Age | Commit message (Collapse) | Author | Files | Lines |
|
* [neurun] DataflowExecutor indentation fix
Remove the extra block in DataflowExecutor implementation.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
* Remove another block
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
|
|
* [neurun] Add CMakeLists.txt for NEON
Add CMakeLists.txt for NEON backend
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* [neurun] Replicate the acl_cl backend code
- This is just a replication of acl_cl backend code
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* [neurun] Add acl_neon sub dir
Add acl_neon sub dir for NEON backend
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* [neurun] Fix build failure
[neurun] Fix build failure in NEON
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* Add acl_neon backend into Config file
Add acl_neon backend into Config file
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* Use NEReshapeLayer for runinng mobilenet model
With this patch, neurun can run mobilenet
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* Fix build error after rebase
Fix build error after rebase
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* Fix copyright year and macros
Fix copyright year and macros
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* Remove unsupported op from StageGenerator.cc
Remove unsupported op from StageGenerator.cc
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* Use NEFullyConnectedLayer instead of GenericFCLayer
Use NEFullyConnectedLayer instead of GenericFCLayer
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* [neurun] Fix build error
[neurun] Fix build error #2
Signed-off-by: Chunseok Lee <chunseok.lee@samsung.com>
* [neurun] Reuse cl version of convert and Swizzle
Also fixed INETensor class: it didn't implemented one abstract member of ITensor
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
* [neurun] Rebase onto master
[neurun] Rebase changes onto master
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
* [neurun] Phase 2: Reuse other backends codes where needed
Reuse other backends codes where needed
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
* [neurun] Fix linking errors of reused codes
Fix linking errors of reused codes
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
* Separate acl common files into single dir
Pull all files, that are used in both acl_cl and acl_neon into separate
dir
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
* Cleanup virtual function declarations in neurun backends
Cleanup virtual function declarations in neurun backends.
This is borrowed from PR 4791
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
* Rename make_acl_function as asAclFunction
Fix coding style of make_acl_function
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
* Move cpu::Object into neurun_core::backend
Move cpu::Object into neurun_core since it is used in NEON also
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
* Rebase onto master: conflict because of TensorInfo
Rebase onto master: conflict because of TensorInfo
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
* Fix copyright year of acl_common/TemplTensorBuilder.h
Fix copyright year of acl_common/TemplTensorBuilder.h
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
* Merge onto master: fixStageGenerator
Merge onto master: fixStageGenerator
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
|
|
* [pureacl] Use fused_act parameter in Conv and DepthwiseConv acl layer
acl v19.03 support fused_act parameter in Conv and DepthwiseConv layer.
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
* Revert DepthwiseConvolution changes
Activation layer information only works for QASYMM8 in DepthwiseConvolutionLayer.
* Add comments for configure of ConvolutionLayer
|
|
Allocate tensor for constant operand
Assign constant tensor to exec-env
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Introduce average pool kernel from tflite
Use kernel in neurun cpu backend
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Introduce concat kernel from tflite
Use concat kernel in cpu backend
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Before this was the only Executor so its name was "Executor", now that
we have more Executors so rename it to prevent from confusion.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
|
|
Define ExecEnv to gather interpreter environment - status, memory, etc
Assign input and output tensor into ExecEnv
Fix unittest
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
This commit Enable setting the data type of made operand::Object.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
|
|
Fix build error by operationFactory parameter change
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Move backend unittest into each backend
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Related : #3708
This commit enables `LocalResponseNormalization`, `DepthToSpace`, `ReduceMin` ops for `acl_cl`.
Signed-off-by: sjsujinkim <sjsujin.kim@samsung.com>
|
|
Reuse setting inputs and outputs part with template method pattern.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
|
|
This commit adds passing a param of operand::Set when creating the Node in order to adjust operand information for each operation.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
|
|
Due to not defining DefBackend for model inputs, there was a problem
when running with multiple backends.
Fix #4828
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
|
|
As all the executors have OperandContext in common, we can promote it to
its base class `ExecutorBase`.
Part of #4823
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
|
|
Implement NNAPI to set input and output using external memory
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
To build neurun cpu backend, check tensorflow lite internal source code instead of gbs build flag
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
(#4838)
This commit moves out of the graph of neurun to collect the model's infomation.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
|
|
To build neurun acl_cl backend, check acl support instead of target architecture name
Fix arm architecture flag bug
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
* Check invalid NNAPI input and output index
Check too large index by return invalid index value in wrppaer function
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
* Fix validation check log
|
|
* [neurun] Enable ArgMax, Dequantize, Mean ops
This commit enables `ArgMax`, `Dequantize`, `Mean` ops for `acl_cl`.
* fix build break
Signed-off-by: sjsujinkim <sjsujin.kim@samsung.com>
|
|
Check correctness of length for input and output tensor setting based on NNAPI spec
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Extract OperandContext from Plan so OperandContext can be merged in
ExecutorBase. (Both Executor and DataflowExecutor contains
OperandContext so we can promote it to ExecutorBase, later.)
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
|
|
* Move util for frontend only
Move util implementation for NNAPI frontend only
* Define static function for convert
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Change compiler::TensorInfo to model::operand::Info
All data fields in TensorInfo is model operand's info
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Redo TensorBuilderSet as unordered_set since we don't care if it is
sorted or not
Signed-off-by: Poshshoev Dilshodzhon <d.poshshoev@samsung.com>
|
|
Introduce tensor interface for interpreter
Introduce ROTensor (read-only tensor) and Tensor (writable tensor) class
Generate interpreter tensor when input/output is set
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
* [neurun] Abstract compilation into functions
Abstract compilation procedure as a function for each Executors.
They are introduced as static, private and has a parameter and a return
value. It is intended to explicitly show which is input and output for
compilation.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
* Rename functions prefix - compile to create
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
|
|
operand::Set and operation::Set have hardcoded uint32_t and assertion,
which is based on wrong assumption that Index is always uint32_t.
This patch replaces uint32_t with Index.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
|
|
Cleanup virtual function declarations in neurun cpu and acl_cl backends
Signed-off-by: Andrei Shedko <a.shedko@samsung.com>
|
|
Looks like it is a remnant of times when neurun was being built as a single library (now it is divided into several parts, such as core, front end and back end).
Signed-off-by: Sergei Barannikov <s.barannikov@samsung.com>
|
|
* Enable NNAPI compilation validation test
Enable ValidationTestCompilation.Finish test
Define compile state STARTED
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
* Rename NONE to CREATED
|
|
Remve simple operation in ARMComputeEx except pack/unpack
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
- Set input and output tensor info for interpreter
- Fix TensorInfo bug
- Remove sink/source implementation
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Remove deprecated NENormalizationLayer and NENormalizationLayerExKernel implementation
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
This commit enables `Gather`, `Neg`, `Abs` ops for `acl_cl`.
Signed-off-by: sjsujinkim <sjsujin.kim@samsung.com>
|
|
Enable ValidationTestIdentify.DuplicateInputs and ValidationTestIdentify.DuplicateOutputs for acl_cl kernel
Enable all ValidationTestIdentify test for cpu kernel
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Introduce parameter checker. As a first commit,
- Check add node only
- Check activation paramter is constant only
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Rename NNAPI Ex operation for reduce max
TENSORFLOW_MAX_EX -> REDUCE_MAX_EX
Update NeuralNetworksEx.h comment
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
This commit removes const keyword from converting functions in NNAPIConvert.
In general, const is used on a return value of RVO because const affect move semantics.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
|
|
Remove redundant `virtual` declarations and use `override` instead( as per Core Guidelines C.128 )
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
|
|
In https://github.sec.samsung.net/STAR/nnfw/pull/4738#discussion_r156180, we've discussed to group all operations in `::arm_compute::ComparisonOperation`.
This commit introduces `ComparisonNode` that includes operations in `::::arm_compute::ComparisonOperation` and intergrate `Equal` and `NotEqual` to `ComparisonNode`
`ComparisonNode` can be support below.
```
enum class ComparisonType
{
Equal,
NotEqual,
Greater,
GreaterEqual,
Less,
LessEqual
};
```
Signed-off-by: sjsujinkim <sjsujin.kim@samsung.com>
|
|
* [neurun] Enable Equal, SquaredDifference, TopKV2 ops
This commit enables `Equal`, `SquaredDifference`, `TopKV2` ops for `acl_cl`.
Signed-off-by: sjsujinkim <sjsujin.kim@samsung.com>
* Fix build break
|
|
Remove `target_include_directories` for NEURUN_INCLUDE_DIR.
However variable NEURUN_INCLUDE_DIR cannot be removed for now since
cpu/acl_cl kernel module improperly depend on this dir.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
|
|
* Fix missing symbol in logging backend
Add missing ANeuralNetworksCompilation_free in logging backend
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
* Format fix
Signed-off-by: Vladimir Plazun <v.plazun@samsung.com>
|
|
Extract CMakeLists.txt file for neurun/core.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
|
|
Save backend dlopen handle in map member variable
It uses to avoid warning by handle lost without dlclose(), but it will be used to other purpose later
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
* [neurun] Enable LogicalOr, LogicalNot ops
Related : #3708
This commit enables `LogicalOr`, `LogicalNot` ops for `acl_cl`.
Signed-off-by: sjsujinkim <sjsujin.kim@samsung.com>
* Change LogicalOr kernel to CLBitwiseOr
* Fix build break
|
|
Introduce TensorInfo for interpreter
Current implementation is copy of compiler::TensorInfo
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|