Browse Source

Merge pull request #193 from Hoikas/sse

SSE3 Optimizations
Branan Purvine-Riley 13 years ago
parent
commit
a038af101d
  1. 6
      CMakeLists.txt
  2. 28
      Sources/Plasma/Apps/plClient/winmain.cpp
  3. 48
      Sources/Plasma/CoreLib/hsMatrix44.cpp
  4. 6
      Sources/Plasma/CoreLib/hsUtils.h
  5. 4
      Sources/Plasma/NucleusLib/pnSceneObject/plCoordinateInterface.cpp
  6. 7
      Sources/Plasma/PubUtilLib/plDrawable/plDrawableSpans.cpp
  7. 108
      Sources/Plasma/PubUtilLib/plPipeline/plDXPipeline.cpp

6
CMakeLists.txt

@ -84,6 +84,12 @@ if(MSVC)
add_definitions(-D_SCL_SECURE_NO_WARNINGS) add_definitions(-D_SCL_SECURE_NO_WARNINGS)
endif(MSVC) endif(MSVC)
# TODO: Maybe some kind of automated test here?
option(PLASMA_USE_SSE "Enable SSE optimizations?" ON)
if(PLASMA_USE_SSE)
add_definitions(-DHAVE_SSE)
endif(PLASMA_USE_SSE)
#TODO: Make the OpenSSL includes less promiscuous so this isn't needed #TODO: Make the OpenSSL includes less promiscuous so this isn't needed
include_directories(${OPENSSL_INCLUDE_DIR}) include_directories(${OPENSSL_INCLUDE_DIR})

28
Sources/Plasma/Apps/plClient/winmain.cpp

@ -49,6 +49,10 @@ You can contact Cyan Worlds, Inc. by email legal@cyan.com
#include <dmdfm.h> // Windows Load EXE into memory suff #include <dmdfm.h> // Windows Load EXE into memory suff
#endif #endif
#ifdef HAVE_SSE
# include <intrin.h>
#endif
#include <curl/curl.h> #include <curl/curl.h>
#include "HeadSpin.h" #include "HeadSpin.h"
@ -1440,11 +1444,35 @@ LONG WINAPI plCustomUnhandledExceptionFilter( struct _EXCEPTION_POINTERS *Except
return EXCEPTION_EXECUTE_HANDLER; return EXCEPTION_EXECUTE_HANDLER;
} }
bool CheckCPU()
{
const unsigned int sse3_flag = 0x00000001;
// (any other CPU features...)
int cpu_info[4];
__cpuid(cpu_info, 1);
#ifdef HAVE_SSE
if(cpu_info[2] & sse3_flag == 0)
return false;
#endif
// Insert additional feature checks here
return true;
}
#include "pfConsoleCore/pfConsoleEngine.h" #include "pfConsoleCore/pfConsoleEngine.h"
PF_CONSOLE_LINK_ALL() PF_CONSOLE_LINK_ALL()
int WINAPI WinMain(HINSTANCE hInst, HINSTANCE hPrevInst, LPSTR lpCmdLine, int nCmdShow) int WINAPI WinMain(HINSTANCE hInst, HINSTANCE hPrevInst, LPSTR lpCmdLine, int nCmdShow)
{ {
// Check to make sure we have a good CPU before getting started
if (!CheckCPU())
{
plString msg = plString::Format("Your processor does not support all of the features required to play %S", ProductLongName());
hsMessageBox(msg.c_str(), "Error", hsMessageBoxNormal, hsMessageBoxIconError);
return PARABLE_NORMAL_EXIT;
}
PF_CONSOLE_INIT_ALL() PF_CONSOLE_INIT_ALL()
// Set global handle // Set global handle

48
Sources/Plasma/CoreLib/hsMatrix44.cpp

@ -47,6 +47,10 @@ You can contact Cyan Worlds, Inc. by email legal@cyan.com
#include "hsStream.h" #include "hsStream.h"
#include <math.h> #include <math.h>
#ifdef HAVE_SSE
# include <smmintrin.h>
#endif
static hsMatrix44 myIdent = hsMatrix44().Reset(); static hsMatrix44 myIdent = hsMatrix44().Reset();
const hsMatrix44& hsMatrix44::IdentityMatrix() { return myIdent; } const hsMatrix44& hsMatrix44::IdentityMatrix() { return myIdent; }
@ -92,6 +96,18 @@ void hsMatrix44::DecompRigid(hsScalarTriple &translate, hsQuat &rotate) const
rotate.QuatFromMatrix44(*this); rotate.QuatFromMatrix44(*this);
} }
#ifdef HAVE_SSE
# define MULTBEGIN(i) \
xmm[0] = _mm_loadu_ps(fMap[i]);
# define MULTCELL(i, j) \
xmm[1] = _mm_set_ps(b.fMap[3][j], b.fMap[2][j], b.fMap[1][j], b.fMap[0][j]); \
xmm[j+2] = _mm_mul_ps(xmm[0], xmm[1]);
# define MULTFINISH(i) \
xmm[6] = _mm_hadd_ps(xmm[2], xmm[3]); \
xmm[7] = _mm_hadd_ps(xmm[4], xmm[5]); \
xmm[1] = _mm_hadd_ps(xmm[6], xmm[7]); \
_mm_storeu_ps(c.fMap[i], xmm[1]);
#endif
hsMatrix44 hsMatrix44::operator*(const hsMatrix44& b) const hsMatrix44 hsMatrix44::operator*(const hsMatrix44& b) const
{ {
@ -108,6 +124,37 @@ hsMatrix44 hsMatrix44::operator*(const hsMatrix44& b) const
if( b.fFlags & hsMatrix44::kIsIdent ) if( b.fFlags & hsMatrix44::kIsIdent )
return *this; return *this;
#ifdef HAVE_SSE
__m128 xmm[8];
MULTBEGIN(0);
MULTCELL(0, 0);
MULTCELL(0, 1);
MULTCELL(0, 2);
MULTCELL(0, 3);
MULTFINISH(0);
MULTBEGIN(1);
MULTCELL(1, 0);
MULTCELL(1, 1);
MULTCELL(1, 2);
MULTCELL(1, 3);
MULTFINISH(1);
MULTBEGIN(2);
MULTCELL(2, 0);
MULTCELL(2, 1);
MULTCELL(2, 2);
MULTCELL(2, 3);
MULTFINISH(2);
MULTBEGIN(3);
MULTCELL(3, 0);
MULTCELL(3, 1);
MULTCELL(3, 2);
MULTCELL(3, 3);
MULTFINISH(3);
#else
c.fMap[0][0] = (fMap[0][0] * b.fMap[0][0]) + (fMap[0][1] * b.fMap[1][0]) + (fMap[0][2] * b.fMap[2][0]) + (fMap[0][3] * b.fMap[3][0]); c.fMap[0][0] = (fMap[0][0] * b.fMap[0][0]) + (fMap[0][1] * b.fMap[1][0]) + (fMap[0][2] * b.fMap[2][0]) + (fMap[0][3] * b.fMap[3][0]);
c.fMap[0][1] = (fMap[0][0] * b.fMap[0][1]) + (fMap[0][1] * b.fMap[1][1]) + (fMap[0][2] * b.fMap[2][1]) + (fMap[0][3] * b.fMap[3][1]); c.fMap[0][1] = (fMap[0][0] * b.fMap[0][1]) + (fMap[0][1] * b.fMap[1][1]) + (fMap[0][2] * b.fMap[2][1]) + (fMap[0][3] * b.fMap[3][1]);
c.fMap[0][2] = (fMap[0][0] * b.fMap[0][2]) + (fMap[0][1] * b.fMap[1][2]) + (fMap[0][2] * b.fMap[2][2]) + (fMap[0][3] * b.fMap[3][2]); c.fMap[0][2] = (fMap[0][0] * b.fMap[0][2]) + (fMap[0][1] * b.fMap[1][2]) + (fMap[0][2] * b.fMap[2][2]) + (fMap[0][3] * b.fMap[3][2]);
@ -127,6 +174,7 @@ hsMatrix44 hsMatrix44::operator*(const hsMatrix44& b) const
c.fMap[3][1] = (fMap[3][0] * b.fMap[0][1]) + (fMap[3][1] * b.fMap[1][1]) + (fMap[3][2] * b.fMap[2][1]) + (fMap[3][3] * b.fMap[3][1]); c.fMap[3][1] = (fMap[3][0] * b.fMap[0][1]) + (fMap[3][1] * b.fMap[1][1]) + (fMap[3][2] * b.fMap[2][1]) + (fMap[3][3] * b.fMap[3][1]);
c.fMap[3][2] = (fMap[3][0] * b.fMap[0][2]) + (fMap[3][1] * b.fMap[1][2]) + (fMap[3][2] * b.fMap[2][2]) + (fMap[3][3] * b.fMap[3][2]); c.fMap[3][2] = (fMap[3][0] * b.fMap[0][2]) + (fMap[3][1] * b.fMap[1][2]) + (fMap[3][2] * b.fMap[2][2]) + (fMap[3][3] * b.fMap[3][2]);
c.fMap[3][3] = (fMap[3][0] * b.fMap[0][3]) + (fMap[3][1] * b.fMap[1][3]) + (fMap[3][2] * b.fMap[2][3]) + (fMap[3][3] * b.fMap[3][3]); c.fMap[3][3] = (fMap[3][0] * b.fMap[0][3]) + (fMap[3][1] * b.fMap[1][3]) + (fMap[3][2] * b.fMap[2][3]) + (fMap[3][3] * b.fMap[3][3]);
#endif
return c; return c;
} }

6
Sources/Plasma/CoreLib/hsUtils.h

@ -177,6 +177,12 @@ inline float hsRadiansToDegrees(float rad) { return float(rad * (180 / M_PI)); }
#include <new> #include <new>
#define NEWZERO(t) new(calloc(sizeof(t), 1)) t #define NEWZERO(t) new(calloc(sizeof(t), 1)) t
#ifdef _MSC_VER
# define ALIGN(n) __declspec(align(n))
#else
# define ALIGN(n) __atribute__(aligned(n))
#endif
///////////////////////////// /////////////////////////////
// Physical memory functions // Physical memory functions
///////////////////////////// /////////////////////////////

4
Sources/Plasma/NucleusLib/pnSceneObject/plCoordinateInterface.cpp

@ -380,6 +380,7 @@ plProfile_CreateTimer(" CIRecalcT", "Object", CIRecalcT);
plProfile_CreateTimer(" CIDirtyT", "Object", CIDirtyT); plProfile_CreateTimer(" CIDirtyT", "Object", CIDirtyT);
plProfile_CreateTimer(" CISetT", "Object", CISetT); plProfile_CreateTimer(" CISetT", "Object", CISetT);
#ifndef HAVE_SSE
static inline hsMatrix44 IMatrixMul34(const hsMatrix44& lhs, const hsMatrix44& rhs) static inline hsMatrix44 IMatrixMul34(const hsMatrix44& lhs, const hsMatrix44& rhs)
{ {
hsMatrix44 ret; hsMatrix44 ret;
@ -440,6 +441,7 @@ static inline hsMatrix44 IMatrixMul34(const hsMatrix44& lhs, const hsMatrix44& r
return ret; return ret;
} }
#endif // HAVE_SSE
void plCoordinateInterface::IRecalcTransforms() void plCoordinateInterface::IRecalcTransforms()
{ {
@ -447,7 +449,7 @@ void plCoordinateInterface::IRecalcTransforms()
plProfile_BeginTiming(CIRecalcT); plProfile_BeginTiming(CIRecalcT);
if( fParent ) if( fParent )
{ {
#if 0 #ifdef HAVE_SSE
fLocalToWorld = fParent->GetLocalToWorld() * fLocalToParent; fLocalToWorld = fParent->GetLocalToWorld() * fLocalToParent;
fWorldToLocal = fParentToLocal * fParent->GetWorldToLocal(); fWorldToLocal = fParentToLocal * fParent->GetWorldToLocal();
#else #else

7
Sources/Plasma/PubUtilLib/plDrawable/plDrawableSpans.cpp

@ -416,7 +416,7 @@ hsBool plDrawableSpans::IBoundsInvalid(const hsBounds3Ext& bnd) const
} }
//// SetTransform //////////////////////////////////////////////////////////// //// SetTransform ////////////////////////////////////////////////////////////
#ifndef HAVE_SSE
static inline hsMatrix44 IMatrixMul34(const hsMatrix44& lhs, const hsMatrix44& rhs) static inline hsMatrix44 IMatrixMul34(const hsMatrix44& lhs, const hsMatrix44& rhs)
{ {
hsMatrix44 ret; hsMatrix44 ret;
@ -477,6 +477,7 @@ static inline hsMatrix44 IMatrixMul34(const hsMatrix44& lhs, const hsMatrix44& r
return ret; return ret;
} }
#endif
#ifdef MF_TEST_UPDATE #ifdef MF_TEST_UPDATE
plProfile_CreateCounter("DSSetTrans", "Update", DSSetTrans); plProfile_CreateCounter("DSSetTrans", "Update", DSSetTrans);
@ -520,13 +521,13 @@ plDrawable& plDrawableSpans::SetTransform( uint32_t index, const hsMatrix44& l2w
#endif // MF_TEST_UPDATE #endif // MF_TEST_UPDATE
for( i = 0; i < spans->GetCount(); i++ ) for( i = 0; i < spans->GetCount(); i++ )
{ {
#if 0 #ifdef HAVE_SSE
fLocalToWorlds[ (*spans)[ i ] ] = l2w * fLocalToBones[ (*spans)[ i ] ]; fLocalToWorlds[ (*spans)[ i ] ] = l2w * fLocalToBones[ (*spans)[ i ] ];
fWorldToLocals[ (*spans)[ i ] ] = fBoneToLocals[ (*spans)[ i ] ] * w2l; fWorldToLocals[ (*spans)[ i ] ] = fBoneToLocals[ (*spans)[ i ] ] * w2l;
#else #else
fLocalToWorlds[ (*spans)[ i ] ] = IMatrixMul34(l2w, fLocalToBones[ (*spans)[ i ] ]); fLocalToWorlds[ (*spans)[ i ] ] = IMatrixMul34(l2w, fLocalToBones[ (*spans)[ i ] ]);
fWorldToLocals[ (*spans)[ i ] ] = IMatrixMul34(fBoneToLocals[ (*spans)[ i ] ], w2l); fWorldToLocals[ (*spans)[ i ] ] = IMatrixMul34(fBoneToLocals[ (*spans)[ i ] ], w2l);
#endif #endif // HAVE_SSE
} }
#ifdef MF_TEST_UPDATE #ifdef MF_TEST_UPDATE
plProfile_EndTiming(DSMatTransT); plProfile_EndTiming(DSMatTransT);

108
Sources/Plasma/PubUtilLib/plPipeline/plDXPipeline.cpp

@ -163,6 +163,10 @@ You can contact Cyan Worlds, Inc. by email legal@cyan.com
#include <algorithm> #include <algorithm>
#ifdef HAVE_SSE
# include <smmintrin.h>
#endif
//#define MF_TOSSER //#define MF_TOSSER
int mfCurrentTest = 100; int mfCurrentTest = 100;
@ -10523,39 +10527,77 @@ void plDXPipeline::LoadResources()
// Sorry about this, but it really did speed up the skinning. // Sorry about this, but it really did speed up the skinning.
// Just some macros for the inner loop of IBlendVertsIntoBuffer. // Just some macros for the inner loop of IBlendVertsIntoBuffer.
#define MATRIXMULTBEGIN(xfm, wgt) \ #ifdef HAVE_SSE
register float m00 = xfm.fMap[0][0]; \ # define MATRIXMULTBEGIN(xfm, wgt) \
register float m01 = xfm.fMap[0][1]; \ __m128 mc0, mc1, mc2, mwt, msr, _x, _y, _z, hbuf; \
register float m02 = xfm.fMap[0][2]; \ ALIGN(16) float hack[4]; \
register float m03 = xfm.fMap[0][3]; \ mc0 = _mm_loadu_ps(xfm.fMap[0]); \
register float m10 = xfm.fMap[1][0]; \ mc1 = _mm_loadu_ps(xfm.fMap[1]); \
register float m11 = xfm.fMap[1][1]; \ mc2 = _mm_loadu_ps(xfm.fMap[2]); \
register float m12 = xfm.fMap[1][2]; \ mwt = _mm_set_ps1(wgt);
register float m13 = xfm.fMap[1][3]; \ # define MATRIXMULTPOINTADD(dst, src) \
register float m20 = xfm.fMap[2][0]; \ msr = _mm_set_ps(1.f, src.fZ, src.fY, src.fX); \
register float m21 = xfm.fMap[2][1]; \ _x = _mm_mul_ps(_mm_mul_ps(mc0, msr), mwt); \
register float m22 = xfm.fMap[2][2]; \ _y = _mm_mul_ps(_mm_mul_ps(mc1, msr), mwt); \
register float m23 = xfm.fMap[2][3]; \ _z = _mm_mul_ps(_mm_mul_ps(mc2, msr), mwt); \
register float m_wgt = wgt; \ \
register float srcX, srcY, srcZ; hbuf = _mm_hadd_ps(_x, _y); \
hbuf = _mm_hadd_ps(hbuf, hbuf); \
#define MATRIXMULTPOINTADD(dst, src) \ _mm_store_ps(hack, hbuf); \
srcX = src.fX; \ dst.fX += hack[0]; \
srcY = src.fY; \ dst.fY += hack[1]; \
srcZ = src.fZ; \ hbuf = _mm_hadd_ps(_z, _z); \
\ hbuf = _mm_hadd_ps(hbuf, hbuf); \
dst.fX += (srcX * m00 + srcY * m01 + srcZ * m02 + m03) * m_wgt; \ _mm_store_ps(hack, hbuf); \
dst.fY += (srcX * m10 + srcY * m11 + srcZ * m12 + m13) * m_wgt; \ dst.fZ += hack[0];
dst.fZ += (srcX * m20 + srcY * m21 + srcZ * m22 + m23) * m_wgt; # define MATRIXMULTVECTORADD(dst, src) \
msr = _mm_set_ps(0.f, src.fZ, src.fY, src.fX); \
#define MATRIXMULTVECTORADD(dst, src) \ _x = _mm_mul_ps(_mm_mul_ps(mc0, msr), mwt); \
srcX = src.fX; \ _y = _mm_mul_ps(_mm_mul_ps(mc1, msr), mwt); \
srcY = src.fY; \ _z = _mm_mul_ps(_mm_mul_ps(mc2, msr), mwt); \
srcZ = src.fZ; \ \
\ hbuf = _mm_hadd_ps(_x, _y); \
dst.fX += (srcX * m00 + srcY * m01 + srcZ * m02) * m_wgt; \ hbuf = _mm_hadd_ps(hbuf, hbuf); \
dst.fY += (srcX * m10 + srcY * m11 + srcZ * m12) * m_wgt; \ _mm_store_ps(hack, hbuf); \
dst.fZ += (srcX * m20 + srcY * m21 + srcZ * m22) * m_wgt; dst.fX += hack[0]; \
dst.fY += hack[1]; \
hbuf = _mm_hadd_ps(_z, _z); \
hbuf = _mm_hadd_ps(hbuf, hbuf); \
_mm_store_ps(hack, hbuf); \
dst.fZ += hack[0];
#else
# define MATRIXMULTBEGIN(xfm, wgt) \
float m00 = xfm.fMap[0][0]; \
float m01 = xfm.fMap[0][1]; \
float m02 = xfm.fMap[0][2]; \
float m03 = xfm.fMap[0][3]; \
float m10 = xfm.fMap[1][0]; \
float m11 = xfm.fMap[1][1]; \
float m12 = xfm.fMap[1][2]; \
float m13 = xfm.fMap[1][3]; \
float m20 = xfm.fMap[2][0]; \
float m21 = xfm.fMap[2][1]; \
float m22 = xfm.fMap[2][2]; \
float m23 = xfm.fMap[2][3]; \
float m_wgt = wgt; \
float srcX, srcY, srcZ;
# define MATRIXMULTPOINTADD(dst, src) \
srcX = src.fX; \
srcY = src.fY; \
srcZ = src.fZ; \
\
dst.fX += (srcX * m00 + srcY * m01 + srcZ * m02 + m03) * m_wgt; \
dst.fY += (srcX * m10 + srcY * m11 + srcZ * m12 + m13) * m_wgt; \
dst.fZ += (srcX * m20 + srcY * m21 + srcZ * m22 + m23) * m_wgt;
# define MATRIXMULTVECTORADD(dst, src) \
srcX = src.fX; \
srcY = src.fY; \
srcZ = src.fZ; \
\
dst.fX += (srcX * m00 + srcY * m01 + srcZ * m02) * m_wgt; \
dst.fY += (srcX * m10 + srcY * m11 + srcZ * m12) * m_wgt; \
dst.fZ += (srcX * m20 + srcY * m21 + srcZ * m22) * m_wgt;
#endif // HAVE_SSE
// inlTESTPOINT ///////////////////////////////////////// // inlTESTPOINT /////////////////////////////////////////
// Update mins and maxs if destP is outside. // Update mins and maxs if destP is outside.

Loading…
Cancel
Save