Revert "Revert "JPEG decoding arm64 optimization""

And fix clang build break on
.arch armv8-a+fp+simd

This reverts commit 0e087c4.

Change-Id: I1447da5403c55a0079e46f052b18005b9bc5a438

Android.mk

@@ -32,6 +32,10 @@ ifeq ($(TARGET_ARCH),x86)
   LOCAL_SRC_FILES += jidctintelsse.c
 endif
 
+LOCAL_SRC_FILES_arm64 += \
+        jsimd_arm64_neon.S \
+        jsimd_neon.c
+
 ifeq ($(strip $(TARGET_ARCH)),arm)
   ifeq ($(ARCH_ARM_HAVE_NEON),true)
     #use NEON accelerations

jdcolor.c

@@ -11,7 +11,7 @@
 #define JPEG_INTERNALS
 #include "jinclude.h"
 #include "jpeglib.h"
-#ifdef NV_ARM_NEON
+#if defined(NV_ARM_NEON) || defined(__aarch64__)
 #include "jsimd_neon.h"
 #endif
 
@@ -804,8 +804,17 @@ jinit_color_deconverter (j_decompress_ptr cinfo)
   case JCS_RGB:
     cinfo->out_color_components = RGB_PIXELSIZE;
     if (cinfo->jpeg_color_space == JCS_YCbCr) {
+#if defined(__aarch64__)
+      if (cap_neon_ycc_rgb()) {
+        cconvert->pub.color_convert = jsimd_ycc_rgb_convert;
+      } else {
+        cconvert->pub.color_convert = ycc_rgb_convert;
+        build_ycc_rgb_table(cinfo);
+      }
+#else
       cconvert->pub.color_convert = ycc_rgb_convert;
       build_ycc_rgb_table(cinfo);
+#endif
     } else if (cinfo->jpeg_color_space == JCS_GRAYSCALE) {
       cconvert->pub.color_convert = gray_rgb_convert;
     } else if (cinfo->jpeg_color_space == JCS_RGB && RGB_PIXELSIZE == 3) {
@@ -818,16 +827,17 @@ jinit_color_deconverter (j_decompress_ptr cinfo)
   case JCS_RGBA_8888:
     cinfo->out_color_components = 4;
     if (cinfo->jpeg_color_space == JCS_YCbCr) {
-#if defined(NV_ARM_NEON) && defined(__ARM_HAVE_NEON)
+#if (defined(NV_ARM_NEON) && defined(__ARM_HAVE_NEON)) || defined(__aarch64__)
       if (cap_neon_ycc_rgb()) {
         cconvert->pub.color_convert = jsimd_ycc_rgba8888_convert;
       } else {
         cconvert->pub.color_convert = ycc_rgba_8888_convert;
+        build_ycc_rgb_table(cinfo);
       }
 #else
       cconvert->pub.color_convert = ycc_rgba_8888_convert;
-#endif
       build_ycc_rgb_table(cinfo);
+#endif
     } else if (cinfo->jpeg_color_space == JCS_GRAYSCALE) {
       cconvert->pub.color_convert = gray_rgba_8888_convert;
     } else if (cinfo->jpeg_color_space == JCS_RGB) {
@@ -840,16 +850,17 @@ jinit_color_deconverter (j_decompress_ptr cinfo)
     cinfo->out_color_components = RGB_PIXELSIZE;
     if (cinfo->dither_mode == JDITHER_NONE) {
       if (cinfo->jpeg_color_space == JCS_YCbCr) {
-#if defined(NV_ARM_NEON) && defined(__ARM_HAVE_NEON)
+#if (defined(NV_ARM_NEON) && defined(__ARM_HAVE_NEON)) || defined(__aarch64__)
         if (cap_neon_ycc_rgb())  {
           cconvert->pub.color_convert = jsimd_ycc_rgb565_convert;
         } else {
           cconvert->pub.color_convert = ycc_rgb_565_convert;
+          build_ycc_rgb_table(cinfo);
         }
 #else
         cconvert->pub.color_convert = ycc_rgb_565_convert;
-#endif
         build_ycc_rgb_table(cinfo);
+#endif
       } else if (cinfo->jpeg_color_space == JCS_GRAYSCALE) {
         cconvert->pub.color_convert = gray_rgb_565_convert;
       } else if (cinfo->jpeg_color_space == JCS_RGB) {

jddctmgr.c

@@ -20,7 +20,7 @@
 #include "jpeglib.h"
 #include "jdct.h"		/* Private declarations for DCT subsystem */
 
-#ifdef NV_ARM_NEON
+#if defined(NV_ARM_NEON) || defined(__aarch64__)
 #include "jsimd_neon.h"
 #endif
 
@@ -137,7 +137,7 @@ start_pass (j_decompress_ptr cinfo)
       method = JDCT_ISLOW;	/* jidctred uses islow-style table */
       break;
     case 2:
-#if defined(NV_ARM_NEON) && defined(__ARM_HAVE_NEON)
+#if (defined(NV_ARM_NEON) && defined(__ARM_HAVE_NEON)) || defined(__aarch64__)
       if (cap_neon_idct_2x2()) {
         method_ptr = jsimd_idct_2x2;
       } else {
@@ -149,7 +149,7 @@ start_pass (j_decompress_ptr cinfo)
       method = JDCT_ISLOW;	/* jidctred uses islow-style table */
       break;
     case 4:
-#if defined(NV_ARM_NEON) && defined(__ARM_HAVE_NEON)
+#if (defined(NV_ARM_NEON) && defined(__ARM_HAVE_NEON)) || defined(__aarch64__)
 	  if (cap_neon_idct_4x4()) {
         method_ptr = jsimd_idct_4x4;
       } else {
@@ -186,13 +186,18 @@ start_pass (j_decompress_ptr cinfo)
 #else /* ANDROID_MIPS_IDCT */
 #ifdef DCT_ISLOW_SUPPORTED
       case JDCT_ISLOW:
+#if defined(__aarch64__)
+        if (cap_neon_idct_islow())
+          method_ptr = jsimd_idct_islow;
+        else
+#endif
 	method_ptr = jpeg_idct_islow;
 	method = JDCT_ISLOW;
 	break;
 #endif
 #ifdef DCT_IFAST_SUPPORTED
       case JDCT_IFAST:
-#if defined(NV_ARM_NEON) && defined(__ARM_HAVE_NEON)
+#if (defined(NV_ARM_NEON) && defined(__ARM_HAVE_NEON)) || defined(__aarch64__)
         if (cap_neon_idct_ifast()) {
           method_ptr = jsimd_idct_ifast;
         } else {

jmorecfg.h

@@ -373,7 +373,7 @@ typedef int boolean;
   #define MULTIPLIER short
 #elif ANDROID_MIPS_IDCT
   #define MULTIPLIER  short
-#elif NV_ARM_NEON
+#elif defined(NV_ARM_NEON) || defined(__aarch64__)
   #define MULTIPLIER short
 #else
   #define MULTIPLIER  int		/* type for fastest integer multiply */