1// Copyright 2017 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// Utilities for processing transparent channel, NEON version.
11//
12// Author: Skal (pascal.massimino@gmail.com)
13
14#include "./dsp.h"
15
16#if defined(WEBP_USE_NEON)
17
18#include "./neon.h"
19
20//------------------------------------------------------------------------------
21
22#define MULTIPLIER(a) ((a) * 0x8081)
23#define PREMULTIPLY(x, m) (((x) * (m)) >> 23)
24
25#define MULTIPLY_BY_ALPHA(V, ALPHA, OTHER) do {                        \
26  const uint8x8_t alpha = (V).val[(ALPHA)];                            \
27  const uint16x8_t r1 = vmull_u8((V).val[1], alpha);                   \
28  const uint16x8_t g1 = vmull_u8((V).val[2], alpha);                   \
29  const uint16x8_t b1 = vmull_u8((V).val[(OTHER)], alpha);             \
30  /* we use: v / 255 = (v + 1 + (v >> 8)) >> 8 */                      \
31  const uint16x8_t r2 = vsraq_n_u16(r1, r1, 8);                        \
32  const uint16x8_t g2 = vsraq_n_u16(g1, g1, 8);                        \
33  const uint16x8_t b2 = vsraq_n_u16(b1, b1, 8);                        \
34  const uint16x8_t r3 = vaddq_u16(r2, kOne);                           \
35  const uint16x8_t g3 = vaddq_u16(g2, kOne);                           \
36  const uint16x8_t b3 = vaddq_u16(b2, kOne);                           \
37  (V).val[1] = vshrn_n_u16(r3, 8);                                     \
38  (V).val[2] = vshrn_n_u16(g3, 8);                                     \
39  (V).val[(OTHER)] = vshrn_n_u16(b3, 8);                               \
40} while (0)
41
42static void ApplyAlphaMultiply_NEON(uint8_t* rgba, int alpha_first,
43                                    int w, int h, int stride) {
44  const uint16x8_t kOne = vdupq_n_u16(1u);
45  while (h-- > 0) {
46    uint32_t* const rgbx = (uint32_t*)rgba;
47    int i = 0;
48    if (alpha_first) {
49      for (; i + 8 <= w; i += 8) {
50        // load aaaa...|rrrr...|gggg...|bbbb...
51        uint8x8x4_t RGBX = vld4_u8((const uint8_t*)(rgbx + i));
52        MULTIPLY_BY_ALPHA(RGBX, 0, 3);
53        vst4_u8((uint8_t*)(rgbx + i), RGBX);
54      }
55    } else {
56      for (; i + 8 <= w; i += 8) {
57        uint8x8x4_t RGBX = vld4_u8((const uint8_t*)(rgbx + i));
58        MULTIPLY_BY_ALPHA(RGBX, 3, 0);
59        vst4_u8((uint8_t*)(rgbx + i), RGBX);
60      }
61    }
62    // Finish with left-overs.
63    for (; i < w; ++i) {
64      uint8_t* const rgb = rgba + (alpha_first ? 1 : 0);
65      const uint8_t* const alpha = rgba + (alpha_first ? 0 : 3);
66      const uint32_t a = alpha[4 * i];
67      if (a != 0xff) {
68        const uint32_t mult = MULTIPLIER(a);
69        rgb[4 * i + 0] = PREMULTIPLY(rgb[4 * i + 0], mult);
70        rgb[4 * i + 1] = PREMULTIPLY(rgb[4 * i + 1], mult);
71        rgb[4 * i + 2] = PREMULTIPLY(rgb[4 * i + 2], mult);
72      }
73    }
74    rgba += stride;
75  }
76}
77#undef MULTIPLY_BY_ALPHA
78#undef MULTIPLIER
79#undef PREMULTIPLY
80
81//------------------------------------------------------------------------------
82
83static int DispatchAlpha_NEON(const uint8_t* alpha, int alpha_stride,
84                              int width, int height,
85                              uint8_t* dst, int dst_stride) {
86  uint32_t alpha_mask = 0xffffffffu;
87  uint8x8_t mask8 = vdup_n_u8(0xff);
88  uint32_t tmp[2];
89  int i, j;
90  for (j = 0; j < height; ++j) {
91    // We don't know if alpha is first or last in dst[] (depending on rgbA/Argb
92    // mode). So we must be sure dst[4*i + 8 - 1] is writable for the store.
93    // Hence the test with 'width - 1' instead of just 'width'.
94    for (i = 0; i + 8 <= width - 1; i += 8) {
95      uint8x8x4_t rgbX = vld4_u8((const uint8_t*)(dst + 4 * i));
96      const uint8x8_t alphas = vld1_u8(alpha + i);
97      rgbX.val[0] = alphas;
98      vst4_u8((uint8_t*)(dst + 4 * i), rgbX);
99      mask8 = vand_u8(mask8, alphas);
100    }
101    for (; i < width; ++i) {
102      const uint32_t alpha_value = alpha[i];
103      dst[4 * i] = alpha_value;
104      alpha_mask &= alpha_value;
105    }
106    alpha += alpha_stride;
107    dst += dst_stride;
108  }
109  vst1_u8((uint8_t*)tmp, mask8);
110  alpha_mask &= tmp[0];
111  alpha_mask &= tmp[1];
112  return (alpha_mask != 0xffffffffu);
113}
114
115static void DispatchAlphaToGreen_NEON(const uint8_t* alpha, int alpha_stride,
116                                      int width, int height,
117                                      uint32_t* dst, int dst_stride) {
118  int i, j;
119  uint8x8x4_t greens;   // leave A/R/B channels zero'd.
120  greens.val[0] = vdup_n_u8(0);
121  greens.val[2] = vdup_n_u8(0);
122  greens.val[3] = vdup_n_u8(0);
123  for (j = 0; j < height; ++j) {
124    for (i = 0; i + 8 <= width; i += 8) {
125      greens.val[1] = vld1_u8(alpha + i);
126      vst4_u8((uint8_t*)(dst + i), greens);
127    }
128    for (; i < width; ++i) dst[i] = alpha[i] << 8;
129    alpha += alpha_stride;
130    dst += dst_stride;
131  }
132}
133
134static int ExtractAlpha_NEON(const uint8_t* argb, int argb_stride,
135                             int width, int height,
136                             uint8_t* alpha, int alpha_stride) {
137  uint32_t alpha_mask = 0xffffffffu;
138  uint8x8_t mask8 = vdup_n_u8(0xff);
139  uint32_t tmp[2];
140  int i, j;
141  for (j = 0; j < height; ++j) {
142    // We don't know if alpha is first or last in dst[] (depending on rgbA/Argb
143    // mode). So we must be sure dst[4*i + 8 - 1] is writable for the store.
144    // Hence the test with 'width - 1' instead of just 'width'.
145    for (i = 0; i + 8 <= width - 1; i += 8) {
146      const uint8x8x4_t rgbX = vld4_u8((const uint8_t*)(argb + 4 * i));
147      const uint8x8_t alphas = rgbX.val[0];
148      vst1_u8((uint8_t*)(alpha + i), alphas);
149      mask8 = vand_u8(mask8, alphas);
150    }
151    for (; i < width; ++i) {
152      alpha[i] = argb[4 * i];
153      alpha_mask &= alpha[i];
154    }
155    argb += argb_stride;
156    alpha += alpha_stride;
157  }
158  vst1_u8((uint8_t*)tmp, mask8);
159  alpha_mask &= tmp[0];
160  alpha_mask &= tmp[1];
161  return (alpha_mask == 0xffffffffu);
162}
163
164static void ExtractGreen_NEON(const uint32_t* argb,
165                              uint8_t* alpha, int size) {
166  int i;
167  for (i = 0; i + 16 <= size; i += 16) {
168    const uint8x16x4_t rgbX = vld4q_u8((const uint8_t*)(argb + i));
169    const uint8x16_t greens = rgbX.val[1];
170    vst1q_u8(alpha + i, greens);
171  }
172  for (; i < size; ++i) alpha[i] = (argb[i] >> 8) & 0xff;
173}
174
175//------------------------------------------------------------------------------
176
177extern void WebPInitAlphaProcessingNEON(void);
178
179WEBP_TSAN_IGNORE_FUNCTION void WebPInitAlphaProcessingNEON(void) {
180  WebPApplyAlphaMultiply = ApplyAlphaMultiply_NEON;
181  WebPDispatchAlpha = DispatchAlpha_NEON;
182  WebPDispatchAlphaToGreen = DispatchAlphaToGreen_NEON;
183  WebPExtractAlpha = ExtractAlpha_NEON;
184  WebPExtractGreen = ExtractGreen_NEON;
185}
186
187#else  // !WEBP_USE_NEON
188
189WEBP_DSP_INIT_STUB(WebPInitAlphaProcessingNEON)
190
191#endif  // WEBP_USE_NEON
192