1 // Copyright 2022 Google Inc. All Rights Reserved. 2 // 3 // Use of this source code is governed by a BSD-style license 4 // that can be found in the COPYING file in the root of the source 5 // tree. An additional intellectual property rights grant can be found 6 // in the file PATENTS. All contributing project authors may 7 // be found in the AUTHORS file in the root of the source tree. 8 // ----------------------------------------------------------------------------- 9 // 10 // CPU detection functions and macros. 11 // 12 // Author: Skal (pascal.massimino@gmail.com) 13 14 #ifndef WEBP_DSP_CPU_H_ 15 #define WEBP_DSP_CPU_H_ 16 17 #include <stddef.h> 18 19 #ifdef HAVE_CONFIG_H 20 #include "src/webp/config.h" 21 #endif 22 23 #include "src/webp/types.h" 24 25 #if defined(__GNUC__) 26 #define LOCAL_GCC_VERSION ((__GNUC__ << 8) | __GNUC_MINOR__) 27 #define LOCAL_GCC_PREREQ(maj, min) (LOCAL_GCC_VERSION >= (((maj) << 8) | (min))) 28 #else 29 #define LOCAL_GCC_VERSION 0 30 #define LOCAL_GCC_PREREQ(maj, min) 0 31 #endif 32 33 #if defined(__clang__) 34 #define LOCAL_CLANG_VERSION ((__clang_major__ << 8) | __clang_minor__) 35 #define LOCAL_CLANG_PREREQ(maj, min) \ 36 (LOCAL_CLANG_VERSION >= (((maj) << 8) | (min))) 37 #else 38 #define LOCAL_CLANG_VERSION 0 39 #define LOCAL_CLANG_PREREQ(maj, min) 0 40 #endif 41 42 #ifndef __has_builtin 43 #define __has_builtin(x) 0 44 #endif 45 46 #if !defined(HAVE_CONFIG_H) 47 #if defined(_MSC_VER) && _MSC_VER > 1310 && \ 48 (defined(_M_X64) || defined(_M_IX86)) 49 #define WEBP_MSC_SSE2 // Visual C++ SSE2 targets 50 #endif 51 52 #if defined(_MSC_VER) && _MSC_VER >= 1500 && \ 53 (defined(_M_X64) || defined(_M_IX86)) 54 #define WEBP_MSC_SSE41 // Visual C++ SSE4.1 targets 55 #endif 56 #endif 57 58 // WEBP_HAVE_* are used to indicate the presence of the instruction set in dsp 59 // files without intrinsics, allowing the corresponding Init() to be called. 60 // Files containing intrinsics will need to be built targeting the instruction 61 // set so should succeed on one of the earlier tests. 62 #if (defined(__SSE2__) || defined(WEBP_MSC_SSE2)) && \ 63 (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_SSE2)) 64 #define WEBP_USE_SSE2 65 #endif 66 67 #if defined(WEBP_USE_SSE2) && !defined(WEBP_HAVE_SSE2) 68 #define WEBP_HAVE_SSE2 69 #endif 70 71 #if (defined(__SSE4_1__) || defined(WEBP_MSC_SSE41)) && \ 72 (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_SSE41)) 73 #define WEBP_USE_SSE41 74 #endif 75 76 #if defined(WEBP_USE_SSE41) && !defined(WEBP_HAVE_SSE41) 77 #define WEBP_HAVE_SSE41 78 #endif 79 80 #undef WEBP_MSC_SSE41 81 #undef WEBP_MSC_SSE2 82 83 // The intrinsics currently cause compiler errors with arm-nacl-gcc and the 84 // inline assembly would need to be modified for use with Native Client. 85 #if ((defined(__ARM_NEON__) || defined(__aarch64__)) && \ 86 (!defined(HAVE_CONFIG_H) || defined(WEBP_HAVE_NEON))) && \ 87 !defined(__native_client__) 88 #define WEBP_USE_NEON 89 #endif 90 91 #if !defined(WEBP_USE_NEON) && defined(__ANDROID__) && \ 92 defined(__ARM_ARCH_7A__) && defined(HAVE_CPU_FEATURES_H) 93 #define WEBP_ANDROID_NEON // Android targets that may have NEON 94 #define WEBP_USE_NEON 95 #endif 96 97 // Note: ARM64 is supported in Visual Studio 2017, but requires the direct 98 // inclusion of arm64_neon.h; Visual Studio 2019 includes this file in 99 // arm_neon.h. Compile errors were seen with Visual Studio 2019 16.4 with 100 // vtbl4_u8(); a fix was made in 16.6. 101 #if defined(_MSC_VER) && ((_MSC_VER >= 1700 && defined(_M_ARM)) || \ 102 (_MSC_VER >= 1926 && defined(_M_ARM64))) 103 #define WEBP_USE_NEON 104 #define WEBP_USE_INTRINSICS 105 #endif 106 107 #if defined(WEBP_USE_NEON) && !defined(WEBP_HAVE_NEON) 108 #define WEBP_HAVE_NEON 109 #endif 110 111 #if defined(__mips__) && !defined(__mips64) && defined(__mips_isa_rev) && \ 112 (__mips_isa_rev >= 1) && (__mips_isa_rev < 6) 113 #define WEBP_USE_MIPS32 114 #if (__mips_isa_rev >= 2) 115 #define WEBP_USE_MIPS32_R2 116 #if defined(__mips_dspr2) || (defined(__mips_dsp_rev) && __mips_dsp_rev >= 2) 117 #define WEBP_USE_MIPS_DSP_R2 118 #endif 119 #endif 120 #endif 121 122 #if defined(__mips_msa) && defined(__mips_isa_rev) && (__mips_isa_rev >= 5) 123 #define WEBP_USE_MSA 124 #endif 125 126 #ifndef WEBP_DSP_OMIT_C_CODE 127 #define WEBP_DSP_OMIT_C_CODE 1 128 #endif 129 130 #if defined(WEBP_USE_NEON) && WEBP_DSP_OMIT_C_CODE 131 #define WEBP_NEON_OMIT_C_CODE 1 132 #else 133 #define WEBP_NEON_OMIT_C_CODE 0 134 #endif 135 136 #if !(LOCAL_CLANG_PREREQ(3, 8) || LOCAL_GCC_PREREQ(4, 8) || \ 137 defined(__aarch64__)) 138 #define WEBP_NEON_WORK_AROUND_GCC 1 139 #else 140 #define WEBP_NEON_WORK_AROUND_GCC 0 141 #endif 142 143 // This macro prevents thread_sanitizer from reporting known concurrent writes. 144 #define WEBP_TSAN_IGNORE_FUNCTION 145 #if defined(__has_feature) 146 #if __has_feature(thread_sanitizer) 147 #undef WEBP_TSAN_IGNORE_FUNCTION 148 #define WEBP_TSAN_IGNORE_FUNCTION __attribute__((no_sanitize_thread)) 149 #endif 150 #endif 151 152 #if defined(__has_feature) 153 #if __has_feature(memory_sanitizer) 154 #define WEBP_MSAN 155 #endif 156 #endif 157 158 #if defined(WEBP_USE_THREAD) && !defined(_WIN32) 159 #include <pthread.h> // NOLINT 160 161 #define WEBP_DSP_INIT(func) \ 162 do { \ 163 static volatile VP8CPUInfo func##_last_cpuinfo_used = \ 164 (VP8CPUInfo)&func##_last_cpuinfo_used; \ 165 static pthread_mutex_t func##_lock = PTHREAD_MUTEX_INITIALIZER; \ 166 if (pthread_mutex_lock(&func##_lock)) break; \ 167 if (func##_last_cpuinfo_used != VP8GetCPUInfo) func(); \ 168 func##_last_cpuinfo_used = VP8GetCPUInfo; \ 169 (void)pthread_mutex_unlock(&func##_lock); \ 170 } while (0) 171 #else // !(defined(WEBP_USE_THREAD) && !defined(_WIN32)) 172 #define WEBP_DSP_INIT(func) \ 173 do { \ 174 static volatile VP8CPUInfo func##_last_cpuinfo_used = \ 175 (VP8CPUInfo)&func##_last_cpuinfo_used; \ 176 if (func##_last_cpuinfo_used == VP8GetCPUInfo) break; \ 177 func(); \ 178 func##_last_cpuinfo_used = VP8GetCPUInfo; \ 179 } while (0) 180 #endif // defined(WEBP_USE_THREAD) && !defined(_WIN32) 181 182 // Defines an Init + helper function that control multiple initialization of 183 // function pointers / tables. 184 /* Usage: 185 WEBP_DSP_INIT_FUNC(InitFunc) { 186 ...function body 187 } 188 */ 189 #define WEBP_DSP_INIT_FUNC(name) \ 190 static WEBP_TSAN_IGNORE_FUNCTION void name##_body(void); \ 191 WEBP_TSAN_IGNORE_FUNCTION void name(void) { WEBP_DSP_INIT(name##_body); } \ 192 static WEBP_TSAN_IGNORE_FUNCTION void name##_body(void) 193 194 #define WEBP_UBSAN_IGNORE_UNDEF 195 #define WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW 196 #if defined(__clang__) && defined(__has_attribute) 197 #if __has_attribute(no_sanitize) 198 // This macro prevents the undefined behavior sanitizer from reporting 199 // failures. This is only meant to silence unaligned loads on platforms that 200 // are known to support them. 201 #undef WEBP_UBSAN_IGNORE_UNDEF 202 #define WEBP_UBSAN_IGNORE_UNDEF __attribute__((no_sanitize("undefined"))) 203 204 // This macro prevents the undefined behavior sanitizer from reporting 205 // failures related to unsigned integer overflows. This is only meant to 206 // silence cases where this well defined behavior is expected. 207 #undef WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW 208 #define WEBP_UBSAN_IGNORE_UNSIGNED_OVERFLOW \ 209 __attribute__((no_sanitize("unsigned-integer-overflow"))) 210 #endif 211 #endif 212 213 // If 'ptr' is NULL, returns NULL. Otherwise returns 'ptr + off'. 214 // Prevents undefined behavior sanitizer nullptr-with-nonzero-offset warning. 215 #if !defined(WEBP_OFFSET_PTR) 216 #define WEBP_OFFSET_PTR(ptr, off) (((ptr) == NULL) ? NULL : ((ptr) + (off))) 217 #endif 218 219 // Regularize the definition of WEBP_SWAP_16BIT_CSP (backward compatibility) 220 #if !defined(WEBP_SWAP_16BIT_CSP) 221 #define WEBP_SWAP_16BIT_CSP 0 222 #endif 223 224 // some endian fix (e.g.: mips-gcc doesn't define __BIG_ENDIAN__) 225 #if !defined(WORDS_BIGENDIAN) && \ 226 (defined(__BIG_ENDIAN__) || defined(_M_PPC) || \ 227 (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__))) 228 #define WORDS_BIGENDIAN 229 #endif 230 231 typedef enum { 232 kSSE2, 233 kSSE3, 234 kSlowSSSE3, // special feature for slow SSSE3 architectures 235 kSSE4_1, 236 kAVX, 237 kAVX2, 238 kNEON, 239 kMIPS32, 240 kMIPSdspR2, 241 kMSA 242 } CPUFeature; 243 244 #ifdef __cplusplus 245 extern "C" { 246 #endif 247 248 // returns true if the CPU supports the feature. 249 typedef int (*VP8CPUInfo)(CPUFeature feature); 250 WEBP_EXTERN VP8CPUInfo VP8GetCPUInfo; 251 252 #ifdef __cplusplus 253 } // extern "C" 254 #endif 255 256 #endif // WEBP_DSP_CPU_H_ 257