2 /* palette_neon_intrinsics.c - NEON optimised palette expansion functions
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.
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
13 #include "../pngpriv.h"
15 #if PNG_ARM_NEON_IMPLEMENTATION == 1
17 #if defined(_MSC_VER) && defined(_M_ARM64)
18 # include <arm64_neon.h>
20 # include <arm_neon.h>
23 /* Build an RGBA palette from the RGB and separate alpha palettes. */
25 png_riffle_palette_rgba(png_structrp png_ptr, png_row_infop row_info)
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;
33 /* Initially black, opaque. */
41 if (row_info->bit_depth != 8)
43 png_error(png_ptr, "bit_depth must be 8 for png_riffle_palette_rgba");
47 /* First, riffle the RGB colours into a RGBA palette, the A value is
48 * set to opaque for now.
50 for (i = 0; i < (1 << row_info->bit_depth); i += 16)
52 uint8x16x3_t v = vld3q_u8((png_const_bytep)(palette + i));
56 vst4q_u8(riffled_palette + (i << 2), w);
59 /* Fix up the missing transparency values. */
60 for (i = 0; i < num_trans; i++)
61 riffled_palette[(i << 2) + 3] = trans_alpha[i];
64 /* Expands a palettized row into RGBA. */
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)
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;
75 if (row_width < pixels_per_chunk)
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.
82 *ddp = *ddp - ((pixels_per_chunk * sizeof(png_uint_32)) - 1);
84 for (i = 0; i < row_width; i += pixels_per_chunk)
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);
96 /* Remove the amount that wasn't processed. */
97 i -= pixels_per_chunk;
100 /* Decrement output pointers. */
102 *ddp = *ddp - (i << 2);
106 /* Expands a palettized row into RGB format. */
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)
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;
116 if (row_width <= pixels_per_chunk)
119 /* Seeking this back by 8 pixels x 3 bytes. */
120 *ddp = *ddp - ((pixels_per_chunk * sizeof(png_color)) - 1);
122 for (i = 0; i < row_width; i += pixels_per_chunk)
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);
139 /* Remove the amount that wasn't processed. */
140 i -= pixels_per_chunk;
143 /* Decrement output pointers. */
145 *ddp = *ddp - ((i << 1) + i);
149 #endif /* PNG_ARM_NEON_IMPLEMENTATION */