|
1 ; |
|
2 ; Copyright (c) 2010 The WebM project authors. All Rights Reserved. |
|
3 ; |
|
4 ; Use of this source code is governed by a BSD-style license |
|
5 ; that can be found in the LICENSE file in the root of the source |
|
6 ; tree. An additional intellectual property rights grant can be found |
|
7 ; in the file PATENTS. All contributing project authors may |
|
8 ; be found in the AUTHORS file in the root of the source tree. |
|
9 ; |
|
10 |
|
11 |
|
12 %include "vpx_ports/x86_abi_support.asm" |
|
13 |
|
14 |
|
15 ;void vp8_dequantize_b_impl_mmx(short *sq, short *dq, short *q) |
|
16 global sym(vp8_dequantize_b_impl_mmx) PRIVATE |
|
17 sym(vp8_dequantize_b_impl_mmx): |
|
18 push rbp |
|
19 mov rbp, rsp |
|
20 SHADOW_ARGS_TO_STACK 3 |
|
21 push rsi |
|
22 push rdi |
|
23 ; end prolog |
|
24 |
|
25 mov rsi, arg(0) ;sq |
|
26 mov rdi, arg(1) ;dq |
|
27 mov rax, arg(2) ;q |
|
28 |
|
29 movq mm1, [rsi] |
|
30 pmullw mm1, [rax+0] ; mm4 *= kernel 0 modifiers. |
|
31 movq [rdi], mm1 |
|
32 |
|
33 movq mm1, [rsi+8] |
|
34 pmullw mm1, [rax+8] ; mm4 *= kernel 0 modifiers. |
|
35 movq [rdi+8], mm1 |
|
36 |
|
37 movq mm1, [rsi+16] |
|
38 pmullw mm1, [rax+16] ; mm4 *= kernel 0 modifiers. |
|
39 movq [rdi+16], mm1 |
|
40 |
|
41 movq mm1, [rsi+24] |
|
42 pmullw mm1, [rax+24] ; mm4 *= kernel 0 modifiers. |
|
43 movq [rdi+24], mm1 |
|
44 |
|
45 ; begin epilog |
|
46 pop rdi |
|
47 pop rsi |
|
48 UNSHADOW_ARGS |
|
49 pop rbp |
|
50 ret |
|
51 |
|
52 |
|
53 ;void dequant_idct_add_mmx( |
|
54 ;short *input, 0 |
|
55 ;short *dq, 1 |
|
56 ;unsigned char *dest, 2 |
|
57 ;int stride) 3 |
|
58 global sym(vp8_dequant_idct_add_mmx) PRIVATE |
|
59 sym(vp8_dequant_idct_add_mmx): |
|
60 push rbp |
|
61 mov rbp, rsp |
|
62 SHADOW_ARGS_TO_STACK 4 |
|
63 GET_GOT rbx |
|
64 push rdi |
|
65 ; end prolog |
|
66 |
|
67 mov rax, arg(0) ;input |
|
68 mov rdx, arg(1) ;dq |
|
69 |
|
70 |
|
71 movq mm0, [rax ] |
|
72 pmullw mm0, [rdx] |
|
73 |
|
74 movq mm1, [rax +8] |
|
75 pmullw mm1, [rdx +8] |
|
76 |
|
77 movq mm2, [rax+16] |
|
78 pmullw mm2, [rdx+16] |
|
79 |
|
80 movq mm3, [rax+24] |
|
81 pmullw mm3, [rdx+24] |
|
82 |
|
83 mov rdx, arg(2) ;dest |
|
84 |
|
85 pxor mm7, mm7 |
|
86 |
|
87 |
|
88 movq [rax], mm7 |
|
89 movq [rax+8], mm7 |
|
90 |
|
91 movq [rax+16],mm7 |
|
92 movq [rax+24],mm7 |
|
93 |
|
94 |
|
95 movsxd rdi, dword ptr arg(3) ;stride |
|
96 |
|
97 psubw mm0, mm2 ; b1= 0-2 |
|
98 paddw mm2, mm2 ; |
|
99 |
|
100 movq mm5, mm1 |
|
101 paddw mm2, mm0 ; a1 =0+2 |
|
102 |
|
103 pmulhw mm5, [GLOBAL(x_s1sqr2)]; |
|
104 paddw mm5, mm1 ; ip1 * sin(pi/8) * sqrt(2) |
|
105 |
|
106 movq mm7, mm3 ; |
|
107 pmulhw mm7, [GLOBAL(x_c1sqr2less1)]; |
|
108 |
|
109 paddw mm7, mm3 ; ip3 * cos(pi/8) * sqrt(2) |
|
110 psubw mm7, mm5 ; c1 |
|
111 |
|
112 movq mm5, mm1 |
|
113 movq mm4, mm3 |
|
114 |
|
115 pmulhw mm5, [GLOBAL(x_c1sqr2less1)] |
|
116 paddw mm5, mm1 |
|
117 |
|
118 pmulhw mm3, [GLOBAL(x_s1sqr2)] |
|
119 paddw mm3, mm4 |
|
120 |
|
121 paddw mm3, mm5 ; d1 |
|
122 movq mm6, mm2 ; a1 |
|
123 |
|
124 movq mm4, mm0 ; b1 |
|
125 paddw mm2, mm3 ;0 |
|
126 |
|
127 paddw mm4, mm7 ;1 |
|
128 psubw mm0, mm7 ;2 |
|
129 |
|
130 psubw mm6, mm3 ;3 |
|
131 |
|
132 movq mm1, mm2 ; 03 02 01 00 |
|
133 movq mm3, mm4 ; 23 22 21 20 |
|
134 |
|
135 punpcklwd mm1, mm0 ; 11 01 10 00 |
|
136 punpckhwd mm2, mm0 ; 13 03 12 02 |
|
137 |
|
138 punpcklwd mm3, mm6 ; 31 21 30 20 |
|
139 punpckhwd mm4, mm6 ; 33 23 32 22 |
|
140 |
|
141 movq mm0, mm1 ; 11 01 10 00 |
|
142 movq mm5, mm2 ; 13 03 12 02 |
|
143 |
|
144 punpckldq mm0, mm3 ; 30 20 10 00 |
|
145 punpckhdq mm1, mm3 ; 31 21 11 01 |
|
146 |
|
147 punpckldq mm2, mm4 ; 32 22 12 02 |
|
148 punpckhdq mm5, mm4 ; 33 23 13 03 |
|
149 |
|
150 movq mm3, mm5 ; 33 23 13 03 |
|
151 |
|
152 psubw mm0, mm2 ; b1= 0-2 |
|
153 paddw mm2, mm2 ; |
|
154 |
|
155 movq mm5, mm1 |
|
156 paddw mm2, mm0 ; a1 =0+2 |
|
157 |
|
158 pmulhw mm5, [GLOBAL(x_s1sqr2)]; |
|
159 paddw mm5, mm1 ; ip1 * sin(pi/8) * sqrt(2) |
|
160 |
|
161 movq mm7, mm3 ; |
|
162 pmulhw mm7, [GLOBAL(x_c1sqr2less1)]; |
|
163 |
|
164 paddw mm7, mm3 ; ip3 * cos(pi/8) * sqrt(2) |
|
165 psubw mm7, mm5 ; c1 |
|
166 |
|
167 movq mm5, mm1 |
|
168 movq mm4, mm3 |
|
169 |
|
170 pmulhw mm5, [GLOBAL(x_c1sqr2less1)] |
|
171 paddw mm5, mm1 |
|
172 |
|
173 pmulhw mm3, [GLOBAL(x_s1sqr2)] |
|
174 paddw mm3, mm4 |
|
175 |
|
176 paddw mm3, mm5 ; d1 |
|
177 paddw mm0, [GLOBAL(fours)] |
|
178 |
|
179 paddw mm2, [GLOBAL(fours)] |
|
180 movq mm6, mm2 ; a1 |
|
181 |
|
182 movq mm4, mm0 ; b1 |
|
183 paddw mm2, mm3 ;0 |
|
184 |
|
185 paddw mm4, mm7 ;1 |
|
186 psubw mm0, mm7 ;2 |
|
187 |
|
188 psubw mm6, mm3 ;3 |
|
189 psraw mm2, 3 |
|
190 |
|
191 psraw mm0, 3 |
|
192 psraw mm4, 3 |
|
193 |
|
194 psraw mm6, 3 |
|
195 |
|
196 movq mm1, mm2 ; 03 02 01 00 |
|
197 movq mm3, mm4 ; 23 22 21 20 |
|
198 |
|
199 punpcklwd mm1, mm0 ; 11 01 10 00 |
|
200 punpckhwd mm2, mm0 ; 13 03 12 02 |
|
201 |
|
202 punpcklwd mm3, mm6 ; 31 21 30 20 |
|
203 punpckhwd mm4, mm6 ; 33 23 32 22 |
|
204 |
|
205 movq mm0, mm1 ; 11 01 10 00 |
|
206 movq mm5, mm2 ; 13 03 12 02 |
|
207 |
|
208 punpckldq mm0, mm3 ; 30 20 10 00 |
|
209 punpckhdq mm1, mm3 ; 31 21 11 01 |
|
210 |
|
211 punpckldq mm2, mm4 ; 32 22 12 02 |
|
212 punpckhdq mm5, mm4 ; 33 23 13 03 |
|
213 |
|
214 pxor mm7, mm7 |
|
215 |
|
216 movd mm4, [rdx] |
|
217 punpcklbw mm4, mm7 |
|
218 paddsw mm0, mm4 |
|
219 packuswb mm0, mm7 |
|
220 movd [rdx], mm0 |
|
221 |
|
222 movd mm4, [rdx+rdi] |
|
223 punpcklbw mm4, mm7 |
|
224 paddsw mm1, mm4 |
|
225 packuswb mm1, mm7 |
|
226 movd [rdx+rdi], mm1 |
|
227 |
|
228 movd mm4, [rdx+2*rdi] |
|
229 punpcklbw mm4, mm7 |
|
230 paddsw mm2, mm4 |
|
231 packuswb mm2, mm7 |
|
232 movd [rdx+rdi*2], mm2 |
|
233 |
|
234 add rdx, rdi |
|
235 |
|
236 movd mm4, [rdx+2*rdi] |
|
237 punpcklbw mm4, mm7 |
|
238 paddsw mm5, mm4 |
|
239 packuswb mm5, mm7 |
|
240 movd [rdx+rdi*2], mm5 |
|
241 |
|
242 ; begin epilog |
|
243 pop rdi |
|
244 RESTORE_GOT |
|
245 UNSHADOW_ARGS |
|
246 pop rbp |
|
247 ret |
|
248 |
|
249 SECTION_RODATA |
|
250 align 16 |
|
251 x_s1sqr2: |
|
252 times 4 dw 0x8A8C |
|
253 align 16 |
|
254 x_c1sqr2less1: |
|
255 times 4 dw 0x4E7B |
|
256 align 16 |
|
257 fours: |
|
258 times 4 dw 0x0004 |