diff options
| author | Chris Robinson <[email protected]> | 2019-03-18 20:14:40 -0700 |
|---|---|---|
| committer | Chris Robinson <[email protected]> | 2019-03-18 20:14:40 -0700 |
| commit | 2de1d17bc6b07ec4e7680f1c46e147eeaa269f05 (patch) | |
| tree | 69165764c92d16727267153ba5b2acd51a69381d | |
| parent | 6a0b2ed0ca2b7628771ee450567c2c62c575db36 (diff) | |
Use SSE intrinsics to set FTZ and DAZ
| -rw-r--r-- | Alc/fpu_modes.h | 2 | ||||
| -rw-r--r-- | Alc/helpers.cpp | 19 |
2 files changed, 18 insertions, 3 deletions
diff --git a/Alc/fpu_modes.h b/Alc/fpu_modes.h index 1afd6618..8261692e 100644 --- a/Alc/fpu_modes.h +++ b/Alc/fpu_modes.h @@ -2,7 +2,7 @@ #define FPU_MODES_H class FPUCtl { -#if defined(__GNUC__) && defined(HAVE_SSE) +#if defined(HAVE_SSE_INTRINSICS) || (defined(__GNUC__) && defined(HAVE_SSE)) unsigned int sse_state{}; #elif defined(HAVE___CONTROL87_2) unsigned int state{}; diff --git a/Alc/helpers.cpp b/Alc/helpers.cpp index 13ba3a2a..d5bbde1d 100644 --- a/Alc/helpers.cpp +++ b/Alc/helpers.cpp @@ -89,6 +89,9 @@ DEFINE_PROPERTYKEY(PKEY_AudioEndpoint_GUID, 0x1da5d803, 0xd492, 0x4edd, 0x8c, 0x #ifdef HAVE_CPUID_H #include <cpuid.h> #endif +#ifdef HAVE_SSE_INTRINSICS +#include <xmmintrin.h> +#endif #ifdef HAVE_SYS_SYSCONF_H #include <sys/sysconf.h> #endif @@ -261,7 +264,15 @@ void FillCPUCaps(int capfilter) FPUCtl::FPUCtl() noexcept { -#if defined(__GNUC__) && defined(HAVE_SSE) +#if defined(HAVE_SSE_INTRINSICS) + this->sse_state = _mm_getcsr(); + unsigned int sseState = this->sse_state; + sseState |= 0x8000; /* set flush-to-zero */ + sseState |= 0x0040; /* set denormals-are-zero */ + _mm_setcsr(sseState); + +#elif defined(__GNUC__) && defined(HAVE_SSE) + if((CPUCapFlags&CPU_CAP_SSE)) { __asm__ __volatile__("stmxcsr %0" : "=m" (*&this->sse_state)); @@ -290,7 +301,11 @@ void FPUCtl::leave() noexcept { if(!this->in_mode) return; -#if defined(__GNUC__) && defined(HAVE_SSE) +#if defined(HAVE_SSE_INTRINSICS) + _mm_setcsr(this->sse_state); + +#elif defined(__GNUC__) && defined(HAVE_SSE) + if((CPUCapFlags&CPU_CAP_SSE)) __asm__ __volatile__("ldmxcsr %0" : : "m" (*&this->sse_state)); |