19 void aom_highbd_h_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_h_predictor_4x4_sse2()  argument
29   _mm_storel_epi64((__m128i *)dst, row0);  in aom_highbd_h_predictor_4x4_sse2()
30   dst += stride;  in aom_highbd_h_predictor_4x4_sse2()
31   _mm_storel_epi64((__m128i *)dst, row1);  in aom_highbd_h_predictor_4x4_sse2()
32   dst += stride;  in aom_highbd_h_predictor_4x4_sse2()
33   _mm_storel_epi64((__m128i *)dst, row2);  in aom_highbd_h_predictor_4x4_sse2()
34   dst += stride;  in aom_highbd_h_predictor_4x4_sse2()
35   _mm_storel_epi64((__m128i *)dst, row3);  in aom_highbd_h_predictor_4x4_sse2()
38 void aom_highbd_h_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_h_predictor_4x8_sse2()  argument
41   aom_highbd_h_predictor_4x4_sse2(dst, stride, above, left, bd);  in aom_highbd_h_predictor_4x8_sse2()
42   dst += stride << 2;  in aom_highbd_h_predictor_4x8_sse2()
44   aom_highbd_h_predictor_4x4_sse2(dst, stride, above, left, bd);  in aom_highbd_h_predictor_4x8_sse2()
47 void aom_highbd_h_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_h_predictor_8x4_sse2()  argument
57   _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row0, row0));  in aom_highbd_h_predictor_8x4_sse2()
58   dst += stride;  in aom_highbd_h_predictor_8x4_sse2()
59   _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row1, row1));  in aom_highbd_h_predictor_8x4_sse2()
60   dst += stride;  in aom_highbd_h_predictor_8x4_sse2()
61   _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row2, row2));  in aom_highbd_h_predictor_8x4_sse2()
62   dst += stride;  in aom_highbd_h_predictor_8x4_sse2()
63   _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row3, row3));  in aom_highbd_h_predictor_8x4_sse2()
66 void aom_highbd_h_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_h_predictor_8x8_sse2()  argument
80   _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row0, row0));  in aom_highbd_h_predictor_8x8_sse2()
81   dst += stride;  in aom_highbd_h_predictor_8x8_sse2()
82   _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row1, row1));  in aom_highbd_h_predictor_8x8_sse2()
83   dst += stride;  in aom_highbd_h_predictor_8x8_sse2()
84   _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row2, row2));  in aom_highbd_h_predictor_8x8_sse2()
85   dst += stride;  in aom_highbd_h_predictor_8x8_sse2()
86   _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row3, row3));  in aom_highbd_h_predictor_8x8_sse2()
87   dst += stride;  in aom_highbd_h_predictor_8x8_sse2()
88   _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row4, row4));  in aom_highbd_h_predictor_8x8_sse2()
89   dst += stride;  in aom_highbd_h_predictor_8x8_sse2()
90   _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row5, row5));  in aom_highbd_h_predictor_8x8_sse2()
91   dst += stride;  in aom_highbd_h_predictor_8x8_sse2()
92   _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row6, row6));  in aom_highbd_h_predictor_8x8_sse2()
93   dst += stride;  in aom_highbd_h_predictor_8x8_sse2()
94   _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row7, row7));  in aom_highbd_h_predictor_8x8_sse2()
97 void aom_highbd_h_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_h_predictor_8x16_sse2()  argument
100   aom_highbd_h_predictor_8x8_sse2(dst, stride, above, left, bd);  in aom_highbd_h_predictor_8x16_sse2()
101   dst += stride << 3;  in aom_highbd_h_predictor_8x16_sse2()
103   aom_highbd_h_predictor_8x8_sse2(dst, stride, above, left, bd);  in aom_highbd_h_predictor_8x16_sse2()
106 static INLINE void h_store_16_unpacklo(uint16_t **dst, const ptrdiff_t stride,  in h_store_16_unpacklo()  argument
109   _mm_store_si128((__m128i *)*dst, val);  in h_store_16_unpacklo()
110   _mm_store_si128((__m128i *)(*dst + 8), val);  in h_store_16_unpacklo()
111   *dst += stride;  in h_store_16_unpacklo()
114 static INLINE void h_store_16_unpackhi(uint16_t **dst, const ptrdiff_t stride,  in h_store_16_unpackhi()  argument
117   _mm_store_si128((__m128i *)(*dst), val);  in h_store_16_unpackhi()
118   _mm_store_si128((__m128i *)(*dst + 8), val);  in h_store_16_unpackhi()
119   *dst += stride;  in h_store_16_unpackhi()
122 static INLINE void h_predictor_16x8(uint16_t *dst, ptrdiff_t stride,  in h_predictor_16x8()  argument
133   h_store_16_unpacklo(&dst, stride, &row0);  in h_predictor_16x8()
134   h_store_16_unpacklo(&dst, stride, &row1);  in h_predictor_16x8()
135   h_store_16_unpacklo(&dst, stride, &row2);  in h_predictor_16x8()
136   h_store_16_unpacklo(&dst, stride, &row3);  in h_predictor_16x8()
137   h_store_16_unpackhi(&dst, stride, &row4);  in h_predictor_16x8()
138   h_store_16_unpackhi(&dst, stride, &row5);  in h_predictor_16x8()
139   h_store_16_unpackhi(&dst, stride, &row6);  in h_predictor_16x8()
140   h_store_16_unpackhi(&dst, stride, &row7);  in h_predictor_16x8()
143 void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_h_predictor_16x8_sse2()  argument
148   h_predictor_16x8(dst, stride, left);  in aom_highbd_h_predictor_16x8_sse2()
151 void aom_highbd_h_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_h_predictor_16x16_sse2()  argument
159     h_predictor_16x8(dst, stride, left);  in aom_highbd_h_predictor_16x16_sse2()
160     dst += stride << 3;  in aom_highbd_h_predictor_16x16_sse2()
164 void aom_highbd_h_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_h_predictor_16x32_sse2()  argument
172     h_predictor_16x8(dst, stride, left);  in aom_highbd_h_predictor_16x32_sse2()
173     dst += stride << 3;  in aom_highbd_h_predictor_16x32_sse2()
177 static INLINE void h_store_32_unpacklo(uint16_t **dst, const ptrdiff_t stride,  in h_store_32_unpacklo()  argument
180   _mm_store_si128((__m128i *)(*dst), val);  in h_store_32_unpacklo()
181   _mm_store_si128((__m128i *)(*dst + 8), val);  in h_store_32_unpacklo()
182   _mm_store_si128((__m128i *)(*dst + 16), val);  in h_store_32_unpacklo()
183   _mm_store_si128((__m128i *)(*dst + 24), val);  in h_store_32_unpacklo()
184   *dst += stride;  in h_store_32_unpacklo()
187 static INLINE void h_store_32_unpackhi(uint16_t **dst, const ptrdiff_t stride,  in h_store_32_unpackhi()  argument
190   _mm_store_si128((__m128i *)(*dst), val);  in h_store_32_unpackhi()
191   _mm_store_si128((__m128i *)(*dst + 8), val);  in h_store_32_unpackhi()
192   _mm_store_si128((__m128i *)(*dst + 16), val);  in h_store_32_unpackhi()
193   _mm_store_si128((__m128i *)(*dst + 24), val);  in h_store_32_unpackhi()
194   *dst += stride;  in h_store_32_unpackhi()
197 static INLINE void h_predictor_32x8(uint16_t *dst, ptrdiff_t stride,  in h_predictor_32x8()  argument
208   h_store_32_unpacklo(&dst, stride, &row0);  in h_predictor_32x8()
209   h_store_32_unpacklo(&dst, stride, &row1);  in h_predictor_32x8()
210   h_store_32_unpacklo(&dst, stride, &row2);  in h_predictor_32x8()
211   h_store_32_unpacklo(&dst, stride, &row3);  in h_predictor_32x8()
212   h_store_32_unpackhi(&dst, stride, &row4);  in h_predictor_32x8()
213   h_store_32_unpackhi(&dst, stride, &row5);  in h_predictor_32x8()
214   h_store_32_unpackhi(&dst, stride, &row6);  in h_predictor_32x8()
215   h_store_32_unpackhi(&dst, stride, &row7);  in h_predictor_32x8()
218 void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_h_predictor_32x16_sse2()  argument
226     h_predictor_32x8(dst, stride, left);  in aom_highbd_h_predictor_32x16_sse2()
227     dst += stride << 3;  in aom_highbd_h_predictor_32x16_sse2()
231 void aom_highbd_h_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_h_predictor_32x32_sse2()  argument
239     h_predictor_32x8(dst, stride, left);  in aom_highbd_h_predictor_32x32_sse2()
240     dst += stride << 3;  in aom_highbd_h_predictor_32x32_sse2()
256 static INLINE void dc_store_4x4(uint16_t *dst, ptrdiff_t stride,  in dc_store_4x4()  argument
260   for (i = 0; i < 4; ++i, dst += stride) {  in dc_store_4x4()
261     _mm_storel_epi64((__m128i *)dst, dc_dup);  in dc_store_4x4()
265 void aom_highbd_dc_left_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_left_predictor_4x4_sse2()  argument
273   dc_store_4x4(dst, stride, &dc);  in aom_highbd_dc_left_predictor_4x4_sse2()
276 void aom_highbd_dc_top_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_top_predictor_4x4_sse2()  argument
284   dc_store_4x4(dst, stride, &dc);  in aom_highbd_dc_top_predictor_4x4_sse2()
287 void aom_highbd_dc_128_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_128_predictor_4x4_sse2()  argument
294   dc_store_4x4(dst, stride, &dc_dup);  in aom_highbd_dc_128_predictor_4x4_sse2()
300 static INLINE void dc_store_4x8(uint16_t *dst, ptrdiff_t stride,  in dc_store_4x8()  argument
304   for (i = 0; i < 8; ++i, dst += stride) {  in dc_store_4x8()
305     _mm_storel_epi64((__m128i *)dst, dc_dup);  in dc_store_4x8()
319 void aom_highbd_dc_left_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_left_predictor_4x8_sse2()  argument
327   dc_store_4x8(dst, stride, &dc);  in aom_highbd_dc_left_predictor_4x8_sse2()
330 void aom_highbd_dc_top_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_top_predictor_4x8_sse2()  argument
338   dc_store_4x8(dst, stride, &dc);  in aom_highbd_dc_top_predictor_4x8_sse2()
341 void aom_highbd_dc_128_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_128_predictor_4x8_sse2()  argument
348   dc_store_4x8(dst, stride, &dc_dup);  in aom_highbd_dc_128_predictor_4x8_sse2()
354 static INLINE void dc_store_8xh(uint16_t *dst, ptrdiff_t stride, int height,  in dc_store_8xh()  argument
359   for (i = 0; i < height; ++i, dst += stride) {  in dc_store_8xh()
360     _mm_store_si128((__m128i *)dst, dc_dup);  in dc_store_8xh()
367 static INLINE void dc_top_predictor_8xh(uint16_t *dst, ptrdiff_t stride,  in dc_top_predictor_8xh()  argument
372   dc_store_8xh(dst, stride, height, &dc);  in dc_top_predictor_8xh()
375 void aom_highbd_dc_top_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_top_predictor_8x4_sse2()  argument
380   dc_top_predictor_8xh(dst, stride, 4, above);  in aom_highbd_dc_top_predictor_8x4_sse2()
383 void aom_highbd_dc_top_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_top_predictor_8x8_sse2()  argument
388   dc_top_predictor_8xh(dst, stride, 8, above);  in aom_highbd_dc_top_predictor_8x8_sse2()
391 void aom_highbd_dc_top_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_top_predictor_8x16_sse2()  argument
396   dc_top_predictor_8xh(dst, stride, 16, above);  in aom_highbd_dc_top_predictor_8x16_sse2()
402 void aom_highbd_dc_left_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_left_predictor_8x4_sse2()  argument
410   dc_store_8xh(dst, stride, 4, &dc);  in aom_highbd_dc_left_predictor_8x4_sse2()
413 void aom_highbd_dc_left_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_left_predictor_8x8_sse2()  argument
421   dc_store_8xh(dst, stride, 8, &dc);  in aom_highbd_dc_left_predictor_8x8_sse2()
431 void aom_highbd_dc_left_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_left_predictor_8x16_sse2()  argument
439   dc_store_8xh(dst, stride, 16, &dc);  in aom_highbd_dc_left_predictor_8x16_sse2()
445 static INLINE void dc_128_predictor_8xh(uint16_t *dst, ptrdiff_t stride,  in dc_128_predictor_8xh()  argument
449   dc_store_8xh(dst, stride, height, &dc_dup);  in dc_128_predictor_8xh()
452 void aom_highbd_dc_128_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_128_predictor_8x4_sse2()  argument
457   dc_128_predictor_8xh(dst, stride, 4, bd);  in aom_highbd_dc_128_predictor_8x4_sse2()
460 void aom_highbd_dc_128_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_128_predictor_8x8_sse2()  argument
465   dc_128_predictor_8xh(dst, stride, 8, bd);  in aom_highbd_dc_128_predictor_8x8_sse2()
468 void aom_highbd_dc_128_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_128_predictor_8x16_sse2()  argument
473   dc_128_predictor_8xh(dst, stride, 16, bd);  in aom_highbd_dc_128_predictor_8x16_sse2()
479 static INLINE void dc_store_16xh(uint16_t *dst, ptrdiff_t stride, int height,  in dc_store_16xh()  argument
484   for (i = 0; i < height; ++i, dst += stride) {  in dc_store_16xh()
485     _mm_store_si128((__m128i *)dst, dc_dup);  in dc_store_16xh()
486     _mm_store_si128((__m128i *)(dst + 8), dc_dup);  in dc_store_16xh()
493 void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_left_predictor_16x8_sse2()  argument
501   dc_store_16xh(dst, stride, 8, &dc);  in aom_highbd_dc_left_predictor_16x8_sse2()
504 void aom_highbd_dc_left_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_left_predictor_16x16_sse2()  argument
512   dc_store_16xh(dst, stride, 16, &dc);  in aom_highbd_dc_left_predictor_16x16_sse2()
525 void aom_highbd_dc_left_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_left_predictor_16x32_sse2()  argument
533   dc_store_16xh(dst, stride, 32, &dc);  in aom_highbd_dc_left_predictor_16x32_sse2()
539 void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_top_predictor_16x8_sse2()  argument
547   dc_store_16xh(dst, stride, 8, &dc);  in aom_highbd_dc_top_predictor_16x8_sse2()
550 void aom_highbd_dc_top_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_top_predictor_16x16_sse2()  argument
558   dc_store_16xh(dst, stride, 16, &dc);  in aom_highbd_dc_top_predictor_16x16_sse2()
561 void aom_highbd_dc_top_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_top_predictor_16x32_sse2()  argument
569   dc_store_16xh(dst, stride, 32, &dc);  in aom_highbd_dc_top_predictor_16x32_sse2()
575 void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_128_predictor_16x8_sse2()  argument
582   dc_store_16xh(dst, stride, 8, &dc_dup);  in aom_highbd_dc_128_predictor_16x8_sse2()
585 void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_128_predictor_16x16_sse2()  argument
592   dc_store_16xh(dst, stride, 16, &dc_dup);  in aom_highbd_dc_128_predictor_16x16_sse2()
595 void aom_highbd_dc_128_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_128_predictor_16x32_sse2()  argument
602   dc_store_16xh(dst, stride, 32, &dc_dup);  in aom_highbd_dc_128_predictor_16x32_sse2()
608 static INLINE void dc_store_32xh(uint16_t *dst, ptrdiff_t stride, int height,  in dc_store_32xh()  argument
613   for (i = 0; i < height; ++i, dst += stride) {  in dc_store_32xh()
614     _mm_store_si128((__m128i *)dst, dc_dup);  in dc_store_32xh()
615     _mm_store_si128((__m128i *)(dst + 8), dc_dup);  in dc_store_32xh()
616     _mm_store_si128((__m128i *)(dst + 16), dc_dup);  in dc_store_32xh()
617     _mm_store_si128((__m128i *)(dst + 24), dc_dup);  in dc_store_32xh()
621 void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_left_predictor_32x16_sse2()  argument
629   dc_store_32xh(dst, stride, 16, &dc);  in aom_highbd_dc_left_predictor_32x16_sse2()
632 void aom_highbd_dc_left_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_left_predictor_32x32_sse2()  argument
640   dc_store_32xh(dst, stride, 32, &dc);  in aom_highbd_dc_left_predictor_32x32_sse2()
643 void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_top_predictor_32x16_sse2()  argument
651   dc_store_32xh(dst, stride, 16, &dc);  in aom_highbd_dc_top_predictor_32x16_sse2()
654 void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_128_predictor_32x16_sse2()  argument
661   dc_store_32xh(dst, stride, 16, &dc_dup);  in aom_highbd_dc_128_predictor_32x16_sse2()
664 void aom_highbd_dc_top_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_top_predictor_32x32_sse2()  argument
672   dc_store_32xh(dst, stride, 32, &dc);  in aom_highbd_dc_top_predictor_32x32_sse2()
675 void aom_highbd_dc_128_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_128_predictor_32x32_sse2()  argument
682   dc_store_32xh(dst, stride, 32, &dc_dup);  in aom_highbd_dc_128_predictor_32x32_sse2()
688 void aom_highbd_v_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_v_predictor_4x8_sse2()  argument
696     _mm_storel_epi64((__m128i *)dst, above_u16);  in aom_highbd_v_predictor_4x8_sse2()
697     _mm_storel_epi64((__m128i *)(dst + stride), above_u16);  in aom_highbd_v_predictor_4x8_sse2()
698     _mm_storel_epi64((__m128i *)(dst + 2 * stride), above_u16);  in aom_highbd_v_predictor_4x8_sse2()
699     _mm_storel_epi64((__m128i *)(dst + 3 * stride), above_u16);  in aom_highbd_v_predictor_4x8_sse2()
700     dst += stride << 2;  in aom_highbd_v_predictor_4x8_sse2()
704 void aom_highbd_v_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_v_predictor_8x4_sse2()  argument
710   _mm_store_si128((__m128i *)dst, above_u16);  in aom_highbd_v_predictor_8x4_sse2()
711   _mm_store_si128((__m128i *)(dst + stride), above_u16);  in aom_highbd_v_predictor_8x4_sse2()
712   _mm_store_si128((__m128i *)(dst + 2 * stride), above_u16);  in aom_highbd_v_predictor_8x4_sse2()
713   _mm_store_si128((__m128i *)(dst + 3 * stride), above_u16);  in aom_highbd_v_predictor_8x4_sse2()
716 void aom_highbd_v_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_v_predictor_8x16_sse2()  argument
724     _mm_store_si128((__m128i *)dst, above_u16);  in aom_highbd_v_predictor_8x16_sse2()
725     _mm_store_si128((__m128i *)(dst + stride), above_u16);  in aom_highbd_v_predictor_8x16_sse2()
726     _mm_store_si128((__m128i *)(dst + 2 * stride), above_u16);  in aom_highbd_v_predictor_8x16_sse2()
727     _mm_store_si128((__m128i *)(dst + 3 * stride), above_u16);  in aom_highbd_v_predictor_8x16_sse2()
728     dst += stride << 2;  in aom_highbd_v_predictor_8x16_sse2()
732 void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_v_predictor_16x8_sse2()  argument
741     _mm_store_si128((__m128i *)dst, above0_u16);  in aom_highbd_v_predictor_16x8_sse2()
742     _mm_store_si128((__m128i *)(dst + 8), above1_u16);  in aom_highbd_v_predictor_16x8_sse2()
743     dst += stride;  in aom_highbd_v_predictor_16x8_sse2()
744     _mm_store_si128((__m128i *)dst, above0_u16);  in aom_highbd_v_predictor_16x8_sse2()
745     _mm_store_si128((__m128i *)(dst + 8), above1_u16);  in aom_highbd_v_predictor_16x8_sse2()
746     dst += stride;  in aom_highbd_v_predictor_16x8_sse2()
747     _mm_store_si128((__m128i *)dst, above0_u16);  in aom_highbd_v_predictor_16x8_sse2()
748     _mm_store_si128((__m128i *)(dst + 8), above1_u16);  in aom_highbd_v_predictor_16x8_sse2()
749     dst += stride;  in aom_highbd_v_predictor_16x8_sse2()
750     _mm_store_si128((__m128i *)dst, above0_u16);  in aom_highbd_v_predictor_16x8_sse2()
751     _mm_store_si128((__m128i *)(dst + 8), above1_u16);  in aom_highbd_v_predictor_16x8_sse2()
752     dst += stride;  in aom_highbd_v_predictor_16x8_sse2()
756 void aom_highbd_v_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_v_predictor_16x32_sse2()  argument
765     _mm_store_si128((__m128i *)dst, above0_u16);  in aom_highbd_v_predictor_16x32_sse2()
766     _mm_store_si128((__m128i *)(dst + 8), above1_u16);  in aom_highbd_v_predictor_16x32_sse2()
767     dst += stride;  in aom_highbd_v_predictor_16x32_sse2()
768     _mm_store_si128((__m128i *)dst, above0_u16);  in aom_highbd_v_predictor_16x32_sse2()
769     _mm_store_si128((__m128i *)(dst + 8), above1_u16);  in aom_highbd_v_predictor_16x32_sse2()
770     dst += stride;  in aom_highbd_v_predictor_16x32_sse2()
771     _mm_store_si128((__m128i *)dst, above0_u16);  in aom_highbd_v_predictor_16x32_sse2()
772     _mm_store_si128((__m128i *)(dst + 8), above1_u16);  in aom_highbd_v_predictor_16x32_sse2()
773     dst += stride;  in aom_highbd_v_predictor_16x32_sse2()
774     _mm_store_si128((__m128i *)dst, above0_u16);  in aom_highbd_v_predictor_16x32_sse2()
775     _mm_store_si128((__m128i *)(dst + 8), above1_u16);  in aom_highbd_v_predictor_16x32_sse2()
776     dst += stride;  in aom_highbd_v_predictor_16x32_sse2()
780 void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_v_predictor_32x16_sse2()  argument
791     _mm_store_si128((__m128i *)dst, above0_u16);  in aom_highbd_v_predictor_32x16_sse2()
792     _mm_store_si128((__m128i *)(dst + 8), above1_u16);  in aom_highbd_v_predictor_32x16_sse2()
793     _mm_store_si128((__m128i *)(dst + 16), above2_u16);  in aom_highbd_v_predictor_32x16_sse2()
794     _mm_store_si128((__m128i *)(dst + 24), above3_u16);  in aom_highbd_v_predictor_32x16_sse2()
795     dst += stride;  in aom_highbd_v_predictor_32x16_sse2()
796     _mm_store_si128((__m128i *)dst, above0_u16);  in aom_highbd_v_predictor_32x16_sse2()
797     _mm_store_si128((__m128i *)(dst + 8), above1_u16);  in aom_highbd_v_predictor_32x16_sse2()
798     _mm_store_si128((__m128i *)(dst + 16), above2_u16);  in aom_highbd_v_predictor_32x16_sse2()
799     _mm_store_si128((__m128i *)(dst + 24), above3_u16);  in aom_highbd_v_predictor_32x16_sse2()
800     dst += stride;  in aom_highbd_v_predictor_32x16_sse2()
801     _mm_store_si128((__m128i *)dst, above0_u16);  in aom_highbd_v_predictor_32x16_sse2()
802     _mm_store_si128((__m128i *)(dst + 8), above1_u16);  in aom_highbd_v_predictor_32x16_sse2()
803     _mm_store_si128((__m128i *)(dst + 16), above2_u16);  in aom_highbd_v_predictor_32x16_sse2()
804     _mm_store_si128((__m128i *)(dst + 24), above3_u16);  in aom_highbd_v_predictor_32x16_sse2()
805     dst += stride;  in aom_highbd_v_predictor_32x16_sse2()
806     _mm_store_si128((__m128i *)dst, above0_u16);  in aom_highbd_v_predictor_32x16_sse2()
807     _mm_store_si128((__m128i *)(dst + 8), above1_u16);  in aom_highbd_v_predictor_32x16_sse2()
808     _mm_store_si128((__m128i *)(dst + 16), above2_u16);  in aom_highbd_v_predictor_32x16_sse2()
809     _mm_store_si128((__m128i *)(dst + 24), above3_u16);  in aom_highbd_v_predictor_32x16_sse2()
810     dst += stride;  in aom_highbd_v_predictor_32x16_sse2()
817 void aom_highbd_dc_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_predictor_4x8_sse2()  argument
831     _mm_storel_epi64((__m128i *)dst, row);  in aom_highbd_dc_predictor_4x8_sse2()
832     dst += stride;  in aom_highbd_dc_predictor_4x8_sse2()
833     _mm_storel_epi64((__m128i *)dst, row);  in aom_highbd_dc_predictor_4x8_sse2()
834     dst += stride;  in aom_highbd_dc_predictor_4x8_sse2()
838 void aom_highbd_dc_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_predictor_8x4_sse2()  argument
851   _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_8x4_sse2()
852   dst += stride;  in aom_highbd_dc_predictor_8x4_sse2()
853   _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_8x4_sse2()
854   dst += stride;  in aom_highbd_dc_predictor_8x4_sse2()
855   _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_8x4_sse2()
856   dst += stride;  in aom_highbd_dc_predictor_8x4_sse2()
857   _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_8x4_sse2()
860 void aom_highbd_dc_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_predictor_8x16_sse2()  argument
876     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_8x16_sse2()
877     dst += stride;  in aom_highbd_dc_predictor_8x16_sse2()
878     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_8x16_sse2()
879     dst += stride;  in aom_highbd_dc_predictor_8x16_sse2()
880     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_8x16_sse2()
881     dst += stride;  in aom_highbd_dc_predictor_8x16_sse2()
882     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_8x16_sse2()
883     dst += stride;  in aom_highbd_dc_predictor_8x16_sse2()
887 void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_predictor_16x8_sse2()  argument
903     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_16x8_sse2()
904     _mm_store_si128((__m128i *)(dst + 8), row);  in aom_highbd_dc_predictor_16x8_sse2()
905     dst += stride;  in aom_highbd_dc_predictor_16x8_sse2()
906     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_16x8_sse2()
907     _mm_store_si128((__m128i *)(dst + 8), row);  in aom_highbd_dc_predictor_16x8_sse2()
908     dst += stride;  in aom_highbd_dc_predictor_16x8_sse2()
909     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_16x8_sse2()
910     _mm_store_si128((__m128i *)(dst + 8), row);  in aom_highbd_dc_predictor_16x8_sse2()
911     dst += stride;  in aom_highbd_dc_predictor_16x8_sse2()
912     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_16x8_sse2()
913     _mm_store_si128((__m128i *)(dst + 8), row);  in aom_highbd_dc_predictor_16x8_sse2()
914     dst += stride;  in aom_highbd_dc_predictor_16x8_sse2()
918 void aom_highbd_dc_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_predictor_16x32_sse2()  argument
933     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_16x32_sse2()
934     _mm_store_si128((__m128i *)(dst + 8), row);  in aom_highbd_dc_predictor_16x32_sse2()
935     dst += stride;  in aom_highbd_dc_predictor_16x32_sse2()
936     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_16x32_sse2()
937     _mm_store_si128((__m128i *)(dst + 8), row);  in aom_highbd_dc_predictor_16x32_sse2()
938     dst += stride;  in aom_highbd_dc_predictor_16x32_sse2()
939     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_16x32_sse2()
940     _mm_store_si128((__m128i *)(dst + 8), row);  in aom_highbd_dc_predictor_16x32_sse2()
941     dst += stride;  in aom_highbd_dc_predictor_16x32_sse2()
942     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_16x32_sse2()
943     _mm_store_si128((__m128i *)(dst + 8), row);  in aom_highbd_dc_predictor_16x32_sse2()
944     dst += stride;  in aom_highbd_dc_predictor_16x32_sse2()
948 void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,  in aom_highbd_dc_predictor_32x16_sse2()  argument
963     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_32x16_sse2()
964     _mm_store_si128((__m128i *)(dst + 8), row);  in aom_highbd_dc_predictor_32x16_sse2()
965     _mm_store_si128((__m128i *)(dst + 16), row);  in aom_highbd_dc_predictor_32x16_sse2()
966     _mm_store_si128((__m128i *)(dst + 24), row);  in aom_highbd_dc_predictor_32x16_sse2()
967     dst += stride;  in aom_highbd_dc_predictor_32x16_sse2()
968     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_32x16_sse2()
969     _mm_store_si128((__m128i *)(dst + 8), row);  in aom_highbd_dc_predictor_32x16_sse2()
970     _mm_store_si128((__m128i *)(dst + 16), row);  in aom_highbd_dc_predictor_32x16_sse2()
971     _mm_store_si128((__m128i *)(dst + 24), row);  in aom_highbd_dc_predictor_32x16_sse2()
972     dst += stride;  in aom_highbd_dc_predictor_32x16_sse2()
973     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_32x16_sse2()
974     _mm_store_si128((__m128i *)(dst + 8), row);  in aom_highbd_dc_predictor_32x16_sse2()
975     _mm_store_si128((__m128i *)(dst + 16), row);  in aom_highbd_dc_predictor_32x16_sse2()
976     _mm_store_si128((__m128i *)(dst + 24), row);  in aom_highbd_dc_predictor_32x16_sse2()
977     dst += stride;  in aom_highbd_dc_predictor_32x16_sse2()
978     _mm_store_si128((__m128i *)dst, row);  in aom_highbd_dc_predictor_32x16_sse2()
979     _mm_store_si128((__m128i *)(dst + 8), row);  in aom_highbd_dc_predictor_32x16_sse2()
980     _mm_store_si128((__m128i *)(dst + 16), row);  in aom_highbd_dc_predictor_32x16_sse2()
981     _mm_store_si128((__m128i *)(dst + 24), row);  in aom_highbd_dc_predictor_32x16_sse2()
982     dst += stride;  in aom_highbd_dc_predictor_32x16_sse2()