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
SubtractGreenFromBlueAndRed_SSE2(uint32_t * argb_data,int num_pixels)29 static 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
TransformColor_SSE2(const VP8LMultipliers * const m,uint32_t * argb_data,int num_pixels)52 static 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(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
CollectColorBlueTransforms_SSE2(const uint32_t * argb,int stride,int tile_width,int tile_height,int green_to_blue,int red_to_blue,int histo[])82 static 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
CollectColorRedTransforms_SSE2(const uint32_t * argb,int stride,int tile_width,int tile_height,int green_to_red,int histo[])129 static 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 #define LINE_SIZE 16 // 8 or 16
AddVector_SSE2(const uint32_t * a,const uint32_t * b,uint32_t * out,int size)174 static void AddVector_SSE2(const uint32_t* a, const uint32_t* b, uint32_t* out,
175 int size) {
176 int i;
177 assert(size % LINE_SIZE == 0);
178 for (i = 0; i < size; i += LINE_SIZE) {
179 const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]);
180 const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
181 #if (LINE_SIZE == 16)
182 const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]);
183 const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
184 #endif
185 const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i + 0]);
186 const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i + 4]);
187 #if (LINE_SIZE == 16)
188 const __m128i b2 = _mm_loadu_si128((const __m128i*)&b[i + 8]);
189 const __m128i b3 = _mm_loadu_si128((const __m128i*)&b[i + 12]);
190 #endif
191 _mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0));
192 _mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1));
193 #if (LINE_SIZE == 16)
194 _mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2));
195 _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
196 #endif
197 }
198 }
199
AddVectorEq_SSE2(const uint32_t * a,uint32_t * out,int size)200 static void AddVectorEq_SSE2(const uint32_t* a, uint32_t* out, int size) {
201 int i;
202 assert(size % LINE_SIZE == 0);
203 for (i = 0; i < size; i += LINE_SIZE) {
204 const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i + 0]);
205 const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i + 4]);
206 #if (LINE_SIZE == 16)
207 const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i + 8]);
208 const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
209 #endif
210 const __m128i b0 = _mm_loadu_si128((const __m128i*)&out[i + 0]);
211 const __m128i b1 = _mm_loadu_si128((const __m128i*)&out[i + 4]);
212 #if (LINE_SIZE == 16)
213 const __m128i b2 = _mm_loadu_si128((const __m128i*)&out[i + 8]);
214 const __m128i b3 = _mm_loadu_si128((const __m128i*)&out[i + 12]);
215 #endif
216 _mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0));
217 _mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1));
218 #if (LINE_SIZE == 16)
219 _mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2));
220 _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
221 #endif
222 }
223 }
224 #undef LINE_SIZE
225
226 // Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But
227 // that's ok since the histogram values are less than 1<<28 (max picture size).
HistogramAdd_SSE2(const VP8LHistogram * const a,const VP8LHistogram * const b,VP8LHistogram * const out)228 static void HistogramAdd_SSE2(const VP8LHistogram* const a,
229 const VP8LHistogram* const b,
230 VP8LHistogram* const out) {
231 int i;
232 const int literal_size = VP8LHistogramNumCodes(a->palette_code_bits_);
233 assert(a->palette_code_bits_ == b->palette_code_bits_);
234 if (b != out) {
235 AddVector_SSE2(a->literal_, b->literal_, out->literal_, NUM_LITERAL_CODES);
236 AddVector_SSE2(a->red_, b->red_, out->red_, NUM_LITERAL_CODES);
237 AddVector_SSE2(a->blue_, b->blue_, out->blue_, NUM_LITERAL_CODES);
238 AddVector_SSE2(a->alpha_, b->alpha_, out->alpha_, NUM_LITERAL_CODES);
239 } else {
240 AddVectorEq_SSE2(a->literal_, out->literal_, NUM_LITERAL_CODES);
241 AddVectorEq_SSE2(a->red_, out->red_, NUM_LITERAL_CODES);
242 AddVectorEq_SSE2(a->blue_, out->blue_, NUM_LITERAL_CODES);
243 AddVectorEq_SSE2(a->alpha_, out->alpha_, NUM_LITERAL_CODES);
244 }
245 for (i = NUM_LITERAL_CODES; i < literal_size; ++i) {
246 out->literal_[i] = a->literal_[i] + b->literal_[i];
247 }
248 for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
249 out->distance_[i] = a->distance_[i] + b->distance_[i];
250 }
251 }
252
253 //------------------------------------------------------------------------------
254 // Entropy
255
256 // Checks whether the X or Y contribution is worth computing and adding.
257 // Used in loop unrolling.
258 #define ANALYZE_X_OR_Y(x_or_y, j) \
259 do { \
260 if ((x_or_y)[i + (j)] != 0) retval -= VP8LFastSLog2((x_or_y)[i + (j)]); \
261 } while (0)
262
263 // Checks whether the X + Y contribution is worth computing and adding.
264 // Used in loop unrolling.
265 #define ANALYZE_XY(j) \
266 do { \
267 if (tmp[j] != 0) { \
268 retval -= VP8LFastSLog2(tmp[j]); \
269 ANALYZE_X_OR_Y(X, j); \
270 } \
271 } while (0)
272
CombinedShannonEntropy_SSE2(const int X[256],const int Y[256])273 static float CombinedShannonEntropy_SSE2(const int X[256], const int Y[256]) {
274 int i;
275 double retval = 0.;
276 int sumX, sumXY;
277 int32_t tmp[4];
278 __m128i zero = _mm_setzero_si128();
279 // Sums up X + Y, 4 ints at a time (and will merge it at the end for sumXY).
280 __m128i sumXY_128 = zero;
281 __m128i sumX_128 = zero;
282
283 for (i = 0; i < 256; i += 4) {
284 const __m128i x = _mm_loadu_si128((const __m128i*)(X + i));
285 const __m128i y = _mm_loadu_si128((const __m128i*)(Y + i));
286
287 // Check if any X is non-zero: this actually provides a speedup as X is
288 // usually sparse.
289 if (_mm_movemask_epi8(_mm_cmpeq_epi32(x, zero)) != 0xFFFF) {
290 const __m128i xy_128 = _mm_add_epi32(x, y);
291 sumXY_128 = _mm_add_epi32(sumXY_128, xy_128);
292
293 sumX_128 = _mm_add_epi32(sumX_128, x);
294
295 // Analyze the different X + Y.
296 _mm_storeu_si128((__m128i*)tmp, xy_128);
297
298 ANALYZE_XY(0);
299 ANALYZE_XY(1);
300 ANALYZE_XY(2);
301 ANALYZE_XY(3);
302 } else {
303 // X is fully 0, so only deal with Y.
304 sumXY_128 = _mm_add_epi32(sumXY_128, y);
305
306 ANALYZE_X_OR_Y(Y, 0);
307 ANALYZE_X_OR_Y(Y, 1);
308 ANALYZE_X_OR_Y(Y, 2);
309 ANALYZE_X_OR_Y(Y, 3);
310 }
311 }
312
313 // Sum up sumX_128 to get sumX.
314 _mm_storeu_si128((__m128i*)tmp, sumX_128);
315 sumX = tmp[3] + tmp[2] + tmp[1] + tmp[0];
316
317 // Sum up sumXY_128 to get sumXY.
318 _mm_storeu_si128((__m128i*)tmp, sumXY_128);
319 sumXY = tmp[3] + tmp[2] + tmp[1] + tmp[0];
320
321 retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY);
322 return (float)retval;
323 }
324 #undef ANALYZE_X_OR_Y
325 #undef ANALYZE_XY
326
327 //------------------------------------------------------------------------------
328
VectorMismatch_SSE2(const uint32_t * const array1,const uint32_t * const array2,int length)329 static int VectorMismatch_SSE2(const uint32_t* const array1,
330 const uint32_t* const array2, int length) {
331 int match_len;
332
333 if (length >= 12) {
334 __m128i A0 = _mm_loadu_si128((const __m128i*)&array1[0]);
335 __m128i A1 = _mm_loadu_si128((const __m128i*)&array2[0]);
336 match_len = 0;
337 do {
338 // Loop unrolling and early load both provide a speedup of 10% for the
339 // current function. Also, max_limit can be MAX_LENGTH=4096 at most.
340 const __m128i cmpA = _mm_cmpeq_epi32(A0, A1);
341 const __m128i B0 =
342 _mm_loadu_si128((const __m128i*)&array1[match_len + 4]);
343 const __m128i B1 =
344 _mm_loadu_si128((const __m128i*)&array2[match_len + 4]);
345 if (_mm_movemask_epi8(cmpA) != 0xffff) break;
346 match_len += 4;
347
348 {
349 const __m128i cmpB = _mm_cmpeq_epi32(B0, B1);
350 A0 = _mm_loadu_si128((const __m128i*)&array1[match_len + 4]);
351 A1 = _mm_loadu_si128((const __m128i*)&array2[match_len + 4]);
352 if (_mm_movemask_epi8(cmpB) != 0xffff) break;
353 match_len += 4;
354 }
355 } while (match_len + 12 < length);
356 } else {
357 match_len = 0;
358 // Unroll the potential first two loops.
359 if (length >= 4 &&
360 _mm_movemask_epi8(_mm_cmpeq_epi32(
361 _mm_loadu_si128((const __m128i*)&array1[0]),
362 _mm_loadu_si128((const __m128i*)&array2[0]))) == 0xffff) {
363 match_len = 4;
364 if (length >= 8 &&
365 _mm_movemask_epi8(_mm_cmpeq_epi32(
366 _mm_loadu_si128((const __m128i*)&array1[4]),
367 _mm_loadu_si128((const __m128i*)&array2[4]))) == 0xffff) {
368 match_len = 8;
369 }
370 }
371 }
372
373 while (match_len < length && array1[match_len] == array2[match_len]) {
374 ++match_len;
375 }
376 return match_len;
377 }
378
379 // Bundles multiple (1, 2, 4 or 8) pixels into a single pixel.
BundleColorMap_SSE2(const uint8_t * const row,int width,int xbits,uint32_t * dst)380 static void BundleColorMap_SSE2(const uint8_t* const row, int width, int xbits,
381 uint32_t* dst) {
382 int x;
383 assert(xbits >= 0);
384 assert(xbits <= 3);
385 switch (xbits) {
386 case 0: {
387 const __m128i ff = _mm_set1_epi16(0xff00);
388 const __m128i zero = _mm_setzero_si128();
389 // Store 0xff000000 | (row[x] << 8).
390 for (x = 0; x + 16 <= width; x += 16, dst += 16) {
391 const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
392 const __m128i in_lo = _mm_unpacklo_epi8(zero, in);
393 const __m128i dst0 = _mm_unpacklo_epi16(in_lo, ff);
394 const __m128i dst1 = _mm_unpackhi_epi16(in_lo, ff);
395 const __m128i in_hi = _mm_unpackhi_epi8(zero, in);
396 const __m128i dst2 = _mm_unpacklo_epi16(in_hi, ff);
397 const __m128i dst3 = _mm_unpackhi_epi16(in_hi, ff);
398 _mm_storeu_si128((__m128i*)&dst[0], dst0);
399 _mm_storeu_si128((__m128i*)&dst[4], dst1);
400 _mm_storeu_si128((__m128i*)&dst[8], dst2);
401 _mm_storeu_si128((__m128i*)&dst[12], dst3);
402 }
403 break;
404 }
405 case 1: {
406 const __m128i ff = _mm_set1_epi16(0xff00);
407 const __m128i mul = _mm_set1_epi16(0x110);
408 for (x = 0; x + 16 <= width; x += 16, dst += 8) {
409 // 0a0b | (where a/b are 4 bits).
410 const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
411 const __m128i tmp = _mm_mullo_epi16(in, mul); // aba0
412 const __m128i pack = _mm_and_si128(tmp, ff); // ab00
413 const __m128i dst0 = _mm_unpacklo_epi16(pack, ff);
414 const __m128i dst1 = _mm_unpackhi_epi16(pack, ff);
415 _mm_storeu_si128((__m128i*)&dst[0], dst0);
416 _mm_storeu_si128((__m128i*)&dst[4], dst1);
417 }
418 break;
419 }
420 case 2: {
421 const __m128i mask_or = _mm_set1_epi32(0xff000000);
422 const __m128i mul_cst = _mm_set1_epi16(0x0104);
423 const __m128i mask_mul = _mm_set1_epi16(0x0f00);
424 for (x = 0; x + 16 <= width; x += 16, dst += 4) {
425 // 000a000b000c000d | (where a/b/c/d are 2 bits).
426 const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
427 const __m128i mul = _mm_mullo_epi16(in, mul_cst); // 00ab00b000cd00d0
428 const __m128i tmp = _mm_and_si128(mul, mask_mul); // 00ab000000cd0000
429 const __m128i shift = _mm_srli_epi32(tmp, 12); // 00000000ab000000
430 const __m128i pack = _mm_or_si128(shift, tmp); // 00000000abcd0000
431 // Convert to 0xff00**00.
432 const __m128i res = _mm_or_si128(pack, mask_or);
433 _mm_storeu_si128((__m128i*)dst, res);
434 }
435 break;
436 }
437 default: {
438 assert(xbits == 3);
439 for (x = 0; x + 16 <= width; x += 16, dst += 2) {
440 // 0000000a00000000b... | (where a/b are 1 bit).
441 const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
442 const __m128i shift = _mm_slli_epi64(in, 7);
443 const uint32_t move = _mm_movemask_epi8(shift);
444 dst[0] = 0xff000000 | ((move & 0xff) << 8);
445 dst[1] = 0xff000000 | (move & 0xff00);
446 }
447 break;
448 }
449 }
450 if (x != width) {
451 VP8LBundleColorMap_C(row + x, width - x, xbits, dst);
452 }
453 }
454
455 //------------------------------------------------------------------------------
456 // Batch version of Predictor Transform subtraction
457
Average2_m128i(const __m128i * const a0,const __m128i * const a1,__m128i * const avg)458 static WEBP_INLINE void Average2_m128i(const __m128i* const a0,
459 const __m128i* const a1,
460 __m128i* const avg) {
461 // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1)
462 const __m128i ones = _mm_set1_epi8(1);
463 const __m128i avg1 = _mm_avg_epu8(*a0, *a1);
464 const __m128i one = _mm_and_si128(_mm_xor_si128(*a0, *a1), ones);
465 *avg = _mm_sub_epi8(avg1, one);
466 }
467
468 // Predictor0: ARGB_BLACK.
PredictorSub0_SSE2(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)469 static void PredictorSub0_SSE2(const uint32_t* in, const uint32_t* upper,
470 int num_pixels, uint32_t* out) {
471 int i;
472 const __m128i black = _mm_set1_epi32(ARGB_BLACK);
473 for (i = 0; i + 4 <= num_pixels; i += 4) {
474 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
475 const __m128i res = _mm_sub_epi8(src, black);
476 _mm_storeu_si128((__m128i*)&out[i], res);
477 }
478 if (i != num_pixels) {
479 VP8LPredictorsSub_C[0](in + i, upper + i, num_pixels - i, out + i);
480 }
481 }
482
483 #define GENERATE_PREDICTOR_1(X, IN) \
484 static void PredictorSub##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
485 int num_pixels, uint32_t* out) { \
486 int i; \
487 for (i = 0; i + 4 <= num_pixels; i += 4) { \
488 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \
489 const __m128i pred = _mm_loadu_si128((const __m128i*)&(IN)); \
490 const __m128i res = _mm_sub_epi8(src, pred); \
491 _mm_storeu_si128((__m128i*)&out[i], res); \
492 } \
493 if (i != num_pixels) { \
494 VP8LPredictorsSub_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
495 } \
496 }
497
498 GENERATE_PREDICTOR_1(1, in[i - 1]) // Predictor1: L
499 GENERATE_PREDICTOR_1(2, upper[i]) // Predictor2: T
500 GENERATE_PREDICTOR_1(3, upper[i + 1]) // Predictor3: TR
501 GENERATE_PREDICTOR_1(4, upper[i - 1]) // Predictor4: TL
502 #undef GENERATE_PREDICTOR_1
503
504 // Predictor5: avg2(avg2(L, TR), T)
PredictorSub5_SSE2(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)505 static void PredictorSub5_SSE2(const uint32_t* in, const uint32_t* upper,
506 int num_pixels, uint32_t* out) {
507 int i;
508 for (i = 0; i + 4 <= num_pixels; i += 4) {
509 const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
510 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
511 const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]);
512 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
513 __m128i avg, pred, res;
514 Average2_m128i(&L, &TR, &avg);
515 Average2_m128i(&avg, &T, &pred);
516 res = _mm_sub_epi8(src, pred);
517 _mm_storeu_si128((__m128i*)&out[i], res);
518 }
519 if (i != num_pixels) {
520 VP8LPredictorsSub_C[5](in + i, upper + i, num_pixels - i, out + i);
521 }
522 }
523
524 #define GENERATE_PREDICTOR_2(X, A, B) \
525 static void PredictorSub##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
526 int num_pixels, uint32_t* out) { \
527 int i; \
528 for (i = 0; i + 4 <= num_pixels; i += 4) { \
529 const __m128i tA = _mm_loadu_si128((const __m128i*)&(A)); \
530 const __m128i tB = _mm_loadu_si128((const __m128i*)&(B)); \
531 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \
532 __m128i pred, res; \
533 Average2_m128i(&tA, &tB, &pred); \
534 res = _mm_sub_epi8(src, pred); \
535 _mm_storeu_si128((__m128i*)&out[i], res); \
536 } \
537 if (i != num_pixels) { \
538 VP8LPredictorsSub_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
539 } \
540 }
541
542 GENERATE_PREDICTOR_2(6, in[i - 1], upper[i - 1]) // Predictor6: avg(L, TL)
543 GENERATE_PREDICTOR_2(7, in[i - 1], upper[i]) // Predictor7: avg(L, T)
544 GENERATE_PREDICTOR_2(8, upper[i - 1], upper[i]) // Predictor8: avg(TL, T)
545 GENERATE_PREDICTOR_2(9, upper[i], upper[i + 1]) // Predictor9: average(T, TR)
546 #undef GENERATE_PREDICTOR_2
547
548 // Predictor10: avg(avg(L,TL), avg(T, TR)).
PredictorSub10_SSE2(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)549 static void PredictorSub10_SSE2(const uint32_t* in, const uint32_t* upper,
550 int num_pixels, uint32_t* out) {
551 int i;
552 for (i = 0; i + 4 <= num_pixels; i += 4) {
553 const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
554 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
555 const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
556 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
557 const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]);
558 __m128i avgTTR, avgLTL, avg, res;
559 Average2_m128i(&T, &TR, &avgTTR);
560 Average2_m128i(&L, &TL, &avgLTL);
561 Average2_m128i(&avgTTR, &avgLTL, &avg);
562 res = _mm_sub_epi8(src, avg);
563 _mm_storeu_si128((__m128i*)&out[i], res);
564 }
565 if (i != num_pixels) {
566 VP8LPredictorsSub_C[10](in + i, upper + i, num_pixels - i, out + i);
567 }
568 }
569
570 // Predictor11: select.
GetSumAbsDiff32_SSE2(const __m128i * const A,const __m128i * const B,__m128i * const out)571 static void GetSumAbsDiff32_SSE2(const __m128i* const A, const __m128i* const B,
572 __m128i* const out) {
573 // We can unpack with any value on the upper 32 bits, provided it's the same
574 // on both operands (to that their sum of abs diff is zero). Here we use *A.
575 const __m128i A_lo = _mm_unpacklo_epi32(*A, *A);
576 const __m128i B_lo = _mm_unpacklo_epi32(*B, *A);
577 const __m128i A_hi = _mm_unpackhi_epi32(*A, *A);
578 const __m128i B_hi = _mm_unpackhi_epi32(*B, *A);
579 const __m128i s_lo = _mm_sad_epu8(A_lo, B_lo);
580 const __m128i s_hi = _mm_sad_epu8(A_hi, B_hi);
581 *out = _mm_packs_epi32(s_lo, s_hi);
582 }
583
PredictorSub11_SSE2(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)584 static void PredictorSub11_SSE2(const uint32_t* in, const uint32_t* upper,
585 int num_pixels, uint32_t* out) {
586 int i;
587 for (i = 0; i + 4 <= num_pixels; i += 4) {
588 const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
589 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
590 const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
591 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
592 __m128i pa, pb;
593 GetSumAbsDiff32_SSE2(&T, &TL, &pa); // pa = sum |T-TL|
594 GetSumAbsDiff32_SSE2(&L, &TL, &pb); // pb = sum |L-TL|
595 {
596 const __m128i mask = _mm_cmpgt_epi32(pb, pa);
597 const __m128i A = _mm_and_si128(mask, L);
598 const __m128i B = _mm_andnot_si128(mask, T);
599 const __m128i pred = _mm_or_si128(A, B); // pred = (L > T)? L : T
600 const __m128i res = _mm_sub_epi8(src, pred);
601 _mm_storeu_si128((__m128i*)&out[i], res);
602 }
603 }
604 if (i != num_pixels) {
605 VP8LPredictorsSub_C[11](in + i, upper + i, num_pixels - i, out + i);
606 }
607 }
608
609 // Predictor12: ClampedSubSubtractFull.
PredictorSub12_SSE2(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)610 static void PredictorSub12_SSE2(const uint32_t* in, const uint32_t* upper,
611 int num_pixels, uint32_t* out) {
612 int i;
613 const __m128i zero = _mm_setzero_si128();
614 for (i = 0; i + 4 <= num_pixels; i += 4) {
615 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
616 const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
617 const __m128i L_lo = _mm_unpacklo_epi8(L, zero);
618 const __m128i L_hi = _mm_unpackhi_epi8(L, zero);
619 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
620 const __m128i T_lo = _mm_unpacklo_epi8(T, zero);
621 const __m128i T_hi = _mm_unpackhi_epi8(T, zero);
622 const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
623 const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero);
624 const __m128i TL_hi = _mm_unpackhi_epi8(TL, zero);
625 const __m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo);
626 const __m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi);
627 const __m128i pred_lo = _mm_add_epi16(L_lo, diff_lo);
628 const __m128i pred_hi = _mm_add_epi16(L_hi, diff_hi);
629 const __m128i pred = _mm_packus_epi16(pred_lo, pred_hi);
630 const __m128i res = _mm_sub_epi8(src, pred);
631 _mm_storeu_si128((__m128i*)&out[i], res);
632 }
633 if (i != num_pixels) {
634 VP8LPredictorsSub_C[12](in + i, upper + i, num_pixels - i, out + i);
635 }
636 }
637
638 // Predictors13: ClampedAddSubtractHalf
PredictorSub13_SSE2(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)639 static void PredictorSub13_SSE2(const uint32_t* in, const uint32_t* upper,
640 int num_pixels, uint32_t* out) {
641 int i;
642 const __m128i zero = _mm_setzero_si128();
643 for (i = 0; i + 2 <= num_pixels; i += 2) {
644 // we can only process two pixels at a time
645 const __m128i L = _mm_loadl_epi64((const __m128i*)&in[i - 1]);
646 const __m128i src = _mm_loadl_epi64((const __m128i*)&in[i]);
647 const __m128i T = _mm_loadl_epi64((const __m128i*)&upper[i]);
648 const __m128i TL = _mm_loadl_epi64((const __m128i*)&upper[i - 1]);
649 const __m128i L_lo = _mm_unpacklo_epi8(L, zero);
650 const __m128i T_lo = _mm_unpacklo_epi8(T, zero);
651 const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero);
652 const __m128i sum = _mm_add_epi16(T_lo, L_lo);
653 const __m128i avg = _mm_srli_epi16(sum, 1);
654 const __m128i A1 = _mm_sub_epi16(avg, TL_lo);
655 const __m128i bit_fix = _mm_cmpgt_epi16(TL_lo, avg);
656 const __m128i A2 = _mm_sub_epi16(A1, bit_fix);
657 const __m128i A3 = _mm_srai_epi16(A2, 1);
658 const __m128i A4 = _mm_add_epi16(avg, A3);
659 const __m128i pred = _mm_packus_epi16(A4, A4);
660 const __m128i res = _mm_sub_epi8(src, pred);
661 _mm_storel_epi64((__m128i*)&out[i], res);
662 }
663 if (i != num_pixels) {
664 VP8LPredictorsSub_C[13](in + i, upper + i, num_pixels - i, out + i);
665 }
666 }
667
668 //------------------------------------------------------------------------------
669 // Entry point
670
671 extern void VP8LEncDspInitSSE2(void);
672
VP8LEncDspInitSSE2(void)673 WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) {
674 VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed_SSE2;
675 VP8LTransformColor = TransformColor_SSE2;
676 VP8LCollectColorBlueTransforms = CollectColorBlueTransforms_SSE2;
677 VP8LCollectColorRedTransforms = CollectColorRedTransforms_SSE2;
678 VP8LHistogramAdd = HistogramAdd_SSE2;
679 VP8LCombinedShannonEntropy = CombinedShannonEntropy_SSE2;
680 VP8LVectorMismatch = VectorMismatch_SSE2;
681 VP8LBundleColorMap = BundleColorMap_SSE2;
682
683 VP8LPredictorsSub[0] = PredictorSub0_SSE2;
684 VP8LPredictorsSub[1] = PredictorSub1_SSE2;
685 VP8LPredictorsSub[2] = PredictorSub2_SSE2;
686 VP8LPredictorsSub[3] = PredictorSub3_SSE2;
687 VP8LPredictorsSub[4] = PredictorSub4_SSE2;
688 VP8LPredictorsSub[5] = PredictorSub5_SSE2;
689 VP8LPredictorsSub[6] = PredictorSub6_SSE2;
690 VP8LPredictorsSub[7] = PredictorSub7_SSE2;
691 VP8LPredictorsSub[8] = PredictorSub8_SSE2;
692 VP8LPredictorsSub[9] = PredictorSub9_SSE2;
693 VP8LPredictorsSub[10] = PredictorSub10_SSE2;
694 VP8LPredictorsSub[11] = PredictorSub11_SSE2;
695 VP8LPredictorsSub[12] = PredictorSub12_SSE2;
696 VP8LPredictorsSub[13] = PredictorSub13_SSE2;
697 VP8LPredictorsSub[14] = PredictorSub0_SSE2; // <- padding security sentinels
698 VP8LPredictorsSub[15] = PredictorSub0_SSE2;
699 }
700
701 #else // !WEBP_USE_SSE2
702
703 WEBP_DSP_INIT_STUB(VP8LEncDspInitSSE2)
704
705 #endif // WEBP_USE_SSE2
706