Adding optimization for brief and android using NEON SIMD intrinsics

android/Android.mk.in

@@ -3,7 +3,11 @@ LOCAL_PATH := ${CMAKE_CURRENT_SOURCE_DIR}
 include $(CLEAR_VARS)
 
 LOCAL_MODULE := ${android_module_name}
- 
+
+ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
+LOCAL_ARM_NEON := true
+endif
+
 LOCAL_SRC_FILES :=  ${android_srcs}
 
 LOCAL_CFLAGS := ${android_defs}

android/android-jni/jni/Android.mk

@@ -16,7 +16,16 @@ LOCAL_C_INCLUDES +=  $(OPENCV_INCLUDES)
 LOCAL_MODULE    := android-opencv
 
 LOCAL_SRC_FILES := gen/android_cv_wrap.cpp image_pool.cpp \
-    yuv420sp2rgb.c gl_code.cpp Calibration.cpp
+    gl_code.cpp Calibration.cpp
+    
+
+#ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
+#    LOCAL_CFLAGS := -DHAVE_NEON=1
+#    LOCAL_SRC_FILES += yuv2rgb_neon.c.neon
+#else
+	LOCAL_SRC_FILES +=  yuv420sp2rgb.c
+#endif
+    
 
 include $(BUILD_SHARED_LIBRARY)
 

android/android-jni/jni/image_pool.cpp

@@ -18,17 +18,17 @@ JNIEXPORT jint JNI_OnLoad(JavaVM* vm, void* reserved)
   return JNI_VERSION_1_4;
 }
 
-JNIEXPORT void JNICALL Java_com_opencv_jni_opencvJNI_addYUVtoPool(JNIEnv * env,
-    jclass thiz, jlong ppool, jobject _jpool, jbyteArray jbuffer,
-    jint jidx, jint jwidth, jint jheight, jboolean jgrey)
+JNIEXPORT void JNICALL Java_com_opencv_jni_opencvJNI_addYUVtoPool(JNIEnv * env, jclass thiz, jlong ppool,
+                                                                  jobject _jpool, jbyteArray jbuffer, jint jidx,
+                                                                  jint jwidth, jint jheight, jboolean jgrey)
 {
-  int buff_height = jheight + (jheight/2);
-  Size buff_size(jwidth,buff_height);
-  image_pool *pool = (image_pool *) ppool;
+  int buff_height = jheight + (jheight / 2);
+  Size buff_size(jwidth, buff_height);
+  image_pool *pool = (image_pool *)ppool;
 
   Mat mat = pool->getYUV(jidx);
 
-  if (mat.empty() || mat.size() != buff_size )
+  if (mat.empty() || mat.size() != buff_size)
   {
     mat.create(buff_size, CV_8UC1);
   }
@@ -36,7 +36,7 @@ JNIEXPORT void JNICALL Java_com_opencv_jni_opencvJNI_addYUVtoPool(JNIEnv * env,
   jsize sz = env->GetArrayLength(jbuffer);
   uchar* buff = mat.ptr<uchar> (0);
 
-  env->GetByteArrayRegion(jbuffer, 0, sz, (jbyte*) buff);
+  env->GetByteArrayRegion(jbuffer, 0, sz, (jbyte*)buff);
 
   pool->addYUVMat(jidx, mat);
 
@@ -51,8 +51,7 @@ JNIEXPORT void JNICALL Java_com_opencv_jni_opencvJNI_addYUVtoPool(JNIEnv * env,
     }
     //doesn't work unfortunately..
     //TODO cvtColor(mat,color, CV_YCrCb2RGB);
-    color_convert_common(buff, buff + jwidth * jheight, jwidth, jheight,
-        color.ptr<uchar> (0), false);
+    color_convert_common(buff, buff + jwidth * jheight, jwidth, jheight, color.ptr<uchar> (0), false);
   }
 
   if (jgrey)
@@ -84,7 +83,7 @@ Mat image_pool::getGrey(int i)
   Mat tm = yuvImagesMap[i];
   if (tm.empty())
     return tm;
-  return tm(Range(0, tm.rows * (2.0f/3)), Range::all());
+  return tm(Range(0, tm.rows * (2.0f / 3)), Range::all());
 }
 Mat image_pool::getYUV(int i)
 {
@@ -99,3 +98,19 @@ void image_pool::addImage(int i, Mat mat)
   imagesmap[i] = mat;
 }
 
+void image_pool::convertYUVtoColor(int i, cv::Mat& out)
+{
+
+  Mat yuv = getYUV(i);
+
+  if (yuv.empty())
+    return;
+  int width = yuv.cols;
+  int height = yuv.rows * (2.0f / 3);
+  out.create(height, width, CV_8UC3);
+  const unsigned char* buff = yuv.ptr<unsigned char> (0);
+  unsigned char* out_buff = out.ptr<unsigned char> (0);
+  //doesn't work unfortunately..
+  //TODO cvtColor(mat,color, CV_YCrCb2RGB);
+  color_convert_common(buff, buff + width * height, width, height, out_buff, false);
+}

android/android-jni/jni/image_pool.h

@@ -53,6 +53,8 @@ public:
    */
   void addYUVMat(int i, cv::Mat mat);
 
+  void convertYUVtoColor(int i, cv::Mat& out);
+
   //	int addYUV(uchar* buffer, int size, int width, int height, bool grey,int idx);
   //
   //	void getBitmap(int * outintarray, int size, int idx);

android/android-jni/jni/yuv420sp2rgb.c

@@ -25,7 +25,7 @@
 #endif
 
 const int bytes_per_pixel = 2;
-void color_convert_common(unsigned char *pY, unsigned char *pUV, int width, int height, unsigned char *buffer, int grey)
+void color_convert_common(const unsigned char *pY, const unsigned char *pUV, int width, int height, unsigned char *buffer, int grey)
 {
 
   int i, j;

android/android-jni/jni/yuv420sp2rgb.h

@@ -7,7 +7,7 @@ extern "C" {
 #endif
 
 void color_convert_common(
-    unsigned char *pY, unsigned char *pUV,
+    const unsigned char *pY, const unsigned char *pUV,
     int width, int height, unsigned char *buffer,
     int grey);
 

android/android-jni/src/com/opencv/camera/CameraConfig.java

@@ -24,7 +24,7 @@ public class CameraConfig extends Activity {
 		// Restore preferences
 		SharedPreferences settings = ctx.getSharedPreferences(CAMERA_SETTINGS,
 				0);
-		int mode = settings.getInt(CAMERA_MODE, CAMERA_MODE_COLOR);
+		int mode = settings.getInt(CAMERA_MODE, CAMERA_MODE_BW);
 		return mode;
 	}
 

modules/features2d/src/brief.cpp

@@ -44,6 +44,11 @@
 #include <algorithm>
 #include <vector>
 
+#if ANDROID  && HAVE_NEON
+#include <cpu-features.h>
+#include <arm_neon.h>
+#endif
+
 using namespace cv;
 
 inline int smoothedSum(const Mat& sum, const KeyPoint& pt, int y, int x)
@@ -106,16 +111,39 @@ HammingLUT::ResultType HammingLUT::operator()( const unsigned char* a, const uns
 Hamming::ResultType Hamming::operator()(const unsigned char* a, const unsigned char* b, int size) const
 {
 #if __GNUC__
-    ResultType result = 0;
-    for (int i = 0; i < size; i += sizeof(unsigned long))
+  ResultType result = 0;
+#if ANDROID && HAVE_NEON
+  static uint64_t features = android_getCpuFeatures();
+  if ((features & ANDROID_CPU_ARM_FEATURE_NEON))
+  {
+    for (int i = 0; i < size; i += 16)
     {
-        unsigned long a2 = *reinterpret_cast<const unsigned long*> (a + i);
-        unsigned long b2 = *reinterpret_cast<const unsigned long*> (b + i);
-        result += __builtin_popcountl(a2 ^ b2);
+      uint8x16_t A_vec = vld1q_u8 (a + i);
+      uint8x16_t B_vec = vld1q_u8 (b + i);
+      //uint8x16_t veorq_u8 (uint8x16_t, uint8x16_t)
+      uint8x16_t AxorB = veorq_u8 (A_vec, B_vec);
+
+      uint8x16_t bitsSet += vcntq_u8 (AxorB);
+      //uint16x8_t vpadalq_u8 (uint16x8_t, uint8x16_t)
+      uint16x8_t bitSet8 = vpaddlq_u8 (bitsSet);
+      uint32x4_t bitSet4 = vpaddlq_u16 (bitSet8);
+
+      uint64x2_t bitSet2 = vpaddlq_u32 (bitSet4);
+      result += vgetq_lane_u64 (bitSet2,0);
+      result += vgetq_lane_u64 (bitSet2,1);
     }
-    return result;
+  }
+  else
+#endif
+  for (int i = 0; i < size; i += sizeof(unsigned long))
+  {
+    unsigned long a2 = *reinterpret_cast<const unsigned long*> (a + i);
+    unsigned long b2 = *reinterpret_cast<const unsigned long*> (b + i);
+    result += __builtin_popcountl(a2 ^ b2);
+  }
+  return result;
 #else
-    return HammingLUT()(a,b,size);
+  return HammingLUT()(a,b,size);
 #endif
 }