frameworks/base: vectorize Bitmap operations

Enable gcc vectorizing optimizations for Bitmap operations within
libandroid_runtime.so.  Adds -ftree-vectorize -mvectorize-with-neon-quad
-fprefetch-loop-arrays to functions and compiles Bitmap.cpp in ARM mode
(necessary to enable vectorizing optimizations) for appropriate targets.

Change-Id: I12defef6f47adf41af849df6e750097f5906d92d

2 files changed, 30 insertions, 11 deletions

diff --git a/core/jni/Android.mk b/core/jni/Android.mk
index de1fb4e..8edf042 100644
--- a/core/jni/Android.mk
+++ b/core/jni/Android.mk
@@ -88,7 +88,6 @@ LOCAL_SRC_FILES:= \
 	android_util_XmlBlock.cpp \
 	android_util_PackageRedirectionMap.cpp \
 	android/graphics/AutoDecodeCancel.cpp \
-	android/graphics/Bitmap.cpp \
 	android/graphics/BitmapFactory.cpp \
 	android/graphics/Camera.cpp \
 	android/graphics/Canvas.cpp \
@@ -226,6 +225,20 @@ LOCAL_SHARED_LIBRARIES += libselinux
 LOCAL_CFLAGS += -DHAVE_SELINUX
 endif # HAVE_SELINUX
 
+ifeq ($(TARGET_ARCH), arm)
+  ifeq ($(TARGET_USE_KRAIT_BIONIC_OPTIMIZATION), true)
+    TARGET_arm_CFLAGS += -DUSE_NEON_BITMAP_OPTS -mvectorize-with-neon-quad
+    LOCAL_SRC_FILES+= \
+		android/graphics/Bitmap.cpp.arm
+  else
+    LOCAL_SRC_FILES+= \
+		android/graphics/Bitmap.cpp
+  endif
+else
+    LOCAL_SRC_FILES+= \
+		android/graphics/Bitmap.cpp
+endif
+
 ifeq ($(USE_OPENGL_RENDERER),true)
 	LOCAL_SHARED_LIBRARIES += libhwui
 endif
diff --git a/core/jni/android/graphics/Bitmap.cpp b/core/jni/android/graphics/Bitmap.cpp
index 63683b4..7c5da26 100644
--- a/core/jni/android/graphics/Bitmap.cpp
+++ b/core/jni/android/graphics/Bitmap.cpp
@@ -22,6 +22,12 @@
     #define TRACE_BITMAP(code)

 #endif

 

+#ifdef USE_NEON_BITMAP_OPTS

+    #define __BITMAP_OPTS __attribute__((optimize("-ftree-vectorize", "-fprefetch-loop-arrays")))

+#else

+    #define __BITMAP_OPTS

+#endif

+

 ///////////////////////////////////////////////////////////////////////////////

 // Conversions to/from SkColor, for get/setPixels, and the create method, which

 // is basically like setPixels

@@ -29,7 +35,7 @@
 typedef void (*FromColorProc)(void* dst, const SkColor src[], int width,

                               int x, int y);

 

-static void FromColor_D32(void* dst, const SkColor src[], int width,

+static void __BITMAP_OPTS FromColor_D32(void* dst, const SkColor src[], int width,

                           int, int) {

     SkPMColor* d = (SkPMColor*)dst;

 

@@ -38,7 +44,7 @@ static void FromColor_D32(void* dst, const SkColor src[], int width,
     }

 }

 

-static void FromColor_D565(void* dst, const SkColor src[], int width,

+static void __BITMAP_OPTS FromColor_D565(void* dst, const SkColor src[], int width,

                            int x, int y) {

     uint16_t* d = (uint16_t*)dst;

 

@@ -50,7 +56,7 @@ static void FromColor_D565(void* dst, const SkColor src[], int width,
     }

 }

 

-static void FromColor_D4444(void* dst, const SkColor src[], int width,

+static void __BITMAP_OPTS FromColor_D4444(void* dst, const SkColor src[], int width,

                             int x, int y) {

     SkPMColor16* d = (SkPMColor16*)dst;

 

@@ -113,7 +119,7 @@ bool GraphicsJNI::SetPixels(JNIEnv* env, jintArray srcColors,
 typedef void (*ToColorProc)(SkColor dst[], const void* src, int width,

                             SkColorTable*);

 

-static void ToColor_S32_Alpha(SkColor dst[], const void* src, int width,

+static void __BITMAP_OPTS ToColor_S32_Alpha(SkColor dst[], const void* src, int width,

                               SkColorTable*) {

     SkASSERT(width > 0);

     const SkPMColor* s = (const SkPMColor*)src;

@@ -122,7 +128,7 @@ static void ToColor_S32_Alpha(SkColor dst[], const void* src, int width,
     } while (--width != 0);

 }

 

-static void ToColor_S32_Opaque(SkColor dst[], const void* src, int width,

+static void __BITMAP_OPTS ToColor_S32_Opaque(SkColor dst[], const void* src, int width,

                                SkColorTable*) {

     SkASSERT(width > 0);

     const SkPMColor* s = (const SkPMColor*)src;

@@ -133,7 +139,7 @@ static void ToColor_S32_Opaque(SkColor dst[], const void* src, int width,
     } while (--width != 0);

 }

 

-static void ToColor_S4444_Alpha(SkColor dst[], const void* src, int width,

+static void __BITMAP_OPTS ToColor_S4444_Alpha(SkColor dst[], const void* src, int width,

                                 SkColorTable*) {

     SkASSERT(width > 0);

     const SkPMColor16* s = (const SkPMColor16*)src;

@@ -142,7 +148,7 @@ static void ToColor_S4444_Alpha(SkColor dst[], const void* src, int width,
     } while (--width != 0);

 }

 

-static void ToColor_S4444_Opaque(SkColor dst[], const void* src, int width,

+static void __BITMAP_OPTS ToColor_S4444_Opaque(SkColor dst[], const void* src, int width,

                                  SkColorTable*) {

     SkASSERT(width > 0);

     const SkPMColor16* s = (const SkPMColor16*)src;

@@ -153,7 +159,7 @@ static void ToColor_S4444_Opaque(SkColor dst[], const void* src, int width,
     } while (--width != 0);

 }

 

-static void ToColor_S565(SkColor dst[], const void* src, int width,

+static void __BITMAP_OPTS ToColor_S565(SkColor dst[], const void* src, int width,

                          SkColorTable*) {

     SkASSERT(width > 0);

     const uint16_t* s = (const uint16_t*)src;

@@ -164,7 +170,7 @@ static void ToColor_S565(SkColor dst[], const void* src, int width,
     } while (--width != 0);

 }

 

-static void ToColor_SI8_Alpha(SkColor dst[], const void* src, int width,

+static void __BITMAP_OPTS ToColor_SI8_Alpha(SkColor dst[], const void* src, int width,

                               SkColorTable* ctable) {

     SkASSERT(width > 0);

     const uint8_t* s = (const uint8_t*)src;

@@ -175,7 +181,7 @@ static void ToColor_SI8_Alpha(SkColor dst[], const void* src, int width,
     ctable->unlockColors(false);

 }

 

-static void ToColor_SI8_Opaque(SkColor dst[], const void* src, int width,

+static void __BITMAP_OPTS ToColor_SI8_Opaque(SkColor dst[], const void* src, int width,

                                SkColorTable* ctable) {

     SkASSERT(width > 0);

     const uint8_t* s = (const uint8_t*)src;