Merge pull request #329 from Tilka/rounding

Small FPU settings changes
This commit is contained in:
Ryan Houdek 2014-05-02 12:11:09 -05:00
commit a1374dd4ba
1 changed files with 14 additions and 39 deletions

View File

@ -2,6 +2,8 @@
// Licensed under GPLv2 // Licensed under GPLv2
// Refer to the license.txt file included. // Refer to the license.txt file included.
#include <cfenv>
#include "Common/Common.h" #include "Common/Common.h"
#include "Common/CPUDetect.h" #include "Common/CPUDetect.h"
#include "Common/FPURoundMode.h" #include "Common/FPURoundMode.h"
@ -20,41 +22,18 @@ namespace FPURoundMode
void SetRoundMode(int mode) void SetRoundMode(int mode)
{ {
// Set FPU rounding mode to mimic the PowerPC's // Convert PowerPC to native rounding mode.
#ifdef _M_X86_32 const int rounding_mode_lut[] = {
// This shouldn't really be needed anymore since we use SSE FE_TONEAREST,
#ifdef _WIN32 FE_TOWARDZERO,
const int table[4] = FE_UPWARD,
{ FE_DOWNWARD
_RC_NEAR, };
_RC_CHOP, fesetround(rounding_mode_lut[mode]);
_RC_UP,
_RC_DOWN
};
_set_controlfp(_MCW_RC, table[mode]);
#else
const unsigned short X87_ROUND_MASK = 3 << 10;
const unsigned short x87_rounding_table[] =
{
0 << 10, // nearest
3 << 10, // zero
2 << 10, // +inf
1 << 10, // -inf
};
unsigned short _mode;
asm ("fstcw %0" : "=m" (_mode));
_mode = (_mode & ~X87_ROUND_MASK) | x87_rounding_table[mode];
asm ("fldcw %0" : : "m" (_mode));
#endif
#endif
} }
void SetPrecisionMode(PrecisionMode mode) void SetPrecisionMode(PrecisionMode mode)
{ {
#ifdef _M_X86_32
// sets the floating-point lib to 53-bit
// PowerPC has a 53bit floating pipeline only
// eg: sscanf is very sensitive
#ifdef _WIN32 #ifdef _WIN32
_control87(_PC_53, MCW_PC); _control87(_PC_53, MCW_PC);
#else #else
@ -64,14 +43,10 @@ namespace FPURoundMode
2 << 8, // 53 bits 2 << 8, // 53 bits
3 << 8, // 64 bits 3 << 8, // 64 bits
}; };
unsigned short _mode; unsigned short cw;
asm ("fstcw %0" : "=m" (_mode)); asm ("fnstcw %0" : "=m" (cw));
_mode = (_mode & ~PRECISION_MASK) | precision_table[mode]; cw = (cw & ~PRECISION_MASK) | precision_table[mode];
asm ("fldcw %0" : : "m" (_mode)); asm ("fldcw %0" : : "m" (cw));
#endif
#else
//x64 doesn't need this - fpu is done with SSE
//but still - set any useful sse options here
#endif #endif
} }