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 Rescaling functions
11//
12// Author: Skal (pascal.massimino@gmail.com)
13
14#include "src/dsp/dsp.h"
15
16#if defined(WEBP_USE_SSE2) && !defined(WEBP_REDUCE_SIZE)
17#include <emmintrin.h>
18
19#include <assert.h>
20#include "src/utils/rescaler_utils.h"
21#include "src/utils/utils.h"
22
23//------------------------------------------------------------------------------
24// Implementations of critical functions ImportRow / ExportRow
25
26#define ROUNDER (WEBP_RESCALER_ONE >> 1)
27#define MULT_FIX(x, y) (((uint64_t)(x) * (y) + ROUNDER) >> WEBP_RESCALER_RFIX)
28#define MULT_FIX_FLOOR(x, y) (((uint64_t)(x) * (y)) >> WEBP_RESCALER_RFIX)
29
30// input: 8 bytes ABCDEFGH -> output: A0E0B0F0C0G0D0H0
31static void LoadTwoPixels_SSE2(const uint8_t* const src, __m128i* out) {
32 const __m128i zero = _mm_setzero_si128();
33 const __m128i A = _mm_loadl_epi64((const __m128i*)(src)); // ABCDEFGH
34 const __m128i B = _mm_unpacklo_epi8(A, zero); // A0B0C0D0E0F0G0H0
35 const __m128i C = _mm_srli_si128(B, 8); // E0F0G0H0
36 *out = _mm_unpacklo_epi16(B, C);
37}
38
39// input: 8 bytes ABCDEFGH -> output: A0B0C0D0E0F0G0H0
40static void LoadEightPixels_SSE2(const uint8_t* const src, __m128i* out) {
41 const __m128i zero = _mm_setzero_si128();
42 const __m128i A = _mm_loadl_epi64((const __m128i*)(src)); // ABCDEFGH
43 *out = _mm_unpacklo_epi8(A, zero);
44}
45
46static void RescalerImportRowExpand_SSE2(WebPRescaler* const wrk,
47 const uint8_t* src) {
48 rescaler_t* frow = wrk->frow;
49 const rescaler_t* const frow_end = frow + wrk->dst_width * wrk->num_channels;
50 const int x_add = wrk->x_add;
51 int accum = x_add;
52 __m128i cur_pixels;
53
54 // SSE2 implementation only works with 16b signed arithmetic at max.
55 if (wrk->src_width < 8 || accum >= (1 << 15)) {
56 WebPRescalerImportRowExpand_C(wrk, src);
57 return;
58 }
59
60 assert(!WebPRescalerInputDone(wrk));
61 assert(wrk->x_expand);
62 if (wrk->num_channels == 4) {
63 LoadTwoPixels_SSE2(src, &cur_pixels);
64 src += 4;
65 while (1) {
66 const __m128i mult = _mm_set1_epi32(((x_add - accum) << 16) | accum);
67 const __m128i out = _mm_madd_epi16(cur_pixels, mult);
68 _mm_storeu_si128((__m128i*)frow, out);
69 frow += 4;
70 if (frow >= frow_end) break;
71 accum -= wrk->x_sub;
72 if (accum < 0) {
73 LoadTwoPixels_SSE2(src, &cur_pixels);
74 src += 4;
75 accum += x_add;
76 }
77 }
78 } else {
79 int left;
80 const uint8_t* const src_limit = src + wrk->src_width - 8;
81 LoadEightPixels_SSE2(src, &cur_pixels);
82 src += 7;
83 left = 7;
84 while (1) {
85 const __m128i mult = _mm_cvtsi32_si128(((x_add - accum) << 16) | accum);
86 const __m128i out = _mm_madd_epi16(cur_pixels, mult);
87 assert(sizeof(*frow) == sizeof(uint32_t));
88 WebPUint32ToMem((uint8_t*)frow, _mm_cvtsi128_si32(out));
89 frow += 1;
90 if (frow >= frow_end) break;
91 accum -= wrk->x_sub;
92 if (accum < 0) {
93 if (--left) {
94 cur_pixels = _mm_srli_si128(cur_pixels, 2);
95 } else if (src <= src_limit) {
96 LoadEightPixels_SSE2(src, &cur_pixels);
97 src += 7;
98 left = 7;
99 } else { // tail
100 cur_pixels = _mm_srli_si128(cur_pixels, 2);
101 cur_pixels = _mm_insert_epi16(cur_pixels, src[1], 1);
102 src += 1;
103 left = 1;
104 }
105 accum += x_add;
106 }
107 }
108 }
109 assert(accum == 0);
110}
111
112static void RescalerImportRowShrink_SSE2(WebPRescaler* const wrk,
113 const uint8_t* src) {
114 const int x_sub = wrk->x_sub;
115 int accum = 0;
116 const __m128i zero = _mm_setzero_si128();
117 const __m128i mult0 = _mm_set1_epi16(x_sub);
118 const __m128i mult1 = _mm_set1_epi32(wrk->fx_scale);
119 const __m128i rounder = _mm_set_epi32(0, ROUNDER, 0, ROUNDER);
120 __m128i sum = zero;
121 rescaler_t* frow = wrk->frow;
122 const rescaler_t* const frow_end = wrk->frow + 4 * wrk->dst_width;
123
124 if (wrk->num_channels != 4 || wrk->x_add > (x_sub << 7)) {
125 WebPRescalerImportRowShrink_C(wrk, src);
126 return;
127 }
128 assert(!WebPRescalerInputDone(wrk));
129 assert(!wrk->x_expand);
130
131 for (; frow < frow_end; frow += 4) {
132 __m128i base = zero;
133 accum += wrk->x_add;
134 while (accum > 0) {
135 const __m128i A = _mm_cvtsi32_si128(WebPMemToUint32(src));
136 src += 4;
137 base = _mm_unpacklo_epi8(A, zero);
138 // To avoid overflow, we need: base * x_add / x_sub < 32768
139 // => x_add < x_sub << 7. That's a 1/128 reduction ratio limit.
140 sum = _mm_add_epi16(sum, base);
141 accum -= x_sub;
142 }
143 { // Emit next horizontal pixel.
144 const __m128i mult = _mm_set1_epi16(-accum);
145 const __m128i frac0 = _mm_mullo_epi16(base, mult); // 16b x 16b -> 32b
146 const __m128i frac1 = _mm_mulhi_epu16(base, mult);
147 const __m128i frac = _mm_unpacklo_epi16(frac0, frac1); // frac is 32b
148 const __m128i A0 = _mm_mullo_epi16(sum, mult0);
149 const __m128i A1 = _mm_mulhi_epu16(sum, mult0);
150 const __m128i B0 = _mm_unpacklo_epi16(A0, A1); // sum * x_sub
151 const __m128i frow_out = _mm_sub_epi32(B0, frac); // sum * x_sub - frac
152 const __m128i D0 = _mm_srli_epi64(frac, 32);
153 const __m128i D1 = _mm_mul_epu32(frac, mult1); // 32b x 16b -> 64b
154 const __m128i D2 = _mm_mul_epu32(D0, mult1);
155 const __m128i E1 = _mm_add_epi64(D1, rounder);
156 const __m128i E2 = _mm_add_epi64(D2, rounder);
157 const __m128i F1 = _mm_shuffle_epi32(E1, 1 | (3 << 2));
158 const __m128i F2 = _mm_shuffle_epi32(E2, 1 | (3 << 2));
159 const __m128i G = _mm_unpacklo_epi32(F1, F2);
160 sum = _mm_packs_epi32(G, zero);
161 _mm_storeu_si128((__m128i*)frow, frow_out);
162 }
163 }
164 assert(accum == 0);
165}
166
167//------------------------------------------------------------------------------
168// Row export
169
170// load *src as epi64, multiply by mult and store result in [out0 ... out3]
171static WEBP_INLINE void LoadDispatchAndMult_SSE2(const rescaler_t* const src,
172 const __m128i* const mult,
173 __m128i* const out0,
174 __m128i* const out1,
175 __m128i* const out2,
176 __m128i* const out3) {
177 const __m128i A0 = _mm_loadu_si128((const __m128i*)(src + 0));
178 const __m128i A1 = _mm_loadu_si128((const __m128i*)(src + 4));
179 const __m128i A2 = _mm_srli_epi64(A0, 32);
180 const __m128i A3 = _mm_srli_epi64(A1, 32);
181 if (mult != NULL) {
182 *out0 = _mm_mul_epu32(A0, *mult);
183 *out1 = _mm_mul_epu32(A1, *mult);
184 *out2 = _mm_mul_epu32(A2, *mult);
185 *out3 = _mm_mul_epu32(A3, *mult);
186 } else {
187 *out0 = A0;
188 *out1 = A1;
189 *out2 = A2;
190 *out3 = A3;
191 }
192}
193
194static WEBP_INLINE void ProcessRow_SSE2(const __m128i* const A0,
195 const __m128i* const A1,
196 const __m128i* const A2,
197 const __m128i* const A3,
198 const __m128i* const mult,
199 uint8_t* const dst) {
200 const __m128i rounder = _mm_set_epi32(0, ROUNDER, 0, ROUNDER);
201 const __m128i mask = _mm_set_epi32(0xffffffffu, 0, 0xffffffffu, 0);
202 const __m128i B0 = _mm_mul_epu32(*A0, *mult);
203 const __m128i B1 = _mm_mul_epu32(*A1, *mult);
204 const __m128i B2 = _mm_mul_epu32(*A2, *mult);
205 const __m128i B3 = _mm_mul_epu32(*A3, *mult);
206 const __m128i C0 = _mm_add_epi64(B0, rounder);
207 const __m128i C1 = _mm_add_epi64(B1, rounder);
208 const __m128i C2 = _mm_add_epi64(B2, rounder);
209 const __m128i C3 = _mm_add_epi64(B3, rounder);
210 const __m128i D0 = _mm_srli_epi64(C0, WEBP_RESCALER_RFIX);
211 const __m128i D1 = _mm_srli_epi64(C1, WEBP_RESCALER_RFIX);
212#if (WEBP_RESCALER_RFIX < 32)
213 const __m128i D2 =
214 _mm_and_si128(_mm_slli_epi64(C2, 32 - WEBP_RESCALER_RFIX), mask);
215 const __m128i D3 =
216 _mm_and_si128(_mm_slli_epi64(C3, 32 - WEBP_RESCALER_RFIX), mask);
217#else
218 const __m128i D2 = _mm_and_si128(C2, mask);
219 const __m128i D3 = _mm_and_si128(C3, mask);
220#endif
221 const __m128i E0 = _mm_or_si128(D0, D2);
222 const __m128i E1 = _mm_or_si128(D1, D3);
223 const __m128i F = _mm_packs_epi32(E0, E1);
224 const __m128i G = _mm_packus_epi16(F, F);
225 _mm_storel_epi64((__m128i*)dst, G);
226}
227
228static void RescalerExportRowExpand_SSE2(WebPRescaler* const wrk) {
229 int x_out;
230 uint8_t* const dst = wrk->dst;
231 rescaler_t* const irow = wrk->irow;
232 const int x_out_max = wrk->dst_width * wrk->num_channels;
233 const rescaler_t* const frow = wrk->frow;
234 const __m128i mult = _mm_set_epi32(0, wrk->fy_scale, 0, wrk->fy_scale);
235
236 assert(!WebPRescalerOutputDone(wrk));
237 assert(wrk->y_accum <= 0 && wrk->y_sub + wrk->y_accum >= 0);
238 assert(wrk->y_expand);
239 if (wrk->y_accum == 0) {
240 for (x_out = 0; x_out + 8 <= x_out_max; x_out += 8) {
241 __m128i A0, A1, A2, A3;
242 LoadDispatchAndMult_SSE2(frow + x_out, NULL, &A0, &A1, &A2, &A3);
243 ProcessRow_SSE2(&A0, &A1, &A2, &A3, &mult, dst + x_out);
244 }
245 for (; x_out < x_out_max; ++x_out) {
246 const uint32_t J = frow[x_out];
247 const int v = (int)MULT_FIX(J, wrk->fy_scale);
248 dst[x_out] = (v > 255) ? 255u : (uint8_t)v;
249 }
250 } else {
251 const uint32_t B = WEBP_RESCALER_FRAC(-wrk->y_accum, wrk->y_sub);
252 const uint32_t A = (uint32_t)(WEBP_RESCALER_ONE - B);
253 const __m128i mA = _mm_set_epi32(0, A, 0, A);
254 const __m128i mB = _mm_set_epi32(0, B, 0, B);
255 const __m128i rounder = _mm_set_epi32(0, ROUNDER, 0, ROUNDER);
256 for (x_out = 0; x_out + 8 <= x_out_max; x_out += 8) {
257 __m128i A0, A1, A2, A3, B0, B1, B2, B3;
258 LoadDispatchAndMult_SSE2(frow + x_out, &mA, &A0, &A1, &A2, &A3);
259 LoadDispatchAndMult_SSE2(irow + x_out, &mB, &B0, &B1, &B2, &B3);
260 {
261 const __m128i C0 = _mm_add_epi64(A0, B0);
262 const __m128i C1 = _mm_add_epi64(A1, B1);
263 const __m128i C2 = _mm_add_epi64(A2, B2);
264 const __m128i C3 = _mm_add_epi64(A3, B3);
265 const __m128i D0 = _mm_add_epi64(C0, rounder);
266 const __m128i D1 = _mm_add_epi64(C1, rounder);
267 const __m128i D2 = _mm_add_epi64(C2, rounder);
268 const __m128i D3 = _mm_add_epi64(C3, rounder);
269 const __m128i E0 = _mm_srli_epi64(D0, WEBP_RESCALER_RFIX);
270 const __m128i E1 = _mm_srli_epi64(D1, WEBP_RESCALER_RFIX);
271 const __m128i E2 = _mm_srli_epi64(D2, WEBP_RESCALER_RFIX);
272 const __m128i E3 = _mm_srli_epi64(D3, WEBP_RESCALER_RFIX);
273 ProcessRow_SSE2(&E0, &E1, &E2, &E3, &mult, dst + x_out);
274 }
275 }
276 for (; x_out < x_out_max; ++x_out) {
277 const uint64_t I = (uint64_t)A * frow[x_out]
278 + (uint64_t)B * irow[x_out];
279 const uint32_t J = (uint32_t)((I + ROUNDER) >> WEBP_RESCALER_RFIX);
280 const int v = (int)MULT_FIX(J, wrk->fy_scale);
281 dst[x_out] = (v > 255) ? 255u : (uint8_t)v;
282 }
283 }
284}
285
286static void RescalerExportRowShrink_SSE2(WebPRescaler* const wrk) {
287 int x_out;
288 uint8_t* const dst = wrk->dst;
289 rescaler_t* const irow = wrk->irow;
290 const int x_out_max = wrk->dst_width * wrk->num_channels;
291 const rescaler_t* const frow = wrk->frow;
292 const uint32_t yscale = wrk->fy_scale * (-wrk->y_accum);
293 assert(!WebPRescalerOutputDone(wrk));
294 assert(wrk->y_accum <= 0);
295 assert(!wrk->y_expand);
296 if (yscale) {
297 const int scale_xy = wrk->fxy_scale;
298 const __m128i mult_xy = _mm_set_epi32(0, scale_xy, 0, scale_xy);
299 const __m128i mult_y = _mm_set_epi32(0, yscale, 0, yscale);
300 for (x_out = 0; x_out + 8 <= x_out_max; x_out += 8) {
301 __m128i A0, A1, A2, A3, B0, B1, B2, B3;
302 LoadDispatchAndMult_SSE2(irow + x_out, NULL, &A0, &A1, &A2, &A3);
303 LoadDispatchAndMult_SSE2(frow + x_out, &mult_y, &B0, &B1, &B2, &B3);
304 {
305 const __m128i D0 = _mm_srli_epi64(B0, WEBP_RESCALER_RFIX); // = frac
306 const __m128i D1 = _mm_srli_epi64(B1, WEBP_RESCALER_RFIX);
307 const __m128i D2 = _mm_srli_epi64(B2, WEBP_RESCALER_RFIX);
308 const __m128i D3 = _mm_srli_epi64(B3, WEBP_RESCALER_RFIX);
309 const __m128i E0 = _mm_sub_epi64(A0, D0); // irow[x] - frac
310 const __m128i E1 = _mm_sub_epi64(A1, D1);
311 const __m128i E2 = _mm_sub_epi64(A2, D2);
312 const __m128i E3 = _mm_sub_epi64(A3, D3);
313 const __m128i F2 = _mm_slli_epi64(D2, 32);
314 const __m128i F3 = _mm_slli_epi64(D3, 32);
315 const __m128i G0 = _mm_or_si128(D0, F2);
316 const __m128i G1 = _mm_or_si128(D1, F3);
317 _mm_storeu_si128((__m128i*)(irow + x_out + 0), G0);
318 _mm_storeu_si128((__m128i*)(irow + x_out + 4), G1);
319 ProcessRow_SSE2(&E0, &E1, &E2, &E3, &mult_xy, dst + x_out);
320 }
321 }
322 for (; x_out < x_out_max; ++x_out) {
323 const uint32_t frac = (int)MULT_FIX_FLOOR(frow[x_out], yscale);
324 const int v = (int)MULT_FIX(irow[x_out] - frac, wrk->fxy_scale);
325 dst[x_out] = (v > 255) ? 255u : (uint8_t)v;
326 irow[x_out] = frac; // new fractional start
327 }
328 } else {
329 const uint32_t scale = wrk->fxy_scale;
330 const __m128i mult = _mm_set_epi32(0, scale, 0, scale);
331 const __m128i zero = _mm_setzero_si128();
332 for (x_out = 0; x_out + 8 <= x_out_max; x_out += 8) {
333 __m128i A0, A1, A2, A3;
334 LoadDispatchAndMult_SSE2(irow + x_out, NULL, &A0, &A1, &A2, &A3);
335 _mm_storeu_si128((__m128i*)(irow + x_out + 0), zero);
336 _mm_storeu_si128((__m128i*)(irow + x_out + 4), zero);
337 ProcessRow_SSE2(&A0, &A1, &A2, &A3, &mult, dst + x_out);
338 }
339 for (; x_out < x_out_max; ++x_out) {
340 const int v = (int)MULT_FIX(irow[x_out], scale);
341 dst[x_out] = (v > 255) ? 255u : (uint8_t)v;
342 irow[x_out] = 0;
343 }
344 }
345}
346
347#undef MULT_FIX_FLOOR
348#undef MULT_FIX
349#undef ROUNDER
350
351//------------------------------------------------------------------------------
352
353extern void WebPRescalerDspInitSSE2(void);
354
355WEBP_TSAN_IGNORE_FUNCTION void WebPRescalerDspInitSSE2(void) {
356 WebPRescalerImportRowExpand = RescalerImportRowExpand_SSE2;
357 WebPRescalerImportRowShrink = RescalerImportRowShrink_SSE2;
358 WebPRescalerExportRowExpand = RescalerExportRowExpand_SSE2;
359 WebPRescalerExportRowShrink = RescalerExportRowShrink_SSE2;
360}
361
362#else // !WEBP_USE_SSE2
363
364WEBP_DSP_INIT_STUB(WebPRescalerDspInitSSE2)
365
366#endif // WEBP_USE_SSE2
367