| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
- Introduce operand ready checker
- Introduce operation stack to save prepared operation
- Manage operand stack and operation stack and show log
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Collect operand usage info in frontend
Clear when model building is finished
Prepare removing nnapi-dependent usage info in neurun model operand
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
* Change default runtime to neurun
Change default runtime: pacl -> neurun
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
* Enable loggin runtime default
|
|
- Remove unused isModelInput() method
- Remove usage() method. Use hasData() and model's input/output
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Chagne ARGMAX definition same with current NNAPI
- Allow one axis reduce
- Output rank = input rank - 1
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Apply code for merging node into subgraph to Graph instead of 1 subrgaph
1 node code
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
|
|
Prepare operand stack in interpreter to save prepared operands to use
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Apply SubgraphContext(1 subgraph 1 node) to Graph, Linear, Compiler and
{Dataflow|Linear}Executor
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
|
|
Interpreter ITensor Inherit backend ITensor for more compatibility with other backends
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
This is for reducing namespace operation.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
|
|
* Fix a couple of noticeable clang static analyzer warnings
Fixed passing vector by value, lack of default switch case, dead code
Signed-off-by: Alexey Grebenkin <a.grebenkin@samsung.com>
* Correcting for PR comments
* Correcting for more PR comments
|
|
It fixes build-break caused by Subgraph's use of model::operation::Index,
which is renamed to model::OperationIndex.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
|
|
* Introduce interpreter class
Introduce interpreter class: run on assigned ExecEnv
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
* Fix define for header
|
|
This class would be used as a context class which has Subgraphs instead
of SubgraphSet in Graph
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
|
|
This is for reducing namespace operation.
Part of #4985
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
|
|
Remove axis conversion in libs/ARMComputeEx because conversion should done before configuration
Fix axis conversion in pack/unpack acl-cl extend kernel
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Change the name SubgraphSet. Actually SubgraphSet has elements in order
so that SubgraphSequence is proper name.
Signed-off-by: Yongseop Kim <yons.kim@samsung.com>
|
|
Instead of using interpreter's own structure, use tensor info structure in model
Move comment and copy constructor
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Remove install static kernel library because they are linked at each backend
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Move backend cmake variable setting into each backend
Remove target check in backend
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Add unittest for specifying unspecified dimension with setInput and setOutput.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
|
|
Remove header path and linking dependency with tflite in cpu backend build cmake
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
setOutput (#4957)
IExecutor, ExecutorBase, ExecManager will have setInput and setOutput that don't require
`type` and `shape` parameters. In this case, `type` and `shape` from model will be used.
ExecManager test is also modified to use newly introduced setInput and setOutput.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
|
|
Remove unused util function in cpu backend: type conversion to tflite internal type
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Introduce cpu quantized int8 softmax kernel from tflite and gemmlowp
Use kernel in neurun cpu backend
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Fix copyright using introduced script
Add copyright in libs/ARMComputeEx/src/runtime/misc/functions/Utils.cpp, runtimes/neurun/backend/acl_neon/operand/INETensor.cc
Fix libs/ARMComputeEx/arm_compute/runtime/misc/functions/Utils.h year based on file created date
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
I removed `virtual` when I am sure it is our code.
Otherwise, I did not removed `virtual`.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
|
|
Remove deprecated header include in neurun cpu fully connected kernel
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
- Merge duplicated calls into one
model->operands.at(operand_activation).usage(Usage::CONSTANT);
model->operands.at(operand_activation).usage(operand::Usage::CONSTANT);
- Group the related statements by operand and operator
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
|
|
This commit eliminate the concept of scalar type.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
|
|
This commit fixes wrong setting of block_size and padding size for SpaceToBatch in NHWC.
- Change setting of them to only support 4-dimensional input and the 2-dimensional spatial shape.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
|
|
This commit fixes a type.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
|
|
Fix wrong added output operand. (operand_rhs to operand_result)
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
|
|
`scale` and `offset` are used only in case of `TENSOR_QUANT8_ASYMM`.
For most cases, we don't need the values at all.
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
|
|
* Fix some errors from Clang compiler
* Fix abs on unsigned char, fix hidden overloaded virtual function, rename predeclared class on struct, remove unused private variable, remove std::move when he called in return statement
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
* Fix unused variables, struct/class mismatching
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
* Fix extern C++, unused variables, explicit overriding, remove std::move in return statement
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
* Fix file for format-check and extern C
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
* Fix library information removing from neurun
Signed-off-by: Pavel Iliutchenko <p.iliutchenk@samsung.com>
|
|
* Append nnapi tests of GREATER_EQUAL and LESS
This commit appends nnapi tests of of GREATER_EQUAL and LESS.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
* [neurun] Support GREATER_EQUAL and LESS
This commit supports GREATER_EQUAL and LESS operations.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
|
|
This patch enables initilization of neurun::operand::Shape with list.
Before:
```
operand::Shape shape{4};
shape.dim(0) = 1;
shape.dim(1) = 2;
shape.dim(2) = 2;
shape.dim(3) = 1;
```
After:
```
operand::Shape shape{1, 2, 2, 1};
```
Signed-off-by: Sanggyu Lee <sg5.lee@samsung.com>
|
|
Remove unused header in cpu backend ReshapeLayer to remove dependency with tensorflow lite
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Introduce cpu quantized int8 fully connected kernel from tflite
Use kernel in neurun cpu backend
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
* Introduce float softrmax cpu kernel
Introduce float softmax cpu kernel from tflite
Use kernel in neurun cpu backend
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
* Fix release build bug
|
|
Introduce cpu quantized int8 convolution kernel from tflite and gemmlowp
Use kernel in neurun cpu backend
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Introduce float fully connected kernel from tflite
Use kernel in neurun cpu backend
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Introduce cpu float depthwise convolution kernel from tflite
Implement neurun cpu backend for depthwise convolution
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
CPU backend assume that neurun's internal shape is always same or less than 4.
It is true on current neurun implementation, but it may change.
So add assertion to check this.
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
This commit adds validation of ComparisonNode.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
|
|
C API was 3 functions that creates each components of a Backend. This
commit revises it to have only one API function `neurun_backend_create`
which packs all the components.
Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
|
|
Introduce cpu maxpool kernel from tflite
Use kernel in neurun cpu backend
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
Introduce float convolution kernel from tflite
Use kernel in neurun cpu backend
Signed-off-by: Hyeongseok Oh <hseok82.oh@samsung.com>
|
|
This commit enable operations using boolean type.
Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
|
|
(#4354)" (#4840)
This reverts commit 34f49de36ae41f10735fd570e006981671e13a68.
The same changes were merged in acl v19.02.
Related PR : https://review.mlplatform.org/#/c/ml/ComputeLibrary/+/511/
Signed-off-by: Jiyoung Yun <jy910.yun@samsung.com>
|
