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_SSE2)
17 #include <stdlib.h>
18 #include <emmintrin.h>
19
clip_SSE2(int v,int max)20 static uint16_t clip_SSE2(int v, int max) {
21 return (v < 0) ? 0 : (v > max) ? max : (uint16_t)v;
22 }
23
SharpYuvUpdateY_SSE2(const uint16_t * ref,const uint16_t * src,uint16_t * dst,int len,int bit_depth)24 static uint64_t SharpYuvUpdateY_SSE2(const uint16_t* ref, const uint16_t* src,
25 uint16_t* dst, int len, int bit_depth) {
26 const int max_y = (1 << bit_depth) - 1;
27 uint64_t diff = 0;
28 uint32_t tmp[4];
29 int i;
30 const __m128i zero = _mm_setzero_si128();
31 const __m128i max = _mm_set1_epi16(max_y);
32 const __m128i one = _mm_set1_epi16(1);
33 __m128i sum = zero;
34
35 for (i = 0; i + 8 <= len; i += 8) {
36 const __m128i A = _mm_loadu_si128((const __m128i*)(ref + i));
37 const __m128i B = _mm_loadu_si128((const __m128i*)(src + i));
38 const __m128i C = _mm_loadu_si128((const __m128i*)(dst + i));
39 const __m128i D = _mm_sub_epi16(A, B); // diff_y
40 const __m128i E = _mm_cmpgt_epi16(zero, D); // sign (-1 or 0)
41 const __m128i F = _mm_add_epi16(C, D); // new_y
42 const __m128i G = _mm_or_si128(E, one); // -1 or 1
43 const __m128i H = _mm_max_epi16(_mm_min_epi16(F, max), zero);
44 const __m128i I = _mm_madd_epi16(D, G); // sum(abs(...))
45 _mm_storeu_si128((__m128i*)(dst + i), H);
46 sum = _mm_add_epi32(sum, I);
47 }
48 _mm_storeu_si128((__m128i*)tmp, sum);
49 diff = tmp[3] + tmp[2] + tmp[1] + tmp[0];
50 for (; i < len; ++i) {
51 const int diff_y = ref[i] - src[i];
52 const int new_y = (int)dst[i] + diff_y;
53 dst[i] = clip_SSE2(new_y, max_y);
54 diff += (uint64_t)abs(diff_y);
55 }
56 return diff;
57 }
58
SharpYuvUpdateRGB_SSE2(const int16_t * ref,const int16_t * src,int16_t * dst,int len)59 static void SharpYuvUpdateRGB_SSE2(const int16_t* ref, const int16_t* src,
60 int16_t* dst, int len) {
61 int i = 0;
62 for (i = 0; i + 8 <= len; i += 8) {
63 const __m128i A = _mm_loadu_si128((const __m128i*)(ref + i));
64 const __m128i B = _mm_loadu_si128((const __m128i*)(src + i));
65 const __m128i C = _mm_loadu_si128((const __m128i*)(dst + i));
66 const __m128i D = _mm_sub_epi16(A, B); // diff_uv
67 const __m128i E = _mm_add_epi16(C, D); // new_uv
68 _mm_storeu_si128((__m128i*)(dst + i), E);
69 }
70 for (; i < len; ++i) {
71 const int diff_uv = ref[i] - src[i];
72 dst[i] += diff_uv;
73 }
74 }
75
SharpYuvFilterRow16_SSE2(const int16_t * A,const int16_t * B,int len,const uint16_t * best_y,uint16_t * out,int bit_depth)76 static void SharpYuvFilterRow16_SSE2(const int16_t* A, const int16_t* B,
77 int len, const uint16_t* best_y,
78 uint16_t* out, int bit_depth) {
79 const int max_y = (1 << bit_depth) - 1;
80 int i;
81 const __m128i kCst8 = _mm_set1_epi16(8);
82 const __m128i max = _mm_set1_epi16(max_y);
83 const __m128i zero = _mm_setzero_si128();
84 for (i = 0; i + 8 <= len; i += 8) {
85 const __m128i a0 = _mm_loadu_si128((const __m128i*)(A + i + 0));
86 const __m128i a1 = _mm_loadu_si128((const __m128i*)(A + i + 1));
87 const __m128i b0 = _mm_loadu_si128((const __m128i*)(B + i + 0));
88 const __m128i b1 = _mm_loadu_si128((const __m128i*)(B + i + 1));
89 const __m128i a0b1 = _mm_add_epi16(a0, b1);
90 const __m128i a1b0 = _mm_add_epi16(a1, b0);
91 const __m128i a0a1b0b1 = _mm_add_epi16(a0b1, a1b0); // A0+A1+B0+B1
92 const __m128i a0a1b0b1_8 = _mm_add_epi16(a0a1b0b1, kCst8);
93 const __m128i a0b1_2 = _mm_add_epi16(a0b1, a0b1); // 2*(A0+B1)
94 const __m128i a1b0_2 = _mm_add_epi16(a1b0, a1b0); // 2*(A1+B0)
95 const __m128i c0 = _mm_srai_epi16(_mm_add_epi16(a0b1_2, a0a1b0b1_8), 3);
96 const __m128i c1 = _mm_srai_epi16(_mm_add_epi16(a1b0_2, a0a1b0b1_8), 3);
97 const __m128i d0 = _mm_add_epi16(c1, a0);
98 const __m128i d1 = _mm_add_epi16(c0, a1);
99 const __m128i e0 = _mm_srai_epi16(d0, 1);
100 const __m128i e1 = _mm_srai_epi16(d1, 1);
101 const __m128i f0 = _mm_unpacklo_epi16(e0, e1);
102 const __m128i f1 = _mm_unpackhi_epi16(e0, e1);
103 const __m128i g0 = _mm_loadu_si128((const __m128i*)(best_y + 2 * i + 0));
104 const __m128i g1 = _mm_loadu_si128((const __m128i*)(best_y + 2 * i + 8));
105 const __m128i h0 = _mm_add_epi16(g0, f0);
106 const __m128i h1 = _mm_add_epi16(g1, f1);
107 const __m128i i0 = _mm_max_epi16(_mm_min_epi16(h0, max), zero);
108 const __m128i i1 = _mm_max_epi16(_mm_min_epi16(h1, max), zero);
109 _mm_storeu_si128((__m128i*)(out + 2 * i + 0), i0);
110 _mm_storeu_si128((__m128i*)(out + 2 * i + 8), i1);
111 }
112 for (; i < len; ++i) {
113 // (9 * A0 + 3 * A1 + 3 * B0 + B1 + 8) >> 4 =
114 // = (8 * A0 + 2 * (A1 + B0) + (A0 + A1 + B0 + B1 + 8)) >> 4
115 // We reuse the common sub-expressions.
116 const int a0b1 = A[i + 0] + B[i + 1];
117 const int a1b0 = A[i + 1] + B[i + 0];
118 const int a0a1b0b1 = a0b1 + a1b0 + 8;
119 const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4;
120 const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4;
121 out[2 * i + 0] = clip_SSE2(best_y[2 * i + 0] + v0, max_y);
122 out[2 * i + 1] = clip_SSE2(best_y[2 * i + 1] + v1, max_y);
123 }
124 }
125
s16_to_s32(__m128i in)126 static WEBP_INLINE __m128i s16_to_s32(__m128i in) {
127 return _mm_srai_epi32(_mm_unpacklo_epi16(in, in), 16);
128 }
129
SharpYuvFilterRow32_SSE2(const int16_t * A,const int16_t * B,int len,const uint16_t * best_y,uint16_t * out,int bit_depth)130 static void SharpYuvFilterRow32_SSE2(const int16_t* A, const int16_t* B,
131 int len, const uint16_t* best_y,
132 uint16_t* out, int bit_depth) {
133 const int max_y = (1 << bit_depth) - 1;
134 int i;
135 const __m128i kCst8 = _mm_set1_epi32(8);
136 const __m128i max = _mm_set1_epi16(max_y);
137 const __m128i zero = _mm_setzero_si128();
138 for (i = 0; i + 4 <= len; i += 4) {
139 const __m128i a0 = s16_to_s32(_mm_loadl_epi64((const __m128i*)(A + i + 0)));
140 const __m128i a1 = s16_to_s32(_mm_loadl_epi64((const __m128i*)(A + i + 1)));
141 const __m128i b0 = s16_to_s32(_mm_loadl_epi64((const __m128i*)(B + i + 0)));
142 const __m128i b1 = s16_to_s32(_mm_loadl_epi64((const __m128i*)(B + i + 1)));
143 const __m128i a0b1 = _mm_add_epi32(a0, b1);
144 const __m128i a1b0 = _mm_add_epi32(a1, b0);
145 const __m128i a0a1b0b1 = _mm_add_epi32(a0b1, a1b0); // A0+A1+B0+B1
146 const __m128i a0a1b0b1_8 = _mm_add_epi32(a0a1b0b1, kCst8);
147 const __m128i a0b1_2 = _mm_add_epi32(a0b1, a0b1); // 2*(A0+B1)
148 const __m128i a1b0_2 = _mm_add_epi32(a1b0, a1b0); // 2*(A1+B0)
149 const __m128i c0 = _mm_srai_epi32(_mm_add_epi32(a0b1_2, a0a1b0b1_8), 3);
150 const __m128i c1 = _mm_srai_epi32(_mm_add_epi32(a1b0_2, a0a1b0b1_8), 3);
151 const __m128i d0 = _mm_add_epi32(c1, a0);
152 const __m128i d1 = _mm_add_epi32(c0, a1);
153 const __m128i e0 = _mm_srai_epi32(d0, 1);
154 const __m128i e1 = _mm_srai_epi32(d1, 1);
155 const __m128i f0 = _mm_unpacklo_epi32(e0, e1);
156 const __m128i f1 = _mm_unpackhi_epi32(e0, e1);
157 const __m128i g = _mm_loadu_si128((const __m128i*)(best_y + 2 * i + 0));
158 const __m128i h_16 = _mm_add_epi16(g, _mm_packs_epi32(f0, f1));
159 const __m128i final = _mm_max_epi16(_mm_min_epi16(h_16, max), zero);
160 _mm_storeu_si128((__m128i*)(out + 2 * i + 0), final);
161 }
162 for (; i < len; ++i) {
163 // (9 * A0 + 3 * A1 + 3 * B0 + B1 + 8) >> 4 =
164 // = (8 * A0 + 2 * (A1 + B0) + (A0 + A1 + B0 + B1 + 8)) >> 4
165 // We reuse the common sub-expressions.
166 const int a0b1 = A[i + 0] + B[i + 1];
167 const int a1b0 = A[i + 1] + B[i + 0];
168 const int a0a1b0b1 = a0b1 + a1b0 + 8;
169 const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4;
170 const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4;
171 out[2 * i + 0] = clip_SSE2(best_y[2 * i + 0] + v0, max_y);
172 out[2 * i + 1] = clip_SSE2(best_y[2 * i + 1] + v1, max_y);
173 }
174 }
175
SharpYuvFilterRow_SSE2(const int16_t * A,const int16_t * B,int len,const uint16_t * best_y,uint16_t * out,int bit_depth)176 static void SharpYuvFilterRow_SSE2(const int16_t* A, const int16_t* B, int len,
177 const uint16_t* best_y, uint16_t* out,
178 int bit_depth) {
179 if (bit_depth <= 10) {
180 SharpYuvFilterRow16_SSE2(A, B, len, best_y, out, bit_depth);
181 } else {
182 SharpYuvFilterRow32_SSE2(A, B, len, best_y, out, bit_depth);
183 }
184 }
185
186 //------------------------------------------------------------------------------
187
188 extern void InitSharpYuvSSE2(void);
189
InitSharpYuvSSE2(void)190 WEBP_TSAN_IGNORE_FUNCTION void InitSharpYuvSSE2(void) {
191 SharpYuvUpdateY = SharpYuvUpdateY_SSE2;
192 SharpYuvUpdateRGB = SharpYuvUpdateRGB_SSE2;
193 SharpYuvFilterRow = SharpYuvFilterRow_SSE2;
194 }
195 #else // !WEBP_USE_SSE2
196
197 extern void InitSharpYuvSSE2(void);
198
InitSharpYuvSSE2(void)199 void InitSharpYuvSSE2(void) {}
200
201 #endif // WEBP_USE_SSE2
202