Add files needed when building libpng for the arm64 architecture.

These files can be compiled-in for other architectures but produce no binary code.

5 files changed, 692 insertions, 1 deletions

diff --git a/png/CMakeLists.txt b/png/CMakeLists.txt
index 6a94c6c28..b9988695b 100644
--- a/png/CMakeLists.txt
+++ b/png/CMakeLists.txt
@@ -31,6 +31,9 @@ set(PNG_SRCS
   pngwrite.c
   pngwtran.c
   pngwutil.c
+  arm/arm_init.c
+  arm/filter_neon_intrinsics.c
+  arm/palette_neon_intrinsics.c
 )
 
 #######################################################################
diff --git a/png/Makefile b/png/Makefile
index 80b134950..f444ebd36 100644
--- a/png/Makefile
+++ b/png/Makefile
@@ -23,7 +23,8 @@ include ../makeinclude
 
 OBJS	=	png.o pngset.o pngget.o pngrutil.o pngtrans.o pngwutil.o \
 		pngread.o pngrio.o pngwio.o pngwrite.o pngrtran.o \
-		pngwtran.o pngmem.o pngerror.o pngpread.o
+		pngwtran.o pngmem.o pngerror.o pngpread.o \
+		arm/arm_init.o arm/filter_neon_intrinsics.o arm/palette_neon_intrinsics.o
 
 LIBPNG	=	../lib/libfltk_png$(LIBEXT)
 
diff --git a/png/arm/arm_init.c b/png/arm/arm_init.c
new file mode 100644
index 000000000..a34ecdbef
--- /dev/null
+++ b/png/arm/arm_init.c
@@ -0,0 +1,136 @@
+
+/* arm_init.c - NEON optimised filter functions
+ *
+ * Copyright (c) 2018 Cosmin Truta
+ * Copyright (c) 2014,2016 Glenn Randers-Pehrson
+ * Written by Mans Rullgard, 2011.
+ *
+ * This code is released under the libpng license.
+ * For conditions of distribution and use, see the disclaimer
+ * and license in png.h
+ */
+
+/* Below, after checking __linux__, various non-C90 POSIX 1003.1 functions are
+ * called.
+ */
+#define _POSIX_SOURCE 1
+
+#include "../pngpriv.h"
+
+#ifdef PNG_READ_SUPPORTED
+
+#if PNG_ARM_NEON_OPT > 0
+#ifdef PNG_ARM_NEON_CHECK_SUPPORTED /* Do run-time checks */
+/* 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
+ * presence of the required support, therefore run-time detection is extremely
+ * OS specific.
+ *
+ * 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.
+ */
+#ifndef PNG_ARM_NEON_FILE
+#  ifdef __linux__
+#     define PNG_ARM_NEON_FILE "contrib/arm-neon/linux.c"
+#  endif
+#endif
+
+#ifdef PNG_ARM_NEON_FILE
+
+#include <signal.h> /* for sig_atomic_t */
+static int png_have_neon(png_structp png_ptr);
+#include PNG_ARM_NEON_FILE
+
+#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
+#  error "ALIGNED_MEMORY is required; set: -DPNG_ALIGNED_MEMORY_SUPPORTED"
+#endif
+
+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
+    */
+   png_debug(1, "in png_init_filter_functions_neon");
+#ifdef PNG_ARM_NEON_API_SUPPORTED
+   switch ((pp->options >> PNG_ARM_NEON) & 3)
+   {
+      case PNG_OPTION_UNSET:
+         /* Allow the run-time check to execute if it has been enabled -
+          * thus both API and CHECK can be turned on.  If it isn't supported
+          * this case will fall through to the 'default' below, which just
+          * returns.
+          */
+#endif /* PNG_ARM_NEON_API_SUPPORTED */
+#ifdef PNG_ARM_NEON_CHECK_SUPPORTED
+         {
+            static volatile sig_atomic_t no_neon = -1; /* not checked */
+
+            if (no_neon < 0)
+               no_neon = !png_have_neon(pp);
+
+            if (no_neon)
+               return;
+         }
+#ifdef PNG_ARM_NEON_API_SUPPORTED
+         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;
+   }
+#endif
+
+   /* IMPORTANT: any new external functions used here must be declared using
+    * PNG_INTERNAL_FUNCTION in ../pngpriv.h.  This is required so that the
+    * 'prefix' option to configure works:
+    *
+    *    ./configure --with-libpng-prefix=foobar_
+    *
+    * Verify you have got this right by running the above command, doing a build
+    * and examining pngprefix.h; it must contain a #define for every external
+    * function you add.  (Notice that this happens automatically for the
+    * initialization function.)
+    */
+   pp->read_filter[PNG_FILTER_VALUE_UP-1] = png_read_filter_row_up_neon;
+
+   if (bpp == 3)
+   {
+      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub3_neon;
+      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg3_neon;
+      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
+         png_read_filter_row_paeth3_neon;
+   }
+
+   else if (bpp == 4)
+   {
+      pp->read_filter[PNG_FILTER_VALUE_SUB-1] = png_read_filter_row_sub4_neon;
+      pp->read_filter[PNG_FILTER_VALUE_AVG-1] = png_read_filter_row_avg4_neon;
+      pp->read_filter[PNG_FILTER_VALUE_PAETH-1] =
+          png_read_filter_row_paeth4_neon;
+   }
+}
+#endif /* PNG_ARM_NEON_OPT > 0 */
+#endif /* READ */
diff --git a/png/arm/filter_neon_intrinsics.c b/png/arm/filter_neon_intrinsics.c
new file mode 100644
index 000000000..553c0be21
--- /dev/null
+++ b/png/arm/filter_neon_intrinsics.c
@@ -0,0 +1,402 @@
+
+/* filter_neon_intrinsics.c - NEON optimised filter functions
+ *
+ * Copyright (c) 2018 Cosmin Truta
+ * Copyright (c) 2014,2016 Glenn Randers-Pehrson
+ * Written by James Yu <james.yu at linaro.org>, October 2013.
+ * Based on filter_neon.S, written by Mans Rullgard, 2011.
+ *
+ * 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"
+
+#ifdef PNG_READ_SUPPORTED
+
+/* This code requires -mfpu=neon on the command line: */
+#if PNG_ARM_NEON_IMPLEMENTATION == 1 /* intrinsics code from pngpriv.h */
+
+#if defined(_MSC_VER) && defined(_M_ARM64)
+#  include <arm64_neon.h>
+#else
+#  include <arm_neon.h>
+#endif
+
+/* 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.
+ *
+ * When compiling with MSVC ARM64, the png_ldr macro can't be passed directly
+ * to vst4_lane_u32, because of an internal compiler error inside MSVC.
+ * To avoid this compiler bug, we use a temporary variable (vdest_val) to store
+ * the result of png_ldr.
+ */
+#define png_ldr(type,pointer)\
+   (temp_pointer = png_ptr(type,pointer), *temp_pointer)
+
+#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;
+
+   png_debug(1, "in png_read_filter_row_up_neon");
+
+   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);
+
+   png_debug(1, "in png_read_filter_row_sub3_neon");
+
+   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);
+
+   png_debug(1, "in png_read_filter_row_sub4_neon");
+
+   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;
+      uint32x2x4_t vdest_val;
+
+      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]);
+
+      vdest_val = png_ldr(uint32x2x4_t, &vdest);
+      vst4_lane_u32(png_ptr(uint32_t,rp), vdest_val, 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;
+
+   png_debug(1, "in png_read_filter_row_avg3_neon");
+
+   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);
+
+   png_debug(1, "in png_read_filter_row_avg4_neon");
+
+   for (; rp < rp_stop; rp += 16, pp += 16)
+   {
+      uint32x2x4_t vtmp;
+      uint8x8x4_t *vrpt, *vppt;
+      uint8x8x4_t vrp, vpp;
+      uint32x2x4_t *temp_pointer;
+      uint32x2x4_t vdest_val;
+
+      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]);
+
+      vdest_val = png_ldr(uint32x2x4_t, &vdest);
+      vst4_lane_u32(png_ptr(uint32_t,rp), vdest_val, 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;
+
+   png_debug(1, "in png_read_filter_row_paeth3_neon");
+
+   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);
+
+   png_debug(1, "in png_read_filter_row_paeth4_neon");
+
+   for (; rp < rp_stop; rp += 16, pp += 16)
+   {
+      uint32x2x4_t vtmp;
+      uint8x8x4_t *vrpt, *vppt;
+      uint8x8x4_t vrp, vpp;
+      uint32x2x4_t *temp_pointer;
+      uint32x2x4_t vdest_val;
+
+      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];
+
+      vdest_val = png_ldr(uint32x2x4_t, &vdest);
+      vst4_lane_u32(png_ptr(uint32_t,rp), vdest_val, 0);
+   }
+}
+
+#endif /* PNG_ARM_NEON_OPT > 0 */
+#endif /* PNG_ARM_NEON_IMPLEMENTATION == 1 (intrinsics) */
+#endif /* READ */
diff --git a/png/arm/palette_neon_intrinsics.c b/png/arm/palette_neon_intrinsics.c
new file mode 100644
index 000000000..b4d1fd2ab
--- /dev/null
+++ b/png/arm/palette_neon_intrinsics.c
@@ -0,0 +1,149 @@
+
+/* palette_neon_intrinsics.c - NEON optimised palette expansion functions
+ *
+ * Copyright (c) 2018-2019 Cosmin Truta
+ * Copyright (c) 2017-2018 Arm Holdings. All rights reserved.
+ * Written by Richard Townsend <Richard.Townsend@arm.com>, February 2017.
+ *
+ * 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"
+
+#if PNG_ARM_NEON_IMPLEMENTATION == 1
+
+#if defined(_MSC_VER) && defined(_M_ARM64)
+#  include <arm64_neon.h>
+#else
+#  include <arm_neon.h>
+#endif
+
+/* Build an RGBA8 palette from the separate RGB and alpha palettes. */
+void
+png_riffle_palette_neon(png_structrp png_ptr)
+{
+   png_const_colorp palette = png_ptr->palette;
+   png_bytep riffled_palette = png_ptr->riffled_palette;
+   png_const_bytep trans_alpha = png_ptr->trans_alpha;
+   int num_trans = png_ptr->num_trans;
+   int i;
+
+   png_debug(1, "in png_riffle_palette_neon");
+
+   /* Initially black, opaque. */
+   uint8x16x4_t w = {{
+      vdupq_n_u8(0x00),
+      vdupq_n_u8(0x00),
+      vdupq_n_u8(0x00),
+      vdupq_n_u8(0xff),
+   }};
+
+   /* First, riffle the RGB colours into an RGBA8 palette.
+    * The alpha component is set to opaque for now.
+    */
+   for (i = 0; i < 256; i += 16)
+   {
+      uint8x16x3_t v = vld3q_u8((png_const_bytep)(palette + i));
+      w.val[0] = v.val[0];
+      w.val[1] = v.val[1];
+      w.val[2] = v.val[2];
+      vst4q_u8(riffled_palette + (i << 2), w);
+   }
+
+   /* Fix up the missing transparency values. */
+   for (i = 0; i < num_trans; i++)
+      riffled_palette[(i << 2) + 3] = trans_alpha[i];
+}
+
+/* Expands a palettized row into RGBA8. */
+int
+png_do_expand_palette_rgba8_neon(png_structrp png_ptr, png_row_infop row_info,
+    png_const_bytep row, png_bytepp ssp, png_bytepp ddp)
+{
+   png_uint_32 row_width = row_info->width;
+   const png_uint_32 *riffled_palette =
+      (const png_uint_32 *)png_ptr->riffled_palette;
+   const png_int_32 pixels_per_chunk = 4;
+   int i;
+
+   png_debug(1, "in png_do_expand_palette_rgba8_neon");
+
+   if (row_width < pixels_per_chunk)
+      return 0;
+
+   /* This function originally gets the last byte of the output row.
+    * The NEON part writes forward from a given position, so we have
+    * to seek this back by 4 pixels x 4 bytes.
+    */
+   *ddp = *ddp - ((pixels_per_chunk * sizeof(png_uint_32)) - 1);
+
+   for (i = 0; i < row_width; i += pixels_per_chunk)
+   {
+      uint32x4_t cur;
+      png_bytep sp = *ssp - i, dp = *ddp - (i << 2);
+      cur = vld1q_dup_u32 (riffled_palette + *(sp - 3));
+      cur = vld1q_lane_u32(riffled_palette + *(sp - 2), cur, 1);
+      cur = vld1q_lane_u32(riffled_palette + *(sp - 1), cur, 2);
+      cur = vld1q_lane_u32(riffled_palette + *(sp - 0), cur, 3);
+      vst1q_u32((void *)dp, cur);
+   }
+   if (i != row_width)
+   {
+      /* Remove the amount that wasn't processed. */
+      i -= pixels_per_chunk;
+   }
+
+   /* Decrement output pointers. */
+   *ssp = *ssp - i;
+   *ddp = *ddp - (i << 2);
+   return i;
+}
+
+/* Expands a palettized row into RGB8. */
+int
+png_do_expand_palette_rgb8_neon(png_structrp png_ptr, png_row_infop row_info,
+    png_const_bytep row, png_bytepp ssp, png_bytepp ddp)
+{
+   png_uint_32 row_width = row_info->width;
+   png_const_bytep palette = (png_const_bytep)png_ptr->palette;
+   const png_uint_32 pixels_per_chunk = 8;
+   int i;
+
+   png_debug(1, "in png_do_expand_palette_rgb8_neon");
+
+   if (row_width <= pixels_per_chunk)
+      return 0;
+
+   /* Seeking this back by 8 pixels x 3 bytes. */
+   *ddp = *ddp - ((pixels_per_chunk * sizeof(png_color)) - 1);
+
+   for (i = 0; i < row_width; i += pixels_per_chunk)
+   {
+      uint8x8x3_t cur;
+      png_bytep sp = *ssp - i, dp = *ddp - ((i << 1) + i);
+      cur = vld3_dup_u8(palette + sizeof(png_color) * (*(sp - 7)));
+      cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 6)), cur, 1);
+      cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 5)), cur, 2);
+      cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 4)), cur, 3);
+      cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 3)), cur, 4);
+      cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 2)), cur, 5);
+      cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 1)), cur, 6);
+      cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 0)), cur, 7);
+      vst3_u8((void *)dp, cur);
+   }
+
+   if (i != row_width)
+   {
+      /* Remove the amount that wasn't processed. */
+      i -= pixels_per_chunk;
+   }
+
+   /* Decrement output pointers. */
+   *ssp = *ssp - i;
+   *ddp = *ddp - ((i << 1) + i);
+   return i;
+}
+
+#endif /* PNG_ARM_NEON_IMPLEMENTATION */