1 |
;/************************************************************************** |
;/**************************************************************************** |
2 |
; * |
; * |
3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * - mmx quantization/dequantization - |
; * - 3dne Quantization/Dequantization - |
5 |
; * |
; * |
6 |
; * Copyright(C) 2001-2003 XviD Team <xvid-devel@xvid.org> |
; * Copyright (C) 2002-2003 Peter Ross <pross@xvid.org> |
7 |
|
; * 2002 Jaan Kalda |
8 |
; * |
; * |
9 |
; * This program is free software ; you can redistribute it and/or modify |
; * This program is free software ; you can redistribute it and/or modify |
10 |
; * 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 |
20 |
; * along with this program ; if not, write to the Free Software |
; * along with this program ; if not, write to the Free Software |
21 |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
22 |
; * |
; * |
23 |
; * $Id: quantize_mpeg_xmm.asm,v 1.1.2.1 2003-10-07 13:02:35 edgomez Exp $ |
; * $Id: quantize_mpeg_xmm.asm,v 1.1.2.5 2003-11-30 16:13:16 edgomez Exp $ |
24 |
; * |
; * |
25 |
; *************************************************************************/ |
; ***************************************************************************/ |
26 |
;/************************************************************************** |
|
|
; * quant4 bugs have been fixed: (a) overflow bug for matrix elements |
|
|
; * equal to 1 or 2 is fixed by substituting pmulhw with pmulhuw (iSSE) |
|
|
; * and using multiplier 0ffffh instead of 10001h (for matrix element = 1; |
|
|
; * in that case, 1 is added before multiplying, that additional 1 comes |
|
|
; * from intra_matrix1; (b) rounding error for large coefficients and matrix |
|
|
; * elements is fixed by two-step approach: first approximation (rounded |
|
|
; * down) is found as usual; the result is multiplied by the matrix element |
|
|
; * and mismatch is used to calculate the correction. |
|
|
; *************************************************************************/ |
|
27 |
; _3dne functions are compatible with iSSE, but are optimized specifically |
; _3dne functions are compatible with iSSE, but are optimized specifically |
28 |
; for K7 pipelines |
; for K7 pipelines |
|
; |
|
|
;--------------------------------------------------------------------------- |
|
|
; 09.12.2002 Athlon optimizations contributed by Jaan Kalda |
|
|
;--------------------------------------------------------------------------- |
|
|
|
|
29 |
|
|
|
; data/text alignment |
|
|
%define ALIGN 8 |
|
30 |
%define SATURATE |
%define SATURATE |
31 |
|
|
32 |
bits 32 |
BITS 32 |
|
|
|
|
%ifdef FORMAT_COFF |
|
|
SECTION .data data |
|
|
%else |
|
|
SECTION .data data align=8 |
|
|
%endif |
|
33 |
|
|
34 |
%macro cglobal 1 |
%macro cglobal 1 |
35 |
%ifdef PREFIX |
%ifdef PREFIX |
48 |
extern %1 |
extern %1 |
49 |
%endif |
%endif |
50 |
%endmacro |
%endmacro |
|
align 8 |
|
|
mmzero dd 0,0 |
|
51 |
|
|
52 |
mmx_one times 4 dw 1 |
;============================================================================= |
53 |
|
; Local data |
54 |
|
;============================================================================= |
55 |
|
|
56 |
;=========================================================================== |
%ifdef FORMAT_COFF |
57 |
; |
SECTION .rodata data |
58 |
|
%else |
59 |
|
SECTION .rodata data align=16 |
60 |
|
%endif |
61 |
|
|
62 |
|
ALIGN 8 |
63 |
|
mmzero: |
64 |
|
dd 0,0 |
65 |
|
mmx_one: |
66 |
|
times 4 dw 1 |
67 |
|
|
68 |
|
;----------------------------------------------------------------------------- |
69 |
; divide by 2Q table |
; divide by 2Q table |
70 |
; |
;----------------------------------------------------------------------------- |
|
;=========================================================================== |
|
71 |
|
|
72 |
align ALIGN |
ALIGN 16 |
73 |
mmx_divs ;i>2 |
mmx_divs: ;i>2 |
74 |
%assign i 1 |
%assign i 1 |
75 |
%rep 31 |
%rep 31 |
76 |
times 4 dw ((1 << 15) / i + 1) |
times 4 dw ((1 << 15) / i + 1) |
77 |
%assign i i+1 |
%assign i i+1 |
78 |
%endrep |
%endrep |
79 |
|
|
80 |
align ALIGN |
ALIGN 16 |
81 |
mmx_div ;i>2 |
mmx_div: ;quant>2 |
82 |
%assign i 1 |
times 4 dw 65535 ; the div by 2 formula will overflow for the case |
83 |
|
; quant=1 but we don't care much because quant=1 |
84 |
|
; is handled by a different piece of code that |
85 |
|
; doesn't use this table. |
86 |
|
%assign quant 2 |
87 |
%rep 31 |
%rep 31 |
88 |
times 4 dw ((1 << 16) / i + 1) |
times 4 dw ((1 << 16) / quant + 1) |
89 |
%assign i i+1 |
%assign quant quant+1 |
90 |
%endrep |
%endrep |
91 |
|
|
|
|
|
|
;=========================================================================== |
|
|
; |
|
|
; intra matrix |
|
|
; |
|
|
;=========================================================================== |
|
|
|
|
92 |
%macro FIXX 1 |
%macro FIXX 1 |
93 |
dw (1 << 16) / (%1) + 1 |
dw (1 << 16) / (%1) + 1 |
94 |
%endmacro |
%endmacro |
95 |
|
|
|
cextern intra_matrix_fixl |
|
|
cextern intra_matrix_fix |
|
|
cextern intra_matrix1 |
|
|
cextern intra_matrix |
|
|
|
|
|
;=========================================================================== |
|
|
; |
|
|
; inter matrix |
|
|
; |
|
|
;=========================================================================== |
|
|
|
|
|
cextern inter_matrix1 |
|
|
cextern inter_matrix |
|
|
cextern inter_matrix_fix |
|
|
cextern inter_matrix_fixl |
|
|
|
|
|
|
|
|
%define VM18P 3 |
|
|
%define VM18Q 4 |
|
96 |
%define nop4 db 08Dh,074h,026h,0 |
%define nop4 db 08Dh,074h,026h,0 |
97 |
%define nop3 add esp,byte 0 |
%define nop3 add esp,byte 0 |
98 |
%define nop2 mov esp,esp |
%define nop2 mov esp,esp |
99 |
%define nop7 db 08dh,02ch,02dh,0,0,0,0 |
%define nop7 db 08dh,02ch,02dh,0,0,0,0 |
100 |
%define nop6 add ebp,dword 0 |
%define nop6 add ebp,dword 0 |
101 |
|
|
102 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
|
; |
|
103 |
; quantd table |
; quantd table |
104 |
; |
;----------------------------------------------------------------------------- |
|
;=========================================================================== |
|
105 |
|
|
106 |
|
%define VM18P 3 |
107 |
|
%define VM18Q 4 |
108 |
|
|
109 |
quantd |
ALIGN 16 |
110 |
|
quantd: |
111 |
%assign i 1 |
%assign i 1 |
112 |
%rep 31 |
%rep 31 |
113 |
times 4 dw (((VM18P*i) + (VM18Q/2)) / VM18Q) |
times 4 dw (((VM18P*i) + (VM18Q/2)) / VM18Q) |
114 |
%assign i i+1 |
%assign i i+1 |
115 |
%endrep |
%endrep |
116 |
|
|
117 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
|
; |
|
118 |
; multiple by 2Q table |
; multiple by 2Q table |
119 |
; |
;----------------------------------------------------------------------------- |
|
;=========================================================================== |
|
120 |
|
|
121 |
|
ALIGN 16 |
122 |
mmx_mul_quant |
mmx_mul_quant: |
123 |
%assign i 1 |
%assign i 1 |
124 |
%rep 31 |
%rep 31 |
125 |
times 4 dw i |
times 4 dw i |
126 |
%assign i i+1 |
%assign i i+1 |
127 |
%endrep |
%endrep |
128 |
|
|
129 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
|
; |
|
130 |
; saturation limits |
; saturation limits |
131 |
; |
;----------------------------------------------------------------------------- |
|
;=========================================================================== |
|
|
|
|
|
align 16 |
|
132 |
|
|
133 |
mmx_32767_minus_2047 times 4 dw (32767-2047) |
ALIGN 16 |
134 |
mmx_32768_minus_2048 times 4 dw (32768-2048) |
mmx_32767_minus_2047: |
135 |
mmx_2047 times 4 dw 2047 |
times 4 dw (32767-2047) |
136 |
mmx_minus_2048 times 4 dw (-2048) |
mmx_32768_minus_2048: |
137 |
zero times 4 dw 0 |
times 4 dw (32768-2048) |
138 |
|
mmx_2047: |
139 |
|
times 4 dw 2047 |
140 |
|
mmx_minus_2048: |
141 |
|
times 4 dw (-2048) |
142 |
|
zero: |
143 |
|
times 4 dw 0 |
144 |
|
|
145 |
int_div |
int_div: |
146 |
dd 0 |
dd 0 |
147 |
%assign i 1 |
%assign i 1 |
148 |
%rep 255 |
%rep 255 |
150 |
%assign i i+1 |
%assign i i+1 |
151 |
%endrep |
%endrep |
152 |
|
|
153 |
section .text |
;============================================================================= |
154 |
|
; Code |
155 |
|
;============================================================================= |
156 |
|
|
157 |
|
SECTION .text |
158 |
|
|
159 |
;=========================================================================== |
cglobal quant_mpeg_intra_xmm |
160 |
|
cglobal quant_mpeg_inter_xmm |
161 |
|
cglobal dequant_mpeg_intra_3dne |
162 |
|
cglobal dequant_mpeg_inter_3dne |
163 |
|
|
164 |
|
;----------------------------------------------------------------------------- |
165 |
; |
; |
166 |
; void quant4_intra_xmm(int16_t * coeff, |
; uint32_t quant_mpeg_intra_xmm(int16_t * coeff, |
167 |
; const int16_t const * data, |
; const int16_t const * data, |
168 |
; const uint32_t quant, |
; const uint32_t quant, |
169 |
; const uint32_t dcscalar); |
; const uint32_t dcscalar, |
170 |
|
; const uint16_t *mpeg_matrices); |
171 |
; |
; |
172 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
173 |
|
|
174 |
align ALIGN |
ALIGN 16 |
|
cglobal quant_mpeg_intra_xmm |
|
175 |
quant_mpeg_intra_xmm: |
quant_mpeg_intra_xmm: |
176 |
mov eax, [esp + 8] ; data |
mov eax, [esp + 8] ; data |
177 |
mov ecx, [esp + 12] ; quant |
mov ecx, [esp + 12] ; quant |
180 |
push edi |
push edi |
181 |
push ebx |
push ebx |
182 |
nop |
nop |
183 |
mov edi,mmzero |
mov edi, [esp + 12 + 20] ; mpeg_quant_matrices |
184 |
mov esi,-14 |
mov esi,-14 |
185 |
pxor mm0,mm0 |
pxor mm0,mm0 |
186 |
pxor mm3,mm3 |
pxor mm3,mm3 |
190 |
jg near .lloop |
jg near .lloop |
191 |
nop6 |
nop6 |
192 |
|
|
193 |
|
ALIGN 16 |
|
align ALIGN |
|
194 |
.loop |
.loop |
195 |
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
196 |
psubw mm0, mm1 ;-mm1 |
psubw mm0, mm1 ;-mm1 |
203 |
psraw mm4, 15 |
psraw mm4, 15 |
204 |
psllw mm0, 4 ;level << 4 ; |
psllw mm0, 4 ;level << 4 ; |
205 |
psllw mm3, 4 |
psllw mm3, 4 |
206 |
paddw mm0, [intra_matrix1 + 8*esi+112] |
paddw mm0, [edi + 128 + 8*esi+112] |
207 |
paddw mm3, [intra_matrix1 + 8*esi+120] |
paddw mm3, [edi + 128 + 8*esi+120] |
208 |
movq mm5, [intra_matrix_fixl + 8*esi+112] |
movq mm5, [edi + 384 + 8*esi+112] |
209 |
movq mm7, [intra_matrix_fixl + 8*esi+120] |
movq mm7, [edi + 384 + 8*esi+120] |
210 |
pmulhuw mm5, mm0 |
pmulhuw mm5, mm0 |
211 |
pmulhuw mm7, mm3 |
pmulhuw mm7, mm3 |
212 |
mov esp, esp |
mov esp, esp |
213 |
movq mm2, [intra_matrix + 8*esi+112] |
movq mm2, [edi + 8*esi+112] |
214 |
movq mm6, [intra_matrix + 8*esi+120] |
movq mm6, [edi + 8*esi+120] |
215 |
pmullw mm2, mm5 |
pmullw mm2, mm5 |
216 |
pmullw mm6, mm7 |
pmullw mm6, mm7 |
217 |
psubw mm0, mm2 |
psubw mm0, mm2 |
222 |
paddw mm5, mm2 |
paddw mm5, mm2 |
223 |
paddw mm7, mm2 |
paddw mm7, mm2 |
224 |
mov esp, esp |
mov esp, esp |
225 |
pmulhuw mm0, [intra_matrix_fix + 8*esi+112] |
pmulhuw mm0, [edi + 256 + 8*esi+112] |
226 |
pmulhuw mm3, [intra_matrix_fix + 8*esi+120] |
pmulhuw mm3, [edi + 256 + 8*esi+120] |
227 |
paddw mm5, mm0 |
paddw mm5, mm0 |
228 |
paddw mm7, mm3 |
paddw mm7, mm3 |
229 |
movq mm0, [edi] |
pxor mm0, mm0 |
230 |
movq mm3, [edi] |
pxor mm3, mm3 |
231 |
pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
232 |
pmulhuw mm7, mm6 ; (level + quantd) / quant (0<quant<32) |
pmulhuw mm7, mm6 ; (level + quantd) / quant (0<quant<32) |
233 |
pxor mm5, mm1 ; mm0 *= sign(mm0) |
pxor mm5, mm1 ; mm0 *= sign(mm0) |
264 |
add esp, byte 12 |
add esp, byte 12 |
265 |
mov [edx], cx ; coeff[0] = ax |
mov [edx], cx ; coeff[0] = ax |
266 |
|
|
267 |
|
xor eax, eax |
268 |
ret |
ret |
269 |
|
|
270 |
align ALIGN |
ALIGN 16 |
271 |
.q1loop |
.q1loop |
272 |
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
273 |
psubw mm0, mm1 ;-mm1 |
psubw mm0, mm1 ;-mm1 |
280 |
psraw mm4, 15 |
psraw mm4, 15 |
281 |
psllw mm0, 4 ; level << 4 |
psllw mm0, 4 ; level << 4 |
282 |
psllw mm3, 4 |
psllw mm3, 4 |
283 |
paddw mm0, [intra_matrix1 + 8*esi+112] ;mm0 is to be divided |
paddw mm0, [edi + 128 + 8*esi+112] ;mm0 is to be divided |
284 |
paddw mm3, [intra_matrix1 + 8*esi+120] ;intra1 contains fix for division by 1 |
paddw mm3, [edi + 128 + 8*esi+120] ;intra1 contains fix for division by 1 |
285 |
movq mm5, [intra_matrix_fixl + 8*esi+112] ;with rounding down |
movq mm5, [edi + 384 + 8*esi+112] ;with rounding down |
286 |
movq mm7, [intra_matrix_fixl + 8*esi+120] |
movq mm7, [edi + 384 + 8*esi+120] |
287 |
pmulhuw mm5, mm0 |
pmulhuw mm5, mm0 |
288 |
pmulhuw mm7, mm3 ;mm7: first approx of division |
pmulhuw mm7, mm3 ;mm7: first approx of division |
289 |
mov esp, esp |
mov esp, esp |
290 |
movq mm2, [intra_matrix + 8*esi+112] |
movq mm2, [edi + 8*esi+112] |
291 |
movq mm6, [intra_matrix + 8*esi+120] ; divs for q<=16 |
movq mm6, [edi + 8*esi+120] ; divs for q<=16 |
292 |
pmullw mm2, mm5 ;test value <= original |
pmullw mm2, mm5 ;test value <= original |
293 |
pmullw mm6, mm7 |
pmullw mm6, mm7 |
294 |
psubw mm0, mm2 ;mismatch |
psubw mm0, mm2 ;mismatch |
298 |
paddw mm5, mm2 ;first approx with quantd |
paddw mm5, mm2 ;first approx with quantd |
299 |
paddw mm7, mm2 |
paddw mm7, mm2 |
300 |
mov esp, esp |
mov esp, esp |
301 |
pmulhuw mm0, [intra_matrix_fix + 8*esi+112] ;correction |
pmulhuw mm0, [edi + 256 + 8*esi+112] ;correction |
302 |
pmulhuw mm3, [intra_matrix_fix + 8*esi+120] |
pmulhuw mm3, [edi + 256 + 8*esi+120] |
303 |
paddw mm5, mm0 ;final result with quantd |
paddw mm5, mm0 ;final result with quantd |
304 |
paddw mm7, mm3 |
paddw mm7, mm3 |
305 |
movq mm0, [edi] |
pxor mm0, mm0 |
306 |
movq mm3, [edi] |
pxor mm3, mm3 |
307 |
mov esp, esp |
mov esp, esp |
308 |
psrlw mm5, 1 ; (level + quantd) /2 (quant = 1) |
psrlw mm5, 1 ; (level + quantd) /2 (quant = 1) |
309 |
psrlw mm7, 1 |
psrlw mm7, 1 |
317 |
jng near .q1loop |
jng near .q1loop |
318 |
jmp near .done |
jmp near .done |
319 |
|
|
320 |
align 8 |
ALIGN 8 |
321 |
.lloop |
.lloop |
322 |
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
323 |
psubw mm0, mm1 ;-mm1 |
psubw mm0, mm1 ;-mm1 |
324 |
movq mm4, [eax + 8*esi+120] ; |
movq mm4, [eax + 8*esi+120] |
325 |
psubw mm3, mm4 ;-mm4 |
psubw mm3, mm4 ;-mm4 |
326 |
pmaxsw mm0, mm1 ;|src| |
pmaxsw mm0, mm1 ;|src| |
327 |
pmaxsw mm3, mm4 |
pmaxsw mm3, mm4 |
330 |
psraw mm4, 15 |
psraw mm4, 15 |
331 |
psllw mm0, 4 ; level << 4 |
psllw mm0, 4 ; level << 4 |
332 |
psllw mm3, 4 ; |
psllw mm3, 4 ; |
333 |
paddw mm0, [intra_matrix1 + 8*esi+112] ;mm0 is to be divided intra1 contains fix for division by 1 |
paddw mm0, [edi + 128 + 8*esi+112] ;mm0 is to be divided intra1 contains fix for division by 1 |
334 |
paddw mm3, [intra_matrix1 + 8*esi+120] |
paddw mm3, [edi + 128 + 8*esi+120] |
335 |
movq mm5, [intra_matrix_fixl + 8*esi+112] |
movq mm5, [edi + 384 + 8*esi+112] |
336 |
movq mm7, [intra_matrix_fixl + 8*esi+120] |
movq mm7, [edi + 384 + 8*esi+120] |
337 |
pmulhuw mm5, mm0 |
pmulhuw mm5, mm0 |
338 |
pmulhuw mm7, mm3 ;mm7: first approx of division |
pmulhuw mm7, mm3 ;mm7: first approx of division |
339 |
mov esp, esp |
mov esp, esp |
340 |
movq mm2, [intra_matrix + 8*esi+112] |
movq mm2, [edi + 8*esi+112] |
341 |
movq mm6, [intra_matrix + 8*esi+120] |
movq mm6, [edi + 8*esi+120] |
342 |
pmullw mm2, mm5 ;test value <= original |
pmullw mm2, mm5 ;test value <= original |
343 |
pmullw mm6, mm7 |
pmullw mm6, mm7 |
344 |
psubw mm0, mm2 ;mismatch |
psubw mm0, mm2 ;mismatch |
349 |
paddw mm5, mm2 ;first approx with quantd |
paddw mm5, mm2 ;first approx with quantd |
350 |
paddw mm7, mm2 |
paddw mm7, mm2 |
351 |
mov esp, esp |
mov esp, esp |
352 |
pmulhuw mm0, [intra_matrix_fix + 8*esi+112] ;correction |
pmulhuw mm0, [edi + 256 + 8*esi+112] ;correction |
353 |
pmulhuw mm3, [intra_matrix_fix + 8*esi+120] |
pmulhuw mm3, [edi + 256 + 8*esi+120] |
354 |
paddw mm5, mm0 ;final result with quantd |
paddw mm5, mm0 ;final result with quantd |
355 |
paddw mm7, mm3 |
paddw mm7, mm3 |
356 |
movq mm0, [edi] |
pxor mm0, mm0 |
357 |
movq mm3, [edi] |
pxor mm3, mm3 |
358 |
mov esp, esp |
mov esp, esp |
359 |
pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
360 |
pmulhuw mm7, mm6 ; (level + quantd) / quant (0<quant<32) |
pmulhuw mm7, mm6 ; (level + quantd) / quant (0<quant<32) |
370 |
jng near .lloop |
jng near .lloop |
371 |
jmp near .done |
jmp near .done |
372 |
|
|
373 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
374 |
; |
; |
375 |
; uint32_t quant4_inter_xmm(int16_t * coeff, |
; uint32_t quant_mpeg_inter_xmm(int16_t * coeff, |
376 |
; const int16_t const * data, |
; const int16_t const * data, |
377 |
; const uint32_t quant); |
; const uint32_t quant, |
378 |
|
; const uint16_t *mpeg_matrices); |
379 |
; |
; |
380 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
381 |
|
|
382 |
align ALIGN |
ALIGN 16 |
|
cglobal quant_mpeg_inter_xmm |
|
383 |
quant_mpeg_inter_xmm: |
quant_mpeg_inter_xmm: |
384 |
mov eax, [esp + 8] ; data |
mov eax, [esp + 8] ; data |
385 |
mov ecx, [esp + 12] ; quant |
mov ecx, [esp + 12] ; quant |
388 |
push edi |
push edi |
389 |
push ebx |
push ebx |
390 |
nop |
nop |
391 |
mov edi,mmzero |
mov edi, [esp + 12 + 16] |
392 |
mov esi,-14 |
mov esi,-14 |
393 |
mov ebx,esp |
mov ebx,esp |
394 |
sub esp,byte 24 |
sub esp,byte 24 |
395 |
lea ebx,[esp+8] |
lea ebx,[esp+8] |
396 |
and ebx,byte -8 ;align 8 |
and ebx, byte -8 ;ALIGN 8 |
397 |
pxor mm0,mm0 |
pxor mm0,mm0 |
398 |
pxor mm3,mm3 |
pxor mm3,mm3 |
399 |
movq [byte ebx],mm0 |
movq [byte ebx],mm0 |
404 |
jg near .lloop |
jg near .lloop |
405 |
nop |
nop |
406 |
|
|
407 |
align ALIGN |
ALIGN 16 |
408 |
.loop |
.loop |
409 |
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
410 |
psubw mm0,mm1 ;-mm1 |
psubw mm0,mm1 ;-mm1 |
417 |
psraw mm4,15 |
psraw mm4,15 |
418 |
psllw mm0, 4 ; level << 4 |
psllw mm0, 4 ; level << 4 |
419 |
psllw mm3, 4 ; |
psllw mm3, 4 ; |
420 |
paddw mm0, [inter_matrix1 + 8*esi+112] |
paddw mm0, [edi + 640 + 8*esi+112] |
421 |
paddw mm3, [inter_matrix1 + 8*esi+120] |
paddw mm3, [edi + 640 + 8*esi+120] |
422 |
movq mm5,[inter_matrix_fixl + 8*esi+112] |
movq mm5, [edi + 896 + 8*esi+112] |
423 |
movq mm7,[inter_matrix_fixl + 8*esi+120] |
movq mm7, [edi + 896 + 8*esi+120] |
424 |
pmulhuw mm5,mm0 |
pmulhuw mm5,mm0 |
425 |
pmulhuw mm7,mm3 |
pmulhuw mm7,mm3 |
426 |
mov esp,esp |
mov esp,esp |
427 |
movq mm2,[inter_matrix + 8*esi+112] |
movq mm2, [edi + 512 + 8*esi+112] |
428 |
movq mm6,[inter_matrix + 8*esi+120] |
movq mm6, [edi + 512 + 8*esi+120] |
429 |
pmullw mm2,mm5 |
pmullw mm2,mm5 |
430 |
pmullw mm6,mm7 |
pmullw mm6,mm7 |
431 |
psubw mm0,mm2 |
psubw mm0,mm2 |
432 |
psubw mm3,mm6 |
psubw mm3,mm6 |
433 |
movq mm2,[byte ebx] |
movq mm2,[byte ebx] |
434 |
movq mm6,[mmx_divs + ecx * 8 - 8] |
movq mm6,[mmx_divs + ecx * 8 - 8] |
435 |
pmulhuw mm0,[inter_matrix_fix + 8*esi+112] |
pmulhuw mm0, [edi + 768 + 8*esi+112] |
436 |
pmulhuw mm3,[inter_matrix_fix + 8*esi+120] |
pmulhuw mm3, [edi + 768 + 8*esi+120] |
437 |
paddw mm2,[ebx+8] ;sum |
paddw mm2,[ebx+8] ;sum |
438 |
paddw mm5,mm0 |
paddw mm5,mm0 |
439 |
paddw mm7,mm3 |
paddw mm7,mm3 |
440 |
movq mm0,[edi] |
pxor mm0, mm0 |
441 |
movq mm3,[edi] |
pxor mm3, mm3 |
442 |
pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
443 |
pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) |
pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) |
444 |
add esi,byte 2 |
add esi,byte 2 |
468 |
|
|
469 |
ret |
ret |
470 |
|
|
471 |
align ALIGN |
ALIGN 16 |
472 |
.q1loop |
.q1loop |
473 |
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
474 |
psubw mm0,mm1 ;-mm1 |
psubw mm0,mm1 ;-mm1 |
475 |
movq mm4, [eax + 8*esi+120] ; |
movq mm4, [eax + 8*esi+120] |
476 |
psubw mm3,mm4 ;-mm4 |
psubw mm3,mm4 ;-mm4 |
477 |
pmaxsw mm0,mm1 ;|src| |
pmaxsw mm0,mm1 ;|src| |
478 |
pmaxsw mm3,mm4 |
pmaxsw mm3,mm4 |
481 |
psraw mm4,15 |
psraw mm4,15 |
482 |
psllw mm0, 4 ; level << 4 |
psllw mm0, 4 ; level << 4 |
483 |
psllw mm3, 4 |
psllw mm3, 4 |
484 |
paddw mm0, [inter_matrix1 + 8*esi+112] ;mm0 is to be divided |
paddw mm0, [edi + 640 + 8*esi+112] ;mm0 is to be divided |
485 |
paddw mm3, [inter_matrix1 + 8*esi+120] ; inter1 contains fix for division by 1 |
paddw mm3, [edi + 640 + 8*esi+120] ; inter1 contains fix for division by 1 |
486 |
movq mm5,[inter_matrix_fixl + 8*esi+112] ;with rounding down |
movq mm5, [edi + 896 + 8*esi+112] ;with rounding down |
487 |
movq mm7,[inter_matrix_fixl + 8*esi+120] |
movq mm7, [edi + 896 + 8*esi+120] |
488 |
pmulhuw mm5,mm0 |
pmulhuw mm5,mm0 |
489 |
pmulhuw mm7,mm3 ;mm7: first approx of division |
pmulhuw mm7,mm3 ;mm7: first approx of division |
490 |
mov esp,esp |
mov esp,esp |
491 |
movq mm2,[inter_matrix + 8*esi+112] |
movq mm2, [edi + 512 + 8*esi+112] |
492 |
movq mm6,[inter_matrix + 8*esi+120] ; divs for q<=16 |
movq mm6, [edi + 512 + 8*esi+120] ; divs for q<=16 |
493 |
pmullw mm2,mm5 ;test value <= original |
pmullw mm2,mm5 ;test value <= original |
494 |
pmullw mm6,mm7 |
pmullw mm6,mm7 |
495 |
psubw mm0,mm2 ;mismatch |
psubw mm0,mm2 ;mismatch |
496 |
psubw mm3,mm6 |
psubw mm3,mm6 |
497 |
movq mm2,[byte ebx] |
movq mm2,[byte ebx] |
498 |
pmulhuw mm0,[inter_matrix_fix + 8*esi+112] ;correction |
pmulhuw mm0, [edi + 768 + 8*esi+112] ;correction |
499 |
pmulhuw mm3,[inter_matrix_fix + 8*esi+120] |
pmulhuw mm3, [edi + 768 + 8*esi+120] |
500 |
paddw mm2,[ebx+8] ;sum |
paddw mm2,[ebx+8] ;sum |
501 |
paddw mm5,mm0 ;final result |
paddw mm5,mm0 ;final result |
502 |
paddw mm7,mm3 |
paddw mm7,mm3 |
503 |
movq mm0,[edi] |
pxor mm0, mm0 |
504 |
movq mm3,[edi] |
pxor mm3, mm3 |
505 |
psrlw mm5, 1 ; (level ) /2 (quant = 1) |
psrlw mm5, 1 ; (level ) /2 (quant = 1) |
506 |
psrlw mm7, 1 |
psrlw mm7, 1 |
507 |
add esi,byte 2 |
add esi,byte 2 |
517 |
jng near .q1loop |
jng near .q1loop |
518 |
jmp near .done |
jmp near .done |
519 |
|
|
520 |
align 8 |
ALIGN 8 |
521 |
.lloop |
.lloop |
522 |
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
523 |
psubw mm0,mm1 ;-mm1 |
psubw mm0,mm1 ;-mm1 |
524 |
movq mm4, [eax + 8*esi+120] ; |
movq mm4, [eax + 8*esi+120] |
525 |
psubw mm3,mm4 ;-mm4 |
psubw mm3,mm4 ;-mm4 |
526 |
pmaxsw mm0,mm1 ;|src| |
pmaxsw mm0,mm1 ;|src| |
527 |
pmaxsw mm3,mm4 |
pmaxsw mm3,mm4 |
530 |
psraw mm4,15 |
psraw mm4,15 |
531 |
psllw mm0, 4 ; level << 4 |
psllw mm0, 4 ; level << 4 |
532 |
psllw mm3, 4 ; |
psllw mm3, 4 ; |
533 |
paddw mm0, [inter_matrix1 + 8*esi+112] ;mm0 is to be divided inter1 contains fix for division by 1 |
paddw mm0, [edi + 640 + 8*esi+112] ;mm0 is to be divided inter1 contains fix for division by 1 |
534 |
paddw mm3, [inter_matrix1 + 8*esi+120] |
paddw mm3, [edi + 640 + 8*esi+120] |
535 |
movq mm5,[inter_matrix_fixl + 8*esi+112] |
movq mm5,[edi + 896 + 8*esi+112] |
536 |
movq mm7,[inter_matrix_fixl + 8*esi+120] |
movq mm7,[edi + 896 + 8*esi+120] |
537 |
pmulhuw mm5,mm0 |
pmulhuw mm5,mm0 |
538 |
pmulhuw mm7,mm3 ;mm7: first approx of division |
pmulhuw mm7,mm3 ;mm7: first approx of division |
539 |
mov esp,esp |
mov esp,esp |
540 |
movq mm2,[inter_matrix + 8*esi+112] |
movq mm2,[edi + 512 + 8*esi+112] |
541 |
movq mm6,[inter_matrix + 8*esi+120] |
movq mm6,[edi + 512 + 8*esi+120] |
542 |
pmullw mm2,mm5 ;test value <= original |
pmullw mm2,mm5 ;test value <= original |
543 |
pmullw mm6,mm7 |
pmullw mm6,mm7 |
544 |
psubw mm0,mm2 ;mismatch |
psubw mm0,mm2 ;mismatch |
545 |
psubw mm3,mm6 |
psubw mm3,mm6 |
546 |
movq mm2,[byte ebx] |
movq mm2,[byte ebx] |
547 |
movq mm6,[mmx_div + ecx * 8 - 8] ; divs for q<=16 |
movq mm6,[mmx_div + ecx * 8 - 8] ; divs for q<=16 |
548 |
pmulhuw mm0,[inter_matrix_fix + 8*esi+112] ;correction |
pmulhuw mm0,[edi + 768 + 8*esi+112] ;correction |
549 |
pmulhuw mm3,[inter_matrix_fix + 8*esi+120] |
pmulhuw mm3,[edi + 768 + 8*esi+120] |
550 |
paddw mm2,[ebx+8] ;sum |
paddw mm2,[ebx+8] ;sum |
551 |
paddw mm5,mm0 ;final result |
paddw mm5,mm0 ;final result |
552 |
paddw mm7,mm3 |
paddw mm7,mm3 |
553 |
movq mm0,[edi] |
pxor mm0,mm0 |
554 |
movq mm3,[edi] |
pxor mm3,mm3 |
555 |
pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
556 |
pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) |
pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) |
557 |
add esi,byte 2 |
add esi,byte 2 |
570 |
jmp near .done |
jmp near .done |
571 |
|
|
572 |
|
|
573 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
574 |
; |
; |
575 |
; void dequant4_intra_mmx(int16_t *data, |
; uint32_t dequant_mpeg_intra_3dne(int16_t *data, |
576 |
; const int16_t const *coeff, |
; const int16_t const *coeff, |
577 |
; const uint32_t quant, |
; const uint32_t quant, |
578 |
; const uint32_t dcscalar); |
; const uint32_t dcscalar, |
579 |
|
; const uint16_t *mpeg_matrices); |
580 |
; |
; |
581 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
582 |
|
|
583 |
; Note: in order to saturate 'easily', we pre-shift the quantifier |
; Note: in order to saturate 'easily', we pre-shift the quantifier |
584 |
; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to |
; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to |
589 |
; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a |
; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a |
590 |
; and quant in [1..31]. |
; and quant in [1..31]. |
591 |
; |
; |
592 |
;******************************************************************** |
|
593 |
%macro DEQUANT4INTRAMMX 1 |
%macro DEQUANT4INTRAMMX 1 |
594 |
movq mm1, [byte ecx+ 16 * %1] ; mm0 = c = coeff[i] |
movq mm1, [byte ecx+ 16 * %1] ; mm0 = c = coeff[i] |
595 |
movq mm4, [ecx+ 16 * %1 +8]; mm3 = c' = coeff[i+1] |
movq mm4, [ecx+ 16 * %1 +8]; mm3 = c' = coeff[i+1] |
603 |
movq mm2,[eax+8] ;preshifted quant |
movq mm2,[eax+8] ;preshifted quant |
604 |
movq mm7,[eax+8] |
movq mm7,[eax+8] |
605 |
%endif |
%endif |
606 |
pmullw mm2, [intra_matrix + 16 * %1 ] ; matrix[i]*quant |
pmullw mm2, [edi + 16 * %1 ] ; matrix[i]*quant |
607 |
pmullw mm7, [intra_matrix + 16 * %1 +8] ; matrix[i+1]*quant |
pmullw mm7, [edi + 16 * %1 +8] ; matrix[i+1]*quant |
608 |
movq mm5,mm0 |
movq mm5,mm0 |
609 |
movq mm6,mm3 |
movq mm6,mm3 |
610 |
pmulhw mm0, mm2 ; high of coeff*(matrix*quant) |
pmulhw mm0, mm2 ; high of coeff*(matrix*quant) |
629 |
movq [edx + 16 * %1 +8], mm7 ; data[i+1] |
movq [edx + 16 * %1 +8], mm7 ; data[i+1] |
630 |
%endmacro |
%endmacro |
631 |
|
|
632 |
align 16 |
ALIGN 16 |
|
cglobal dequant_mpeg_intra_3dne |
|
633 |
dequant_mpeg_intra_3dne: |
dequant_mpeg_intra_3dne: |
634 |
mov eax, [esp+12] ; quant |
mov eax, [esp+12] ; quant |
635 |
mov ecx, [esp+8] ; coeff |
mov ecx, [esp+8] ; coeff |
643 |
push esi |
push esi |
644 |
lea eax,[esp-28] |
lea eax,[esp-28] |
645 |
sub esp,byte 32 |
sub esp,byte 32 |
646 |
and eax,byte -8 ;points to qword aligned space on stack |
and eax, byte -8 ;points to qword ALIGNed space on stack |
647 |
movq [eax],mm0 |
movq [eax],mm0 |
648 |
movq [eax+8],mm7 |
movq [eax+8],mm7 |
649 |
imul ebx,[esp+16+8+32] ; dcscalar |
imul ebx,[esp+16+8+32] ; dcscalar |
650 |
movq mm2,mm7 |
movq mm2,mm7 |
651 |
|
push edi |
652 |
|
mov edi, [esp + 32 + 12 + 20] ; mpeg_quant_matrices |
653 |
align 4 |
ALIGN 4 |
654 |
|
|
655 |
DEQUANT4INTRAMMX 0 |
DEQUANT4INTRAMMX 0 |
656 |
|
|
672 |
|
|
673 |
DEQUANT4INTRAMMX 3 |
DEQUANT4INTRAMMX 3 |
674 |
|
|
675 |
mov esi, [esp+32] |
mov esi, [esp+36] |
676 |
mov [byte edx], bx |
mov [byte edx], bx |
677 |
mov ebx, [esp+32+4] |
mov ebx, [esp+36+4] |
678 |
|
|
679 |
DEQUANT4INTRAMMX 4 |
DEQUANT4INTRAMMX 4 |
680 |
DEQUANT4INTRAMMX 5 |
DEQUANT4INTRAMMX 5 |
681 |
DEQUANT4INTRAMMX 6 |
DEQUANT4INTRAMMX 6 |
682 |
DEQUANT4INTRAMMX 7 |
DEQUANT4INTRAMMX 7 |
683 |
|
|
684 |
|
pop edi |
685 |
|
|
686 |
add esp, byte 32+8 |
add esp, byte 32+8 |
687 |
|
|
688 |
|
xor eax, eax |
689 |
ret |
ret |
690 |
|
|
691 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
692 |
; |
; |
693 |
; void dequant4_inter_3dne(int16_t * data, |
; uint32_t dequant_mpeg_inter_3dne(int16_t * data, |
694 |
; const int16_t * const coeff, |
; const int16_t * const coeff, |
695 |
; const uint32_t quant); |
; const uint32_t quant, |
696 |
|
; const uint16_t *mpeg_matrices); |
697 |
; |
; |
698 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
699 |
|
|
700 |
; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier |
; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier |
701 |
; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. |
; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. |
702 |
; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. |
; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. |
703 |
; It's mixed with the extraction of the absolute value. |
; It's mixed with the extraction of the absolute value. |
704 |
|
|
705 |
align 16 |
ALIGN 16 |
|
cglobal dequant_mpeg_inter_3dne |
|
706 |
dequant_mpeg_inter_3dne: |
dequant_mpeg_inter_3dne: |
707 |
mov edx, [esp+ 4] ; data |
mov edx, [esp+ 4] ; data |
708 |
mov ecx, [esp+ 8] ; coeff |
mov ecx, [esp+ 8] ; coeff |
712 |
paddw mm7, mm7 ; << 1 |
paddw mm7, mm7 ; << 1 |
713 |
pxor mm6, mm6 ; mismatch sum |
pxor mm6, mm6 ; mismatch sum |
714 |
push esi |
push esi |
715 |
|
push edi |
716 |
mov esi,mmzero |
mov esi,mmzero |
717 |
pxor mm1,mm1 |
pxor mm1,mm1 |
718 |
pxor mm3,mm3 |
pxor mm3,mm3 |
719 |
|
mov edi, [esp + 8 + 16] ; mpeg_quant_matrices |
720 |
nop |
nop |
721 |
nop4 |
nop4 |
722 |
|
|
723 |
align 16 |
ALIGN 16 |
724 |
.loop |
.loop |
725 |
movq mm0, [ecx+8*eax + 7*16 ] ; mm0 = coeff[i] |
movq mm0, [ecx+8*eax + 7*16 ] ; mm0 = coeff[i] |
726 |
pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) |
pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) |
747 |
|
|
748 |
movq mm4, mm7 ; (matrix*quant) |
movq mm4, mm7 ; (matrix*quant) |
749 |
nop |
nop |
750 |
pmullw mm4, [inter_matrix + 8*eax + 7*16] |
pmullw mm4, [edi + 512 + 8*eax + 7*16] |
751 |
movq mm5, mm4 |
movq mm5, mm4 |
752 |
pmulhw mm5, mm0 ; high of c*(matrix*quant) |
pmulhw mm5, mm0 ; high of c*(matrix*quant) |
753 |
pmullw mm0, mm4 ; low of c*(matrix*quant) |
pmullw mm0, mm4 ; low of c*(matrix*quant) |
754 |
|
|
755 |
movq mm4, mm7 ; (matrix*quant) |
movq mm4, mm7 ; (matrix*quant) |
756 |
pmullw mm4, [inter_matrix + 8*eax + 7*16 + 8] |
pmullw mm4, [edi + 512 + 8*eax + 7*16 + 8] |
757 |
add eax,byte 2 |
add eax,byte 2 |
758 |
|
|
759 |
pcmpgtw mm5, [esi] |
pcmpgtw mm5, [esi] |
792 |
pxor mm1, mm2 |
pxor mm1, mm2 |
793 |
pxor mm6, mm1 |
pxor mm6, mm1 |
794 |
movd eax, mm6 |
movd eax, mm6 |
795 |
|
pop edi |
796 |
and eax,byte 1 |
and eax,byte 1 |
797 |
xor eax,byte 1 |
xor eax,byte 1 |
798 |
mov esi,[esp] |
mov esi,[esp] |
799 |
add esp,byte 4 |
add esp,byte 4 |
800 |
xor word [edx + 2*63], ax |
xor word [edx + 2*63], ax |
801 |
|
|
802 |
|
xor eax, eax |
803 |
ret |
ret |