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 * rgb,uint8_t * y,int width)49 static void ConvertRGB24ToY_NEON(const uint8_t* rgb, uint8_t* y, int width) {
50 int i;
51 for (i = 0; i + 8 <= width; i += 8, rgb += 3 * 8) {
52 const uint8x8x3_t RGB = vld3_u8(rgb);
53 const uint8x8_t Y = ConvertRGBToY_NEON(RGB.val[0], RGB.val[1], RGB.val[2]);
54 vst1_u8(y + i, Y);
55 }
56 for (; i < width; ++i, rgb += 3) { // left-over
57 y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF);
58 }
59 }
60
ConvertBGR24ToY_NEON(const uint8_t * bgr,uint8_t * y,int width)61 static void ConvertBGR24ToY_NEON(const uint8_t* bgr, uint8_t* y, int width) {
62 int i;
63 for (i = 0; i + 8 <= width; i += 8, bgr += 3 * 8) {
64 const uint8x8x3_t BGR = vld3_u8(bgr);
65 const uint8x8_t Y = ConvertRGBToY_NEON(BGR.val[2], BGR.val[1], BGR.val[0]);
66 vst1_u8(y + i, Y);
67 }
68 for (; i < width; ++i, bgr += 3) { // left-over
69 y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF);
70 }
71 }
72
ConvertARGBToY_NEON(const uint32_t * argb,uint8_t * y,int width)73 static void ConvertARGBToY_NEON(const uint32_t* argb, uint8_t* y, int width) {
74 int i;
75 for (i = 0; i + 8 <= width; i += 8) {
76 const uint8x8x4_t RGB = vld4_u8((const uint8_t*)&argb[i]);
77 const uint8x8_t Y = ConvertRGBToY_NEON(RGB.val[2], RGB.val[1], RGB.val[0]);
78 vst1_u8(y + i, Y);
79 }
80 for (; i < width; ++i) { // left-over
81 const uint32_t p = argb[i];
82 y[i] = VP8RGBToY((p >> 16) & 0xff, (p >> 8) & 0xff, (p >> 0) & 0xff,
83 YUV_HALF);
84 }
85 }
86
87 //-----------------------------------------------------------------------------
88
89 // computes: DST_s16 = [(C0 * r + C1 * g + C2 * b) >> 16] + CST
90 #define MULTIPLY_16b_PREAMBLE(r, g, b) \
91 const int16x4_t r_lo = vreinterpret_s16_u16(vget_low_u16(r)); \
92 const int16x4_t r_hi = vreinterpret_s16_u16(vget_high_u16(r)); \
93 const int16x4_t g_lo = vreinterpret_s16_u16(vget_low_u16(g)); \
94 const int16x4_t g_hi = vreinterpret_s16_u16(vget_high_u16(g)); \
95 const int16x4_t b_lo = vreinterpret_s16_u16(vget_low_u16(b)); \
96 const int16x4_t b_hi = vreinterpret_s16_u16(vget_high_u16(b))
97
98 #define MULTIPLY_16b(C0, C1, C2, CST, DST_s16) do { \
99 const int32x4_t tmp0_lo = vmull_n_s16( r_lo, C0); \
100 const int32x4_t tmp0_hi = vmull_n_s16( r_hi, C0); \
101 const int32x4_t tmp1_lo = vmlal_n_s16(tmp0_lo, g_lo, C1); \
102 const int32x4_t tmp1_hi = vmlal_n_s16(tmp0_hi, g_hi, C1); \
103 const int32x4_t tmp2_lo = vmlal_n_s16(tmp1_lo, b_lo, C2); \
104 const int32x4_t tmp2_hi = vmlal_n_s16(tmp1_hi, b_hi, C2); \
105 const int16x8_t tmp3 = vcombine_s16(vshrn_n_s32(tmp2_lo, 16), \
106 vshrn_n_s32(tmp2_hi, 16)); \
107 DST_s16 = vaddq_s16(tmp3, vdupq_n_s16(CST)); \
108 } while (0)
109
110 // This needs to be a macro, since (128 << SHIFT) needs to be an immediate.
111 #define CONVERT_RGB_TO_UV(r, g, b, SHIFT, U_DST, V_DST) do { \
112 MULTIPLY_16b_PREAMBLE(r, g, b); \
113 MULTIPLY_16b(-9719, -19081, 28800, 128 << SHIFT, U_DST); \
114 MULTIPLY_16b(28800, -24116, -4684, 128 << SHIFT, V_DST); \
115 } while (0)
116
ConvertRGBA32ToUV_NEON(const uint16_t * rgb,uint8_t * u,uint8_t * v,int width)117 static void ConvertRGBA32ToUV_NEON(const uint16_t* rgb,
118 uint8_t* u, uint8_t* v, int width) {
119 int i;
120 for (i = 0; i + 8 <= width; i += 8, rgb += 4 * 8) {
121 const uint16x8x4_t RGB = vld4q_u16((const uint16_t*)rgb);
122 int16x8_t U, V;
123 CONVERT_RGB_TO_UV(RGB.val[0], RGB.val[1], RGB.val[2], 2, U, V);
124 vst1_u8(u + i, vqrshrun_n_s16(U, 2));
125 vst1_u8(v + i, vqrshrun_n_s16(V, 2));
126 }
127 for (; i < width; i += 1, rgb += 4) {
128 const int r = rgb[0], g = rgb[1], b = rgb[2];
129 u[i] = VP8RGBToU(r, g, b, YUV_HALF << 2);
130 v[i] = VP8RGBToV(r, g, b, YUV_HALF << 2);
131 }
132 }
133
ConvertARGBToUV_NEON(const uint32_t * argb,uint8_t * u,uint8_t * v,int src_width,int do_store)134 static void ConvertARGBToUV_NEON(const uint32_t* argb, uint8_t* u, uint8_t* v,
135 int src_width, int do_store) {
136 int i;
137 for (i = 0; i + 16 <= src_width; i += 16, u += 8, v += 8) {
138 const uint8x16x4_t RGB = vld4q_u8((const uint8_t*)&argb[i]);
139 const uint16x8_t R = vpaddlq_u8(RGB.val[2]); // pair-wise adds
140 const uint16x8_t G = vpaddlq_u8(RGB.val[1]);
141 const uint16x8_t B = vpaddlq_u8(RGB.val[0]);
142 int16x8_t U_tmp, V_tmp;
143 CONVERT_RGB_TO_UV(R, G, B, 1, U_tmp, V_tmp);
144 {
145 const uint8x8_t U = vqrshrun_n_s16(U_tmp, 1);
146 const uint8x8_t V = vqrshrun_n_s16(V_tmp, 1);
147 if (do_store) {
148 vst1_u8(u, U);
149 vst1_u8(v, V);
150 } else {
151 const uint8x8_t prev_u = vld1_u8(u);
152 const uint8x8_t prev_v = vld1_u8(v);
153 vst1_u8(u, vrhadd_u8(U, prev_u));
154 vst1_u8(v, vrhadd_u8(V, prev_v));
155 }
156 }
157 }
158 if (i < src_width) { // left-over
159 WebPConvertARGBToUV_C(argb + i, u, v, src_width - i, do_store);
160 }
161 }
162
163
164 //------------------------------------------------------------------------------
165
166 extern void WebPInitConvertARGBToYUVNEON(void);
167
WebPInitConvertARGBToYUVNEON(void)168 WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUVNEON(void) {
169 WebPConvertRGB24ToY = ConvertRGB24ToY_NEON;
170 WebPConvertBGR24ToY = ConvertBGR24ToY_NEON;
171 WebPConvertARGBToY = ConvertARGBToY_NEON;
172 WebPConvertARGBToUV = ConvertARGBToUV_NEON;
173 WebPConvertRGBA32ToUV = ConvertRGBA32ToUV_NEON;
174 }
175
176 //------------------------------------------------------------------------------
177
178 #define MAX_Y ((1 << 10) - 1) // 10b precision over 16b-arithmetic
clip_y_NEON(int v)179 static uint16_t clip_y_NEON(int v) {
180 return (v < 0) ? 0 : (v > MAX_Y) ? MAX_Y : (uint16_t)v;
181 }
182
SharpYUVUpdateY_NEON(const uint16_t * ref,const uint16_t * src,uint16_t * dst,int len)183 static uint64_t SharpYUVUpdateY_NEON(const uint16_t* ref, const uint16_t* src,
184 uint16_t* dst, int len) {
185 int i;
186 const int16x8_t zero = vdupq_n_s16(0);
187 const int16x8_t max = vdupq_n_s16(MAX_Y);
188 uint64x2_t sum = vdupq_n_u64(0);
189 uint64_t diff;
190
191 for (i = 0; i + 8 <= len; i += 8) {
192 const int16x8_t A = vreinterpretq_s16_u16(vld1q_u16(ref + i));
193 const int16x8_t B = vreinterpretq_s16_u16(vld1q_u16(src + i));
194 const int16x8_t C = vreinterpretq_s16_u16(vld1q_u16(dst + i));
195 const int16x8_t D = vsubq_s16(A, B); // diff_y
196 const int16x8_t F = vaddq_s16(C, D); // new_y
197 const uint16x8_t H =
198 vreinterpretq_u16_s16(vmaxq_s16(vminq_s16(F, max), zero));
199 const int16x8_t I = vabsq_s16(D); // abs(diff_y)
200 vst1q_u16(dst + i, H);
201 sum = vpadalq_u32(sum, vpaddlq_u16(vreinterpretq_u16_s16(I)));
202 }
203 diff = vgetq_lane_u64(sum, 0) + vgetq_lane_u64(sum, 1);
204 for (; i < len; ++i) {
205 const int diff_y = ref[i] - src[i];
206 const int new_y = (int)(dst[i]) + diff_y;
207 dst[i] = clip_y_NEON(new_y);
208 diff += (uint64_t)(abs(diff_y));
209 }
210 return diff;
211 }
212
SharpYUVUpdateRGB_NEON(const int16_t * ref,const int16_t * src,int16_t * dst,int len)213 static void SharpYUVUpdateRGB_NEON(const int16_t* ref, const int16_t* src,
214 int16_t* dst, int len) {
215 int i;
216 for (i = 0; i + 8 <= len; i += 8) {
217 const int16x8_t A = vld1q_s16(ref + i);
218 const int16x8_t B = vld1q_s16(src + i);
219 const int16x8_t C = vld1q_s16(dst + i);
220 const int16x8_t D = vsubq_s16(A, B); // diff_uv
221 const int16x8_t E = vaddq_s16(C, D); // new_uv
222 vst1q_s16(dst + i, E);
223 }
224 for (; i < len; ++i) {
225 const int diff_uv = ref[i] - src[i];
226 dst[i] += diff_uv;
227 }
228 }
229
SharpYUVFilterRow_NEON(const int16_t * A,const int16_t * B,int len,const uint16_t * best_y,uint16_t * out)230 static void SharpYUVFilterRow_NEON(const int16_t* A, const int16_t* B, int len,
231 const uint16_t* best_y, uint16_t* out) {
232 int i;
233 const int16x8_t max = vdupq_n_s16(MAX_Y);
234 const int16x8_t zero = vdupq_n_s16(0);
235 for (i = 0; i + 8 <= len; i += 8) {
236 const int16x8_t a0 = vld1q_s16(A + i + 0);
237 const int16x8_t a1 = vld1q_s16(A + i + 1);
238 const int16x8_t b0 = vld1q_s16(B + i + 0);
239 const int16x8_t b1 = vld1q_s16(B + i + 1);
240 const int16x8_t a0b1 = vaddq_s16(a0, b1);
241 const int16x8_t a1b0 = vaddq_s16(a1, b0);
242 const int16x8_t a0a1b0b1 = vaddq_s16(a0b1, a1b0); // A0+A1+B0+B1
243 const int16x8_t a0b1_2 = vaddq_s16(a0b1, a0b1); // 2*(A0+B1)
244 const int16x8_t a1b0_2 = vaddq_s16(a1b0, a1b0); // 2*(A1+B0)
245 const int16x8_t c0 = vshrq_n_s16(vaddq_s16(a0b1_2, a0a1b0b1), 3);
246 const int16x8_t c1 = vshrq_n_s16(vaddq_s16(a1b0_2, a0a1b0b1), 3);
247 const int16x8_t d0 = vaddq_s16(c1, a0);
248 const int16x8_t d1 = vaddq_s16(c0, a1);
249 const int16x8_t e0 = vrshrq_n_s16(d0, 1);
250 const int16x8_t e1 = vrshrq_n_s16(d1, 1);
251 const int16x8x2_t f = vzipq_s16(e0, e1);
252 const int16x8_t g0 = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i + 0));
253 const int16x8_t g1 = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i + 8));
254 const int16x8_t h0 = vaddq_s16(g0, f.val[0]);
255 const int16x8_t h1 = vaddq_s16(g1, f.val[1]);
256 const int16x8_t i0 = vmaxq_s16(vminq_s16(h0, max), zero);
257 const int16x8_t i1 = vmaxq_s16(vminq_s16(h1, max), zero);
258 vst1q_u16(out + 2 * i + 0, vreinterpretq_u16_s16(i0));
259 vst1q_u16(out + 2 * i + 8, vreinterpretq_u16_s16(i1));
260 }
261 for (; i < len; ++i) {
262 const int a0b1 = A[i + 0] + B[i + 1];
263 const int a1b0 = A[i + 1] + B[i + 0];
264 const int a0a1b0b1 = a0b1 + a1b0 + 8;
265 const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4;
266 const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4;
267 out[2 * i + 0] = clip_y_NEON(best_y[2 * i + 0] + v0);
268 out[2 * i + 1] = clip_y_NEON(best_y[2 * i + 1] + v1);
269 }
270 }
271 #undef MAX_Y
272
273 //------------------------------------------------------------------------------
274
275 extern void WebPInitSharpYUVNEON(void);
276
WebPInitSharpYUVNEON(void)277 WEBP_TSAN_IGNORE_FUNCTION void WebPInitSharpYUVNEON(void) {
278 WebPSharpYUVUpdateY = SharpYUVUpdateY_NEON;
279 WebPSharpYUVUpdateRGB = SharpYUVUpdateRGB_NEON;
280 WebPSharpYUVFilterRow = SharpYUVFilterRow_NEON;
281 }
282
283 #else // !WEBP_USE_NEON
284
285 WEBP_DSP_INIT_STUB(WebPInitConvertARGBToYUVNEON)
286 WEBP_DSP_INIT_STUB(WebPInitSharpYUVNEON)
287
288 #endif // WEBP_USE_NEON
289