1 |
;/************************************************************************** |
;/***************************************************************************** |
2 |
; * |
; * |
3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * 3dnow 8x8 block-based halfpel interpolation |
; * - 3dnow 8x8 block-based halfpel interpolation - |
5 |
|
; * |
6 |
|
; * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
7 |
|
; * 2002 Michael Militzer <isibaar@xvid.org> |
8 |
|
; * 2002 Pascal Massimino <skal@planet-d.net> |
9 |
; * |
; * |
10 |
; * This program is free software; you can redistribute it and/or modify |
; * This program is free software; you can redistribute it and/or modify |
11 |
; * it under the terms of the GNU General Public License as published by |
; * it under the terms of the GNU General Public License as published by |
19 |
; * |
; * |
20 |
; * You should have received a copy of the GNU General Public License |
; * You should have received a copy of the GNU General Public License |
21 |
; * along with this program; if not, write to the Free Software |
; * along with this program; if not, write to the Free Software |
22 |
; * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
|
; * |
|
|
; *************************************************************************/ |
|
|
|
|
|
;/************************************************************************** |
|
|
; * |
|
|
; * History: |
|
23 |
; * |
; * |
24 |
; * 06.07.2002 applied Skal's XMM changes to the 3dnow code (Isibaar) |
; ****************************************************************************/ |
|
; * 04.02.2002 additional 3dnow optimizations (Isibaar) |
|
|
; * 22.12.2001 inital version; (c)2001 peter ross <pross@cs.rmit.edu.au> |
|
|
; * |
|
|
; *************************************************************************/ |
|
|
|
|
25 |
|
|
26 |
bits 32 |
BITS 32 |
27 |
|
|
28 |
%macro cglobal 1 |
%macro cglobal 1 |
29 |
%ifdef PREFIX |
%ifdef PREFIX |
30 |
|
%ifdef MARK_FUNCS |
31 |
|
global _%1:function %1.endfunc-%1 |
32 |
|
%define %1 _%1:function %1.endfunc-%1 |
33 |
|
%else |
34 |
global _%1 |
global _%1 |
35 |
%define %1 _%1 |
%define %1 _%1 |
36 |
|
%endif |
37 |
|
%else |
38 |
|
%ifdef MARK_FUNCS |
39 |
|
global %1:function %1.endfunc-%1 |
40 |
%else |
%else |
41 |
global %1 |
global %1 |
42 |
%endif |
%endif |
43 |
|
%endif |
44 |
%endmacro |
%endmacro |
45 |
|
|
46 |
section .data |
;============================================================================= |
47 |
|
; Read Only data |
48 |
|
;============================================================================= |
49 |
|
|
50 |
|
%ifdef FORMAT_COFF |
51 |
|
SECTION .rodata |
52 |
|
%else |
53 |
|
SECTION .rodata align=16 |
54 |
|
%endif |
55 |
|
|
56 |
align 16 |
ALIGN 16 |
57 |
|
mmx_one: |
|
mmx_one |
|
58 |
times 8 db 1 |
times 8 db 1 |
59 |
|
|
60 |
section .text |
;============================================================================= |
61 |
|
; Code |
62 |
|
;============================================================================= |
63 |
|
|
64 |
|
SECTION .text |
65 |
|
|
66 |
cglobal interpolate8x8_halfpel_h_3dn |
cglobal interpolate8x8_halfpel_h_3dn |
67 |
cglobal interpolate8x8_halfpel_v_3dn |
cglobal interpolate8x8_halfpel_v_3dn |
68 |
cglobal interpolate8x8_halfpel_hv_3dn |
cglobal interpolate8x8_halfpel_hv_3dn |
69 |
|
|
70 |
|
cglobal interpolate8x4_halfpel_h_3dn |
71 |
|
cglobal interpolate8x4_halfpel_v_3dn |
72 |
|
cglobal interpolate8x4_halfpel_hv_3dn |
73 |
|
|
74 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
75 |
; |
; |
76 |
; void interpolate8x8_halfpel_h_3dn(uint8_t * const dst, |
; void interpolate8x8_halfpel_h_3dn(uint8_t * const dst, |
77 |
; const uint8_t * const src, |
; const uint8_t * const src, |
78 |
; const uint32_t stride, |
; const uint32_t stride, |
79 |
; const uint32_t rounding); |
; const uint32_t rounding); |
80 |
; |
; |
81 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
82 |
|
|
83 |
%macro COPY_H_3DN_RND0 0 |
%macro COPY_H_3DN_RND0 0 |
84 |
movq mm0, [eax] |
movq mm0, [eax] |
110 |
movq [ecx+edx], mm1 |
movq [ecx+edx], mm1 |
111 |
%endmacro |
%endmacro |
112 |
|
|
113 |
align 16 |
ALIGN 16 |
114 |
interpolate8x8_halfpel_h_3dn: |
interpolate8x8_halfpel_h_3dn: |
115 |
|
|
116 |
mov eax, [esp+16] ; rounding |
mov eax, [esp+16] ; rounding |
141 |
lea ecx,[ecx+2*edx] |
lea ecx,[ecx+2*edx] |
142 |
COPY_H_3DN_RND1 |
COPY_H_3DN_RND1 |
143 |
ret |
ret |
144 |
|
.endfunc |
145 |
|
|
146 |
|
|
147 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
148 |
; |
; |
149 |
; void interpolate8x8_halfpel_v_3dn(uint8_t * const dst, |
; void interpolate8x8_halfpel_v_3dn(uint8_t * const dst, |
150 |
; const uint8_t * const src, |
; const uint8_t * const src, |
151 |
; const uint32_t stride, |
; const uint32_t stride, |
152 |
; const uint32_t rounding); |
; const uint32_t rounding); |
153 |
; |
; |
154 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
155 |
|
|
156 |
%macro COPY_V_3DN_RND0 0 |
%macro COPY_V_3DN_RND0 0 |
157 |
movq mm0, [eax] |
movq mm0, [eax] |
182 |
movq [ecx+edx], mm1 |
movq [ecx+edx], mm1 |
183 |
%endmacro |
%endmacro |
184 |
|
|
185 |
align 16 |
ALIGN 16 |
186 |
interpolate8x8_halfpel_v_3dn: |
interpolate8x8_halfpel_v_3dn: |
187 |
|
|
188 |
mov eax, [esp+16] ; rounding |
mov eax, [esp+16] ; rounding |
218 |
lea ecx,[ecx+2*edx] |
lea ecx,[ecx+2*edx] |
219 |
COPY_V_3DN_RND1 |
COPY_V_3DN_RND1 |
220 |
ret |
ret |
221 |
|
.endfunc |
222 |
|
|
223 |
|
|
224 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
225 |
; |
; |
226 |
; void interpolate8x8_halfpel_hv_3dn(uint8_t * const dst, |
; void interpolate8x8_halfpel_hv_3dn(uint8_t * const dst, |
227 |
; const uint8_t * const src, |
; const uint8_t * const src, |
229 |
; const uint32_t rounding); |
; const uint32_t rounding); |
230 |
; |
; |
231 |
; |
; |
232 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
233 |
|
|
234 |
; The trick is to correct the result of 'pavgusb' with some combination of the |
; The trick is to correct the result of 'pavgusb' with some combination of the |
235 |
; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgusb' (s and t). |
; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgusb' (s and t). |
321 |
movq [ecx], mm0 |
movq [ecx], mm0 |
322 |
%endmacro |
%endmacro |
323 |
|
|
324 |
align 16 |
ALIGN 16 |
325 |
interpolate8x8_halfpel_hv_3dn |
interpolate8x8_halfpel_hv_3dn |
326 |
mov eax, [esp+16] ; rounding |
mov eax, [esp+16] ; rounding |
327 |
mov ecx, [esp+ 4] ; Dst |
mov ecx, [esp+ 4] ; Dst |
358 |
add ecx, edx |
add ecx, edx |
359 |
COPY_HV_3DN_RND1 |
COPY_HV_3DN_RND1 |
360 |
ret |
ret |
361 |
|
.endfunc |
362 |
|
|
363 |
|
;----------------------------------------------------------------------------- |
364 |
|
; |
365 |
|
; void interpolate8x4_halfpel_h_3dn(uint8_t * const dst, |
366 |
|
; const uint8_t * const src, |
367 |
|
; const uint32_t stride, |
368 |
|
; const uint32_t rounding); |
369 |
|
; |
370 |
|
;----------------------------------------------------------------------------- |
371 |
|
|
372 |
|
ALIGN 16 |
373 |
|
interpolate8x4_halfpel_h_3dn: |
374 |
|
|
375 |
|
mov eax, [esp+16] ; rounding |
376 |
|
mov ecx, [esp+ 4] ; Dst |
377 |
|
test eax, eax |
378 |
|
mov eax, [esp+ 8] ; Src |
379 |
|
mov edx, [esp+12] ; stride |
380 |
|
|
381 |
|
jnz near .rounding1 |
382 |
|
|
383 |
|
COPY_H_3DN_RND0 |
384 |
|
lea ecx, [ecx+2*edx] |
385 |
|
COPY_H_3DN_RND0 |
386 |
|
ret |
387 |
|
|
388 |
|
.rounding1 |
389 |
|
; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
390 |
|
movq mm7, [mmx_one] |
391 |
|
COPY_H_3DN_RND1 |
392 |
|
lea ecx, [ecx+2*edx] |
393 |
|
COPY_H_3DN_RND1 |
394 |
|
ret |
395 |
|
.endfunc |
396 |
|
|
397 |
|
|
398 |
|
;----------------------------------------------------------------------------- |
399 |
|
; |
400 |
|
; void interpolate8x4_halfpel_v_3dn(uint8_t * const dst, |
401 |
|
; const uint8_t * const src, |
402 |
|
; const uint32_t stride, |
403 |
|
; const uint32_t rounding); |
404 |
|
; |
405 |
|
;----------------------------------------------------------------------------- |
406 |
|
|
407 |
|
ALIGN 16 |
408 |
|
interpolate8x4_halfpel_v_3dn: |
409 |
|
|
410 |
|
mov eax, [esp+16] ; rounding |
411 |
|
mov ecx, [esp+ 4] ; Dst |
412 |
|
test eax,eax |
413 |
|
mov eax, [esp+ 8] ; Src |
414 |
|
mov edx, [esp+12] ; stride |
415 |
|
|
416 |
|
; we process 2 line at a time |
417 |
|
|
418 |
|
jnz near .rounding1 |
419 |
|
|
420 |
|
COPY_V_3DN_RND0 |
421 |
|
lea ecx, [ecx+2*edx] |
422 |
|
COPY_V_3DN_RND0 |
423 |
|
ret |
424 |
|
|
425 |
|
.rounding1 |
426 |
|
; we use: (i+j)/2 = ( i+j+1 )/2 - (i^j)&1 |
427 |
|
movq mm7, [mmx_one] |
428 |
|
movq mm2, [eax] ; loop invariant |
429 |
|
add eax, edx |
430 |
|
|
431 |
|
COPY_V_3DN_RND1 |
432 |
|
lea ecx, [ecx+2*edx] |
433 |
|
COPY_V_3DN_RND1 |
434 |
|
ret |
435 |
|
.endfunc |
436 |
|
|
437 |
|
|
438 |
|
;----------------------------------------------------------------------------- |
439 |
|
; |
440 |
|
; void interpolate8x4_halfpel_hv_3dn(uint8_t * const dst, |
441 |
|
; const uint8_t * const src, |
442 |
|
; const uint32_t stride, |
443 |
|
; const uint32_t rounding); |
444 |
|
; |
445 |
|
; |
446 |
|
;----------------------------------------------------------------------------- |
447 |
|
|
448 |
|
; The trick is to correct the result of 'pavgusb' with some combination of the |
449 |
|
; lsb's of the 4 input values i,j,k,l, and their intermediate 'pavgusb' (s and t). |
450 |
|
; The boolean relations are: |
451 |
|
; (i+j+k+l+3)/4 = (s+t+1)/2 - (ij&kl)&st |
452 |
|
; (i+j+k+l+2)/4 = (s+t+1)/2 - (ij|kl)&st |
453 |
|
; (i+j+k+l+1)/4 = (s+t+1)/2 - (ij&kl)|st |
454 |
|
; (i+j+k+l+0)/4 = (s+t+1)/2 - (ij|kl)|st |
455 |
|
; with s=(i+j+1)/2, t=(k+l+1)/2, ij = i^j, kl = k^l, st = s^t. |
456 |
|
|
457 |
|
ALIGN 16 |
458 |
|
interpolate8x4_halfpel_hv_3dn |
459 |
|
mov eax, [esp+16] ; rounding |
460 |
|
mov ecx, [esp+ 4] ; Dst |
461 |
|
test eax, eax |
462 |
|
mov eax, [esp+ 8] ; Src |
463 |
|
mov edx, [esp+12] ; stride |
464 |
|
|
465 |
|
movq mm7, [mmx_one] |
466 |
|
|
467 |
|
; loop invariants: mm2=(i+j+1)/2 and mm3= i^j |
468 |
|
movq mm2, [eax] |
469 |
|
movq mm3, [eax+1] |
470 |
|
movq mm6, mm2 |
471 |
|
pavgusb mm2, mm3 |
472 |
|
pxor mm3, mm6 ; mm2/mm3 ready |
473 |
|
|
474 |
|
jnz near .rounding1 |
475 |
|
|
476 |
|
COPY_HV_3DN_RND0 |
477 |
|
add ecx, edx |
478 |
|
COPY_HV_3DN_RND0 |
479 |
|
ret |
480 |
|
|
481 |
|
.rounding1 |
482 |
|
COPY_HV_3DN_RND1 |
483 |
|
add ecx, edx |
484 |
|
COPY_HV_3DN_RND1 |
485 |
|
ret |
486 |
|
.endfunc |
487 |
|
|
488 |
|
|
489 |
|
%ifidn __OUTPUT_FORMAT__,elf |
490 |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
491 |
|
%endif |
492 |
|
|