Started to optimize OpenCV for the ESP32

- Using float instead of double for floating point matrix multiplications (in core/src/matmul.simd.hpp) reduces a lot the computation time

CMakeLists.txt

@@ -202,6 +202,9 @@ endif()
 # ESP32 target option. Enable it with -DESP32=ON will disable/modify parts of the code for compilation to work
 OCV_OPTION(ESP32 "Compilation for esp32 target" OFF)
 
+# ESP32 optimization. Enable it with -DESP32_OPTIMIZATION=ON will use code optimized for esp32 target
+OCV_OPTION(ESP32_OPTIMIZATION "Optimization for the esp32 target" OFF)
+
 OCV_OPTION(OPENCV_ENABLE_NONFREE "Enable non-free algorithms" OFF)
 
 # 3rd party libs
@@ -497,6 +500,12 @@ if(ESP32)
   add_definitions(-DESP32)
 endif()
 
+if(ESP32_OPTIMIZATION)
+  message(STATUS "Enabled ESP32 target optimized code")
+  add_definitions(-DESP32_OPTIMIZATION)
+endif()
+
+
 if(ENABLE_IMPL_COLLECTION)
   add_definitions(-DCV_COLLECT_IMPL_DATA)
 endif()

esp32/doc/optimization.md

@@ -0,0 +1,57 @@
+# Optimization
+
+This doc details some optimizations done for OpenCV to run faster on the ESP32.
+
+
+
+Activating optimization
+----------------------------
+
+To activate the optimization for the ESP32, the CMake parameter `-DESP32_OPTIMIZATION=ON` must be enabled. Every optimization done will be disabled if this parameter is OFF.
+
+
+
+## Floating point support
+
+The ESP32 only have a single precision Floating Point Unit (no double precision). Therefore, OpenCV functions using double types are very slow.
+
+
+
+### Matrix  multiplications
+
+In files `core/matmul.dispatch.cp` and `core/matmul.simd.hpp`.
+
+Results by multiplying 100x6 * 6x100 matrices:
+
+1. Initial test : 60 ms
+2. Changing `alpha` and `beta` from double to float in `GEMMsingleMult()` function: 12ms
+3. Changing `alpha` and `beta` from double to float in `gemmImpl()` function: 4.6ms
+
+
+
+Results by multiplying 150x100 * 100x150 matrices: 
+
+1. Initial test: 2757ms
+2. Changing double in `GEMMStore()` function: 888ms
+
+
+
+
+
+## Esp-dsp library
+
+The ESP32 processor has the following hardware: 
+
+* 16/24-bit Instruction Set
+* Support for FPU (Floating Point Unit) 
+* Support for DSP instructions 
+  * 32-bit integer multiplier
+  * 32-bit integer divider 
+  * 40-bit MAC (Multiply-Accumulate)
+
+
+
+The *esp-dsp* library (https://github.com/espressif/esp-dsp) provides functions written in assembly to use this hardware.
+
+This part describes which functions are used where in OpenCV for better performances. 
+

esp32/scripts/build_opencv_for_esp32.sh

@@ -16,7 +16,7 @@ TOOLCHAIN_CMAKE_PATH=$HOME/esp/esp-idf/tools/cmake/toolchain-esp32.cmake
 LIB_INSTALL_PATH=$SCRIPTDIR/../lib
 
 # list of modules to compile
-OPENCV_MODULES_LIST=core,imgproc,imgcodecs
+OPENCV_MODULES_LIST=core,imgproc,imgcodecs,features2d,calib3d
 
 echo "################################################################################"
 echo "######################## build_opencv_for_esp32 script #########################"
@@ -40,7 +40,7 @@ else
 fi
 
 
-CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Release -DESP32=ON -DBUILD_SHARED_LIBS=OFF -DCV_DISABLE_OPTIMIZATION=OFF -DWITH_IPP=OFF -DWITH_TBB=OFF -DWITH_OPENMP=OFF -DWITH_PTHREADS_PF=OFF -DWITH_QUIRC=OFF -DWITH_1394=OFF -DWITH_CUDA=OFF -DWITH_OPENCL=OFF -DWITH_OPENCLAMDFFT=OFF -DWITH_OPENCLAMDBLAS=OFF -DWITH_VA_INTEL=OFF -DWITH_EIGEN=OFF -DWITH_GSTREAMER=OFF -DWITH_GTK=OFF -DWITH_JASPER=OFF -DWITH_JPEG=OFF -DWITH_WEBP=OFF -DBUILD_ZLIB=ON -DBUILD_PNG=ON -DWITH_TIFF=OFF -DWITH_V4L=OFF -DWITH_LAPACK=OFF -DWITH_ITT=OFF -DWITH_PROTOBUF=OFF -DWITH_IMGCODEC_HDR=OFF -DWITH_IMGCODEC_SUNRASTER=OFF -DWITH_IMGCODEC_PXM=OFF -DWITH_IMGCODEC_PFM=OFF -DBUILD_LIST=${OPENCV_MODULES_LIST} -DBUILD_JAVA=OFF -DBUILD_opencv_python=OFF -DBUILD_opencv_java=OFF -DBUILD_opencv_apps=OFF -DBUILD_PACKAGE=OFF -DBUILD_PERF_TESTS=OFF -DBUILD_TESTS=OFF -DCV_ENABLE_INTRINSICS=OFF -DCV_TRACE=OFF -DOPENCV_ENABLE_MEMALIGN=OFF -DCMAKE_TOOLCHAIN_FILE=${TOOLCHAIN_CMAKE_PATH}"
+CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Release -DESP32=ON -DESP32_OPTIMIZATION=ON -DBUILD_SHARED_LIBS=OFF -DCV_DISABLE_OPTIMIZATION=OFF -DWITH_IPP=OFF -DWITH_TBB=OFF -DWITH_OPENMP=OFF -DWITH_PTHREADS_PF=OFF -DWITH_QUIRC=OFF -DWITH_1394=OFF -DWITH_CUDA=OFF -DWITH_OPENCL=OFF -DWITH_OPENCLAMDFFT=OFF -DWITH_OPENCLAMDBLAS=OFF -DWITH_VA_INTEL=OFF -DWITH_EIGEN=OFF -DWITH_GSTREAMER=OFF -DWITH_GTK=OFF -DWITH_JASPER=OFF -DWITH_JPEG=OFF -DWITH_WEBP=OFF -DBUILD_ZLIB=ON -DBUILD_PNG=ON -DWITH_TIFF=OFF -DWITH_V4L=OFF -DWITH_LAPACK=OFF -DWITH_ITT=OFF -DWITH_PROTOBUF=OFF -DWITH_IMGCODEC_HDR=OFF -DWITH_IMGCODEC_SUNRASTER=OFF -DWITH_IMGCODEC_PXM=OFF -DWITH_IMGCODEC_PFM=OFF -DBUILD_LIST=${OPENCV_MODULES_LIST} -DBUILD_JAVA=OFF -DBUILD_opencv_python=OFF -DBUILD_opencv_java=OFF -DBUILD_opencv_apps=OFF -DBUILD_PACKAGE=OFF -DBUILD_PERF_TESTS=OFF -DBUILD_TESTS=OFF -DCV_ENABLE_INTRINSICS=OFF -DCV_TRACE=OFF -DOPENCV_ENABLE_MEMALIGN=OFF -DCMAKE_TOOLCHAIN_FILE=${TOOLCHAIN_CMAKE_PATH}"
 
 
 ### configure and build opencv ###
@@ -67,7 +67,7 @@ cp $SCRIPTDIR/resources/alloc_fix.cpp ./3rdparty/ade/ade-0.1.1f/sources/ade/sour
 echo "================================================================================"
 echo "Compiling with make -j"
 echo "================================================================================"
-make -j
+make -j3
 
 ### installing in output directory ###
 echo "================================================================================"