윤지영/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 9 Sep 2019 09:53:30 +0000 (18:53 +0900)]
[backend/srcn] Add transpose conv kernel (#7248)
* [backend/srcn] Add transpose conv kernel
Srcn library supports raw-major(NCHW) and col-major(NHWC) image layout.
However this commit only supports col-major image layout.
And Srcn library takes HWOI kernel filter layout when the image layout is NHWC.
But in neurun, it takes OHWI kernel filter layout when using the same image layout.
So it need to permute codes for TransposeConv weight.
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Change the Permute function to be limited to kernel layout
* Use original data array without copying to new address
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 9 Sep 2019 09:37:11 +0000 (18:37 +0900)]
[locomotiv] Use const if possible (#7264)
"annot_data" and "annot_domain" do not change the node itself.
This commit revies the signature of these helpers to express this
property explicitly.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 07:55:29 +0000 (16:55 +0900)]
[moco-tf] Use shape_inference_done for helpers (#7288)
This will change to use shape_inference_done() for helpers in FixShapeTransform
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 07:40:54 +0000 (16:40 +0900)]
[logo] Test helper (#7289)
* [logo] Test helper
This will introduce TestHelper for logo with find_first_node_bytype() method
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* apply comments
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 9 Sep 2019 07:10:35 +0000 (16:10 +0900)]
[loco] Add LEP_000 document (#7128)
* [loco] Add LEP_000 document
This commit adds LEP_000 document which proposes a design change over
loco IR.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Fix typos
* Rename and add heading
* Add a comment
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 9 Sep 2019 07:08:55 +0000 (16:08 +0900)]
[loco] Implement TensorBoradcast::Mapping methods (#7279)
This commit adds missing TensorBroadcast::Mapping method implementations
with related tests.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 9 Sep 2019 07:08:07 +0000 (16:08 +0900)]
Introduce fipe library (#7281)
* Introduce fipe library
This commit introduces fipe, which permits function chaining with
UNIX pipe-like syntax.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Add License Notice
* Add a comment
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 9 Sep 2019 07:07:08 +0000 (16:07 +0900)]
[logo] Use GTest_AddTest helper (#7285)
This commit simplifies logo's CMakeLists.txt using GTest_AddTest.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 07:02:30 +0000 (16:02 +0900)]
[exo-tflite] Move Convert.h/cpp, GraphBlock.h/cpp to src (#7291)
Convert.h/cpp, GraphBlock.h/cpp are moved into src.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
김용섭/On-Device Lab(SR)/Engineer/삼성전자 [Mon, 9 Sep 2019 06:07:33 +0000 (15:07 +0900)]
[neurun] Apply TensorManager instead of MemoryManager (#7251)
* [neurun] Apply TensorManager instead of MemoryManager
Apply TensorManager instead of MemoryManager. It looks huge but there
are the only many changes instead TensorManager of MemoryManager
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
* Fix format
* Adding finalize() and moving wrap_tensor() in dataflow
* Fix wrong code in dataflow
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 04:32:36 +0000 (13:32 +0900)]
[exo-tflite] Adding test assert into TFLTypeInferenceRuleTest (#7284)
This adds more asserts into the test body of TFLTypeInferenceRuleTest.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 04:20:04 +0000 (13:20 +0900)]
[exo-tflite] Adding test assert into TFLShapeInferenceRuleTest (#7283)
This adds more asserts into the test body of TFLShapeInferenceRuleTest.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 03:42:48 +0000 (12:42 +0900)]
[exo-tflite] Attaching progress reporter into convert_to_TFLNodes() (#7280)
This attaches progress reporter into the phase of convert_to_TFLNodes().
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 03:38:09 +0000 (12:38 +0900)]
[exo-tflite] Helper macro and class for test (#7255)
* [exo-tflite] Helper macro and class for test
Adding a macro to count the number of node and a class to easily run a pass to test.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* phase -> _phase
Ivan Vagin/AI Tools Lab /SRR/Engineer/삼성전자 [Mon, 9 Sep 2019 03:04:41 +0000 (06:04 +0300)]
[tflite_loader] Implemented TransposeConv node conversion (#7268)
* [tflite_loader] Implemented TransposeConv node conversion
Implemented transpose conv node conversion
Signed-off-by: Ivan Vagin <ivan.vagin@samsung.com>
* Fixed formatting
Signed-off-by: Ivan Vagin <ivan.vagin@samsung.com>
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Mon, 9 Sep 2019 01:48:42 +0000 (04:48 +0300)]
[custom op] Fix custom kernel configuration (#7238)
Previously tensor address was accidently passed instead of tensor buffer
Correctly store userdata_size insude CustomKernel::configure
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
Ivan Vagin/AI Tools Lab /SRR/Engineer/삼성전자 [Mon, 9 Sep 2019 01:45:41 +0000 (04:45 +0300)]
[tflite_loader] Implemented Mean node conversion (#7267)
* [tflite_loader] Implemented Mean node conversion
* implemented Mean node conversion
* changed MeanNode keep_dims param type to bool
Signed-off-by: Ivan Vagin <ivan.vagin@samsung.com>
* Fixed formatting
Signed-off-by: Ivan Vagin <ivan.vagin@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 9 Sep 2019 01:43:53 +0000 (10:43 +0900)]
[loco] Introduce ShapeInferenceRule v2 API (#7265)
This commit extends ShapeInferenceRule with a new version of infer API.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Mon, 9 Sep 2019 01:38:17 +0000 (10:38 +0900)]
[loco] Introduce TensorBroadcast node (#7262)
This commit extends loco.canonical with TensorBroadcast node.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Mon, 9 Sep 2019 01:34:46 +0000 (10:34 +0900)]
[moco-tf] Introduce shape_inference_done (#7276)
This will introduce shape_inference_done() method in FixShapeTransform as a helper function not to directly access ShapeInferenceData annotation to check shape inference is done or not
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
[mir] Refactor graph dot dumper (#7273)
Refactor graph dot dumper to loosen dependencies.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Use dumpGraph function (#7243)
Use `dumpGraph` function instead of `IrDotDumper` class.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 6 Sep 2019 10:25:35 +0000 (19:25 +0900)]
[moco-tf] Remove unused code in FixShapeTransform (#7266)
This will remove unused code in FixShapeTransform
- Shape inference for canonical dialect nodes are handled by loco service
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 6 Sep 2019 10:25:20 +0000 (19:25 +0900)]
[exo-tflite] Put "exo::" prefix (to easily distinguish log from exo) (#7258)
It is difficult to tell which (moco? exo-tflite?) prints log. So, "exo::" prefix is added for exo-tflite.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 6 Sep 2019 10:15:57 +0000 (19:15 +0900)]
Check OOR exception intflite_benchmark (#7260)
Add try-catch to catch OOR exception by substr
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 6 Sep 2019 09:00:47 +0000 (18:00 +0900)]
[exo-tflite] Adding ReluConverter into convert_to_TFLNodes() (#7257)
* [exo-tflite] Adding ReluConverter into convert_to_TFLNodes()
This adds ReluConverter and type/shape inference into convert_to_TFLNodes().
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* saturate -> restart
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 6 Sep 2019 08:58:09 +0000 (17:58 +0900)]
[res] TF test for Add broadcast with const scala (#7250)
This will add TF test material of Add operator with rank 4 tensor with const scala
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 6 Sep 2019 08:57:46 +0000 (17:57 +0900)]
[moco-tf] Skip canonical node fix shape (#7244)
This will skip canonical nodes fix shapes that is done as loco service
- Pull and Push nodes are still called as they exist in import
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 6 Sep 2019 08:57:22 +0000 (17:57 +0900)]
[moco-tf] Fix SquaredDifferenceCanonicalizer (#7245)
This will fix SquaredDifferenceCanonicalizer to check input nodes shape
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 6 Sep 2019 08:57:00 +0000 (17:57 +0900)]
[moco-tf] Remove unused ForwardShapeInferenceAlgorithm (#7246)
This will remove not used anymore ForwardShapeInferenceAlgorithm
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 6 Sep 2019 08:55:56 +0000 (17:55 +0900)]
[res] Conv2DBackpropInput with 3x3 stride 2 (#7194)
This will add TensorFlow test material for Conv2DBackpropInput with 3x3 filter and stride 2
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 6 Sep 2019 08:29:06 +0000 (17:29 +0900)]
Introduce bino (#7247)
This commit introduces bino, a library that facilitates std::pair
manipulation.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 6 Sep 2019 07:46:11 +0000 (16:46 +0900)]
[exo-tflite] change TODO (TFL node list to canonical node list) (#7256)
Changing TODO list from (wrong) TFL node list to canonical node list.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Fri, 6 Sep 2019 07:33:46 +0000 (16:33 +0900)]
[res] TF Test for Instance normalization (#7198)
This will add TF network test material of small input size of Instance normalization
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 6 Sep 2019 07:27:46 +0000 (16:27 +0900)]
[loco] Do NOT invoke shape_get directly (#7231)
This commit replaces all the loco::shape_get calls in ForwardShapeInferenceAlgorithm
implemention with "node_shape" method to facilitate refactoring.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 6 Sep 2019 06:45:21 +0000 (15:45 +0900)]
[locop] Introduce ExampleGraph (#7229)
Let's reuse common test code.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 6 Sep 2019 05:53:44 +0000 (14:53 +0900)]
Remove CLPermuteEx layer and kernel (#7253)
CLPermuteEx is not used anymore in neurun
- Remove this implementation
- Fix pacl to use CLPermute (used in neurun acl-cl backend)
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
Ivan Vagin/AI Tools Lab /SRR/Engineer/삼성전자 [Fri, 6 Sep 2019 05:53:07 +0000 (08:53 +0300)]
[neurun] Implemented test for scheduler profiling mode (#7034)
Implemented test for scheduler with enabled profiling mode
Fixed scheduling order by increasing et of one branch
Signed-off-by: Ivan Vagin <ivan.vagin@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 6 Sep 2019 05:38:23 +0000 (14:38 +0900)]
Remove GenericFullyConnectedLayer implementation (#7252)
GenericFullyConnectedLayer is not used anymore in neurun
- Remove this implementation
- Fix pacl to use CLFullyconnectedReshpingLayer (used in neurun acl-cl backend)
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 6 Sep 2019 04:26:13 +0000 (13:26 +0900)]
Add nnpkg test in coverage test (#7249)
Run nnpkg test for coverage test if nnpkg test suite exists
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Fri, 6 Sep 2019 03:40:17 +0000 (12:40 +0900)]
Use option for srcn kernel build (#7222)
Change to use srcn kernel build option BUILD_SRCN_KERNEL
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 23:03:38 +0000 (08:03 +0900)]
[moco-tf] Remove canonical concat axis test (#7230)
This will remove Canonical Concat IR test for axis
- Canonical IRs will not be handled in shape inference anymore
- Current Concat axis calculation is done in shape inference step
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
[mir] Add dumpGraph function (#7242)
Add `dumpGraph` function as a future replacement of `IrDotDumper`.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir] Revive IVisitor in proper way (#6657)
Return IVisitor as a strict version of Visitor
Propose to not use IVisitor outside of MIR component
Add missing includes and licenses
Signed-off-by: Efimov Alexander <a.efimov@samsung.com>
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Add missing include (#7241)
Add missing include of `Visitor.h`.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir2loco] Add missing include (#7240)
Add missing include of `Visitor.h`.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[mir] Make Shape constructor explicit (#7239)
Make `Shape` constructor explicit again.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
[nnc] Remove conversions of strides vector to Shape (#7237)
Remove implicit conversions of strides to `Shape`.
Add helper functions to work with strides.
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 5 Sep 2019 12:33:07 +0000 (21:33 +0900)]
Enable some tests for acl neon (#7235)
This commit eanbles some tests for acl neon.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Thu, 5 Sep 2019 12:13:19 +0000 (15:13 +0300)]
[custom op] (fix) Implement missing ShapeFixer method in cpu backend (#7236)
Provide empty imlpementation for ShapeFixer::visit(CustomNode&)
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
이성재/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 11:24:29 +0000 (04:24 -0700)]
Update `tflite_convert.sh` script (#6869)
- Change to use the 'tensorflow' module installed on the system by
default.
Signed-off-by: Sung-Jae Lee <sj925.lee@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 5 Sep 2019 11:16:37 +0000 (20:16 +0900)]
Enable Split op for ACL neon (#7226)
This commit enables to support Split op for ACL neon.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 5 Sep 2019 11:16:03 +0000 (20:16 +0900)]
Use type convert helper in tflite loader (#7232)
Use activation type convert helper for sub operation
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 5 Sep 2019 11:15:52 +0000 (20:15 +0900)]
Enable binary Logical ops for ACL neon (#7233)
This commit enables to support binary Logical ops for ACL neon.
- Implement kernel of acl_neon for binary logical ops
- Enable to support these ops for acl_neon on the neurun
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Thu, 5 Sep 2019 10:14:25 +0000 (13:14 +0300)]
[custom op] Implement custom operation loading in tflite_loader (#7182)
Load custom operations during model input
Loader assumes custom operations were already registered before model loading
Pass userdata size to CustomNode constructor
update custom op creation
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 09:36:53 +0000 (18:36 +0900)]
[mocotest-tf] Enable UNIT_StopGradient_001 (#7224)
This will enable UNIT_StopGradient_001 that is rank 4 test of StopGradient
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 09:36:33 +0000 (18:36 +0900)]
[tf2tflite] Enable UNIT_StopGradient_001 (#7223)
This will enable UNIT_StopGradient_001 that is rank 4 test of StopGradient
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 09:35:41 +0000 (18:35 +0900)]
[exo-tflite] introducing TypeInferencePass (#7225)
TypeInferencePass annotates dtype of nodes (canonical/TFL/COp nodes).
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 09:34:19 +0000 (18:34 +0900)]
[exo-tflite] Introducing ProgressReporter listner (#7221)
* [exo-tflite] Introducing ProgressReporter listner
ProgressReporter is copied from moco-tf to print information before/after phase.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
* CI error -> remove unused parameter
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 09:33:45 +0000 (18:33 +0900)]
[exo-tflite] Adding list of converters as TODO (#7228)
This adds list of converters as TODO.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 09:07:15 +0000 (18:07 +0900)]
[exo-tflite] Prepare slots for template initialization (#7227)
This adds slots for template initialization in CanonicalNodeConverter, which prevents code conflict.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 09:06:06 +0000 (18:06 +0900)]
[cusop] replace new with make_shared (#7211)
custom::KernelRegistry will be created with make_shared.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Thu, 5 Sep 2019 08:35:07 +0000 (11:35 +0300)]
[custom op] Build custom kernels (#6964)
* [custom op] Build custom kernels
Update cpu backend's `KernelGenerator` to build custom operation kernels
* Fix allocations in custom node kernel generator
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
윤지영/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 5 Sep 2019 08:32:18 +0000 (17:32 +0900)]
Add new backend for srcn (#7132)
* [neurun] Add new backend for srcn
This patch has a basic framework for adding new backend.
In practice, normal behavior is not implemented.
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Fix an issue where arm_neon.h file is not always found
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 08:21:55 +0000 (17:21 +0900)]
[moco-tf] Return for ShapeInferencePass (#7220)
This will return value from loco shape inference in ShapeInferencePass
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 08:19:36 +0000 (17:19 +0900)]
[moco-tf] Use loco::shape in import (#7217)
This will revise import to use loco shape instead of ShapeInferenceData
- also add guards for safety
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 08:10:12 +0000 (17:10 +0900)]
[loco] Enhancing TypeInference::to() (#7206)
Now 1) TypeInference::to()returns bool 2) checks previously inferred dtype
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 08:07:23 +0000 (17:07 +0900)]
[exo-tflite] Introducing ReluConverter (#7216)
This converter converts loco::Relu to locoex::TFLRelu.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 5 Sep 2019 08:02:19 +0000 (17:02 +0900)]
[moco-tf] Skip FuseBinaryIntoPreceding on Scalar (#7219)
The current implementation of FuseBinaryIntoPreceding gets stuck when
it encounters scalar constants.
Let's make FuseBinaryIntoPreceding skip on this case.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Thu, 5 Sep 2019 07:54:48 +0000 (10:54 +0300)]
[custom op] Free custom op userdata in CustomNode destructor (#7213)
- Makes userdata sized as this is required to use tflite flexbuffers format
- Changes userdata type to char* in order to properly delete it
- initialize custom kernel fields
- free memory using delete[]
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
이상규/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 07:50:36 +0000 (16:50 +0900)]
[cus_op] remove unnecessary header dependency in nnfw_dev.hpp (#7209)
Move the include to implementation to remove depedencies.
Forwarding declarations are enough:
- Execution.h
- Graph.h
- CustomKernelRegistry.h
Unnecessary dependency:
- compiler/Compiler.h
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 5 Sep 2019 07:40:25 +0000 (16:40 +0900)]
[locop] Use shared FormattedTensorShape (#7212)
The current implementation of CanonicalNodeSummaryBuilder implements its
own TensorShape formatter.
Let's replace this local formater with the shared one.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 5 Sep 2019 07:38:51 +0000 (16:38 +0900)]
[locop] Show value domain (#7202)
LinearV1 Graph Formatter now shows domain if domain(shape) is known.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 07:21:13 +0000 (16:21 +0900)]
[moco_tf] Do shape and type inference in TFOptimizer (#7215)
This will run Shape and Type inference while TF dialect optimizations
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 07:07:26 +0000 (16:07 +0900)]
[exo-tflite] introducing ShapeInferencePass (#7214)
ShapeInferencePass annotates shape of nodes (canonical/TFL/COp nodes).
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 06:55:08 +0000 (15:55 +0900)]
[exo-tflite] type inference for TFLRelu (#7208)
This adds type inference for TFLRelu.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 06:53:12 +0000 (15:53 +0900)]
[exo-tflite] shape inference for TFLRelu (#7210)
This adds shape inferene for TFLRelu.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 5 Sep 2019 06:19:55 +0000 (15:19 +0900)]
Scripts to make nnpkg testsuite and test (#7199)
- Introduce script to make nnpkg testsuite in CI build server
- Introduce script to test nnpkg testsuite in xu4 farm
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 5 Sep 2019 06:15:21 +0000 (15:15 +0900)]
Enable Pad op for ACL neon (#7205)
This commit enables to support Pad op for ACL neon.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 05:51:09 +0000 (14:51 +0900)]
[locoex] Enhance COpTypeInference test (#7204)
This includes CanonicalTypeInferenceRule check. This is required for future change of loco::TypeInference, which looks into the dtype of all input nodes. So, the type of loco::Pull should be inferred to infer the type of locoex::COpCall.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 5 Sep 2019 05:27:21 +0000 (14:27 +0900)]
Update armcl-ex library cmake (#7167)
Use finding arm compute package instead of checking target arch
Use target_compile_definition for preprocessor and merge
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 05:26:44 +0000 (14:26 +0900)]
[res] TF test for SquaredDifference broadcast (#7190)
This will add SquaredDifference test material for broadcast operation
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 05:26:12 +0000 (14:26 +0900)]
[res] TF test for Add broadcast with scala (#7195)
This will add TF test material of Add operator with rank 4 tensor with scala
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 05:25:35 +0000 (14:25 +0900)]
[moco-tf] Use ShapeInferenceData for rule (#7184)
* [moco-tf] Use ShapeInferenceData for rule
This will revise TFShapeInferenceRule infer to directly use ShapeInferenceData
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* change scope
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 05:25:07 +0000 (14:25 +0900)]
[res] TF test for rank 4 StopGradient (#7191)
This will add TensorFlow test material for rank 4 StopGradient operator
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 05:24:40 +0000 (14:24 +0900)]
[res_tf] TF test for Sub broadcast (#7192)
This will add TensorFlow test material for Sub broadcast operation
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 05:24:21 +0000 (14:24 +0900)]
[exo-tflite] Added TODO for each TFL Nodes in shape & type inference code (#7203)
This creates a slots for visit() of each TFL node.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 5 Sep 2019 05:22:56 +0000 (14:22 +0900)]
[loco] Introduce TensorAxis header (#7188)
Currently, PermutingCodec internally declares TensorAxis alias.
Let's extract this as a separate header and reuse it!
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 5 Sep 2019 05:22:52 +0000 (14:22 +0900)]
Enable LSTM op for ACL neon (#7201)
This commit enables to support LSTM op for ACL neon.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 5 Sep 2019 05:22:34 +0000 (14:22 +0900)]
[locop] NodeSummary with Comments (#7193)
This commit allows usersr to add multi-line comments to NodeSummary, and
revises LinearV1 Graph Formatter to show these comments properly.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
장지섭/On-Device Lab(SR)/Engineer/삼성전자 [Thu, 5 Sep 2019 05:22:17 +0000 (14:22 +0900)]
Enable LocalResponseNormalization op for ACL neon (#7149)
This commit enables to support LocalResponseNormalization op for ACL neon.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 5 Sep 2019 05:11:44 +0000 (14:11 +0900)]
Introduce tfts v0.1 (#7165)
* Introduce tfts v0.1
This commit introduces TensorFlow Testcase Service (tfts) v0.1.
The current version supports basic shape check on TensorFlow testcases
through nnkit-tf.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
* Use simple tfts_check instead of tfts_shape_check
윤현식/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 05:09:23 +0000 (14:09 +0900)]
[exo-tflite] Remove TFLShapeAnnot and TFLTypeAnnot (#7197)
Instead of using TFLShapeAnnot and TFLTypeAnnot, we will have type and shape inference for each TFL Nodes. So these annot classes is removed.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
박종현/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 5 Sep 2019 04:31:51 +0000 (13:31 +0900)]
[locop] Remove symbol_lookup helper (#7196)
This helper is no longer meaningful.
Signed-off-by: Jonghyun Park <jh1302.park@samsung.com>
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Thu, 5 Sep 2019 04:28:40 +0000 (07:28 +0300)]
[custom op] Implement custom op registration using public API (#7181)
* [custom op] Implement custom op registration using public API
Adds public custom operation registration method
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
* format fix
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 04:25:51 +0000 (13:25 +0900)]
[res] TF test for Mul broadcast (#7187)
This will add two Mul test materials for broadcast operation
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Thu, 5 Sep 2019 04:25:31 +0000 (13:25 +0900)]
[res] TF test for Add broadcast (#7186)
This will add two Add TF test materials for broadcast operation
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
Vladimir Plazun/AI Tools Lab /SRR/Engineer/삼성전자 [Thu, 5 Sep 2019 04:12:52 +0000 (07:12 +0300)]
[custom op] overload getName in CustomNode implementation (#7180)
Fixes CustomNode class being abstract
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 [Thu, 5 Sep 2019 02:13:11 +0000 (11:13 +0900)]
Update srcn kernel library cmake (#7135)
Use finding arm_neon.h instead of checking target arch
Link libraries OpenMP_CXX_LIBRARIES
Prepare compile definition NCNN and comment out
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
박세희/On-Device Lab(SR)/Principal Engineer/삼성전자 [Wed, 4 Sep 2019 22:30:46 +0000 (07:30 +0900)]
[loco] Revise shape inference condition (#7161)
* [loco] Revise shape inference condition
This will add condition to ShapeInference to infer when target node shape is unknown and input(s) shape is ready
Signed-off-by: SaeHie Park <saehie.park@samsung.com>
* remove print code