Menu
▾
▴
[OMPT-svn-commit] SF.net SVN: modplug:[2311] trunk/OpenMPT
|
From: <man...@us...> - 2013-06-07 00:18:26
|
Revision: 2311
http://sourceforge.net/p/modplug/code/2311
Author: manxorist
Date: 2013-06-07 00:18:17 +0000 (Fri, 07 Jun 2013)
Log Message:
-----------
[Ref] Rewrite cpuid feature detection in more readable C with intrinsics and move it from soundlib/Mmx_mix.cpp into common/misc_util.cpp.
[Fix] Fix the semantics and naming of the feature flags and feature flag macros to more closely match commonly used feature names.
Modified Paths:
--------------
trunk/OpenMPT/common/BuildSettings.h
trunk/OpenMPT/common/misc_util.cpp
trunk/OpenMPT/common/misc_util.h
trunk/OpenMPT/mptrack/Mptrack.cpp
trunk/OpenMPT/mptrack/View_smp.cpp
trunk/OpenMPT/mptrack/Vstplug.cpp
trunk/OpenMPT/sounddsp/EQ.cpp
trunk/OpenMPT/soundlib/Mmx_mix.cpp
trunk/OpenMPT/soundlib/Snd_defs.h
trunk/OpenMPT/soundlib/Sndfile.h
Modified: trunk/OpenMPT/common/BuildSettings.h
===================================================================
--- trunk/OpenMPT/common/BuildSettings.h 2013-06-06 23:16:28 UTC (rev 2310)
+++ trunk/OpenMPT/common/BuildSettings.h 2013-06-07 00:18:17 UTC (rev 2311)
@@ -44,12 +44,12 @@
// Generate inline assembly using MMX instructions (only used when the CPU supports it).
#define ENABLE_MMX
-// Generate inline assembly using 3DNOW instructions (only used when the CPU supports it).
-#define ENABLE_3DNOW
-
// Generate inline assembly using SSE instructions (only used when the CPU supports it).
#define ENABLE_SSE
+// Generate inline assembly using AMD specific instruction set extensions (only used when the CPU supports it).
+#define ENABLE_X86_AMD
+
#endif // ENABLE_ASM
Modified: trunk/OpenMPT/common/misc_util.cpp
===================================================================
--- trunk/OpenMPT/common/misc_util.cpp 2013-06-06 23:16:28 UTC (rev 2310)
+++ trunk/OpenMPT/common/misc_util.cpp 2013-06-07 00:18:17 UTC (rev 2311)
@@ -10,3 +10,124 @@
#include "stdafx.h"
#include "misc_util.h"
+
+
+#if defined(ENABLE_ASM)
+
+
+uint32 ProcSupport = 0;
+
+
+#if MPT_COMPILER_MSVC && defined(ENABLE_X86)
+
+
+#include <intrin.h>
+
+
+typedef char cpuid_result_string[12];
+
+
+struct cpuid_result {
+ uint32 a;
+ uint32 b;
+ uint32 c;
+ uint32 d;
+ std::string as_string() const
+ {
+ cpuid_result_string result;
+ result[0+0] = (b >> 0) & 0xff;
+ result[0+1] = (b >> 8) & 0xff;
+ result[0+2] = (b >>16) & 0xff;
+ result[0+3] = (b >>24) & 0xff;
+ result[4+0] = (d >> 0) & 0xff;
+ result[4+1] = (d >> 8) & 0xff;
+ result[4+2] = (d >>16) & 0xff;
+ result[4+3] = (d >>24) & 0xff;
+ result[8+0] = (c >> 0) & 0xff;
+ result[8+1] = (c >> 8) & 0xff;
+ result[8+2] = (c >>16) & 0xff;
+ result[8+3] = (c >>24) & 0xff;
+ return std::string(result, result + 12);
+ }
+};
+
+
+static cpuid_result cpuid(uint32 function)
+//----------------------------------------
+{
+ cpuid_result result;
+ int CPUInfo[4];
+ __cpuid(CPUInfo, function);
+ result.a = CPUInfo[0];
+ result.b = CPUInfo[1];
+ result.c = CPUInfo[2];
+ result.d = CPUInfo[3];
+ return result;
+}
+
+
+static bool has_cpuid()
+//---------------------
+{
+ const uint32 eflags_cpuid = 1<<21;
+ uint32 old_eflags = __readeflags();
+ __writeeflags(old_eflags ^ eflags_cpuid);
+ bool result = ((__readeflags() ^ old_eflags) & eflags_cpuid) ? true : false;
+ __writeeflags(old_eflags);
+ return result;
+}
+
+
+void InitProcSupport()
+//--------------------
+{
+
+ ProcSupport = 0;
+
+ if(has_cpuid())
+ {
+
+ cpuid_result VendorString = cpuid(0x00000000u);
+
+ cpuid_result StandardFeatureFlags = cpuid(0x00000001u);
+ if(StandardFeatureFlags.d & (1<<23)) ProcSupport |= PROCSUPPORT_MMX;
+ if(StandardFeatureFlags.d & (1<<25)) ProcSupport |= PROCSUPPORT_SSE;
+ if(StandardFeatureFlags.d & (1<<26)) ProcSupport |= PROCSUPPORT_SSE2;
+ if(StandardFeatureFlags.c & (1<< 0)) ProcSupport |= PROCSUPPORT_SSE3;
+
+ if(VendorString.as_string() == "AuthenticAMD")
+ {
+
+ cpuid_result ExtendedVendorString = cpuid(0x80000000u);
+ if(ExtendedVendorString.a >= 0x80000001u)
+ {
+
+ cpuid_result ExtendedFeatureFlags = cpuid(0x80000001u);
+ if(ExtendedFeatureFlags.d & (1<<22)) ProcSupport |= PROCSUPPORT_AMD_MMXEXT;
+ if(ExtendedFeatureFlags.d & (1<<31)) ProcSupport |= PROCSUPPORT_AMD_3DNOW;
+ if(ExtendedFeatureFlags.d & (1<<30)) ProcSupport |= PROCSUPPORT_AMD_3DNOW2;
+
+ }
+
+ }
+
+ }
+
+}
+
+
+#else // !( MPT_COMPILER_MSVC && ENABLE_X86 )
+
+
+void InitProcSupport()
+//--------------------
+{
+ ProcSupport = 0;
+}
+
+
+#endif // MPT_COMPILER_MSVC && ENABLE_X86
+
+
+#endif // ENABLE_ASM
+
Modified: trunk/OpenMPT/common/misc_util.h
===================================================================
--- trunk/OpenMPT/common/misc_util.h 2013-06-06 23:16:28 UTC (rev 2310)
+++ trunk/OpenMPT/common/misc_util.h 2013-06-07 00:18:17 UTC (rev 2311)
@@ -441,3 +441,21 @@
};
} // namespace Util
+
+
+#ifdef ENABLE_ASM
+#define PROCSUPPORT_MMX 0x00001 // Processor supports MMX instructions
+#define PROCSUPPORT_SSE 0x00010 // Processor supports SSE instructions
+#define PROCSUPPORT_SSE2 0x00020 // Processor supports SSE2 instructions
+#define PROCSUPPORT_SSE3 0x00040 // Processor supports SSE3 instructions
+#define PROCSUPPORT_AMD_MMXEXT 0x10000 // Processor supports AMD MMX extensions
+#define PROCSUPPORT_AMD_3DNOW 0x20000 // Processor supports AMD 3DNow! instructions
+#define PROCSUPPORT_AMD_3DNOW2 0x40000 // Processor supports AMD 3DNow!2 instructions
+extern uint32 ProcSupport;
+void InitProcSupport();
+static inline uint32 GetProcSupport()
+{
+ return ProcSupport;
+}
+#endif // ENABLE_ASM
+
Modified: trunk/OpenMPT/mptrack/Mptrack.cpp
===================================================================
--- trunk/OpenMPT/mptrack/Mptrack.cpp 2013-06-06 23:16:28 UTC (rev 2310)
+++ trunk/OpenMPT/mptrack/Mptrack.cpp 2013-06-07 00:18:17 UTC (rev 2311)
@@ -874,7 +874,7 @@
{
TrackerSettings::Instance().m_ResamplerSettings.SrcMode = SRCMODE_SPLINE;
}
- if(GetProcSupport() & PROCSUPPORT_MMXEX)
+ if(GetProcSupport() & PROCSUPPORT_SSE)
{
TrackerSettings::Instance().m_ResamplerSettings.SrcMode = SRCMODE_POLYPHASE;
}
Modified: trunk/OpenMPT/mptrack/View_smp.cpp
===================================================================
--- trunk/OpenMPT/mptrack/View_smp.cpp 2013-06-06 23:16:28 UTC (rev 2310)
+++ trunk/OpenMPT/mptrack/View_smp.cpp 2013-06-07 00:18:17 UTC (rev 2311)
@@ -521,10 +521,10 @@
}
-#ifdef ENABLE_MMX
+#if defined(ENABLE_X86_AMD) || defined(ENABLE_SSE)
-static void mmxex_findminmax16(void *p, int scanlen, int smplsize, int *smin, int *smax)
-//--------------------------------------------------------------------------------------
+static void amdmmxext_or_sse_findminmax16(void *p, int scanlen, int smplsize, int *smin, int *smax)
+//-------------------------------------------------------------------------------------------------
{
_asm {
mov ebx, p
@@ -585,8 +585,8 @@
}
-static void mmxex_findminmax8(void *p, int scanlen, int smplsize, int *smin, int *smax)
-//-------------------------------------------------------------------------------------
+static void amdmmxext_or_sse_findminmax8(void *p, int scanlen, int smplsize, int *smin, int *smax)
+//------------------------------------------------------------------------------------------------
{
_asm {
mov ebx, p
@@ -715,10 +715,10 @@
signed short *p = (signed short *)(psample + poshi*smplsize);
smin = 32767;
smax = -32768;
-#ifdef ENABLE_MMX
- if(GetProcSupport() & PROCSUPPORT_MMXEX)
+#if defined(ENABLE_X86_AMD) || defined(ENABLE_SSE)
+ if(GetProcSupport() & (PROCSUPPORT_AMD_MMXEXT|PROCSUPPORT_SSE))
{
- mmxex_findminmax16(p, scanlen, smplsize, &smin, &smax);
+ amdmmxext_or_sse_findminmax16(p, scanlen, smplsize, &smin, &smax);
} else
#endif
{
@@ -738,10 +738,10 @@
signed char *p = psample + poshi * smplsize;
smin = 127;
smax = -128;
-#ifdef ENABLE_MMX
- if(GetProcSupport() & PROCSUPPORT_MMXEX)
+#if defined(ENABLE_X86_AMD) || defined(ENABLE_SSE)
+ if(GetProcSupport() & (PROCSUPPORT_AMD_MMXEXT|PROCSUPPORT_SSE))
{
- mmxex_findminmax8(p, scanlen, smplsize, &smin, &smax);
+ amdmmxext_or_sse_findminmax8(p, scanlen, smplsize, &smin, &smax);
} else
#endif
{
Modified: trunk/OpenMPT/mptrack/Vstplug.cpp
===================================================================
--- trunk/OpenMPT/mptrack/Vstplug.cpp 2013-06-06 23:16:28 UTC (rev 2310)
+++ trunk/OpenMPT/mptrack/Vstplug.cpp 2013-06-07 00:18:17 UTC (rev 2311)
@@ -3413,7 +3413,6 @@
return;
}
-#ifdef ENABLE_MMX
#ifdef ENABLE_SSE
if(GetProcSupport() & PROCSUPPORT_SSE)
{
@@ -3422,7 +3421,6 @@
SSEDeinterleaveInt16ToFloat(outputs[0], outputs[1], samples);
} else
#endif // ENABLE_SSE
-#endif // ENABLE_MMX
{
InterleaveFloatToInt16(inputs[0], inputs[1], samples);
m_pMediaProcess->Process(samples * 2 * sizeof(int16), reinterpret_cast<BYTE *>(m_pMixBuffer), m_DataTime, DMO_INPLACE_NORMAL);
Modified: trunk/OpenMPT/sounddsp/EQ.cpp
===================================================================
--- trunk/OpenMPT/sounddsp/EQ.cpp 2013-06-06 23:16:28 UTC (rev 2310)
+++ trunk/OpenMPT/sounddsp/EQ.cpp 2013-06-07 00:18:17 UTC (rev 2311)
@@ -109,10 +109,11 @@
}
+#ifdef ENABLE_X86_AMD
+
static void AMD_StereoEQ(EQBANDSTRUCT *pbl, EQBANDSTRUCT *pbr, float32 *pbuffer, UINT nCount)
//-------------------------------------------------------------------------------------------
{
-#ifdef ENABLE_3DNOW
float tmp[16];
_asm {
@@ -189,14 +190,16 @@
movd [edx+EQBANDSTRUCT.y2], mm7
emms
}
-#endif
}
+#endif // ENABLE_X86_AMD
+
+#ifdef ENABLE_SSE
+
static void SSE_StereoEQ(EQBANDSTRUCT *pbl, EQBANDSTRUCT *pbr, float32 *pbuffer, UINT nCount)
//-------------------------------------------------------------------------------------------
{
-#ifdef ENABLE_SSE
static const float gk1 = 1.0f;
_asm {
mov eax, pbl
@@ -284,9 +287,10 @@
movss [edx+EQBANDSTRUCT.y2], xmm1
done:;
}
-#endif // ENABLE_SSE
}
+#endif // ENABLE_SSE
+
#pragma warning(default:4100)
#else
@@ -326,7 +330,6 @@
{
#ifdef ENABLE_SSE
-#ifdef ENABLE_MMX
if(GetProcSupport() & PROCSUPPORT_SSE)
{
@@ -347,12 +350,11 @@
} else
-#endif // ENABLE_MMX
#endif // ENABLE_SSE
-#ifdef ENABLE_3DNOW
+#ifdef ENABLE_X86_AMD
- if(GetProcSupport() & PROCSUPPORT_3DNOW)
+ if(GetProcSupport() & PROCSUPPORT_AMD_3DNOW)
{
MonoMixToFloat(pbuffer, MixFloatBuffer, nCount*2, 1.0f/static_cast<float>(MIXING_CLIPMAX));
@@ -366,7 +368,7 @@
FloatToMonoMix(MixFloatBuffer, pbuffer, nCount*2, static_cast<float>(MIXING_CLIPMAX));
} else
-#endif // ENABLE_3DNOW
+#endif // ENABLE_X86_AMD
{
@@ -390,10 +392,10 @@
CEQ::CEQ()
//--------
{
- #if defined(ENABLE_MMX) || defined(ENABLE_SSE)
+ #if defined(ENABLE_SSE) || defined(ENABLE_X86_AMD)
ALWAYS_ASSERT(((uintptr_t)&(gEQ[0])) % 4 == 0);
ALWAYS_ASSERT(((uintptr_t)&(gEQ[1])) % 4 == 0);
- #endif
+ #endif // ENABLE_SSE || ENABLE_X86_AMD
memcpy(gEQ, gEQDefaults, sizeof(gEQ));
}
Modified: trunk/OpenMPT/soundlib/Mmx_mix.cpp
===================================================================
--- trunk/OpenMPT/soundlib/Mmx_mix.cpp 2013-06-06 23:16:28 UTC (rev 2310)
+++ trunk/OpenMPT/soundlib/Mmx_mix.cpp 2013-06-07 00:18:17 UTC (rev 2311)
@@ -24,75 +24,10 @@
#include "Sndfile.h"
-#ifdef ENABLE_ASM
-
-static uint32 gdwSysInfo = 0;
-
-void InitProcSupport()
-//--------------------
-{
-#ifdef ENABLE_X86
- static unsigned int fProcessorExtensions = 0;
- _asm
- {
- pushfd // Store original EFLAGS on stack
- pop eax // Get original EFLAGS in EAX
- mov ecx, eax // Duplicate original EFLAGS in ECX for toggle check
- xor eax, 0x00200000L // Flip ID bit in EFLAGS
- push eax // Save new EFLAGS value on stack
- popfd // Replace current EFLAGS value
- pushfd // Store new EFLAGS on stack
- pop eax // Get new EFLAGS in EAX
- xor eax, ecx // Can we toggle ID bit?
- jz Done // Jump if no, Processor is older than a Pentium so CPU_ID is not supported
- mov fProcessorExtensions, PROCSUPPORT_CPUID
- mov eax, 1 // Set EAX to tell the CPUID instruction what to return
- push ebx
- cpuid // Get family/model/stepping/features
- pop ebx
- test edx, 0x00800000L // Check if mmx technology available
- jz Done // Jump if no
- // Tests have passed, this machine supports the Intel MultiMedia Instruction Set!
- or fProcessorExtensions, PROCSUPPORT_MMX
- // Check for SSE: PROCSUPPORT_SSE
- test edx, 0x02000000L // check if SSE is present (bit 25)
- jz nosse // done if no
- // else set the correct bit in fProcessorExtensions
- or fProcessorExtensions, (PROCSUPPORT_MMXEX|PROCSUPPORT_SSE)
- jmp Done
- nosse:
- // Check for AMD 3DNow!
- mov eax, 0x80000000
- cpuid
- cmp eax, 0x80000000
- jbe Done
- mov eax, 0x80000001
- cpuid // CPU_ID
- test edx, 0x80000000
- jz Done
- or fProcessorExtensions, PROCSUPPORT_3DNOW // 3DNow! supported
- test edx, (1<<22) // Bit 22: AMD MMX extensions
- jz Done
- or fProcessorExtensions, PROCSUPPORT_MMXEX // MMX extensions supported
- Done:
- }
- gdwSysInfo = fProcessorExtensions;
-#endif
-}
-
-uint32 GetProcSupport()
-//---------------------
-{
- return gdwSysInfo;
-}
-
-#endif
-
-
////////////////////////////////////////////////////////////////////////////////////
// 3DNow! optimizations
-#ifdef ENABLE_3DNOW
+#ifdef ENABLE_X86_AMD
// Convert integer mix to floating-point
static void AMD_StereoMixToFloat(const int *pSrc, float *pOut1, float *pOut2, UINT nCount, const float _i2fc)
@@ -222,7 +157,7 @@
}
}
-#endif
+#endif // ENABLE_X86_AMD
///////////////////////////////////////////////////////////////////////////////////////
// SSE Optimizations
@@ -442,19 +377,19 @@
{
#ifdef ENABLE_SSE
- if(gdwSysInfo & PROCSUPPORT_SSE)
+ if(GetProcSupport() & PROCSUPPORT_SSE)
{
SSE_StereoMixToFloat(pSrc, pOut1, pOut2, nCount, _i2fc);
return;
}
#endif // ENABLE_SSE
- #ifdef ENABLE_3DNOW
- if(gdwSysInfo & PROCSUPPORT_3DNOW)
+ #ifdef ENABLE_X86_AMD
+ if(GetProcSupport() & PROCSUPPORT_AMD_3DNOW)
{
AMD_StereoMixToFloat(pSrc, pOut1, pOut2, nCount, _i2fc);
return;
}
- #endif // ENABLE_3DNOW
+ #endif // ENABLE_X86_AMD
#ifdef ENABLE_X86
X86_StereoMixToFloat(pSrc, pOut1, pOut2, nCount, _i2fc);
@@ -468,13 +403,13 @@
void FloatToStereoMix(const float *pIn1, const float *pIn2, int *pOut, UINT nCount, const float _f2ic)
{
- #ifdef ENABLE_3DNOW
- if(gdwSysInfo & PROCSUPPORT_3DNOW)
+ #ifdef ENABLE_X86_AMD
+ if(GetProcSupport() & PROCSUPPORT_AMD_3DNOW)
{
AMD_FloatToStereoMix(pIn1, pIn2, pOut, nCount, _f2ic);
return;
}
- #endif // ENABLE_3DNOW
+ #endif // ENABLE_X86_AMD
#ifdef ENABLE_X86
X86_FloatToStereoMix(pIn1, pIn2, pOut, nCount, _f2ic);
@@ -489,19 +424,19 @@
{
#ifdef ENABLE_SSE
- if(gdwSysInfo & PROCSUPPORT_SSE)
+ if(GetProcSupport() & PROCSUPPORT_SSE)
{
SSE_MonoMixToFloat(pSrc, pOut, nCount, _i2fc);
return;
}
#endif // ENABLE_SSE
- #ifdef ENABLE_3DNOW
- if(gdwSysInfo & PROCSUPPORT_3DNOW)
+ #ifdef ENABLE_X86_AMD
+ if(GetProcSupport() & PROCSUPPORT_AMD_3DNOW)
{
AMD_MonoMixToFloat(pSrc, pOut, nCount, _i2fc);
return;
}
- #endif // ENABLE_3DNOW
+ #endif // ENABLE_X86_AMD
#ifdef ENABLE_X86
X86_MonoMixToFloat(pSrc, pOut, nCount, _i2fc);
@@ -515,13 +450,13 @@
void FloatToMonoMix(const float *pIn, int *pOut, UINT nCount, const float _f2ic)
{
- #ifdef ENABLE_3DNOW
- if(gdwSysInfo & PROCSUPPORT_3DNOW)
+ #ifdef ENABLE_X86_AMD
+ if(GetProcSupport() & PROCSUPPORT_AMD_3DNOW)
{
AMD_FloatToMonoMix(pIn, pOut, nCount, _f2ic);
return;
}
- #endif // ENABLE_3DNOW
+ #endif // ENABLE_X86_AMD
#ifdef ENABLE_X86
X86_FloatToMonoMix(pIn, pOut, nCount, _f2ic);
Modified: trunk/OpenMPT/soundlib/Snd_defs.h
===================================================================
--- trunk/OpenMPT/soundlib/Snd_defs.h 2013-06-06 23:16:28 UTC (rev 2310)
+++ trunk/OpenMPT/soundlib/Snd_defs.h 2013-06-07 00:18:17 UTC (rev 2311)
@@ -242,15 +242,6 @@
#define DNA_NOTEFADE 2
-#ifdef ENABLE_ASM
-#define PROCSUPPORT_CPUID 0x01
-#define PROCSUPPORT_MMX 0x02 // Processor supports MMX instructions
-#define PROCSUPPORT_MMXEX 0x04 // Processor supports AMD MMX extensions
-#define PROCSUPPORT_3DNOW 0x08 // Processor supports AMD 3DNow! instructions
-#define PROCSUPPORT_SSE 0x10 // Processor supports SSE instructions
-#endif
-
-
// Module flags
enum SongFlags
{
Modified: trunk/OpenMPT/soundlib/Sndfile.h
===================================================================
--- trunk/OpenMPT/soundlib/Sndfile.h 2013-06-06 23:16:28 UTC (rev 2310)
+++ trunk/OpenMPT/soundlib/Sndfile.h 2013-06-07 00:18:17 UTC (rev 2311)
@@ -186,12 +186,6 @@
#endif // MODPLUG_TRACKER
-#ifdef ENABLE_ASM
-void InitProcSupport();
-uint32 GetProcSupport();
-#endif
-
-
void StereoMixToFloat(const int *pSrc, float *pOut1, float *pOut2, UINT nCount, const float _i2fc);
void FloatToStereoMix(const float *pIn1, const float *pIn2, int *pOut, UINT nCount, const float _f2ic);
void MonoMixToFloat(const int *pSrc, float *pOut, UINT nCount, const float _i2fc);
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
