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 |
55 |
|
|
56 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
57 |
; |
; |
58 |
; subtract by Q/2 table |
; quant table |
59 |
; |
; |
60 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
61 |
|
|
62 |
ALIGN 16 |
ALIGN 16 |
63 |
mmx_sub: |
mmx_quant: |
64 |
%assign quant 1 |
%assign quant 0 |
65 |
%rep 31 |
%rep 32 |
66 |
times 4 dw quant / 2 |
times 4 dw quant |
67 |
%assign quant quant+1 |
%assign quant quant+1 |
68 |
%endrep |
%endrep |
69 |
|
|
70 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
71 |
; |
; |
72 |
; divide by 2Q table |
; subtract by Q/2 table |
|
; |
|
|
; use a shift of 16 to take full advantage of _pmulhw_ |
|
|
; for q=1, _pmulhw_ will overflow so it is treated seperately |
|
|
; (3dnow2 provides _pmulhuw_ which wont cause overflow) |
|
73 |
; |
; |
74 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
75 |
|
|
76 |
ALIGN 16 |
ALIGN 16 |
77 |
mmx_div: |
mmx_sub: |
78 |
%assign quant 1 |
%assign quant 1 |
79 |
%rep 31 |
%rep 31 |
80 |
times 4 dw (1<<16) / (quant*2) + 1 |
times 4 dw quant / 2 |
81 |
%assign quant quant+1 |
%assign quant quant+1 |
82 |
%endrep |
%endrep |
83 |
|
|
84 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
85 |
; |
; |
86 |
; add by (odd(Q) ? Q : Q - 1) table |
; divide by 2Q 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 |
|
|
|
|
|
;----------------------------------------------------------------------------- |
|
87 |
; |
; |
88 |
; multiple by 2Q table |
; use a shift of 16 to take full advantage of _pmulhw_ |
89 |
|
; for q=1, _pmulhw_ will overflow so it is treated seperately |
90 |
|
; (3dnow2 provides _pmulhuw_ which wont cause overflow) |
91 |
; |
; |
92 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
93 |
|
|
94 |
ALIGN 16 |
ALIGN 16 |
95 |
mmx_mul: |
mmx_div: |
96 |
%assign quant 1 |
%assign quant 1 |
97 |
%rep 31 |
%rep 31 |
98 |
times 4 dw quant*2 |
times 4 dw (1<<16) / (quant*2) + 1 |
99 |
%assign quant quant+1 |
%assign quant quant+1 |
100 |
%endrep |
%endrep |
101 |
|
|
|
;----------------------------------------------------------------------------- |
|
|
; |
|
|
; 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) |
|
|
|
|
|
|
|
102 |
;============================================================================= |
;============================================================================= |
103 |
; Code |
; Code |
104 |
;============================================================================= |
;============================================================================= |
129 |
ALIGN 16 |
ALIGN 16 |
130 |
quant_h263_intra_mmx: |
quant_h263_intra_mmx: |
131 |
|
|
|
push ecx |
|
132 |
push esi |
push esi |
|
push edi |
|
|
|
|
|
mov edi, [esp + 12 + 4] ; coeff |
|
|
mov esi, [esp + 12 + 8] ; data |
|
|
mov eax, [esp + 12 + 12] ; quant |
|
133 |
|
|
134 |
xor ecx, ecx |
mov esi, [esp + 4 + 8] ; data |
|
cmp al, 1 |
|
|
jz .q1loop |
|
135 |
|
|
136 |
movq mm7, [mmx_div + eax * 8 - 8] |
movsx eax, word [esi] ; data[0] |
137 |
|
|
138 |
ALIGN 16 |
mov ecx,[esp + 4 + 16] ; dcscalar |
139 |
.loop |
mov edx,eax |
140 |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
sar ecx,1 |
141 |
movq mm3, [esi + 8*ecx + 8] |
add eax,ecx |
142 |
pxor mm1, mm1 ; mm1 = 0 |
sub edx,ecx |
143 |
pxor mm4, mm4 ; |
cmovl eax,edx ; +/- dcscalar/2 |
144 |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
mov ecx, [esp + 4 + 12] ; quant |
145 |
pcmpgtw mm4, mm3 ; |
cdq |
146 |
pxor mm0, mm1 ; mm0 = |mm0| |
idiv dword [esp + 4 + 16] ; dcscalar |
147 |
pxor mm3, mm4 ; |
cmp ecx, 1 |
148 |
psubw mm0, mm1 ; displace |
mov edx, [esp + 4 + 4] ; coeff |
149 |
psubw mm3, mm4 ; |
je .low |
|
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
|
|
pmulhw mm3, mm7 ; |
|
|
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 |
|
150 |
|
|
151 |
add ecx, 2 |
movq mm7, [mmx_div+ecx * 8 - 8] |
152 |
cmp ecx, 16 |
mov ecx,4 |
|
jnz .loop |
|
153 |
|
|
154 |
.done |
.loop |
155 |
|
movq mm0, [esi] ; data |
156 |
|
pxor mm4,mm4 |
157 |
|
movq mm1, [esi + 8] |
158 |
|
pcmpgtw mm4,mm0 ; (data<0) |
159 |
|
pxor mm5,mm5 |
160 |
|
pmulhw mm0,mm7 ; /(2*quant) |
161 |
|
pcmpgtw mm5,mm1 |
162 |
|
movq mm2, [esi+16] |
163 |
|
psubw mm0,mm4 ; +(data<0) |
164 |
|
pmulhw mm1,mm7 |
165 |
|
pxor mm4,mm4 |
166 |
|
movq mm3,[esi+24] |
167 |
|
pcmpgtw mm4,mm2 |
168 |
|
psubw mm1,mm5 |
169 |
|
pmulhw mm2,mm7 |
170 |
|
pxor mm5,mm5 |
171 |
|
pcmpgtw mm5,mm3 |
172 |
|
pmulhw mm3,mm7 |
173 |
|
psubw mm2,mm4 |
174 |
|
psubw mm3,mm5 |
175 |
|
movq [edx], mm0 |
176 |
|
lea esi, [esi+32] |
177 |
|
movq [edx + 8], mm1 |
178 |
|
movq [edx + 16], mm2 |
179 |
|
movq [edx + 24], mm3 |
180 |
|
|
181 |
|
dec ecx |
182 |
|
lea edx, [edx+32] |
183 |
|
jne .loop |
184 |
|
jmp .end |
185 |
|
|
186 |
|
.low |
187 |
|
movd mm7,ecx |
188 |
|
mov ecx,4 |
189 |
|
.loop_low |
190 |
|
movq mm0, [esi] |
191 |
|
pxor mm4,mm4 |
192 |
|
movq mm1, [esi + 8] |
193 |
|
pcmpgtw mm4,mm0 |
194 |
|
pxor mm5,mm5 |
195 |
|
psubw mm0,mm4 |
196 |
|
pcmpgtw mm5,mm1 |
197 |
|
psraw mm0,mm7 |
198 |
|
psubw mm1,mm5 |
199 |
|
movq mm2,[esi+16] |
200 |
|
pxor mm4,mm4 |
201 |
|
psraw mm1,mm7 |
202 |
|
pcmpgtw mm4,mm2 |
203 |
|
pxor mm5,mm5 |
204 |
|
psubw mm2,mm4 |
205 |
|
movq mm3,[esi+24] |
206 |
|
pcmpgtw mm5,mm3 |
207 |
|
psraw mm2,mm7 |
208 |
|
psubw mm3,mm5 |
209 |
|
movq [edx], mm0 |
210 |
|
psraw mm3,mm7 |
211 |
|
movq [edx + 8], mm1 |
212 |
|
movq [edx+16],mm2 |
213 |
|
lea esi, [esi+32] |
214 |
|
movq [edx+24],mm3 |
215 |
|
|
216 |
|
dec ecx |
217 |
|
lea edx, [edx+32] |
218 |
|
jne .loop_low |
219 |
|
|
220 |
; caclulate data[0] // (int32_t)dcscalar) |
.end |
221 |
mov ecx, [esp + 12 + 16] ; dcscalar |
mov edx, [esp + 4 + 4] ; coeff |
222 |
mov edx, ecx |
mov [edx],ax |
223 |
movsx eax, word [esi] ; data[0] |
xor eax,eax ; return 0 |
|
shr edx, 1 ; edx = dcscalar /2 |
|
|
cmp eax, 0 |
|
|
jg .gtzero |
|
|
|
|
|
sub eax, edx |
|
|
jmp short .mul |
|
|
|
|
|
.gtzero |
|
|
add eax, edx |
|
|
.mul |
|
|
cdq ; expand eax -> edx:eax |
|
|
idiv ecx ; eax = edx:eax / dcscalar |
|
|
mov [edi], ax ; coeff[0] = ax |
|
224 |
|
|
|
xor eax, eax ; return(0); |
|
|
pop edi |
|
225 |
pop esi |
pop esi |
|
pop ecx |
|
|
|
|
226 |
ret |
ret |
227 |
|
|
|
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 |
|
|
|
|
|
|
|
228 |
|
|
229 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
230 |
; |
; |
240 |
quant_h263_intra_sse2: |
quant_h263_intra_sse2: |
241 |
|
|
242 |
push esi |
push esi |
|
push edi |
|
243 |
|
|
244 |
mov edi, [esp + 8 + 4] ; coeff |
mov esi, [esp + 4 + 8] ; data |
|
mov esi, [esp + 8 + 8] ; data |
|
|
mov eax, [esp + 8 + 12] ; quant |
|
245 |
|
|
246 |
xor ecx, ecx |
movsx eax, word [esi] ; data[0] |
247 |
cmp al, 1 |
|
248 |
jz near .qas2_q1loop |
mov ecx,[esp + 4 + 16] ; dcscalar |
249 |
|
mov edx,eax |
250 |
|
sar ecx,1 |
251 |
|
add eax,ecx |
252 |
|
sub edx,ecx |
253 |
|
cmovl eax,edx ; +/- dcscalar/2 |
254 |
|
mov ecx, [esp + 4 + 12] ; quant |
255 |
|
cdq |
256 |
|
idiv dword [esp + 4 + 16] ; dcscalar |
257 |
|
cmp ecx, 1 |
258 |
|
mov edx, [esp + 4 + 4] ; coeff |
259 |
|
movq xmm7, [mmx_div+ecx * 8 - 8] |
260 |
|
je .low |
261 |
|
|
262 |
.qas2_not1 |
mov ecx,2 |
|
movq mm7, [mmx_div + eax*8 - 8] |
|
|
movq2dq xmm7, mm7 |
|
263 |
movlhps xmm7, xmm7 |
movlhps xmm7, xmm7 |
264 |
|
|
265 |
ALIGN 16 |
.loop |
266 |
.qas2_loop |
movdqa xmm0, [esi] |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
|
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
|
|
pxor xmm1, xmm1 |
|
267 |
pxor xmm4, xmm4 |
pxor xmm4, xmm4 |
268 |
pcmpgtw xmm1, xmm0 |
movdqa xmm1, [esi + 16] |
269 |
pcmpgtw xmm4, xmm3 |
pcmpgtw xmm4,xmm0 |
270 |
pxor xmm0, xmm1 |
pxor xmm5,xmm5 |
|
pxor xmm3, xmm4 |
|
|
psubw xmm0, xmm1 |
|
|
psubw xmm3, xmm4 |
|
271 |
pmulhw xmm0, xmm7 |
pmulhw xmm0, xmm7 |
272 |
|
pcmpgtw xmm5,xmm1 |
273 |
|
movdqa xmm2, [esi+32] |
274 |
|
psubw xmm0,xmm4 |
275 |
|
pmulhw xmm1,xmm7 |
276 |
|
pxor xmm4,xmm4 |
277 |
|
movdqa xmm3,[esi+48] |
278 |
|
pcmpgtw xmm4,xmm2 |
279 |
|
psubw xmm1,xmm5 |
280 |
|
pmulhw xmm2,xmm7 |
281 |
|
pxor xmm5,xmm5 |
282 |
|
pcmpgtw xmm5,xmm3 |
283 |
pmulhw xmm3, xmm7 |
pmulhw xmm3, xmm7 |
284 |
pxor xmm0, xmm1 |
psubw xmm2,xmm4 |
285 |
pxor xmm3, xmm4 |
psubw xmm3,xmm5 |
286 |
psubw xmm0, xmm1 |
movdqa [edx], xmm0 |
287 |
psubw xmm3, xmm4 |
lea esi, [esi+64] |
288 |
movdqa [edi + ecx*8], xmm0 |
movdqa [edx + 16], xmm1 |
289 |
movdqa [edi + ecx*8 + 16], xmm3 |
movdqa [edx + 32], xmm2 |
290 |
|
movdqa [edx + 48], xmm3 |
291 |
add ecx, 4 |
|
292 |
cmp ecx, 16 |
dec ecx |
293 |
jnz .qas2_loop |
lea edx, [edx+64] |
294 |
|
jne .loop |
295 |
.qas2_done |
jmp .end |
296 |
mov ecx, [esp + 8 + 16] ; dcscalar |
|
297 |
mov edx, ecx |
.low |
298 |
movsx eax, word [esi] |
movd xmm7,ecx |
299 |
shr edx, 1 |
mov ecx,2 |
300 |
cmp eax, 0 |
.loop_low |
301 |
jg .qas2_gtzero |
movdqa xmm0, [esi] |
302 |
|
pxor xmm4,xmm4 |
303 |
sub eax, edx |
movdqa xmm1, [esi + 16] |
304 |
jmp short .qas2_mul |
pcmpgtw xmm4,xmm0 |
305 |
|
pxor xmm5,xmm5 |
306 |
.qas2_gtzero |
psubw xmm0,xmm4 |
307 |
add eax, edx |
pcmpgtw xmm5,xmm1 |
308 |
|
psraw xmm0,xmm7 |
309 |
.qas2_mul |
psubw xmm1,xmm5 |
310 |
cdq |
movdqa xmm2,[esi+32] |
311 |
idiv ecx |
pxor xmm4,xmm4 |
312 |
|
psraw xmm1,xmm7 |
313 |
|
pcmpgtw xmm4,xmm2 |
314 |
|
pxor xmm5,xmm5 |
315 |
|
psubw xmm2,xmm4 |
316 |
|
movdqa xmm3,[esi+48] |
317 |
|
pcmpgtw xmm5,xmm3 |
318 |
|
psraw xmm2,xmm7 |
319 |
|
psubw xmm3,xmm5 |
320 |
|
movdqa [edx], xmm0 |
321 |
|
psraw xmm3,xmm7 |
322 |
|
movdqa [edx+16], xmm1 |
323 |
|
movdqa [edx+32],xmm2 |
324 |
|
lea esi, [esi+64] |
325 |
|
movdqa [edx+48],xmm3 |
326 |
|
|
327 |
|
dec ecx |
328 |
|
lea edx, [edx+64] |
329 |
|
jne .loop_low |
330 |
|
|
331 |
mov [edi], ax |
.end |
332 |
|
mov edx, [esp + 4 + 4] ; coeff |
333 |
|
mov [edx],ax |
334 |
|
xor eax,eax ; return 0 |
335 |
|
|
|
xor eax, eax ; return(0); |
|
|
pop edi |
|
336 |
pop esi |
pop esi |
|
|
|
337 |
ret |
ret |
338 |
|
|
|
ALIGN 16 |
|
|
.qas2_q1loop |
|
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
|
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
|
|
pxor xmm1, xmm1 |
|
|
pxor xmm4, xmm4 |
|
|
pcmpgtw xmm1, xmm0 |
|
|
pcmpgtw xmm4, xmm3 |
|
|
pxor xmm0, xmm1 |
|
|
pxor xmm3, xmm4 |
|
|
psubw xmm0, xmm1 |
|
|
psubw xmm3, xmm4 |
|
|
psrlw xmm0, 1 |
|
|
psrlw xmm3, 1 |
|
|
pxor xmm0, xmm1 |
|
|
pxor xmm3, xmm4 |
|
|
psubw xmm0, xmm1 |
|
|
psubw xmm3, xmm4 |
|
|
movdqa [edi + ecx*8], xmm0 |
|
|
movdqa [edi + ecx*8 + 16], xmm3 |
|
|
|
|
|
add ecx, 4 |
|
|
cmp ecx, 16 |
|
|
jnz .qas2_q1loop |
|
|
jmp near .qas2_done |
|
|
|
|
|
|
|
|
|
|
339 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
340 |
; |
; |
341 |
; uint32_t quant_h263_inter_mmx(int16_t * coeff, |
; uint32_t quant_h263_inter_mmx(int16_t * coeff, |
564 |
; |
; |
565 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
566 |
|
|
|
; note: we only saturate to +2047 *before* restoring the sign. |
|
|
; Hence, final clamp really is [-2048,2047] |
|
|
|
|
567 |
ALIGN 16 |
ALIGN 16 |
568 |
dequant_h263_intra_mmx: |
dequant_h263_intra_mmx: |
569 |
|
|
570 |
|
mov ecx, [esp+12] ; quant |
571 |
|
mov eax, [esp+ 8] ; coeff |
572 |
|
pcmpeqw mm0,mm0 |
573 |
|
movq mm6, [mmx_quant + ecx*8] ; quant |
574 |
|
shl ecx,31 ; quant & 1 ? 0 : - 1 |
575 |
|
movq mm7,mm6 |
576 |
|
movq mm5,mm0 |
577 |
|
movd mm1,ecx |
578 |
mov edx, [esp+ 4] ; data |
mov edx, [esp+ 4] ; data |
579 |
mov ecx, [esp+ 8] ; coeff |
psllw mm0,mm1 |
580 |
mov eax, [esp+12] ; quant |
paddw mm7,mm7 ; 2*quant |
581 |
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
paddw mm6,mm0 ; quant-1 |
582 |
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
psllw mm5,12 |
583 |
mov eax, -16 |
mov ecx,8 |
584 |
|
psrlw mm5,1 |
585 |
|
|
586 |
ALIGN 16 |
.loop: |
587 |
.loop |
movq mm0,[eax] |
588 |
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
pxor mm2,mm2 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
|
|
pxor mm1, mm1 |
|
589 |
pxor mm4, mm4 |
pxor mm4, mm4 |
590 |
pcmpgtw mm1, mm0 ; sign(c) |
pcmpgtw mm2,mm0 |
591 |
pcmpgtw mm4, mm3 ; sign(c') |
pcmpeqw mm4,mm0 |
592 |
|
pmullw mm0,mm7 ; * 2 * quant |
593 |
|
movq mm1,[eax+8] |
594 |
|
psubw mm0,mm2 |
595 |
|
pxor mm2,mm6 |
596 |
|
pxor mm3,mm3 |
597 |
|
pandn mm4,mm2 |
598 |
pxor mm2, mm2 |
pxor mm2, mm2 |
599 |
pxor mm5, mm5 |
pcmpgtw mm3,mm1 |
600 |
pcmpeqw mm2, mm0 ; c is zero |
pcmpeqw mm2,mm1 |
601 |
pcmpeqw mm5, mm3 ; c' is zero |
pmullw mm1,mm7 |
602 |
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
paddw mm0,mm4 |
603 |
pandn mm5, mm6 |
psubw mm1,mm3 |
604 |
pxor mm0, mm1 ; negate if negative |
pxor mm3,mm6 |
605 |
pxor mm3, mm4 ; negate if negative |
pandn mm2,mm3 |
606 |
psubw mm0, mm1 |
paddsw mm0, mm5 ; saturate |
607 |
psubw mm3, mm4 |
paddw mm1,mm2 |
608 |
pmullw mm0, mm7 ; *= 2Q |
|
609 |
pmullw mm3, mm7 ; *= 2Q |
paddsw mm1, mm5 |
610 |
paddw mm0, mm2 ; + offset |
psubsw mm0, mm5 |
611 |
paddw mm3, mm5 ; + offset |
psubsw mm1, mm5 |
612 |
paddw mm0, mm1 ; negate back |
psubsw mm0, mm5 |
613 |
paddw mm3, mm4 ; negate back |
psubsw mm1, mm5 |
614 |
|
paddsw mm0, mm5 |
615 |
; saturates to +2047 |
paddsw mm1, mm5 |
616 |
movq mm2, [mmx_32767_minus_2047] |
|
617 |
add eax, 2 |
movq [edx],mm0 |
618 |
paddsw mm0, mm2 |
lea eax,[eax+16] |
619 |
paddsw mm3, mm2 |
movq [edx+8],mm1 |
620 |
psubsw mm0, mm2 |
|
621 |
psubsw mm3, mm2 |
dec ecx |
622 |
|
lea edx,[edx+16] |
623 |
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 |
|
624 |
|
|
625 |
; deal with DC |
; deal with DC |
626 |
movd mm0, [ecx] |
mov eax, [esp+ 8] ; coeff |
627 |
pmullw mm0, [esp+16] ; dcscalar |
movsx eax,word [eax] |
628 |
movq mm2, [mmx_32767_minus_2047] |
mov ecx,2047 |
629 |
paddsw mm0, mm2 |
imul dword [esp+16] ; dcscalar |
630 |
psubsw mm0, mm2 |
mov edx, [esp+ 4] ; data |
631 |
movq mm3, [mmx_32768_minus_2048] |
cmp eax,ecx |
632 |
psubsw mm0, mm3 |
cmovg eax,ecx |
633 |
paddsw mm0, mm3 |
not ecx |
634 |
movd eax, mm0 |
cmp eax,ecx |
635 |
|
cmovl eax,ecx |
636 |
|
|
637 |
mov [edx], ax |
mov [edx], ax |
638 |
|
|
639 |
xor eax, eax ; return(0); |
xor eax, eax ; return 0 |
640 |
ret |
ret |
641 |
|
|
642 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
649 |
; |
; |
650 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
651 |
|
|
|
; this is the same as dequant_inter_mmx, except that we're |
|
|
; saturating using 'pminsw' (saves 2 cycles/loop => ~5% faster) |
|
652 |
|
|
653 |
ALIGN 16 |
ALIGN 16 |
654 |
dequant_h263_intra_xmm: |
dequant_h263_intra_xmm: |
655 |
|
|
656 |
|
mov ecx, [esp+12] ; quant |
657 |
|
mov eax, [esp+ 8] ; coeff |
658 |
|
|
659 |
|
movd mm6,ecx ; quant |
660 |
|
pcmpeqw mm0,mm0 |
661 |
|
pshufw mm6,mm6,0 ; all quant |
662 |
|
shl ecx,31 |
663 |
|
movq mm5,mm0 |
664 |
|
movq mm7,mm6 |
665 |
|
movd mm1,ecx |
666 |
mov edx, [esp+ 4] ; data |
mov edx, [esp+ 4] ; data |
667 |
mov ecx, [esp+ 8] ; coeff |
psllw mm0,mm1 ; quant & 1 ? 0 : - 1 |
668 |
mov eax, [esp+12] ; quant |
movq mm4,mm5 |
669 |
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
paddw mm7,mm7 ; quant*2 |
670 |
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
paddw mm6,mm0 ; quant-1 |
671 |
mov eax, -16 |
psrlw mm4,5 ; mm4=2047 |
672 |
|
mov ecx,8 |
673 |
|
pxor mm5,mm4 ; mm5=-2048 |
674 |
|
|
675 |
ALIGN 16 |
.loop: |
676 |
.loop |
movq mm0,[eax] |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
|
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
|
|
pxor mm1, mm1 |
|
|
pxor mm4, mm4 |
|
|
pcmpgtw mm1, mm0 ; sign(c) |
|
|
pcmpgtw mm4, mm3 ; sign(c') |
|
677 |
pxor mm2, mm2 |
pxor mm2, mm2 |
678 |
pxor mm5, mm5 |
pxor mm3,mm3 |
679 |
pcmpeqw mm2, mm0 ; c is zero |
|
680 |
pcmpeqw mm5, mm3 ; c' is zero |
pcmpgtw mm2,mm0 |
681 |
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
pcmpeqw mm3,mm0 ; if coeff==0... |
682 |
pandn mm5, mm6 |
pmullw mm0,mm7 ; * 2 * quant |
683 |
pxor mm0, mm1 ; negate if negative |
movq mm1,[eax+8] |
684 |
pxor mm3, mm4 ; negate if negative |
|
685 |
psubw mm0, mm1 |
psubw mm0,mm2 |
686 |
psubw mm3, mm4 |
pxor mm2,mm6 |
687 |
pmullw mm0, mm7 ; *= 2Q |
pandn mm3,mm2 ; ...then data=0 |
688 |
pmullw mm3, mm7 ; *= 2Q |
pxor mm2,mm2 |
689 |
paddw mm0, mm2 ; + offset |
paddw mm0,mm3 |
690 |
paddw mm3, mm5 ; + offset |
pxor mm3,mm3 |
691 |
paddw mm0, mm1 ; negate back |
pcmpeqw mm2,mm1 |
692 |
paddw mm3, mm4 ; negate back |
pcmpgtw mm3,mm1 |
693 |
|
pmullw mm1,mm7 |
694 |
; saturates to +2047 |
|
695 |
movq mm2, [mmx_2047] |
pminsw mm0,mm4 |
696 |
pminsw mm0, mm2 |
psubw mm1,mm3 |
697 |
add eax, 2 |
pxor mm3,mm6 |
698 |
pminsw mm3, mm2 |
pandn mm2,mm3 |
699 |
|
paddw mm1,mm2 |
700 |
pxor mm0, mm1 |
|
701 |
pxor mm3, mm4 |
pmaxsw mm0,mm5 |
702 |
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
pminsw mm1,mm4 |
703 |
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
movq [edx],mm0 |
704 |
jnz near .loop |
pmaxsw mm1,mm5 |
705 |
|
lea eax,[eax+16] |
706 |
|
movq [edx+8],mm1 |
707 |
|
|
708 |
|
dec ecx |
709 |
|
lea edx,[edx+16] |
710 |
|
jne .loop |
711 |
|
|
712 |
; deal with DC |
; deal with DC |
713 |
movd mm0, [ecx] |
mov eax, [esp+ 8] ; coeff |
714 |
|
mov edx, [esp+ 4] ; data |
715 |
|
movd mm0, [eax] |
716 |
pmullw mm0, [esp+16] ; dcscalar |
pmullw mm0, [esp+16] ; dcscalar |
717 |
movq mm2, [mmx_32767_minus_2047] |
pminsw mm0,mm4 |
718 |
paddsw mm0, mm2 |
pmaxsw mm0,mm5 |
|
psubsw mm0, mm2 |
|
|
movq mm2, [mmx_32768_minus_2048] |
|
|
psubsw mm0, mm2 |
|
|
paddsw mm0, mm2 |
|
719 |
movd eax, mm0 |
movd eax, mm0 |
720 |
mov [edx], ax |
mov [edx], ax |
721 |
|
|
722 |
xor eax, eax |
xor eax, eax ; return 0 |
723 |
ret |
ret |
724 |
|
|
725 |
|
|
735 |
|
|
736 |
ALIGN 16 |
ALIGN 16 |
737 |
dequant_h263_intra_sse2: |
dequant_h263_intra_sse2: |
738 |
|
|
739 |
|
mov ecx, [esp+12] ; quant |
740 |
|
mov eax, [esp+ 8] ; coeff |
741 |
|
|
742 |
|
movd xmm6,ecx ; quant |
743 |
|
; shr ecx,1 |
744 |
|
shl ecx,31 |
745 |
|
pshuflw xmm6,xmm6,0 |
746 |
|
pcmpeqw xmm0,xmm0 |
747 |
|
movlhps xmm6,xmm6 ; all quant |
748 |
|
movd xmm1,ecx |
749 |
|
movdqa xmm5,xmm0 |
750 |
|
movdqa xmm7,xmm6 |
751 |
mov edx, [esp+ 4] ; data |
mov edx, [esp+ 4] ; data |
752 |
mov ecx, [esp+ 8] ; coeff |
paddw xmm7,xmm7 ; quant *2 |
753 |
mov eax, [esp+12] ; quant |
psllw xmm0,xmm1 ; quant & 1 ? 0 : - 1 |
754 |
movq mm6, [mmx_add + eax * 8 - 8] |
movdqa xmm4,xmm5 |
755 |
movq mm7, [mmx_mul + eax * 8 - 8] |
paddw xmm6,xmm0 ; quant-1 |
756 |
movq2dq xmm6, mm6 |
psrlw xmm4,5 ; 2047 |
757 |
movq2dq xmm7, mm7 |
mov ecx,4 |
758 |
movlhps xmm6, xmm6 |
pxor xmm5,xmm4 ; mm5=-2048 |
|
movlhps xmm7, xmm7 |
|
|
mov eax, -16 |
|
759 |
|
|
760 |
ALIGN 16 |
.loop: |
761 |
.loop |
movdqa xmm0,[eax] |
|
movdqa xmm0, [ecx + 8*16 + 8*eax] ; c = coeff[i] |
|
|
movdqa xmm3, [ecx + 8*16 + 8*eax+ 16] |
|
|
pxor xmm1, xmm1 |
|
|
pxor xmm4, xmm4 |
|
|
pcmpgtw xmm1, xmm0 ; sign(c) |
|
|
pcmpgtw xmm4, xmm3 |
|
762 |
pxor xmm2, xmm2 |
pxor xmm2, xmm2 |
763 |
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 |
|
764 |
|
|
765 |
pxor xmm0, xmm1 |
pcmpgtw xmm2,xmm0 |
766 |
pxor xmm3, xmm4 |
pcmpeqw xmm3,xmm0 |
767 |
movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0 |
pmullw xmm0,xmm7 ; * 2 * quant |
768 |
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
movdqa xmm1,[eax+16] |
769 |
jnz near .loop |
|
770 |
|
psubw xmm0,xmm2 |
771 |
|
pxor xmm2,xmm6 |
772 |
|
pandn xmm3,xmm2 |
773 |
|
pxor xmm2,xmm2 |
774 |
|
paddw xmm0,xmm3 |
775 |
|
pxor xmm3,xmm3 |
776 |
|
pcmpeqw xmm2,xmm1 |
777 |
|
pcmpgtw xmm3,xmm1 |
778 |
|
pmullw xmm1,xmm7 |
779 |
|
|
780 |
|
pminsw xmm0,xmm4 |
781 |
|
psubw xmm1,xmm3 |
782 |
|
pxor xmm3,xmm6 |
783 |
|
pandn xmm2,xmm3 |
784 |
|
paddw xmm1,xmm2 |
785 |
|
|
786 |
|
pmaxsw xmm0,xmm5 |
787 |
|
pminsw xmm1,xmm4 |
788 |
|
movdqa [edx],xmm0 |
789 |
|
pmaxsw xmm1,xmm5 |
790 |
|
lea eax,[eax+32] |
791 |
|
movdqa [edx+16],xmm1 |
792 |
|
|
793 |
|
dec ecx |
794 |
|
lea edx,[edx+32] |
795 |
|
jne .loop |
796 |
|
|
797 |
; deal with DC |
; deal with DC |
798 |
movd mm0, [ecx] |
|
799 |
pmullw mm0, [esp+16] ; dcscalar |
mov eax, [esp+ 8] ; coeff |
800 |
movq mm2, [mmx_32767_minus_2047] |
movsx eax,word [eax] |
801 |
paddsw mm0, mm2 |
imul dword [esp+16] ; dcscalar |
802 |
psubsw mm0, mm2 |
mov edx, [esp+ 4] ; data |
803 |
movq mm2, [mmx_32768_minus_2048] |
movd xmm0,eax |
804 |
psubsw mm0, mm2 |
pminsw xmm0,xmm4 |
805 |
paddsw mm0, mm2 |
pmaxsw xmm0,xmm5 |
806 |
movd eax, mm0 |
movd eax,xmm0 |
807 |
|
|
808 |
mov [edx], ax |
mov [edx], ax |
809 |
|
|
810 |
xor eax, eax |
xor eax, eax ; return 0 |
811 |
ret |
ret |
812 |
|
|
813 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
822 |
ALIGN 16 |
ALIGN 16 |
823 |
dequant_h263_inter_mmx: |
dequant_h263_inter_mmx: |
824 |
|
|
825 |
|
mov ecx, [esp+12] ; quant |
826 |
|
mov eax, [esp+ 8] ; coeff |
827 |
|
pcmpeqw mm0,mm0 |
828 |
|
movq mm6, [mmx_quant + ecx*8] ; quant |
829 |
|
shl ecx,31 ; odd/even |
830 |
|
movq mm7,mm6 |
831 |
|
movd mm1,ecx |
832 |
mov edx, [esp+ 4] ; data |
mov edx, [esp+ 4] ; data |
833 |
mov ecx, [esp+ 8] ; coeff |
movq mm5,mm0 |
834 |
mov eax, [esp+12] ; quant |
psllw mm0,mm1 ; quant & 1 ? 0 : - 1 |
835 |
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
paddw mm7,mm7 ; quant*2 |
836 |
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
paddw mm6,mm0 ; quant & 1 ? quant : quant - 1 |
837 |
mov eax, -16 |
psllw mm5,12 |
838 |
|
mov ecx,8 |
839 |
|
psrlw mm5,1 ; 32767-2047 (32768-2048) |
840 |
|
|
841 |
ALIGN 16 |
.loop: |
842 |
.loop |
movq mm0,[eax] |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
|
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
|
|
pxor mm1, mm1 |
|
843 |
pxor mm4, mm4 |
pxor mm4, mm4 |
|
pcmpgtw mm1, mm0 ; sign(c) |
|
|
pcmpgtw mm4, mm3 ; sign(c') |
|
844 |
pxor mm2, mm2 |
pxor mm2, mm2 |
845 |
pxor mm5, mm5 |
pcmpeqw mm4,mm0 ; if coeff==0... |
846 |
pcmpeqw mm2, mm0 ; c is zero |
pcmpgtw mm2,mm0 |
847 |
pcmpeqw mm5, mm3 ; c' is zero |
pmullw mm0,mm7 ; * 2 * quant |
848 |
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
pxor mm3,mm3 |
849 |
pandn mm5, mm6 |
psubw mm0,mm2 |
850 |
pxor mm0, mm1 ; negate if negative |
movq mm1,[eax+8] |
851 |
pxor mm3, mm4 ; negate if negative |
pxor mm2,mm6 |
852 |
psubw mm0, mm1 |
pcmpgtw mm3,mm1 |
853 |
psubw mm3, mm4 |
pandn mm4,mm2 ; ... then data==0 |
854 |
pmullw mm0, mm7 ; *= 2Q |
pmullw mm1,mm7 |
855 |
pmullw mm3, mm7 ; *= 2Q |
pxor mm2,mm2 |
856 |
paddw mm0, mm2 ; + offset |
pcmpeqw mm2,mm1 |
857 |
paddw mm3, mm5 ; + offset |
psubw mm1,mm3 |
858 |
paddw mm0, mm1 ; negate back |
pxor mm3,mm6 |
859 |
paddw mm3, mm4 ; negate back |
pandn mm2,mm3 |
860 |
; saturates to +2047 |
paddw mm0,mm4 |
861 |
movq mm2, [mmx_32767_minus_2047] |
paddw mm1,mm2 |
862 |
add eax, 2 |
|
863 |
paddsw mm0, mm2 |
paddsw mm0, mm5 ; saturate |
864 |
paddsw mm3, mm2 |
paddsw mm1, mm5 |
865 |
psubsw mm0, mm2 |
psubsw mm0, mm5 |
866 |
psubsw mm3, mm2 |
psubsw mm1, mm5 |
867 |
|
psubsw mm0, mm5 |
868 |
pxor mm0, mm1 |
psubsw mm1, mm5 |
869 |
pxor mm3, mm4 |
paddsw mm0, mm5 |
870 |
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
paddsw mm1, mm5 |
871 |
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
|
872 |
jnz near .loop |
movq [edx],mm0 |
873 |
|
lea eax,[eax+16] |
874 |
|
movq [edx+8],mm1 |
875 |
|
|
876 |
|
dec ecx |
877 |
|
lea edx,[edx+16] |
878 |
|
jne .loop |
879 |
|
|
880 |
xor eax, eax |
xor eax, eax ; return 0 |
881 |
ret |
ret |
882 |
|
|
883 |
|
|
884 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
885 |
; |
; |
886 |
; uint32_t dequant_h263_inter_xmm(int16_t * data, |
; uint32_t dequant_h263_inter_xmm(int16_t * data, |
889 |
; const uint16_t *mpeg_matrices); |
; const uint16_t *mpeg_matrices); |
890 |
; |
; |
891 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
|
|
|
|
; this is the same as dequant_inter_mmx, |
|
|
; except that we're saturating using 'pminsw' (saves 2 cycles/loop) |
|
|
|
|
892 |
ALIGN 16 |
ALIGN 16 |
893 |
dequant_h263_inter_xmm: |
dequant_h263_inter_xmm: |
894 |
|
|
895 |
|
mov ecx, [esp+12] ; quant |
896 |
|
mov eax, [esp+ 8] ; coeff |
897 |
|
pcmpeqw mm0,mm0 |
898 |
|
movq mm6, [mmx_quant + ecx*8] ; quant |
899 |
|
shl ecx,31 |
900 |
|
movq mm5,mm0 |
901 |
|
movd mm1,ecx |
902 |
|
movq mm7,mm6 |
903 |
|
psllw mm0,mm1 |
904 |
mov edx, [esp+ 4] ; data |
mov edx, [esp+ 4] ; data |
905 |
mov ecx, [esp+ 8] ; coeff |
movq mm4,mm5 |
906 |
mov eax, [esp+12] ; quant |
paddw mm7,mm7 |
907 |
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
paddw mm6,mm0 ; quant-1 |
908 |
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
|
909 |
mov eax, -16 |
psrlw mm4,5 |
910 |
|
mov ecx,8 |
911 |
ALIGN 16 |
pxor mm5,mm4 ; mm5=-2048 |
912 |
.loop |
|
913 |
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
.loop: |
914 |
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
movq mm0,[eax] |
915 |
pxor mm1, mm1 |
pxor mm3,mm3 |
|
pxor mm4, mm4 |
|
|
pcmpgtw mm1, mm0 ; sign(c) |
|
|
pcmpgtw mm4, mm3 ; sign(c') |
|
916 |
pxor mm2, mm2 |
pxor mm2, mm2 |
917 |
pxor mm5, mm5 |
pcmpeqw mm3,mm0 |
918 |
pcmpeqw mm2, mm0 ; c is zero |
pcmpgtw mm2,mm0 |
919 |
pcmpeqw mm5, mm3 ; c' is zero |
pmullw mm0,mm7 ; * 2 * quant |
920 |
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
pandn mm3,mm6 |
921 |
pandn mm5, mm6 |
movq mm1,[eax+8] |
922 |
pxor mm0, mm1 ; negate if negative |
psubw mm0,mm2 |
923 |
pxor mm3, mm4 ; negate if negative |
pxor mm2,mm3 |
924 |
psubw mm0, mm1 |
pxor mm3,mm3 |
925 |
psubw mm3, mm4 |
paddw mm0,mm2 |
926 |
pmullw mm0, mm7 ; *= 2Q |
pxor mm2,mm2 |
927 |
pmullw mm3, mm7 ; *= 2Q |
pcmpgtw mm3,mm1 |
928 |
paddw mm0, mm2 ; + offset |
pcmpeqw mm2,mm1 |
929 |
paddw mm3, mm5 ; + offset |
pmullw mm1,mm7 |
930 |
paddw mm0, mm1 ; start restoring sign |
pandn mm2,mm6 |
931 |
paddw mm3, mm4 ; start restoring sign |
psubw mm1,mm3 |
932 |
; saturates to +2047 |
pxor mm3,mm2 |
933 |
movq mm2, [mmx_2047] |
paddw mm1,mm3 |
934 |
pminsw mm0, mm2 |
|
935 |
add eax, 2 |
pminsw mm0,mm4 |
936 |
pminsw mm3, mm2 |
pminsw mm1,mm4 |
937 |
|
pmaxsw mm0,mm5 |
938 |
pxor mm0, mm1 ; finish restoring sign |
pmaxsw mm1,mm5 |
939 |
pxor mm3, mm4 ; finish restoring sign |
|
940 |
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
movq [edx],mm0 |
941 |
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
lea eax,[eax+16] |
942 |
jnz near .loop |
movq [edx+8],mm1 |
943 |
|
|
944 |
|
dec ecx |
945 |
|
lea edx,[edx+16] |
946 |
|
jne .loop |
947 |
|
|
948 |
xor eax, eax |
xor eax, eax ; return 0 |
949 |
ret |
ret |
950 |
|
|
951 |
|
|
952 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
953 |
; |
; |
954 |
; uint32_t dequant_h263_inter_sse2(int16_t * data, |
; uint32_t dequant_h263_inter_sse2(int16_t * data, |
960 |
|
|
961 |
ALIGN 16 |
ALIGN 16 |
962 |
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 |
|
963 |
|
|
964 |
ALIGN 16 |
mov ecx, [esp+12] ; quant |
965 |
.loop |
mov eax, [esp+ 8] ; coeff |
|
movdqa xmm0, [ecx + 8*16 + 8*eax] ; c = coeff[i] |
|
|
movdqa xmm3, [ecx + 8*16 + 8*eax + 16] |
|
966 |
|
|
967 |
pxor xmm1, xmm1 |
movq xmm6, [mmx_quant + ecx*8] ; quant |
968 |
pxor xmm4, xmm4 |
inc ecx |
969 |
pcmpgtw xmm1, xmm0 ; sign(c) |
pcmpeqw xmm5,xmm5 |
970 |
pcmpgtw xmm4, xmm3 |
and ecx,1 |
971 |
|
movlhps xmm6,xmm6 |
972 |
|
movd xmm0,ecx |
973 |
|
movdqa xmm7,xmm6 |
974 |
|
pshuflw xmm0,xmm0,0 |
975 |
|
movdqa xmm4,xmm5 |
976 |
|
mov edx, [esp+ 4] ; data |
977 |
|
movlhps xmm0,xmm0 |
978 |
|
paddw xmm7,xmm7 |
979 |
|
psubw xmm6,xmm0 |
980 |
|
psrlw xmm4,5 ; 2047 |
981 |
|
mov ecx,4 |
982 |
|
pxor xmm5,xmm4 ; mm5=-2048 |
983 |
|
|
984 |
|
.loop: |
985 |
|
movdqa xmm0,[eax] |
986 |
|
pxor xmm3,xmm3 |
987 |
pxor xmm2, xmm2 |
pxor xmm2, xmm2 |
988 |
pxor xmm5, xmm5 |
pcmpeqw xmm3,xmm0 |
989 |
pcmpeqw xmm2, xmm0 ; c is zero |
pcmpgtw xmm2,xmm0 |
990 |
pcmpeqw xmm5, xmm3 |
pmullw xmm0,xmm7 ; * 2 * quant |
991 |
|
pandn xmm3,xmm6 |
992 |
|
movdqa xmm1,[eax+16] |
993 |
|
psubw xmm0,xmm2 |
994 |
|
pxor xmm2,xmm3 |
995 |
|
pxor xmm3,xmm3 |
996 |
|
paddw xmm0,xmm2 |
997 |
|
pxor xmm2,xmm2 |
998 |
|
pcmpgtw xmm3,xmm1 |
999 |
|
pcmpeqw xmm2,xmm1 |
1000 |
|
pmullw xmm1,xmm7 |
1001 |
pandn xmm2, xmm6 |
pandn xmm2, xmm6 |
1002 |
pandn xmm5, xmm6 |
psubw xmm1,xmm3 |
1003 |
pxor xmm0, xmm1 ; negate if negative |
pxor xmm3,xmm2 |
1004 |
pxor xmm3, xmm4 |
paddw xmm1,xmm3 |
1005 |
psubw xmm0, xmm1 |
|
1006 |
psubw xmm3, xmm4 |
pminsw xmm0,xmm4 |
1007 |
pmullw xmm0, xmm7 ; *= 2Q |
pminsw xmm1,xmm4 |
1008 |
pmullw xmm3, xmm7 |
pmaxsw xmm0,xmm5 |
1009 |
paddw xmm0, xmm2 ; + offset |
pmaxsw xmm1,xmm5 |
1010 |
paddw xmm3, xmm5 |
|
1011 |
|
movdqa [edx],xmm0 |
1012 |
paddw xmm0, xmm1 ; start restoring sign |
lea eax,[eax+32] |
1013 |
paddw xmm3, xmm4 |
movdqa [edx+16],xmm1 |
1014 |
|
|
1015 |
; saturates to +2047 |
dec ecx |
1016 |
movdqa xmm2, [sse2_2047] |
lea edx,[edx+32] |
1017 |
pminsw xmm0, xmm2 |
jne .loop |
|
add eax, 4 |
|
|
pminsw xmm3, xmm2 |
|
|
|
|
|
pxor xmm0, xmm1 ; finish restoring sign |
|
|
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 |
|
1018 |
|
|
1019 |
xor eax, eax |
xor eax, eax ; return 0 |
1020 |
ret |
ret |
1021 |
|
|