xref: /aosp_15_r20/external/libaom/aom_dsp/x86/intrapred_sse4.c (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
1 /*
2  * Copyright (c) 2021, 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 <emmintrin.h>  // SSE2
13 #include <smmintrin.h>  /* SSE4.1 */
14 
15 #include "config/av1_rtcd.h"
16 #include "aom_dsp/x86/intrapred_x86.h"
17 #include "aom_dsp/x86/intrapred_utils.h"
18 #include "aom_dsp/x86/lpf_common_sse2.h"
19 
20 // Low bit depth functions
21 static DECLARE_ALIGNED(16, uint8_t, Mask[2][33][16]) = {
22   { { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
23     { 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
24     { 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
25     { 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
26     { 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
27     { 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
28     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
29     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
30     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0 },
31     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0,
32       0 },
33     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0,
34       0 },
35     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0,
36       0, 0 },
37     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0,
38       0, 0, 0 },
39     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
40       0xff, 0, 0, 0 },
41     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
42       0xff, 0xff, 0, 0 },
43     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
44       0xff, 0xff, 0xff, 0 },
45     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
46       0xff, 0xff, 0xff, 0xff },
47     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
48       0xff, 0xff, 0xff, 0xff },
49     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
50       0xff, 0xff, 0xff, 0xff },
51     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
52       0xff, 0xff, 0xff, 0xff },
53     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
54       0xff, 0xff, 0xff, 0xff },
55     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
56       0xff, 0xff, 0xff, 0xff },
57     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
58       0xff, 0xff, 0xff, 0xff },
59     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
60       0xff, 0xff, 0xff, 0xff },
61     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
62       0xff, 0xff, 0xff, 0xff },
63     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
64       0xff, 0xff, 0xff, 0xff },
65     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
66       0xff, 0xff, 0xff, 0xff },
67     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
68       0xff, 0xff, 0xff, 0xff },
69     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
70       0xff, 0xff, 0xff, 0xff },
71     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
72       0xff, 0xff, 0xff, 0xff },
73     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
74       0xff, 0xff, 0xff, 0xff },
75     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
76       0xff, 0xff, 0xff, 0xff },
77     { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
78       0xff, 0xff, 0xff, 0xff } },
79   {
80       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
81       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
82       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
83       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
84       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
85       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
86       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
87       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
88       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
89       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
90       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
91       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
92       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
93       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
94       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
95       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
96       { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
97       { 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
98       { 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
99       { 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
100       { 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
101       { 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
102       { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
103       { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
104       { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0,
105         0 },
106       { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0,
107         0 },
108       { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0,
109         0, 0 },
110       { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0,
111         0, 0, 0 },
112       { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
113         0, 0, 0, 0 },
114       { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
115         0xff, 0, 0, 0 },
116       { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
117         0xff, 0xff, 0, 0 },
118       { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
119         0xff, 0xff, 0xff, 0 },
120       { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
121         0xff, 0xff, 0xff, 0xff },
122   },
123 };
124 
125 /* clang-format on */
dr_prediction_z1_HxW_internal_sse4_1(int H,int W,__m128i * dst,const uint8_t * above,int upsample_above,int dx)126 static AOM_FORCE_INLINE void dr_prediction_z1_HxW_internal_sse4_1(
127     int H, int W, __m128i *dst, const uint8_t *above, int upsample_above,
128     int dx) {
129   const int frac_bits = 6 - upsample_above;
130   const int max_base_x = ((W + H) - 1) << upsample_above;
131 
132   assert(dx > 0);
133   // pre-filter above pixels
134   // store in temp buffers:
135   //   above[x] * 32 + 16
136   //   above[x+1] - above[x]
137   // final pixels will be calculated as:
138   //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
139   __m128i a0, a1, a32, a16;
140   __m128i diff, c3f;
141   __m128i a_mbase_x;
142 
143   a16 = _mm_set1_epi16(16);
144   a_mbase_x = _mm_set1_epi8((char)above[max_base_x]);
145   c3f = _mm_set1_epi16(0x3f);
146 
147   int x = dx;
148   for (int r = 0; r < W; r++) {
149     __m128i b, res, res1, shift;
150     __m128i a0_above, a1_above;
151 
152     int base = x >> frac_bits;
153     int base_max_diff = (max_base_x - base) >> upsample_above;
154     if (base_max_diff <= 0) {
155       for (int i = r; i < W; ++i) {
156         dst[i] = a_mbase_x;  // save 4 values
157       }
158       return;
159     }
160     if (base_max_diff > H) base_max_diff = H;
161     a0_above = _mm_loadu_si128((__m128i *)(above + base));
162     a1_above = _mm_loadu_si128((__m128i *)(above + base + 1));
163 
164     if (upsample_above) {
165       a0_above = _mm_shuffle_epi8(a0_above, *(__m128i *)EvenOddMaskx[0]);
166       a1_above = _mm_srli_si128(a0_above, 8);
167 
168       shift = _mm_srli_epi16(
169           _mm_and_si128(_mm_slli_epi16(_mm_set1_epi16(x), upsample_above), c3f),
170           1);
171     } else {
172       shift = _mm_srli_epi16(_mm_and_si128(_mm_set1_epi16(x), c3f), 1);
173     }
174     // lower half
175     a0 = _mm_cvtepu8_epi16(a0_above);
176     a1 = _mm_cvtepu8_epi16(a1_above);
177 
178     diff = _mm_sub_epi16(a1, a0);   // a[x+1] - a[x]
179     a32 = _mm_slli_epi16(a0, 5);    // a[x] * 32
180     a32 = _mm_add_epi16(a32, a16);  // a[x] * 32 + 16
181 
182     b = _mm_mullo_epi16(diff, shift);
183     res = _mm_add_epi16(a32, b);
184     res = _mm_srli_epi16(res, 5);
185 
186     // uppar half
187     a0 = _mm_cvtepu8_epi16(_mm_srli_si128(a0_above, 8));
188     a1 = _mm_cvtepu8_epi16(_mm_srli_si128(a1_above, 8));
189 
190     diff = _mm_sub_epi16(a1, a0);   // a[x+1] - a[x]
191     a32 = _mm_slli_epi16(a0, 5);    // a[x] * 32
192     a32 = _mm_add_epi16(a32, a16);  // a[x] * 32 + 16
193 
194     b = _mm_mullo_epi16(diff, shift);
195     res1 = _mm_add_epi16(a32, b);
196     res1 = _mm_srli_epi16(res1, 5);
197 
198     res = _mm_packus_epi16(res, res1);
199 
200     dst[r] =
201         _mm_blendv_epi8(a_mbase_x, res, *(__m128i *)Mask[0][base_max_diff]);
202     x += dx;
203   }
204 }
205 
dr_prediction_z1_4xN_sse4_1(int N,uint8_t * dst,ptrdiff_t stride,const uint8_t * above,int upsample_above,int dx)206 static void dr_prediction_z1_4xN_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
207                                         const uint8_t *above,
208                                         int upsample_above, int dx) {
209   __m128i dstvec[16];
210 
211   dr_prediction_z1_HxW_internal_sse4_1(4, N, dstvec, above, upsample_above, dx);
212   for (int i = 0; i < N; i++) {
213     *(int *)(dst + stride * i) = _mm_cvtsi128_si32(dstvec[i]);
214   }
215 }
216 
dr_prediction_z1_8xN_sse4_1(int N,uint8_t * dst,ptrdiff_t stride,const uint8_t * above,int upsample_above,int dx)217 static void dr_prediction_z1_8xN_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
218                                         const uint8_t *above,
219                                         int upsample_above, int dx) {
220   __m128i dstvec[32];
221 
222   dr_prediction_z1_HxW_internal_sse4_1(8, N, dstvec, above, upsample_above, dx);
223   for (int i = 0; i < N; i++) {
224     _mm_storel_epi64((__m128i *)(dst + stride * i), dstvec[i]);
225   }
226 }
227 
dr_prediction_z1_16xN_sse4_1(int N,uint8_t * dst,ptrdiff_t stride,const uint8_t * above,int upsample_above,int dx)228 static void dr_prediction_z1_16xN_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
229                                          const uint8_t *above,
230                                          int upsample_above, int dx) {
231   __m128i dstvec[64];
232 
233   dr_prediction_z1_HxW_internal_sse4_1(16, N, dstvec, above, upsample_above,
234                                        dx);
235   for (int i = 0; i < N; i++) {
236     _mm_storeu_si128((__m128i *)(dst + stride * i), dstvec[i]);
237   }
238 }
239 
dr_prediction_z1_32xN_internal_sse4_1(int N,__m128i * dstvec,__m128i * dstvec_h,const uint8_t * above,int upsample_above,int dx)240 static AOM_FORCE_INLINE void dr_prediction_z1_32xN_internal_sse4_1(
241     int N, __m128i *dstvec, __m128i *dstvec_h, const uint8_t *above,
242     int upsample_above, int dx) {
243   // here upsample_above is 0 by design of av1_use_intra_edge_upsample
244   (void)upsample_above;
245   const int frac_bits = 6;
246   const int max_base_x = ((32 + N) - 1);
247 
248   // pre-filter above pixels
249   // store in temp buffers:
250   //   above[x] * 32 + 16
251   //   above[x+1] - above[x]
252   // final pixels will be calculated as:
253   //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
254   __m128i a0, a1, a32, a16;
255   __m128i a_mbase_x, diff, c3f;
256 
257   a16 = _mm_set1_epi16(16);
258   a_mbase_x = _mm_set1_epi8((char)above[max_base_x]);
259   c3f = _mm_set1_epi16(0x3f);
260 
261   int x = dx;
262   for (int r = 0; r < N; r++) {
263     __m128i b, res, res1, res16[2];
264     __m128i a0_above, a1_above;
265 
266     int base = x >> frac_bits;
267     int base_max_diff = (max_base_x - base);
268     if (base_max_diff <= 0) {
269       for (int i = r; i < N; ++i) {
270         dstvec[i] = a_mbase_x;  // save 32 values
271         dstvec_h[i] = a_mbase_x;
272       }
273       return;
274     }
275     if (base_max_diff > 32) base_max_diff = 32;
276     __m128i shift = _mm_srli_epi16(_mm_and_si128(_mm_set1_epi16(x), c3f), 1);
277 
278     for (int j = 0, jj = 0; j < 32; j += 16, jj++) {
279       int mdiff = base_max_diff - j;
280       if (mdiff <= 0) {
281         res16[jj] = a_mbase_x;
282       } else {
283         a0_above = _mm_loadu_si128((__m128i *)(above + base + j));
284         a1_above = _mm_loadu_si128((__m128i *)(above + base + j + 1));
285 
286         // lower half
287         a0 = _mm_cvtepu8_epi16(a0_above);
288         a1 = _mm_cvtepu8_epi16(a1_above);
289 
290         diff = _mm_sub_epi16(a1, a0);   // a[x+1] - a[x]
291         a32 = _mm_slli_epi16(a0, 5);    // a[x] * 32
292         a32 = _mm_add_epi16(a32, a16);  // a[x] * 32 + 16
293         b = _mm_mullo_epi16(diff, shift);
294 
295         res = _mm_add_epi16(a32, b);
296         res = _mm_srli_epi16(res, 5);
297 
298         // uppar half
299         a0 = _mm_cvtepu8_epi16(_mm_srli_si128(a0_above, 8));
300         a1 = _mm_cvtepu8_epi16(_mm_srli_si128(a1_above, 8));
301 
302         diff = _mm_sub_epi16(a1, a0);   // a[x+1] - a[x]
303         a32 = _mm_slli_epi16(a0, 5);    // a[x] * 32
304         a32 = _mm_add_epi16(a32, a16);  // a[x] * 32 + 16
305 
306         b = _mm_mullo_epi16(diff, shift);
307         res1 = _mm_add_epi16(a32, b);
308         res1 = _mm_srli_epi16(res1, 5);
309 
310         res16[jj] = _mm_packus_epi16(res, res1);  // 16 8bit values
311       }
312     }
313 
314     dstvec[r] =
315         _mm_blendv_epi8(a_mbase_x, res16[0],
316                         *(__m128i *)Mask[0][base_max_diff]);  // 16 8bit values
317 
318     dstvec_h[r] =
319         _mm_blendv_epi8(a_mbase_x, res16[1],
320                         *(__m128i *)Mask[1][base_max_diff]);  // 16 8bit values
321     x += dx;
322   }
323 }
324 
dr_prediction_z1_32xN_sse4_1(int N,uint8_t * dst,ptrdiff_t stride,const uint8_t * above,int upsample_above,int dx)325 static void dr_prediction_z1_32xN_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
326                                          const uint8_t *above,
327                                          int upsample_above, int dx) {
328   __m128i dstvec[64], dstvec_h[64];
329   dr_prediction_z1_32xN_internal_sse4_1(N, dstvec, dstvec_h, above,
330                                         upsample_above, dx);
331   for (int i = 0; i < N; i++) {
332     _mm_storeu_si128((__m128i *)(dst + stride * i), dstvec[i]);
333     _mm_storeu_si128((__m128i *)(dst + stride * i + 16), dstvec_h[i]);
334   }
335 }
336 
dr_prediction_z1_64xN_sse4_1(int N,uint8_t * dst,ptrdiff_t stride,const uint8_t * above,int upsample_above,int dx)337 static void dr_prediction_z1_64xN_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
338                                          const uint8_t *above,
339                                          int upsample_above, int dx) {
340   // here upsample_above is 0 by design of av1_use_intra_edge_upsample
341   (void)upsample_above;
342   const int frac_bits = 6;
343   const int max_base_x = ((64 + N) - 1);
344 
345   // pre-filter above pixels
346   // store in temp buffers:
347   //   above[x] * 32 + 16
348   //   above[x+1] - above[x]
349   // final pixels will be calculated as:
350   //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
351   __m128i a0, a1, a32, a16;
352   __m128i a_mbase_x, diff, c3f;
353   __m128i max_base, base_inc, mask;
354 
355   a16 = _mm_set1_epi16(16);
356   a_mbase_x = _mm_set1_epi8((char)above[max_base_x]);
357   max_base = _mm_set1_epi8(max_base_x);
358   c3f = _mm_set1_epi16(0x3f);
359 
360   int x = dx;
361   for (int r = 0; r < N; r++, dst += stride) {
362     __m128i b, res, res1;
363     int base = x >> frac_bits;
364     if (base >= max_base_x) {
365       for (int i = r; i < N; ++i) {
366         _mm_storeu_si128((__m128i *)dst, a_mbase_x);  // save 32 values
367         _mm_storeu_si128((__m128i *)(dst + 16), a_mbase_x);
368         _mm_storeu_si128((__m128i *)(dst + 32), a_mbase_x);
369         _mm_storeu_si128((__m128i *)(dst + 48), a_mbase_x);
370         dst += stride;
371       }
372       return;
373     }
374 
375     __m128i shift =
376         _mm_srli_epi16(_mm_and_si128(_mm_set1_epi16(x), c3f), 1);  // 8 element
377 
378     __m128i a0_above, a1_above, res_val;
379     for (int j = 0; j < 64; j += 16) {
380       int mdif = max_base_x - (base + j);
381       if (mdif <= 0) {
382         _mm_storeu_si128((__m128i *)(dst + j), a_mbase_x);
383       } else {
384         a0_above =
385             _mm_loadu_si128((__m128i *)(above + base + j));  // load 16 element
386         a1_above = _mm_loadu_si128((__m128i *)(above + base + 1 + j));
387 
388         // lower half
389         a0 = _mm_cvtepu8_epi16(a0_above);
390         a1 = _mm_cvtepu8_epi16(a1_above);
391 
392         diff = _mm_sub_epi16(a1, a0);   // a[x+1] - a[x]
393         a32 = _mm_slli_epi16(a0, 5);    // a[x] * 32
394         a32 = _mm_add_epi16(a32, a16);  // a[x] * 32 + 16
395         b = _mm_mullo_epi16(diff, shift);
396 
397         res = _mm_add_epi16(a32, b);
398         res = _mm_srli_epi16(res, 5);
399 
400         // uppar half
401         a0 = _mm_cvtepu8_epi16(_mm_srli_si128(a0_above, 8));
402         a1 = _mm_cvtepu8_epi16(_mm_srli_si128(a1_above, 8));
403 
404         diff = _mm_sub_epi16(a1, a0);   // a[x+1] - a[x]
405         a32 = _mm_slli_epi16(a0, 5);    // a[x] * 32
406         a32 = _mm_add_epi16(a32, a16);  // a[x] * 32 + 16
407 
408         b = _mm_mullo_epi16(diff, shift);
409         res1 = _mm_add_epi16(a32, b);
410         res1 = _mm_srli_epi16(res1, 5);
411 
412         res = _mm_packus_epi16(res, res1);  // 16 8bit values
413 
414         base_inc =
415             _mm_setr_epi8((int8_t)(base + j), (int8_t)(base + j + 1),
416                           (int8_t)(base + j + 2), (int8_t)(base + j + 3),
417                           (int8_t)(base + j + 4), (int8_t)(base + j + 5),
418                           (int8_t)(base + j + 6), (int8_t)(base + j + 7),
419                           (int8_t)(base + j + 8), (int8_t)(base + j + 9),
420                           (int8_t)(base + j + 10), (int8_t)(base + j + 11),
421                           (int8_t)(base + j + 12), (int8_t)(base + j + 13),
422                           (int8_t)(base + j + 14), (int8_t)(base + j + 15));
423 
424         mask = _mm_cmpgt_epi8(_mm_subs_epu8(max_base, base_inc),
425                               _mm_setzero_si128());
426         res_val = _mm_blendv_epi8(a_mbase_x, res, mask);
427         _mm_storeu_si128((__m128i *)(dst + j), res_val);
428       }
429     }
430     x += dx;
431   }
432 }
433 
434 // Directional prediction, zone 1: 0 < angle < 90
av1_dr_prediction_z1_sse4_1(uint8_t * dst,ptrdiff_t stride,int bw,int bh,const uint8_t * above,const uint8_t * left,int upsample_above,int dx,int dy)435 void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh,
436                                  const uint8_t *above, const uint8_t *left,
437                                  int upsample_above, int dx, int dy) {
438   (void)left;
439   (void)dy;
440   switch (bw) {
441     case 4:
442       dr_prediction_z1_4xN_sse4_1(bh, dst, stride, above, upsample_above, dx);
443       break;
444     case 8:
445       dr_prediction_z1_8xN_sse4_1(bh, dst, stride, above, upsample_above, dx);
446       break;
447     case 16:
448       dr_prediction_z1_16xN_sse4_1(bh, dst, stride, above, upsample_above, dx);
449       break;
450     case 32:
451       dr_prediction_z1_32xN_sse4_1(bh, dst, stride, above, upsample_above, dx);
452       break;
453     case 64:
454       dr_prediction_z1_64xN_sse4_1(bh, dst, stride, above, upsample_above, dx);
455       break;
456     default: assert(0 && "Invalid block size");
457   }
458   return;
459 }
460 
dr_prediction_z2_Nx4_sse4_1(int N,uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left,int upsample_above,int upsample_left,int dx,int dy)461 static void dr_prediction_z2_Nx4_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
462                                         const uint8_t *above,
463                                         const uint8_t *left, int upsample_above,
464                                         int upsample_left, int dx, int dy) {
465   const int min_base_x = -(1 << upsample_above);
466   const int min_base_y = -(1 << upsample_left);
467   const int frac_bits_x = 6 - upsample_above;
468   const int frac_bits_y = 6 - upsample_left;
469 
470   assert(dx > 0);
471   // pre-filter above pixels
472   // store in temp buffers:
473   //   above[x] * 32 + 16
474   //   above[x+1] - above[x]
475   // final pixels will be calculated as:
476   //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
477   __m128i a0_x, a1_x, a32, diff;
478 
479   const __m128i c3f = _mm_set1_epi16(0x3f);
480   const __m128i min_y_base = _mm_set1_epi16(min_base_y);
481   const __m128i c1234 = _mm_setr_epi16(0, 1, 2, 3, 4, 0, 0, 0);
482   const __m128i dy_reg = _mm_set1_epi16(dy);
483   const __m128i a16 = _mm_set1_epi16(16);
484 
485   for (int r = 0; r < N; r++) {
486     __m128i b, res, shift, r6, ydx;
487     __m128i resx, resy, resxy;
488     __m128i a0_above, a1_above;
489     int y = r + 1;
490     int base_x = (-y * dx) >> frac_bits_x;
491     int base_shift = 0;
492     if (base_x < (min_base_x - 1)) {
493       base_shift = (min_base_x - base_x - 1) >> upsample_above;
494     }
495     int base_min_diff =
496         (min_base_x - base_x + upsample_above) >> upsample_above;
497     if (base_min_diff > 4) {
498       base_min_diff = 4;
499     } else {
500       if (base_min_diff < 0) base_min_diff = 0;
501     }
502 
503     if (base_shift > 3) {
504       a0_x = _mm_setzero_si128();
505       a1_x = _mm_setzero_si128();
506       shift = _mm_setzero_si128();
507     } else {
508       a0_above = _mm_loadu_si128((__m128i *)(above + base_x + base_shift));
509       ydx = _mm_set1_epi16(y * dx);
510       r6 = _mm_slli_epi16(c1234, 6);
511 
512       if (upsample_above) {
513         a0_above =
514             _mm_shuffle_epi8(a0_above, *(__m128i *)EvenOddMaskx[base_shift]);
515         a1_above = _mm_srli_si128(a0_above, 8);
516 
517         shift = _mm_srli_epi16(
518             _mm_and_si128(
519                 _mm_slli_epi16(_mm_sub_epi16(r6, ydx), upsample_above), c3f),
520             1);
521       } else {
522         a0_above =
523             _mm_shuffle_epi8(a0_above, *(__m128i *)LoadMaskx[base_shift]);
524         a1_above = _mm_srli_si128(a0_above, 1);
525 
526         shift = _mm_srli_epi16(_mm_and_si128(_mm_sub_epi16(r6, ydx), c3f), 1);
527       }
528       a0_x = _mm_cvtepu8_epi16(a0_above);
529       a1_x = _mm_cvtepu8_epi16(a1_above);
530     }
531     // y calc
532     __m128i a0_y, a1_y, shifty;
533     if (base_x < min_base_x) {
534       DECLARE_ALIGNED(32, int16_t, base_y_c[8]);
535       __m128i y_c, base_y_c_reg, mask, c1234_;
536       c1234_ = _mm_srli_si128(c1234, 2);
537       r6 = _mm_set1_epi16(r << 6);
538       y_c = _mm_sub_epi16(r6, _mm_mullo_epi16(c1234_, dy_reg));
539       base_y_c_reg = _mm_srai_epi16(y_c, frac_bits_y);
540       mask = _mm_cmpgt_epi16(min_y_base, base_y_c_reg);
541       base_y_c_reg = _mm_andnot_si128(mask, base_y_c_reg);
542       _mm_store_si128((__m128i *)base_y_c, base_y_c_reg);
543 
544       a0_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
545                             left[base_y_c[2]], left[base_y_c[3]], 0, 0, 0, 0);
546       base_y_c_reg = _mm_add_epi16(base_y_c_reg, _mm_srli_epi16(a16, 4));
547       _mm_store_si128((__m128i *)base_y_c, base_y_c_reg);
548       a1_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
549                             left[base_y_c[2]], left[base_y_c[3]], 0, 0, 0, 0);
550 
551       if (upsample_left) {
552         shifty = _mm_srli_epi16(
553             _mm_and_si128(_mm_slli_epi16(y_c, upsample_left), c3f), 1);
554       } else {
555         shifty = _mm_srli_epi16(_mm_and_si128(y_c, c3f), 1);
556       }
557       a0_x = _mm_unpacklo_epi64(a0_x, a0_y);
558       a1_x = _mm_unpacklo_epi64(a1_x, a1_y);
559       shift = _mm_unpacklo_epi64(shift, shifty);
560     }
561 
562     diff = _mm_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
563     a32 = _mm_slli_epi16(a0_x, 5);     // a[x] * 32
564     a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
565 
566     b = _mm_mullo_epi16(diff, shift);
567     res = _mm_add_epi16(a32, b);
568     res = _mm_srli_epi16(res, 5);
569 
570     resx = _mm_packus_epi16(res, res);
571     resy = _mm_srli_si128(resx, 4);
572 
573     resxy = _mm_blendv_epi8(resx, resy, *(__m128i *)Mask[0][base_min_diff]);
574     *(int *)(dst) = _mm_cvtsi128_si32(resxy);
575     dst += stride;
576   }
577 }
578 
dr_prediction_z2_Nx8_sse4_1(int N,uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left,int upsample_above,int upsample_left,int dx,int dy)579 static void dr_prediction_z2_Nx8_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
580                                         const uint8_t *above,
581                                         const uint8_t *left, int upsample_above,
582                                         int upsample_left, int dx, int dy) {
583   const int min_base_x = -(1 << upsample_above);
584   const int min_base_y = -(1 << upsample_left);
585   const int frac_bits_x = 6 - upsample_above;
586   const int frac_bits_y = 6 - upsample_left;
587 
588   // pre-filter above pixels
589   // store in temp buffers:
590   //   above[x] * 32 + 16
591   //   above[x+1] - above[x]
592   // final pixels will be calculated as:
593   //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
594   __m128i diff, a32;
595   __m128i a0_x, a1_x, a0_y, a1_y;
596   __m128i a0_above, a1_above;
597 
598   const __m128i a16 = _mm_set1_epi16(16);
599   const __m128i c3f = _mm_set1_epi16(0x3f);
600   const __m128i min_y_base = _mm_set1_epi16(min_base_y);
601   const __m128i dy_reg = _mm_set1_epi16(dy);
602   const __m128i c1234 = _mm_setr_epi16(1, 2, 3, 4, 5, 6, 7, 8);
603 
604   for (int r = 0; r < N; r++) {
605     __m128i b, res, res1, shift;
606     __m128i resx, resy, resxy, r6, ydx;
607 
608     int y = r + 1;
609     int base_x = (-y * dx) >> frac_bits_x;
610     int base_shift = 0;
611     if (base_x < (min_base_x - 1)) {
612       base_shift = (min_base_x - base_x - 1) >> upsample_above;
613     }
614     int base_min_diff =
615         (min_base_x - base_x + upsample_above) >> upsample_above;
616     if (base_min_diff > 8) {
617       base_min_diff = 8;
618     } else {
619       if (base_min_diff < 0) base_min_diff = 0;
620     }
621 
622     if (base_shift > 7) {
623       resx = _mm_setzero_si128();
624     } else {
625       a0_above = _mm_loadu_si128((__m128i *)(above + base_x + base_shift));
626       ydx = _mm_set1_epi16(y * dx);
627       r6 = _mm_slli_epi16(_mm_srli_si128(c1234, 2), 6);
628       if (upsample_above) {
629         a0_above =
630             _mm_shuffle_epi8(a0_above, *(__m128i *)EvenOddMaskx[base_shift]);
631         a1_above = _mm_srli_si128(a0_above, 8);
632 
633         shift = _mm_srli_epi16(
634             _mm_and_si128(
635                 _mm_slli_epi16(_mm_sub_epi16(r6, ydx), upsample_above), c3f),
636             1);
637       } else {
638         a1_above = _mm_srli_si128(a0_above, 1);
639         a0_above =
640             _mm_shuffle_epi8(a0_above, *(__m128i *)LoadMaskx[base_shift]);
641         a1_above =
642             _mm_shuffle_epi8(a1_above, *(__m128i *)LoadMaskx[base_shift]);
643 
644         shift = _mm_srli_epi16(_mm_and_si128(_mm_sub_epi16(r6, ydx), c3f), 1);
645       }
646       a0_x = _mm_cvtepu8_epi16(a0_above);
647       a1_x = _mm_cvtepu8_epi16(a1_above);
648 
649       diff = _mm_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
650       a32 = _mm_slli_epi16(a0_x, 5);     // a[x] * 32
651       a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
652 
653       b = _mm_mullo_epi16(diff, shift);
654       res = _mm_add_epi16(a32, b);
655       res = _mm_srli_epi16(res, 5);
656       resx = _mm_packus_epi16(res, res);
657     }
658 
659     // y calc
660     if (base_x < min_base_x) {
661       DECLARE_ALIGNED(32, int16_t, base_y_c[16]);
662       __m128i y_c, base_y_c_reg, mask;
663       r6 = _mm_set1_epi16(r << 6);
664       y_c = _mm_sub_epi16(r6, _mm_mullo_epi16(c1234, dy_reg));
665       base_y_c_reg = _mm_srai_epi16(y_c, frac_bits_y);
666       mask = _mm_cmpgt_epi16(min_y_base, base_y_c_reg);
667       base_y_c_reg = _mm_andnot_si128(mask, base_y_c_reg);
668       _mm_store_si128((__m128i *)base_y_c, base_y_c_reg);
669 
670       a0_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
671                             left[base_y_c[2]], left[base_y_c[3]],
672                             left[base_y_c[4]], left[base_y_c[5]],
673                             left[base_y_c[6]], left[base_y_c[7]]);
674       base_y_c_reg = _mm_add_epi16(base_y_c_reg, _mm_srli_epi16(a16, 4));
675       _mm_store_si128((__m128i *)base_y_c, base_y_c_reg);
676 
677       a1_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
678                             left[base_y_c[2]], left[base_y_c[3]],
679                             left[base_y_c[4]], left[base_y_c[5]],
680                             left[base_y_c[6]], left[base_y_c[7]]);
681 
682       if (upsample_left) {
683         shift = _mm_srli_epi16(
684             _mm_and_si128(_mm_slli_epi16(y_c, upsample_left), c3f), 1);
685       } else {
686         shift = _mm_srli_epi16(_mm_and_si128(y_c, c3f), 1);
687       }
688 
689       diff = _mm_sub_epi16(a1_y, a0_y);  // a[x+1] - a[x]
690       a32 = _mm_slli_epi16(a0_y, 5);     // a[x] * 32
691       a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
692 
693       b = _mm_mullo_epi16(diff, shift);
694       res1 = _mm_add_epi16(a32, b);
695       res1 = _mm_srli_epi16(res1, 5);
696 
697       resy = _mm_packus_epi16(res1, res1);
698       resxy = _mm_blendv_epi8(resx, resy, *(__m128i *)Mask[0][base_min_diff]);
699       _mm_storel_epi64((__m128i *)dst, resxy);
700     } else {
701       _mm_storel_epi64((__m128i *)dst, resx);
702     }
703 
704     dst += stride;
705   }
706 }
707 
dr_prediction_z2_HxW_sse4_1(int H,int W,uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left,int upsample_above,int upsample_left,int dx,int dy)708 static void dr_prediction_z2_HxW_sse4_1(int H, int W, uint8_t *dst,
709                                         ptrdiff_t stride, const uint8_t *above,
710                                         const uint8_t *left, int upsample_above,
711                                         int upsample_left, int dx, int dy) {
712   // here upsample_above and upsample_left are 0 by design of
713   // av1_use_intra_edge_upsample
714   const int min_base_x = -1;
715   const int min_base_y = -1;
716   (void)upsample_above;
717   (void)upsample_left;
718   const int frac_bits_x = 6;
719   const int frac_bits_y = 6;
720 
721   __m128i a0_x, a1_x, a0_y, a1_y, a0_y_h, a1_y_h, a32;
722   __m128i diff, shifty, shifty_h;
723   __m128i a0_above, a1_above;
724 
725   DECLARE_ALIGNED(32, int16_t, base_y_c[16]);
726   const __m128i a16 = _mm_set1_epi16(16);
727   const __m128i c1 = _mm_srli_epi16(a16, 4);
728   const __m128i min_y_base = _mm_set1_epi16(min_base_y);
729   const __m128i c3f = _mm_set1_epi16(0x3f);
730   const __m128i dy256 = _mm_set1_epi16(dy);
731   const __m128i c0123 = _mm_setr_epi16(0, 1, 2, 3, 4, 5, 6, 7);
732   const __m128i c0123_h = _mm_setr_epi16(8, 9, 10, 11, 12, 13, 14, 15);
733   const __m128i c1234 = _mm_add_epi16(c0123, c1);
734   const __m128i c1234_h = _mm_add_epi16(c0123_h, c1);
735 
736   for (int r = 0; r < H; r++) {
737     __m128i b, res, res1, shift, reg_j, r6, ydx;
738     __m128i resx, resy;
739     __m128i resxy;
740     int y = r + 1;
741     ydx = _mm_set1_epi16((int16_t)(y * dx));
742 
743     int base_x = (-y * dx) >> frac_bits_x;
744     for (int j = 0; j < W; j += 16) {
745       reg_j = _mm_set1_epi16(j);
746       int base_shift = 0;
747       if ((base_x + j) < (min_base_x - 1)) {
748         base_shift = (min_base_x - (base_x + j) - 1);
749       }
750       int base_min_diff = (min_base_x - base_x - j);
751       if (base_min_diff > 16) {
752         base_min_diff = 16;
753       } else {
754         if (base_min_diff < 0) base_min_diff = 0;
755       }
756 
757       if (base_shift < 16) {
758         a0_above =
759             _mm_loadu_si128((__m128i *)(above + base_x + base_shift + j));
760         a1_above =
761             _mm_loadu_si128((__m128i *)(above + base_x + base_shift + 1 + j));
762         a0_above =
763             _mm_shuffle_epi8(a0_above, *(__m128i *)LoadMaskx[base_shift]);
764         a1_above =
765             _mm_shuffle_epi8(a1_above, *(__m128i *)LoadMaskx[base_shift]);
766 
767         a0_x = _mm_cvtepu8_epi16(a0_above);
768         a1_x = _mm_cvtepu8_epi16(a1_above);
769 
770         r6 = _mm_slli_epi16(_mm_add_epi16(c0123, reg_j), 6);
771         shift = _mm_srli_epi16(_mm_and_si128(_mm_sub_epi16(r6, ydx), c3f), 1);
772 
773         diff = _mm_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
774         a32 = _mm_slli_epi16(a0_x, 5);     // a[x] * 32
775         a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
776 
777         b = _mm_mullo_epi16(diff, shift);
778         res = _mm_add_epi16(a32, b);
779         res = _mm_srli_epi16(res, 5);  // 16 16-bit values
780 
781         a0_x = _mm_cvtepu8_epi16(_mm_srli_si128(a0_above, 8));
782         a1_x = _mm_cvtepu8_epi16(_mm_srli_si128(a1_above, 8));
783 
784         r6 = _mm_slli_epi16(_mm_add_epi16(c0123_h, reg_j), 6);
785         shift = _mm_srli_epi16(_mm_and_si128(_mm_sub_epi16(r6, ydx), c3f), 1);
786 
787         diff = _mm_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
788         a32 = _mm_slli_epi16(a0_x, 5);     // a[x] * 32
789         a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
790 
791         b = _mm_mullo_epi16(diff, shift);
792         res1 = _mm_add_epi16(a32, b);
793         res1 = _mm_srli_epi16(res1, 5);  // 16 16-bit values
794 
795         resx = _mm_packus_epi16(res, res1);
796       } else {
797         resx = _mm_setzero_si128();
798       }
799 
800       // y calc
801       if (base_x < min_base_x) {
802         __m128i c_reg, c_reg_h, y_reg, y_reg_h, base_y, base_y_h;
803         __m128i mask, mask_h, mul16, mul16_h;
804         r6 = _mm_set1_epi16(r << 6);
805         c_reg = _mm_add_epi16(reg_j, c1234);
806         c_reg_h = _mm_add_epi16(reg_j, c1234_h);
807         mul16 = _mm_min_epu16(_mm_mullo_epi16(c_reg, dy256),
808                               _mm_srli_epi16(min_y_base, 1));
809         mul16_h = _mm_min_epu16(_mm_mullo_epi16(c_reg_h, dy256),
810                                 _mm_srli_epi16(min_y_base, 1));
811         y_reg = _mm_sub_epi16(r6, mul16);
812         y_reg_h = _mm_sub_epi16(r6, mul16_h);
813 
814         base_y = _mm_srai_epi16(y_reg, frac_bits_y);
815         base_y_h = _mm_srai_epi16(y_reg_h, frac_bits_y);
816         mask = _mm_cmpgt_epi16(min_y_base, base_y);
817         mask_h = _mm_cmpgt_epi16(min_y_base, base_y_h);
818 
819         base_y = _mm_blendv_epi8(base_y, min_y_base, mask);
820         base_y_h = _mm_blendv_epi8(base_y_h, min_y_base, mask_h);
821         int16_t min_y = (int16_t)_mm_extract_epi16(base_y_h, 7);
822         int16_t max_y = (int16_t)_mm_extract_epi16(base_y, 0);
823         int16_t offset_diff = max_y - min_y;
824 
825         if (offset_diff < 16) {
826           __m128i min_y_reg = _mm_set1_epi16(min_y);
827 
828           __m128i base_y_offset = _mm_sub_epi16(base_y, min_y_reg);
829           __m128i base_y_offset_h = _mm_sub_epi16(base_y_h, min_y_reg);
830           __m128i y_offset = _mm_packs_epi16(base_y_offset, base_y_offset_h);
831 
832           __m128i a0_mask = _mm_loadu_si128((__m128i *)(left + min_y));
833           __m128i a1_mask = _mm_loadu_si128((__m128i *)(left + min_y + 1));
834           __m128i LoadMask =
835               _mm_loadu_si128((__m128i *)(LoadMaskz2[offset_diff / 4]));
836 
837           a0_mask = _mm_and_si128(a0_mask, LoadMask);
838           a1_mask = _mm_and_si128(a1_mask, LoadMask);
839 
840           a0_mask = _mm_shuffle_epi8(a0_mask, y_offset);
841           a1_mask = _mm_shuffle_epi8(a1_mask, y_offset);
842           a0_y = _mm_cvtepu8_epi16(a0_mask);
843           a1_y = _mm_cvtepu8_epi16(a1_mask);
844           a0_y_h = _mm_cvtepu8_epi16(_mm_srli_si128(a0_mask, 8));
845           a1_y_h = _mm_cvtepu8_epi16(_mm_srli_si128(a1_mask, 8));
846         } else {
847           base_y = _mm_andnot_si128(mask, base_y);
848           base_y_h = _mm_andnot_si128(mask_h, base_y_h);
849           _mm_store_si128((__m128i *)base_y_c, base_y);
850           _mm_store_si128((__m128i *)&base_y_c[8], base_y_h);
851 
852           a0_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
853                                 left[base_y_c[2]], left[base_y_c[3]],
854                                 left[base_y_c[4]], left[base_y_c[5]],
855                                 left[base_y_c[6]], left[base_y_c[7]]);
856           a0_y_h = _mm_setr_epi16(left[base_y_c[8]], left[base_y_c[9]],
857                                   left[base_y_c[10]], left[base_y_c[11]],
858                                   left[base_y_c[12]], left[base_y_c[13]],
859                                   left[base_y_c[14]], left[base_y_c[15]]);
860           base_y = _mm_add_epi16(base_y, c1);
861           base_y_h = _mm_add_epi16(base_y_h, c1);
862           _mm_store_si128((__m128i *)base_y_c, base_y);
863           _mm_store_si128((__m128i *)&base_y_c[8], base_y_h);
864 
865           a1_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
866                                 left[base_y_c[2]], left[base_y_c[3]],
867                                 left[base_y_c[4]], left[base_y_c[5]],
868                                 left[base_y_c[6]], left[base_y_c[7]]);
869           a1_y_h = _mm_setr_epi16(left[base_y_c[8]], left[base_y_c[9]],
870                                   left[base_y_c[10]], left[base_y_c[11]],
871                                   left[base_y_c[12]], left[base_y_c[13]],
872                                   left[base_y_c[14]], left[base_y_c[15]]);
873         }
874         shifty = _mm_srli_epi16(_mm_and_si128(y_reg, c3f), 1);
875         shifty_h = _mm_srli_epi16(_mm_and_si128(y_reg_h, c3f), 1);
876 
877         diff = _mm_sub_epi16(a1_y, a0_y);  // a[x+1] - a[x]
878         a32 = _mm_slli_epi16(a0_y, 5);     // a[x] * 32
879         a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
880 
881         b = _mm_mullo_epi16(diff, shifty);
882         res = _mm_add_epi16(a32, b);
883         res = _mm_srli_epi16(res, 5);  // 16 16-bit values
884 
885         diff = _mm_sub_epi16(a1_y_h, a0_y_h);  // a[x+1] - a[x]
886         a32 = _mm_slli_epi16(a0_y_h, 5);       // a[x] * 32
887         a32 = _mm_add_epi16(a32, a16);         // a[x] * 32 + 16
888 
889         b = _mm_mullo_epi16(diff, shifty_h);
890         res1 = _mm_add_epi16(a32, b);
891         res1 = _mm_srli_epi16(res1, 5);  // 16 16-bit values
892         resy = _mm_packus_epi16(res, res1);
893       } else {
894         resy = _mm_setzero_si128();
895       }
896       resxy = _mm_blendv_epi8(resx, resy, *(__m128i *)Mask[0][base_min_diff]);
897       _mm_storeu_si128((__m128i *)(dst + j), resxy);
898     }  // for j
899     dst += stride;
900   }
901 }
902 
903 // Directional prediction, zone 2: 90 < angle < 180
av1_dr_prediction_z2_sse4_1(uint8_t * dst,ptrdiff_t stride,int bw,int bh,const uint8_t * above,const uint8_t * left,int upsample_above,int upsample_left,int dx,int dy)904 void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh,
905                                  const uint8_t *above, const uint8_t *left,
906                                  int upsample_above, int upsample_left, int dx,
907                                  int dy) {
908   assert(dx > 0);
909   assert(dy > 0);
910   switch (bw) {
911     case 4:
912       dr_prediction_z2_Nx4_sse4_1(bh, dst, stride, above, left, upsample_above,
913                                   upsample_left, dx, dy);
914       break;
915     case 8:
916       dr_prediction_z2_Nx8_sse4_1(bh, dst, stride, above, left, upsample_above,
917                                   upsample_left, dx, dy);
918       break;
919     default:
920       dr_prediction_z2_HxW_sse4_1(bh, bw, dst, stride, above, left,
921                                   upsample_above, upsample_left, dx, dy);
922   }
923   return;
924 }
925 
926 // z3 functions
dr_prediction_z3_4x4_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)927 static void dr_prediction_z3_4x4_sse4_1(uint8_t *dst, ptrdiff_t stride,
928                                         const uint8_t *left, int upsample_left,
929                                         int dy) {
930   __m128i dstvec[4], d[4];
931 
932   dr_prediction_z1_HxW_internal_sse4_1(4, 4, dstvec, left, upsample_left, dy);
933   transpose4x8_8x4_low_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3],
934                             &d[0], &d[1], &d[2], &d[3]);
935 
936   *(int *)(dst + stride * 0) = _mm_cvtsi128_si32(d[0]);
937   *(int *)(dst + stride * 1) = _mm_cvtsi128_si32(d[1]);
938   *(int *)(dst + stride * 2) = _mm_cvtsi128_si32(d[2]);
939   *(int *)(dst + stride * 3) = _mm_cvtsi128_si32(d[3]);
940   return;
941 }
942 
dr_prediction_z3_8x8_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)943 static void dr_prediction_z3_8x8_sse4_1(uint8_t *dst, ptrdiff_t stride,
944                                         const uint8_t *left, int upsample_left,
945                                         int dy) {
946   __m128i dstvec[8], d[8];
947 
948   dr_prediction_z1_HxW_internal_sse4_1(8, 8, dstvec, left, upsample_left, dy);
949   transpose8x8_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &dstvec[4],
950                     &dstvec[5], &dstvec[6], &dstvec[7], &d[0], &d[1], &d[2],
951                     &d[3]);
952 
953   _mm_storel_epi64((__m128i *)(dst + 0 * stride), d[0]);
954   _mm_storel_epi64((__m128i *)(dst + 1 * stride), _mm_srli_si128(d[0], 8));
955   _mm_storel_epi64((__m128i *)(dst + 2 * stride), d[1]);
956   _mm_storel_epi64((__m128i *)(dst + 3 * stride), _mm_srli_si128(d[1], 8));
957   _mm_storel_epi64((__m128i *)(dst + 4 * stride), d[2]);
958   _mm_storel_epi64((__m128i *)(dst + 5 * stride), _mm_srli_si128(d[2], 8));
959   _mm_storel_epi64((__m128i *)(dst + 6 * stride), d[3]);
960   _mm_storel_epi64((__m128i *)(dst + 7 * stride), _mm_srli_si128(d[3], 8));
961 }
962 
dr_prediction_z3_4x8_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)963 static void dr_prediction_z3_4x8_sse4_1(uint8_t *dst, ptrdiff_t stride,
964                                         const uint8_t *left, int upsample_left,
965                                         int dy) {
966   __m128i dstvec[4], d[8];
967 
968   dr_prediction_z1_HxW_internal_sse4_1(8, 4, dstvec, left, upsample_left, dy);
969   transpose4x8_8x4_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &d[0],
970                         &d[1], &d[2], &d[3], &d[4], &d[5], &d[6], &d[7]);
971   for (int i = 0; i < 8; i++) {
972     *(int *)(dst + stride * i) = _mm_cvtsi128_si32(d[i]);
973   }
974 }
975 
dr_prediction_z3_8x4_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)976 static void dr_prediction_z3_8x4_sse4_1(uint8_t *dst, ptrdiff_t stride,
977                                         const uint8_t *left, int upsample_left,
978                                         int dy) {
979   __m128i dstvec[8], d[4];
980 
981   dr_prediction_z1_HxW_internal_sse4_1(4, 8, dstvec, left, upsample_left, dy);
982   transpose8x8_low_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3],
983                         &dstvec[4], &dstvec[5], &dstvec[6], &dstvec[7], &d[0],
984                         &d[1], &d[2], &d[3]);
985   _mm_storel_epi64((__m128i *)(dst + 0 * stride), d[0]);
986   _mm_storel_epi64((__m128i *)(dst + 1 * stride), d[1]);
987   _mm_storel_epi64((__m128i *)(dst + 2 * stride), d[2]);
988   _mm_storel_epi64((__m128i *)(dst + 3 * stride), d[3]);
989 }
990 
dr_prediction_z3_8x16_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)991 static void dr_prediction_z3_8x16_sse4_1(uint8_t *dst, ptrdiff_t stride,
992                                          const uint8_t *left, int upsample_left,
993                                          int dy) {
994   __m128i dstvec[8], d[8];
995 
996   dr_prediction_z1_HxW_internal_sse4_1(16, 8, dstvec, left, upsample_left, dy);
997   transpose8x16_16x8_sse2(dstvec, dstvec + 1, dstvec + 2, dstvec + 3,
998                           dstvec + 4, dstvec + 5, dstvec + 6, dstvec + 7, d,
999                           d + 1, d + 2, d + 3, d + 4, d + 5, d + 6, d + 7);
1000   for (int i = 0; i < 8; i++) {
1001     _mm_storel_epi64((__m128i *)(dst + i * stride), d[i]);
1002     _mm_storel_epi64((__m128i *)(dst + (i + 8) * stride),
1003                      _mm_srli_si128(d[i], 8));
1004   }
1005 }
1006 
dr_prediction_z3_16x8_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1007 static void dr_prediction_z3_16x8_sse4_1(uint8_t *dst, ptrdiff_t stride,
1008                                          const uint8_t *left, int upsample_left,
1009                                          int dy) {
1010   __m128i dstvec[16], d[16];
1011 
1012   dr_prediction_z1_HxW_internal_sse4_1(8, 16, dstvec, left, upsample_left, dy);
1013   transpose16x8_8x16_sse2(
1014       &dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &dstvec[4], &dstvec[5],
1015       &dstvec[6], &dstvec[7], &dstvec[8], &dstvec[9], &dstvec[10], &dstvec[11],
1016       &dstvec[12], &dstvec[13], &dstvec[14], &dstvec[15], &d[0], &d[1], &d[2],
1017       &d[3], &d[4], &d[5], &d[6], &d[7]);
1018 
1019   for (int i = 0; i < 8; i++) {
1020     _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
1021   }
1022 }
1023 
1024 #if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
dr_prediction_z3_4x16_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1025 static void dr_prediction_z3_4x16_sse4_1(uint8_t *dst, ptrdiff_t stride,
1026                                          const uint8_t *left, int upsample_left,
1027                                          int dy) {
1028   __m128i dstvec[4], d[16];
1029 
1030   dr_prediction_z1_HxW_internal_sse4_1(16, 4, dstvec, left, upsample_left, dy);
1031   transpose4x16_sse2(dstvec, d);
1032   for (int i = 0; i < 16; i++) {
1033     *(int *)(dst + stride * i) = _mm_cvtsi128_si32(d[i]);
1034   }
1035 }
1036 
dr_prediction_z3_16x4_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1037 static void dr_prediction_z3_16x4_sse4_1(uint8_t *dst, ptrdiff_t stride,
1038                                          const uint8_t *left, int upsample_left,
1039                                          int dy) {
1040   __m128i dstvec[16], d[8];
1041 
1042   dr_prediction_z1_HxW_internal_sse4_1(4, 16, dstvec, left, upsample_left, dy);
1043   for (int i = 4; i < 8; i++) {
1044     d[i] = _mm_setzero_si128();
1045   }
1046   transpose16x8_8x16_sse2(
1047       &dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &dstvec[4], &dstvec[5],
1048       &dstvec[6], &dstvec[7], &dstvec[8], &dstvec[9], &dstvec[10], &dstvec[11],
1049       &dstvec[12], &dstvec[13], &dstvec[14], &dstvec[15], &d[0], &d[1], &d[2],
1050       &d[3], &d[4], &d[5], &d[6], &d[7]);
1051 
1052   for (int i = 0; i < 4; i++) {
1053     _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
1054   }
1055 }
1056 
dr_prediction_z3_8x32_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1057 static void dr_prediction_z3_8x32_sse4_1(uint8_t *dst, ptrdiff_t stride,
1058                                          const uint8_t *left, int upsample_left,
1059                                          int dy) {
1060   __m128i dstvec[16], d[16], dstvec_h[16], d_h[16];
1061 
1062   dr_prediction_z1_32xN_internal_sse4_1(8, dstvec, dstvec_h, left,
1063                                         upsample_left, dy);
1064   for (int i = 8; i < 16; i++) {
1065     dstvec[i] = _mm_setzero_si128();
1066     dstvec_h[i] = _mm_setzero_si128();
1067   }
1068   transpose16x16_sse2(dstvec, d);
1069   transpose16x16_sse2(dstvec_h, d_h);
1070 
1071   for (int i = 0; i < 16; i++) {
1072     _mm_storel_epi64((__m128i *)(dst + i * stride), d[i]);
1073   }
1074   for (int i = 0; i < 16; i++) {
1075     _mm_storel_epi64((__m128i *)(dst + (i + 16) * stride), d_h[i]);
1076   }
1077 }
1078 
dr_prediction_z3_32x8_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1079 static void dr_prediction_z3_32x8_sse4_1(uint8_t *dst, ptrdiff_t stride,
1080                                          const uint8_t *left, int upsample_left,
1081                                          int dy) {
1082   __m128i dstvec[32], d[16];
1083 
1084   dr_prediction_z1_HxW_internal_sse4_1(8, 32, dstvec, left, upsample_left, dy);
1085 
1086   transpose16x8_8x16_sse2(
1087       &dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &dstvec[4], &dstvec[5],
1088       &dstvec[6], &dstvec[7], &dstvec[8], &dstvec[9], &dstvec[10], &dstvec[11],
1089       &dstvec[12], &dstvec[13], &dstvec[14], &dstvec[15], &d[0], &d[1], &d[2],
1090       &d[3], &d[4], &d[5], &d[6], &d[7]);
1091   transpose16x8_8x16_sse2(
1092       &dstvec[0 + 16], &dstvec[1 + 16], &dstvec[2 + 16], &dstvec[3 + 16],
1093       &dstvec[4 + 16], &dstvec[5 + 16], &dstvec[6 + 16], &dstvec[7 + 16],
1094       &dstvec[8 + 16], &dstvec[9 + 16], &dstvec[10 + 16], &dstvec[11 + 16],
1095       &dstvec[12 + 16], &dstvec[13 + 16], &dstvec[14 + 16], &dstvec[15 + 16],
1096       &d[0 + 8], &d[1 + 8], &d[2 + 8], &d[3 + 8], &d[4 + 8], &d[5 + 8],
1097       &d[6 + 8], &d[7 + 8]);
1098 
1099   for (int i = 0; i < 8; i++) {
1100     _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
1101     _mm_storeu_si128((__m128i *)(dst + i * stride + 16), d[i + 8]);
1102   }
1103 }
1104 #endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1105 
dr_prediction_z3_16x16_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1106 static void dr_prediction_z3_16x16_sse4_1(uint8_t *dst, ptrdiff_t stride,
1107                                           const uint8_t *left,
1108                                           int upsample_left, int dy) {
1109   __m128i dstvec[16], d[16];
1110 
1111   dr_prediction_z1_HxW_internal_sse4_1(16, 16, dstvec, left, upsample_left, dy);
1112   transpose16x16_sse2(dstvec, d);
1113 
1114   for (int i = 0; i < 16; i++) {
1115     _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
1116   }
1117 }
1118 
dr_prediction_z3_32x32_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1119 static void dr_prediction_z3_32x32_sse4_1(uint8_t *dst, ptrdiff_t stride,
1120                                           const uint8_t *left,
1121                                           int upsample_left, int dy) {
1122   __m128i dstvec[32], d[32], dstvec_h[32], d_h[32];
1123 
1124   dr_prediction_z1_32xN_internal_sse4_1(32, dstvec, dstvec_h, left,
1125                                         upsample_left, dy);
1126   transpose16x16_sse2(dstvec, d);
1127   transpose16x16_sse2(dstvec_h, d_h);
1128   transpose16x16_sse2(dstvec + 16, d + 16);
1129   transpose16x16_sse2(dstvec_h + 16, d_h + 16);
1130   for (int j = 0; j < 16; j++) {
1131     _mm_storeu_si128((__m128i *)(dst + j * stride), d[j]);
1132     _mm_storeu_si128((__m128i *)(dst + j * stride + 16), d[j + 16]);
1133   }
1134   for (int j = 0; j < 16; j++) {
1135     _mm_storeu_si128((__m128i *)(dst + (j + 16) * stride), d_h[j]);
1136     _mm_storeu_si128((__m128i *)(dst + (j + 16) * stride + 16), d_h[j + 16]);
1137   }
1138 }
1139 
dr_prediction_z3_64x64_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1140 static void dr_prediction_z3_64x64_sse4_1(uint8_t *dst, ptrdiff_t stride,
1141                                           const uint8_t *left,
1142                                           int upsample_left, int dy) {
1143   uint8_t dstT[64 * 64];
1144   dr_prediction_z1_64xN_sse4_1(64, dstT, 64, left, upsample_left, dy);
1145   transpose(dstT, 64, dst, stride, 64, 64);
1146 }
1147 
dr_prediction_z3_16x32_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1148 static void dr_prediction_z3_16x32_sse4_1(uint8_t *dst, ptrdiff_t stride,
1149                                           const uint8_t *left,
1150                                           int upsample_left, int dy) {
1151   __m128i dstvec[16], d[16], dstvec_h[16], d_h[16];
1152 
1153   dr_prediction_z1_32xN_internal_sse4_1(16, dstvec, dstvec_h, left,
1154                                         upsample_left, dy);
1155   transpose16x16_sse2(dstvec, d);
1156   transpose16x16_sse2(dstvec_h, d_h);
1157   // store
1158   for (int j = 0; j < 16; j++) {
1159     _mm_storeu_si128((__m128i *)(dst + j * stride), d[j]);
1160     _mm_storeu_si128((__m128i *)(dst + (j + 16) * stride), d_h[j]);
1161   }
1162 }
1163 
dr_prediction_z3_32x16_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1164 static void dr_prediction_z3_32x16_sse4_1(uint8_t *dst, ptrdiff_t stride,
1165                                           const uint8_t *left,
1166                                           int upsample_left, int dy) {
1167   __m128i dstvec[32], d[16];
1168 
1169   dr_prediction_z1_HxW_internal_sse4_1(16, 32, dstvec, left, upsample_left, dy);
1170   for (int i = 0; i < 32; i += 16) {
1171     transpose16x16_sse2((dstvec + i), d);
1172     for (int j = 0; j < 16; j++) {
1173       _mm_storeu_si128((__m128i *)(dst + j * stride + i), d[j]);
1174     }
1175   }
1176 }
1177 
dr_prediction_z3_32x64_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1178 static void dr_prediction_z3_32x64_sse4_1(uint8_t *dst, ptrdiff_t stride,
1179                                           const uint8_t *left,
1180                                           int upsample_left, int dy) {
1181   uint8_t dstT[64 * 32];
1182   dr_prediction_z1_64xN_sse4_1(32, dstT, 64, left, upsample_left, dy);
1183   transpose(dstT, 64, dst, stride, 32, 64);
1184 }
1185 
dr_prediction_z3_64x32_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1186 static void dr_prediction_z3_64x32_sse4_1(uint8_t *dst, ptrdiff_t stride,
1187                                           const uint8_t *left,
1188                                           int upsample_left, int dy) {
1189   uint8_t dstT[32 * 64];
1190   dr_prediction_z1_32xN_sse4_1(64, dstT, 32, left, upsample_left, dy);
1191   transpose(dstT, 32, dst, stride, 64, 32);
1192   return;
1193 }
1194 
1195 #if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
dr_prediction_z3_16x64_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1196 static void dr_prediction_z3_16x64_sse4_1(uint8_t *dst, ptrdiff_t stride,
1197                                           const uint8_t *left,
1198                                           int upsample_left, int dy) {
1199   uint8_t dstT[64 * 16];
1200   dr_prediction_z1_64xN_sse4_1(16, dstT, 64, left, upsample_left, dy);
1201   transpose(dstT, 64, dst, stride, 16, 64);
1202 }
1203 
dr_prediction_z3_64x16_sse4_1(uint8_t * dst,ptrdiff_t stride,const uint8_t * left,int upsample_left,int dy)1204 static void dr_prediction_z3_64x16_sse4_1(uint8_t *dst, ptrdiff_t stride,
1205                                           const uint8_t *left,
1206                                           int upsample_left, int dy) {
1207   __m128i dstvec[64], d[16];
1208 
1209   dr_prediction_z1_HxW_internal_sse4_1(16, 64, dstvec, left, upsample_left, dy);
1210   for (int i = 0; i < 64; i += 16) {
1211     transpose16x16_sse2(dstvec + i, d);
1212     for (int j = 0; j < 16; j++) {
1213       _mm_storeu_si128((__m128i *)(dst + j * stride + i), d[j]);
1214     }
1215   }
1216 }
1217 #endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1218 
av1_dr_prediction_z3_sse4_1(uint8_t * dst,ptrdiff_t stride,int bw,int bh,const uint8_t * above,const uint8_t * left,int upsample_left,int dx,int dy)1219 void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh,
1220                                  const uint8_t *above, const uint8_t *left,
1221                                  int upsample_left, int dx, int dy) {
1222   (void)above;
1223   (void)dx;
1224   assert(dx == 1);
1225   assert(dy > 0);
1226 
1227   if (bw == bh) {
1228     switch (bw) {
1229       case 4:
1230         dr_prediction_z3_4x4_sse4_1(dst, stride, left, upsample_left, dy);
1231         break;
1232       case 8:
1233         dr_prediction_z3_8x8_sse4_1(dst, stride, left, upsample_left, dy);
1234         break;
1235       case 16:
1236         dr_prediction_z3_16x16_sse4_1(dst, stride, left, upsample_left, dy);
1237         break;
1238       case 32:
1239         dr_prediction_z3_32x32_sse4_1(dst, stride, left, upsample_left, dy);
1240         break;
1241       case 64:
1242         dr_prediction_z3_64x64_sse4_1(dst, stride, left, upsample_left, dy);
1243         break;
1244       default: assert(0 && "Invalid block size");
1245     }
1246   } else {
1247     if (bw < bh) {
1248       if (bw + bw == bh) {
1249         switch (bw) {
1250           case 4:
1251             dr_prediction_z3_4x8_sse4_1(dst, stride, left, upsample_left, dy);
1252             break;
1253           case 8:
1254             dr_prediction_z3_8x16_sse4_1(dst, stride, left, upsample_left, dy);
1255             break;
1256           case 16:
1257             dr_prediction_z3_16x32_sse4_1(dst, stride, left, upsample_left, dy);
1258             break;
1259           case 32:
1260             dr_prediction_z3_32x64_sse4_1(dst, stride, left, upsample_left, dy);
1261             break;
1262           default: assert(0 && "Invalid block size");
1263         }
1264       } else {
1265         switch (bw) {
1266 #if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1267           case 4:
1268             dr_prediction_z3_4x16_sse4_1(dst, stride, left, upsample_left, dy);
1269             break;
1270           case 8:
1271             dr_prediction_z3_8x32_sse4_1(dst, stride, left, upsample_left, dy);
1272             break;
1273           case 16:
1274             dr_prediction_z3_16x64_sse4_1(dst, stride, left, upsample_left, dy);
1275             break;
1276           default: assert(0 && "Invalid block size");
1277 #endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1278         }
1279       }
1280     } else {
1281       if (bh + bh == bw) {
1282         switch (bh) {
1283           case 4:
1284             dr_prediction_z3_8x4_sse4_1(dst, stride, left, upsample_left, dy);
1285             break;
1286           case 8:
1287             dr_prediction_z3_16x8_sse4_1(dst, stride, left, upsample_left, dy);
1288             break;
1289           case 16:
1290             dr_prediction_z3_32x16_sse4_1(dst, stride, left, upsample_left, dy);
1291             break;
1292           case 32:
1293             dr_prediction_z3_64x32_sse4_1(dst, stride, left, upsample_left, dy);
1294             break;
1295           default: assert(0 && "Invalid block size");
1296         }
1297       } else {
1298         switch (bh) {
1299 #if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1300           case 4:
1301             dr_prediction_z3_16x4_sse4_1(dst, stride, left, upsample_left, dy);
1302             break;
1303           case 8:
1304             dr_prediction_z3_32x8_sse4_1(dst, stride, left, upsample_left, dy);
1305             break;
1306           case 16:
1307             dr_prediction_z3_64x16_sse4_1(dst, stride, left, upsample_left, dy);
1308             break;
1309           default: assert(0 && "Invalid block size");
1310 #endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1311         }
1312       }
1313     }
1314   }
1315 }
1316