IVE Library

This is an Image Processing library using TPU on CVI1835.

Note:

1. bmtap2 docker environment is recomended.

2. Run ./clang-format.sh before pushing the code.

How to build

Requirements

1. Middleware headers
2. MLIR SDK
3. Tracer lib

Tracer lib can be get from GitLab at http://10.34.33.3:8480/sys_app/tracer. Put the Tracer lib under the 3rdparty folder.

$ mkdir build
$ cd build
$ CC=clang CXX=clang++ \
  cmake -G Ninja .. -DMLIR_SDK_ROOT=${PWD}/../../cvitek_mlir \
                    -DMIDDLEWARE_SDK_ROOT=${PWD}/../../middleware \
                    -DCMAKE_BUILD_TYPE=Release
$ ninja -j8

SOC mode

1. 64-bit

$ mkdir build_soc
$ cd build
$ cmake -G Ninja .. -DENABLE_SYSTRACE=ON \
                    -DCVI_TARGET=soc \
                    -DTOOLCHAIN_ROOT_DIR=${PWD}/../../gcc-linaro-6.3.1-2017.05-x86_64_aarch64-linux-gnu \
                    -DCMAKE_TOOLCHAIN_FILE=${PWD}/../toolchain/toolchain-aarch64-linux.cmake \
                    -DMLIR_SDK_ROOT=${PWD}/../../mlir/sdk/cvitek_tpu_sdk \
                    -DMIDDLEWARE_SDK_ROOT=${PWD}/../prebuilt/middleware \
                    -DCMAKE_BUILD_TYPE=Release
$ ninja -j8

2. 32-bit

$ mkdir build_soc
$ cd build
$ cmake -G Ninja .. -DCVI_TARGET=soc \
                    -DTOOLCHAIN_ROOT_DIR=${PWD}/../../gcc-linaro-6.3.1-2017.05-x86_64_arm-linux-gnueabihf \
                    -DCMAKE_TOOLCHAIN_FILE=${PWD}/../toolchain/toolchain-gnueabihf-linux.cmake \
                    -DMLIR_SDK_ROOT=${PWD}/../../cvitek_tpu_sdk \
                    -DMIDDLEWARE_SDK_ROOT=${PWD}/../../middleware \
                    -DCMAKE_BUILD_TYPE=Release
$ ninja -j8

Note: To make sure everytime you configure cmake correctly, you can set the compiler flags to empty with command.

cmake -G Ninja .. -DCMAKE_C_COMPILER="" \
                  -DCMAKE_CXX_COMPILER="" \
                  ......

You may install the library with the following command.

$ninja install

Currently supported TPU operations

1. Add
2. And
3. Block
4. Copy
5. Filter
6. HOG
7. Morphology- Binary dilate/ erode
8. Normalize gradient
9. Or
10. SAD
11. Sigmoid
12. Sobel X/ Y Gradient
13. Sub
14. a - b
15. abs(a - b)
16. Threshold
17. Binary threshold w/o high low value
18. Slope
19. Xor

Currently supported NEON operations

1. U16 related image conversion.
2. U16, S16 to U8, S8 threshold.

BMKernel known issues

1. Currently TPU API does not support TL(BF16) to TG(FP32).
2. Some required API is missing,
   1. div

Workaround table

Chip	Flags
cv1835	WORKAROUND_SCALAR_4096_ALIGN_BUG