1 /*
2  *  Copyright 2011 The LibYuv 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 #include "libyuv/cpu_id.h"
12 
13 #ifdef _MSC_VER
14 #include <intrin.h>  // For __cpuid()
15 #endif
16 #if !defined(__CLR_VER) && defined(_M_X64) && \
17     defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
18 #include <immintrin.h>  // For _xgetbv()
19 #endif
20 
21 #include <stdlib.h>  // For getenv()
22 
23 // For ArmCpuCaps() but unittested on all platforms
24 #include <stdio.h>
25 #include <string.h>
26 
27 #include "libyuv/basic_types.h"  // For CPU_X86
28 
29 // TODO(fbarchard): Use cpuid.h when gcc 4.4 is used on OSX and Linux.
30 #if (defined(__pic__) || defined(__APPLE__)) && defined(__i386__)
__cpuid(int cpu_info[4],int info_type)31 static __inline void __cpuid(int cpu_info[4], int info_type) {
32   asm volatile (  // NOLINT
33     "mov %%ebx, %%edi                          \n"
34     "cpuid                                     \n"
35     "xchg %%edi, %%ebx                         \n"
36     : "=a"(cpu_info[0]), "=D"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
37     : "a"(info_type));
38 }
39 #elif defined(__i386__) || defined(__x86_64__)
__cpuid(int cpu_info[4],int info_type)40 static __inline void __cpuid(int cpu_info[4], int info_type) {
41   asm volatile (  // NOLINT
42     "cpuid                                     \n"
43     : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
44     : "a"(info_type));
45 }
46 #endif
47 
48 #ifdef __cplusplus
49 namespace libyuv {
50 extern "C" {
51 #endif
52 
53 // Low level cpuid for X86. Returns zeros on other CPUs.
54 #if !defined(__CLR_VER) && (defined(_M_IX86) || defined(_M_X64) || \
55     defined(__i386__) || defined(__x86_64__))
56 LIBYUV_API
CpuId(int cpu_info[4],int info_type)57 void CpuId(int cpu_info[4], int info_type) {
58   __cpuid(cpu_info, info_type);
59 }
60 #else
61 LIBYUV_API
62 void CpuId(int cpu_info[4], int) {
63   cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0;
64 }
65 #endif
66 
67 // X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
68 #if !defined(__CLR_VER) && defined(_M_X64) && \
69     defined(_MSC_VER) && (_MSC_FULL_VER >= 160040219)
70 #define HAS_XGETBV
XGetBV(unsigned int xcr)71 static uint32 XGetBV(unsigned int xcr) {
72   return static_cast<uint32>(_xgetbv(xcr));
73 }
74 #elif !defined(__CLR_VER) && defined(_M_IX86)
75 #define HAS_XGETBV
76 __declspec(naked) __declspec(align(16))
XGetBV(unsigned int xcr)77 static uint32 XGetBV(unsigned int xcr) {
78   __asm {
79     mov        ecx, [esp + 4]    // xcr
80     _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0  // xgetbv for vs2005.
81     ret
82   }
83 }
84 #elif defined(__i386__) || defined(__x86_64__)
85 #define HAS_XGETBV
XGetBV(unsigned int xcr)86 static uint32 XGetBV(unsigned int xcr) {
87   uint32 xcr_feature_mask;
88   asm volatile (  // NOLINT
89     ".byte 0x0f, 0x01, 0xd0\n"
90     : "=a"(xcr_feature_mask)
91     : "c"(xcr)
92     : "memory", "cc", "edx");  // edx unused.
93   return xcr_feature_mask;
94 }
95 #endif
96 #ifdef HAS_XGETBV
97 static const int kXCR_XFEATURE_ENABLED_MASK = 0;
98 #endif
99 
100 // based on libvpx arm_cpudetect.c
101 // For Arm, but public to allow testing on any CPU
102 LIBYUV_API
ArmCpuCaps(const char * cpuinfo_name)103 int ArmCpuCaps(const char* cpuinfo_name) {
104   int flags = 0;
105   FILE* fin = fopen(cpuinfo_name, "r");
106   if (fin) {
107     char buf[512];
108     while (fgets(buf, 511, fin)) {
109       if (memcmp(buf, "Features", 8) == 0) {
110         flags |= kCpuInitialized;
111         char* p = strstr(buf, " neon");
112         if (p && (p[5] == ' ' || p[5] == '\n')) {
113           flags |= kCpuHasNEON;
114           break;
115         }
116       }
117     }
118     fclose(fin);
119   }
120   return flags;
121 }
122 
123 // CPU detect function for SIMD instruction sets.
124 LIBYUV_API
125 int cpu_info_ = 0;
126 
127 // Test environment variable for disabling CPU features. Any non-zero value
128 // to disable. Zero ignored to make it easy to set the variable on/off.
TestEnv(const char * name)129 static bool TestEnv(const char* name) {
130   const char* var = getenv(name);
131   if (var) {
132     if (var[0] != '0') {
133       return true;
134     }
135   }
136   return false;
137 }
138 
139 LIBYUV_API
InitCpuFlags(void)140 int InitCpuFlags(void) {
141 #if !defined(__CLR_VER) && defined(CPU_X86)
142   int cpu_info[4];
143   __cpuid(cpu_info, 1);
144   cpu_info_ = ((cpu_info[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
145               ((cpu_info[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
146               ((cpu_info[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
147               ((cpu_info[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
148               (((cpu_info[2] & 0x18000000) == 0x18000000) ? kCpuHasAVX : 0) |
149               kCpuInitialized | kCpuHasX86;
150 #ifdef HAS_XGETBV
151   if (cpu_info_ & kCpuHasAVX) {
152     __cpuid(cpu_info, 7);
153     if ((cpu_info[1] & 0x00000020) &&
154         ((XGetBV(kXCR_XFEATURE_ENABLED_MASK) & 0x06) == 0x06)) {
155       cpu_info_ |= kCpuHasAVX2;
156     }
157   }
158 #endif
159   // environment variable overrides for testing.
160   if (TestEnv("LIBYUV_DISABLE_X86")) {
161     cpu_info_ &= ~kCpuHasX86;
162   }
163   if (TestEnv("LIBYUV_DISABLE_SSE2")) {
164     cpu_info_ &= ~kCpuHasSSE2;
165   }
166   if (TestEnv("LIBYUV_DISABLE_SSSE3")) {
167     cpu_info_ &= ~kCpuHasSSSE3;
168   }
169   if (TestEnv("LIBYUV_DISABLE_SSE41")) {
170     cpu_info_ &= ~kCpuHasSSE41;
171   }
172   if (TestEnv("LIBYUV_DISABLE_SSE42")) {
173     cpu_info_ &= ~kCpuHasSSE42;
174   }
175   if (TestEnv("LIBYUV_DISABLE_AVX")) {
176     cpu_info_ &= ~kCpuHasAVX;
177   }
178   if (TestEnv("LIBYUV_DISABLE_AVX2")) {
179     cpu_info_ &= ~kCpuHasAVX2;
180   }
181   if (TestEnv("LIBYUV_DISABLE_ASM")) {
182     cpu_info_ = kCpuInitialized;
183   }
184 #elif defined(__arm__)
185 #if defined(__linux__) && (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
186   // linux arm parse text file for neon detect.
187   cpu_info_ = ArmCpuCaps("/proc/cpuinfo");
188 #elif defined(__ARM_NEON__)
189   // gcc -mfpu=neon defines __ARM_NEON__
190   // Enable Neon if you want support for Neon and Arm, and use MaskCpuFlags
191   // to disable Neon on devices that do not have it.
192   cpu_info_ = kCpuHasNEON;
193 #endif
194   cpu_info_ |= kCpuInitialized | kCpuHasARM;
195   if (TestEnv("LIBYUV_DISABLE_NEON")) {
196     cpu_info_ &= ~kCpuHasNEON;
197   }
198   if (TestEnv("LIBYUV_DISABLE_ASM")) {
199     cpu_info_ = kCpuInitialized;
200   }
201 #endif  // __arm__
202   return cpu_info_;
203 }
204 
205 LIBYUV_API
MaskCpuFlags(int enable_flags)206 void MaskCpuFlags(int enable_flags) {
207   InitCpuFlags();
208   cpu_info_ = (cpu_info_ & enable_flags) | kCpuInitialized;
209 }
210 
211 #ifdef __cplusplus
212 }  // extern "C"
213 }  // namespace libyuv
214 #endif
215