1 
2 /* palette_neon_intrinsics.c - NEON optimised palette expansion functions
3  *
4  * Copyright (c) 2018-2019 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 RGBA8 palette from the separate RGB and alpha palettes. */
24 void
png_riffle_palette_neon(png_structrp png_ptr)25 png_riffle_palette_neon(png_structrp png_ptr)
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    png_debug(1, "in png_riffle_palette_neon");
34 
35    /* Initially black, opaque. */
36    uint8x16x4_t w = {{
37       vdupq_n_u8(0x00),
38       vdupq_n_u8(0x00),
39       vdupq_n_u8(0x00),
40       vdupq_n_u8(0xff),
41    }};
42 
43    /* First, riffle the RGB colours into an RGBA8 palette.
44     * The alpha component is set to opaque for now.
45     */
46    for (i = 0; i < 256; i += 16)
47    {
48       uint8x16x3_t v = vld3q_u8((png_const_bytep)(palette + i));
49       w.val[0] = v.val[0];
50       w.val[1] = v.val[1];
51       w.val[2] = v.val[2];
52       vst4q_u8(riffled_palette + (i << 2), w);
53    }
54 
55    /* Fix up the missing transparency values. */
56    for (i = 0; i < num_trans; i++)
57       riffled_palette[(i << 2) + 3] = trans_alpha[i];
58 }
59 
60 /* Expands a palettized row into RGBA8. */
61 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)62 png_do_expand_palette_rgba8_neon(png_structrp png_ptr, png_row_infop row_info,
63     png_const_bytep row, png_bytepp ssp, png_bytepp ddp)
64 {
65    png_uint_32 row_width = row_info->width;
66    const png_uint_32 *riffled_palette =
67       (const png_uint_32 *)png_ptr->riffled_palette;
68    const png_int_32 pixels_per_chunk = 4;
69    int i;
70 
71    png_debug(1, "in png_do_expand_palette_rgba8_neon");
72 
73    if (row_width < pixels_per_chunk)
74       return 0;
75 
76    /* This function originally gets the last byte of the output row.
77     * The NEON part writes forward from a given position, so we have
78     * to seek this back by 4 pixels x 4 bytes.
79     */
80    *ddp = *ddp - ((pixels_per_chunk * sizeof(png_uint_32)) - 1);
81 
82    for (i = 0; i < row_width; i += pixels_per_chunk)
83    {
84       uint32x4_t cur;
85       png_bytep sp = *ssp - i, dp = *ddp - (i << 2);
86       cur = vld1q_dup_u32 (riffled_palette + *(sp - 3));
87       cur = vld1q_lane_u32(riffled_palette + *(sp - 2), cur, 1);
88       cur = vld1q_lane_u32(riffled_palette + *(sp - 1), cur, 2);
89       cur = vld1q_lane_u32(riffled_palette + *(sp - 0), cur, 3);
90       vst1q_u32((void *)dp, cur);
91    }
92    if (i != row_width)
93    {
94       /* Remove the amount that wasn't processed. */
95       i -= pixels_per_chunk;
96    }
97 
98    /* Decrement output pointers. */
99    *ssp = *ssp - i;
100    *ddp = *ddp - (i << 2);
101    return i;
102 }
103 
104 /* Expands a palettized row into RGB8. */
105 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)106 png_do_expand_palette_rgb8_neon(png_structrp png_ptr, png_row_infop row_info,
107     png_const_bytep row, png_bytepp ssp, png_bytepp ddp)
108 {
109    png_uint_32 row_width = row_info->width;
110    png_const_bytep palette = (png_const_bytep)png_ptr->palette;
111    const png_uint_32 pixels_per_chunk = 8;
112    int i;
113 
114    png_debug(1, "in png_do_expand_palette_rgb8_neon");
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 */
150