2013-04-17 23:09:55 -04:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
								// Copyright 2013 Dolphin Emulator Project
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								// Licensed under GPLv2
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								// Refer to the license.txt file included.
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
								
							 | 
						
					
						
							
								
									
										
										
										
											2014-03-28 19:06:12 +01:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
								#include <cfenv>
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								
							 | 
						
					
						
							
								
									
										
										
										
											2014-02-17 05:18:15 -05:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
								#include "Common/Common.h"
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								#include "Common/CPUDetect.h"
							 | 
						
					
						
							
								
									
										
										
										
											2014-02-23 14:57:31 +01:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
								#include "Common/FPURoundMode.h"
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
								
							 | 
						
					
						
							
								
									
										
										
										
											2014-02-23 14:57:31 +01:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
								#ifdef _WIN32
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								#	include <mmintrin.h>
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								#else
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								#	include <xmmintrin.h>
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
								#endif
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								namespace FPURoundMode
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								{
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									// Get the default SSE states here.
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									static u32 saved_sse_state = _mm_getcsr();
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									static const u32 default_sse_state = _mm_getcsr();
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								
							 | 
						
					
						
							
								
									
										
										
										
											2014-03-09 15:21:50 +01:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
									void SetRoundMode(int mode)
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
									{
							 | 
						
					
						
							
								
									
										
										
										
											2014-03-28 19:06:12 +01:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
										// Convert PowerPC to native rounding mode.
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
										const int rounding_mode_lut[] = {
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											FE_TONEAREST,
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											FE_TOWARDZERO,
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											FE_UPWARD,
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											FE_DOWNWARD
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
										};
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
										fesetround(rounding_mode_lut[mode]);
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
									}
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								
							 | 
						
					
						
							
								
									
										
										
										
											2014-03-09 15:21:50 +01:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
									void SetPrecisionMode(PrecisionMode mode)
							 | 
						
					
						
							
								
									
										
										
										
											2013-03-19 21:51:12 -04:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
									{
							 | 
						
					
						
							
								
									
										
										
										
											2014-05-10 20:21:07 -07:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
										#ifdef _M_X86_32
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											// sets the floating-point lib to 53-bit
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											// PowerPC has a 53bit floating pipeline only
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											// eg: sscanf is very sensitive
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
										#ifdef _WIN32
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											_control87(_PC_53, MCW_PC);
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
										#else
							 | 
						
					
						
							
								
									
										
										
										
											2014-02-23 14:57:31 +01:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
											const unsigned short PRECISION_MASK = 3 << 8;
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											const unsigned short precision_table[] = {
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
												0 << 8, // 24 bits
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
												2 << 8, // 53 bits
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
												3 << 8, // 64 bits
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
											};
							 | 
						
					
						
							
								
									
										
										
										
											2014-05-10 20:21:07 -07:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
											unsigned short _mode;
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											asm ("fstcw %0" : "=m" (_mode));
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											_mode = (_mode & ~PRECISION_MASK) | precision_table[mode];
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											asm ("fldcw %0" : : "m" (_mode));
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
										#endif
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
										#else
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											//x64 doesn't need this - fpu is done with SSE
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											//but still - set any useful sse options here
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
										#endif
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									}
							 | 
						
					
						
							
								
									
										
										
										
											2013-10-24 13:52:22 +02:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
								
							 | 
						
					
						
							
								
									
										
										
										
											2014-03-09 15:21:50 +01:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
									void SetSIMDMode(int rounding_mode, bool non_ieee_mode)
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
									{
							 | 
						
					
						
							
								
									
										
										
										
											2014-02-23 14:57:31 +01:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
										// OR-mask for disabling FPU exceptions (bits 7-12 in the MXCSR register)
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
										const u32 EXCEPTION_MASK = 0x1F80;
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
										// Flush-To-Zero (non-IEEE mode: denormal outputs are set to +/- 0)
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
										const u32 FTZ = 0x8000;
							 | 
						
					
						
							
								
									
										
										
										
											2013-10-24 13:52:22 +02:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
										// lookup table for FPSCR.RN-to-MXCSR.RC translation
							 | 
						
					
						
							
								
									
										
										
										
											2014-02-23 14:57:31 +01:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
										static const u32 simd_rounding_table[] =
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
										{
							 | 
						
					
						
							
								
									
										
										
										
											2013-10-24 13:52:22 +02:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
											(0 << 13) | EXCEPTION_MASK, // nearest
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											(3 << 13) | EXCEPTION_MASK, // -inf
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											(2 << 13) | EXCEPTION_MASK, // +inf
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
											(1 << 13) | EXCEPTION_MASK, // zero
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
										};
							 | 
						
					
						
							
								
									
										
										
										
											2014-02-23 14:57:31 +01:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
										u32 csr = simd_rounding_table[rounding_mode];
							 | 
						
					
						
							
								
									
										
										
										
											2013-10-24 13:52:22 +02:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
								
							 | 
						
					
						
							
								
									
										
										
										
											2014-02-23 14:57:31 +01:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
										if (non_ieee_mode)
							 | 
						
					
						
							
								
									
										
										
										
											2013-10-24 13:52:22 +02:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
										{
							 | 
						
					
						
							
								
									
										
										
										
											2014-06-14 21:22:34 -07:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
											csr |= FTZ;
							 | 
						
					
						
							
								
									
										
										
										
											2013-10-24 13:52:22 +02:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
										}
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
										_mm_setcsr(csr);
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									}
							 | 
						
					
						
							
								
									
										
										
										
											2013-03-19 21:51:12 -04:00
										 
									 
								 
							 | 
							
								
									
										
									
								
							 | 
							
								
							 | 
							
							
								
							 | 
						
					
						
							
								
									
										
										
										
											2013-02-26 13:49:00 -06:00
										 
									 
								 
							 | 
							
								
							 | 
							
								
							 | 
							
							
									void SaveSIMDState()
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									{
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
										saved_sse_state = _mm_getcsr();
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									}
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									void LoadSIMDState()
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									{
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
										_mm_setcsr(saved_sse_state);
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									}
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									void LoadDefaultSIMDState()
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									{
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
										_mm_setcsr(default_sse_state);
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
									}
							 | 
						
					
						
							| 
								
							 | 
							
								
							 | 
							
								
							 | 
							
							
								}
							 |