1#ifndef SkColor_opts_neon_DEFINED
2#define SkColor_opts_neon_DEFINED
3
4#include "SkTypes.h"
5#include "SkColorPriv.h"
6
7#include <arm_neon.h>
8
9#define NEON_A (SK_A32_SHIFT / 8)
10#define NEON_R (SK_R32_SHIFT / 8)
11#define NEON_G (SK_G32_SHIFT / 8)
12#define NEON_B (SK_B32_SHIFT / 8)
13
14static inline uint16x8_t SkAlpha255To256_neon8(uint8x8_t alpha) {
15    return vaddw_u8(vdupq_n_u16(1), alpha);
16}
17
18static inline uint8x8_t SkAlphaMul_neon8(uint8x8_t color, uint16x8_t scale) {
19    return vshrn_n_u16(vmovl_u8(color) * scale, 8);
20}
21
22static inline uint8x8x4_t SkAlphaMulQ_neon8(uint8x8x4_t color, uint16x8_t scale) {
23    uint8x8x4_t ret;
24
25    ret.val[NEON_A] = SkAlphaMul_neon8(color.val[NEON_A], scale);
26    ret.val[NEON_R] = SkAlphaMul_neon8(color.val[NEON_R], scale);
27    ret.val[NEON_G] = SkAlphaMul_neon8(color.val[NEON_G], scale);
28    ret.val[NEON_B] = SkAlphaMul_neon8(color.val[NEON_B], scale);
29
30    return ret;
31}
32
33/* This function expands 8 pixels from RGB565 (R, G, B from high to low) to
34 * SkPMColor (all possible configurations supported) in the exact same way as
35 * SkPixel16ToPixel32.
36 */
37static inline uint8x8x4_t SkPixel16ToPixel32_neon8(uint16x8_t vsrc) {
38
39    uint8x8x4_t ret;
40    uint8x8_t vr, vg, vb;
41
42    vr = vmovn_u16(vshrq_n_u16(vsrc, SK_R16_SHIFT));
43    vg = vmovn_u16(vshrq_n_u16(vshlq_n_u16(vsrc, SK_R16_BITS), SK_R16_BITS + SK_B16_BITS));
44    vb = vmovn_u16(vsrc & vdupq_n_u16(SK_B16_MASK));
45
46    ret.val[NEON_A] = vdup_n_u8(0xFF);
47    ret.val[NEON_R] = vshl_n_u8(vr, 8 - SK_R16_BITS) | vshr_n_u8(vr, 2 * SK_R16_BITS - 8);
48    ret.val[NEON_G] = vshl_n_u8(vg, 8 - SK_G16_BITS) | vshr_n_u8(vg, 2 * SK_G16_BITS - 8);
49    ret.val[NEON_B] = vshl_n_u8(vb, 8 - SK_B16_BITS) | vshr_n_u8(vb, 2 * SK_B16_BITS - 8);
50
51    return ret;
52}
53
54/* This function packs 8 pixels from SkPMColor (all possible configurations
55 * supported) to RGB565 (R, G, B from high to low) in the exact same way as
56 * SkPixel32ToPixel16.
57 */
58static inline uint16x8_t SkPixel32ToPixel16_neon8(uint8x8x4_t vsrc) {
59
60    uint16x8_t ret;
61
62    ret = vshll_n_u8(vsrc.val[NEON_R], 8);
63    ret = vsriq_n_u16(ret, vshll_n_u8(vsrc.val[NEON_G], 8), SK_R16_BITS);
64    ret = vsriq_n_u16(ret, vshll_n_u8(vsrc.val[NEON_B], 8), SK_R16_BITS + SK_G16_BITS);
65
66    return ret;
67}
68
69/* This function blends 8 pixels of the same channel in the exact same way as
70 * SkBlend32.
71 */
72static inline uint8x8_t SkBlend32_neon8(uint8x8_t src, uint8x8_t dst, uint16x8_t scale) {
73    int16x8_t src_wide, dst_wide;
74
75    src_wide = vreinterpretq_s16_u16(vmovl_u8(src));
76    dst_wide = vreinterpretq_s16_u16(vmovl_u8(dst));
77
78    src_wide = (src_wide - dst_wide) * vreinterpretq_s16_u16(scale);
79
80    dst_wide += vshrq_n_s16(src_wide, 5);
81
82    return vmovn_u16(vreinterpretq_u16_s16(dst_wide));
83}
84
85static inline SkPMColor SkFourByteInterp256_neon(SkPMColor src, SkPMColor dst,
86                                                 unsigned srcScale) {
87    SkASSERT(srcScale <= 256);
88    int16x8_t vscale = vdupq_n_s16(srcScale);
89    int16x8_t vsrc_wide, vdst_wide, vdiff;
90    uint8x8_t res;
91
92    vsrc_wide = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(vdup_n_u32(src))));
93    vdst_wide = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(vdup_n_u32(dst))));
94
95    vdiff = vsrc_wide - vdst_wide;
96    vdiff *= vscale;
97
98    vdiff = vshrq_n_s16(vdiff, 8);
99
100    vdst_wide += vdiff;
101
102    res = vmovn_u16(vreinterpretq_u16_s16(vdst_wide));
103
104    return vget_lane_u32(vreinterpret_u32_u8(res), 0);
105}
106
107static inline SkPMColor SkFourByteInterp_neon(SkPMColor src, SkPMColor dst,
108                                              U8CPU srcWeight) {
109    SkASSERT(srcWeight <= 255);
110    unsigned scale = SkAlpha255To256(srcWeight);
111    return SkFourByteInterp256_neon(src, dst, scale);
112}
113
114#endif /* #ifndef SkColor_opts_neon_DEFINED */
115