[master] Imported from libpng-1.6.36.tar

arm/arm_init.c

@@ -1,14 +1,15 @@
 
 /* arm_init.c - NEON optimised filter functions
  *
+ * Copyright (c) 2018 Cosmin Truta
  * Copyright (c) 2014,2016 Glenn Randers-Pehrson
  * Written by Mans Rullgard, 2011.
- * Last changed in libpng 1.6.22 [May 26, 2016]
  *
  * 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.
  */

arm/filter_neon.S

@@ -1,9 +1,9 @@
 
 /* filter_neon.S - NEON optimised filter functions
  *
+ * Copyright (c) 2018 Cosmin Truta
  * Copyright (c) 2014,2017 Glenn Randers-Pehrson
  * Written by Mans Rullgard, 2011.
- * Last changed in libpng 1.6.31 [July 27, 2017]
  *
  * This code is released under the libpng license.
  * For conditions of distribution and use, see the disclaimer

arm/filter_neon_intrinsics.c

@@ -1,12 +1,11 @@
 
 /* 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.
  *
- * Last changed in libpng 1.6.22 [May 26, 2016]
- *
  * This code is released under the libpng license.
  * For conditions of distribution and use, see the disclaimer
  * and license in png.h
@@ -19,7 +18,11 @@
 /* This code requires -mfpu=neon on the command line: */
 #if PNG_ARM_NEON_IMPLEMENTATION == 1 /* intrinsics code from pngpriv.h */
 
-#include <arm_neon.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
@@ -33,6 +36,11 @@
  * '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)
@@ -125,12 +133,15 @@ png_read_filter_row_sub4_neon(png_row_infop row_info, png_bytep row,
       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]);
-      vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
+
+      vdest_val = png_ldr(uint32x2x4_t, &vdest);
+      vst4_lane_u32(png_ptr(uint32_t,rp), vdest_val, 0);
    }
 
    PNG_UNUSED(prev_row)
@@ -223,6 +234,7 @@ png_read_filter_row_avg4_neon(png_row_infop row_info, png_bytep row,
       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);
@@ -240,7 +252,8 @@ png_read_filter_row_avg4_neon(png_row_infop row_info, png_bytep row,
       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);
+      vdest_val = png_ldr(uint32x2x4_t, &vdest);
+      vst4_lane_u32(png_ptr(uint32_t,rp), vdest_val, 0);
    }
 }
 
@@ -359,6 +372,7 @@ png_read_filter_row_paeth4_neon(png_row_infop row_info, png_bytep row,
       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);
@@ -378,7 +392,8 @@ png_read_filter_row_paeth4_neon(png_row_infop row_info, png_bytep row,
 
       vlast = vpp.val[3];
 
-      vst4_lane_u32(png_ptr(uint32_t,rp), png_ldr(uint32x2x4_t,&vdest), 0);
+      vdest_val = png_ldr(uint32x2x4_t, &vdest);
+      vst4_lane_u32(png_ptr(uint32_t,rp), vdest_val, 0);
    }
 }
 

arm/palette_neon_intrinsics.c

@@ -0,0 +1,149 @@
+
+/* palette_neon_intrinsics.c - NEON optimised palette expansion functions
+ *
+ * Copyright (c) 2018 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 RGBA palette from the RGB and separate alpha palettes. */
+void
+png_riffle_palette_rgba(png_structrp png_ptr, png_row_infop row_info)
+{
+   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;
+
+   /* Initially black, opaque. */
+   uint8x16x4_t w = {{
+      vdupq_n_u8(0x00),
+      vdupq_n_u8(0x00),
+      vdupq_n_u8(0x00),
+      vdupq_n_u8(0xff),
+   }};
+
+   if (row_info->bit_depth != 8)
+   {
+      png_error(png_ptr, "bit_depth must be 8 for png_riffle_palette_rgba");
+      return;
+   }
+
+   /* First, riffle the RGB colours into a RGBA palette, the A value is
+    * set to opaque for now.
+    */
+   for (i = 0; i < (1 << row_info->bit_depth); 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 RGBA. */
+int
+png_do_expand_palette_neon_rgba(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;
+
+   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 RGB format. */
+int
+png_do_expand_palette_neon_rgb(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;
+
+   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 */