1// Copyright 2015 Google Inc. All Rights Reserved.
2//
3// Use of this source code is governed by a BSD-style license
4// that can be found in the COPYING file in the root of the source
5// tree. An additional intellectual property rights grant can be found
6// in the file PATENTS. All contributing project authors may
7// be found in the AUTHORS file in the root of the source tree.
8// -----------------------------------------------------------------------------
9//
10// NEON variant of methods for lossless encoder
11//
12// Author: Skal (pascal.massimino@gmail.com)
13
14#include "./dsp.h"
15
16#if defined(WEBP_USE_NEON)
17
18#include <arm_neon.h>
19
20#include "./lossless.h"
21#include "./neon.h"
22
23//------------------------------------------------------------------------------
24// Subtract-Green Transform
25
26// vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use
27// non-standard versions there.
28#if defined(__APPLE__) && defined(__aarch64__) && \
29    defined(__apple_build_version__) && (__apple_build_version__< 6020037)
30#define USE_VTBLQ
31#endif
32
33#ifdef USE_VTBLQ
34// 255 = byte will be zeroed
35static const uint8_t kGreenShuffle[16] = {
36  1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255
37};
38
39static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb,
40                                             const uint8x16_t shuffle) {
41  return vcombine_u8(vtbl1q_u8(argb, vget_low_u8(shuffle)),
42                     vtbl1q_u8(argb, vget_high_u8(shuffle)));
43}
44#else  // !USE_VTBLQ
45// 255 = byte will be zeroed
46static const uint8_t kGreenShuffle[8] = { 1, 255, 1, 255, 5, 255, 5, 255  };
47
48static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb,
49                                             const uint8x8_t shuffle) {
50  return vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle),
51                     vtbl1_u8(vget_high_u8(argb), shuffle));
52}
53#endif  // USE_VTBLQ
54
55static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
56  const uint32_t* const end = argb_data + (num_pixels & ~3);
57#ifdef USE_VTBLQ
58  const uint8x16_t shuffle = vld1q_u8(kGreenShuffle);
59#else
60  const uint8x8_t shuffle = vld1_u8(kGreenShuffle);
61#endif
62  for (; argb_data < end; argb_data += 4) {
63    const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data);
64    const uint8x16_t greens = DoGreenShuffle(argb, shuffle);
65    vst1q_u8((uint8_t*)argb_data, vsubq_u8(argb, greens));
66  }
67  // fallthrough and finish off with plain-C
68  VP8LSubtractGreenFromBlueAndRed_C(argb_data, num_pixels & 3);
69}
70
71//------------------------------------------------------------------------------
72// Color Transform
73
74static void TransformColor(const VP8LMultipliers* const m,
75                           uint32_t* argb_data, int num_pixels) {
76  // sign-extended multiplying constants, pre-shifted by 6.
77#define CST(X)  (((int16_t)(m->X << 8)) >> 6)
78  const int16_t rb[8] = {
79    CST(green_to_blue_), CST(green_to_red_),
80    CST(green_to_blue_), CST(green_to_red_),
81    CST(green_to_blue_), CST(green_to_red_),
82    CST(green_to_blue_), CST(green_to_red_)
83  };
84  const int16x8_t mults_rb = vld1q_s16(rb);
85  const int16_t b2[8] = {
86    0, CST(red_to_blue_), 0, CST(red_to_blue_),
87    0, CST(red_to_blue_), 0, CST(red_to_blue_),
88  };
89  const int16x8_t mults_b2 = vld1q_s16(b2);
90#undef CST
91#ifdef USE_VTBLQ
92  static const uint8_t kg0g0[16] = {
93    255, 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13
94  };
95  const uint8x16_t shuffle = vld1q_u8(kg0g0);
96#else
97  static const uint8_t k0g0g[8] = { 255, 1, 255, 1, 255, 5, 255, 5 };
98  const uint8x8_t shuffle = vld1_u8(k0g0g);
99#endif
100  const uint32x4_t mask_rb = vdupq_n_u32(0x00ff00ffu);  // red-blue masks
101  int i;
102  for (i = 0; i + 4 <= num_pixels; i += 4) {
103    const uint8x16_t in = vld1q_u8((uint8_t*)(argb_data + i));
104    // 0 g 0 g
105    const uint8x16_t greens = DoGreenShuffle(in, shuffle);
106    // x dr  x db1
107    const int16x8_t A = vqdmulhq_s16(vreinterpretq_s16_u8(greens), mults_rb);
108    // r 0   b   0
109    const int16x8_t B = vshlq_n_s16(vreinterpretq_s16_u8(in), 8);
110    // x db2 0   0
111    const int16x8_t C = vqdmulhq_s16(B, mults_b2);
112    // 0 0   x db2
113    const uint32x4_t D = vshrq_n_u32(vreinterpretq_u32_s16(C), 16);
114    // x dr  x  db
115    const int8x16_t E = vaddq_s8(vreinterpretq_s8_u32(D),
116                                 vreinterpretq_s8_s16(A));
117    // 0 dr  0  db
118    const uint32x4_t F = vandq_u32(vreinterpretq_u32_s8(E), mask_rb);
119    const int8x16_t out = vsubq_s8(vreinterpretq_s8_u8(in),
120                                   vreinterpretq_s8_u32(F));
121    vst1q_s8((int8_t*)(argb_data + i), out);
122  }
123  // fallthrough and finish off with plain-C
124  VP8LTransformColor_C(m, argb_data + i, num_pixels - i);
125}
126
127#undef USE_VTBLQ
128
129//------------------------------------------------------------------------------
130// Entry point
131
132extern void VP8LEncDspInitNEON(void);
133
134WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitNEON(void) {
135  VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
136  VP8LTransformColor = TransformColor;
137}
138
139#else  // !WEBP_USE_NEON
140
141WEBP_DSP_INIT_STUB(VP8LEncDspInitNEON)
142
143#endif  // WEBP_USE_NEON
144