt to import Sqrt node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* rename var
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 07:29:40 +0000 (16:29 +0900)]
[loco/Service] FixedReshape shape inference (#6594)
This commit introduces shape inference logic of FixedReshape and its
test
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 06:58:50 +0000 (15:58 +0900)]
[loco/service] ShapeInferenceRule for multiple dialects (#6587)
* [loco/service] ShapeInferenceRule for multiple dialects
This adds a loco service, a ShapeInferenceRule for multiple dialects.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* fix typo
* modified include
* add #include <map>
* Remove always true assert
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 06:43:50 +0000 (15:43 +0900)]
[loco] remove assert cond that is always true in MultiDialectTypeInferenceRule (#6603)
This commit removes assert cond that is always true in MultiDialectTypeInferenceRule
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 06:10:42 +0000 (15:10 +0900)]
[tf2tflite] enable Const_000 (#6598)
This will enable Const_000 that was fixed in exo-tflite
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 05:43:32 +0000 (14:43 +0900)]
[moco-tf] GraphBuilder for TFRsqrt (#6579)
* [moco-tf] GraphBuilder for TFRsqrt
This will introduce GraphBuilder for TFRsqrt to import Rsqrt node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* rename var
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 05:39:02 +0000 (14:39 +0900)]
[loco/Service] FixedReshape type inference (#6595)
This commit introduces type inference of FixedReshape
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 05:23:30 +0000 (14:23 +0900)]
[locoex/customop] ShapeInference rule for custom op (#6585)
* [locoex/customop] ShapeInference rule for custom op
This commit adds COpShapeInference rule, which is shape inferring rule for custom op.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* add TODO
* wrong comment
* remove 'canonical' from comment
* wrong comment again: Type -> Shape
* wrong comment again and again... COP -> COp
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 05:21:05 +0000 (14:21 +0900)]
[exo-exporter] Remove unused header (#6596)
Let's remove unused header in OperationExporter.cpp
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 05:20:31 +0000 (14:20 +0900)]
[moco-tf] Introduce ResolveRedundantReshape transform (#6571)
* [moco-tf] Introduce ResolveRedundantReshape transform
This commit introduces ResolveRedundantReshape transform of canonical
optimize stage.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Use const arg
* Review fix: fine grained logic
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 05:18:05 +0000 (14:18 +0900)]
[moco-tf] Introduce TFSquaredDifference IR (#6580)
This will introduce TFSquaredDifference IR and required minimum changes
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 04:49:07 +0000 (13:49 +0900)]
[exo-tflite] Fix buffer export with adding empty (#6576)
* [exo-tflite] Fix buffer export with adding empty
This will add one empty buffer as first buffer to follow TF lite behavior
- This makes unit Const_000 test pass
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* remove test list change
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 03:12:54 +0000 (12:12 +0900)]
[moco-tf] Fix missing changed tag (#6583)
This commit fixes missing 'changed' setting on ResolveDuplicateReshape
transform
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 03:11:42 +0000 (12:11 +0900)]
[moco-tf] Comment for TFRsqrt IR (#6578)
This will add comment for TFRsqrt IR
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Aug 2019 03:10:40 +0000 (12:10 +0900)]
[locomotiv] EltwiseSqrt (#6532)
* [locomotiv] EltwiseSqrt
This will introduce execution code for EltwiseSqrt node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* add test
* update comment
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Wed, 14 Aug 2019 02:05:49 +0000 (05:05 +0300)]
[custom op] api addition (#6219)
* [custom op] Api addition to support custom op registration
Add nnfw_dev.h header with types required to define custom operations
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
* Roll back changes to nnfw.h, add separate header for custom kernel development
* format fix
* use limited dim count for typeinfo struct
* remove _t suffix for types
* fix missed _t suffixes for function types
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 01:56:31 +0000 (10:56 +0900)]
[loco] FixedReshape graph testcase (#6545)
This commit adds graph builder layer and testcase for FixedReshape
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 01:53:42 +0000 (10:53 +0900)]
[exo-tflite] Export FixedReshape operation (#6565)
This commit introduces export of FixedReshape operation
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 01:44:33 +0000 (10:44 +0900)]
[res] Introduce Softmax testcases (#6562)
* [res] Introduce Softmax testcases
This commit will introduce `Softmax` testcases
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
* remove versions
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 00:45:11 +0000 (09:45 +0900)]
[xtrace] Hoist loop invariant code (#6566)
Hoist loop invariant code in `benchmark_runner.cc` for readability.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 00:43:01 +0000 (09:43 +0900)]
Make acl_cl KernelGenerator supports NCHW frontend layout (#6552)
This commit makes acl_cl KernelGenerator supports NCHW frontend layout.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Wed, 14 Aug 2019 00:41:41 +0000 (09:41 +0900)]
Make acl_neon KernelGenerator supports NCHW frontend layout (#6551)
This commit makes acl_neon KernelGenerator supports NCHW frontend layout.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 13 Aug 2019 22:30:41 +0000 (07:30 +0900)]
[loco] Introduce services for TensorSoftmax (#6564)
This commit will introduce loco services for `TensorSoftmax`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 22:24:22 +0000 (07:24 +0900)]
[moco-tf] Clear ConcatV2 graph builder (#6547)
This will erase ConcatV2 graph builder implementation to reduce complexity of diff for migration to using VariadicArityNode
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 22:23:49 +0000 (07:23 +0900)]
[moco-tf] Introduce TFSqrt IR (#6553)
This will introduce TFSqrt IR to represent TensorFlow Sqrt node and required minimum implementation
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
[mir_onnx] Remove uses of DropoutOp (#6560)
This operation is a no-op.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
Павел Ильютченко/AI Tools Lab /SRR/Engineer/삼성전자 [Tue, 13 Aug 2019 19:54:47 +0000 (22:54 +0300)]
[mir2loco] Support Softmax operation in transformer (#6575)
* Added SoftmaxOp convertion and test for this
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
Павел Ильютченко/AI Tools Lab /SRR/Engineer/삼성전자 [Tue, 13 Aug 2019 18:50:30 +0000 (21:50 +0300)]
[mir2loco] Support Mul operation in transformer (#6573)
* Added MulOp convertion and test for this
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
Павел Ильютченко/AI Tools Lab /SRR/Engineer/삼성전자 [Tue, 13 Aug 2019 18:43:52 +0000 (21:43 +0300)]
[mir_onnx] Fix output name for softbackend (#6572)
* Setting empty output node name
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
[mir] Introduce data format (#6570)
* Add `DataFormat` enum class.
* Use it in convolutional and pooling operations.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir_caffe2] Remove uses of DropoutOp (#6561)
This operation is a no-op.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir_caffe] Remove uses of DropoutOp (#6558)
This operation is a no-op.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Remove uses of DropoutOp (#6556)
This operation is a no-op.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Eliminate IVisitor in nnc (#6519)
Eliminate abstract IVisitor uses in nnc componet and use Visitor instead
Signed-off-by: Efimov Alexander <a.efimov@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 10:40:01 +0000 (19:40 +0900)]
Don't build nnpackage_run if neurun is not built (#6557)
nnpackage_run is for neurun.
pacl should not try to build nnpackage_run.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 13 Aug 2019 10:18:46 +0000 (19:18 +0900)]
[loco] Introduce DepthwiseFilterEncode to CanonicalShapeInferenceRule (#6548)
This commit introduces DepthwiseFilterEncode to
CanonicalShapeInferenceRule.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 13 Aug 2019 09:59:28 +0000 (18:59 +0900)]
[loco/Service] Forward node shape inference (#6542)
This commit implements canonical shape inference for Forward node
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 09:28:02 +0000 (18:28 +0900)]
[neurun, nnpackage] Introduce nnpackage_run (#5720)
Initial version of nnpackage.
It supports only `run` with random values.
( store, compare, input are not supported yet. )
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 13 Aug 2019 08:35:35 +0000 (17:35 +0900)]
Make acl_cl KernelGenerator supports NCHW frontend layout (#6544)
This commit makes acl_cl KernelGenerator supports NCHW frontend layout.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 13 Aug 2019 08:01:41 +0000 (17:01 +0900)]
[loco] Remove "shape" setter from Push class (#6527)
This commit removes deprecated "shape" setter from Push class.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 07:40:22 +0000 (16:40 +0900)]
[moco-tf] Clear ConcatV2 Canonicalizer (#6521)
This will erase ConcatV2 Canonicalier implementation to reduce complexcity of diff for migration to using VariadicArityNode
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 07:36:43 +0000 (16:36 +0900)]
[moco-tf] Turning on the knob for ReorderDecodeTensorBiasAdd (#6535)
This commit turns on the knob for ReorderDecodeTensorBiasAdd
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 13 Aug 2019 07:33:10 +0000 (16:33 +0900)]
[exo-tflite] Introduce DepthwiseFilterEncode with exporting as reshape (#6524)
* [exo-tflite] Introduce DepthwiseFilterEncode with exporting as reshape
This commit introcude DepthwiseFilterEncode with exporting as reshape to
exo-tflite.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
* remove unused function.
* add some comment.
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 13 Aug 2019 07:29:22 +0000 (16:29 +0900)]
[moco-tf] Use official FixedReshape alias (#6541)
Now loco::FixedReshape alias is official, so this commit starts to use
the alias in moco-tf module.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 07:24:36 +0000 (16:24 +0900)]
[res] TF_SMALL_NET_0015 with 1.0 / Sqrt (#6539)
This will introduce TF_SMALL_NET_0015 model with Sqrt+RealDiv that works like Rsqrt
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 05:57:57 +0000 (14:57 +0900)]
[neurun] Implement nnfw_input/output_tensorinfo (#6530)
nnfw_input_tensorinfo and nnfw_output_tensorinfo are implemented.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 13 Aug 2019 05:28:47 +0000 (14:28 +0900)]
Apply layout for helper functions converting axis values (#6528)
This commit applies layout for helper functions converting axis values.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 04:52:01 +0000 (13:52 +0900)]
[moco-tf] Clear ConcatV2 fix shape (#6523)
This will erase ConcatV2 fix shape to reduce complexity of diff for migration to using VariadicArityNode
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 13 Aug 2019 04:50:54 +0000 (13:50 +0900)]
[mocotest-tf] Test script runnable from any directory (#6529)
Previously test script partially assumes it runs on WORKDIR, which makes
no harm at all. This commit just explicitly call file path including
WORKDIR to be matched with other lines convention and to be runnable from
any directory.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 03:51:51 +0000 (12:51 +0900)]
[loco] Inference service for EltwiseSqrt (#6526)
This will add Shape and Type inference for EltwiseSqrt
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 01:51:50 +0000 (10:51 +0900)]
[neurun] Implement nnfw_input_size and nnfw_output_size (#6495)
* [neurun] Implement nnfw_input_size and nnfw_output_size
This patch implements nnfw_input_size and nnfw_output_size.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
* [neurun] Update int to uint32_t for input/output_size
This patch updates int to uint32_t for input/output_size.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 13 Aug 2019 01:48:24 +0000 (10:48 +0900)]
Set lcov's gcov-tool param (#6525)
For cross build, set lcov's gcov-tool by GCOV_PATH environment variable. Default is gcov (native build).
nnfw's gen-coverage-report has same feature
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 01:47:12 +0000 (10:47 +0900)]
[moco-tf] Clear ConcatV2 import test (#6522)
This will erase ConcatV2 TF dialect import test to reduce complexity of diff for migration to using VariadicArityNode
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 13 Aug 2019 01:41:07 +0000 (10:41 +0900)]
[tf2tflite] Test script runnable from any directory (#6514)
Previously test script partially assumes it runs on WORKDIR, which makes
no harm at all. This commit just explicitly call file path including WORKDIR
to be matched with other lines convention and to be runnable from any directory.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Tue, 13 Aug 2019 01:34:00 +0000 (10:34 +0900)]
[loco] Using FixedReshape (#6496)
This commit introduces alias for Reshape<ReshapeType::Fixed> loco node.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Tue, 13 Aug 2019 01:29:47 +0000 (10:29 +0900)]
[loco] Introduce EltwiseSqrt (#6520)
This will introduce IR for elementwise Sqrt
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 13 Aug 2019 01:05:28 +0000 (10:05 +0900)]
[mlapse-tfl] Implement a command-line argument parser (#6512)
This commit implements a command-line argument parse for mlapse-tfl.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Tue, 13 Aug 2019 00:40:56 +0000 (09:40 +0900)]
Fix deadlock when access two acl_cl at once (#6500)
To avoid deadlock when operation need to access two acl_cl at once (source, target), introduce access function to access two tensor
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 12 Aug 2019 23:42:58 +0000 (08:42 +0900)]
[moco-tf] Introduce VariadicArityNode (#6472)
This will introduce TF dialect VariadicArityNode to manage Concat node
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 12 Aug 2019 23:33:30 +0000 (08:33 +0900)]
[locomotiv] supporting FeatureBiasAdd (#6504)
This code enables locomotiv to support FeatureBiasAdd.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 12 Aug 2019 22:10:09 +0000 (07:10 +0900)]
[moco-tf] Fix FuseBinaryIntoProceeding logic (#6511)
This commit fixes FuseBinaryIntoProceeding logic. Assertion was
misplace, so that normal addition blocked to this assertion.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 12 Aug 2019 17:23:51 +0000 (02:23 +0900)]
[mir2loco] Remove Push node's shape setter (#6487)
This commit revises all the code that invokes deprecated shape setter.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
Павел Ильютченко/AI Tools Lab /SRR/Engineer/삼성전자 [Mon, 12 Aug 2019 17:15:36 +0000 (20:15 +0300)]
[mir2loco] Remove unused visitors and use default visit_fallback (#6515)
* Remove unused visitors
* Override visit_fallback
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
[mir_caffe] Refactor importing of BatchNorm (#6467)
Structurize the code, fix comments.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
Павел Ильютченко/AI Tools Lab /SRR/Engineer/삼성전자 [Mon, 12 Aug 2019 17:11:17 +0000 (20:11 +0300)]
[mir2loco] Supported other types in ConstantOp transformer (#6516)
* Supported F32, F64, S32, S64 data types
* Temporary implementations for F64 and S64
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 12 Aug 2019 11:03:03 +0000 (20:03 +0900)]
[locop] Ignore the shape attribute for Push nodes (#6485)
CanonicalNodeSummaryBuilder now ignores (deprecated) tensor shape
attributes when it generates a node summary for Push nodes.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 12 Aug 2019 11:00:06 +0000 (20:00 +0900)]
[loco] Introduce Softmax node in loco (#6492)
This commit will introduce `Softmax` node in `loco`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 12 Aug 2019 10:59:29 +0000 (19:59 +0900)]
[tf2tflite] Loading customop.conf file into ModelSignature (#6406)
* [tf2tflite] Loading customop.conf file into ModelSignature
Adds code for loading customop.conf into ModelSignature
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* 1) class to function 2) remove loco from cmake file 3) remove some headers
* fix wrong lambda comment // use argv[5] directly
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 12 Aug 2019 10:34:39 +0000 (19:34 +0900)]
[neurun] Do not collect TensorBuilders one by one (#6458)
* [neurun] Do not collect TensorBuilders one by one
To get the full set of tensor builders currently using it used to
collect from all the `LowerInfo`s. Now that `BackendResolver` manages
them all we can just use it.
Allow allocation of 0-byte object
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 12 Aug 2019 10:14:17 +0000 (19:14 +0900)]
[exo-tflite] support DetphwiseFilterShape in shape_desc (#6491)
* [exo-tflite] support DetphwiseFilterShape in shape_desc
This commit supports DethwiseFilterShape in shape_desc.
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
* apply comment.
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 12 Aug 2019 10:13:45 +0000 (19:13 +0900)]
[neurun] change int to uint32_t for size of input/output (#6502)
neurun uses uint32_t internally. Runtime API uses int.
This patch will change Runtime API to use uint32_t.
(same type to `neurun` internal).
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 12 Aug 2019 10:04:31 +0000 (19:04 +0900)]
[moco-tf] Clean up trivials in FuseBinaryIntoPreceding (#6497)
This commit will clean up some codes
- Remove deprecated function
- Replace and correct comments
- Modify bracket formats
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
박천교/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 12 Aug 2019 10:02:26 +0000 (19:02 +0900)]
[moco-tf] Introduce ResolveDuplicateReshape (#6499)
ResolvedDuplicateReshape transform introduced. This transform finds
duplicate FixedReshape pair, and let the later one to bypass the
former.
Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 12 Aug 2019 10:00:58 +0000 (19:00 +0900)]
[mocotest-tf] Enable TF_SMALL_NET_0004 (#6498)
Since `FuseBinaryIntoPreceding` now supports for `DepthwiseConv2dNative`,
this commit will enable `TF_SMALL_NET_0004`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
채성우/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 12 Aug 2019 09:58:50 +0000 (18:58 +0900)]
[loco] fix a window_of function in CanonicalShapeInferenceRule (#6489)
This commit fix a window_of function.
Related issue:#6486
Signed-off-by: seongwoo <sw4670.chae@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 12 Aug 2019 09:13:41 +0000 (18:13 +0900)]
Use nnfw_coverage for gcov build (#6445)
For gcov build release target, use nnfw_coverage link
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 12 Aug 2019 09:13:09 +0000 (18:13 +0900)]
Replace coverage report command (#6476)
- Replace nncc's coverage report command into infra/command
- Introduce build.configuration setting
- Default workspace: build/
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 12 Aug 2019 07:50:00 +0000 (16:50 +0900)]
[neurun] Use nnfw api wrapper's return value (#6493)
Currently nnfw implementation returns no error.
Now, it will use the return value of wrapper.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 12 Aug 2019 07:49:45 +0000 (16:49 +0900)]
Remove index parameter from nnfw_input_size (#6488)
index parameter is removed that is not necessary.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
남궁석/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 12 Aug 2019 07:27:49 +0000 (16:27 +0900)]
[moco-tf] Support DepthwiseConv2dNative in FuseBinaryIntoPreceding (#6490)
This commit will enable supporting `DepthwiseConv2dNative` in `FuseBinaryIntoPreceding`
Signed-off-by: Seok NamKoong <sk.namkoong@samsung.com>
이한종/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 12 Aug 2019 07:21:09 +0000 (16:21 +0900)]
[neurun] Remove class `Plan` (#6450)
`Plan` is nothing more than a wrapping `FunctionSequence` so it is
redundant. It has been replaced with `FunctionSequence`.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 12 Aug 2019 05:05:02 +0000 (14:05 +0900)]
Print build error on TensorFlowVersionChecker build error (#6477)
TensorFlowConfig.cmake checks running result only.
When prebuilt libtensorflow library is built from Ubuntu 18.04,
and nncc is going to be built on Ubuntu 16.04, it says:
"you need tensorflow version 1.12.0".
This patch will check build error also.
Then, it prints error message to guide the developers
to use libtensorflow built on same version of Ubuntu.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
이춘석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 12 Aug 2019 04:56:19 +0000 (13:56 +0900)]
Disable downlaoding NNPACK on aarch64-tizen (#6438)
- DOWNLOAD_NNPACK=OFF on aarch64 tizen build
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 12 Aug 2019 04:28:10 +0000 (13:28 +0900)]
[moco-tf] Do not reset input for BiasAdd Canonicalizer (#6470)
This will remove reset input for old replaced BiasAdd node in Canonicalization and update the diagram
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 12 Aug 2019 04:27:47 +0000 (13:27 +0900)]
[moco-tf] Do not reset input for Conv2D Canonicalizier (#6471)
This will remove reset input for old replaced Conv2D node in Canonicalization and update the diagram
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 12 Aug 2019 03:33:52 +0000 (12:33 +0900)]
Introduce mlapse toolkit (#6441)
This commit introduces the inital version of mlapse toolkit.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
이춘석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 12 Aug 2019 03:17:15 +0000 (12:17 +0900)]
Fix build error (#6484)
- add missing semicolon on nnfw.h
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
이춘석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 12 Aug 2019 02:20:32 +0000 (11:20 +0900)]
API for Input/Output Tensor information (#6153)
* API for Input/Output Tensor information
- Add nnfw_tensor_info
- input/output size getter
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* Fix typo, tensor_info -> tensorinfo