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 // YUV->RGB conversion functions
11 //
12 // Author: Skal (pascal.massimino@gmail.com)
13
14 #include "src/dsp/yuv.h"
15
16 #if defined(WEBP_USE_NEON)
17
18 #include <assert.h>
19 #include <stdlib.h>
20
21 #include "src/dsp/neon.h"
22
23 //-----------------------------------------------------------------------------
24
ConvertRGBToY_NEON(const uint8x8_t R,const uint8x8_t G,const uint8x8_t B)25 static uint8x8_t ConvertRGBToY_NEON(const uint8x8_t R,
26 const uint8x8_t G,
27 const uint8x8_t B) {
28 const uint16x8_t r = vmovl_u8(R);
29 const uint16x8_t g = vmovl_u8(G);
30 const uint16x8_t b = vmovl_u8(B);
31 const uint16x4_t r_lo = vget_low_u16(r);
32 const uint16x4_t r_hi = vget_high_u16(r);
33 const uint16x4_t g_lo = vget_low_u16(g);
34 const uint16x4_t g_hi = vget_high_u16(g);
35 const uint16x4_t b_lo = vget_low_u16(b);
36 const uint16x4_t b_hi = vget_high_u16(b);
37 const uint32x4_t tmp0_lo = vmull_n_u16( r_lo, 16839u);
38 const uint32x4_t tmp0_hi = vmull_n_u16( r_hi, 16839u);
39 const uint32x4_t tmp1_lo = vmlal_n_u16(tmp0_lo, g_lo, 33059u);
40 const uint32x4_t tmp1_hi = vmlal_n_u16(tmp0_hi, g_hi, 33059u);
41 const uint32x4_t tmp2_lo = vmlal_n_u16(tmp1_lo, b_lo, 6420u);
42 const uint32x4_t tmp2_hi = vmlal_n_u16(tmp1_hi, b_hi, 6420u);
43 const uint16x8_t Y1 = vcombine_u16(vrshrn_n_u32(tmp2_lo, 16),
44 vrshrn_n_u32(tmp2_hi, 16));
45 const uint16x8_t Y2 = vaddq_u16(Y1, vdupq_n_u16(16));
46 return vqmovn_u16(Y2);
47 }
48
ConvertRGB24ToY_NEON(const uint8_t * WEBP_RESTRICT rgb,uint8_t * WEBP_RESTRICT y,int width)49 static void ConvertRGB24ToY_NEON(const uint8_t* WEBP_RESTRICT rgb,
50 uint8_t* WEBP_RESTRICT y, int width) {
51 int i;
52 for (i = 0; i + 8 <= width; i += 8, rgb += 3 * 8) {
53 const uint8x8x3_t RGB = vld3_u8(rgb);
54 const uint8x8_t Y = ConvertRGBToY_NEON(RGB.val[0], RGB.val[1], RGB.val[2]);
55 vst1_u8(y + i, Y);
56 }
57 for (; i < width; ++i, rgb += 3) { // left-over
58 y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF);
59 }
60 }
61
ConvertBGR24ToY_NEON(const uint8_t * WEBP_RESTRICT bgr,uint8_t * WEBP_RESTRICT y,int width)62 static void ConvertBGR24ToY_NEON(const uint8_t* WEBP_RESTRICT bgr,
63 uint8_t* WEBP_RESTRICT y, int width) {
64 int i;
65 for (i = 0; i + 8 <= width; i += 8, bgr += 3 * 8) {
66 const uint8x8x3_t BGR = vld3_u8(bgr);
67 const uint8x8_t Y = ConvertRGBToY_NEON(BGR.val[2], BGR.val[1], BGR.val[0]);
68 vst1_u8(y + i, Y);
69 }
70 for (; i < width; ++i, bgr += 3) { // left-over
71 y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF);
72 }
73 }
74
ConvertARGBToY_NEON(const uint32_t * WEBP_RESTRICT argb,uint8_t * WEBP_RESTRICT y,int width)75 static void ConvertARGBToY_NEON(const uint32_t* WEBP_RESTRICT argb,
76 uint8_t* WEBP_RESTRICT y, int width) {
77 int i;
78 for (i = 0; i + 8 <= width; i += 8) {
79 const uint8x8x4_t RGB = vld4_u8((const uint8_t*)&argb[i]);
80 const uint8x8_t Y = ConvertRGBToY_NEON(RGB.val[2], RGB.val[1], RGB.val[0]);
81 vst1_u8(y + i, Y);
82 }
83 for (; i < width; ++i) { // left-over
84 const uint32_t p = argb[i];
85 y[i] = VP8RGBToY((p >> 16) & 0xff, (p >> 8) & 0xff, (p >> 0) & 0xff,
86 YUV_HALF);
87 }
88 }
89
90 //-----------------------------------------------------------------------------
91
92 // computes: DST_s16 = [(C0 * r + C1 * g + C2 * b) >> 16] + CST
93 #define MULTIPLY_16b_PREAMBLE(r, g, b) \
94 const int16x4_t r_lo = vreinterpret_s16_u16(vget_low_u16(r)); \
95 const int16x4_t r_hi = vreinterpret_s16_u16(vget_high_u16(r)); \
96 const int16x4_t g_lo = vreinterpret_s16_u16(vget_low_u16(g)); \
97 const int16x4_t g_hi = vreinterpret_s16_u16(vget_high_u16(g)); \
98 const int16x4_t b_lo = vreinterpret_s16_u16(vget_low_u16(b)); \
99 const int16x4_t b_hi = vreinterpret_s16_u16(vget_high_u16(b))
100
101 #define MULTIPLY_16b(C0, C1, C2, CST, DST_s16) do { \
102 const int32x4_t tmp0_lo = vmull_n_s16( r_lo, C0); \
103 const int32x4_t tmp0_hi = vmull_n_s16( r_hi, C0); \
104 const int32x4_t tmp1_lo = vmlal_n_s16(tmp0_lo, g_lo, C1); \
105 const int32x4_t tmp1_hi = vmlal_n_s16(tmp0_hi, g_hi, C1); \
106 const int32x4_t tmp2_lo = vmlal_n_s16(tmp1_lo, b_lo, C2); \
107 const int32x4_t tmp2_hi = vmlal_n_s16(tmp1_hi, b_hi, C2); \
108 const int16x8_t tmp3 = vcombine_s16(vshrn_n_s32(tmp2_lo, 16), \
109 vshrn_n_s32(tmp2_hi, 16)); \
110 DST_s16 = vaddq_s16(tmp3, vdupq_n_s16(CST)); \
111 } while (0)
112
113 // This needs to be a macro, since (128 << SHIFT) needs to be an immediate.
114 #define CONVERT_RGB_TO_UV(r, g, b, SHIFT, U_DST, V_DST) do { \
115 MULTIPLY_16b_PREAMBLE(r, g, b); \
116 MULTIPLY_16b(-9719, -19081, 28800, 128 << SHIFT, U_DST); \
117 MULTIPLY_16b(28800, -24116, -4684, 128 << SHIFT, V_DST); \
118 } while (0)
119
ConvertRGBA32ToUV_NEON(const uint16_t * WEBP_RESTRICT rgb,uint8_t * WEBP_RESTRICT u,uint8_t * WEBP_RESTRICT v,int width)120 static void ConvertRGBA32ToUV_NEON(const uint16_t* WEBP_RESTRICT rgb,
121 uint8_t* WEBP_RESTRICT u,
122 uint8_t* WEBP_RESTRICT v, int width) {
123 int i;
124 for (i = 0; i + 8 <= width; i += 8, rgb += 4 * 8) {
125 const uint16x8x4_t RGB = vld4q_u16((const uint16_t*)rgb);
126 int16x8_t U, V;
127 CONVERT_RGB_TO_UV(RGB.val[0], RGB.val[1], RGB.val[2], 2, U, V);
128 vst1_u8(u + i, vqrshrun_n_s16(U, 2));
129 vst1_u8(v + i, vqrshrun_n_s16(V, 2));
130 }
131 for (; i < width; i += 1, rgb += 4) {
132 const int r = rgb[0], g = rgb[1], b = rgb[2];
133 u[i] = VP8RGBToU(r, g, b, YUV_HALF << 2);
134 v[i] = VP8RGBToV(r, g, b, YUV_HALF << 2);
135 }
136 }
137
ConvertARGBToUV_NEON(const uint32_t * WEBP_RESTRICT argb,uint8_t * WEBP_RESTRICT u,uint8_t * WEBP_RESTRICT v,int src_width,int do_store)138 static void ConvertARGBToUV_NEON(const uint32_t* WEBP_RESTRICT argb,
139 uint8_t* WEBP_RESTRICT u,
140 uint8_t* WEBP_RESTRICT v,
141 int src_width, int do_store) {
142 int i;
143 for (i = 0; i + 16 <= src_width; i += 16, u += 8, v += 8) {
144 const uint8x16x4_t RGB = vld4q_u8((const uint8_t*)&argb[i]);
145 const uint16x8_t R = vpaddlq_u8(RGB.val[2]); // pair-wise adds
146 const uint16x8_t G = vpaddlq_u8(RGB.val[1]);
147 const uint16x8_t B = vpaddlq_u8(RGB.val[0]);
148 int16x8_t U_tmp, V_tmp;
149 CONVERT_RGB_TO_UV(R, G, B, 1, U_tmp, V_tmp);
150 {
151 const uint8x8_t U = vqrshrun_n_s16(U_tmp, 1);
152 const uint8x8_t V = vqrshrun_n_s16(V_tmp, 1);
153 if (do_store) {
154 vst1_u8(u, U);
155 vst1_u8(v, V);
156 } else {
157 const uint8x8_t prev_u = vld1_u8(u);
158 const uint8x8_t prev_v = vld1_u8(v);
159 vst1_u8(u, vrhadd_u8(U, prev_u));
160 vst1_u8(v, vrhadd_u8(V, prev_v));
161 }
162 }
163 }
164 if (i < src_width) { // left-over
165 WebPConvertARGBToUV_C(argb + i, u, v, src_width - i, do_store);
166 }
167 }
168
169
170 //------------------------------------------------------------------------------
171
172 extern void WebPInitConvertARGBToYUVNEON(void);
173
WebPInitConvertARGBToYUVNEON(void)174 WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUVNEON(void) {
175 WebPConvertRGB24ToY = ConvertRGB24ToY_NEON;
176 WebPConvertBGR24ToY = ConvertBGR24ToY_NEON;
177 WebPConvertARGBToY = ConvertARGBToY_NEON;
178 WebPConvertARGBToUV = ConvertARGBToUV_NEON;
179 WebPConvertRGBA32ToUV = ConvertRGBA32ToUV_NEON;
180 }
181
182 #else // !WEBP_USE_NEON
183
184 WEBP_DSP_INIT_STUB(WebPInitConvertARGBToYUVNEON)
185
186 #endif // WEBP_USE_NEON
187