diff --git a/arm/arm_init.c b/arm/arm_init.c
index 15f5162..952b802 100644
--- a/arm/arm_init.c
+++ b/arm/arm_init.c
@@ -1,9 +1,9 @@
 
 /* arm_init.c - NEON optimised filter functions
  *
- * Copyright (c) 2013 Glenn Randers-Pehrson
+ * Copyright (c) 2014 Glenn Randers-Pehrson
  * Written by Mans Rullgard, 2011.
- * Last changed in libpng 1.5.17 [July 18, 2013]
+ * Last changed in libpng 1.6.10 [March 6, 2014]
  *
  * This code is released under the libpng license.
  * For conditions of distribution and use, see the disclaimer
@@ -19,132 +19,33 @@
 #ifdef PNG_READ_SUPPORTED
 #if PNG_ARM_NEON_OPT > 0
 #ifdef PNG_ARM_NEON_CHECK_SUPPORTED /* Do run-time checks */
-#include <signal.h> /* for sig_atomic_t */
-
-#ifdef __ANDROID__
-/* Linux provides access to information about CPU capabilites via
- * /proc/self/auxv, however Android blocks this while still claiming to be
- * Linux.  The Andoid NDK, however, provides appropriate support.
+/* WARNING: it is strongly recommended that you do not build libpng with
+ * run-time checks for CPU features if at all possible.  In the case of the ARM
+ * NEON instructions there is no processor-specific way of detecting the
+ * presense of the required support, therefore run-time detectioon is extremely
+ * OS specific.
  *
- * Documentation: http://www.kandroid.org/ndk/docs/CPU-ARM-NEON.html
+ * You may set the macro PNG_ARM_NEON_FILE to the file name of file containing
+ * a fragment of C source code which defines the png_have_neon function.  There
+ * are a number of implementations in contrib/arm-neon, but the only one that
+ * has partial support is contrib/arm-neon/linux.c - a generic Linux
+ * implementation which reads /proc/cpufino.
  */
-#include <cpu-features.h>
+#ifndef PNG_ARM_NEON_FILE
+#  ifdef __linux__
+#     define PNG_ARM_NEON_FILE "contrib/arm-neon/linux.c"
+#  endif
+#endif
 
-static int
-png_have_neon(png_structp png_ptr)
-{
-   /* This is a whole lot easier than the mess below, however it is probably
-    * implemented as below, therefore it is better to cache the result (these
-    * function calls may be slow!)
-    */
-   PNG_UNUSED(png_ptr)
-   return android_getCpuFamily() == ANDROID_CPU_FAMILY_ARM &&
-      (android_getCpuFeatures() & ANDROID_CPU_ARM_FEATURE_NEON) != 0;
-}
-#elif defined(__linux__)
-/* The generic __linux__ implementation requires reading /proc/self/auxv and
- * looking at each element for one that records NEON capabilities.
- */
-#include <unistd.h> /* for POSIX 1003.1 */
-#include <errno.h>  /* for EINTR */
+#ifdef PNG_ARM_NEON_FILE
 
-#include <sys/types.h>
-#include <sys/stat.h>
-#include <fcntl.h>
-#include <elf.h>
-#include <asm/hwcap.h>
+#include <signal.h> /* for sig_atomic_t */
+static int png_have_neon(png_structp png_ptr);
+#include PNG_ARM_NEON_FILE
 
-/* A read call may be interrupted, in which case it returns -1 and sets errno to
- * EINTR if nothing was done, otherwise (if something was done) a partial read
- * may result.
- */
-static size_t
-safe_read(png_structp png_ptr, int fd, void *buffer_in, size_t nbytes)
-{
-   size_t ntotal = 0;
-   char *buffer = png_voidcast(char*, buffer_in);
-
-   while (nbytes > 0)
-   {
-      unsigned int nread;
-      int iread;
-
-      /* Passing nread > INT_MAX to read is implementation defined in POSIX
-       * 1003.1, therefore despite the unsigned argument portable code must
-       * limit the value to INT_MAX!
-       */
-      if (nbytes > INT_MAX)
-         nread = INT_MAX;
-
-      else
-         nread = (unsigned int)/*SAFE*/nbytes;
-
-      iread = read(fd, buffer, nread);
-
-      if (iread == -1)
-      {
-         /* This is the devil in the details, a read can terminate early with 0
-          * bytes read because of EINTR, yet it still returns -1 otherwise end
-          * of file cannot be distinguished.
-          */
-         if (errno != EINTR)
-         {
-            png_warning(png_ptr, "/proc read failed");
-            return 0; /* I.e. a permanent failure */
-         }
-      }
-
-      else if (iread < 0)
-      {
-         /* Not a valid 'read' result: */
-         png_warning(png_ptr, "OS /proc read bug");
-         return 0;
-      }
-
-      else if (iread > 0)
-      {
-         /* Continue reading until a permanent failure, or EOF */
-         buffer += iread;
-         nbytes -= (unsigned int)/*SAFE*/iread;
-         ntotal += (unsigned int)/*SAFE*/iread;
-      }
-
-      else
-         return ntotal;
-   }
-
-   return ntotal; /* nbytes == 0 */
-}
-
-static int
-png_have_neon(png_structp png_ptr)
-{
-   int fd = open("/proc/self/auxv", O_RDONLY);
-   Elf32_auxv_t aux;
-
-   /* Failsafe: failure to open means no NEON */
-   if (fd == -1)
-   {
-      png_warning(png_ptr, "/proc/self/auxv open failed");
-      return 0;
-   }
-
-   while (safe_read(png_ptr, fd, &aux, sizeof aux) == sizeof aux)
-   {
-      if (aux.a_type == AT_HWCAP && (aux.a_un.a_val & HWCAP_NEON) != 0)
-      {
-         close(fd);
-         return 1;
-      }
-   }
-
-   close(fd);
-   return 0;
-}
-#else
-   /* We don't know how to do a run-time check on this system */
-#  error "no support for run-time ARM NEON checks"
-#endif /* OS checks */
+#else  /* PNG_ARM_NEON_FILE */
+#  error "PNG_ARM_NEON_FILE undefined: no support for run-time ARM NEON checks"
+#endif /* PNG_ARM_NEON_FILE */
 #endif /* PNG_ARM_NEON_CHECK_SUPPORTED */
 
 #ifndef PNG_ALIGNED_MEMORY_SUPPORTED
@@ -154,6 +55,16 @@
 void
 png_init_filter_functions_neon(png_structp pp, unsigned int bpp)
 {
+   /* The switch statement is compiled in for ARM_NEON_API, the call to
+    * png_have_neon is compiled in for ARM_NEON_CHECK.  If both are defined
+    * the check is only performed if the API has not set the NEON option on
+    * or off explicitly.  In this case the check controls what happens.
+    *
+    * If the CHECK is not compiled in and the option is UNSET the behavior prior
+    * to 1.6.7 was to use the NEON code - this was a bug caused by having the
+    * wrong order of the 'ON' and 'default' cases.  UNSET now defaults to OFF,
+    * as documented in png.h
+    */
 #ifdef PNG_ARM_NEON_API_SUPPORTED
    switch ((pp->options >> PNG_ARM_NEON) & 3)
    {
@@ -178,13 +89,14 @@
          break;
 #endif
 #endif /* PNG_ARM_NEON_CHECK_SUPPORTED */
+
 #ifdef PNG_ARM_NEON_API_SUPPORTED
+      default: /* OFF or INVALID */
+         return;
+
       case PNG_OPTION_ON:
          /* Option turned on */
          break;
-
-      default: /* OFF or INVALID */
-         return;
    }
 #endif
 
diff --git a/arm/filter_neon.S b/arm/filter_neon.S
index b8aef10..40d9777 100644
--- a/arm/filter_neon.S
+++ b/arm/filter_neon.S
@@ -3,7 +3,7 @@
  *
  * Copyright (c) 2013 Glenn Randers-Pehrson
  * Written by Mans Rullgard, 2011.
- * Last changed in libpng 1.5.17 [July 18, 2013]
+ * Last changed in libpng 1.6.8 [December 19, 2013]
  *
  * This code is released under the libpng license.
  * For conditions of distribution and use, see the disclaimer
@@ -11,7 +11,8 @@
  */
 
 /* This is required to get the symbol renames, which are #defines, and also
- * includes the definition (or not) of PNG_ARM_NEON_OPT.
+ * includes the definition (or not) of PNG_ARM_NEON_OPT and
+ * PNG_ARM_NEON_IMPLEMENTATION.
  */
 #define PNG_VERSION_INFO_ONLY
 #include "../pngpriv.h"
@@ -20,6 +21,13 @@
 .section .note.GNU-stack,"",%progbits /* mark stack as non-executable */
 #endif
 
+/* Assembler NEON support - only works for 32-bit ARM (i.e. it does not work for
+ * ARM64).  The code in arm/filter_neon_intrinsics.c supports ARM64, however it
+ * only works if -mfpu=neon is specified on the GCC command line.  See pngpriv.h
+ * for the logic which sets PNG_USE_ARM_NEON_ASM:
+ */
+#if PNG_ARM_NEON_IMPLEMENTATION == 2 /* hand-coded assembler */
+
 #ifdef PNG_READ_SUPPORTED
 #if PNG_ARM_NEON_OPT > 0
 
@@ -235,3 +243,4 @@
 endfunc
 #endif /* PNG_ARM_NEON_OPT > 0 */
 #endif /* PNG_READ_SUPPORTED */
+#endif /* PNG_ARM_NEON_IMPLEMENTATION == 2 (assembler) */
diff --git a/arm/filter_neon_intrinsics.c b/arm/filter_neon_intrinsics.c
new file mode 100644
index 0000000..e7a09f4
--- /dev/null
+++ b/arm/filter_neon_intrinsics.c
@@ -0,0 +1,372 @@
+
+/* filter_neon_intrinsics.c - NEON optimised filter functions
+ *
+ * Copyright (c) 2013 Glenn Randers-Pehrson
+ * Written by James Yu <james.yu at linaro.org>, October 2013.
+ * Based on filter_neon.S, written by Mans Rullgard, 2011.
+ *
+ * Last changed in libpng 1.6.8 [December 19, 2013]
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+#include "../pngpriv.h"
+
+/* This code requires -mfpu=neon on the command line: */
+#if PNG_ARM_NEON_IMPLEMENTATION == 1 /* intrinsics code from pngpriv.h */
+
+#include <arm_neon.h>
+
+/* libpng row pointers are not necessarily aligned to any particular boundary,
+ * however this code will only work with appropriate alignment.  arm/arm_init.c
+ * checks for this (and will not compile unless it is done). This code uses
+ * variants of png_aligncast to avoid compiler warnings.
+ */
+#define png_ptr(type,pointer) png_aligncast(type *,pointer)
+#define png_ptrc(type,pointer) png_aligncastconst(const type *,pointer)
+
+/* The following relies on a variable 'temp_pointer' being declared with type
+ * 'type'.  This is written this way just to hide the GCC strict aliasing
+ * warning; note that the code is safe because there never is an alias between
+ * the input and output pointers.
+ */
+#define png_ldr(type,pointer)\
+   (temp_pointer = png_ptr(type,pointer), *temp_pointer)
+
+#ifdef PNG_READ_SUPPORTED
+#if PNG_ARM_NEON_OPT > 0
+
+void
+png_read_filter_row_up_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+   png_const_bytep pp = prev_row;
+
+   for (; rp < rp_stop; rp += 16, pp += 16)
+   {
+      uint8x16_t qrp, qpp;
+
+      qrp = vld1q_u8(rp);
+      qpp = vld1q_u8(pp);
+      qrp = vaddq_u8(qrp, qpp);
+      vst1q_u8(rp, qrp);
+   }
+}
+
+void
+png_read_filter_row_sub3_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+
+   uint8x16_t vtmp = vld1q_u8(rp);
+   uint8x8x2_t *vrpt = png_ptr(uint8x8x2_t, &vtmp);
+   uint8x8x2_t vrp = *vrpt;
+
+   uint8x8x4_t vdest;
+   vdest.val[3] = vdup_n_u8(0);
+
+   for (; rp < rp_stop;)
+   {
+      uint8x8_t vtmp1, vtmp2;
+      uint32x2_t *temp_pointer;
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
+      vdest.val[0] = vadd_u8(vdest.val[3], vrp.val[0]);
+      vtmp2 = vext_u8(vrp.val[0], vrp.val[1], 6);
+      vdest.val[1] = vadd_u8(vdest.val[0], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
+      vdest.val[2] = vadd_u8(vdest.val[1], vtmp2);
+      vdest.val[3] = vadd_u8(vdest.val[2], vtmp1);
+
+      vtmp = vld1q_u8(rp + 12);
+      vrpt = png_ptr(uint8x8x2_t, &vtmp);
+      vrp = *vrpt;
+
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
+      rp += 3;
+   }
+
+   PNG_UNUSED(prev_row)
+}
+
+void
+png_read_filter_row_sub4_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+
+   uint8x8x4_t vdest;
+   vdest.val[3] = vdup_n_u8(0);
+
+   for (; rp < rp_stop; rp += 16)
+   {
+      uint32x2x4_t vtmp = vld4_u32(png_ptr(uint32_t,rp));
+      uint8x8x4_t *vrpt = png_ptr(uint8x8x4_t,&vtmp);
+      uint8x8x4_t vrp = *vrpt;
+      uint32x2x4_t *temp_pointer;
+
+      vdest.val[0] = vadd_u8(vdest.val[3], vrp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[0], vrp.val[1]);
+      vdest.val[2] = vadd_u8(vdest.val[1], vrp.val[2]);
+      vdest.val[3] = vadd_u8(vdest.val[2], vrp.val[3]);
+      vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
+   }
+
+   PNG_UNUSED(prev_row)
+}
+
+void
+png_read_filter_row_avg3_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+
+   uint8x16_t vtmp;
+   uint8x8x2_t *vrpt;
+   uint8x8x2_t vrp;
+   uint8x8x4_t vdest;
+   vdest.val[3] = vdup_n_u8(0);
+
+   vtmp = vld1q_u8(rp);
+   vrpt = png_ptr(uint8x8x2_t,&vtmp);
+   vrp = *vrpt;
+
+   for (; rp < rp_stop; pp += 12)
+   {
+      uint8x8_t vtmp1, vtmp2, vtmp3;
+
+      uint8x8x2_t *vppt;
+      uint8x8x2_t vpp;
+
+      uint32x2_t *temp_pointer;
+
+      vtmp = vld1q_u8(pp);
+      vppt = png_ptr(uint8x8x2_t,&vtmp);
+      vpp = *vppt;
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
+      vdest.val[0] = vhadd_u8(vdest.val[3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 3);
+      vtmp3 = vext_u8(vrp.val[0], vrp.val[1], 6);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vtmp2);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 6);
+      vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
+
+      vtmp = vld1q_u8(rp + 12);
+      vrpt = png_ptr(uint8x8x2_t,&vtmp);
+      vrp = *vrpt;
+
+      vdest.val[2] = vhadd_u8(vdest.val[1], vtmp2);
+      vdest.val[2] = vadd_u8(vdest.val[2], vtmp3);
+
+      vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
+
+      vdest.val[3] = vhadd_u8(vdest.val[2], vtmp2);
+      vdest.val[3] = vadd_u8(vdest.val[3], vtmp1);
+
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
+      rp += 3;
+   }
+}
+
+void
+png_read_filter_row_avg4_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+   png_const_bytep pp = prev_row;
+
+   uint8x8x4_t vdest;
+   vdest.val[3] = vdup_n_u8(0);
+
+   for (; rp < rp_stop; rp += 16, pp += 16)
+   {
+      uint32x2x4_t vtmp;
+      uint8x8x4_t *vrpt, *vppt;
+      uint8x8x4_t vrp, vpp;
+      uint32x2x4_t *temp_pointer;
+
+      vtmp = vld4_u32(png_ptr(uint32_t,rp));
+      vrpt = png_ptr(uint8x8x4_t,&vtmp);
+      vrp = *vrpt;
+      vtmp = vld4_u32(png_ptrc(uint32_t,pp));
+      vppt = png_ptr(uint8x8x4_t,&vtmp);
+      vpp = *vppt;
+
+      vdest.val[0] = vhadd_u8(vdest.val[3], vpp.val[0]);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+      vdest.val[1] = vhadd_u8(vdest.val[0], vpp.val[1]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]);
+      vdest.val[2] = vhadd_u8(vdest.val[1], vpp.val[2]);
+      vdest.val[2] = vadd_u8(vdest.val[2], vrp.val[2]);
+      vdest.val[3] = vhadd_u8(vdest.val[2], vpp.val[3]);
+      vdest.val[3] = vadd_u8(vdest.val[3], vrp.val[3]);
+
+      vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
+   }
+}
+
+static uint8x8_t
+paeth(uint8x8_t a, uint8x8_t b, uint8x8_t c)
+{
+   uint8x8_t d, e;
+   uint16x8_t p1, pa, pb, pc;
+
+   p1 = vaddl_u8(a, b); /* a + b */
+   pc = vaddl_u8(c, c); /* c * 2 */
+   pa = vabdl_u8(b, c); /* pa */
+   pb = vabdl_u8(a, c); /* pb */
+   pc = vabdq_u16(p1, pc); /* pc */
+
+   p1 = vcleq_u16(pa, pb); /* pa <= pb */
+   pa = vcleq_u16(pa, pc); /* pa <= pc */
+   pb = vcleq_u16(pb, pc); /* pb <= pc */
+
+   p1 = vandq_u16(p1, pa); /* pa <= pb && pa <= pc */
+
+   d = vmovn_u16(pb);
+   e = vmovn_u16(p1);
+
+   d = vbsl_u8(d, b, c);
+   e = vbsl_u8(e, a, d);
+
+   return e;
+}
+
+void
+png_read_filter_row_paeth3_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_const_bytep pp = prev_row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+
+   uint8x16_t vtmp;
+   uint8x8x2_t *vrpt;
+   uint8x8x2_t vrp;
+   uint8x8_t vlast = vdup_n_u8(0);
+   uint8x8x4_t vdest;
+   vdest.val[3] = vdup_n_u8(0);
+
+   vtmp = vld1q_u8(rp);
+   vrpt = png_ptr(uint8x8x2_t,&vtmp);
+   vrp = *vrpt;
+
+   for (; rp < rp_stop; pp += 12)
+   {
+      uint8x8x2_t *vppt;
+      uint8x8x2_t vpp;
+      uint8x8_t vtmp1, vtmp2, vtmp3;
+      uint32x2_t *temp_pointer;
+
+      vtmp = vld1q_u8(pp);
+      vppt = png_ptr(uint8x8x2_t,&vtmp);
+      vpp = *vppt;
+
+      vdest.val[0] = paeth(vdest.val[3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 3);
+      vtmp2 = vext_u8(vpp.val[0], vpp.val[1], 3);
+      vdest.val[1] = paeth(vdest.val[0], vtmp2, vpp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[0], vrp.val[1], 6);
+      vtmp3 = vext_u8(vpp.val[0], vpp.val[1], 6);
+      vdest.val[2] = paeth(vdest.val[1], vtmp3, vtmp2);
+      vdest.val[2] = vadd_u8(vdest.val[2], vtmp1);
+
+      vtmp1 = vext_u8(vrp.val[1], vrp.val[1], 1);
+      vtmp2 = vext_u8(vpp.val[1], vpp.val[1], 1);
+
+      vtmp = vld1q_u8(rp + 12);
+      vrpt = png_ptr(uint8x8x2_t,&vtmp);
+      vrp = *vrpt;
+
+      vdest.val[3] = paeth(vdest.val[2], vtmp2, vtmp3);
+      vdest.val[3] = vadd_u8(vdest.val[3], vtmp1);
+
+      vlast = vtmp2;
+
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[0]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[1]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[2]), 0);
+      rp += 3;
+      vst1_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2_t,&vdest.val[3]), 0);
+      rp += 3;
+   }
+}
+
+void
+png_read_filter_row_paeth4_neon(png_row_infop row_info, png_bytep row,
+   png_const_bytep prev_row)
+{
+   png_bytep rp = row;
+   png_bytep rp_stop = row + row_info->rowbytes;
+   png_const_bytep pp = prev_row;
+
+   uint8x8_t vlast = vdup_n_u8(0);
+   uint8x8x4_t vdest;
+   vdest.val[3] = vdup_n_u8(0);
+
+   for (; rp < rp_stop; rp += 16, pp += 16)
+   {
+      uint32x2x4_t vtmp;
+      uint8x8x4_t *vrpt, *vppt;
+      uint8x8x4_t vrp, vpp;
+      uint32x2x4_t *temp_pointer;
+
+      vtmp = vld4_u32(png_ptr(uint32_t,rp));
+      vrpt = png_ptr(uint8x8x4_t,&vtmp);
+      vrp = *vrpt;
+      vtmp = vld4_u32(png_ptrc(uint32_t,pp));
+      vppt = png_ptr(uint8x8x4_t,&vtmp);
+      vpp = *vppt;
+
+      vdest.val[0] = paeth(vdest.val[3], vpp.val[0], vlast);
+      vdest.val[0] = vadd_u8(vdest.val[0], vrp.val[0]);
+      vdest.val[1] = paeth(vdest.val[0], vpp.val[1], vpp.val[0]);
+      vdest.val[1] = vadd_u8(vdest.val[1], vrp.val[1]);
+      vdest.val[2] = paeth(vdest.val[1], vpp.val[2], vpp.val[1]);
+      vdest.val[2] = vadd_u8(vdest.val[2], vrp.val[2]);
+      vdest.val[3] = paeth(vdest.val[2], vpp.val[3], vpp.val[2]);
+      vdest.val[3] = vadd_u8(vdest.val[3], vrp.val[3]);
+
+      vlast = vpp.val[3];
+
+      vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
+   }
+}
+
+#endif /* PNG_ARM_NEON_OPT > 0 */
+#endif /* PNG_READ_SUPPORTED */
+#endif /* PNG_ARM_NEON_IMPLEMENTATION == 1 (intrinsics) */