1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
|
/*
* Copyright 2012 The Android Open Source Project
*
* Use of this source code is governed by a BSD-style license that can be
* found in the LICENSE file.
*/
#include "SkBlitRow_opts_arm_neon.h"
#include "SkBlitMask.h"
#include "SkBlitRow.h"
#include "SkColorData.h"
#include "SkDither.h"
#include "SkMathPriv.h"
#include "SkUtils.h"
#include "SkColor_opts_neon.h"
#include <arm_neon.h>
/* Neon version of S32_Blend_BlitRow32()
* portable version is in src/core/SkBlitRow_D32.cpp
*/
void S32_Blend_BlitRow32_neon(SkPMColor* SK_RESTRICT dst,
const SkPMColor* SK_RESTRICT src,
int count, U8CPU alpha) {
SkASSERT(alpha <= 255);
if (count <= 0) {
return;
}
uint16_t src_scale = SkAlpha255To256(alpha);
uint16_t dst_scale = 256 - src_scale;
while (count >= 2) {
uint8x8_t vsrc, vdst, vres;
uint16x8_t vsrc_wide, vdst_wide;
/* These commented prefetches are a big win for count
* values > 64 on an A9 (Pandaboard) but hurt by 10% for count = 4.
* They also hurt a little (<5%) on an A15
*/
//__builtin_prefetch(src+32);
//__builtin_prefetch(dst+32);
// Load
vsrc = vreinterpret_u8_u32(vld1_u32(src));
vdst = vreinterpret_u8_u32(vld1_u32(dst));
// Process src
vsrc_wide = vmovl_u8(vsrc);
vsrc_wide = vmulq_u16(vsrc_wide, vdupq_n_u16(src_scale));
// Process dst
vdst_wide = vmull_u8(vdst, vdup_n_u8(dst_scale));
// Combine
vdst_wide += vsrc_wide;
vres = vshrn_n_u16(vdst_wide, 8);
// Store
vst1_u32(dst, vreinterpret_u32_u8(vres));
src += 2;
dst += 2;
count -= 2;
}
if (count == 1) {
uint8x8_t vsrc = vdup_n_u8(0), vdst = vdup_n_u8(0), vres;
uint16x8_t vsrc_wide, vdst_wide;
// Load
vsrc = vreinterpret_u8_u32(vld1_lane_u32(src, vreinterpret_u32_u8(vsrc), 0));
vdst = vreinterpret_u8_u32(vld1_lane_u32(dst, vreinterpret_u32_u8(vdst), 0));
// Process
vsrc_wide = vmovl_u8(vsrc);
vsrc_wide = vmulq_u16(vsrc_wide, vdupq_n_u16(src_scale));
vdst_wide = vmull_u8(vdst, vdup_n_u8(dst_scale));
vdst_wide += vsrc_wide;
vres = vshrn_n_u16(vdst_wide, 8);
// Store
vst1_lane_u32(dst, vreinterpret_u32_u8(vres), 0);
}
}
#ifdef SK_CPU_ARM32
void S32A_Blend_BlitRow32_neon(SkPMColor* SK_RESTRICT dst,
const SkPMColor* SK_RESTRICT src,
int count, U8CPU alpha) {
SkASSERT(255 > alpha);
if (count <= 0) {
return;
}
unsigned alpha256 = SkAlpha255To256(alpha);
// First deal with odd counts
if (count & 1) {
uint8x8_t vsrc = vdup_n_u8(0), vdst = vdup_n_u8(0), vres;
uint16x8_t vdst_wide, vsrc_wide;
unsigned dst_scale;
// Load
vsrc = vreinterpret_u8_u32(vld1_lane_u32(src, vreinterpret_u32_u8(vsrc), 0));
vdst = vreinterpret_u8_u32(vld1_lane_u32(dst, vreinterpret_u32_u8(vdst), 0));
// Calc dst_scale
dst_scale = vget_lane_u8(vsrc, 3);
dst_scale = SkAlphaMulInv256(dst_scale, alpha256);
// Process src
vsrc_wide = vmovl_u8(vsrc);
vsrc_wide = vmulq_n_u16(vsrc_wide, alpha256);
// Process dst
vdst_wide = vmovl_u8(vdst);
vdst_wide = vmulq_n_u16(vdst_wide, dst_scale);
// Combine
vdst_wide += vsrc_wide;
vres = vshrn_n_u16(vdst_wide, 8);
vst1_lane_u32(dst, vreinterpret_u32_u8(vres), 0);
dst++;
src++;
count--;
}
if (count) {
uint8x8_t alpha_mask;
static const uint8_t alpha_mask_setup[] = {3,3,3,3,7,7,7,7};
alpha_mask = vld1_u8(alpha_mask_setup);
do {
uint8x8_t vsrc, vdst, vres, vsrc_alphas;
uint16x8_t vdst_wide, vsrc_wide, vsrc_scale, vdst_scale;
__builtin_prefetch(src+32);
__builtin_prefetch(dst+32);
// Load
vsrc = vreinterpret_u8_u32(vld1_u32(src));
vdst = vreinterpret_u8_u32(vld1_u32(dst));
// Prepare src_scale
vsrc_scale = vdupq_n_u16(alpha256);
// Calc dst_scale
vsrc_alphas = vtbl1_u8(vsrc, alpha_mask);
vdst_scale = vmovl_u8(vsrc_alphas);
// Calculate SkAlphaMulInv256(vdst_scale, vsrc_scale).
// A 16-bit lane would overflow if we used 0xFFFF here,
// so use an approximation with 0xFF00 that is off by 1,
// and add back 1 after to get the correct value.
// This is valid if alpha256 <= 255.
vdst_scale = vmlsq_u16(vdupq_n_u16(0xFF00), vdst_scale, vsrc_scale);
vdst_scale = vsraq_n_u16(vdst_scale, vdst_scale, 8);
vdst_scale = vsraq_n_u16(vdupq_n_u16(1), vdst_scale, 8);
// Process src
vsrc_wide = vmovl_u8(vsrc);
vsrc_wide *= vsrc_scale;
// Process dst
vdst_wide = vmovl_u8(vdst);
vdst_wide *= vdst_scale;
// Combine
vdst_wide += vsrc_wide;
vres = vshrn_n_u16(vdst_wide, 8);
vst1_u32(dst, vreinterpret_u32_u8(vres));
src += 2;
dst += 2;
count -= 2;
} while(count);
}
}
///////////////////////////////////////////////////////////////////////////////
#endif // #ifdef SK_CPU_ARM32
///////////////////////////////////////////////////////////////////////////////
const SkBlitRow::Proc32 sk_blitrow_platform_32_procs_arm_neon[] = {
nullptr, // S32_Opaque,
S32_Blend_BlitRow32_neon, // S32_Blend,
nullptr, // Ported to SkOpts
#ifdef SK_CPU_ARM32
S32A_Blend_BlitRow32_neon // S32A_Blend
#else
nullptr
#endif
};
|
