Line data Source code
1 : /*
2 : * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 : *
4 : * This source code is subject to the terms of the BSD 2 Clause License and
5 : * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 : * was not distributed with this source code in the LICENSE file, you can
7 : * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 : * Media Patent License 1.0 was not distributed with this source code in the
9 : * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 : */
11 :
12 : #include <assert.h>
13 : #include <emmintrin.h> // SSE2
14 :
15 : #include "./aom_config.h"
16 : #include "./aom_dsp_rtcd.h"
17 :
18 : #include "aom_ports/mem.h"
19 :
20 : typedef void (*getNxMvar_fn_t)(const unsigned char *src, int src_stride,
21 : const unsigned char *ref, int ref_stride,
22 : unsigned int *sse, int *sum);
23 :
24 0 : unsigned int aom_get_mb_ss_sse2(const int16_t *src) {
25 0 : __m128i vsum = _mm_setzero_si128();
26 : int i;
27 :
28 0 : for (i = 0; i < 32; ++i) {
29 0 : const __m128i v = _mm_loadu_si128((const __m128i *)src);
30 0 : vsum = _mm_add_epi32(vsum, _mm_madd_epi16(v, v));
31 0 : src += 8;
32 : }
33 :
34 0 : vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8));
35 0 : vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4));
36 0 : return _mm_cvtsi128_si32(vsum);
37 : }
38 :
39 : #define READ64(p, stride, i) \
40 : _mm_unpacklo_epi8( \
41 : _mm_cvtsi32_si128(*(const uint32_t *)(p + i * stride)), \
42 : _mm_cvtsi32_si128(*(const uint32_t *)(p + (i + 1) * stride)))
43 :
44 0 : static void get4x4var_sse2(const uint8_t *src, int src_stride,
45 : const uint8_t *ref, int ref_stride,
46 : unsigned int *sse, int *sum) {
47 0 : const __m128i zero = _mm_setzero_si128();
48 0 : const __m128i src0 = _mm_unpacklo_epi8(READ64(src, src_stride, 0), zero);
49 0 : const __m128i src1 = _mm_unpacklo_epi8(READ64(src, src_stride, 2), zero);
50 0 : const __m128i ref0 = _mm_unpacklo_epi8(READ64(ref, ref_stride, 0), zero);
51 0 : const __m128i ref1 = _mm_unpacklo_epi8(READ64(ref, ref_stride, 2), zero);
52 0 : const __m128i diff0 = _mm_sub_epi16(src0, ref0);
53 0 : const __m128i diff1 = _mm_sub_epi16(src1, ref1);
54 :
55 : // sum
56 0 : __m128i vsum = _mm_add_epi16(diff0, diff1);
57 0 : vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
58 0 : vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
59 0 : vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2));
60 0 : *sum = (int16_t)_mm_extract_epi16(vsum, 0);
61 :
62 : // sse
63 0 : vsum =
64 0 : _mm_add_epi32(_mm_madd_epi16(diff0, diff0), _mm_madd_epi16(diff1, diff1));
65 0 : vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8));
66 0 : vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4));
67 0 : *sse = _mm_cvtsi128_si32(vsum);
68 0 : }
69 :
70 0 : void aom_get8x8var_sse2(const uint8_t *src, int src_stride, const uint8_t *ref,
71 : int ref_stride, unsigned int *sse, int *sum) {
72 0 : const __m128i zero = _mm_setzero_si128();
73 0 : __m128i vsum = _mm_setzero_si128();
74 0 : __m128i vsse = _mm_setzero_si128();
75 : int i;
76 :
77 0 : for (i = 0; i < 8; i += 2) {
78 0 : const __m128i src0 = _mm_unpacklo_epi8(
79 0 : _mm_loadl_epi64((const __m128i *)(src + i * src_stride)), zero);
80 0 : const __m128i ref0 = _mm_unpacklo_epi8(
81 0 : _mm_loadl_epi64((const __m128i *)(ref + i * ref_stride)), zero);
82 0 : const __m128i diff0 = _mm_sub_epi16(src0, ref0);
83 :
84 0 : const __m128i src1 = _mm_unpacklo_epi8(
85 0 : _mm_loadl_epi64((const __m128i *)(src + (i + 1) * src_stride)), zero);
86 0 : const __m128i ref1 = _mm_unpacklo_epi8(
87 0 : _mm_loadl_epi64((const __m128i *)(ref + (i + 1) * ref_stride)), zero);
88 0 : const __m128i diff1 = _mm_sub_epi16(src1, ref1);
89 :
90 0 : vsum = _mm_add_epi16(vsum, diff0);
91 0 : vsum = _mm_add_epi16(vsum, diff1);
92 0 : vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
93 0 : vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
94 : }
95 :
96 : // sum
97 0 : vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
98 0 : vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
99 0 : vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2));
100 0 : *sum = (int16_t)_mm_extract_epi16(vsum, 0);
101 :
102 : // sse
103 0 : vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8));
104 0 : vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4));
105 0 : *sse = _mm_cvtsi128_si32(vsse);
106 0 : }
107 :
108 0 : void aom_get16x16var_sse2(const uint8_t *src, int src_stride,
109 : const uint8_t *ref, int ref_stride, unsigned int *sse,
110 : int *sum) {
111 0 : const __m128i zero = _mm_setzero_si128();
112 0 : __m128i vsum = _mm_setzero_si128();
113 0 : __m128i vsse = _mm_setzero_si128();
114 : int i;
115 :
116 0 : for (i = 0; i < 16; ++i) {
117 0 : const __m128i s = _mm_loadu_si128((const __m128i *)src);
118 0 : const __m128i r = _mm_loadu_si128((const __m128i *)ref);
119 :
120 0 : const __m128i src0 = _mm_unpacklo_epi8(s, zero);
121 0 : const __m128i ref0 = _mm_unpacklo_epi8(r, zero);
122 0 : const __m128i diff0 = _mm_sub_epi16(src0, ref0);
123 :
124 0 : const __m128i src1 = _mm_unpackhi_epi8(s, zero);
125 0 : const __m128i ref1 = _mm_unpackhi_epi8(r, zero);
126 0 : const __m128i diff1 = _mm_sub_epi16(src1, ref1);
127 :
128 0 : vsum = _mm_add_epi16(vsum, diff0);
129 0 : vsum = _mm_add_epi16(vsum, diff1);
130 0 : vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
131 0 : vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
132 :
133 0 : src += src_stride;
134 0 : ref += ref_stride;
135 : }
136 :
137 : // sum
138 0 : vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
139 0 : vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
140 0 : *sum =
141 0 : (int16_t)_mm_extract_epi16(vsum, 0) + (int16_t)_mm_extract_epi16(vsum, 1);
142 :
143 : // sse
144 0 : vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8));
145 0 : vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4));
146 0 : *sse = _mm_cvtsi128_si32(vsse);
147 0 : }
148 :
149 0 : static void variance_sse2(const unsigned char *src, int src_stride,
150 : const unsigned char *ref, int ref_stride, int w,
151 : int h, unsigned int *sse, int *sum,
152 : getNxMvar_fn_t var_fn, int block_size) {
153 : int i, j;
154 :
155 0 : *sse = 0;
156 0 : *sum = 0;
157 :
158 0 : for (i = 0; i < h; i += block_size) {
159 0 : for (j = 0; j < w; j += block_size) {
160 : unsigned int sse0;
161 : int sum0;
162 0 : var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
163 : ref_stride, &sse0, &sum0);
164 0 : *sse += sse0;
165 0 : *sum += sum0;
166 : }
167 : }
168 0 : }
169 :
170 0 : unsigned int aom_variance4x4_sse2(const uint8_t *src, int src_stride,
171 : const uint8_t *ref, int ref_stride,
172 : unsigned int *sse) {
173 : int sum;
174 0 : get4x4var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
175 0 : assert(sum <= 255 * 4 * 4);
176 0 : assert(sum >= -255 * 4 * 4);
177 0 : return *sse - ((sum * sum) >> 4);
178 : }
179 :
180 0 : unsigned int aom_variance8x4_sse2(const uint8_t *src, int src_stride,
181 : const uint8_t *ref, int ref_stride,
182 : unsigned int *sse) {
183 : int sum;
184 0 : variance_sse2(src, src_stride, ref, ref_stride, 8, 4, sse, &sum,
185 : get4x4var_sse2, 4);
186 0 : assert(sum <= 255 * 8 * 4);
187 0 : assert(sum >= -255 * 8 * 4);
188 0 : return *sse - ((sum * sum) >> 5);
189 : }
190 :
191 0 : unsigned int aom_variance4x8_sse2(const uint8_t *src, int src_stride,
192 : const uint8_t *ref, int ref_stride,
193 : unsigned int *sse) {
194 : int sum;
195 0 : variance_sse2(src, src_stride, ref, ref_stride, 4, 8, sse, &sum,
196 : get4x4var_sse2, 4);
197 0 : assert(sum <= 255 * 8 * 4);
198 0 : assert(sum >= -255 * 8 * 4);
199 0 : return *sse - ((sum * sum) >> 5);
200 : }
201 :
202 0 : unsigned int aom_variance8x8_sse2(const unsigned char *src, int src_stride,
203 : const unsigned char *ref, int ref_stride,
204 : unsigned int *sse) {
205 : int sum;
206 0 : aom_get8x8var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
207 0 : assert(sum <= 255 * 8 * 8);
208 0 : assert(sum >= -255 * 8 * 8);
209 0 : return *sse - ((sum * sum) >> 6);
210 : }
211 :
212 0 : unsigned int aom_variance16x8_sse2(const unsigned char *src, int src_stride,
213 : const unsigned char *ref, int ref_stride,
214 : unsigned int *sse) {
215 : int sum;
216 0 : variance_sse2(src, src_stride, ref, ref_stride, 16, 8, sse, &sum,
217 : aom_get8x8var_sse2, 8);
218 0 : assert(sum <= 255 * 16 * 8);
219 0 : assert(sum >= -255 * 16 * 8);
220 0 : return *sse - ((sum * sum) >> 7);
221 : }
222 :
223 0 : unsigned int aom_variance8x16_sse2(const unsigned char *src, int src_stride,
224 : const unsigned char *ref, int ref_stride,
225 : unsigned int *sse) {
226 : int sum;
227 0 : variance_sse2(src, src_stride, ref, ref_stride, 8, 16, sse, &sum,
228 : aom_get8x8var_sse2, 8);
229 0 : assert(sum <= 255 * 16 * 8);
230 0 : assert(sum >= -255 * 16 * 8);
231 0 : return *sse - ((sum * sum) >> 7);
232 : }
233 :
234 0 : unsigned int aom_variance16x16_sse2(const unsigned char *src, int src_stride,
235 : const unsigned char *ref, int ref_stride,
236 : unsigned int *sse) {
237 : int sum;
238 0 : aom_get16x16var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
239 0 : assert(sum <= 255 * 16 * 16);
240 0 : assert(sum >= -255 * 16 * 16);
241 0 : return *sse - ((uint32_t)((int64_t)sum * sum) >> 8);
242 : }
243 :
244 0 : unsigned int aom_variance32x32_sse2(const uint8_t *src, int src_stride,
245 : const uint8_t *ref, int ref_stride,
246 : unsigned int *sse) {
247 : int sum;
248 0 : variance_sse2(src, src_stride, ref, ref_stride, 32, 32, sse, &sum,
249 : aom_get16x16var_sse2, 16);
250 0 : assert(sum <= 255 * 32 * 32);
251 0 : assert(sum >= -255 * 32 * 32);
252 0 : return *sse - (unsigned int)(((int64_t)sum * sum) >> 10);
253 : }
254 :
255 0 : unsigned int aom_variance32x16_sse2(const uint8_t *src, int src_stride,
256 : const uint8_t *ref, int ref_stride,
257 : unsigned int *sse) {
258 : int sum;
259 0 : variance_sse2(src, src_stride, ref, ref_stride, 32, 16, sse, &sum,
260 : aom_get16x16var_sse2, 16);
261 0 : assert(sum <= 255 * 32 * 16);
262 0 : assert(sum >= -255 * 32 * 16);
263 0 : return *sse - (unsigned int)(((int64_t)sum * sum) >> 9);
264 : }
265 :
266 0 : unsigned int aom_variance16x32_sse2(const uint8_t *src, int src_stride,
267 : const uint8_t *ref, int ref_stride,
268 : unsigned int *sse) {
269 : int sum;
270 0 : variance_sse2(src, src_stride, ref, ref_stride, 16, 32, sse, &sum,
271 : aom_get16x16var_sse2, 16);
272 0 : assert(sum <= 255 * 32 * 16);
273 0 : assert(sum >= -255 * 32 * 16);
274 0 : return *sse - (unsigned int)(((int64_t)sum * sum) >> 9);
275 : }
276 :
277 0 : unsigned int aom_variance64x64_sse2(const uint8_t *src, int src_stride,
278 : const uint8_t *ref, int ref_stride,
279 : unsigned int *sse) {
280 : int sum;
281 0 : variance_sse2(src, src_stride, ref, ref_stride, 64, 64, sse, &sum,
282 : aom_get16x16var_sse2, 16);
283 0 : assert(sum <= 255 * 64 * 64);
284 0 : assert(sum >= -255 * 64 * 64);
285 0 : return *sse - (unsigned int)(((int64_t)sum * sum) >> 12);
286 : }
287 :
288 0 : unsigned int aom_variance64x32_sse2(const uint8_t *src, int src_stride,
289 : const uint8_t *ref, int ref_stride,
290 : unsigned int *sse) {
291 : int sum;
292 0 : variance_sse2(src, src_stride, ref, ref_stride, 64, 32, sse, &sum,
293 : aom_get16x16var_sse2, 16);
294 0 : assert(sum <= 255 * 64 * 32);
295 0 : assert(sum >= -255 * 64 * 32);
296 0 : return *sse - (unsigned int)(((int64_t)sum * sum) >> 11);
297 : }
298 :
299 0 : unsigned int aom_variance32x64_sse2(const uint8_t *src, int src_stride,
300 : const uint8_t *ref, int ref_stride,
301 : unsigned int *sse) {
302 : int sum;
303 0 : variance_sse2(src, src_stride, ref, ref_stride, 32, 64, sse, &sum,
304 : aom_get16x16var_sse2, 16);
305 0 : assert(sum <= 255 * 64 * 32);
306 0 : assert(sum >= -255 * 64 * 32);
307 0 : return *sse - (unsigned int)(((int64_t)sum * sum) >> 11);
308 : }
309 :
310 0 : unsigned int aom_mse8x8_sse2(const uint8_t *src, int src_stride,
311 : const uint8_t *ref, int ref_stride,
312 : unsigned int *sse) {
313 0 : aom_variance8x8_sse2(src, src_stride, ref, ref_stride, sse);
314 0 : return *sse;
315 : }
316 :
317 0 : unsigned int aom_mse8x16_sse2(const uint8_t *src, int src_stride,
318 : const uint8_t *ref, int ref_stride,
319 : unsigned int *sse) {
320 0 : aom_variance8x16_sse2(src, src_stride, ref, ref_stride, sse);
321 0 : return *sse;
322 : }
323 :
324 0 : unsigned int aom_mse16x8_sse2(const uint8_t *src, int src_stride,
325 : const uint8_t *ref, int ref_stride,
326 : unsigned int *sse) {
327 0 : aom_variance16x8_sse2(src, src_stride, ref, ref_stride, sse);
328 0 : return *sse;
329 : }
330 :
331 0 : unsigned int aom_mse16x16_sse2(const uint8_t *src, int src_stride,
332 : const uint8_t *ref, int ref_stride,
333 : unsigned int *sse) {
334 0 : aom_variance16x16_sse2(src, src_stride, ref, ref_stride, sse);
335 0 : return *sse;
336 : }
337 :
338 : // The 2 unused parameters are place holders for PIC enabled build.
339 : // These definitions are for functions defined in subpel_variance.asm
340 : #define DECL(w, opt) \
341 : int aom_sub_pixel_variance##w##xh_##opt( \
342 : const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \
343 : const uint8_t *dst, ptrdiff_t dst_stride, int height, unsigned int *sse, \
344 : void *unused0, void *unused)
345 : #define DECLS(opt1, opt2) \
346 : DECL(4, opt1); \
347 : DECL(8, opt1); \
348 : DECL(16, opt1)
349 :
350 : DECLS(sse2, sse2);
351 : DECLS(ssse3, ssse3);
352 : #undef DECLS
353 : #undef DECL
354 :
355 : #define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast) \
356 : unsigned int aom_sub_pixel_variance##w##x##h##_##opt( \
357 : const uint8_t *src, int src_stride, int x_offset, int y_offset, \
358 : const uint8_t *dst, int dst_stride, unsigned int *sse_ptr) { \
359 : unsigned int sse; \
360 : int se = aom_sub_pixel_variance##wf##xh_##opt(src, src_stride, x_offset, \
361 : y_offset, dst, dst_stride, \
362 : h, &sse, NULL, NULL); \
363 : if (w > wf) { \
364 : unsigned int sse2; \
365 : int se2 = aom_sub_pixel_variance##wf##xh_##opt( \
366 : src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, h, \
367 : &sse2, NULL, NULL); \
368 : se += se2; \
369 : sse += sse2; \
370 : if (w > wf * 2) { \
371 : se2 = aom_sub_pixel_variance##wf##xh_##opt( \
372 : src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, h, \
373 : &sse2, NULL, NULL); \
374 : se += se2; \
375 : sse += sse2; \
376 : se2 = aom_sub_pixel_variance##wf##xh_##opt( \
377 : src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, h, \
378 : &sse2, NULL, NULL); \
379 : se += se2; \
380 : sse += sse2; \
381 : } \
382 : } \
383 : *sse_ptr = sse; \
384 : return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2)); \
385 : }
386 :
387 : #define FNS(opt1, opt2) \
388 : FN(64, 64, 16, 6, 6, opt1, (int64_t), (int64_t)); \
389 : FN(64, 32, 16, 6, 5, opt1, (int64_t), (int64_t)); \
390 : FN(32, 64, 16, 5, 6, opt1, (int64_t), (int64_t)); \
391 : FN(32, 32, 16, 5, 5, opt1, (int64_t), (int64_t)); \
392 : FN(32, 16, 16, 5, 4, opt1, (int64_t), (int64_t)); \
393 : FN(16, 32, 16, 4, 5, opt1, (int64_t), (int64_t)); \
394 : FN(16, 16, 16, 4, 4, opt1, (uint32_t), (int64_t)); \
395 : FN(16, 8, 16, 4, 3, opt1, (int32_t), (int32_t)); \
396 : FN(8, 16, 8, 3, 4, opt1, (int32_t), (int32_t)); \
397 : FN(8, 8, 8, 3, 3, opt1, (int32_t), (int32_t)); \
398 : FN(8, 4, 8, 3, 2, opt1, (int32_t), (int32_t)); \
399 : FN(4, 8, 4, 2, 3, opt1, (int32_t), (int32_t)); \
400 : FN(4, 4, 4, 2, 2, opt1, (int32_t), (int32_t))
401 :
402 0 : FNS(sse2, sse2);
403 0 : FNS(ssse3, ssse3);
404 :
405 : #undef FNS
406 : #undef FN
407 :
408 : // The 2 unused parameters are place holders for PIC enabled build.
409 : #define DECL(w, opt) \
410 : int aom_sub_pixel_avg_variance##w##xh_##opt( \
411 : const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \
412 : const uint8_t *dst, ptrdiff_t dst_stride, const uint8_t *sec, \
413 : ptrdiff_t sec_stride, int height, unsigned int *sse, void *unused0, \
414 : void *unused)
415 : #define DECLS(opt1, opt2) \
416 : DECL(4, opt1); \
417 : DECL(8, opt1); \
418 : DECL(16, opt1)
419 :
420 : DECLS(sse2, sse2);
421 : DECLS(ssse3, ssse3);
422 : #undef DECL
423 : #undef DECLS
424 :
425 : #define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast) \
426 : unsigned int aom_sub_pixel_avg_variance##w##x##h##_##opt( \
427 : const uint8_t *src, int src_stride, int x_offset, int y_offset, \
428 : const uint8_t *dst, int dst_stride, unsigned int *sseptr, \
429 : const uint8_t *sec) { \
430 : unsigned int sse; \
431 : int se = aom_sub_pixel_avg_variance##wf##xh_##opt( \
432 : src, src_stride, x_offset, y_offset, dst, dst_stride, sec, w, h, &sse, \
433 : NULL, NULL); \
434 : if (w > wf) { \
435 : unsigned int sse2; \
436 : int se2 = aom_sub_pixel_avg_variance##wf##xh_##opt( \
437 : src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, \
438 : sec + 16, w, h, &sse2, NULL, NULL); \
439 : se += se2; \
440 : sse += sse2; \
441 : if (w > wf * 2) { \
442 : se2 = aom_sub_pixel_avg_variance##wf##xh_##opt( \
443 : src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, \
444 : sec + 32, w, h, &sse2, NULL, NULL); \
445 : se += se2; \
446 : sse += sse2; \
447 : se2 = aom_sub_pixel_avg_variance##wf##xh_##opt( \
448 : src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, \
449 : sec + 48, w, h, &sse2, NULL, NULL); \
450 : se += se2; \
451 : sse += sse2; \
452 : } \
453 : } \
454 : *sseptr = sse; \
455 : return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2)); \
456 : }
457 :
458 : #define FNS(opt1, opt2) \
459 : FN(64, 64, 16, 6, 6, opt1, (int64_t), (int64_t)); \
460 : FN(64, 32, 16, 6, 5, opt1, (int64_t), (int64_t)); \
461 : FN(32, 64, 16, 5, 6, opt1, (int64_t), (int64_t)); \
462 : FN(32, 32, 16, 5, 5, opt1, (int64_t), (int64_t)); \
463 : FN(32, 16, 16, 5, 4, opt1, (int64_t), (int64_t)); \
464 : FN(16, 32, 16, 4, 5, opt1, (int64_t), (int64_t)); \
465 : FN(16, 16, 16, 4, 4, opt1, (uint32_t), (int64_t)); \
466 : FN(16, 8, 16, 4, 3, opt1, (uint32_t), (int32_t)); \
467 : FN(8, 16, 8, 3, 4, opt1, (uint32_t), (int32_t)); \
468 : FN(8, 8, 8, 3, 3, opt1, (uint32_t), (int32_t)); \
469 : FN(8, 4, 8, 3, 2, opt1, (uint32_t), (int32_t)); \
470 : FN(4, 8, 4, 2, 3, opt1, (uint32_t), (int32_t)); \
471 : FN(4, 4, 4, 2, 2, opt1, (uint32_t), (int32_t))
472 :
473 0 : FNS(sse2, sse);
474 0 : FNS(ssse3, ssse3);
475 :
476 : #undef FNS
477 : #undef FN
478 :
479 0 : void aom_upsampled_pred_sse2(uint8_t *comp_pred, int width, int height,
480 : const uint8_t *ref, int ref_stride) {
481 : int i, j;
482 0 : int stride = ref_stride << 3;
483 :
484 0 : if (width >= 16) {
485 : // read 16 points at one time
486 0 : for (i = 0; i < height; i++) {
487 0 : for (j = 0; j < width; j += 16) {
488 0 : __m128i s0 = _mm_loadu_si128((const __m128i *)ref);
489 0 : __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16));
490 0 : __m128i s2 = _mm_loadu_si128((const __m128i *)(ref + 32));
491 0 : __m128i s3 = _mm_loadu_si128((const __m128i *)(ref + 48));
492 0 : __m128i s4 = _mm_loadu_si128((const __m128i *)(ref + 64));
493 0 : __m128i s5 = _mm_loadu_si128((const __m128i *)(ref + 80));
494 0 : __m128i s6 = _mm_loadu_si128((const __m128i *)(ref + 96));
495 0 : __m128i s7 = _mm_loadu_si128((const __m128i *)(ref + 112));
496 : __m128i t0, t1, t2, t3;
497 :
498 0 : t0 = _mm_unpacklo_epi8(s0, s1);
499 0 : s1 = _mm_unpackhi_epi8(s0, s1);
500 0 : t1 = _mm_unpacklo_epi8(s2, s3);
501 0 : s3 = _mm_unpackhi_epi8(s2, s3);
502 0 : t2 = _mm_unpacklo_epi8(s4, s5);
503 0 : s5 = _mm_unpackhi_epi8(s4, s5);
504 0 : t3 = _mm_unpacklo_epi8(s6, s7);
505 0 : s7 = _mm_unpackhi_epi8(s6, s7);
506 :
507 0 : s0 = _mm_unpacklo_epi8(t0, s1);
508 0 : s2 = _mm_unpacklo_epi8(t1, s3);
509 0 : s4 = _mm_unpacklo_epi8(t2, s5);
510 0 : s6 = _mm_unpacklo_epi8(t3, s7);
511 0 : s0 = _mm_unpacklo_epi32(s0, s2);
512 0 : s4 = _mm_unpacklo_epi32(s4, s6);
513 0 : s0 = _mm_unpacklo_epi64(s0, s4);
514 :
515 : _mm_storeu_si128((__m128i *)(comp_pred), s0);
516 0 : comp_pred += 16;
517 0 : ref += 16 * 8;
518 : }
519 0 : ref += stride - (width << 3);
520 : }
521 0 : } else if (width >= 8) {
522 : // read 8 points at one time
523 0 : for (i = 0; i < height; i++) {
524 0 : for (j = 0; j < width; j += 8) {
525 0 : __m128i s0 = _mm_loadu_si128((const __m128i *)ref);
526 0 : __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16));
527 0 : __m128i s2 = _mm_loadu_si128((const __m128i *)(ref + 32));
528 0 : __m128i s3 = _mm_loadu_si128((const __m128i *)(ref + 48));
529 : __m128i t0, t1;
530 :
531 0 : t0 = _mm_unpacklo_epi8(s0, s1);
532 0 : s1 = _mm_unpackhi_epi8(s0, s1);
533 0 : t1 = _mm_unpacklo_epi8(s2, s3);
534 0 : s3 = _mm_unpackhi_epi8(s2, s3);
535 :
536 0 : s0 = _mm_unpacklo_epi8(t0, s1);
537 0 : s2 = _mm_unpacklo_epi8(t1, s3);
538 0 : s0 = _mm_unpacklo_epi32(s0, s2);
539 :
540 : _mm_storel_epi64((__m128i *)(comp_pred), s0);
541 0 : comp_pred += 8;
542 0 : ref += 8 * 8;
543 : }
544 0 : ref += stride - (width << 3);
545 : }
546 : } else {
547 : // read 4 points at one time
548 0 : for (i = 0; i < height; i++) {
549 0 : for (j = 0; j < width; j += 4) {
550 0 : __m128i s0 = _mm_loadu_si128((const __m128i *)ref);
551 0 : __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16));
552 : __m128i t0;
553 :
554 0 : t0 = _mm_unpacklo_epi8(s0, s1);
555 0 : s1 = _mm_unpackhi_epi8(s0, s1);
556 0 : s0 = _mm_unpacklo_epi8(t0, s1);
557 :
558 0 : *(int *)comp_pred = _mm_cvtsi128_si32(s0);
559 0 : comp_pred += 4;
560 0 : ref += 4 * 8;
561 : }
562 0 : ref += stride - (width << 3);
563 : }
564 : }
565 0 : }
566 :
567 0 : void aom_comp_avg_upsampled_pred_sse2(uint8_t *comp_pred, const uint8_t *pred,
568 : int width, int height, const uint8_t *ref,
569 : int ref_stride) {
570 0 : const __m128i zero = _mm_set1_epi16(0);
571 0 : const __m128i one = _mm_set1_epi16(1);
572 : int i, j;
573 0 : int stride = ref_stride << 3;
574 :
575 0 : if (width >= 16) {
576 : // read 16 points at one time
577 0 : for (i = 0; i < height; i++) {
578 0 : for (j = 0; j < width; j += 16) {
579 0 : __m128i s0 = _mm_loadu_si128((const __m128i *)ref);
580 0 : __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16));
581 0 : __m128i s2 = _mm_loadu_si128((const __m128i *)(ref + 32));
582 0 : __m128i s3 = _mm_loadu_si128((const __m128i *)(ref + 48));
583 0 : __m128i s4 = _mm_loadu_si128((const __m128i *)(ref + 64));
584 0 : __m128i s5 = _mm_loadu_si128((const __m128i *)(ref + 80));
585 0 : __m128i s6 = _mm_loadu_si128((const __m128i *)(ref + 96));
586 0 : __m128i s7 = _mm_loadu_si128((const __m128i *)(ref + 112));
587 0 : __m128i p0 = _mm_loadu_si128((const __m128i *)pred);
588 : __m128i p1;
589 : __m128i t0, t1, t2, t3;
590 :
591 0 : t0 = _mm_unpacklo_epi8(s0, s1);
592 0 : s1 = _mm_unpackhi_epi8(s0, s1);
593 0 : t1 = _mm_unpacklo_epi8(s2, s3);
594 0 : s3 = _mm_unpackhi_epi8(s2, s3);
595 0 : t2 = _mm_unpacklo_epi8(s4, s5);
596 0 : s5 = _mm_unpackhi_epi8(s4, s5);
597 0 : t3 = _mm_unpacklo_epi8(s6, s7);
598 0 : s7 = _mm_unpackhi_epi8(s6, s7);
599 :
600 0 : s0 = _mm_unpacklo_epi8(t0, s1);
601 0 : s2 = _mm_unpacklo_epi8(t1, s3);
602 0 : s4 = _mm_unpacklo_epi8(t2, s5);
603 0 : s6 = _mm_unpacklo_epi8(t3, s7);
604 :
605 0 : s0 = _mm_unpacklo_epi32(s0, s2);
606 0 : s4 = _mm_unpacklo_epi32(s4, s6);
607 0 : s0 = _mm_unpacklo_epi8(s0, zero);
608 0 : s4 = _mm_unpacklo_epi8(s4, zero);
609 :
610 0 : p1 = _mm_unpackhi_epi8(p0, zero);
611 0 : p0 = _mm_unpacklo_epi8(p0, zero);
612 0 : p0 = _mm_adds_epu16(s0, p0);
613 0 : p1 = _mm_adds_epu16(s4, p1);
614 0 : p0 = _mm_adds_epu16(p0, one);
615 0 : p1 = _mm_adds_epu16(p1, one);
616 :
617 0 : p0 = _mm_srli_epi16(p0, 1);
618 0 : p1 = _mm_srli_epi16(p1, 1);
619 0 : p0 = _mm_packus_epi16(p0, p1);
620 :
621 : _mm_storeu_si128((__m128i *)(comp_pred), p0);
622 0 : comp_pred += 16;
623 0 : pred += 16;
624 0 : ref += 16 * 8;
625 : }
626 0 : ref += stride - (width << 3);
627 : }
628 0 : } else if (width >= 8) {
629 : // read 8 points at one time
630 0 : for (i = 0; i < height; i++) {
631 0 : for (j = 0; j < width; j += 8) {
632 0 : __m128i s0 = _mm_loadu_si128((const __m128i *)ref);
633 0 : __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16));
634 0 : __m128i s2 = _mm_loadu_si128((const __m128i *)(ref + 32));
635 0 : __m128i s3 = _mm_loadu_si128((const __m128i *)(ref + 48));
636 0 : __m128i p0 = _mm_loadl_epi64((const __m128i *)pred);
637 : __m128i t0, t1;
638 :
639 0 : t0 = _mm_unpacklo_epi8(s0, s1);
640 0 : s1 = _mm_unpackhi_epi8(s0, s1);
641 0 : t1 = _mm_unpacklo_epi8(s2, s3);
642 0 : s3 = _mm_unpackhi_epi8(s2, s3);
643 :
644 0 : s0 = _mm_unpacklo_epi8(t0, s1);
645 0 : s2 = _mm_unpacklo_epi8(t1, s3);
646 0 : s0 = _mm_unpacklo_epi32(s0, s2);
647 0 : s0 = _mm_unpacklo_epi8(s0, zero);
648 :
649 0 : p0 = _mm_unpacklo_epi8(p0, zero);
650 0 : p0 = _mm_adds_epu16(s0, p0);
651 0 : p0 = _mm_adds_epu16(p0, one);
652 0 : p0 = _mm_srli_epi16(p0, 1);
653 0 : p0 = _mm_packus_epi16(p0, zero);
654 :
655 : _mm_storel_epi64((__m128i *)(comp_pred), p0);
656 0 : comp_pred += 8;
657 0 : pred += 8;
658 0 : ref += 8 * 8;
659 : }
660 0 : ref += stride - (width << 3);
661 : }
662 : } else {
663 : // read 4 points at one time
664 0 : for (i = 0; i < height; i++) {
665 0 : for (j = 0; j < width; j += 4) {
666 0 : __m128i s0 = _mm_loadu_si128((const __m128i *)ref);
667 0 : __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16));
668 0 : __m128i p0 = _mm_cvtsi32_si128(*(const uint32_t *)pred);
669 : __m128i t0;
670 :
671 0 : t0 = _mm_unpacklo_epi8(s0, s1);
672 0 : s1 = _mm_unpackhi_epi8(s0, s1);
673 0 : s0 = _mm_unpacklo_epi8(t0, s1);
674 0 : s0 = _mm_unpacklo_epi8(s0, zero);
675 :
676 0 : p0 = _mm_unpacklo_epi8(p0, zero);
677 0 : p0 = _mm_adds_epu16(s0, p0);
678 0 : p0 = _mm_adds_epu16(p0, one);
679 0 : p0 = _mm_srli_epi16(p0, 1);
680 0 : p0 = _mm_packus_epi16(p0, zero);
681 :
682 0 : *(int *)comp_pred = _mm_cvtsi128_si32(p0);
683 0 : comp_pred += 4;
684 0 : pred += 4;
685 0 : ref += 4 * 8;
686 : }
687 0 : ref += stride - (width << 3);
688 : }
689 : }
690 0 : }
|
