5 |
; * |
; * |
6 |
; * Copyright(C) 2001-2003 Peter Ross <pross@xvid.org> |
; * Copyright(C) 2001-2003 Peter Ross <pross@xvid.org> |
7 |
; * 2002-2003 Pascal Massimino <skal@planet-d.net> |
; * 2002-2003 Pascal Massimino <skal@planet-d.net> |
8 |
|
; * 2004 Jean-Marc Bastide <jmtest@voila.fr> |
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 |
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., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
23 |
; * |
; * |
24 |
; * $Id: quantize_h263_mmx.asm,v 1.2 2004-03-22 22:36:24 edgomez Exp $ |
; * $Id: quantize_h263_mmx.asm,v 1.11 2008-11-26 23:35:50 Isibaar Exp $ |
25 |
; * |
; * |
26 |
; ****************************************************************************/ |
; ****************************************************************************/ |
27 |
|
|
28 |
; enable dequant saturate [-2048,2047], test purposes only. |
; enable dequant saturate [-2048,2047], test purposes only. |
29 |
%define SATURATE |
%define SATURATE |
30 |
|
|
31 |
BITS 32 |
%include "nasm.inc" |
|
|
|
|
%macro cglobal 1 |
|
|
%ifdef PREFIX |
|
|
global _%1 |
|
|
%define %1 _%1 |
|
|
%else |
|
|
global %1 |
|
|
%endif |
|
|
%endmacro |
|
32 |
|
|
33 |
;============================================================================= |
;============================================================================= |
34 |
; Read only Local data |
; Read only Local data |
35 |
;============================================================================= |
;============================================================================= |
36 |
|
|
37 |
%ifdef FORMAT_COFF |
DATA |
|
SECTION .rodata data |
|
|
%else |
|
|
SECTION .rodata data align=16 |
|
|
%endif |
|
38 |
|
|
39 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
40 |
plus_one: |
plus_one: |
41 |
times 8 dw 1 |
times 8 dw 1 |
42 |
|
|
43 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
44 |
; |
; |
45 |
|
; quant table |
46 |
|
; |
47 |
|
;----------------------------------------------------------------------------- |
48 |
|
|
49 |
|
ALIGN SECTION_ALIGN |
50 |
|
mmx_quant: |
51 |
|
%assign quant 0 |
52 |
|
%rep 32 |
53 |
|
times 4 dw quant |
54 |
|
%assign quant quant+1 |
55 |
|
%endrep |
56 |
|
|
57 |
|
;----------------------------------------------------------------------------- |
58 |
|
; |
59 |
; subtract by Q/2 table |
; subtract by Q/2 table |
60 |
; |
; |
61 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
62 |
|
|
63 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
64 |
mmx_sub: |
mmx_sub: |
65 |
%assign quant 1 |
%assign quant 1 |
66 |
%rep 31 |
%rep 31 |
78 |
; |
; |
79 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
80 |
|
|
81 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
82 |
mmx_div: |
mmx_div: |
83 |
%assign quant 1 |
%assign quant 1 |
84 |
%rep 31 |
%rep 31 |
86 |
%assign quant quant+1 |
%assign quant quant+1 |
87 |
%endrep |
%endrep |
88 |
|
|
|
;----------------------------------------------------------------------------- |
|
|
; |
|
|
; add by (odd(Q) ? Q : Q - 1) table |
|
|
; |
|
|
;----------------------------------------------------------------------------- |
|
|
|
|
|
ALIGN 16 |
|
|
mmx_add: |
|
|
%assign quant 1 |
|
|
%rep 31 |
|
|
%if quant % 2 != 0 |
|
|
times 4 dw quant |
|
|
%else |
|
|
times 4 dw quant - 1 |
|
|
%endif |
|
|
%assign quant quant+1 |
|
|
%endrep |
|
|
|
|
|
;----------------------------------------------------------------------------- |
|
|
; |
|
|
; multiple by 2Q table |
|
|
; |
|
|
;----------------------------------------------------------------------------- |
|
|
|
|
|
ALIGN 16 |
|
|
mmx_mul: |
|
|
%assign quant 1 |
|
|
%rep 31 |
|
|
times 4 dw quant*2 |
|
|
%assign quant quant+1 |
|
|
%endrep |
|
|
|
|
|
;----------------------------------------------------------------------------- |
|
|
; |
|
|
; saturation limits |
|
|
; |
|
|
;----------------------------------------------------------------------------- |
|
|
|
|
|
ALIGN 16 |
|
|
sse2_2047: |
|
|
times 8 dw 2047 |
|
|
|
|
|
ALIGN 16 |
|
|
mmx_2047: |
|
|
times 4 dw 2047 |
|
|
|
|
|
ALIGN 8 |
|
|
mmx_32768_minus_2048: |
|
|
times 4 dw (32768-2048) |
|
|
|
|
|
mmx_32767_minus_2047: |
|
|
times 4 dw (32767-2047) |
|
|
|
|
|
|
|
89 |
;============================================================================= |
;============================================================================= |
90 |
; Code |
; Code |
91 |
;============================================================================= |
;============================================================================= |
92 |
|
|
93 |
SECTION .text |
SECTION .rotext align=SECTION_ALIGN |
94 |
|
|
95 |
cglobal quant_h263_intra_mmx |
cglobal quant_h263_intra_mmx |
96 |
cglobal quant_h263_intra_sse2 |
cglobal quant_h263_intra_sse2 |
113 |
; |
; |
114 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
115 |
|
|
116 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
117 |
quant_h263_intra_mmx: |
quant_h263_intra_mmx: |
118 |
|
|
119 |
push ecx |
mov _EAX, prm2 ; data |
120 |
push esi |
mov TMP0, prm4 ; dcscalar |
121 |
push edi |
movsx _EAX, word [_EAX] ; data[0] |
122 |
|
|
123 |
mov edi, [esp + 12 + 4] ; coeff |
sar TMP0, 1 ; dcscalar /2 |
124 |
mov esi, [esp + 12 + 8] ; data |
mov TMP1, _EAX |
125 |
mov eax, [esp + 12 + 12] ; quant |
sar TMP1, 31 ; sgn(data[0]) |
126 |
|
xor TMP0,TMP1 ; *sgn(data[0]) |
127 |
xor ecx, ecx |
sub _EAX,TMP1 |
128 |
cmp al, 1 |
add _EAX,TMP0 ; + (dcscalar/2)*sgn(data[0]) |
129 |
jz .q1loop |
|
130 |
|
mov TMP0, prm3 ; quant |
131 |
movq mm7, [mmx_div + eax * 8 - 8] |
lea TMP1, [mmx_div] |
132 |
|
movq mm7, [TMP1+TMP0 * 8 - 8] |
133 |
ALIGN 16 |
%ifdef ARCH_IS_X86_64 |
134 |
.loop |
%ifdef WINDOWS |
135 |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
mov TMP1, prm2 |
136 |
movq mm3, [esi + 8*ecx + 8] |
%endif |
137 |
pxor mm1, mm1 ; mm1 = 0 |
%endif |
138 |
pxor mm4, mm4 ; |
cdq |
139 |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
idiv prm4d ; dcscalar |
140 |
pcmpgtw mm4, mm3 ; |
%ifdef ARCH_IS_X86_64 |
141 |
pxor mm0, mm1 ; mm0 = |mm0| |
%ifdef WINDOWS |
142 |
pxor mm3, mm4 ; |
mov prm2, TMP1 |
143 |
psubw mm0, mm1 ; displace |
%endif |
144 |
psubw mm3, mm4 ; |
%endif |
145 |
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
cmp TMP0, 1 |
146 |
pmulhw mm3, mm7 ; |
mov TMP1, prm1 ; coeff |
147 |
pxor mm0, mm1 ; mm0 *= sign(mm0) |
je .low |
148 |
pxor mm3, mm4 ; |
|
149 |
psubw mm0, mm1 ; undisplace |
mov TMP0, prm2 ; data |
150 |
psubw mm3, mm4 ; |
push _EAX ; DC |
151 |
movq [edi + 8*ecx], mm0 |
mov _EAX, TMP0 |
|
movq [edi + 8*ecx + 8], mm3 |
|
|
|
|
|
add ecx, 2 |
|
|
cmp ecx, 16 |
|
|
jnz .loop |
|
152 |
|
|
153 |
.done |
mov TMP0,4 |
154 |
|
|
155 |
; caclulate data[0] // (int32_t)dcscalar) |
.loop: |
156 |
mov ecx, [esp + 12 + 16] ; dcscalar |
movq mm0, [_EAX] ; data |
157 |
mov edx, ecx |
pxor mm4,mm4 |
158 |
movsx eax, word [esi] ; data[0] |
movq mm1, [_EAX + 8] |
159 |
shr edx, 1 ; edx = dcscalar /2 |
pcmpgtw mm4,mm0 ; (data<0) |
160 |
cmp eax, 0 |
pxor mm5,mm5 |
161 |
jg .gtzero |
pmulhw mm0,mm7 ; /(2*quant) |
162 |
|
pcmpgtw mm5,mm1 |
163 |
sub eax, edx |
movq mm2, [_EAX+16] |
164 |
jmp short .mul |
psubw mm0,mm4 ; +(data<0) |
165 |
|
pmulhw mm1,mm7 |
166 |
.gtzero |
pxor mm4,mm4 |
167 |
add eax, edx |
movq mm3,[_EAX+24] |
168 |
.mul |
pcmpgtw mm4,mm2 |
169 |
cdq ; expand eax -> edx:eax |
psubw mm1,mm5 |
170 |
idiv ecx ; eax = edx:eax / dcscalar |
pmulhw mm2,mm7 |
171 |
mov [edi], ax ; coeff[0] = ax |
pxor mm5,mm5 |
172 |
|
pcmpgtw mm5,mm3 |
173 |
xor eax, eax ; return(0); |
pmulhw mm3,mm7 |
174 |
pop edi |
psubw mm2,mm4 |
175 |
pop esi |
psubw mm3,mm5 |
176 |
pop ecx |
movq [TMP1], mm0 |
177 |
|
lea _EAX, [_EAX+32] |
178 |
|
movq [TMP1 + 8], mm1 |
179 |
|
movq [TMP1 + 16], mm2 |
180 |
|
movq [TMP1 + 24], mm3 |
181 |
|
|
182 |
|
dec TMP0 |
183 |
|
lea TMP1, [TMP1+32] |
184 |
|
jne .loop |
185 |
|
jmp .end |
186 |
|
|
187 |
|
.low: |
188 |
|
movd mm7,TMP0d |
189 |
|
|
190 |
|
mov TMP0, prm2 |
191 |
|
push _EAX |
192 |
|
mov _EAX, TMP0 |
193 |
|
|
194 |
|
mov TMP0,4 |
195 |
|
.loop_low: |
196 |
|
movq mm0, [_EAX] |
197 |
|
pxor mm4,mm4 |
198 |
|
movq mm1, [_EAX + 8] |
199 |
|
pcmpgtw mm4,mm0 |
200 |
|
pxor mm5,mm5 |
201 |
|
psubw mm0,mm4 |
202 |
|
pcmpgtw mm5,mm1 |
203 |
|
psraw mm0,mm7 |
204 |
|
psubw mm1,mm5 |
205 |
|
movq mm2,[_EAX+16] |
206 |
|
pxor mm4,mm4 |
207 |
|
psraw mm1,mm7 |
208 |
|
pcmpgtw mm4,mm2 |
209 |
|
pxor mm5,mm5 |
210 |
|
psubw mm2,mm4 |
211 |
|
movq mm3,[_EAX+24] |
212 |
|
pcmpgtw mm5,mm3 |
213 |
|
psraw mm2,mm7 |
214 |
|
psubw mm3,mm5 |
215 |
|
movq [TMP1], mm0 |
216 |
|
psraw mm3,mm7 |
217 |
|
movq [TMP1 + 8], mm1 |
218 |
|
movq [TMP1+16],mm2 |
219 |
|
lea _EAX, [_EAX+32] |
220 |
|
movq [TMP1+24],mm3 |
221 |
|
|
222 |
|
dec TMP0 |
223 |
|
lea TMP1, [TMP1+32] |
224 |
|
jne .loop_low |
225 |
|
|
226 |
|
.end: |
227 |
|
|
228 |
|
pop _EAX |
229 |
|
|
230 |
|
mov TMP1, prm1 ; coeff |
231 |
|
mov [TMP1],ax |
232 |
|
xor _EAX,_EAX ; return 0 |
233 |
|
|
234 |
ret |
ret |
235 |
|
ENDFUNC |
|
ALIGN 16 |
|
|
.q1loop |
|
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
|
|
movq mm3, [esi + 8*ecx + 8] |
|
|
pxor mm1, mm1 ; mm1 = 0 |
|
|
pxor mm4, mm4 ; |
|
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
|
|
pcmpgtw mm4, mm3 ; |
|
|
pxor mm0, mm1 ; mm0 = |mm0| |
|
|
pxor mm3, mm4 ; |
|
|
psubw mm0, mm1 ; displace |
|
|
psubw mm3, mm4 ; |
|
|
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
|
|
psrlw mm3, 1 ; |
|
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
|
|
pxor mm3, mm4 |
|
|
psubw mm0, mm1 ; undisplace |
|
|
psubw mm3, mm4 ; |
|
|
movq [edi + 8*ecx], mm0 |
|
|
movq [edi + 8*ecx + 8], mm3 |
|
|
|
|
|
add ecx, 2 |
|
|
cmp ecx, 16 |
|
|
jnz .q1loop |
|
|
jmp short .done |
|
|
|
|
236 |
|
|
237 |
|
|
238 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
245 |
; |
; |
246 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
247 |
|
|
248 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
249 |
quant_h263_intra_sse2: |
quant_h263_intra_sse2: |
250 |
|
|
251 |
push esi |
mov _EAX, prm2 ; data |
|
push edi |
|
252 |
|
|
253 |
mov edi, [esp + 8 + 4] ; coeff |
movsx _EAX, word [_EAX] ; data[0] |
|
mov esi, [esp + 8 + 8] ; data |
|
|
mov eax, [esp + 8 + 12] ; quant |
|
254 |
|
|
255 |
xor ecx, ecx |
mov TMP0,prm4 ; dcscalar |
256 |
cmp al, 1 |
mov TMP1,_EAX |
257 |
jz near .qas2_q1loop |
sar TMP0,1 |
258 |
|
add _EAX,TMP0 |
259 |
|
sub TMP1,TMP0 |
260 |
|
cmovl _EAX,TMP1 ; +/- dcscalar/2 |
261 |
|
mov TMP0, prm3 ; quant |
262 |
|
lea TMP1, [mmx_div] |
263 |
|
movq xmm7, [TMP1+TMP0 * 8 - 8] |
264 |
|
|
265 |
|
%ifdef ARCH_IS_X86_64 |
266 |
|
%ifdef WINDOWS |
267 |
|
mov TMP1, prm2 |
268 |
|
%endif |
269 |
|
%endif |
270 |
|
cdq |
271 |
|
idiv prm4d ; dcscalar |
272 |
|
%ifdef ARCH_IS_X86_64 |
273 |
|
%ifdef WINDOWS |
274 |
|
mov prm2, TMP1 |
275 |
|
%endif |
276 |
|
%endif |
277 |
|
cmp TMP0, 1 |
278 |
|
mov TMP1, prm1 ; coeff |
279 |
|
je near .low |
280 |
|
|
281 |
|
mov TMP0, prm2 |
282 |
|
push _EAX ; DC |
283 |
|
mov _EAX, TMP0 |
284 |
|
|
285 |
.qas2_not1 |
mov TMP0,2 |
|
movq mm7, [mmx_div + eax*8 - 8] |
|
|
movq2dq xmm7, mm7 |
|
286 |
movlhps xmm7, xmm7 |
movlhps xmm7, xmm7 |
287 |
|
|
288 |
ALIGN 16 |
.loop: |
289 |
.qas2_loop |
movdqa xmm0, [_EAX] |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
|
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
|
|
pxor xmm1, xmm1 |
|
290 |
pxor xmm4, xmm4 |
pxor xmm4, xmm4 |
291 |
pcmpgtw xmm1, xmm0 |
movdqa xmm1, [_EAX + 16] |
292 |
pcmpgtw xmm4, xmm3 |
pcmpgtw xmm4,xmm0 |
293 |
pxor xmm0, xmm1 |
pxor xmm5,xmm5 |
|
pxor xmm3, xmm4 |
|
|
psubw xmm0, xmm1 |
|
|
psubw xmm3, xmm4 |
|
294 |
pmulhw xmm0, xmm7 |
pmulhw xmm0, xmm7 |
295 |
|
pcmpgtw xmm5,xmm1 |
296 |
|
movdqa xmm2, [_EAX+32] |
297 |
|
psubw xmm0,xmm4 |
298 |
|
pmulhw xmm1,xmm7 |
299 |
|
pxor xmm4,xmm4 |
300 |
|
movdqa xmm3,[_EAX+48] |
301 |
|
pcmpgtw xmm4,xmm2 |
302 |
|
psubw xmm1,xmm5 |
303 |
|
pmulhw xmm2,xmm7 |
304 |
|
pxor xmm5,xmm5 |
305 |
|
pcmpgtw xmm5,xmm3 |
306 |
pmulhw xmm3, xmm7 |
pmulhw xmm3, xmm7 |
307 |
pxor xmm0, xmm1 |
psubw xmm2,xmm4 |
308 |
pxor xmm3, xmm4 |
psubw xmm3,xmm5 |
309 |
psubw xmm0, xmm1 |
movdqa [TMP1], xmm0 |
310 |
psubw xmm3, xmm4 |
lea _EAX, [_EAX+64] |
311 |
movdqa [edi + ecx*8], xmm0 |
movdqa [TMP1 + 16], xmm1 |
312 |
movdqa [edi + ecx*8 + 16], xmm3 |
movdqa [TMP1 + 32], xmm2 |
313 |
|
movdqa [TMP1 + 48], xmm3 |
314 |
add ecx, 4 |
|
315 |
cmp ecx, 16 |
dec TMP0 |
316 |
jnz .qas2_loop |
lea TMP1, [TMP1+64] |
317 |
|
jne .loop |
318 |
.qas2_done |
jmp .end |
319 |
mov ecx, [esp + 8 + 16] ; dcscalar |
|
320 |
mov edx, ecx |
.low: |
321 |
movsx eax, word [esi] |
movd xmm7,TMP0d |
322 |
shr edx, 1 |
|
323 |
cmp eax, 0 |
mov TMP0, prm2 |
324 |
jg .qas2_gtzero |
push _EAX ; DC |
325 |
|
mov _EAX, TMP0 |
326 |
sub eax, edx |
|
327 |
jmp short .qas2_mul |
mov TMP0,2 |
328 |
|
.loop_low: |
329 |
.qas2_gtzero |
movdqa xmm0, [_EAX] |
|
add eax, edx |
|
|
|
|
|
.qas2_mul |
|
|
cdq |
|
|
idiv ecx |
|
|
|
|
|
mov [edi], ax |
|
|
|
|
|
xor eax, eax ; return(0); |
|
|
pop edi |
|
|
pop esi |
|
|
|
|
|
ret |
|
|
|
|
|
ALIGN 16 |
|
|
.qas2_q1loop |
|
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
|
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
|
|
pxor xmm1, xmm1 |
|
330 |
pxor xmm4, xmm4 |
pxor xmm4, xmm4 |
331 |
pcmpgtw xmm1, xmm0 |
movdqa xmm1, [_EAX + 16] |
332 |
pcmpgtw xmm4, xmm3 |
pcmpgtw xmm4,xmm0 |
333 |
pxor xmm0, xmm1 |
pxor xmm5,xmm5 |
334 |
pxor xmm3, xmm4 |
psubw xmm0,xmm4 |
335 |
psubw xmm0, xmm1 |
pcmpgtw xmm5,xmm1 |
336 |
psubw xmm3, xmm4 |
psraw xmm0,xmm7 |
337 |
psrlw xmm0, 1 |
psubw xmm1,xmm5 |
338 |
psrlw xmm3, 1 |
movdqa xmm2,[_EAX+32] |
339 |
pxor xmm0, xmm1 |
pxor xmm4,xmm4 |
340 |
pxor xmm3, xmm4 |
psraw xmm1,xmm7 |
341 |
psubw xmm0, xmm1 |
pcmpgtw xmm4,xmm2 |
342 |
psubw xmm3, xmm4 |
pxor xmm5,xmm5 |
343 |
movdqa [edi + ecx*8], xmm0 |
psubw xmm2,xmm4 |
344 |
movdqa [edi + ecx*8 + 16], xmm3 |
movdqa xmm3,[_EAX+48] |
345 |
|
pcmpgtw xmm5,xmm3 |
346 |
add ecx, 4 |
psraw xmm2,xmm7 |
347 |
cmp ecx, 16 |
psubw xmm3,xmm5 |
348 |
jnz .qas2_q1loop |
movdqa [TMP1], xmm0 |
349 |
jmp near .qas2_done |
psraw xmm3,xmm7 |
350 |
|
movdqa [TMP1+16], xmm1 |
351 |
|
movdqa [TMP1+32],xmm2 |
352 |
|
lea _EAX, [_EAX+64] |
353 |
|
movdqa [TMP1+48],xmm3 |
354 |
|
|
355 |
|
dec TMP0 |
356 |
|
lea TMP1, [TMP1+64] |
357 |
|
jne .loop_low |
358 |
|
|
359 |
|
.end: |
360 |
|
|
361 |
|
pop _EAX |
362 |
|
|
363 |
|
mov TMP1, prm1 ; coeff |
364 |
|
mov [TMP1],ax |
365 |
|
xor _EAX,_EAX ; return 0 |
366 |
|
|
367 |
|
ret |
368 |
|
ENDFUNC |
369 |
|
|
370 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
371 |
; |
; |
376 |
; |
; |
377 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
378 |
|
|
379 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
380 |
quant_h263_inter_mmx: |
quant_h263_inter_mmx: |
381 |
|
|
382 |
push ecx |
mov TMP1, prm1 ; coeff |
383 |
push esi |
mov _EAX, prm3 ; quant |
|
push edi |
|
|
|
|
|
mov edi, [esp + 12 + 4] ; coeff |
|
|
mov esi, [esp + 12 + 8] ; data |
|
|
mov eax, [esp + 12 + 12] ; quant |
|
|
|
|
|
xor ecx, ecx |
|
384 |
|
|
385 |
pxor mm5, mm5 ; sum |
pxor mm5, mm5 ; sum |
386 |
movq mm6, [mmx_sub + eax * 8 - 8] ; sub |
lea TMP0, [mmx_sub] |
387 |
|
movq mm6, [TMP0 + _EAX * 8 - 8] ; sub |
388 |
|
|
389 |
cmp al, 1 |
cmp al, 1 |
390 |
jz .q1loop |
jz near .q1routine |
391 |
|
|
392 |
movq mm7, [mmx_div + eax * 8 - 8] ; divider |
lea TMP0, [mmx_div] |
393 |
|
movq mm7, [TMP0 + _EAX * 8 - 8] ; divider |
394 |
|
|
395 |
ALIGN 8 |
xor TMP0, TMP0 |
396 |
.loop |
mov _EAX, prm2 ; data |
397 |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
|
398 |
movq mm3, [esi + 8*ecx + 8] |
ALIGN SECTION_ALIGN |
399 |
|
.loop: |
400 |
|
movq mm0, [_EAX + 8*TMP0] ; mm0 = [1st] |
401 |
|
movq mm3, [_EAX + 8*TMP0 + 8] |
402 |
pxor mm1, mm1 ; mm1 = 0 |
pxor mm1, mm1 ; mm1 = 0 |
403 |
pxor mm4, mm4 ; |
pxor mm4, mm4 ; |
404 |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
417 |
pxor mm3, mm4 ; |
pxor mm3, mm4 ; |
418 |
psubw mm0, mm1 ; undisplace |
psubw mm0, mm1 ; undisplace |
419 |
psubw mm3, mm4 |
psubw mm3, mm4 |
420 |
movq [edi + 8*ecx], mm0 |
movq [TMP1 + 8*TMP0], mm0 |
421 |
movq [edi + 8*ecx + 8], mm3 |
movq [TMP1 + 8*TMP0 + 8], mm3 |
422 |
|
|
423 |
add ecx, 2 |
add TMP0, 2 |
424 |
cmp ecx, 16 |
cmp TMP0, 16 |
425 |
jnz .loop |
jnz .loop |
426 |
|
|
427 |
.done |
.done: |
428 |
pmaddwd mm5, [plus_one] |
pmaddwd mm5, [plus_one] |
429 |
movq mm0, mm5 |
movq mm0, mm5 |
430 |
psrlq mm5, 32 |
psrlq mm5, 32 |
431 |
paddd mm0, mm5 |
paddd mm0, mm5 |
432 |
|
|
433 |
movd eax, mm0 ; return sum |
movd eax, mm0 ; return sum |
|
pop edi |
|
|
pop esi |
|
|
pop ecx |
|
434 |
|
|
435 |
ret |
ret |
436 |
|
|
437 |
ALIGN 8 |
.q1routine: |
438 |
.q1loop |
xor TMP0, TMP0 |
439 |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
mov _EAX, prm2 ; data |
440 |
movq mm3, [esi + 8*ecx+ 8] ; |
|
441 |
|
ALIGN SECTION_ALIGN |
442 |
|
.q1loop: |
443 |
|
movq mm0, [_EAX + 8*TMP0] ; mm0 = [1st] |
444 |
|
movq mm3, [_EAX + 8*TMP0+ 8] ; |
445 |
pxor mm1, mm1 ; mm1 = 0 |
pxor mm1, mm1 ; mm1 = 0 |
446 |
pxor mm4, mm4 ; |
pxor mm4, mm4 ; |
447 |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
460 |
pxor mm3, mm4 ; |
pxor mm3, mm4 ; |
461 |
psubw mm0, mm1 ; undisplace |
psubw mm0, mm1 ; undisplace |
462 |
psubw mm3, mm4 |
psubw mm3, mm4 |
463 |
movq [edi + 8*ecx], mm0 |
movq [TMP1 + 8*TMP0], mm0 |
464 |
movq [edi + 8*ecx + 8], mm3 |
movq [TMP1 + 8*TMP0 + 8], mm3 |
465 |
|
|
466 |
add ecx, 2 |
add TMP0, 2 |
467 |
cmp ecx, 16 |
cmp TMP0, 16 |
468 |
jnz .q1loop |
jnz .q1loop |
469 |
|
|
470 |
jmp .done |
jmp .done |
471 |
|
ENDFUNC |
472 |
|
|
473 |
|
|
474 |
|
|
481 |
; |
; |
482 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
483 |
|
|
484 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
485 |
quant_h263_inter_sse2: |
quant_h263_inter_sse2: |
486 |
|
|
487 |
push esi |
mov TMP1, prm1 ; coeff |
488 |
push edi |
mov _EAX, prm3 ; quant |
|
|
|
|
mov edi, [esp + 8 + 4] ; coeff |
|
|
mov esi, [esp + 8 + 8] ; data |
|
|
mov eax, [esp + 8 + 12] ; quant |
|
|
|
|
|
xor ecx, ecx |
|
489 |
|
|
490 |
pxor xmm5, xmm5 ; sum |
pxor xmm5, xmm5 ; sum |
491 |
|
|
492 |
movq mm0, [mmx_sub + eax*8 - 8] ; sub |
lea TMP0, [mmx_sub] |
493 |
|
movq mm0, [TMP0 + _EAX*8 - 8] ; sub |
494 |
movq2dq xmm6, mm0 ; load into low 8 bytes |
movq2dq xmm6, mm0 ; load into low 8 bytes |
495 |
movlhps xmm6, xmm6 ; duplicate into high 8 bytes |
movlhps xmm6, xmm6 ; duplicate into high 8 bytes |
496 |
|
|
497 |
cmp al, 1 |
cmp al, 1 |
498 |
jz near .qes2_q1loop |
jz near .qes2_q1_routine |
499 |
|
|
500 |
|
.qes2_not1: |
501 |
|
lea TMP0, [mmx_div] |
502 |
|
movq mm0, [TMP0 + _EAX*8 - 8] ; divider |
503 |
|
|
504 |
|
xor TMP0, TMP0 |
505 |
|
mov _EAX, prm2 ; data |
506 |
|
|
|
.qes2_not1 |
|
|
movq mm0, [mmx_div + eax*8 - 8] ; divider |
|
507 |
movq2dq xmm7, mm0 |
movq2dq xmm7, mm0 |
508 |
movlhps xmm7, xmm7 |
movlhps xmm7, xmm7 |
509 |
|
|
510 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
511 |
.qes2_loop |
.qes2_loop: |
512 |
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
movdqa xmm0, [_EAX + TMP0*8] ; xmm0 = [1st] |
513 |
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
movdqa xmm3, [_EAX + TMP0*8 + 16] ; xmm3 = [2nd] |
514 |
pxor xmm1, xmm1 |
pxor xmm1, xmm1 |
515 |
pxor xmm4, xmm4 |
pxor xmm4, xmm4 |
516 |
pcmpgtw xmm1, xmm0 |
pcmpgtw xmm1, xmm0 |
529 |
pxor xmm3, xmm4 |
pxor xmm3, xmm4 |
530 |
psubw xmm0, xmm1 |
psubw xmm0, xmm1 |
531 |
psubw xmm3, xmm4 |
psubw xmm3, xmm4 |
532 |
movdqa [edi + ecx*8], xmm0 |
movdqa [TMP1 + TMP0*8], xmm0 |
533 |
movdqa [edi + ecx*8 + 16], xmm3 |
movdqa [TMP1 + TMP0*8 + 16], xmm3 |
534 |
|
|
535 |
add ecx, 4 |
add TMP0, 4 |
536 |
cmp ecx, 16 |
cmp TMP0, 16 |
537 |
jnz .qes2_loop |
jnz .qes2_loop |
538 |
|
|
539 |
.qes2_done |
.qes2_done: |
540 |
movdqu xmm6, [plus_one] |
movdqu xmm6, [plus_one] |
541 |
pmaddwd xmm5, xmm6 |
pmaddwd xmm5, xmm6 |
542 |
movhlps xmm6, xmm5 |
movhlps xmm6, xmm5 |
549 |
|
|
550 |
movd eax, mm0 ; return sum |
movd eax, mm0 ; return sum |
551 |
|
|
|
pop edi |
|
|
pop esi |
|
|
|
|
552 |
ret |
ret |
553 |
|
|
554 |
ALIGN 16 |
.qes2_q1_routine: |
555 |
.qes2_q1loop |
xor TMP0, TMP0 |
556 |
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
mov _EAX, prm2 ; data |
557 |
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
|
558 |
|
ALIGN SECTION_ALIGN |
559 |
|
.qes2_q1loop: |
560 |
|
movdqa xmm0, [_EAX + TMP0*8] ; xmm0 = [1st] |
561 |
|
movdqa xmm3, [_EAX + TMP0*8 + 16] ; xmm3 = [2nd] |
562 |
pxor xmm1, xmm1 |
pxor xmm1, xmm1 |
563 |
pxor xmm4, xmm4 |
pxor xmm4, xmm4 |
564 |
pcmpgtw xmm1, xmm0 |
pcmpgtw xmm1, xmm0 |
577 |
pxor xmm3, xmm4 |
pxor xmm3, xmm4 |
578 |
psubw xmm0, xmm1 |
psubw xmm0, xmm1 |
579 |
psubw xmm3, xmm4 |
psubw xmm3, xmm4 |
580 |
movdqa [edi + ecx*8], xmm0 |
movdqa [TMP1 + TMP0*8], xmm0 |
581 |
movdqa [edi + ecx*8 + 16], xmm3 |
movdqa [TMP1 + TMP0*8 + 16], xmm3 |
582 |
|
|
583 |
add ecx, 4 |
add TMP0, 4 |
584 |
cmp ecx, 16 |
cmp TMP0, 16 |
585 |
jnz .qes2_q1loop |
jnz .qes2_q1loop |
586 |
jmp .qes2_done |
jmp .qes2_done |
587 |
|
ENDFUNC |
588 |
|
|
589 |
|
|
590 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
597 |
; |
; |
598 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
599 |
|
|
600 |
; note: we only saturate to +2047 *before* restoring the sign. |
ALIGN SECTION_ALIGN |
|
; Hence, final clamp really is [-2048,2047] |
|
|
|
|
|
ALIGN 16 |
|
601 |
dequant_h263_intra_mmx: |
dequant_h263_intra_mmx: |
602 |
|
|
603 |
mov edx, [esp+ 4] ; data |
mov TMP0, prm3 ; quant |
604 |
mov ecx, [esp+ 8] ; coeff |
mov _EAX, prm2 ; coeff |
605 |
mov eax, [esp+12] ; quant |
pcmpeqw mm0,mm0 |
606 |
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
lea TMP1, [mmx_quant] |
607 |
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
movq mm6, [TMP1 + TMP0*8] ; quant |
608 |
mov eax, -16 |
shl TMP0,31 ; quant & 1 ? 0 : - 1 |
609 |
|
movq mm7,mm6 |
610 |
ALIGN 16 |
movq mm5,mm0 |
611 |
.loop |
movd mm1,TMP0d |
612 |
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
mov TMP1, prm1 ; data |
613 |
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
psllw mm0,mm1 |
614 |
pxor mm1, mm1 |
paddw mm7,mm7 ; 2*quant |
615 |
|
paddw mm6,mm0 ; quant-1 |
616 |
|
psllw mm5,12 |
617 |
|
mov TMP0,8 |
618 |
|
psrlw mm5,1 |
619 |
|
|
620 |
|
.loop: |
621 |
|
movq mm0,[_EAX] |
622 |
|
pxor mm2,mm2 |
623 |
pxor mm4, mm4 |
pxor mm4, mm4 |
624 |
pcmpgtw mm1, mm0 ; sign(c) |
pcmpgtw mm2,mm0 |
625 |
pcmpgtw mm4, mm3 ; sign(c') |
pcmpeqw mm4,mm0 |
626 |
|
pmullw mm0,mm7 ; * 2 * quant |
627 |
|
movq mm1,[_EAX+8] |
628 |
|
psubw mm0,mm2 |
629 |
|
pxor mm2,mm6 |
630 |
|
pxor mm3,mm3 |
631 |
|
pandn mm4,mm2 |
632 |
pxor mm2, mm2 |
pxor mm2, mm2 |
633 |
pxor mm5, mm5 |
pcmpgtw mm3,mm1 |
634 |
pcmpeqw mm2, mm0 ; c is zero |
pcmpeqw mm2,mm1 |
635 |
pcmpeqw mm5, mm3 ; c' is zero |
pmullw mm1,mm7 |
636 |
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
paddw mm0,mm4 |
637 |
pandn mm5, mm6 |
psubw mm1,mm3 |
638 |
pxor mm0, mm1 ; negate if negative |
pxor mm3,mm6 |
639 |
pxor mm3, mm4 ; negate if negative |
pandn mm2,mm3 |
640 |
psubw mm0, mm1 |
paddsw mm0, mm5 ; saturate |
641 |
psubw mm3, mm4 |
paddw mm1,mm2 |
642 |
pmullw mm0, mm7 ; *= 2Q |
|
643 |
pmullw mm3, mm7 ; *= 2Q |
paddsw mm1, mm5 |
644 |
paddw mm0, mm2 ; + offset |
psubsw mm0, mm5 |
645 |
paddw mm3, mm5 ; + offset |
psubsw mm1, mm5 |
646 |
paddw mm0, mm1 ; negate back |
psubsw mm0, mm5 |
647 |
paddw mm3, mm4 ; negate back |
psubsw mm1, mm5 |
648 |
|
paddsw mm0, mm5 |
649 |
; saturates to +2047 |
paddsw mm1, mm5 |
650 |
movq mm2, [mmx_32767_minus_2047] |
|
651 |
add eax, 2 |
movq [TMP1],mm0 |
652 |
paddsw mm0, mm2 |
lea _EAX,[_EAX+16] |
653 |
paddsw mm3, mm2 |
movq [TMP1+8],mm1 |
654 |
psubsw mm0, mm2 |
|
655 |
psubsw mm3, mm2 |
dec TMP0 |
656 |
|
lea TMP1,[TMP1+16] |
657 |
pxor mm0, mm1 |
jne .loop |
|
pxor mm3, mm4 |
|
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
|
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
|
|
jnz near .loop |
|
658 |
|
|
659 |
; deal with DC |
; deal with DC |
660 |
movd mm0, [ecx] |
mov _EAX, prm2 ; coeff |
661 |
pmullw mm0, [esp+16] ; dcscalar |
movd mm1,prm4d ; dcscalar |
662 |
movq mm2, [mmx_32767_minus_2047] |
movd mm0,[_EAX] ; coeff[0] |
663 |
paddsw mm0, mm2 |
pmullw mm0,mm1 ; * dcscalar |
664 |
psubsw mm0, mm2 |
mov TMP1, prm1 ; data |
665 |
movq mm3, [mmx_32768_minus_2048] |
paddsw mm0, mm5 ; saturate + |
666 |
psubsw mm0, mm3 |
psubsw mm0, mm5 |
667 |
paddsw mm0, mm3 |
psubsw mm0, mm5 ; saturate - |
668 |
|
paddsw mm0, mm5 |
669 |
movd eax, mm0 |
movd eax, mm0 |
670 |
mov [edx], ax |
mov [TMP1], ax |
671 |
|
|
672 |
xor eax, eax ; return(0); |
xor _EAX, _EAX ; return 0 |
673 |
ret |
ret |
674 |
|
ENDFUNC |
675 |
|
|
676 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
677 |
; |
; |
683 |
; |
; |
684 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
685 |
|
|
|
; this is the same as dequant_inter_mmx, except that we're |
|
|
; saturating using 'pminsw' (saves 2 cycles/loop => ~5% faster) |
|
686 |
|
|
687 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
688 |
dequant_h263_intra_xmm: |
dequant_h263_intra_xmm: |
689 |
|
|
690 |
mov edx, [esp+ 4] ; data |
mov TMP0, prm3 ; quant |
691 |
mov ecx, [esp+ 8] ; coeff |
mov _EAX, prm2 ; coeff |
692 |
mov eax, [esp+12] ; quant |
|
693 |
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
movd mm6,TMP0d ; quant |
694 |
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
pcmpeqw mm0,mm0 |
695 |
mov eax, -16 |
pshufw mm6,mm6,0 ; all quant |
696 |
|
shl TMP0,31 |
697 |
ALIGN 16 |
movq mm5,mm0 |
698 |
.loop |
movq mm7,mm6 |
699 |
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
movd mm1,TMP0d |
700 |
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
mov TMP1, prm1 ; data |
701 |
pxor mm1, mm1 |
psllw mm0,mm1 ; quant & 1 ? 0 : - 1 |
702 |
pxor mm4, mm4 |
movq mm4,mm5 |
703 |
pcmpgtw mm1, mm0 ; sign(c) |
paddw mm7,mm7 ; quant*2 |
704 |
pcmpgtw mm4, mm3 ; sign(c') |
paddw mm6,mm0 ; quant-1 |
705 |
|
psrlw mm4,5 ; mm4=2047 |
706 |
|
mov TMP0,8 |
707 |
|
pxor mm5,mm4 ; mm5=-2048 |
708 |
|
|
709 |
|
.loop: |
710 |
|
movq mm0,[_EAX] |
711 |
pxor mm2, mm2 |
pxor mm2, mm2 |
712 |
pxor mm5, mm5 |
pxor mm3,mm3 |
713 |
pcmpeqw mm2, mm0 ; c is zero |
|
714 |
pcmpeqw mm5, mm3 ; c' is zero |
pcmpgtw mm2,mm0 |
715 |
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
pcmpeqw mm3,mm0 ; if coeff==0... |
716 |
pandn mm5, mm6 |
pmullw mm0,mm7 ; * 2 * quant |
717 |
pxor mm0, mm1 ; negate if negative |
movq mm1,[_EAX+8] |
718 |
pxor mm3, mm4 ; negate if negative |
|
719 |
psubw mm0, mm1 |
psubw mm0,mm2 |
720 |
psubw mm3, mm4 |
pxor mm2,mm6 |
721 |
pmullw mm0, mm7 ; *= 2Q |
pandn mm3,mm2 ; ...then data=0 |
722 |
pmullw mm3, mm7 ; *= 2Q |
pxor mm2,mm2 |
723 |
paddw mm0, mm2 ; + offset |
paddw mm0,mm3 |
724 |
paddw mm3, mm5 ; + offset |
pxor mm3,mm3 |
725 |
paddw mm0, mm1 ; negate back |
pcmpeqw mm2,mm1 |
726 |
paddw mm3, mm4 ; negate back |
pcmpgtw mm3,mm1 |
727 |
|
pmullw mm1,mm7 |
728 |
; saturates to +2047 |
|
729 |
movq mm2, [mmx_2047] |
pminsw mm0,mm4 |
730 |
pminsw mm0, mm2 |
psubw mm1,mm3 |
731 |
add eax, 2 |
pxor mm3,mm6 |
732 |
pminsw mm3, mm2 |
pandn mm2,mm3 |
733 |
|
paddw mm1,mm2 |
734 |
pxor mm0, mm1 |
|
735 |
pxor mm3, mm4 |
pmaxsw mm0,mm5 |
736 |
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
pminsw mm1,mm4 |
737 |
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
movq [TMP1],mm0 |
738 |
jnz near .loop |
pmaxsw mm1,mm5 |
739 |
|
lea _EAX,[_EAX+16] |
740 |
|
movq [TMP1+8],mm1 |
741 |
|
|
742 |
|
dec TMP0 |
743 |
|
lea TMP1,[TMP1+16] |
744 |
|
jne .loop |
745 |
|
|
746 |
; deal with DC |
; deal with DC |
747 |
movd mm0, [ecx] |
mov _EAX, prm2 ; coeff |
748 |
pmullw mm0, [esp+16] ; dcscalar |
movd mm1,prm4d ; dcscalar |
749 |
movq mm2, [mmx_32767_minus_2047] |
movd mm0, [_EAX] |
750 |
paddsw mm0, mm2 |
pmullw mm0, mm1 |
751 |
psubsw mm0, mm2 |
mov TMP1, prm1 ; data |
752 |
movq mm2, [mmx_32768_minus_2048] |
pminsw mm0,mm4 |
753 |
psubsw mm0, mm2 |
pmaxsw mm0,mm5 |
|
paddsw mm0, mm2 |
|
754 |
movd eax, mm0 |
movd eax, mm0 |
755 |
mov [edx], ax |
mov [TMP1], ax |
756 |
|
|
757 |
xor eax, eax |
xor _EAX, _EAX ; return 0 |
758 |
ret |
ret |
759 |
|
ENDFUNC |
760 |
|
|
761 |
|
|
762 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
769 |
; |
; |
770 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
771 |
|
|
772 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
773 |
dequant_h263_intra_sse2: |
dequant_h263_intra_sse2: |
|
mov edx, [esp+ 4] ; data |
|
|
mov ecx, [esp+ 8] ; coeff |
|
|
mov eax, [esp+12] ; quant |
|
|
movq mm6, [mmx_add + eax * 8 - 8] |
|
|
movq mm7, [mmx_mul + eax * 8 - 8] |
|
|
movq2dq xmm6, mm6 |
|
|
movq2dq xmm7, mm7 |
|
|
movlhps xmm6, xmm6 |
|
|
movlhps xmm7, xmm7 |
|
|
mov eax, -16 |
|
774 |
|
|
775 |
ALIGN 16 |
mov TMP0, prm3 ; quant |
776 |
.loop |
mov _EAX, prm2 ; coeff |
777 |
movdqa xmm0, [ecx + 8*16 + 8*eax] ; c = coeff[i] |
|
778 |
movdqa xmm3, [ecx + 8*16 + 8*eax+ 16] |
movd xmm6,TMP0d ; quant |
779 |
pxor xmm1, xmm1 |
|
780 |
pxor xmm4, xmm4 |
shl TMP0,31 |
781 |
pcmpgtw xmm1, xmm0 ; sign(c) |
pshuflw xmm6,xmm6,0 |
782 |
pcmpgtw xmm4, xmm3 |
pcmpeqw xmm0,xmm0 |
783 |
|
movlhps xmm6,xmm6 ; all quant |
784 |
|
movd xmm1,TMP0d |
785 |
|
movdqa xmm5,xmm0 |
786 |
|
movdqa xmm7,xmm6 |
787 |
|
mov TMP1, prm1 ; data |
788 |
|
paddw xmm7,xmm7 ; quant *2 |
789 |
|
psllw xmm0,xmm1 ; quant & 1 ? 0 : - 1 |
790 |
|
movdqa xmm4,xmm5 |
791 |
|
paddw xmm6,xmm0 ; quant-1 |
792 |
|
psrlw xmm4,5 ; 2047 |
793 |
|
mov TMP0,4 |
794 |
|
pxor xmm5,xmm4 ; mm5=-2048 |
795 |
|
|
796 |
|
.loop: |
797 |
|
movdqa xmm0,[_EAX] |
798 |
pxor xmm2, xmm2 |
pxor xmm2, xmm2 |
799 |
pxor xmm5, xmm5 |
pxor xmm3,xmm3 |
|
pcmpeqw xmm2, xmm0 ; c is zero |
|
|
pcmpeqw xmm5, xmm3 |
|
|
pandn xmm2, xmm6 ; offset = isZero ? 0 : quant_add |
|
|
pandn xmm5, xmm6 |
|
|
pxor xmm0, xmm1 ; negate if negative |
|
|
pxor xmm3, xmm4 |
|
|
psubw xmm0, xmm1 |
|
|
psubw xmm3, xmm4 |
|
|
pmullw xmm0, xmm7 ; *= 2Q |
|
|
pmullw xmm3, xmm7 |
|
|
paddw xmm0, xmm2 ; + offset |
|
|
paddw xmm3, xmm5 |
|
|
paddw xmm0, xmm1 ; negate back |
|
|
paddw xmm3, xmm4 |
|
|
|
|
|
; saturates to +2047 |
|
|
movdqa xmm2, [sse2_2047] |
|
|
pminsw xmm0, xmm2 |
|
|
add eax, 4 |
|
|
pminsw xmm3, xmm2 |
|
800 |
|
|
801 |
pxor xmm0, xmm1 |
pcmpgtw xmm2,xmm0 |
802 |
pxor xmm3, xmm4 |
pcmpeqw xmm3,xmm0 |
803 |
movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0 |
pmullw xmm0,xmm7 ; * 2 * quant |
804 |
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
movdqa xmm1,[_EAX+16] |
805 |
jnz near .loop |
|
806 |
|
psubw xmm0,xmm2 |
807 |
|
pxor xmm2,xmm6 |
808 |
|
pandn xmm3,xmm2 |
809 |
|
pxor xmm2,xmm2 |
810 |
|
paddw xmm0,xmm3 |
811 |
|
pxor xmm3,xmm3 |
812 |
|
pcmpeqw xmm2,xmm1 |
813 |
|
pcmpgtw xmm3,xmm1 |
814 |
|
pmullw xmm1,xmm7 |
815 |
|
|
816 |
|
pminsw xmm0,xmm4 |
817 |
|
psubw xmm1,xmm3 |
818 |
|
pxor xmm3,xmm6 |
819 |
|
pandn xmm2,xmm3 |
820 |
|
paddw xmm1,xmm2 |
821 |
|
|
822 |
|
pmaxsw xmm0,xmm5 |
823 |
|
pminsw xmm1,xmm4 |
824 |
|
movdqa [TMP1],xmm0 |
825 |
|
pmaxsw xmm1,xmm5 |
826 |
|
lea _EAX,[_EAX+32] |
827 |
|
movdqa [TMP1+16],xmm1 |
828 |
|
|
829 |
|
dec TMP0 |
830 |
|
lea TMP1,[TMP1+32] |
831 |
|
jne .loop |
832 |
|
|
833 |
; deal with DC |
; deal with DC |
|
movd mm0, [ecx] |
|
|
pmullw mm0, [esp+16] ; dcscalar |
|
|
movq mm2, [mmx_32767_minus_2047] |
|
|
paddsw mm0, mm2 |
|
|
psubsw mm0, mm2 |
|
|
movq mm2, [mmx_32768_minus_2048] |
|
|
psubsw mm0, mm2 |
|
|
paddsw mm0, mm2 |
|
|
movd eax, mm0 |
|
|
mov [edx], ax |
|
834 |
|
|
835 |
xor eax, eax |
mov _EAX, prm2 ; coeff |
836 |
|
movsx _EAX,word [_EAX] |
837 |
|
imul prm4d ; dcscalar |
838 |
|
mov TMP1, prm1 ; data |
839 |
|
movd xmm0,eax |
840 |
|
pminsw xmm0,xmm4 |
841 |
|
pmaxsw xmm0,xmm5 |
842 |
|
movd eax,xmm0 |
843 |
|
|
844 |
|
mov [TMP1], ax |
845 |
|
|
846 |
|
xor _EAX, _EAX ; return 0 |
847 |
ret |
ret |
848 |
|
ENDFUNC |
849 |
|
|
850 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
851 |
; |
; |
856 |
; |
; |
857 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
858 |
|
|
859 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
860 |
dequant_h263_inter_mmx: |
dequant_h263_inter_mmx: |
861 |
|
|
862 |
mov edx, [esp+ 4] ; data |
mov TMP0, prm3 ; quant |
863 |
mov ecx, [esp+ 8] ; coeff |
mov _EAX, prm2 ; coeff |
864 |
mov eax, [esp+12] ; quant |
pcmpeqw mm0,mm0 |
865 |
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
lea TMP1, [mmx_quant] |
866 |
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
movq mm6, [TMP1 + TMP0*8] ; quant |
867 |
mov eax, -16 |
shl TMP0,31 ; odd/even |
868 |
|
movq mm7,mm6 |
869 |
ALIGN 16 |
movd mm1,TMP0d |
870 |
.loop |
mov TMP1, prm1 ; data |
871 |
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
movq mm5,mm0 |
872 |
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
psllw mm0,mm1 ; quant & 1 ? 0 : - 1 |
873 |
pxor mm1, mm1 |
paddw mm7,mm7 ; quant*2 |
874 |
|
paddw mm6,mm0 ; quant & 1 ? quant : quant - 1 |
875 |
|
psllw mm5,12 |
876 |
|
mov TMP0,8 |
877 |
|
psrlw mm5,1 ; 32767-2047 (32768-2048) |
878 |
|
|
879 |
|
.loop: |
880 |
|
movq mm0,[_EAX] |
881 |
pxor mm4, mm4 |
pxor mm4, mm4 |
|
pcmpgtw mm1, mm0 ; sign(c) |
|
|
pcmpgtw mm4, mm3 ; sign(c') |
|
882 |
pxor mm2, mm2 |
pxor mm2, mm2 |
883 |
pxor mm5, mm5 |
pcmpeqw mm4,mm0 ; if coeff==0... |
884 |
pcmpeqw mm2, mm0 ; c is zero |
pcmpgtw mm2,mm0 |
885 |
pcmpeqw mm5, mm3 ; c' is zero |
pmullw mm0,mm7 ; * 2 * quant |
886 |
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
pxor mm3,mm3 |
887 |
pandn mm5, mm6 |
psubw mm0,mm2 |
888 |
pxor mm0, mm1 ; negate if negative |
movq mm1,[_EAX+8] |
889 |
pxor mm3, mm4 ; negate if negative |
pxor mm2,mm6 |
890 |
psubw mm0, mm1 |
pcmpgtw mm3,mm1 |
891 |
psubw mm3, mm4 |
pandn mm4,mm2 ; ... then data==0 |
892 |
pmullw mm0, mm7 ; *= 2Q |
pmullw mm1,mm7 |
893 |
pmullw mm3, mm7 ; *= 2Q |
pxor mm2,mm2 |
894 |
paddw mm0, mm2 ; + offset |
pcmpeqw mm2,mm1 |
895 |
paddw mm3, mm5 ; + offset |
psubw mm1,mm3 |
896 |
paddw mm0, mm1 ; negate back |
pxor mm3,mm6 |
897 |
paddw mm3, mm4 ; negate back |
pandn mm2,mm3 |
898 |
; saturates to +2047 |
paddw mm0,mm4 |
899 |
movq mm2, [mmx_32767_minus_2047] |
paddw mm1,mm2 |
900 |
add eax, 2 |
|
901 |
paddsw mm0, mm2 |
paddsw mm0, mm5 ; saturate |
902 |
paddsw mm3, mm2 |
paddsw mm1, mm5 |
903 |
psubsw mm0, mm2 |
psubsw mm0, mm5 |
904 |
psubsw mm3, mm2 |
psubsw mm1, mm5 |
905 |
|
psubsw mm0, mm5 |
906 |
pxor mm0, mm1 |
psubsw mm1, mm5 |
907 |
pxor mm3, mm4 |
paddsw mm0, mm5 |
908 |
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
paddsw mm1, mm5 |
909 |
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
|
910 |
jnz near .loop |
movq [TMP1],mm0 |
911 |
|
lea _EAX,[_EAX+16] |
912 |
|
movq [TMP1+8],mm1 |
913 |
|
|
914 |
|
dec TMP0 |
915 |
|
lea TMP1,[TMP1+16] |
916 |
|
jne .loop |
917 |
|
|
918 |
xor eax, eax |
xor _EAX, _EAX ; return 0 |
919 |
ret |
ret |
920 |
|
ENDFUNC |
921 |
|
|
922 |
|
|
923 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
924 |
; |
; |
928 |
; const uint16_t *mpeg_matrices); |
; const uint16_t *mpeg_matrices); |
929 |
; |
; |
930 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
931 |
|
ALIGN SECTION_ALIGN |
|
; this is the same as dequant_inter_mmx, |
|
|
; except that we're saturating using 'pminsw' (saves 2 cycles/loop) |
|
|
|
|
|
ALIGN 16 |
|
932 |
dequant_h263_inter_xmm: |
dequant_h263_inter_xmm: |
933 |
|
|
934 |
mov edx, [esp+ 4] ; data |
mov TMP0, prm3 ; quant |
935 |
mov ecx, [esp+ 8] ; coeff |
mov _EAX, prm2 ; coeff |
936 |
mov eax, [esp+12] ; quant |
pcmpeqw mm0,mm0 |
937 |
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
lea TMP1, [mmx_quant] |
938 |
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
movq mm6, [TMP1 + TMP0*8] ; quant |
939 |
mov eax, -16 |
shl TMP0,31 |
940 |
|
movq mm5,mm0 |
941 |
ALIGN 16 |
movd mm1,TMP0d |
942 |
.loop |
movq mm7,mm6 |
943 |
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
psllw mm0,mm1 |
944 |
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
mov TMP1, prm1 ; data |
945 |
pxor mm1, mm1 |
movq mm4,mm5 |
946 |
pxor mm4, mm4 |
paddw mm7,mm7 |
947 |
pcmpgtw mm1, mm0 ; sign(c) |
paddw mm6,mm0 ; quant-1 |
948 |
pcmpgtw mm4, mm3 ; sign(c') |
|
949 |
|
psrlw mm4,5 |
950 |
|
mov TMP0,8 |
951 |
|
pxor mm5,mm4 ; mm5=-2048 |
952 |
|
|
953 |
|
.loop: |
954 |
|
movq mm0,[_EAX] |
955 |
|
pxor mm3,mm3 |
956 |
pxor mm2, mm2 |
pxor mm2, mm2 |
957 |
pxor mm5, mm5 |
pcmpeqw mm3,mm0 |
958 |
pcmpeqw mm2, mm0 ; c is zero |
pcmpgtw mm2,mm0 |
959 |
pcmpeqw mm5, mm3 ; c' is zero |
pmullw mm0,mm7 ; * 2 * quant |
960 |
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
pandn mm3,mm6 |
961 |
pandn mm5, mm6 |
movq mm1,[_EAX+8] |
962 |
pxor mm0, mm1 ; negate if negative |
psubw mm0,mm2 |
963 |
pxor mm3, mm4 ; negate if negative |
pxor mm2,mm3 |
964 |
psubw mm0, mm1 |
pxor mm3,mm3 |
965 |
psubw mm3, mm4 |
paddw mm0,mm2 |
966 |
pmullw mm0, mm7 ; *= 2Q |
pxor mm2,mm2 |
967 |
pmullw mm3, mm7 ; *= 2Q |
pcmpgtw mm3,mm1 |
968 |
paddw mm0, mm2 ; + offset |
pcmpeqw mm2,mm1 |
969 |
paddw mm3, mm5 ; + offset |
pmullw mm1,mm7 |
970 |
paddw mm0, mm1 ; start restoring sign |
pandn mm2,mm6 |
971 |
paddw mm3, mm4 ; start restoring sign |
psubw mm1,mm3 |
972 |
; saturates to +2047 |
pxor mm3,mm2 |
973 |
movq mm2, [mmx_2047] |
paddw mm1,mm3 |
974 |
pminsw mm0, mm2 |
|
975 |
add eax, 2 |
pminsw mm0,mm4 |
976 |
pminsw mm3, mm2 |
pminsw mm1,mm4 |
977 |
|
pmaxsw mm0,mm5 |
978 |
pxor mm0, mm1 ; finish restoring sign |
pmaxsw mm1,mm5 |
979 |
pxor mm3, mm4 ; finish restoring sign |
|
980 |
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
movq [TMP1],mm0 |
981 |
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
lea _EAX,[_EAX+16] |
982 |
jnz near .loop |
movq [TMP1+8],mm1 |
983 |
|
|
984 |
|
dec TMP0 |
985 |
|
lea TMP1,[TMP1+16] |
986 |
|
jne .loop |
987 |
|
|
988 |
xor eax, eax |
xor _EAX, _EAX ; return 0 |
989 |
ret |
ret |
990 |
|
ENDFUNC |
991 |
|
|
992 |
|
|
993 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
994 |
; |
; |
999 |
; |
; |
1000 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
1001 |
|
|
1002 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
1003 |
dequant_h263_inter_sse2: |
dequant_h263_inter_sse2: |
|
mov edx, [esp + 4] ; data |
|
|
mov ecx, [esp + 8] ; coeff |
|
|
mov eax, [esp + 12] ; quant |
|
|
movq mm6, [mmx_add + eax * 8 - 8] |
|
|
movq mm7, [mmx_mul + eax * 8 - 8] |
|
|
movq2dq xmm6, mm6 |
|
|
movq2dq xmm7, mm7 |
|
|
movlhps xmm6, xmm6 |
|
|
movlhps xmm7, xmm7 |
|
|
mov eax, -16 |
|
1004 |
|
|
1005 |
ALIGN 16 |
mov TMP0, prm3 ; quant |
1006 |
.loop |
mov _EAX, prm2 ; coeff |
|
movdqa xmm0, [ecx + 8*16 + 8*eax] ; c = coeff[i] |
|
|
movdqa xmm3, [ecx + 8*16 + 8*eax + 16] |
|
1007 |
|
|
1008 |
pxor xmm1, xmm1 |
lea TMP1, [mmx_quant] |
1009 |
pxor xmm4, xmm4 |
movq xmm6, [TMP1 + TMP0*8] ; quant |
1010 |
pcmpgtw xmm1, xmm0 ; sign(c) |
inc TMP0 |
1011 |
pcmpgtw xmm4, xmm3 |
pcmpeqw xmm5,xmm5 |
1012 |
|
and TMP0,1 |
1013 |
|
movlhps xmm6,xmm6 |
1014 |
|
movd xmm0,TMP0d |
1015 |
|
movdqa xmm7,xmm6 |
1016 |
|
pshuflw xmm0,xmm0,0 |
1017 |
|
movdqa xmm4,xmm5 |
1018 |
|
mov TMP1, prm1 ; data |
1019 |
|
movlhps xmm0,xmm0 |
1020 |
|
paddw xmm7,xmm7 |
1021 |
|
psubw xmm6,xmm0 |
1022 |
|
psrlw xmm4,5 ; 2047 |
1023 |
|
mov TMP0,4 |
1024 |
|
pxor xmm5,xmm4 ; mm5=-2048 |
1025 |
|
|
1026 |
|
.loop: |
1027 |
|
movdqa xmm0,[_EAX] |
1028 |
|
pxor xmm3,xmm3 |
1029 |
pxor xmm2, xmm2 |
pxor xmm2, xmm2 |
1030 |
pxor xmm5, xmm5 |
pcmpeqw xmm3,xmm0 |
1031 |
pcmpeqw xmm2, xmm0 ; c is zero |
pcmpgtw xmm2,xmm0 |
1032 |
pcmpeqw xmm5, xmm3 |
pmullw xmm0,xmm7 ; * 2 * quant |
1033 |
|
pandn xmm3,xmm6 |
1034 |
|
movdqa xmm1,[_EAX+16] |
1035 |
|
psubw xmm0,xmm2 |
1036 |
|
pxor xmm2,xmm3 |
1037 |
|
pxor xmm3,xmm3 |
1038 |
|
paddw xmm0,xmm2 |
1039 |
|
pxor xmm2,xmm2 |
1040 |
|
pcmpgtw xmm3,xmm1 |
1041 |
|
pcmpeqw xmm2,xmm1 |
1042 |
|
pmullw xmm1,xmm7 |
1043 |
pandn xmm2, xmm6 |
pandn xmm2, xmm6 |
1044 |
pandn xmm5, xmm6 |
psubw xmm1,xmm3 |
1045 |
pxor xmm0, xmm1 ; negate if negative |
pxor xmm3,xmm2 |
1046 |
pxor xmm3, xmm4 |
paddw xmm1,xmm3 |
1047 |
psubw xmm0, xmm1 |
|
1048 |
psubw xmm3, xmm4 |
pminsw xmm0,xmm4 |
1049 |
pmullw xmm0, xmm7 ; *= 2Q |
pminsw xmm1,xmm4 |
1050 |
pmullw xmm3, xmm7 |
pmaxsw xmm0,xmm5 |
1051 |
paddw xmm0, xmm2 ; + offset |
pmaxsw xmm1,xmm5 |
1052 |
paddw xmm3, xmm5 |
|
1053 |
|
movdqa [TMP1],xmm0 |
1054 |
paddw xmm0, xmm1 ; start restoring sign |
lea _EAX,[_EAX+32] |
1055 |
paddw xmm3, xmm4 |
movdqa [TMP1+16],xmm1 |
1056 |
|
|
1057 |
; saturates to +2047 |
dec TMP0 |
1058 |
movdqa xmm2, [sse2_2047] |
lea TMP1,[TMP1+32] |
1059 |
pminsw xmm0, xmm2 |
jne .loop |
|
add eax, 4 |
|
|
pminsw xmm3, xmm2 |
|
1060 |
|
|
1061 |
pxor xmm0, xmm1 ; finish restoring sign |
xor _EAX, _EAX ; return 0 |
|
pxor xmm3, xmm4 |
|
|
movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0 |
|
|
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
|
|
jnz near .loop |
|
|
|
|
|
xor eax, eax |
|
1062 |
ret |
ret |
1063 |
|
ENDFUNC |
1064 |
|
|
1065 |
|
|
1066 |
|
%ifidn __OUTPUT_FORMAT__,elf |
1067 |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
1068 |
|
%endif |
1069 |
|
|