Update mojo sdk to rev 1dc8a9a5db73d3718d99917fadf31f5fb2ebad4f
[chromium-blink-merge.git] / third_party / libwebp / dsp / dec_sse2.c
blobc37a637f56425e55f9f43182998860ef9f0e3cb9
1 // Copyright 2011 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 // SSE2 version of some decoding functions (idct, loop filtering).
12 // Author: somnath@google.com (Somnath Banerjee)
13 // cduvivier@google.com (Christian Duvivier)
15 #include "./dsp.h"
17 #if defined(WEBP_USE_SSE2)
19 // The 3-coeff sparse transform in SSE2 is not really faster than the plain-C
20 // one it seems => disable it by default. Uncomment the following to enable:
21 // #define USE_TRANSFORM_AC3
23 #include <emmintrin.h>
24 #include "../dec/vp8i.h"
26 //------------------------------------------------------------------------------
27 // Transforms (Paragraph 14.4)
29 static void Transform(const int16_t* in, uint8_t* dst, int do_two) {
30 // This implementation makes use of 16-bit fixed point versions of two
31 // multiply constants:
32 // K1 = sqrt(2) * cos (pi/8) ~= 85627 / 2^16
33 // K2 = sqrt(2) * sin (pi/8) ~= 35468 / 2^16
35 // To be able to use signed 16-bit integers, we use the following trick to
36 // have constants within range:
37 // - Associated constants are obtained by subtracting the 16-bit fixed point
38 // version of one:
39 // k = K - (1 << 16) => K = k + (1 << 16)
40 // K1 = 85267 => k1 = 20091
41 // K2 = 35468 => k2 = -30068
42 // - The multiplication of a variable by a constant become the sum of the
43 // variable and the multiplication of that variable by the associated
44 // constant:
45 // (x * K) >> 16 = (x * (k + (1 << 16))) >> 16 = ((x * k ) >> 16) + x
46 const __m128i k1 = _mm_set1_epi16(20091);
47 const __m128i k2 = _mm_set1_epi16(-30068);
48 __m128i T0, T1, T2, T3;
50 // Load and concatenate the transform coefficients (we'll do two transforms
51 // in parallel). In the case of only one transform, the second half of the
52 // vectors will just contain random value we'll never use nor store.
53 __m128i in0, in1, in2, in3;
55 in0 = _mm_loadl_epi64((__m128i*)&in[0]);
56 in1 = _mm_loadl_epi64((__m128i*)&in[4]);
57 in2 = _mm_loadl_epi64((__m128i*)&in[8]);
58 in3 = _mm_loadl_epi64((__m128i*)&in[12]);
59 // a00 a10 a20 a30 x x x x
60 // a01 a11 a21 a31 x x x x
61 // a02 a12 a22 a32 x x x x
62 // a03 a13 a23 a33 x x x x
63 if (do_two) {
64 const __m128i inB0 = _mm_loadl_epi64((__m128i*)&in[16]);
65 const __m128i inB1 = _mm_loadl_epi64((__m128i*)&in[20]);
66 const __m128i inB2 = _mm_loadl_epi64((__m128i*)&in[24]);
67 const __m128i inB3 = _mm_loadl_epi64((__m128i*)&in[28]);
68 in0 = _mm_unpacklo_epi64(in0, inB0);
69 in1 = _mm_unpacklo_epi64(in1, inB1);
70 in2 = _mm_unpacklo_epi64(in2, inB2);
71 in3 = _mm_unpacklo_epi64(in3, inB3);
72 // a00 a10 a20 a30 b00 b10 b20 b30
73 // a01 a11 a21 a31 b01 b11 b21 b31
74 // a02 a12 a22 a32 b02 b12 b22 b32
75 // a03 a13 a23 a33 b03 b13 b23 b33
79 // Vertical pass and subsequent transpose.
81 // First pass, c and d calculations are longer because of the "trick"
82 // multiplications.
83 const __m128i a = _mm_add_epi16(in0, in2);
84 const __m128i b = _mm_sub_epi16(in0, in2);
85 // c = MUL(in1, K2) - MUL(in3, K1) = MUL(in1, k2) - MUL(in3, k1) + in1 - in3
86 const __m128i c1 = _mm_mulhi_epi16(in1, k2);
87 const __m128i c2 = _mm_mulhi_epi16(in3, k1);
88 const __m128i c3 = _mm_sub_epi16(in1, in3);
89 const __m128i c4 = _mm_sub_epi16(c1, c2);
90 const __m128i c = _mm_add_epi16(c3, c4);
91 // d = MUL(in1, K1) + MUL(in3, K2) = MUL(in1, k1) + MUL(in3, k2) + in1 + in3
92 const __m128i d1 = _mm_mulhi_epi16(in1, k1);
93 const __m128i d2 = _mm_mulhi_epi16(in3, k2);
94 const __m128i d3 = _mm_add_epi16(in1, in3);
95 const __m128i d4 = _mm_add_epi16(d1, d2);
96 const __m128i d = _mm_add_epi16(d3, d4);
98 // Second pass.
99 const __m128i tmp0 = _mm_add_epi16(a, d);
100 const __m128i tmp1 = _mm_add_epi16(b, c);
101 const __m128i tmp2 = _mm_sub_epi16(b, c);
102 const __m128i tmp3 = _mm_sub_epi16(a, d);
104 // Transpose the two 4x4.
105 // a00 a01 a02 a03 b00 b01 b02 b03
106 // a10 a11 a12 a13 b10 b11 b12 b13
107 // a20 a21 a22 a23 b20 b21 b22 b23
108 // a30 a31 a32 a33 b30 b31 b32 b33
109 const __m128i transpose0_0 = _mm_unpacklo_epi16(tmp0, tmp1);
110 const __m128i transpose0_1 = _mm_unpacklo_epi16(tmp2, tmp3);
111 const __m128i transpose0_2 = _mm_unpackhi_epi16(tmp0, tmp1);
112 const __m128i transpose0_3 = _mm_unpackhi_epi16(tmp2, tmp3);
113 // a00 a10 a01 a11 a02 a12 a03 a13
114 // a20 a30 a21 a31 a22 a32 a23 a33
115 // b00 b10 b01 b11 b02 b12 b03 b13
116 // b20 b30 b21 b31 b22 b32 b23 b33
117 const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
118 const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
119 const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
120 const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
121 // a00 a10 a20 a30 a01 a11 a21 a31
122 // b00 b10 b20 b30 b01 b11 b21 b31
123 // a02 a12 a22 a32 a03 a13 a23 a33
124 // b02 b12 a22 b32 b03 b13 b23 b33
125 T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
126 T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
127 T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
128 T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
129 // a00 a10 a20 a30 b00 b10 b20 b30
130 // a01 a11 a21 a31 b01 b11 b21 b31
131 // a02 a12 a22 a32 b02 b12 b22 b32
132 // a03 a13 a23 a33 b03 b13 b23 b33
135 // Horizontal pass and subsequent transpose.
137 // First pass, c and d calculations are longer because of the "trick"
138 // multiplications.
139 const __m128i four = _mm_set1_epi16(4);
140 const __m128i dc = _mm_add_epi16(T0, four);
141 const __m128i a = _mm_add_epi16(dc, T2);
142 const __m128i b = _mm_sub_epi16(dc, T2);
143 // c = MUL(T1, K2) - MUL(T3, K1) = MUL(T1, k2) - MUL(T3, k1) + T1 - T3
144 const __m128i c1 = _mm_mulhi_epi16(T1, k2);
145 const __m128i c2 = _mm_mulhi_epi16(T3, k1);
146 const __m128i c3 = _mm_sub_epi16(T1, T3);
147 const __m128i c4 = _mm_sub_epi16(c1, c2);
148 const __m128i c = _mm_add_epi16(c3, c4);
149 // d = MUL(T1, K1) + MUL(T3, K2) = MUL(T1, k1) + MUL(T3, k2) + T1 + T3
150 const __m128i d1 = _mm_mulhi_epi16(T1, k1);
151 const __m128i d2 = _mm_mulhi_epi16(T3, k2);
152 const __m128i d3 = _mm_add_epi16(T1, T3);
153 const __m128i d4 = _mm_add_epi16(d1, d2);
154 const __m128i d = _mm_add_epi16(d3, d4);
156 // Second pass.
157 const __m128i tmp0 = _mm_add_epi16(a, d);
158 const __m128i tmp1 = _mm_add_epi16(b, c);
159 const __m128i tmp2 = _mm_sub_epi16(b, c);
160 const __m128i tmp3 = _mm_sub_epi16(a, d);
161 const __m128i shifted0 = _mm_srai_epi16(tmp0, 3);
162 const __m128i shifted1 = _mm_srai_epi16(tmp1, 3);
163 const __m128i shifted2 = _mm_srai_epi16(tmp2, 3);
164 const __m128i shifted3 = _mm_srai_epi16(tmp3, 3);
166 // Transpose the two 4x4.
167 // a00 a01 a02 a03 b00 b01 b02 b03
168 // a10 a11 a12 a13 b10 b11 b12 b13
169 // a20 a21 a22 a23 b20 b21 b22 b23
170 // a30 a31 a32 a33 b30 b31 b32 b33
171 const __m128i transpose0_0 = _mm_unpacklo_epi16(shifted0, shifted1);
172 const __m128i transpose0_1 = _mm_unpacklo_epi16(shifted2, shifted3);
173 const __m128i transpose0_2 = _mm_unpackhi_epi16(shifted0, shifted1);
174 const __m128i transpose0_3 = _mm_unpackhi_epi16(shifted2, shifted3);
175 // a00 a10 a01 a11 a02 a12 a03 a13
176 // a20 a30 a21 a31 a22 a32 a23 a33
177 // b00 b10 b01 b11 b02 b12 b03 b13
178 // b20 b30 b21 b31 b22 b32 b23 b33
179 const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
180 const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
181 const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
182 const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
183 // a00 a10 a20 a30 a01 a11 a21 a31
184 // b00 b10 b20 b30 b01 b11 b21 b31
185 // a02 a12 a22 a32 a03 a13 a23 a33
186 // b02 b12 a22 b32 b03 b13 b23 b33
187 T0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
188 T1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
189 T2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
190 T3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
191 // a00 a10 a20 a30 b00 b10 b20 b30
192 // a01 a11 a21 a31 b01 b11 b21 b31
193 // a02 a12 a22 a32 b02 b12 b22 b32
194 // a03 a13 a23 a33 b03 b13 b23 b33
197 // Add inverse transform to 'dst' and store.
199 const __m128i zero = _mm_setzero_si128();
200 // Load the reference(s).
201 __m128i dst0, dst1, dst2, dst3;
202 if (do_two) {
203 // Load eight bytes/pixels per line.
204 dst0 = _mm_loadl_epi64((__m128i*)(dst + 0 * BPS));
205 dst1 = _mm_loadl_epi64((__m128i*)(dst + 1 * BPS));
206 dst2 = _mm_loadl_epi64((__m128i*)(dst + 2 * BPS));
207 dst3 = _mm_loadl_epi64((__m128i*)(dst + 3 * BPS));
208 } else {
209 // Load four bytes/pixels per line.
210 dst0 = _mm_cvtsi32_si128(*(int*)(dst + 0 * BPS));
211 dst1 = _mm_cvtsi32_si128(*(int*)(dst + 1 * BPS));
212 dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS));
213 dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS));
215 // Convert to 16b.
216 dst0 = _mm_unpacklo_epi8(dst0, zero);
217 dst1 = _mm_unpacklo_epi8(dst1, zero);
218 dst2 = _mm_unpacklo_epi8(dst2, zero);
219 dst3 = _mm_unpacklo_epi8(dst3, zero);
220 // Add the inverse transform(s).
221 dst0 = _mm_add_epi16(dst0, T0);
222 dst1 = _mm_add_epi16(dst1, T1);
223 dst2 = _mm_add_epi16(dst2, T2);
224 dst3 = _mm_add_epi16(dst3, T3);
225 // Unsigned saturate to 8b.
226 dst0 = _mm_packus_epi16(dst0, dst0);
227 dst1 = _mm_packus_epi16(dst1, dst1);
228 dst2 = _mm_packus_epi16(dst2, dst2);
229 dst3 = _mm_packus_epi16(dst3, dst3);
230 // Store the results.
231 if (do_two) {
232 // Store eight bytes/pixels per line.
233 _mm_storel_epi64((__m128i*)(dst + 0 * BPS), dst0);
234 _mm_storel_epi64((__m128i*)(dst + 1 * BPS), dst1);
235 _mm_storel_epi64((__m128i*)(dst + 2 * BPS), dst2);
236 _mm_storel_epi64((__m128i*)(dst + 3 * BPS), dst3);
237 } else {
238 // Store four bytes/pixels per line.
239 *(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0);
240 *(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1);
241 *(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2);
242 *(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3);
247 #if defined(USE_TRANSFORM_AC3)
248 #define MUL(a, b) (((a) * (b)) >> 16)
249 static void TransformAC3(const int16_t* in, uint8_t* dst) {
250 static const int kC1 = 20091 + (1 << 16);
251 static const int kC2 = 35468;
252 const __m128i A = _mm_set1_epi16(in[0] + 4);
253 const __m128i c4 = _mm_set1_epi16(MUL(in[4], kC2));
254 const __m128i d4 = _mm_set1_epi16(MUL(in[4], kC1));
255 const int c1 = MUL(in[1], kC2);
256 const int d1 = MUL(in[1], kC1);
257 const __m128i CD = _mm_set_epi16(0, 0, 0, 0, -d1, -c1, c1, d1);
258 const __m128i B = _mm_adds_epi16(A, CD);
259 const __m128i m0 = _mm_adds_epi16(B, d4);
260 const __m128i m1 = _mm_adds_epi16(B, c4);
261 const __m128i m2 = _mm_subs_epi16(B, c4);
262 const __m128i m3 = _mm_subs_epi16(B, d4);
263 const __m128i zero = _mm_setzero_si128();
264 // Load the source pixels.
265 __m128i dst0 = _mm_cvtsi32_si128(*(int*)(dst + 0 * BPS));
266 __m128i dst1 = _mm_cvtsi32_si128(*(int*)(dst + 1 * BPS));
267 __m128i dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS));
268 __m128i dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS));
269 // Convert to 16b.
270 dst0 = _mm_unpacklo_epi8(dst0, zero);
271 dst1 = _mm_unpacklo_epi8(dst1, zero);
272 dst2 = _mm_unpacklo_epi8(dst2, zero);
273 dst3 = _mm_unpacklo_epi8(dst3, zero);
274 // Add the inverse transform.
275 dst0 = _mm_adds_epi16(dst0, _mm_srai_epi16(m0, 3));
276 dst1 = _mm_adds_epi16(dst1, _mm_srai_epi16(m1, 3));
277 dst2 = _mm_adds_epi16(dst2, _mm_srai_epi16(m2, 3));
278 dst3 = _mm_adds_epi16(dst3, _mm_srai_epi16(m3, 3));
279 // Unsigned saturate to 8b.
280 dst0 = _mm_packus_epi16(dst0, dst0);
281 dst1 = _mm_packus_epi16(dst1, dst1);
282 dst2 = _mm_packus_epi16(dst2, dst2);
283 dst3 = _mm_packus_epi16(dst3, dst3);
284 // Store the results.
285 *(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0);
286 *(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1);
287 *(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2);
288 *(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3);
290 #undef MUL
291 #endif // USE_TRANSFORM_AC3
293 //------------------------------------------------------------------------------
294 // Loop Filter (Paragraph 15)
296 // Compute abs(p - q) = subs(p - q) OR subs(q - p)
297 #define MM_ABS(p, q) _mm_or_si128( \
298 _mm_subs_epu8((q), (p)), \
299 _mm_subs_epu8((p), (q)))
301 // Shift each byte of "x" by 3 bits while preserving by the sign bit.
302 static WEBP_INLINE void SignedShift8b(__m128i* const x) {
303 const __m128i zero = _mm_setzero_si128();
304 const __m128i signs = _mm_cmpgt_epi8(zero, *x);
305 const __m128i lo_0 = _mm_unpacklo_epi8(*x, signs); // s8 -> s16 sign extend
306 const __m128i hi_0 = _mm_unpackhi_epi8(*x, signs);
307 const __m128i lo_1 = _mm_srai_epi16(lo_0, 3);
308 const __m128i hi_1 = _mm_srai_epi16(hi_0, 3);
309 *x = _mm_packs_epi16(lo_1, hi_1);
312 #define FLIP_SIGN_BIT2(a, b) { \
313 a = _mm_xor_si128(a, sign_bit); \
314 b = _mm_xor_si128(b, sign_bit); \
317 #define FLIP_SIGN_BIT4(a, b, c, d) { \
318 FLIP_SIGN_BIT2(a, b); \
319 FLIP_SIGN_BIT2(c, d); \
322 // input/output is uint8_t
323 static WEBP_INLINE void GetNotHEV(const __m128i* const p1,
324 const __m128i* const p0,
325 const __m128i* const q0,
326 const __m128i* const q1,
327 int hev_thresh, __m128i* const not_hev) {
328 const __m128i zero = _mm_setzero_si128();
329 const __m128i t_1 = MM_ABS(*p1, *p0);
330 const __m128i t_2 = MM_ABS(*q1, *q0);
332 const __m128i h = _mm_set1_epi8(hev_thresh);
333 const __m128i t_3 = _mm_subs_epu8(t_1, h); // abs(p1 - p0) - hev_tresh
334 const __m128i t_4 = _mm_subs_epu8(t_2, h); // abs(q1 - q0) - hev_tresh
336 *not_hev = _mm_or_si128(t_3, t_4);
337 *not_hev = _mm_cmpeq_epi8(*not_hev, zero); // not_hev <= t1 && not_hev <= t2
340 // input pixels are int8_t
341 static WEBP_INLINE void GetBaseDelta(const __m128i* const p1,
342 const __m128i* const p0,
343 const __m128i* const q0,
344 const __m128i* const q1,
345 __m128i* const delta) {
346 // beware of addition order, for saturation!
347 const __m128i p1_q1 = _mm_subs_epi8(*p1, *q1); // p1 - q1
348 const __m128i q0_p0 = _mm_subs_epi8(*q0, *p0); // q0 - p0
349 const __m128i s1 = _mm_adds_epi8(p1_q1, q0_p0); // p1 - q1 + 1 * (q0 - p0)
350 const __m128i s2 = _mm_adds_epi8(q0_p0, s1); // p1 - q1 + 2 * (q0 - p0)
351 const __m128i s3 = _mm_adds_epi8(q0_p0, s2); // p1 - q1 + 3 * (q0 - p0)
352 *delta = s3;
355 // input and output are int8_t
356 static WEBP_INLINE void DoSimpleFilter(__m128i* const p0, __m128i* const q0,
357 const __m128i* const fl) {
358 const __m128i k3 = _mm_set1_epi8(3);
359 const __m128i k4 = _mm_set1_epi8(4);
360 __m128i v3 = _mm_adds_epi8(*fl, k3);
361 __m128i v4 = _mm_adds_epi8(*fl, k4);
363 SignedShift8b(&v4); // v4 >> 3
364 SignedShift8b(&v3); // v3 >> 3
365 *q0 = _mm_subs_epi8(*q0, v4); // q0 -= v4
366 *p0 = _mm_adds_epi8(*p0, v3); // p0 += v3
369 // Updates values of 2 pixels at MB edge during complex filtering.
370 // Update operations:
371 // q = q - delta and p = p + delta; where delta = [(a_hi >> 7), (a_lo >> 7)]
372 // Pixels 'pi' and 'qi' are int8_t on input, uint8_t on output (sign flip).
373 static WEBP_INLINE void Update2Pixels(__m128i* const pi, __m128i* const qi,
374 const __m128i* const a0_lo,
375 const __m128i* const a0_hi) {
376 const __m128i a1_lo = _mm_srai_epi16(*a0_lo, 7);
377 const __m128i a1_hi = _mm_srai_epi16(*a0_hi, 7);
378 const __m128i delta = _mm_packs_epi16(a1_lo, a1_hi);
379 const __m128i sign_bit = _mm_set1_epi8(0x80);
380 *pi = _mm_adds_epi8(*pi, delta);
381 *qi = _mm_subs_epi8(*qi, delta);
382 FLIP_SIGN_BIT2(*pi, *qi);
385 // input pixels are uint8_t
386 static WEBP_INLINE void NeedsFilter(const __m128i* const p1,
387 const __m128i* const p0,
388 const __m128i* const q0,
389 const __m128i* const q1,
390 int thresh, __m128i* const mask) {
391 const __m128i m_thresh = _mm_set1_epi8(thresh);
392 const __m128i t1 = MM_ABS(*p1, *q1); // abs(p1 - q1)
393 const __m128i kFE = _mm_set1_epi8(0xFE);
394 const __m128i t2 = _mm_and_si128(t1, kFE); // set lsb of each byte to zero
395 const __m128i t3 = _mm_srli_epi16(t2, 1); // abs(p1 - q1) / 2
397 const __m128i t4 = MM_ABS(*p0, *q0); // abs(p0 - q0)
398 const __m128i t5 = _mm_adds_epu8(t4, t4); // abs(p0 - q0) * 2
399 const __m128i t6 = _mm_adds_epu8(t5, t3); // abs(p0-q0)*2 + abs(p1-q1)/2
401 const __m128i t7 = _mm_subs_epu8(t6, m_thresh); // mask <= m_thresh
402 *mask = _mm_cmpeq_epi8(t7, _mm_setzero_si128());
405 //------------------------------------------------------------------------------
406 // Edge filtering functions
408 // Applies filter on 2 pixels (p0 and q0)
409 static WEBP_INLINE void DoFilter2(__m128i* const p1, __m128i* const p0,
410 __m128i* const q0, __m128i* const q1,
411 int thresh) {
412 __m128i a, mask;
413 const __m128i sign_bit = _mm_set1_epi8(0x80);
414 // convert p1/q1 to int8_t (for GetBaseDelta)
415 const __m128i p1s = _mm_xor_si128(*p1, sign_bit);
416 const __m128i q1s = _mm_xor_si128(*q1, sign_bit);
418 NeedsFilter(p1, p0, q0, q1, thresh, &mask);
420 FLIP_SIGN_BIT2(*p0, *q0);
421 GetBaseDelta(&p1s, p0, q0, &q1s, &a);
422 a = _mm_and_si128(a, mask); // mask filter values we don't care about
423 DoSimpleFilter(p0, q0, &a);
424 FLIP_SIGN_BIT2(*p0, *q0);
427 // Applies filter on 4 pixels (p1, p0, q0 and q1)
428 static WEBP_INLINE void DoFilter4(__m128i* const p1, __m128i* const p0,
429 __m128i* const q0, __m128i* const q1,
430 const __m128i* const mask, int hev_thresh) {
431 const __m128i sign_bit = _mm_set1_epi8(0x80);
432 const __m128i k64 = _mm_set1_epi8(0x40);
433 const __m128i zero = _mm_setzero_si128();
434 __m128i not_hev;
435 __m128i t1, t2, t3;
437 // compute hev mask
438 GetNotHEV(p1, p0, q0, q1, hev_thresh, &not_hev);
440 // convert to signed values
441 FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
443 t1 = _mm_subs_epi8(*p1, *q1); // p1 - q1
444 t1 = _mm_andnot_si128(not_hev, t1); // hev(p1 - q1)
445 t2 = _mm_subs_epi8(*q0, *p0); // q0 - p0
446 t1 = _mm_adds_epi8(t1, t2); // hev(p1 - q1) + 1 * (q0 - p0)
447 t1 = _mm_adds_epi8(t1, t2); // hev(p1 - q1) + 2 * (q0 - p0)
448 t1 = _mm_adds_epi8(t1, t2); // hev(p1 - q1) + 3 * (q0 - p0)
449 t1 = _mm_and_si128(t1, *mask); // mask filter values we don't care about
451 t2 = _mm_set1_epi8(3);
452 t3 = _mm_set1_epi8(4);
453 t2 = _mm_adds_epi8(t1, t2); // 3 * (q0 - p0) + (p1 - q1) + 3
454 t3 = _mm_adds_epi8(t1, t3); // 3 * (q0 - p0) + (p1 - q1) + 4
455 SignedShift8b(&t2); // (3 * (q0 - p0) + hev(p1 - q1) + 3) >> 3
456 SignedShift8b(&t3); // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 3
457 *p0 = _mm_adds_epi8(*p0, t2); // p0 += t2
458 *q0 = _mm_subs_epi8(*q0, t3); // q0 -= t3
459 FLIP_SIGN_BIT2(*p0, *q0);
461 // this is equivalent to signed (a + 1) >> 1 calculation
462 t2 = _mm_add_epi8(t3, sign_bit);
463 t3 = _mm_avg_epu8(t2, zero);
464 t3 = _mm_sub_epi8(t3, k64);
466 t3 = _mm_and_si128(not_hev, t3); // if !hev
467 *q1 = _mm_subs_epi8(*q1, t3); // q1 -= t3
468 *p1 = _mm_adds_epi8(*p1, t3); // p1 += t3
469 FLIP_SIGN_BIT2(*p1, *q1);
472 // Applies filter on 6 pixels (p2, p1, p0, q0, q1 and q2)
473 static WEBP_INLINE void DoFilter6(__m128i* const p2, __m128i* const p1,
474 __m128i* const p0, __m128i* const q0,
475 __m128i* const q1, __m128i* const q2,
476 const __m128i* const mask, int hev_thresh) {
477 const __m128i zero = _mm_setzero_si128();
478 const __m128i sign_bit = _mm_set1_epi8(0x80);
479 __m128i a, not_hev;
481 // compute hev mask
482 GetNotHEV(p1, p0, q0, q1, hev_thresh, &not_hev);
484 FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
485 FLIP_SIGN_BIT2(*p2, *q2);
486 GetBaseDelta(p1, p0, q0, q1, &a);
488 { // do simple filter on pixels with hev
489 const __m128i m = _mm_andnot_si128(not_hev, *mask);
490 const __m128i f = _mm_and_si128(a, m);
491 DoSimpleFilter(p0, q0, &f);
494 { // do strong filter on pixels with not hev
495 const __m128i k9 = _mm_set1_epi16(0x0900);
496 const __m128i k63 = _mm_set1_epi16(63);
498 const __m128i m = _mm_and_si128(not_hev, *mask);
499 const __m128i f = _mm_and_si128(a, m);
501 const __m128i f_lo = _mm_unpacklo_epi8(zero, f);
502 const __m128i f_hi = _mm_unpackhi_epi8(zero, f);
504 const __m128i f9_lo = _mm_mulhi_epi16(f_lo, k9); // Filter (lo) * 9
505 const __m128i f9_hi = _mm_mulhi_epi16(f_hi, k9); // Filter (hi) * 9
507 const __m128i a2_lo = _mm_add_epi16(f9_lo, k63); // Filter * 9 + 63
508 const __m128i a2_hi = _mm_add_epi16(f9_hi, k63); // Filter * 9 + 63
510 const __m128i a1_lo = _mm_add_epi16(a2_lo, f9_lo); // Filter * 18 + 63
511 const __m128i a1_hi = _mm_add_epi16(a2_hi, f9_hi); // Filter * 18 + 63
513 const __m128i a0_lo = _mm_add_epi16(a1_lo, f9_lo); // Filter * 27 + 63
514 const __m128i a0_hi = _mm_add_epi16(a1_hi, f9_hi); // Filter * 27 + 63
516 Update2Pixels(p2, q2, &a2_lo, &a2_hi);
517 Update2Pixels(p1, q1, &a1_lo, &a1_hi);
518 Update2Pixels(p0, q0, &a0_lo, &a0_hi);
522 // reads 8 rows across a vertical edge.
524 // TODO(somnath): Investigate _mm_shuffle* also see if it can be broken into
525 // two Load4x4() to avoid code duplication.
526 static WEBP_INLINE void Load8x4(const uint8_t* const b, int stride,
527 __m128i* const p, __m128i* const q) {
528 __m128i t1, t2;
530 // Load 0th, 1st, 4th and 5th rows
531 __m128i r0 = _mm_cvtsi32_si128(*((int*)&b[0 * stride])); // 03 02 01 00
532 __m128i r1 = _mm_cvtsi32_si128(*((int*)&b[1 * stride])); // 13 12 11 10
533 __m128i r4 = _mm_cvtsi32_si128(*((int*)&b[4 * stride])); // 43 42 41 40
534 __m128i r5 = _mm_cvtsi32_si128(*((int*)&b[5 * stride])); // 53 52 51 50
536 r0 = _mm_unpacklo_epi32(r0, r4); // 43 42 41 40 03 02 01 00
537 r1 = _mm_unpacklo_epi32(r1, r5); // 53 52 51 50 13 12 11 10
539 // t1 = 53 43 52 42 51 41 50 40 13 03 12 02 11 01 10 00
540 t1 = _mm_unpacklo_epi8(r0, r1);
542 // Load 2nd, 3rd, 6th and 7th rows
543 r0 = _mm_cvtsi32_si128(*((int*)&b[2 * stride])); // 23 22 21 22
544 r1 = _mm_cvtsi32_si128(*((int*)&b[3 * stride])); // 33 32 31 30
545 r4 = _mm_cvtsi32_si128(*((int*)&b[6 * stride])); // 63 62 61 60
546 r5 = _mm_cvtsi32_si128(*((int*)&b[7 * stride])); // 73 72 71 70
548 r0 = _mm_unpacklo_epi32(r0, r4); // 63 62 61 60 23 22 21 20
549 r1 = _mm_unpacklo_epi32(r1, r5); // 73 72 71 70 33 32 31 30
551 // t2 = 73 63 72 62 71 61 70 60 33 23 32 22 31 21 30 20
552 t2 = _mm_unpacklo_epi8(r0, r1);
554 // t1 = 33 23 13 03 32 22 12 02 31 21 11 01 30 20 10 00
555 // t2 = 73 63 53 43 72 62 52 42 71 61 51 41 70 60 50 40
556 r0 = t1;
557 t1 = _mm_unpacklo_epi16(t1, t2);
558 t2 = _mm_unpackhi_epi16(r0, t2);
560 // *p = 71 61 51 41 31 21 11 01 70 60 50 40 30 20 10 00
561 // *q = 73 63 53 43 33 23 13 03 72 62 52 42 32 22 12 02
562 *p = _mm_unpacklo_epi32(t1, t2);
563 *q = _mm_unpackhi_epi32(t1, t2);
566 static WEBP_INLINE void Load16x4(const uint8_t* const r0,
567 const uint8_t* const r8,
568 int stride,
569 __m128i* const p1, __m128i* const p0,
570 __m128i* const q0, __m128i* const q1) {
571 __m128i t1, t2;
572 // Assume the pixels around the edge (|) are numbered as follows
573 // 00 01 | 02 03
574 // 10 11 | 12 13
575 // ... | ...
576 // e0 e1 | e2 e3
577 // f0 f1 | f2 f3
579 // r0 is pointing to the 0th row (00)
580 // r8 is pointing to the 8th row (80)
582 // Load
583 // p1 = 71 61 51 41 31 21 11 01 70 60 50 40 30 20 10 00
584 // q0 = 73 63 53 43 33 23 13 03 72 62 52 42 32 22 12 02
585 // p0 = f1 e1 d1 c1 b1 a1 91 81 f0 e0 d0 c0 b0 a0 90 80
586 // q1 = f3 e3 d3 c3 b3 a3 93 83 f2 e2 d2 c2 b2 a2 92 82
587 Load8x4(r0, stride, p1, q0);
588 Load8x4(r8, stride, p0, q1);
590 t1 = *p1;
591 t2 = *q0;
592 // p1 = f0 e0 d0 c0 b0 a0 90 80 70 60 50 40 30 20 10 00
593 // p0 = f1 e1 d1 c1 b1 a1 91 81 71 61 51 41 31 21 11 01
594 // q0 = f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02
595 // q1 = f3 e3 d3 c3 b3 a3 93 83 73 63 53 43 33 23 13 03
596 *p1 = _mm_unpacklo_epi64(t1, *p0);
597 *p0 = _mm_unpackhi_epi64(t1, *p0);
598 *q0 = _mm_unpacklo_epi64(t2, *q1);
599 *q1 = _mm_unpackhi_epi64(t2, *q1);
602 static WEBP_INLINE void Store4x4(__m128i* const x, uint8_t* dst, int stride) {
603 int i;
604 for (i = 0; i < 4; ++i, dst += stride) {
605 *((int32_t*)dst) = _mm_cvtsi128_si32(*x);
606 *x = _mm_srli_si128(*x, 4);
610 // Transpose back and store
611 static WEBP_INLINE void Store16x4(const __m128i* const p1,
612 const __m128i* const p0,
613 const __m128i* const q0,
614 const __m128i* const q1,
615 uint8_t* r0, uint8_t* r8,
616 int stride) {
617 __m128i t1, p1_s, p0_s, q0_s, q1_s;
619 // p0 = 71 70 61 60 51 50 41 40 31 30 21 20 11 10 01 00
620 // p1 = f1 f0 e1 e0 d1 d0 c1 c0 b1 b0 a1 a0 91 90 81 80
621 t1 = *p0;
622 p0_s = _mm_unpacklo_epi8(*p1, t1);
623 p1_s = _mm_unpackhi_epi8(*p1, t1);
625 // q0 = 73 72 63 62 53 52 43 42 33 32 23 22 13 12 03 02
626 // q1 = f3 f2 e3 e2 d3 d2 c3 c2 b3 b2 a3 a2 93 92 83 82
627 t1 = *q0;
628 q0_s = _mm_unpacklo_epi8(t1, *q1);
629 q1_s = _mm_unpackhi_epi8(t1, *q1);
631 // p0 = 33 32 31 30 23 22 21 20 13 12 11 10 03 02 01 00
632 // q0 = 73 72 71 70 63 62 61 60 53 52 51 50 43 42 41 40
633 t1 = p0_s;
634 p0_s = _mm_unpacklo_epi16(t1, q0_s);
635 q0_s = _mm_unpackhi_epi16(t1, q0_s);
637 // p1 = b3 b2 b1 b0 a3 a2 a1 a0 93 92 91 90 83 82 81 80
638 // q1 = f3 f2 f1 f0 e3 e2 e1 e0 d3 d2 d1 d0 c3 c2 c1 c0
639 t1 = p1_s;
640 p1_s = _mm_unpacklo_epi16(t1, q1_s);
641 q1_s = _mm_unpackhi_epi16(t1, q1_s);
643 Store4x4(&p0_s, r0, stride);
644 r0 += 4 * stride;
645 Store4x4(&q0_s, r0, stride);
647 Store4x4(&p1_s, r8, stride);
648 r8 += 4 * stride;
649 Store4x4(&q1_s, r8, stride);
652 //------------------------------------------------------------------------------
653 // Simple In-loop filtering (Paragraph 15.2)
655 static void SimpleVFilter16(uint8_t* p, int stride, int thresh) {
656 // Load
657 __m128i p1 = _mm_loadu_si128((__m128i*)&p[-2 * stride]);
658 __m128i p0 = _mm_loadu_si128((__m128i*)&p[-stride]);
659 __m128i q0 = _mm_loadu_si128((__m128i*)&p[0]);
660 __m128i q1 = _mm_loadu_si128((__m128i*)&p[stride]);
662 DoFilter2(&p1, &p0, &q0, &q1, thresh);
664 // Store
665 _mm_storeu_si128((__m128i*)&p[-stride], p0);
666 _mm_storeu_si128((__m128i*)&p[0], q0);
669 static void SimpleHFilter16(uint8_t* p, int stride, int thresh) {
670 __m128i p1, p0, q0, q1;
672 p -= 2; // beginning of p1
674 Load16x4(p, p + 8 * stride, stride, &p1, &p0, &q0, &q1);
675 DoFilter2(&p1, &p0, &q0, &q1, thresh);
676 Store16x4(&p1, &p0, &q0, &q1, p, p + 8 * stride, stride);
679 static void SimpleVFilter16i(uint8_t* p, int stride, int thresh) {
680 int k;
681 for (k = 3; k > 0; --k) {
682 p += 4 * stride;
683 SimpleVFilter16(p, stride, thresh);
687 static void SimpleHFilter16i(uint8_t* p, int stride, int thresh) {
688 int k;
689 for (k = 3; k > 0; --k) {
690 p += 4;
691 SimpleHFilter16(p, stride, thresh);
695 //------------------------------------------------------------------------------
696 // Complex In-loop filtering (Paragraph 15.3)
698 #define MAX_DIFF1(p3, p2, p1, p0, m) do { \
699 m = MM_ABS(p1, p0); \
700 m = _mm_max_epu8(m, MM_ABS(p3, p2)); \
701 m = _mm_max_epu8(m, MM_ABS(p2, p1)); \
702 } while (0)
704 #define MAX_DIFF2(p3, p2, p1, p0, m) do { \
705 m = _mm_max_epu8(m, MM_ABS(p1, p0)); \
706 m = _mm_max_epu8(m, MM_ABS(p3, p2)); \
707 m = _mm_max_epu8(m, MM_ABS(p2, p1)); \
708 } while (0)
710 #define LOAD_H_EDGES4(p, stride, e1, e2, e3, e4) { \
711 e1 = _mm_loadu_si128((__m128i*)&(p)[0 * stride]); \
712 e2 = _mm_loadu_si128((__m128i*)&(p)[1 * stride]); \
713 e3 = _mm_loadu_si128((__m128i*)&(p)[2 * stride]); \
714 e4 = _mm_loadu_si128((__m128i*)&(p)[3 * stride]); \
717 #define LOADUV_H_EDGE(p, u, v, stride) do { \
718 const __m128i U = _mm_loadl_epi64((__m128i*)&(u)[(stride)]); \
719 const __m128i V = _mm_loadl_epi64((__m128i*)&(v)[(stride)]); \
720 p = _mm_unpacklo_epi64(U, V); \
721 } while (0)
723 #define LOADUV_H_EDGES4(u, v, stride, e1, e2, e3, e4) { \
724 LOADUV_H_EDGE(e1, u, v, 0 * stride); \
725 LOADUV_H_EDGE(e2, u, v, 1 * stride); \
726 LOADUV_H_EDGE(e3, u, v, 2 * stride); \
727 LOADUV_H_EDGE(e4, u, v, 3 * stride); \
730 #define STOREUV(p, u, v, stride) { \
731 _mm_storel_epi64((__m128i*)&u[(stride)], p); \
732 p = _mm_srli_si128(p, 8); \
733 _mm_storel_epi64((__m128i*)&v[(stride)], p); \
736 static WEBP_INLINE void ComplexMask(const __m128i* const p1,
737 const __m128i* const p0,
738 const __m128i* const q0,
739 const __m128i* const q1,
740 int thresh, int ithresh,
741 __m128i* const mask) {
742 const __m128i it = _mm_set1_epi8(ithresh);
743 const __m128i diff = _mm_subs_epu8(*mask, it);
744 const __m128i thresh_mask = _mm_cmpeq_epi8(diff, _mm_setzero_si128());
745 __m128i filter_mask;
746 NeedsFilter(p1, p0, q0, q1, thresh, &filter_mask);
747 *mask = _mm_and_si128(thresh_mask, filter_mask);
750 // on macroblock edges
751 static void VFilter16(uint8_t* p, int stride,
752 int thresh, int ithresh, int hev_thresh) {
753 __m128i t1;
754 __m128i mask;
755 __m128i p2, p1, p0, q0, q1, q2;
757 // Load p3, p2, p1, p0
758 LOAD_H_EDGES4(p - 4 * stride, stride, t1, p2, p1, p0);
759 MAX_DIFF1(t1, p2, p1, p0, mask);
761 // Load q0, q1, q2, q3
762 LOAD_H_EDGES4(p, stride, q0, q1, q2, t1);
763 MAX_DIFF2(t1, q2, q1, q0, mask);
765 ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
766 DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
768 // Store
769 _mm_storeu_si128((__m128i*)&p[-3 * stride], p2);
770 _mm_storeu_si128((__m128i*)&p[-2 * stride], p1);
771 _mm_storeu_si128((__m128i*)&p[-1 * stride], p0);
772 _mm_storeu_si128((__m128i*)&p[+0 * stride], q0);
773 _mm_storeu_si128((__m128i*)&p[+1 * stride], q1);
774 _mm_storeu_si128((__m128i*)&p[+2 * stride], q2);
777 static void HFilter16(uint8_t* p, int stride,
778 int thresh, int ithresh, int hev_thresh) {
779 __m128i mask;
780 __m128i p3, p2, p1, p0, q0, q1, q2, q3;
782 uint8_t* const b = p - 4;
783 Load16x4(b, b + 8 * stride, stride, &p3, &p2, &p1, &p0); // p3, p2, p1, p0
784 MAX_DIFF1(p3, p2, p1, p0, mask);
786 Load16x4(p, p + 8 * stride, stride, &q0, &q1, &q2, &q3); // q0, q1, q2, q3
787 MAX_DIFF2(q3, q2, q1, q0, mask);
789 ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
790 DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
792 Store16x4(&p3, &p2, &p1, &p0, b, b + 8 * stride, stride);
793 Store16x4(&q0, &q1, &q2, &q3, p, p + 8 * stride, stride);
796 // on three inner edges
797 static void VFilter16i(uint8_t* p, int stride,
798 int thresh, int ithresh, int hev_thresh) {
799 int k;
800 __m128i p3, p2, p1, p0; // loop invariants
802 LOAD_H_EDGES4(p, stride, p3, p2, p1, p0); // prologue
804 for (k = 3; k > 0; --k) {
805 __m128i mask, tmp1, tmp2;
806 uint8_t* const b = p + 2 * stride; // beginning of p1
807 p += 4 * stride;
809 MAX_DIFF1(p3, p2, p1, p0, mask); // compute partial mask
810 LOAD_H_EDGES4(p, stride, p3, p2, tmp1, tmp2);
811 MAX_DIFF2(p3, p2, tmp1, tmp2, mask);
813 // p3 and p2 are not just temporary variables here: they will be
814 // re-used for next span. And q2/q3 will become p1/p0 accordingly.
815 ComplexMask(&p1, &p0, &p3, &p2, thresh, ithresh, &mask);
816 DoFilter4(&p1, &p0, &p3, &p2, &mask, hev_thresh);
818 // Store
819 _mm_storeu_si128((__m128i*)&b[0 * stride], p1);
820 _mm_storeu_si128((__m128i*)&b[1 * stride], p0);
821 _mm_storeu_si128((__m128i*)&b[2 * stride], p3);
822 _mm_storeu_si128((__m128i*)&b[3 * stride], p2);
824 // rotate samples
825 p1 = tmp1;
826 p0 = tmp2;
830 static void HFilter16i(uint8_t* p, int stride,
831 int thresh, int ithresh, int hev_thresh) {
832 int k;
833 __m128i p3, p2, p1, p0; // loop invariants
835 Load16x4(p, p + 8 * stride, stride, &p3, &p2, &p1, &p0); // prologue
837 for (k = 3; k > 0; --k) {
838 __m128i mask, tmp1, tmp2;
839 uint8_t* const b = p + 2; // beginning of p1
841 p += 4; // beginning of q0 (and next span)
843 MAX_DIFF1(p3, p2, p1, p0, mask); // compute partial mask
844 Load16x4(p, p + 8 * stride, stride, &p3, &p2, &tmp1, &tmp2);
845 MAX_DIFF2(p3, p2, tmp1, tmp2, mask);
847 ComplexMask(&p1, &p0, &p3, &p2, thresh, ithresh, &mask);
848 DoFilter4(&p1, &p0, &p3, &p2, &mask, hev_thresh);
850 Store16x4(&p1, &p0, &p3, &p2, b, b + 8 * stride, stride);
852 // rotate samples
853 p1 = tmp1;
854 p0 = tmp2;
858 // 8-pixels wide variant, for chroma filtering
859 static void VFilter8(uint8_t* u, uint8_t* v, int stride,
860 int thresh, int ithresh, int hev_thresh) {
861 __m128i mask;
862 __m128i t1, p2, p1, p0, q0, q1, q2;
864 // Load p3, p2, p1, p0
865 LOADUV_H_EDGES4(u - 4 * stride, v - 4 * stride, stride, t1, p2, p1, p0);
866 MAX_DIFF1(t1, p2, p1, p0, mask);
868 // Load q0, q1, q2, q3
869 LOADUV_H_EDGES4(u, v, stride, q0, q1, q2, t1);
870 MAX_DIFF2(t1, q2, q1, q0, mask);
872 ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
873 DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
875 // Store
876 STOREUV(p2, u, v, -3 * stride);
877 STOREUV(p1, u, v, -2 * stride);
878 STOREUV(p0, u, v, -1 * stride);
879 STOREUV(q0, u, v, 0 * stride);
880 STOREUV(q1, u, v, 1 * stride);
881 STOREUV(q2, u, v, 2 * stride);
884 static void HFilter8(uint8_t* u, uint8_t* v, int stride,
885 int thresh, int ithresh, int hev_thresh) {
886 __m128i mask;
887 __m128i p3, p2, p1, p0, q0, q1, q2, q3;
889 uint8_t* const tu = u - 4;
890 uint8_t* const tv = v - 4;
891 Load16x4(tu, tv, stride, &p3, &p2, &p1, &p0); // p3, p2, p1, p0
892 MAX_DIFF1(p3, p2, p1, p0, mask);
894 Load16x4(u, v, stride, &q0, &q1, &q2, &q3); // q0, q1, q2, q3
895 MAX_DIFF2(q3, q2, q1, q0, mask);
897 ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
898 DoFilter6(&p2, &p1, &p0, &q0, &q1, &q2, &mask, hev_thresh);
900 Store16x4(&p3, &p2, &p1, &p0, tu, tv, stride);
901 Store16x4(&q0, &q1, &q2, &q3, u, v, stride);
904 static void VFilter8i(uint8_t* u, uint8_t* v, int stride,
905 int thresh, int ithresh, int hev_thresh) {
906 __m128i mask;
907 __m128i t1, t2, p1, p0, q0, q1;
909 // Load p3, p2, p1, p0
910 LOADUV_H_EDGES4(u, v, stride, t2, t1, p1, p0);
911 MAX_DIFF1(t2, t1, p1, p0, mask);
913 u += 4 * stride;
914 v += 4 * stride;
916 // Load q0, q1, q2, q3
917 LOADUV_H_EDGES4(u, v, stride, q0, q1, t1, t2);
918 MAX_DIFF2(t2, t1, q1, q0, mask);
920 ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
921 DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
923 // Store
924 STOREUV(p1, u, v, -2 * stride);
925 STOREUV(p0, u, v, -1 * stride);
926 STOREUV(q0, u, v, 0 * stride);
927 STOREUV(q1, u, v, 1 * stride);
930 static void HFilter8i(uint8_t* u, uint8_t* v, int stride,
931 int thresh, int ithresh, int hev_thresh) {
932 __m128i mask;
933 __m128i t1, t2, p1, p0, q0, q1;
934 Load16x4(u, v, stride, &t2, &t1, &p1, &p0); // p3, p2, p1, p0
935 MAX_DIFF1(t2, t1, p1, p0, mask);
937 u += 4; // beginning of q0
938 v += 4;
939 Load16x4(u, v, stride, &q0, &q1, &t1, &t2); // q0, q1, q2, q3
940 MAX_DIFF2(t2, t1, q1, q0, mask);
942 ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
943 DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
945 u -= 2; // beginning of p1
946 v -= 2;
947 Store16x4(&p1, &p0, &q0, &q1, u, v, stride);
950 #endif // WEBP_USE_SSE2
952 //------------------------------------------------------------------------------
953 // Entry point
955 extern void VP8DspInitSSE2(void);
957 void VP8DspInitSSE2(void) {
958 #if defined(WEBP_USE_SSE2)
959 VP8Transform = Transform;
960 #if defined(USE_TRANSFORM_AC3)
961 VP8TransformAC3 = TransformAC3;
962 #endif
964 VP8VFilter16 = VFilter16;
965 VP8HFilter16 = HFilter16;
966 VP8VFilter8 = VFilter8;
967 VP8HFilter8 = HFilter8;
968 VP8VFilter16i = VFilter16i;
969 VP8HFilter16i = HFilter16i;
970 VP8VFilter8i = VFilter8i;
971 VP8HFilter8i = HFilter8i;
973 VP8SimpleVFilter16 = SimpleVFilter16;
974 VP8SimpleHFilter16 = SimpleHFilter16;
975 VP8SimpleVFilter16i = SimpleVFilter16i;
976 VP8SimpleHFilter16i = SimpleHFilter16i;
977 #endif // WEBP_USE_SSE2