workbench/libs/png/arm/palette_neon_intrinsics.c

   1
   2 /* palette_neon_intrinsics.c - NEON optimised palette expansion functions
   3  *
   4  * Copyright (c) 2018 Cosmin Truta
   5  * Copyright (c) 2017-2018 Arm Holdings. All rights reserved.
   6  * Written by Richard Townsend <Richard.Townsend@arm.com>, February 2017.
   7  *
   8  * This code is released under the libpng license.
   9  * For conditions of distribution and use, see the disclaimer
  10  * and license in png.h
  11  */
  12
  13 #include "../pngpriv.h"
  14
  15 #if PNG_ARM_NEON_IMPLEMENTATION == 1
  16
  17 #if defined(_MSC_VER) && defined(_M_ARM64)
  18 #  include <arm64_neon.h>
  19 #else
  20 #  include <arm_neon.h>
  21 #endif
  22
  23 /* Build an RGBA palette from the RGB and separate alpha palettes. */
  24 void
  25 png_riffle_palette_rgba(png_structrp png_ptr, png_row_infop row_info)
  26 {
  27    png_const_colorp palette = png_ptr->palette;
  28    png_bytep riffled_palette = png_ptr->riffled_palette;
  29    png_const_bytep trans_alpha = png_ptr->trans_alpha;
  30    int num_trans = png_ptr->num_trans;
  31    int i;
  32
  33    /* Initially black, opaque. */
  34    uint8x16x4_t w = {{
  35       vdupq_n_u8(0x00),
  36       vdupq_n_u8(0x00),
  37       vdupq_n_u8(0x00),
  38       vdupq_n_u8(0xff),
  39    }};
  40
  41    if (row_info->bit_depth != 8)
  42    {
  43       png_error(png_ptr, "bit_depth must be 8 for png_riffle_palette_rgba");
  44       return;
  45    }
  46
  47    /* First, riffle the RGB colours into a RGBA palette, the A value is
  48     * set to opaque for now.
  49     */
  50    for (i = 0; i < (1 << row_info->bit_depth); i += 16)
  51    {
  52       uint8x16x3_t v = vld3q_u8((png_const_bytep)(palette + i));
  53       w.val[0] = v.val[0];
  54       w.val[1] = v.val[1];
  55       w.val[2] = v.val[2];
  56       vst4q_u8(riffled_palette + (i << 2), w);
  57    }
  58
  59    /* Fix up the missing transparency values. */
  60    for (i = 0; i < num_trans; i++)
  61       riffled_palette[(i << 2) + 3] = trans_alpha[i];
  62 }
  63
  64 /* Expands a palettized row into RGBA. */
  65 int
  66 png_do_expand_palette_neon_rgba(png_structrp png_ptr, png_row_infop row_info,
  67     png_const_bytep row, png_bytepp ssp, png_bytepp ddp)
  68 {
  69    png_uint_32 row_width = row_info->width;
  70    const png_uint_32 *riffled_palette =
  71       (const png_uint_32 *)png_ptr->riffled_palette;
  72    const png_int_32 pixels_per_chunk = 4;
  73    int i;
  74
  75    if (row_width < pixels_per_chunk)
  76       return 0;
  77
  78    /* This function originally gets the last byte of the output row.
  79     * The NEON part writes forward from a given position, so we have
  80     * to seek this back by 4 pixels x 4 bytes.
  81     */
  82    *ddp = *ddp - ((pixels_per_chunk * sizeof(png_uint_32)) - 1);
  83
  84    for (i = 0; i < row_width; i += pixels_per_chunk)
  85    {
  86       uint32x4_t cur;
  87       png_bytep sp = *ssp - i, dp = *ddp - (i << 2);
  88       cur = vld1q_dup_u32 (riffled_palette + *(sp - 3));
  89       cur = vld1q_lane_u32(riffled_palette + *(sp - 2), cur, 1);
  90       cur = vld1q_lane_u32(riffled_palette + *(sp - 1), cur, 2);
  91       cur = vld1q_lane_u32(riffled_palette + *(sp - 0), cur, 3);
  92       vst1q_u32((void *)dp, cur);
  93    }
  94    if (i != row_width)
  95    {
  96       /* Remove the amount that wasn't processed. */
  97       i -= pixels_per_chunk;
  98    }
  99
 100    /* Decrement output pointers. */
 101    *ssp = *ssp - i;
 102    *ddp = *ddp - (i << 2);
 103    return i;
 104 }
 105
 106 /* Expands a palettized row into RGB format. */
 107 int
 108 png_do_expand_palette_neon_rgb(png_structrp png_ptr, png_row_infop row_info,
 109     png_const_bytep row, png_bytepp ssp, png_bytepp ddp)
 110 {
 111    png_uint_32 row_width = row_info->width;
 112    png_const_bytep palette = (png_const_bytep)png_ptr->palette;
 113    const png_uint_32 pixels_per_chunk = 8;
 114    int i;
 115
 116    if (row_width <= pixels_per_chunk)
 117       return 0;
 118
 119    /* Seeking this back by 8 pixels x 3 bytes. */
 120    *ddp = *ddp - ((pixels_per_chunk * sizeof(png_color)) - 1);
 121
 122    for (i = 0; i < row_width; i += pixels_per_chunk)
 123    {
 124       uint8x8x3_t cur;
 125       png_bytep sp = *ssp - i, dp = *ddp - ((i << 1) + i);
 126       cur = vld3_dup_u8(palette + sizeof(png_color) * (*(sp - 7)));
 127       cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 6)), cur, 1);
 128       cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 5)), cur, 2);
 129       cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 4)), cur, 3);
 130       cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 3)), cur, 4);
 131       cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 2)), cur, 5);
 132       cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 1)), cur, 6);
 133       cur = vld3_lane_u8(palette + sizeof(png_color) * (*(sp - 0)), cur, 7);
 134       vst3_u8((void *)dp, cur);
 135    }
 136
 137    if (i != row_width)
 138    {
 139       /* Remove the amount that wasn't processed. */
 140       i -= pixels_per_chunk;
 141    }
 142
 143    /* Decrement output pointers. */
 144    *ssp = *ssp - i;
 145    *ddp = *ddp - ((i << 1) + i);
 146    return i;
 147 }
 148
 149 #endif /* PNG_ARM_NEON_IMPLEMENTATION */