오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Wed, 21 Nov 2018 08:27:17 +0000 (17:27 +0900)]
Move OperationValidator into compiler (#3662)
Move OperationValidator into compiler directory
Fix namespace
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Wed, 21 Nov 2018 08:00:55 +0000 (17:00 +0900)]
[neurun] Graph holds Model as pointer (#3663)
Graph holds Model as pointer. This is for the next work that `_model` is moved
to Linear.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Wed, 21 Nov 2018 06:29:26 +0000 (15:29 +0900)]
Merge middleend into compiler directory (#3658)
Move SubTensorAnalyzer.cc/h into compiler directory
Fix namespace and cmake
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
Shubham Gupta/SNAP /SRI-Bangalore/Engineer/삼성전자 [Wed, 21 Nov 2018 06:27:35 +0000 (11:57 +0530)]
Removing unused variables and rearranging by name in TransposeConv (#3656)
This patch will remove unused variables and rearranges the variables by their name
Signed-off-by: shubham <shub98.gupta@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Wed, 21 Nov 2018 01:41:19 +0000 (10:41 +0900)]
Introduce Compiler class (#3632)
Introduce neurun::compiler::Compiler class to hide compilation phase from frontend
Introduce src/compiler directory for compile phases
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
Vishal Keshav/SNAP /SRI-Bangalore/Engineer/삼성전자 [Wed, 21 Nov 2018 01:26:50 +0000 (06:56 +0530)]
Added generated test data for CUSTOM unpack ops (#3638)
The input tensor data is treated as buffer.
Along any given axis, the output buffer are to be generated
according to the strides (determined by axis and num_splits)
Signed-off-by: Vishal Keshav <vishal.k1@samsung.com>
Shubham Gupta/SNAP /SRI-Bangalore/Engineer/삼성전자 [Wed, 21 Nov 2018 01:01:17 +0000 (06:31 +0530)]
Introduce LOGICAL_OR_EX in NeuralNetworksEx.h (#3618)
This patch Introduces LOGICAL_AND_EX in NeuralNetworksEx.h.
Related issue: #3459
Signed-off-by: shubham <shub98.gupta@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Tue, 20 Nov 2018 10:09:36 +0000 (19:09 +0900)]
Generate text file to save build path depth in coverage build (#3651)
Generate text file tools/test_driver/build_path_depth.txt to save build path depth.
gcov for cross build need directory strip to match build host directory and target directory.
Build path depth will be used to set GCOV_PREFIX_STRIP environment variable on target.
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Tue, 20 Nov 2018 10:09:14 +0000 (19:09 +0900)]
Fix comment for argmax_ex in NeuralNetworksEx.h (#3606)
Fix incorrect comment for output of argmax_ex in NeuralNetworksEx.h
- Output rank should be same rank with input
- Reduced axes are retained with length 1
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
NeuralNetworkEx code cleanup (#3649)
Just little code cleanup for NeuralNetworkEx.h
Signed-off-by: Tanuj Tekriwal <tanuj.tekri@samsung.com>
[neurun] Fix TensorBuilder to add subtensors to Plan::operands (#3614)
Fix TensorBuilder to add subtensors to Plan::operands
Related issue is #3611
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
Introduce LOGICAL_AND_EX in NeuralNetworksEx.h (#3563)
This patch Introduces LOGICAL_AND_EX in NeuralNetworksEx.h.
Related issue: #3459
Signed-off-by: prasannar <prasanna.r@samsung.com>
Added Generic test file for RSQRT_EX (#3297)
This patch adds the Gtest file for RSQRT_EX.
Signed-off-by: Tanuj Tekriwal <tanuj.tekri@samsung.com>
윤지영/동작제어Lab(SR)/Engineer/삼성전자 [Tue, 20 Nov 2018 05:13:25 +0000 (14:13 +0900)]
Pass all arguments in summarize_tflite tool (#3646)
To use various options of model_parser.py tool,
it allows to pass all arguments to model_parser.py file in summarize_tflite file.
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
윤지영/동작제어Lab(SR)/Engineer/삼성전자 [Tue, 20 Nov 2018 05:13:15 +0000 (14:13 +0900)]
Fix non-string passed to CreateString issue in tfltool (#3645)
If the operator_codes is empty, select_operator.py tool occurs an error.
This patch is fixed the issue.
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Tue, 20 Nov 2018 05:01:14 +0000 (14:01 +0900)]
[neurun] Rename Planner as OperationValidator (#3582)
Now Planner has only have assertions, so rename accordingly.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
Add gtest to test for quant8 input for Pad op (#3620)
This patch will add gtest to test for quant8 input for pad op.
Signed-off-by: shubham <shub98.gupta@samsung.com>
Add GTests for SQRT op (#3608)
This patch addds GTests for SQRT op.
Added files:
```
sqrt_ex_1D_float
sqrt_ex_4D_float
```
Signed-off-by: prasannar <prasanna.r@samsung.com>
Removing unnecessary included header files in NENormalization (#3621)
This patch will remove unnecessary headers included in src and runtime
of NENormalization.
Signed-off-by: shubham <shub98.gupta@samsung.com>
Introduce LogicalOr op in runtime (#3622)
This patch will add node for LogicalOr op in PACL
Signed-off-by: shubham <shub98.gupta@samsung.com>
Correcting the comments in equal and notequal cl (#3627)
This patch will corrects the comments in equal and notequal cl
And Adds the missing comments.
Signed-off-by: shubham <shub98.gupta@samsung.com>
Removing unnecessary headers included in clfiles of Permute op (#3628)
This patch will remove unnecessary headers included in clfiles of Permute op.
Signed-off-by: shubham <shub98.gupta@samsung.com>
Correction in Error checking conditions and namespace missalignment (#3630)
This patch will correct:
1. error checking condition for output, since output is always QASSYM8
input1 and output datatypes cannot be compared. Also, output can omnly be
checked for QASSYM8 in validate function.
2.Since all the non class functions are placed in namespace (format followed
by ACL) hence placing validate the function in the defined namespace.
Signed-off-by: shubham <shub98.gupta@samsung.com>
Minor changes in TransposeConv op (#3619)
* Minor changes in TransposeConv op
Addressing the comments by jh1302-park in #3423
Signed-off-by: b.devansh <b.devansh@samsung.com>
* Minor changes in TransposeConv op
Addressing the comments by jh1302-park in #3423
Signed-off-by: b.devansh <b.devansh@samsung.com>
Introducing LogicalNot op in NeuralNetworksEx (#3598)
This commit introduces LogicalNot op in NeuralNetworksEx.h
Related issues: #3596
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
김용섭/동작제어Lab(SR)/Engineer/삼성전자 [Tue, 20 Nov 2018 02:03:54 +0000 (11:03 +0900)]
[model_parser] Fix wrong printing type of tensor (#3635)
* [model_parser] Fix wrong printing type of tensor
Fixes wrong printing type of tensor. The type names like FLOAT32,
UINT8 not 0 should be printed.
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
* Fix to get type_name from tensor_wrapping
Nnapi delegate for ArgMax (#3558)
The changes introduces the nnapi delegate for ArgMax operation.
Signed-off-by: Praveen D N <praveen.dn@samsung.com>
Add generated test for ARG_MAX (#3554)
The changes adds the generated test's for ARG max operation.
Signed-off-by: Praveen D N <praveen.dn@samsung.com>
장지섭/동작제어Lab(SR)/Engineer/삼성전자 [Mon, 19 Nov 2018 07:44:12 +0000 (16:44 +0900)]
Apply CL Kernel of Embedding_Lookup to PACL (#3616)
This commit applies CL Kernel of Embedding_Lookup to PACL.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
Add gtest to test Transpose op for Tensors with rank < 4 (#3601)
This patch will add gtest to verify Transpose op for input tensors
having rank < 4
Signed-off-by: shubham <shub98.gupta@samsung.com>
윤현식/동작제어Lab(SR)/Principal Engineer/삼성전자 [Mon, 19 Nov 2018 02:02:46 +0000 (11:02 +0900)]
[model_parser] Prints optons for Depthwise_Conv_2D (#3613)
This commit adds feature to print optons for Depthwise_Conv_2D.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Mon, 19 Nov 2018 02:02:35 +0000 (11:02 +0900)]
[model parser] Fix parsing bug when operator don't have option (#3610)
Fix parsing bug invaild exit when operator don't have option
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
장지섭/동작제어Lab(SR)/Engineer/삼성전자 [Mon, 19 Nov 2018 01:22:51 +0000 (10:22 +0900)]
Change SimpleSpaceToBatchND's inputs to arm_compute::ITensor (#3553)
This commit changes SimpleSpaceToBatchND's inputs to arm_compute::ITensor.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
윤지영/동작제어Lab(SR)/Engineer/삼성전자 [Fri, 16 Nov 2018 08:13:41 +0000 (17:13 +0900)]
Add freeze_graph script into tflkit (#3556)
freeze_graph tool will freeze a standalone graphdef with the stored variables.
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
Add GTest for testing the enabled Broadcast of EqualEx op (#3304)
This patch adds GTest for testing the enabled Broadcast of EqualEx op.
Added GTests:
```
equal_ex_broadcast
equal_ex_broadcast_4D_2D
```
Signed-off-by: prasannar <prasanna.r@samsung.com>
Implementation of TransposeConv op in PACL for CPU (#3423)
This patch will add TransposeConv op
Related issues : #2739 #2809
Signed-off-by: b.devansh <b.devansh@samsung.com>
장지섭/동작제어Lab(SR)/Engineer/삼성전자 [Fri, 16 Nov 2018 06:30:26 +0000 (15:30 +0900)]
Implement CL kernel of EmbeddingLookup (#3489)
This commit implements CL kernel of EmbeddingLookup op.
- Add an IFunction class for EmbeddingLookup.
- Add an ICLKernel class for EmbeddingLookup.
- Add cl kernel functions for EmbeddingLookup.
- Append the cl functions to kernel_program_map.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
Introducing LogicalNot op in pure_arm_compute runtime (#3599)
This commit introduces LogicalNot op in in pure_arm_compute runtime
Related PR is #3598
Related issues: #3596 and #3459
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
Pack generated test update (#3557)
The changes updates the pack operation in all_generated_test.
Signed-off-by: Praveen D N <praveen.dn@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Fri, 16 Nov 2018 02:43:23 +0000 (11:43 +0900)]
Revise x86 build script (#3604)
x86 build by non-root anomymous user: same userid with host user
Use docker-run-user command and remove docker setting code
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
윤지영/동작제어Lab(SR)/Engineer/삼성전자 [Thu, 15 Nov 2018 07:27:31 +0000 (16:27 +0900)]
Add transform_graph script into tflkit (#3545)
* Add transform_graph script into tflkit
transform_graph tool will transform a TensorFlow GraphDef applying some transformations.
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Update detailed usage to README.md
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 15 Nov 2018 07:10:18 +0000 (16:10 +0900)]
[tfltool] Update option parsing in select operator for v1.12 (#3594)
Add comment for new options
Fix DepthwiseConv2DOptions for additional field
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Thu, 15 Nov 2018 04:32:30 +0000 (13:32 +0900)]
[neurun] Remove IPlanBuilder::addShapeConstr (#3591)
Since #2741 is merged, we do not need this method. We can remove this
from IPlanBuilder.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Thu, 15 Nov 2018 04:17:01 +0000 (13:17 +0900)]
[neurun] DotDumper shows model outputs (#3590)
Just like model inputs, this commit shows model outputs as a different
shape.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 15 Nov 2018 04:14:16 +0000 (13:14 +0900)]
[tfltool] Update tflite schema API version (#3581)
Update tflite schema API version v1.9 -> v1.12
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 15 Nov 2018 04:03:19 +0000 (13:03 +0900)]
Support docker run as non-root user (#3592)
Run non-root anomymous user: same userid with host user
Used on arm cross build and tizen build
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Thu, 15 Nov 2018 02:41:50 +0000 (11:41 +0900)]
[neurun] Introduce Model class (#3586)
Model class is a container of actual operands/operations and model's
inputs/outputs.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
김용섭/동작제어Lab(SR)/Engineer/삼성전자 [Thu, 15 Nov 2018 02:09:24 +0000 (11:09 +0900)]
[model_parser] Supports printing specific tensors and operators (#3580)
Now model_parser.py supports argument options for features which prints
specific tensors and operators.
./model_parser.py ./tflite -t 1 2 3 # tensor id
./model_parser.py ./tflite -o 1 2 3 # operator id
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
Implementing NotEqual op in runtime (#3515)
This commit implements the CL version of NotEqual op.
Related issues: #3335, #3337
Signed-off-by: Siva Sai <siva.sai@samsung.com>
Removing TODO for mul (#3363)
This patch will remove TODO for mul op.
Signed-off-by: shubham <shub98.gupta@samsung.com>
Add Pad gtests to test padding across any dimention (#3308)
This patch will add gtests to test padding across any dimension.
This will verify padding op is working when in addition to spatial padding,
Padding is done across batch or depth too.
Signed-off-by: shubham <shub98.gupta@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Wed, 14 Nov 2018 09:02:43 +0000 (18:02 +0900)]
Add comment for argmax_ex (#3576)
Add comment for ANEURALNETWORKS_ARGMAX_EX in NeuralNetworksEx.h
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
김용섭/동작제어Lab(SR)/Engineer/삼성전자 [Wed, 14 Nov 2018 08:21:56 +0000 (17:21 +0900)]
[neurun] Remove NDEBUG in planTensors (#3579)
* [neurun] Remove NDEBUG in planTensors
Removes NDEBUG code in Linear::planTensors.
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
* Fix format-checker failed
Define TENSORFLOW_SUM in NeuralNetworksEx.h (#3578)
This patch defines TENSORFLOW_SUM in NeuralNetworksEx.h.
Related issue: #3347.
Signed-off-by: prasannar <prasanna.r@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Wed, 14 Nov 2018 07:22:05 +0000 (16:22 +0900)]
[neurun] Refine DAGChecker implementation (#3538)
Make DAGChecker implementation simpler by using operands' use info.
Cannot make a cyclic test for now, since after creating a cycle then
`finishBuilding` is called, it will cause assert failure by runtime.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
Add gtest for mul op to test 4D tensors (#3452)
This patch will add gtest to test mul op for 4D tensors
Signed-off-by: shubham <shub98.gupta@samsung.com>
윤현식/동작제어Lab(SR)/Principal Engineer/삼성전자 [Wed, 14 Nov 2018 04:34:31 +0000 (13:34 +0900)]
[model_parser] Print options of Conv2D, AvgPool2D, MaxPool2D (#3573)
Through this commit, model parser prints options of Conv2D, AvgPool2D, MaxPool2D such as stride, padding, etc.
Signed-off-by: Hyun Sik Yoon <hyunsik.yoon@samsung.com>
Add missing comments in NodeVisitor.h (#3575)
This patch adds the missing comments in NodeVisitor.h
Signed-off-by: prasannar <prasanna.r@samsung.com>
Implement SQRT op in PACL (#3418)
This patch implements SQRT op in PACL.
Related issue: #3341
Signed-off-by: prasannar <prasanna.r@samsung.com>
윤지영/동작제어Lab(SR)/Engineer/삼성전자 [Wed, 14 Nov 2018 00:46:35 +0000 (09:46 +0900)]
Add optimize_for_inference script into tflkit (#3546)
* Add optimize_for_inference script into tflkit
optimize_for_inference tool optimizes a TensorFlow GraphDef for inference.
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Add detail information to README.md
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
Add NNAPI delegation for SQRT_EX (#3422)
This patch adds NNAPI delegation for SQRT_EX
Signed-off-by: prasannar <prasanna.r@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Tue, 13 Nov 2018 09:34:26 +0000 (18:34 +0900)]
[neurun] Remove Planner implementation (#3560)
Now that we are able to have uniform planning procedure for all Nodes,
We do not need Planner class anymore. The planning procedure is now
unified and moved to `compilation.cc` but Planner class is left for
assertion checks which will be renamed as `OperationAsserter` later.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
Mismatch in generated test files for NOT_EQUAL_EX (#3550)
This patches the mismatch in the GTest files for NOT_EQUAL_EX.
Signed-off-by: prasannar <prasanna.r@samsung.com>
윤지영/동작제어Lab(SR)/Engineer/삼성전자 [Tue, 13 Nov 2018 01:51:24 +0000 (10:51 +0900)]
Add tflite_convert scripts into tflkit (#3544)
* Add tflite_convert scripts into tflkit
tflite_convert tool will convert a TensorFlow GraphDef into TfLite model.
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Add check routine for bazel
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Change the PB_PATH to GRAPHDEF_PATH and add detail usages
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Add detail usage to README.md
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Fixed wrong exitcode
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Tue, 13 Nov 2018 01:51:04 +0000 (10:51 +0900)]
[neurun] Add const iterate method to Linear (#3547)
Add const iterate method to Linear
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Mon, 12 Nov 2018 04:12:55 +0000 (13:12 +0900)]
[neurun] Revise IStageGenerator (#3536)
* [neurun] Revise IStageGenerator
Revise IStageGenerator to be a NodeVisitor. This will make it possible
to unify Planner implementation.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
* Pass Stage as const reference
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Mon, 12 Nov 2018 02:06:02 +0000 (11:06 +0900)]
Use pre-built acl by environment variable (#3533)
* Use pre-built acl by environment variable
When environment variable EXT_ACL_FOLDER is set, use pre-built acl binary
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
* Fix if-else syntax in Makefile
Adding testcases for NotEqual Op (#3516)
This commit adds Gtest testcases for NotEqual op.
Related issues: #3335,#3337
Signed-off-by: Siva Sai <siva.sai@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Fri, 9 Nov 2018 06:50:12 +0000 (15:50 +0900)]
Configure parallel execution job count (#3531)
User can configure parallel execution job count using NPROCS environment variable
- Reduce local compilation overhead when doing multiple job
- Protect excessive CI processor consumtion by one job
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
윤지영/동작제어Lab(SR)/Engineer/삼성전자 [Fri, 9 Nov 2018 05:20:39 +0000 (14:20 +0900)]
Create tflkit directory (#3438)
* Create tflite_easytool directory
This directory is created to make it easier to run frequently used tools in scripts.
Now, it has two functions:
- Summraize PB model
- Summraize TfLite model
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Change the directory name to tflkit
Also changes the relative path to absolute path
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Add Prerequires and TF tool guideline
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Add additional description for prerequires
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Fri, 9 Nov 2018 03:34:47 +0000 (12:34 +0900)]
[neurun] Default implementation for NodeVisitor (#3530)
Implement default implementation for NodeVisitor which is just doing
nothing.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Nov 2018 08:55:29 +0000 (17:55 +0900)]
Enable concat elimination on acl backend (#3528)
Enable concat elimination on acl backend
Change config and remove assertion block
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Thu, 8 Nov 2018 08:16:57 +0000 (17:16 +0900)]
[neurun] Remove IInitializerGenerator (#3525)
Remove IInitializerGenerator and its implementations. We do not need
this backend API anymore.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Nov 2018 08:16:17 +0000 (17:16 +0900)]
Add comment for tensorflow_max_ex (#3526)
Add comment for tensorflow_max_ex in NeuralNetworksEx.h
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Nov 2018 08:15:49 +0000 (17:15 +0900)]
Fix typo in nnapi delegate (#3527)
Fix typo in nnapi delegate: v1.11 -> v1.12
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
김용섭/동작제어Lab(SR)/Engineer/삼성전자 [Thu, 8 Nov 2018 07:48:44 +0000 (16:48 +0900)]
[neurun] Apply IMemoryPlanner to backend/cpu (#3517)
Now IMemoryPlanner and Allocator plays the role of IMemoryAllocator.
This commits includes some fixing for bugs to apply on backend/cpu.
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Nov 2018 07:12:55 +0000 (16:12 +0900)]
Update tensorflow to v1.12 (#3486)
* Update external package
- Update external package version
- Update cmake to build tflite
- Add gitignore for new external code
* Update library to support tflite
- Update nnapi delegate
- Update NeuralNetworksShim.h for new API
- Fix cmake for new files in updated tensorflow
- Workaround: Introduce OBS_BUILD flag to avoid gbs build fail
* Update neurun CPU kernel
- Update neurun CPU kernel to call updated tflite interpreter kernel
- Add build option -fPIC for updated tflite kernel
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Nov 2018 06:49:08 +0000 (15:49 +0900)]
Register subtensor info to tensor builder (#3519)
Register subtensor info to tensor builder if
- Tensor is subtensor of other tensor
- Backend support subtensor allocation
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Nov 2018 06:40:53 +0000 (15:40 +0900)]
Check external directory exists (#3522)
Check external directory exists and download if directory does not exist
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
이춘석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Nov 2018 05:12:20 +0000 (14:12 +0900)]
Fix run_benchmark.sh typo(MODLES->MODELS) (#3523)
- Fix run_benchmark.sh typo( MODLES -> MODELS )
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
이춘석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Nov 2018 04:08:02 +0000 (13:08 +0900)]
Fix CI benchmark failure (#3520)
CI benchmark job fails since it does not reflect model path
change. This commit sets the proper model path.
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Nov 2018 02:44:39 +0000 (11:44 +0900)]
Concat elimination in codegen (#3518)
Concat elimination in codegen: check output tensor is parent of inputs and do nothing if true
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Nov 2018 01:56:03 +0000 (10:56 +0900)]
Build subtensor in acl TensorBuilder (#3513)
* Build subtensor in acl TensorBuilder
Build subtensor in acl TensorBuilder
Change to public function in Convert.cc
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
* Remove deprecated declaration
* Remove diff by space line
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Thu, 8 Nov 2018 01:11:02 +0000 (10:11 +0900)]
[neurun] Remove InitializerGenerator dependency (#3508)
Remove all InitializerGenerator dependent code since we now have
ConstantInitializer for this purpose.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Thu, 8 Nov 2018 01:10:32 +0000 (10:10 +0900)]
Introduce subtensor allocation check API (#3510)
Check API will be used in codegen
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
김용섭/동작제어Lab(SR)/Engineer/삼성전자 [Thu, 8 Nov 2018 01:09:12 +0000 (10:09 +0900)]
[neurun] Introduce MemoryPlanner and Allocator (#3511)
Introduces IMemoryPlanner, BumpPlanner, FirstFitPlanner and
Allocator to replace IMemoryAllocator.
Current IMemoryAllocator has two roles, Planner and Allocator. This
commits separates IMemoryAllocator to the two classes. After this commit
lands, IMemoryAllocator will be removed soon.
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
장지섭/동작제어Lab(SR)/Engineer/삼성전자 [Wed, 7 Nov 2018 09:31:46 +0000 (18:31 +0900)]
Optimize the StridedSlice Kernel for performance improvement (#3230)
This commit optimizes the StridedSlice Kernel for performance inprovement.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
Introducing NotEqual Op in runtime (#3485)
This commit introduces NotEqual op in runtime.
Related issues: #3335, #3337
Signed-off-by: Siva Sai <siva.sai@samsung.com>
ArgMinMax Implementation for ArgMin and ArgMax. (#3331)
This change introduces the cpu kernel for the ArgMin and ArgMax
operations.
Signed-off-by: Praveen D N <praveen.dn@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Wed, 7 Nov 2018 08:01:26 +0000 (17:01 +0900)]
[neurun] Fix build break (#3507)
Caused by #3426 and #3436 collision.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Wed, 7 Nov 2018 07:40:33 +0000 (16:40 +0900)]
Update framework test list (#3499)
Add more framework test for neurun
Change inception test: use simple module instead of full models
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Wed, 7 Nov 2018 07:17:33 +0000 (16:17 +0900)]
Introduce subtensor map (#3493)
- Introduce subtensor map in ACL tensor builder
- TensorBuilder::at() and TensorBuilder::tensorAt() can return subtensor
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
이한종/동작제어Lab(SR)/Engineer/삼성전자 [Wed, 7 Nov 2018 07:02:25 +0000 (16:02 +0900)]
[neurun] Introduce ConstantInitializer (#3426)
* [neurun] Introduce ConstantInitializer
This commit introduces `ConstantIntializer` class that fills operand
objects with the constant values which were specified by
`ANeuralNetworksModel_setOperandValue`. Plus, this class replaces the
InitializerGenerator.
NOTE due to FullyConnected operation's implicit reshaping, it is
implemented with workarounds.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
* Add license
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Wed, 7 Nov 2018 06:34:34 +0000 (15:34 +0900)]
Fix custom operator register macro (#3495)
Fix custom operator register macro for updated TfLiteRegistration field
- First 4 fields may not change
- Can adjust new fields before custom_name
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
이춘석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Wed, 7 Nov 2018 06:34:20 +0000 (15:34 +0900)]
Enable more model tests (#3497)
The following tests are triggered on default on framework/verification
test :
- embedding lookup
- L2 normalization
- L2 pool
- Squeeze
- Tanh
- Transpose
- Logistic
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Wed, 7 Nov 2018 06:18:40 +0000 (15:18 +0900)]
Introduce namespace in runtime load helper (#3494)
Introduce namespace in runtime load helper
Restore function name
It protects naming conflict with upstream tensorflow
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
김용섭/동작제어Lab(SR)/Engineer/삼성전자 [Wed, 7 Nov 2018 03:12:34 +0000 (12:12 +0900)]
[model_parser] Hide instruction counts by verbose option (#3487)
Hides instruction counts by verbose option. Now -v or --verbose can skip
printing instruction counts. You can use this verbose option as follows
- With verbose level 0, prints only summary of tflite
- With verbose level 1(default), prints all like current default except for
instructions
- With verbose level 2, prints all like current default including for
instructions
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
김용섭/동작제어Lab(SR)/Engineer/삼성전자 [Tue, 6 Nov 2018 09:15:59 +0000 (18:15 +0900)]
[neurun] Introduce FirstfitAllocator (#3458)
* [neurun] Introduce FirstfitAllocator
Introduces FirstfitAllocator instead of BumpAllocator. This allocator
allocates a block for tensor at the first-fit offset.
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
* Revise allocate() of FirstfitAllocator
- Adds comment for description of allocate()
- Removes inner lambda allocateBlock() in allocate()
* Change the name from Firstfit to FirstFit
* Fix typo
* Adds comment for _alloc_table
장지섭/동작제어Lab(SR)/Engineer/삼성전자 [Tue, 6 Nov 2018 08:50:30 +0000 (17:50 +0900)]
Introduce a enum class declaration for reduce operations. (#3461)
This commit introduces a enum declaration for reduce operations.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
오형석/동작제어Lab(SR)/Staff Engineer/삼성전자 [Tue, 6 Nov 2018 02:48:23 +0000 (11:48 +0900)]
Introduce registerSubTensorInfo (#3474)
Introduce registerTensorInfo to register subtensor information (not used yet)
Backend allocator will use this info for allocation
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>