1 #include "../dsputil.h"
2 #include "../mpegvideo.h"
3 #include "../avcodec.h"
5 static void dct_unquantize_h263_intra_iwmmxt(MpegEncContext
*s
,
6 DCTELEM
*block
, int n
, int qscale
)
10 DCTELEM
*block_orig
= block
;
12 assert(s
->block_last_index
[n
]>=0);
18 level
= block
[0] * s
->y_dc_scale
;
20 level
= block
[0] * s
->c_dc_scale
;
21 qadd
= (qscale
- 1) | 1;
29 nCoeffs
= s
->inter_scantable
.raster_end
[ s
->block_last_index
[n
] ];
31 __asm__
__volatile__ (
32 /* "movd %1, %%mm6 \n\t" //qmul */
33 /* "packssdw %%mm6, %%mm6 \n\t" */
34 /* "packssdw %%mm6, %%mm6 \n\t" */
35 "tbcsth wr6, %[qmul] \n\t"
36 /* "movd %2, %%mm5 \n\t" //qadd */
37 /* "packssdw %%mm5, %%mm5 \n\t" */
38 /* "packssdw %%mm5, %%mm5 \n\t" */
39 "tbcsth wr5, %[qadd] \n\t"
40 "wzero wr7 \n\t" /* "pxor %%mm7, %%mm7 \n\t" */
41 "wzero wr4 \n\t" /* "pxor %%mm4, %%mm4 \n\t" */
42 "wsubh wr7, wr5, wr7 \n\t" /* "psubw %%mm5, %%mm7 \n\t" */
44 "wldrd wr2, [%[block]] \n\t" /* "movq (%0, %3), %%mm0 \n\t" */
45 "wldrd wr3, [%[block], #8] \n\t" /* "movq 8(%0, %3), %%mm1 \n\t" */
46 "wmulsl wr0, wr6, wr2 \n\t" /* "pmullw %%mm6, %%mm0 \n\t" */
47 "wmulsl wr1, wr6, wr3 \n\t" /* "pmullw %%mm6, %%mm1 \n\t" */
48 /* "movq (%0, %3), %%mm2 \n\t" */
49 /* "movq 8(%0, %3), %%mm3 \n\t" */
50 "wcmpgtsh wr2, wr4, wr2 \n\t" /* "pcmpgtw %%mm4, %%mm2 \n\t" // block[i] < 0 ? -1 : 0 */
51 "wcmpgtsh wr3, wr4, wr2 \n\t" /* "pcmpgtw %%mm4, %%mm3 \n\t" // block[i] < 0 ? -1 : 0 */
52 "wxor wr0, wr2, wr0 \n\t" /* "pxor %%mm2, %%mm0 \n\t" */
53 "wxor wr1, wr3, wr1 \n\t" /* "pxor %%mm3, %%mm1 \n\t" */
54 "waddh wr0, wr7, wr0 \n\t" /* "paddw %%mm7, %%mm0 \n\t" */
55 "waddh wr1, wr7, wr1 \n\t" /* "paddw %%mm7, %%mm1 \n\t" */
56 "wxor wr2, wr0, wr2 \n\t" /* "pxor %%mm0, %%mm2 \n\t" */
57 "wxor wr3, wr1, wr3 \n\t" /* "pxor %%mm1, %%mm3 \n\t" */
58 "wcmpeqh wr0, wr7, wr0 \n\t" /* "pcmpeqw %%mm7, %%mm0 \n\t" // block[i] == 0 ? -1 : 0 */
59 "wcmpeqh wr1, wr7, wr1 \n\t" /* "pcmpeqw %%mm7, %%mm1 \n\t" // block[i] == 0 ? -1 : 0 */
60 "wandn wr0, wr2, wr0 \n\t" /* "pandn %%mm2, %%mm0 \n\t" */
61 "wandn wr1, wr3, wr1 \n\t" /* "pandn %%mm3, %%mm1 \n\t" */
62 "wstrd wr0, [%[block]] \n\t" /* "movq %%mm0, (%0, %3) \n\t" */
63 "wstrd wr1, [%[block], #8] \n\t" /* "movq %%mm1, 8(%0, %3) \n\t" */
64 "add %[block], %[block], #16 \n\t" /* "addl $16, %3 \n\t" */
65 "subs %[i], %[i], #1 \n\t"
66 "bne 1b \n\t" /* "jng 1b \n\t" */
68 :[i
]"r"((nCoeffs
+ 8) / 8), [qmul
]"r"(qmul
), [qadd
]"r"(qadd
)
71 block_orig
[0] = level
;
75 static void dct_unquantize_h263_inter_iwmmxt(MpegEncContext
*s
,
76 DCTELEM
*block
, int n
, int qscale
)
80 assert(s
->block_last_index
[n
]>=0);
85 nCoeffs
= s
->inter_scantable
.raster_end
[ s
->block_last_index
[n
] ];
87 ippiQuantInvInter_Compact_H263_16s_I(block
, nCoeffs
+1, qscale
);
91 void MPV_common_init_iwmmxt(MpegEncContext
*s
)
93 if (!(mm_flags
& MM_IWMMXT
)) return;
95 s
->dct_unquantize_h263_intra
= dct_unquantize_h263_intra_iwmmxt
;
97 s
->dct_unquantize_h263_inter
= dct_unquantize_h263_inter_iwmmxt
;