3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * - CPUID check processors capabilities - |
; * - CPUID check processors capabilities - |
5 |
; * |
; * |
6 |
; * Copyright (C) 2001 Michael Militzer <isibaar@xvid.org> |
; * Copyright (C) 2001-2008 Michael Militzer <michael@xvid.org> |
7 |
; * |
; * |
8 |
; * This program is free software ; you can redistribute it and/or modify |
; * This program is free software ; you can redistribute it and/or modify |
9 |
; * 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 |
23 |
; * |
; * |
24 |
; ***************************************************************************/ |
; ***************************************************************************/ |
25 |
|
|
26 |
BITS 32 |
%include "nasm.inc" |
|
|
|
|
%macro cglobal 1 |
|
|
%ifdef PREFIX |
|
|
global _%1 |
|
|
%define %1 _%1 |
|
|
%else |
|
|
global %1 |
|
|
%endif |
|
|
%endmacro |
|
27 |
|
|
28 |
;============================================================================= |
;============================================================================= |
29 |
; Constants |
; Constants |
33 |
%define CPUID_MMX 0x00800000 |
%define CPUID_MMX 0x00800000 |
34 |
%define CPUID_SSE 0x02000000 |
%define CPUID_SSE 0x02000000 |
35 |
%define CPUID_SSE2 0x04000000 |
%define CPUID_SSE2 0x04000000 |
36 |
|
%define CPUID_SSE3 0x00000001 |
37 |
|
%define CPUID_SSE41 0x00080000 |
38 |
|
|
39 |
%define EXT_CPUID_3DNOW 0x80000000 |
%define EXT_CPUID_3DNOW 0x80000000 |
40 |
%define EXT_CPUID_AMD_3DNOWEXT 0x40000000 |
%define EXT_CPUID_AMD_3DNOWEXT 0x40000000 |
45 |
%define XVID_CPU_MMXEXT (1<< 1) |
%define XVID_CPU_MMXEXT (1<< 1) |
46 |
%define XVID_CPU_SSE (1<< 2) |
%define XVID_CPU_SSE (1<< 2) |
47 |
%define XVID_CPU_SSE2 (1<< 3) |
%define XVID_CPU_SSE2 (1<< 3) |
48 |
|
%define XVID_CPU_SSE3 (1<< 8) |
49 |
|
%define XVID_CPU_SSE41 (1<< 9) |
50 |
%define XVID_CPU_3DNOW (1<< 4) |
%define XVID_CPU_3DNOW (1<< 4) |
51 |
%define XVID_CPU_3DNOWEXT (1<< 5) |
%define XVID_CPU_3DNOWEXT (1<< 5) |
52 |
%define XVID_CPU_TSC (1<< 6) |
%define XVID_CPU_TSC (1<< 6) |
55 |
; Read only data |
; Read only data |
56 |
;============================================================================= |
;============================================================================= |
57 |
|
|
58 |
ALIGN 32 |
ALIGN SECTION_ALIGN |
59 |
%ifdef FORMAT_COFF |
|
60 |
SECTION .rodata data |
DATA |
|
%else |
|
|
SECTION .rodata data align=16 |
|
|
%endif |
|
61 |
|
|
62 |
vendorAMD: |
vendorAMD: |
63 |
db "AuthenticAMD" |
db "AuthenticAMD" |
66 |
; Macros |
; Macros |
67 |
;============================================================================= |
;============================================================================= |
68 |
|
|
69 |
%macro CHECK_FEATURE 3 |
%macro CHECK_FEATURE 4 |
70 |
mov ecx, %1 |
mov eax, %1 |
71 |
and ecx, edx |
and eax, %4 |
72 |
neg ecx |
neg eax |
73 |
sbb ecx, ecx |
sbb eax, eax |
74 |
and ecx, %2 |
and eax, %2 |
75 |
or %3, ecx |
or %3, eax |
76 |
%endmacro |
%endmacro |
77 |
|
|
78 |
;============================================================================= |
;============================================================================= |
79 |
; Code |
; Code |
80 |
;============================================================================= |
;============================================================================= |
81 |
|
|
82 |
SECTION .text |
%ifdef ARCH_IS_X86_64 |
83 |
|
%define XVID_PUSHFD pushfq |
84 |
|
%define XVID_POPFD popfq |
85 |
|
%else |
86 |
|
%define XVID_PUSHFD pushfd |
87 |
|
%define XVID_POPFD popfd |
88 |
|
%endif |
89 |
|
|
90 |
|
SECTION .rotext align=SECTION_ALIGN |
91 |
|
|
92 |
; int check_cpu_feature(void) |
; int check_cpu_feature(void) |
93 |
|
|
94 |
cglobal check_cpu_features |
cglobal check_cpu_features |
95 |
check_cpu_features: |
check_cpu_features: |
96 |
|
|
97 |
push ebx |
push _EBX |
98 |
push esi |
push _ESI |
99 |
push edi |
push _EDI |
100 |
push ebp |
push _EBP |
101 |
|
|
102 |
sub esp, 12 ; Stack space for vendor name |
sub _ESP, 12 ; Stack space for vendor name |
103 |
|
|
104 |
xor ebp, ebp |
xor ebp, ebp |
105 |
|
|
106 |
; CPUID command ? |
; CPUID command ? |
107 |
pushfd |
XVID_PUSHFD |
108 |
pop eax |
pop _EAX |
109 |
mov ecx, eax |
mov ecx, eax |
110 |
xor eax, 0x200000 |
xor eax, 0x200000 |
111 |
push eax |
push _EAX |
112 |
popfd |
XVID_POPFD |
113 |
pushfd |
XVID_PUSHFD |
114 |
pop eax |
pop _EAX |
115 |
cmp eax, ecx |
cmp eax, ecx |
116 |
|
|
117 |
jz near .cpu_quit ; no CPUID command -> exit |
jz near .cpu_quit ; no CPUID command -> exit |
118 |
|
|
|
|
|
119 |
; get vendor string, used later |
; get vendor string, used later |
120 |
xor eax, eax |
xor eax, eax |
121 |
cpuid |
cpuid |
122 |
mov [esp], ebx ; vendor string |
mov [_ESP], ebx ; vendor string |
123 |
mov [esp+4], edx |
mov [_ESP+4], edx |
124 |
mov [esp+8], ecx |
mov [_ESP+8], ecx |
125 |
test eax, eax |
test eax, eax |
126 |
|
|
127 |
jz near .cpu_quit |
jz near .cpu_quit |
130 |
cpuid |
cpuid |
131 |
|
|
132 |
; RDTSC command ? |
; RDTSC command ? |
133 |
CHECK_FEATURE CPUID_TSC, XVID_CPU_TSC, ebp |
CHECK_FEATURE CPUID_TSC, XVID_CPU_TSC, ebp, edx |
134 |
|
|
135 |
; MMX support ? |
; MMX support ? |
136 |
CHECK_FEATURE CPUID_MMX, XVID_CPU_MMX, ebp |
CHECK_FEATURE CPUID_MMX, XVID_CPU_MMX, ebp, edx |
137 |
|
|
138 |
; SSE support ? |
; SSE support ? |
139 |
CHECK_FEATURE CPUID_SSE, (XVID_CPU_MMXEXT|XVID_CPU_SSE), ebp |
CHECK_FEATURE CPUID_SSE, (XVID_CPU_MMXEXT|XVID_CPU_SSE), ebp, edx |
140 |
|
|
141 |
; SSE2 support? |
; SSE2 support? |
142 |
CHECK_FEATURE CPUID_SSE2, XVID_CPU_SSE2, ebp |
CHECK_FEATURE CPUID_SSE2, XVID_CPU_SSE2, ebp, edx |
143 |
|
|
144 |
|
; SSE3 support? |
145 |
|
CHECK_FEATURE CPUID_SSE3, XVID_CPU_SSE3, ebp, ecx |
146 |
|
|
147 |
|
; SSE41 support? |
148 |
|
CHECK_FEATURE CPUID_SSE41, XVID_CPU_SSE41, ebp, ecx |
149 |
|
|
150 |
; extended functions? |
; extended functions? |
151 |
mov eax, 0x80000000 |
mov eax, 0x80000000 |
157 |
cpuid |
cpuid |
158 |
|
|
159 |
; AMD cpu ? |
; AMD cpu ? |
160 |
lea esi, [vendorAMD] |
lea _ESI, [vendorAMD] |
161 |
lea edi, [esp] |
lea _EDI, [_ESP] |
162 |
mov ecx, 12 |
mov ecx, 12 |
163 |
cld |
cld |
164 |
repe cmpsb |
repe cmpsb |
165 |
jnz .cpu_quit |
jnz .cpu_quit |
166 |
|
|
167 |
; 3DNow! support ? |
; 3DNow! support ? |
168 |
CHECK_FEATURE EXT_CPUID_3DNOW, XVID_CPU_3DNOW, ebp |
CHECK_FEATURE EXT_CPUID_3DNOW, XVID_CPU_3DNOW, ebp, edx |
169 |
|
|
170 |
; 3DNOW extended ? |
; 3DNOW extended ? |
171 |
CHECK_FEATURE EXT_CPUID_AMD_3DNOWEXT, XVID_CPU_3DNOWEXT, ebp |
CHECK_FEATURE EXT_CPUID_AMD_3DNOWEXT, XVID_CPU_3DNOWEXT, ebp, edx |
172 |
|
|
173 |
; extended MMX ? |
; extended MMX ? |
174 |
CHECK_FEATURE EXT_CPUID_AMD_MMXEXT, XVID_CPU_MMXEXT, ebp |
CHECK_FEATURE EXT_CPUID_AMD_MMXEXT, XVID_CPU_MMXEXT, ebp, edx |
175 |
|
|
176 |
.cpu_quit: |
.cpu_quit: |
177 |
|
|
178 |
mov eax, ebp |
mov eax, ebp |
179 |
|
|
180 |
add esp, 12 |
add _ESP, 12 |
181 |
|
|
182 |
pop ebp |
pop _EBP |
183 |
pop edi |
pop _EDI |
184 |
pop esi |
pop _ESI |
185 |
pop ebx |
pop _EBX |
186 |
|
|
187 |
ret |
ret |
188 |
|
ENDFUNC |
189 |
|
|
190 |
; sse/sse2 operating support detection routines |
; sse/sse2 operating support detection routines |
191 |
; these will trigger an invalid instruction signal if not supported. |
; these will trigger an invalid instruction signal if not supported. |
192 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
193 |
cglobal sse_os_trigger |
cglobal sse_os_trigger |
194 |
sse_os_trigger: |
sse_os_trigger: |
195 |
xorps xmm0, xmm0 |
xorps xmm0, xmm0 |
196 |
ret |
ret |
197 |
|
ENDFUNC |
198 |
|
|
199 |
|
|
200 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
201 |
cglobal sse2_os_trigger |
cglobal sse2_os_trigger |
202 |
sse2_os_trigger: |
sse2_os_trigger: |
203 |
xorpd xmm0, xmm0 |
xorpd xmm0, xmm0 |
204 |
ret |
ret |
205 |
|
ENDFUNC |
206 |
|
|
207 |
|
|
208 |
; enter/exit mmx state |
; enter/exit mmx state |
209 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
210 |
cglobal emms_mmx |
cglobal emms_mmx |
211 |
emms_mmx: |
emms_mmx: |
212 |
emms |
emms |
213 |
ret |
ret |
214 |
|
ENDFUNC |
215 |
|
|
216 |
; faster enter/exit mmx state |
; faster enter/exit mmx state |
217 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
218 |
cglobal emms_3dn |
cglobal emms_3dn |
219 |
emms_3dn: |
emms_3dn: |
220 |
femms |
femms |
221 |
ret |
ret |
222 |
|
ENDFUNC |
223 |
|
|
224 |
|
|
225 |
|
|
226 |
|
%ifidn __OUTPUT_FORMAT__,elf |
227 |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
228 |
|
%endif |
229 |
|
|