1// Copyright 2015 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 variant of methods for lossless encoder
11//
12// Author: Skal (pascal.massimino@gmail.com)
13
14#include "src/dsp/dsp.h"
15
16#if defined(WEBP_USE_SSE2)
17#include <assert.h>
18#include <emmintrin.h>
19#include "src/dsp/lossless.h"
20#include "src/dsp/common_sse2.h"
21#include "src/dsp/lossless_common.h"
22
23// For sign-extended multiplying constants, pre-shifted by 5:
24#define CST_5b(X) (((int16_t)((uint16_t)(X) << 8)) >> 5)
25
26//------------------------------------------------------------------------------
27// Subtract-Green Transform
28
29static void SubtractGreenFromBlueAndRed_SSE2(uint32_t* argb_data,
30 int num_pixels) {
31 int i;
32 for (i = 0; i + 4 <= num_pixels; i += 4) {
33 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb
34 const __m128i A = _mm_srli_epi16(in, 8); // 0 a 0 g
35 const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0));
36 const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // 0g0g
37 const __m128i out = _mm_sub_epi8(in, C);
38 _mm_storeu_si128((__m128i*)&argb_data[i], out);
39 }
40 // fallthrough and finish off with plain-C
41 if (i != num_pixels) {
42 VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i);
43 }
44}
45
46//------------------------------------------------------------------------------
47// Color Transform
48
49#define MK_CST_16(HI, LO) \
50 _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff)))
51
52static void TransformColor_SSE2(const VP8LMultipliers* const m,
53 uint32_t* argb_data, int num_pixels) {
54 const __m128i mults_rb = MK_CST_16(CST_5b(m->green_to_red_),
55 CST_5b(m->green_to_blue_));
56 const __m128i mults_b2 = MK_CST_16(CST_5b(m->red_to_blue_), 0);
57 const __m128i mask_ag = _mm_set1_epi32((int)0xff00ff00); // alpha-green masks
58 const __m128i mask_rb = _mm_set1_epi32(0x00ff00ff); // red-blue masks
59 int i;
60 for (i = 0; i + 4 <= num_pixels; i += 4) {
61 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb
62 const __m128i A = _mm_and_si128(in, mask_ag); // a 0 g 0
63 const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0));
64 const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // g0g0
65 const __m128i D = _mm_mulhi_epi16(C, mults_rb); // x dr x db1
66 const __m128i E = _mm_slli_epi16(in, 8); // r 0 b 0
67 const __m128i F = _mm_mulhi_epi16(E, mults_b2); // x db2 0 0
68 const __m128i G = _mm_srli_epi32(F, 16); // 0 0 x db2
69 const __m128i H = _mm_add_epi8(G, D); // x dr x db
70 const __m128i I = _mm_and_si128(H, mask_rb); // 0 dr 0 db
71 const __m128i out = _mm_sub_epi8(in, I);
72 _mm_storeu_si128((__m128i*)&argb_data[i], out);
73 }
74 // fallthrough and finish off with plain-C
75 if (i != num_pixels) {
76 VP8LTransformColor_C(m, argb_data + i, num_pixels - i);
77 }
78}
79
80//------------------------------------------------------------------------------
81#define SPAN 8
82static void CollectColorBlueTransforms_SSE2(const uint32_t* argb, int stride,
83 int tile_width, int tile_height,
84 int green_to_blue, int red_to_blue,
85 int histo[]) {
86 const __m128i mults_r = MK_CST_16(CST_5b(red_to_blue), 0);
87 const __m128i mults_g = MK_CST_16(0, CST_5b(green_to_blue));
88 const __m128i mask_g = _mm_set1_epi32(0x00ff00); // green mask
89 const __m128i mask_b = _mm_set1_epi32(0x0000ff); // blue mask
90 int y;
91 for (y = 0; y < tile_height; ++y) {
92 const uint32_t* const src = argb + y * stride;
93 int i, x;
94 for (x = 0; x + SPAN <= tile_width; x += SPAN) {
95 uint16_t values[SPAN];
96 const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]);
97 const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]);
98 const __m128i A0 = _mm_slli_epi16(in0, 8); // r 0 | b 0
99 const __m128i A1 = _mm_slli_epi16(in1, 8);
100 const __m128i B0 = _mm_and_si128(in0, mask_g); // 0 0 | g 0
101 const __m128i B1 = _mm_and_si128(in1, mask_g);
102 const __m128i C0 = _mm_mulhi_epi16(A0, mults_r); // x db | 0 0
103 const __m128i C1 = _mm_mulhi_epi16(A1, mults_r);
104 const __m128i D0 = _mm_mulhi_epi16(B0, mults_g); // 0 0 | x db
105 const __m128i D1 = _mm_mulhi_epi16(B1, mults_g);
106 const __m128i E0 = _mm_sub_epi8(in0, D0); // x x | x b'
107 const __m128i E1 = _mm_sub_epi8(in1, D1);
108 const __m128i F0 = _mm_srli_epi32(C0, 16); // 0 0 | x db
109 const __m128i F1 = _mm_srli_epi32(C1, 16);
110 const __m128i G0 = _mm_sub_epi8(E0, F0); // 0 0 | x b'
111 const __m128i G1 = _mm_sub_epi8(E1, F1);
112 const __m128i H0 = _mm_and_si128(G0, mask_b); // 0 0 | 0 b
113 const __m128i H1 = _mm_and_si128(G1, mask_b);
114 const __m128i I = _mm_packs_epi32(H0, H1); // 0 b' | 0 b'
115 _mm_storeu_si128((__m128i*)values, I);
116 for (i = 0; i < SPAN; ++i) ++histo[values[i]];
117 }
118 }
119 {
120 const int left_over = tile_width & (SPAN - 1);
121 if (left_over > 0) {
122 VP8LCollectColorBlueTransforms_C(argb + tile_width - left_over, stride,
123 left_over, tile_height,
124 green_to_blue, red_to_blue, histo);
125 }
126 }
127}
128
129static void CollectColorRedTransforms_SSE2(const uint32_t* argb, int stride,
130 int tile_width, int tile_height,
131 int green_to_red, int histo[]) {
132 const __m128i mults_g = MK_CST_16(0, CST_5b(green_to_red));
133 const __m128i mask_g = _mm_set1_epi32(0x00ff00); // green mask
134 const __m128i mask = _mm_set1_epi32(0xff);
135
136 int y;
137 for (y = 0; y < tile_height; ++y) {
138 const uint32_t* const src = argb + y * stride;
139 int i, x;
140 for (x = 0; x + SPAN <= tile_width; x += SPAN) {
141 uint16_t values[SPAN];
142 const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x + 0]);
143 const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]);
144 const __m128i A0 = _mm_and_si128(in0, mask_g); // 0 0 | g 0
145 const __m128i A1 = _mm_and_si128(in1, mask_g);
146 const __m128i B0 = _mm_srli_epi32(in0, 16); // 0 0 | x r
147 const __m128i B1 = _mm_srli_epi32(in1, 16);
148 const __m128i C0 = _mm_mulhi_epi16(A0, mults_g); // 0 0 | x dr
149 const __m128i C1 = _mm_mulhi_epi16(A1, mults_g);
150 const __m128i E0 = _mm_sub_epi8(B0, C0); // x x | x r'
151 const __m128i E1 = _mm_sub_epi8(B1, C1);
152 const __m128i F0 = _mm_and_si128(E0, mask); // 0 0 | 0 r'
153 const __m128i F1 = _mm_and_si128(E1, mask);
154 const __m128i I = _mm_packs_epi32(F0, F1);
155 _mm_storeu_si128((__m128i*)values, I);
156 for (i = 0; i < SPAN; ++i) ++histo[values[i]];
157 }
158 }
159 {
160 const int left_over = tile_width & (SPAN - 1);
161 if (left_over > 0) {
162 VP8LCollectColorRedTransforms_C(argb + tile_width - left_over, stride,
163 left_over, tile_height,
164 green_to_red, histo);
165 }
166 }
167}
168#undef SPAN
169#undef MK_CST_16
170
171//------------------------------------------------------------------------------
172
173// Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But
174// that's ok since the histogram values are less than 1<<28 (max picture size).
175#define LINE_SIZE 16 // 8 or 16
176static void AddVector_SSE2(const uint32_t* a, const uint32_t* b, uint32_t* out,
177 int size) {
178 int i;
179 for (i = 0; i + LINE_SIZE <= size; i += LINE_SIZE) {
180 const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]);
181 const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
182#if (LINE_SIZE == 16)
183 const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]);
184 const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
185#endif
186 const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i + 0]);
187 const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i + 4]);
188#if (LINE_SIZE == 16)
189 const __m128i b2 = _mm_loadu_si128((const __m128i*)&b[i + 8]);
190 const __m128i b3 = _mm_loadu_si128((const __m128i*)&b[i + 12]);
191#endif
192 _mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0));
193 _mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1));
194#if (LINE_SIZE == 16)
195 _mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2));
196 _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
197#endif
198 }
199 for (; i < size; ++i) {
200 out[i] = a[i] + b[i];
201 }
202}
203
204static void AddVectorEq_SSE2(const uint32_t* a, uint32_t* out, int size) {
205 int i;
206 for (i = 0; i + LINE_SIZE <= size; i += LINE_SIZE) {
207 const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]);
208 const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
209#if (LINE_SIZE == 16)
210 const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]);
211 const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
212#endif
213 const __m128i b0 = _mm_loadu_si128((const __m128i*)&out[i + 0]);
214 const __m128i b1 = _mm_loadu_si128((const __m128i*)&out[i + 4]);
215#if (LINE_SIZE == 16)
216 const __m128i b2 = _mm_loadu_si128((const __m128i*)&out[i + 8]);
217 const __m128i b3 = _mm_loadu_si128((const __m128i*)&out[i + 12]);
218#endif
219 _mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0));
220 _mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1));
221#if (LINE_SIZE == 16)
222 _mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2));
223 _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
224#endif
225 }
226 for (; i < size; ++i) {
227 out[i] += a[i];
228 }
229}
230#undef LINE_SIZE
231
232//------------------------------------------------------------------------------
233// Entropy
234
235// TODO(https://crbug.com/webp/499): this function produces different results
236// from the C code due to use of double/float resulting in output differences
237// when compared to -noasm.
238#if !(defined(WEBP_HAVE_SLOW_CLZ_CTZ) || defined(__i386__) || defined(_M_IX86))
239
240static float CombinedShannonEntropy_SSE2(const int X[256], const int Y[256]) {
241 int i;
242 float retval = 0.f;
243 int sumX = 0, sumXY = 0;
244 const __m128i zero = _mm_setzero_si128();
245
246 for (i = 0; i < 256; i += 16) {
247 const __m128i x0 = _mm_loadu_si128((const __m128i*)(X + i + 0));
248 const __m128i y0 = _mm_loadu_si128((const __m128i*)(Y + i + 0));
249 const __m128i x1 = _mm_loadu_si128((const __m128i*)(X + i + 4));
250 const __m128i y1 = _mm_loadu_si128((const __m128i*)(Y + i + 4));
251 const __m128i x2 = _mm_loadu_si128((const __m128i*)(X + i + 8));
252 const __m128i y2 = _mm_loadu_si128((const __m128i*)(Y + i + 8));
253 const __m128i x3 = _mm_loadu_si128((const __m128i*)(X + i + 12));
254 const __m128i y3 = _mm_loadu_si128((const __m128i*)(Y + i + 12));
255 const __m128i x4 = _mm_packs_epi16(_mm_packs_epi32(x0, x1),
256 _mm_packs_epi32(x2, x3));
257 const __m128i y4 = _mm_packs_epi16(_mm_packs_epi32(y0, y1),
258 _mm_packs_epi32(y2, y3));
259 const int32_t mx = _mm_movemask_epi8(_mm_cmpgt_epi8(x4, zero));
260 int32_t my = _mm_movemask_epi8(_mm_cmpgt_epi8(y4, zero)) | mx;
261 while (my) {
262 const int32_t j = BitsCtz(my);
263 int xy;
264 if ((mx >> j) & 1) {
265 const int x = X[i + j];
266 sumXY += x;
267 retval -= VP8LFastSLog2(x);
268 }
269 xy = X[i + j] + Y[i + j];
270 sumX += xy;
271 retval -= VP8LFastSLog2(xy);
272 my &= my - 1;
273 }
274 }
275 retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY);
276 return retval;
277}
278
279#else
280
281#define DONT_USE_COMBINED_SHANNON_ENTROPY_SSE2_FUNC // won't be faster
282
283#endif
284
285//------------------------------------------------------------------------------
286
287static int VectorMismatch_SSE2(const uint32_t* const array1,
288 const uint32_t* const array2, int length) {
289 int match_len;
290
291 if (length >= 12) {
292 __m128i A0 = _mm_loadu_si128((const __m128i*)&array1[0]);
293 __m128i A1 = _mm_loadu_si128((const __m128i*)&array2[0]);
294 match_len = 0;
295 do {
296 // Loop unrolling and early load both provide a speedup of 10% for the
297 // current function. Also, max_limit can be MAX_LENGTH=4096 at most.
298 const __m128i cmpA = _mm_cmpeq_epi32(A0, A1);
299 const __m128i B0 =
300 _mm_loadu_si128((const __m128i*)&array1[match_len + 4]);
301 const __m128i B1 =
302 _mm_loadu_si128((const __m128i*)&array2[match_len + 4]);
303 if (_mm_movemask_epi8(cmpA) != 0xffff) break;
304 match_len += 4;
305
306 {
307 const __m128i cmpB = _mm_cmpeq_epi32(B0, B1);
308 A0 = _mm_loadu_si128((const __m128i*)&array1[match_len + 4]);
309 A1 = _mm_loadu_si128((const __m128i*)&array2[match_len + 4]);
310 if (_mm_movemask_epi8(cmpB) != 0xffff) break;
311 match_len += 4;
312 }
313 } while (match_len + 12 < length);
314 } else {
315 match_len = 0;
316 // Unroll the potential first two loops.
317 if (length >= 4 &&
318 _mm_movemask_epi8(_mm_cmpeq_epi32(
319 _mm_loadu_si128((const __m128i*)&array1[0]),
320 _mm_loadu_si128((const __m128i*)&array2[0]))) == 0xffff) {
321 match_len = 4;
322 if (length >= 8 &&
323 _mm_movemask_epi8(_mm_cmpeq_epi32(
324 _mm_loadu_si128((const __m128i*)&array1[4]),
325 _mm_loadu_si128((const __m128i*)&array2[4]))) == 0xffff) {
326 match_len = 8;
327 }
328 }
329 }
330
331 while (match_len < length && array1[match_len] == array2[match_len]) {
332 ++match_len;
333 }
334 return match_len;
335}
336
337// Bundles multiple (1, 2, 4 or 8) pixels into a single pixel.
338static void BundleColorMap_SSE2(const uint8_t* const row, int width, int xbits,
339 uint32_t* dst) {
340 int x;
341 assert(xbits >= 0);
342 assert(xbits <= 3);
343 switch (xbits) {
344 case 0: {
345 const __m128i ff = _mm_set1_epi16((short)0xff00);
346 const __m128i zero = _mm_setzero_si128();
347 // Store 0xff000000 | (row[x] << 8).
348 for (x = 0; x + 16 <= width; x += 16, dst += 16) {
349 const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
350 const __m128i in_lo = _mm_unpacklo_epi8(zero, in);
351 const __m128i dst0 = _mm_unpacklo_epi16(in_lo, ff);
352 const __m128i dst1 = _mm_unpackhi_epi16(in_lo, ff);
353 const __m128i in_hi = _mm_unpackhi_epi8(zero, in);
354 const __m128i dst2 = _mm_unpacklo_epi16(in_hi, ff);
355 const __m128i dst3 = _mm_unpackhi_epi16(in_hi, ff);
356 _mm_storeu_si128((__m128i*)&dst[0], dst0);
357 _mm_storeu_si128((__m128i*)&dst[4], dst1);
358 _mm_storeu_si128((__m128i*)&dst[8], dst2);
359 _mm_storeu_si128((__m128i*)&dst[12], dst3);
360 }
361 break;
362 }
363 case 1: {
364 const __m128i ff = _mm_set1_epi16((short)0xff00);
365 const __m128i mul = _mm_set1_epi16(0x110);
366 for (x = 0; x + 16 <= width; x += 16, dst += 8) {
367 // 0a0b | (where a/b are 4 bits).
368 const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
369 const __m128i tmp = _mm_mullo_epi16(in, mul); // aba0
370 const __m128i pack = _mm_and_si128(tmp, ff); // ab00
371 const __m128i dst0 = _mm_unpacklo_epi16(pack, ff);
372 const __m128i dst1 = _mm_unpackhi_epi16(pack, ff);
373 _mm_storeu_si128((__m128i*)&dst[0], dst0);
374 _mm_storeu_si128((__m128i*)&dst[4], dst1);
375 }
376 break;
377 }
378 case 2: {
379 const __m128i mask_or = _mm_set1_epi32((int)0xff000000);
380 const __m128i mul_cst = _mm_set1_epi16(0x0104);
381 const __m128i mask_mul = _mm_set1_epi16(0x0f00);
382 for (x = 0; x + 16 <= width; x += 16, dst += 4) {
383 // 000a000b000c000d | (where a/b/c/d are 2 bits).
384 const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
385 const __m128i mul = _mm_mullo_epi16(in, mul_cst); // 00ab00b000cd00d0
386 const __m128i tmp = _mm_and_si128(mul, mask_mul); // 00ab000000cd0000
387 const __m128i shift = _mm_srli_epi32(tmp, 12); // 00000000ab000000
388 const __m128i pack = _mm_or_si128(shift, tmp); // 00000000abcd0000
389 // Convert to 0xff00**00.
390 const __m128i res = _mm_or_si128(pack, mask_or);
391 _mm_storeu_si128((__m128i*)dst, res);
392 }
393 break;
394 }
395 default: {
396 assert(xbits == 3);
397 for (x = 0; x + 16 <= width; x += 16, dst += 2) {
398 // 0000000a00000000b... | (where a/b are 1 bit).
399 const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
400 const __m128i shift = _mm_slli_epi64(in, 7);
401 const uint32_t move = _mm_movemask_epi8(shift);
402 dst[0] = 0xff000000 | ((move & 0xff) << 8);
403 dst[1] = 0xff000000 | (move & 0xff00);
404 }
405 break;
406 }
407 }
408 if (x != width) {
409 VP8LBundleColorMap_C(row + x, width - x, xbits, dst);
410 }
411}
412
413//------------------------------------------------------------------------------
414// Batch version of Predictor Transform subtraction
415
416static WEBP_INLINE void Average2_m128i(const __m128i* const a0,
417 const __m128i* const a1,
418 __m128i* const avg) {
419 // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1)
420 const __m128i ones = _mm_set1_epi8(1);
421 const __m128i avg1 = _mm_avg_epu8(*a0, *a1);
422 const __m128i one = _mm_and_si128(_mm_xor_si128(*a0, *a1), ones);
423 *avg = _mm_sub_epi8(avg1, one);
424}
425
426// Predictor0: ARGB_BLACK.
427static void PredictorSub0_SSE2(const uint32_t* in, const uint32_t* upper,
428 int num_pixels, uint32_t* out) {
429 int i;
430 const __m128i black = _mm_set1_epi32((int)ARGB_BLACK);
431 for (i = 0; i + 4 <= num_pixels; i += 4) {
432 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
433 const __m128i res = _mm_sub_epi8(src, black);
434 _mm_storeu_si128((__m128i*)&out[i], res);
435 }
436 if (i != num_pixels) {
437 VP8LPredictorsSub_C[0](in + i, NULL, num_pixels - i, out + i);
438 }
439 (void)upper;
440}
441
442#define GENERATE_PREDICTOR_1(X, IN) \
443 static void PredictorSub##X##_SSE2(const uint32_t* const in, \
444 const uint32_t* const upper, \
445 int num_pixels, uint32_t* const out) { \
446 int i; \
447 for (i = 0; i + 4 <= num_pixels; i += 4) { \
448 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \
449 const __m128i pred = _mm_loadu_si128((const __m128i*)&(IN)); \
450 const __m128i res = _mm_sub_epi8(src, pred); \
451 _mm_storeu_si128((__m128i*)&out[i], res); \
452 } \
453 if (i != num_pixels) { \
454 VP8LPredictorsSub_C[(X)](in + i, WEBP_OFFSET_PTR(upper, i), \
455 num_pixels - i, out + i); \
456 } \
457 }
458
459GENERATE_PREDICTOR_1(1, in[i - 1]) // Predictor1: L
460GENERATE_PREDICTOR_1(2, upper[i]) // Predictor2: T
461GENERATE_PREDICTOR_1(3, upper[i + 1]) // Predictor3: TR
462GENERATE_PREDICTOR_1(4, upper[i - 1]) // Predictor4: TL
463#undef GENERATE_PREDICTOR_1
464
465// Predictor5: avg2(avg2(L, TR), T)
466static void PredictorSub5_SSE2(const uint32_t* in, const uint32_t* upper,
467 int num_pixels, uint32_t* out) {
468 int i;
469 for (i = 0; i + 4 <= num_pixels; i += 4) {
470 const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
471 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
472 const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]);
473 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
474 __m128i avg, pred, res;
475 Average2_m128i(&L, &TR, &avg);
476 Average2_m128i(&avg, &T, &pred);
477 res = _mm_sub_epi8(src, pred);
478 _mm_storeu_si128((__m128i*)&out[i], res);
479 }
480 if (i != num_pixels) {
481 VP8LPredictorsSub_C[5](in + i, upper + i, num_pixels - i, out + i);
482 }
483}
484
485#define GENERATE_PREDICTOR_2(X, A, B) \
486static void PredictorSub##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
487 int num_pixels, uint32_t* out) { \
488 int i; \
489 for (i = 0; i + 4 <= num_pixels; i += 4) { \
490 const __m128i tA = _mm_loadu_si128((const __m128i*)&(A)); \
491 const __m128i tB = _mm_loadu_si128((const __m128i*)&(B)); \
492 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \
493 __m128i pred, res; \
494 Average2_m128i(&tA, &tB, &pred); \
495 res = _mm_sub_epi8(src, pred); \
496 _mm_storeu_si128((__m128i*)&out[i], res); \
497 } \
498 if (i != num_pixels) { \
499 VP8LPredictorsSub_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
500 } \
501}
502
503GENERATE_PREDICTOR_2(6, in[i - 1], upper[i - 1]) // Predictor6: avg(L, TL)
504GENERATE_PREDICTOR_2(7, in[i - 1], upper[i]) // Predictor7: avg(L, T)
505GENERATE_PREDICTOR_2(8, upper[i - 1], upper[i]) // Predictor8: avg(TL, T)
506GENERATE_PREDICTOR_2(9, upper[i], upper[i + 1]) // Predictor9: average(T, TR)
507#undef GENERATE_PREDICTOR_2
508
509// Predictor10: avg(avg(L,TL), avg(T, TR)).
510static void PredictorSub10_SSE2(const uint32_t* in, const uint32_t* upper,
511 int num_pixels, uint32_t* out) {
512 int i;
513 for (i = 0; i + 4 <= num_pixels; i += 4) {
514 const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
515 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
516 const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
517 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
518 const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]);
519 __m128i avgTTR, avgLTL, avg, res;
520 Average2_m128i(&T, &TR, &avgTTR);
521 Average2_m128i(&L, &TL, &avgLTL);
522 Average2_m128i(&avgTTR, &avgLTL, &avg);
523 res = _mm_sub_epi8(src, avg);
524 _mm_storeu_si128((__m128i*)&out[i], res);
525 }
526 if (i != num_pixels) {
527 VP8LPredictorsSub_C[10](in + i, upper + i, num_pixels - i, out + i);
528 }
529}
530
531// Predictor11: select.
532static void GetSumAbsDiff32_SSE2(const __m128i* const A, const __m128i* const B,
533 __m128i* const out) {
534 // We can unpack with any value on the upper 32 bits, provided it's the same
535 // on both operands (to that their sum of abs diff is zero). Here we use *A.
536 const __m128i A_lo = _mm_unpacklo_epi32(*A, *A);
537 const __m128i B_lo = _mm_unpacklo_epi32(*B, *A);
538 const __m128i A_hi = _mm_unpackhi_epi32(*A, *A);
539 const __m128i B_hi = _mm_unpackhi_epi32(*B, *A);
540 const __m128i s_lo = _mm_sad_epu8(A_lo, B_lo);
541 const __m128i s_hi = _mm_sad_epu8(A_hi, B_hi);
542 *out = _mm_packs_epi32(s_lo, s_hi);
543}
544
545static void PredictorSub11_SSE2(const uint32_t* in, const uint32_t* upper,
546 int num_pixels, uint32_t* out) {
547 int i;
548 for (i = 0; i + 4 <= num_pixels; i += 4) {
549 const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
550 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
551 const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
552 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
553 __m128i pa, pb;
554 GetSumAbsDiff32_SSE2(&T, &TL, &pa); // pa = sum |T-TL|
555 GetSumAbsDiff32_SSE2(&L, &TL, &pb); // pb = sum |L-TL|
556 {
557 const __m128i mask = _mm_cmpgt_epi32(pb, pa);
558 const __m128i A = _mm_and_si128(mask, L);
559 const __m128i B = _mm_andnot_si128(mask, T);
560 const __m128i pred = _mm_or_si128(A, B); // pred = (L > T)? L : T
561 const __m128i res = _mm_sub_epi8(src, pred);
562 _mm_storeu_si128((__m128i*)&out[i], res);
563 }
564 }
565 if (i != num_pixels) {
566 VP8LPredictorsSub_C[11](in + i, upper + i, num_pixels - i, out + i);
567 }
568}
569
570// Predictor12: ClampedSubSubtractFull.
571static void PredictorSub12_SSE2(const uint32_t* in, const uint32_t* upper,
572 int num_pixels, uint32_t* out) {
573 int i;
574 const __m128i zero = _mm_setzero_si128();
575 for (i = 0; i + 4 <= num_pixels; i += 4) {
576 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
577 const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
578 const __m128i L_lo = _mm_unpacklo_epi8(L, zero);
579 const __m128i L_hi = _mm_unpackhi_epi8(L, zero);
580 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
581 const __m128i T_lo = _mm_unpacklo_epi8(T, zero);
582 const __m128i T_hi = _mm_unpackhi_epi8(T, zero);
583 const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
584 const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero);
585 const __m128i TL_hi = _mm_unpackhi_epi8(TL, zero);
586 const __m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo);
587 const __m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi);
588 const __m128i pred_lo = _mm_add_epi16(L_lo, diff_lo);
589 const __m128i pred_hi = _mm_add_epi16(L_hi, diff_hi);
590 const __m128i pred = _mm_packus_epi16(pred_lo, pred_hi);
591 const __m128i res = _mm_sub_epi8(src, pred);
592 _mm_storeu_si128((__m128i*)&out[i], res);
593 }
594 if (i != num_pixels) {
595 VP8LPredictorsSub_C[12](in + i, upper + i, num_pixels - i, out + i);
596 }
597}
598
599// Predictors13: ClampedAddSubtractHalf
600static void PredictorSub13_SSE2(const uint32_t* in, const uint32_t* upper,
601 int num_pixels, uint32_t* out) {
602 int i;
603 const __m128i zero = _mm_setzero_si128();
604 for (i = 0; i + 2 <= num_pixels; i += 2) {
605 // we can only process two pixels at a time
606 const __m128i L = _mm_loadl_epi64((const __m128i*)&in[i - 1]);
607 const __m128i src = _mm_loadl_epi64((const __m128i*)&in[i]);
608 const __m128i T = _mm_loadl_epi64((const __m128i*)&upper[i]);
609 const __m128i TL = _mm_loadl_epi64((const __m128i*)&upper[i - 1]);
610 const __m128i L_lo = _mm_unpacklo_epi8(L, zero);
611 const __m128i T_lo = _mm_unpacklo_epi8(T, zero);
612 const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero);
613 const __m128i sum = _mm_add_epi16(T_lo, L_lo);
614 const __m128i avg = _mm_srli_epi16(sum, 1);
615 const __m128i A1 = _mm_sub_epi16(avg, TL_lo);
616 const __m128i bit_fix = _mm_cmpgt_epi16(TL_lo, avg);
617 const __m128i A2 = _mm_sub_epi16(A1, bit_fix);
618 const __m128i A3 = _mm_srai_epi16(A2, 1);
619 const __m128i A4 = _mm_add_epi16(avg, A3);
620 const __m128i pred = _mm_packus_epi16(A4, A4);
621 const __m128i res = _mm_sub_epi8(src, pred);
622 _mm_storel_epi64((__m128i*)&out[i], res);
623 }
624 if (i != num_pixels) {
625 VP8LPredictorsSub_C[13](in + i, upper + i, num_pixels - i, out + i);
626 }
627}
628
629//------------------------------------------------------------------------------
630// Entry point
631
632extern void VP8LEncDspInitSSE2(void);
633
634WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) {
635 VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_SSE2;
636 VP8LTransformColor = TransformColor_SSE2;
637 VP8LCollectColorBlueTransforms = CollectColorBlueTransforms_SSE2;
638 VP8LCollectColorRedTransforms = CollectColorRedTransforms_SSE2;
639 VP8LAddVector = AddVector_SSE2;
640 VP8LAddVectorEq = AddVectorEq_SSE2;
641#if !defined(DONT_USE_COMBINED_SHANNON_ENTROPY_SSE2_FUNC)
642 VP8LCombinedShannonEntropy = CombinedShannonEntropy_SSE2;
643#endif
644 VP8LVectorMismatch = VectorMismatch_SSE2;
645 VP8LBundleColorMap = BundleColorMap_SSE2;
646
647 VP8LPredictorsSub[0] = PredictorSub0_SSE2;
648 VP8LPredictorsSub[1] = PredictorSub1_SSE2;
649 VP8LPredictorsSub[2] = PredictorSub2_SSE2;
650 VP8LPredictorsSub[3] = PredictorSub3_SSE2;
651 VP8LPredictorsSub[4] = PredictorSub4_SSE2;
652 VP8LPredictorsSub[5] = PredictorSub5_SSE2;
653 VP8LPredictorsSub[6] = PredictorSub6_SSE2;
654 VP8LPredictorsSub[7] = PredictorSub7_SSE2;
655 VP8LPredictorsSub[8] = PredictorSub8_SSE2;
656 VP8LPredictorsSub[9] = PredictorSub9_SSE2;
657 VP8LPredictorsSub[10] = PredictorSub10_SSE2;
658 VP8LPredictorsSub[11] = PredictorSub11_SSE2;
659 VP8LPredictorsSub[12] = PredictorSub12_SSE2;
660 VP8LPredictorsSub[13] = PredictorSub13_SSE2;
661 VP8LPredictorsSub[14] = PredictorSub0_SSE2; // <- padding security sentinels
662 VP8LPredictorsSub[15] = PredictorSub0_SSE2;
663}
664
665#else // !WEBP_USE_SSE2
666
667WEBP_DSP_INIT_STUB(VP8LEncDspInitSSE2)
668
669#endif // WEBP_USE_SSE2
670