Menu
▾
▴
[OMPT-svn-commit] SF.net SVN: modplug:[3765] trunk/OpenMPT
|
From: <sag...@us...> - 2014-02-23 22:43:46
|
Revision: 3765
http://sourceforge.net/p/modplug/code/3765
Author: saga-games
Date: 2014-02-23 22:43:39 +0000 (Sun, 23 Feb 2014)
Log Message:
-----------
[Imp] Add SSE3 implementation for sample autotuner (for big samples, this doubles the processing speed for me)
[Mod] Added command line switch /noAssembly which turns off MMX/SSE/3DNow!-accelerated code
Modified Paths:
--------------
trunk/OpenMPT/common/BuildSettings.h
trunk/OpenMPT/mptrack/Autotune.cpp
trunk/OpenMPT/mptrack/Autotune.h
trunk/OpenMPT/mptrack/Mptrack.cpp
Modified: trunk/OpenMPT/common/BuildSettings.h
===================================================================
--- trunk/OpenMPT/common/BuildSettings.h 2014-02-23 20:45:52 UTC (rev 3764)
+++ trunk/OpenMPT/common/BuildSettings.h 2014-02-23 22:43:39 UTC (rev 3765)
@@ -40,7 +40,7 @@
#ifdef _M_IX86
-// Generate general x86 inline assembly.
+// Generate general x86 inline assembly / intrinsics.
#define ENABLE_X86
// Generate inline assembly using MMX instructions (only used when the CPU supports it).
@@ -49,7 +49,7 @@
// Generate inline assembly using SSE instructions (only used when the CPU supports it).
#define ENABLE_SSE
-// Generate inline assembly using SSE2 instructions (only used when the CPU supports it).
+// Generate inline assembly using SSE2/SSE3 instructions (only used when the CPU supports it).
#define ENABLE_SSE2
// Generate inline assembly using AMD specific instruction set extensions (only used when the CPU supports it).
@@ -57,10 +57,10 @@
#elif defined(_M_X64)
-// Generate general x64 inline assembly.
+// Generate general x64 inline assembly / intrinsics.
#define ENABLE_X64
-// Generate inline assembly using SSE2 instructions (only used when the CPU supports it).
+// Generate inline assembly using SSE2/SSE3 instructions (only used when the CPU supports it).
#define ENABLE_SSE2
#endif
Modified: trunk/OpenMPT/mptrack/Autotune.cpp
===================================================================
--- trunk/OpenMPT/mptrack/Autotune.cpp 2014-02-23 20:45:52 UTC (rev 3764)
+++ trunk/OpenMPT/mptrack/Autotune.cpp 2014-02-23 22:43:39 UTC (rev 3765)
@@ -14,6 +14,9 @@
#include "../common/thread.h"
#include "../soundlib/Sndfile.h"
#include "Autotune.h"
+#ifdef ENABLE_SSE2
+#include <emmintrin.h>
+#endif
// The more bins, the more autocorrelations are done and the more precise the result is.
@@ -66,16 +69,16 @@
const T* sample = origSample + pos;
- int16 data = 0; // Enough for 256 channels... :)
+ int32 data = 0; // More than enough for 256 channels... :)
for(uint8 chn = 0; chn < channels; chn++)
{
// We only want the MSB.
- data += static_cast<int16>(sample[chn] >> ((sizeof(T) - 1) * 8));
+ data += static_cast<int32>(sample[chn] >> ((sizeof(T) - 1) * 8));
}
data /= channels;
- sampleData[i] = static_cast<int8>(data);
+ sampleData[i] = static_cast<int16>(data);
}
}
@@ -109,12 +112,13 @@
// We should analyse at least a one second (= GetSampleRate() samples) long sample.
sampleLength = std::max<SmpLength>(sampleLoopEnd, sample.GetSampleRate(modType)) + maxShift;
+ sampleLength = (sampleLength + 7) & ~7;
if(sampleData != nullptr)
{
delete[] sampleData;
}
- sampleData = new (std::nothrow) int8[sampleLength];
+ sampleData = new int16[sampleLength];
if(sampleData == nullptr)
{
return false;
@@ -148,18 +152,21 @@
{
std::vector<uint64> histogram;
double pitchReference;
- int8 *sampleData;
+ int16 *sampleData;
SmpLength processLength;
uint32 sampleFreq;
int startNote, endNote;
-
};
+
DWORD WINAPI Autotune::AutotuneThread(void *i)
//--------------------------------------------
{
AutotuneThreadData &info = *static_cast<AutotuneThreadData *>(i);
info.histogram.resize(HISTORY_BINS, 0);
+#ifdef ENABLE_SSE2
+ const bool useSSE = (ProcSupport & PROCSUPPORT_SSE3) != 0;
+#endif
// Do autocorrelation and save results in a note histogram (restriced to one octave).
for(int note = info.startNote, noteBin = note; note < info.endNote; note++, noteBin++)
@@ -173,13 +180,35 @@
const SmpLength autocorrShift = NoteToShift(info.sampleFreq, note, info.pitchReference);
uint64 autocorrSum = 0;
- const int8 *normalData = info.sampleData;
- const int8 *shiftedData = info.sampleData + autocorrShift;
- // Add up squared differences of all values
- for(SmpLength i = info.processLength; i != 0; i--, normalData++, shiftedData++)
+
+#ifdef ENABLE_SSE2
+ if(useSSE)
{
- autocorrSum += (*normalData - *shiftedData) * (*normalData - *shiftedData);
+ const __m128i *normalData = reinterpret_cast<const __m128i *>(info.sampleData);
+ const __m128i *shiftedData = reinterpret_cast<const __m128i *>(info.sampleData + autocorrShift);
+ for(SmpLength i = info.processLength / 8; i != 0; i--)
+ {
+ __m128i normal = _mm_load_si128(normalData++);
+ __m128i shifted = _mm_loadu_si128(shiftedData++);
+ __m128i diff = _mm_sub_epi16(normal, shifted); // 8 16-bit differences
+ __m128i squares = _mm_madd_epi16(diff, diff); // Multiply and add: 4 32-bit squares
+ squares = _mm_hadd_epi32(squares, squares); // This is SSE3!
+ squares = _mm_hadd_epi32(squares, squares);
+ autocorrSum += _mm_cvtsi128_si32(squares);
+ //autocorrSum += squares.m128i_i32[0] +squares.m128i_i32[1] + squares.m128i_i32[2] + squares.m128i_i32[3]; // For SSE2 only
+ }
+ } else
+#endif
+ {
+ const int16 *normalData = info.sampleData;
+ const int16 *shiftedData = info.sampleData + autocorrShift;
+ // Add up squared differences of all values
+ for(SmpLength i = info.processLength; i != 0; i--, normalData++, shiftedData++)
+ {
+ autocorrSum += (*normalData - *shiftedData) * (*normalData - *shiftedData);
+ }
}
+
info.histogram[noteBin] += autocorrSum;
}
return 0;
Modified: trunk/OpenMPT/mptrack/Autotune.h
===================================================================
--- trunk/OpenMPT/mptrack/Autotune.h 2014-02-23 20:45:52 UTC (rev 3764)
+++ trunk/OpenMPT/mptrack/Autotune.h 2014-02-23 22:43:39 UTC (rev 3765)
@@ -23,7 +23,7 @@
SmpLength selectionStart, selectionEnd;
- int8 *sampleData;
+ int16 *sampleData;
SmpLength sampleLength;
public:
Modified: trunk/OpenMPT/mptrack/Mptrack.cpp
===================================================================
--- trunk/OpenMPT/mptrack/Mptrack.cpp 2014-02-23 20:45:52 UTC (rev 3764)
+++ trunk/OpenMPT/mptrack/Mptrack.cpp 2014-02-23 22:43:39 UTC (rev 3765)
@@ -317,12 +317,13 @@
{
if ((lpszParam) && (bFlag))
{
- if (!lstrcmpi(lpszParam, "nologo")) { m_bShowSplash = FALSE; return; } else
- if (!lstrcmpi(lpszParam, "nodls")) { m_bNoDls = true; return; } else
- if (!lstrcmpi(lpszParam, "noplugs")) { m_bNoPlugins = true; return; } else
- if (!lstrcmpi(lpszParam, "portable")) { m_bPortable = true; return; } else
- if (!lstrcmpi(lpszParam, "noSettingsOnNewVersion")) { m_bNoSettingsOnNewVersion = true; return; }
- if (!lstrcmpi(lpszParam, "fullMemDump")) { ExceptionHandler::fullMemDump = true; return; }
+ if (!lstrcmpi(lpszParam, _T("nologo"))) { m_bShowSplash = FALSE; return; } else
+ if (!lstrcmpi(lpszParam, _T("nodls"))) { m_bNoDls = true; return; } else
+ if (!lstrcmpi(lpszParam, _T("noplugs"))) { m_bNoPlugins = true; return; } else
+ if (!lstrcmpi(lpszParam, _T("portable"))) { m_bPortable = true; return; } else
+ if (!lstrcmpi(lpszParam, _T("noSettingsOnNewVersion"))) { m_bNoSettingsOnNewVersion = true; return; }
+ if (!lstrcmpi(lpszParam, _T("fullMemDump"))) { ExceptionHandler::fullMemDump = true; return; }
+ if (!lstrcmpi(lpszParam, _T("noAssembly"))) { ProcSupport = 0; return; }
}
CCommandLineInfo::ParseParam(lpszParam, bFlag, bLast);
}
This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.
|
