2 Copyright (C) 2012 Paul Davis
4 This program is free software; you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation; either version 2 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
14 You should have received a copy of the GNU General Public License
15 along with this program; if not, write to the Free Software
16 Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
20 #include "libpbd-config.h"
22 #define _XOPEN_SOURCE 600
23 #include <cstring> // for memset
28 #ifdef PLATFORM_WINDOWS
32 #include "pbd/compose.h"
34 #include "pbd/error.h"
41 FPU* FPU::_instance (0);
43 #if ( (defined __x86_64__) || (defined __i386__) || (defined _M_X64) || (defined _M_IX86) ) // ARCH_X86
44 #ifndef PLATFORM_WINDOWS
46 /* use __cpuid() as the name to match the MSVC/mingw intrinsic */
49 __cpuid(int regs[4], int cpuid_leaf)
56 "movl %%eax, (%1);\n\t"
57 "movl %%ebx, 4(%1);\n\t"
58 "movl %%ecx, 8(%1);\n\t"
59 "movl %%edx, 12(%1);\n\t"
63 :"=a" (cpuid_leaf) /* %eax clobbered by CPUID */
64 :"S" (regs), "a" (cpuid_leaf)
66 #if !defined(__i386__)
69 "%ecx", "%edx", "memory");
72 #endif /* !PLATFORM_WINDOWS */
74 #ifndef HAVE_XGETBV // Allow definition by build system
75 #if defined(__MINGW32__) && defined(__MINGW64_VERSION_MAJOR) && __MINGW64_VERSION_MAJOR >= 5
77 #elif defined(_MSC_VER) && _MSC_VER >= 1600
78 // '_xgetbv()' was only available from VC10 onwards
87 // '_xgetbv()' was only available from VC10 onwards
88 __declspec(noinline) static uint64_t
89 _xgetbv (uint32_t xcr)
93 // N.B. The following would probably work for a pre-VC10 build,
94 // although it might suffer from optimization issues. We'd need
95 // to place this function into its own (unoptimized) source file.
98 __asm _emit 0x0f __asm _emit 0x01 __asm _emit 0xd0 /*xgetbv*/
105 _xgetbv (uint32_t xcr)
108 /* it would be nice to make this work on OS X but as long we use veclib,
109 we don't really need to know about SSE/AVX on that platform.
114 __asm__ volatile ("xgetbv" : "=a" (eax), "=d" (edx) : "c" (xcr));
115 return (static_cast<uint64_t>(edx) << 32) | eax;
119 #endif /* !COMPILER_MSVC */
120 #endif /* !HAVE_XGETBV */
121 #endif /* ARCH_X86 */
123 #ifndef _XCR_XFEATURE_ENABLED_MASK
124 #define _XCR_XFEATURE_ENABLED_MASK 0
148 error << _("FPU object instantiated more than once") << endmsg;
151 if (getenv("ARDOUR_FPU_FLAGS")) {
152 _flags = Flags (atoi (getenv("ARDOUR_FPU_FLAGS")));
156 #if !( (defined __x86_64__) || (defined __i386__) || (defined _M_X64) || (defined _M_IX86) ) // !ARCH_X86
157 /* Non-Intel architecture, nothing to do here */
161 /* Get the CPU vendor just for kicks */
163 // __cpuid with an InfoType argument of 0 returns the number of
164 // valid Ids in CPUInfo[0] and the CPU identification string in
165 // the other three array elements. The CPU identification string is
166 // not in linear order. The code below arranges the information
167 // in a human readable form. The human readable order is CPUInfo[1] |
168 // CPUInfo[3] | CPUInfo[2]. CPUInfo[2] and CPUInfo[3] are swapped
169 // before using memcpy to copy these three array elements to cpu_string.
175 __cpuid (cpu_info, 0);
177 int num_ids = cpu_info[0];
178 std::swap(cpu_info[2], cpu_info[3]);
179 memcpy(cpu_string, &cpu_info[1], 3 * sizeof(cpu_info[1]));
180 cpu_vendor.assign(cpu_string, 3 * sizeof(cpu_info[1]));
182 info << string_compose (_("CPU vendor: %1"), cpu_vendor) << endmsg;
186 /* Now get CPU/FPU flags */
188 __cpuid (cpu_info, 1);
190 if ((cpu_info[2] & (1<<27)) /* OSXSAVE */ &&
191 (cpu_info[2] & (1<<28) /* AVX */) &&
192 ((_xgetbv (_XCR_XFEATURE_ENABLED_MASK) & 0x6) == 0x6)) { /* OS really supports XSAVE */
193 info << _("AVX-capable processor") << endmsg;
194 _flags = Flags (_flags | (HasAVX) );
197 if (cpu_info[3] & (1<<25)) {
198 _flags = Flags (_flags | (HasSSE|HasFlushToZero));
201 if (cpu_info[3] & (1<<26)) {
202 _flags = Flags (_flags | HasSSE2);
205 /* Figure out CPU/FPU denormal handling capabilities */
207 if (cpu_info[3] & (1 << 24)) {
211 /* DAZ wasn't available in the first version of SSE. Since
212 setting a reserved bit in MXCSR causes a general protection
213 fault, we need to be able to check the availability of this
214 feature without causing problems. To do this, one needs to
215 set up a 512-byte area of memory to save the SSE state to,
216 using fxsave, and then one needs to inspect bytes 28 through
217 31 for the MXCSR_MASK value. If bit 6 is set, DAZ is
218 supported, otherwise, it isn't.
221 #ifndef HAVE_POSIX_MEMALIGN
222 # ifdef PLATFORM_WINDOWS
223 fxbuf = (char **) _aligned_malloc (sizeof (char *), 16);
225 *fxbuf = (char *) _aligned_malloc (512, 16);
228 # warning using default malloc for aligned memory
229 fxbuf = (char **) malloc (sizeof (char *));
231 *fxbuf = (char *) malloc (512);
235 (void) posix_memalign ((void **) &fxbuf, 16, sizeof (char *));
237 (void) posix_memalign ((void **) fxbuf, 16, 512);
241 memset (*fxbuf, 0, 512);
258 uint32_t mxcsr_mask = *((uint32_t*) &((*fxbuf)[28]));
260 /* if the mask is zero, set its default value (from intel specs) */
262 if (mxcsr_mask == 0) {
266 if (mxcsr_mask & (1<<6)) {
267 _flags = Flags (_flags | HasDenormalsAreZero);
270 #if !defined HAVE_POSIX_MEMALIGN && defined PLATFORM_WINDOWS
271 _aligned_free (*fxbuf);
272 _aligned_free (fxbuf);
279 /* finally get the CPU brand */
281 __cpuid (cpu_info, 0x80000000);
283 const int parameter_end = 0x80000004;
286 if (cpu_info[0] >= parameter_end) {
287 char* cpu_string_ptr = cpu_string;
289 for (int parameter = 0x80000002; parameter <= parameter_end &&
290 cpu_string_ptr < &cpu_string[sizeof(cpu_string)]; parameter++) {
291 __cpuid(cpu_info, parameter);
292 memcpy(cpu_string_ptr, cpu_info, sizeof(cpu_info));
293 cpu_string_ptr += sizeof(cpu_info);
295 cpu_brand.assign(cpu_string, cpu_string_ptr - cpu_string);
296 info << string_compose (_("CPU brand: %1"), cpu_brand) << endmsg;
299 #endif /* !ARCH_X86 */