1// Copyright 2022 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// Speed-critical functions for Sharp YUV.
11//
12// Author: Skal (pascal.massimino@gmail.com)
13
14#include "sharpyuv/sharpyuv_dsp.h"
15
16#if defined(WEBP_USE_NEON)
17#include <assert.h>
18#include <stdlib.h>
19#include <arm_neon.h>
20
21static uint16_t clip_NEON(int v, int max) {
22 return (v < 0) ? 0 : (v > max) ? max : (uint16_t)v;
23}
24
25static uint64_t SharpYuvUpdateY_NEON(const uint16_t* ref, const uint16_t* src,
26 uint16_t* dst, int len, int bit_depth) {
27 const int max_y = (1 << bit_depth) - 1;
28 int i;
29 const int16x8_t zero = vdupq_n_s16(0);
30 const int16x8_t max = vdupq_n_s16(max_y);
31 uint64x2_t sum = vdupq_n_u64(0);
32 uint64_t diff;
33
34 for (i = 0; i + 8 <= len; i += 8) {
35 const int16x8_t A = vreinterpretq_s16_u16(vld1q_u16(ref + i));
36 const int16x8_t B = vreinterpretq_s16_u16(vld1q_u16(src + i));
37 const int16x8_t C = vreinterpretq_s16_u16(vld1q_u16(dst + i));
38 const int16x8_t D = vsubq_s16(A, B); // diff_y
39 const int16x8_t F = vaddq_s16(C, D); // new_y
40 const uint16x8_t H =
41 vreinterpretq_u16_s16(vmaxq_s16(vminq_s16(F, max), zero));
42 const int16x8_t I = vabsq_s16(D); // abs(diff_y)
43 vst1q_u16(dst + i, H);
44 sum = vpadalq_u32(sum, vpaddlq_u16(vreinterpretq_u16_s16(I)));
45 }
46 diff = vgetq_lane_u64(sum, 0) + vgetq_lane_u64(sum, 1);
47 for (; i < len; ++i) {
48 const int diff_y = ref[i] - src[i];
49 const int new_y = (int)(dst[i]) + diff_y;
50 dst[i] = clip_NEON(new_y, max_y);
51 diff += (uint64_t)(abs(diff_y));
52 }
53 return diff;
54}
55
56static void SharpYuvUpdateRGB_NEON(const int16_t* ref, const int16_t* src,
57 int16_t* dst, int len) {
58 int i;
59 for (i = 0; i + 8 <= len; i += 8) {
60 const int16x8_t A = vld1q_s16(ref + i);
61 const int16x8_t B = vld1q_s16(src + i);
62 const int16x8_t C = vld1q_s16(dst + i);
63 const int16x8_t D = vsubq_s16(A, B); // diff_uv
64 const int16x8_t E = vaddq_s16(C, D); // new_uv
65 vst1q_s16(dst + i, E);
66 }
67 for (; i < len; ++i) {
68 const int diff_uv = ref[i] - src[i];
69 dst[i] += diff_uv;
70 }
71}
72
73static void SharpYuvFilterRow16_NEON(const int16_t* A, const int16_t* B,
74 int len, const uint16_t* best_y,
75 uint16_t* out, int bit_depth) {
76 const int max_y = (1 << bit_depth) - 1;
77 int i;
78 const int16x8_t max = vdupq_n_s16(max_y);
79 const int16x8_t zero = vdupq_n_s16(0);
80 for (i = 0; i + 8 <= len; i += 8) {
81 const int16x8_t a0 = vld1q_s16(A + i + 0);
82 const int16x8_t a1 = vld1q_s16(A + i + 1);
83 const int16x8_t b0 = vld1q_s16(B + i + 0);
84 const int16x8_t b1 = vld1q_s16(B + i + 1);
85 const int16x8_t a0b1 = vaddq_s16(a0, b1);
86 const int16x8_t a1b0 = vaddq_s16(a1, b0);
87 const int16x8_t a0a1b0b1 = vaddq_s16(a0b1, a1b0); // A0+A1+B0+B1
88 const int16x8_t a0b1_2 = vaddq_s16(a0b1, a0b1); // 2*(A0+B1)
89 const int16x8_t a1b0_2 = vaddq_s16(a1b0, a1b0); // 2*(A1+B0)
90 const int16x8_t c0 = vshrq_n_s16(vaddq_s16(a0b1_2, a0a1b0b1), 3);
91 const int16x8_t c1 = vshrq_n_s16(vaddq_s16(a1b0_2, a0a1b0b1), 3);
92 const int16x8_t e0 = vrhaddq_s16(c1, a0);
93 const int16x8_t e1 = vrhaddq_s16(c0, a1);
94 const int16x8x2_t f = vzipq_s16(e0, e1);
95 const int16x8_t g0 = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i + 0));
96 const int16x8_t g1 = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i + 8));
97 const int16x8_t h0 = vaddq_s16(g0, f.val[0]);
98 const int16x8_t h1 = vaddq_s16(g1, f.val[1]);
99 const int16x8_t i0 = vmaxq_s16(vminq_s16(h0, max), zero);
100 const int16x8_t i1 = vmaxq_s16(vminq_s16(h1, max), zero);
101 vst1q_u16(out + 2 * i + 0, vreinterpretq_u16_s16(i0));
102 vst1q_u16(out + 2 * i + 8, vreinterpretq_u16_s16(i1));
103 }
104 for (; i < len; ++i) {
105 const int a0b1 = A[i + 0] + B[i + 1];
106 const int a1b0 = A[i + 1] + B[i + 0];
107 const int a0a1b0b1 = a0b1 + a1b0 + 8;
108 const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4;
109 const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4;
110 out[2 * i + 0] = clip_NEON(best_y[2 * i + 0] + v0, max_y);
111 out[2 * i + 1] = clip_NEON(best_y[2 * i + 1] + v1, max_y);
112 }
113}
114
115static void SharpYuvFilterRow32_NEON(const int16_t* A, const int16_t* B,
116 int len, const uint16_t* best_y,
117 uint16_t* out, int bit_depth) {
118 const int max_y = (1 << bit_depth) - 1;
119 int i;
120 const uint16x8_t max = vdupq_n_u16(max_y);
121 for (i = 0; i + 4 <= len; i += 4) {
122 const int16x4_t a0 = vld1_s16(A + i + 0);
123 const int16x4_t a1 = vld1_s16(A + i + 1);
124 const int16x4_t b0 = vld1_s16(B + i + 0);
125 const int16x4_t b1 = vld1_s16(B + i + 1);
126 const int32x4_t a0b1 = vaddl_s16(a0, b1);
127 const int32x4_t a1b0 = vaddl_s16(a1, b0);
128 const int32x4_t a0a1b0b1 = vaddq_s32(a0b1, a1b0); // A0+A1+B0+B1
129 const int32x4_t a0b1_2 = vaddq_s32(a0b1, a0b1); // 2*(A0+B1)
130 const int32x4_t a1b0_2 = vaddq_s32(a1b0, a1b0); // 2*(A1+B0)
131 const int32x4_t c0 = vshrq_n_s32(vaddq_s32(a0b1_2, a0a1b0b1), 3);
132 const int32x4_t c1 = vshrq_n_s32(vaddq_s32(a1b0_2, a0a1b0b1), 3);
133 const int32x4_t e0 = vrhaddq_s32(c1, vmovl_s16(a0));
134 const int32x4_t e1 = vrhaddq_s32(c0, vmovl_s16(a1));
135 const int32x4x2_t f = vzipq_s32(e0, e1);
136
137 const int16x8_t g = vreinterpretq_s16_u16(vld1q_u16(best_y + 2 * i));
138 const int32x4_t h0 = vaddw_s16(f.val[0], vget_low_s16(g));
139 const int32x4_t h1 = vaddw_s16(f.val[1], vget_high_s16(g));
140 const uint16x8_t i_16 = vcombine_u16(vqmovun_s32(h0), vqmovun_s32(h1));
141 const uint16x8_t i_clamped = vminq_u16(i_16, max);
142 vst1q_u16(out + 2 * i + 0, i_clamped);
143 }
144 for (; i < len; ++i) {
145 const int a0b1 = A[i + 0] + B[i + 1];
146 const int a1b0 = A[i + 1] + B[i + 0];
147 const int a0a1b0b1 = a0b1 + a1b0 + 8;
148 const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4;
149 const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4;
150 out[2 * i + 0] = clip_NEON(best_y[2 * i + 0] + v0, max_y);
151 out[2 * i + 1] = clip_NEON(best_y[2 * i + 1] + v1, max_y);
152 }
153}
154
155static void SharpYuvFilterRow_NEON(const int16_t* A, const int16_t* B, int len,
156 const uint16_t* best_y, uint16_t* out,
157 int bit_depth) {
158 if (bit_depth <= 10) {
159 SharpYuvFilterRow16_NEON(A, B, len, best_y, out, bit_depth);
160 } else {
161 SharpYuvFilterRow32_NEON(A, B, len, best_y, out, bit_depth);
162 }
163}
164
165//------------------------------------------------------------------------------
166
167extern void InitSharpYuvNEON(void);
168
169WEBP_TSAN_IGNORE_FUNCTION void InitSharpYuvNEON(void) {
170 SharpYuvUpdateY = SharpYuvUpdateY_NEON;
171 SharpYuvUpdateRGB = SharpYuvUpdateRGB_NEON;
172 SharpYuvFilterRow = SharpYuvFilterRow_NEON;
173}
174
175#else // !WEBP_USE_NEON
176
177extern void InitSharpYuvNEON(void);
178
179void InitSharpYuvNEON(void) {}
180
181#endif // WEBP_USE_NEON
182