/*****************************************************************************
The Dark Mod GPL Source Code

This file is part of the The Dark Mod Source Code, originally based
on the Doom 3 GPL Source Code as published in 2011.

The Dark Mod Source Code is free software: you can redistribute it
and/or modify it under the terms of the GNU General Public License as
published by the Free Software Foundation, either version 3 of the License,
or (at your option) any later version. For details, see LICENSE.TXT.

Project: The Dark Mod (http://www.thedarkmod.com/)

******************************************************************************/

#include "precompiled.h"
#pragma hdrstop

#include "Simd_Generic.h"
#include "Simd_SSE.h"
#include "Simd_SSE2.h"
#include "Simd_SSSE3.h"
#include "Simd_AVX.h"
#include "Simd_AVX2.h"
#include "Simd_IdAsm.h"

idSIMDProcessor *	SIMDProcessor = NULL;


idSIMDProcessor::~idSIMDProcessor() {}

/*
================
idSIMD::Init
================
*/
void idSIMD::Init( void ) {
	SIMDProcessor = CreateProcessor( "generic" ); 
}

/*
============
idSIMD::InitProcessor
============
*/
void idSIMD::InitProcessor( const char *module, const char *forceImpl ) {
	SIMDProcessor = CreateProcessor( forceImpl );
	int cpuid = SIMDProcessor->cpuid;

	// Print what we found to console
	idLib::common->Printf( "Found %s CPU, features:%s%s%s%s%s%s%s%s\n",
		// Vendor
		cpuid & CPUID_AMD ? "AMD" :
		cpuid & CPUID_INTEL ? "Intel" :
		cpuid & CPUID_GENERIC ? "Generic" :
		"Unsupported",
		// Flags
		cpuid & CPUID_SSE ? " SSE" : "",
		cpuid & CPUID_SSE2 ? " SSE2" : "",
		cpuid & CPUID_SSE3 ? " SSE3" : "",
		cpuid & CPUID_SSSE3 ? " SSSE3" : "",
		cpuid & CPUID_SSE41 ? " SSE41" : "",
		cpuid & CPUID_AVX ? " AVX" : "",
		cpuid & CPUID_AVX2 ? " AVX2" : "",
		cpuid & CPUID_FMA3 ? " FMA3" : ""
	);

	idLib::common->Printf( "%s using %s for SIMD processing.\n", module, SIMDProcessor->GetName() );
}

/*
============
idSIMD::CreateProcessor
============
*/
idSIMDProcessor *idSIMD::CreateProcessor( const char *forceImpl ) {
	int cpuid = idLib::sys->GetProcessorId();

	//stgatilov: force cpuid bits for SIMD choice if compiler macros are set
	//this is used for Elbrus compiler, which can cross-compile SSE intrinsics but has no CPUID instruction
	#ifdef __SSE__
		cpuid |= CPUID_SSE;
	#endif
	#ifdef __SSE2__
		cpuid |= CPUID_SSE2;
	#endif
	#ifdef __SSE3__
		cpuid |= CPUID_SSE3;
	#endif
	#ifdef __SSSE3__
		cpuid |= CPUID_SSSE3;
	#endif
	#ifdef __SSE4_1__ 
		cpuid |= CPUID_SSE41;
	#endif
	#ifdef __AVX__
		cpuid |= CPUID_AVX;
	#endif
	#ifdef __AVX2__
		cpuid |= CPUID_AVX2;
		cpuid |= CPUID_FMA3;
	#endif

	bool upToSSE = ( cpuid & CPUID_SSE );
	bool upToSSE2 = upToSSE && ( cpuid & CPUID_SSE2 );
	bool upToSSE3 = upToSSE2 && ( cpuid & CPUID_SSE3 );
	bool upToSSSE3 = upToSSE3 && ( cpuid & CPUID_SSSE3 );
	bool upToSSE41 = upToSSSE3 && ( cpuid & CPUID_SSE41 );
	bool upToAVX = upToSSE41 && ( cpuid & CPUID_AVX );
	bool upToAVX2 = upToAVX && ( cpuid & CPUID_AVX2 ) && ( cpuid & CPUID_FMA3 );

	idSIMDProcessor *result = nullptr;
	if (false) {
#ifdef ENABLE_SSE_PROCESSORS
	} else if ( upToAVX2 && (!forceImpl || idStr::Icmp(forceImpl, "AVX2") == 0) ) {
		result = new idSIMD_AVX2;
	} else if ( upToAVX && (!forceImpl || idStr::Icmp(forceImpl, "AVX") == 0) ) {
		result = new idSIMD_AVX;
#if defined(_MSC_VER) && defined(_M_IX86)	//stgatilov: this processor is defined only on MSVC 32-bit
	} else if ( upToSSE3 && (forceImpl && idStr::Icmp(forceImpl, "IdAsm") == 0) ) {
		result = new idSIMD_IdAsm;
#endif
	} else if ( upToSSSE3 && (!forceImpl || idStr::Icmp(forceImpl, "SSSE3") == 0) ) {
		result = new idSIMD_SSSE3;
	} else if ( upToSSE2 && (!forceImpl || idStr::Icmp(forceImpl, "SSE2") == 0) ) {
		result = new idSIMD_SSE2;
	} else if ( upToSSE && (!forceImpl || idStr::Icmp(forceImpl, "SSE") == 0) ) {
		result = new idSIMD_SSE;
#endif
	} else {
		result = new idSIMD_Generic;
	}

	result->cpuid = (cpuid_t)cpuid;
	return result;
}

/*
================
idSIMD::Shutdown
================
*/
void idSIMD::Shutdown( void ) {
	delete SIMDProcessor;
	SIMDProcessor = NULL;
}


//===============================================================
//
// Test code
//
//===============================================================

#define COUNT		1024		// data count
#define NUMTESTS	2048		// number of tests

#define RANDOM_SEED		1013904223L	//((int)idLib::sys->GetClockTicks())

idSIMDProcessor *p_simd;
idSIMDProcessor *p_generic;
int baseClocks = 0;

#define TIME_TYPE double
#define StartRecordTime( start ) start = sys->GetClockTicks();
#define StopRecordTime( end ) StartRecordTime( end )
#define GetBest( start, end, best )			\
	if ( !best || end - start < best ) {	\
		best = end - start;					\
	}


/*
============
PrintClocks
============
*/
void PrintClocks( const char *string, const int dataCount, TIME_TYPE clocks, TIME_TYPE otherClocks = 0 ) {
	int i;

	idLib::common->Printf( string );
	for ( i = idStr::LengthWithoutColors( string ); i < 48; i++ ) {
		idLib::common->Printf( " " );
	}
	clocks -= baseClocks;
	if ( otherClocks && clocks ) {
		otherClocks -= baseClocks;
		int p = ( int )( ( float )( otherClocks - clocks ) * 100.0f / ( float ) otherClocks );
		idLib::common->Printf( "c = %4d, clcks = %5d, %d%%\n", int(dataCount), int(clocks), p );
	} else {
		idLib::common->Printf( "c = %4d, clcks = %5d\n", int(dataCount), int(clocks) );
	}
}

/*
============
GetBaseClocks
============
*/
void GetBaseClocks( void ) {
	int i, start, end, bestClocks;

	bestClocks = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
	}
	baseClocks = bestClocks;
}

/*
============
TestAdd
============
*/
void TestAdd( void ) {
	int i;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( float fdst0[COUNT] );
	ALIGN16( float fdst1[COUNT] );
	ALIGN16( float fsrc0[COUNT] );
	ALIGN16( float fsrc1[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		fsrc0[i] = srnd.CRandomFloat() * 10.0f;
		fsrc1[i] = srnd.CRandomFloat() * 10.0f;
	}
	idLib::common->Printf( "====================================\n" );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Add( fdst0, 4.0f, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Add( float + float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Add( fdst1, 4.0f, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Add( float + float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Add( fdst0, fsrc0, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Add( float[] + float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Add( fdst1, fsrc0, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Add( float[] + float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestSub
============
*/
void TestSub( void ) {
	int i;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( float fdst0[COUNT] );
	ALIGN16( float fdst1[COUNT] );
	ALIGN16( float fsrc0[COUNT] );
	ALIGN16( float fsrc1[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		fsrc0[i] = srnd.CRandomFloat() * 10.0f;
		fsrc1[i] = srnd.CRandomFloat() * 10.0f;
	}
	idLib::common->Printf( "====================================\n" );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Sub( fdst0, 4.0f, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Sub( float + float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Sub( fdst1, 4.0f, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Sub( float + float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Sub( fdst0, fsrc0, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Sub( float[] + float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Sub( fdst1, fsrc0, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Sub( float[] + float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestMul
============
*/
void TestMul( void ) {
	int i;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( float fdst0[COUNT] );
	ALIGN16( float fdst1[COUNT] );
	ALIGN16( float fsrc0[COUNT] );
	ALIGN16( float fsrc1[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		fsrc0[i] = srnd.CRandomFloat() * 10.0f;
		fsrc1[i] = srnd.CRandomFloat() * 10.0f;
	}
	idLib::common->Printf( "====================================\n" );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Mul( fdst0, 4.0f, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Mul( float * float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Mul( fdst1, 4.0f, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Mul( float * float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );


	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Mul( fdst0, fsrc0, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Mul( float[] * float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Mul( fdst1, fsrc0, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Mul( float[] * float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestDiv
============
*/
void TestDiv( void ) {
	int i;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( float fdst0[COUNT] );
	ALIGN16( float fdst1[COUNT] );
	ALIGN16( float fsrc0[COUNT] );
	ALIGN16( float fsrc1[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		fsrc0[i] = srnd.CRandomFloat() * 10.0f;
		do {
			fsrc1[i] = srnd.CRandomFloat() * 10.0f;
		} while ( idMath::Fabs( fsrc1[i] ) < 0.1f );
	}
	idLib::common->Printf( "====================================\n" );


	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Div( fdst0, 4.0f, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Div( float * float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Div( fdst1, 4.0f, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Div( float * float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Div( fdst0, fsrc0, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Div( float[] * float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Div( fdst1, fsrc0, fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-3f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Div( float[] * float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

#define TEST_VECTOR_LENGTH_START	16
#define TEST_VECTOR_LENGTH_END		25
#define TEST_VECTOR_LENGTH_STEP		3

/*
============
TestMulAdd
============
*/
void TestMulAdd( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( float fdst0[COUNT] );
	ALIGN16( float fdst1[COUNT] );
	ALIGN16( float fsrc0[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		fsrc0[i] = srnd.CRandomFloat() * 10.0f;
	}
	idLib::common->Printf( "====================================\n" );

	for ( j = TEST_VECTOR_LENGTH_START; j < TEST_VECTOR_LENGTH_END && j < COUNT; j += TEST_VECTOR_LENGTH_STEP ) {

		bestClocksGeneric = 0;
		for ( i = 0; i < NUMTESTS; i++ ) {
			for ( int k = 0; k < COUNT; k++ ) {
				fdst0[k] = k;
			}
			StartRecordTime( start );
			p_generic->MulAdd( fdst0, 0.123f, fsrc0, j );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		PrintClocks( va( "generic->MulAdd( float * float[%2d] )", j ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( i = 0; i < NUMTESTS; i++ ) {
			for ( int k = 0; k < COUNT; k++ ) {
				fdst1[k] = k;
			}
			StartRecordTime( start );
			p_simd->MulAdd( fdst1, 0.123f, fsrc0, j );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}

		for ( i = 0; i < COUNT; i++ ) {
			if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
				break;
			}
		}
		result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MulAdd( float * float[%2d] ) %s", j, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
}

/*
============
TestMulSub
============
*/
void TestMulSub( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( float fdst0[COUNT] );
	ALIGN16( float fdst1[COUNT] );
	ALIGN16( float fsrc0[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		fsrc0[i] = srnd.CRandomFloat() * 10.0f;
	}
	idLib::common->Printf( "====================================\n" );

	for ( j = TEST_VECTOR_LENGTH_START; j < TEST_VECTOR_LENGTH_END && j < COUNT; j += TEST_VECTOR_LENGTH_STEP ) {

		bestClocksGeneric = 0;
		for ( i = 0; i < NUMTESTS; i++ ) {
			for ( int k = 0; k < COUNT; k++ ) {
				fdst0[k] = k;
			}
			StartRecordTime( start );
			p_generic->MulSub( fdst0, 0.123f, fsrc0, j );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		PrintClocks( va( "generic->MulSub( float * float[%2d] )", j ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( i = 0; i < NUMTESTS; i++ ) {
			for ( int k = 0; k < COUNT; k++ ) {
				fdst1[k] = k;
			}
			StartRecordTime( start );
			p_simd->MulSub( fdst1, 0.123f, fsrc0, j );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}

		for ( i = 0; i < COUNT; i++ ) {
			if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
				break;
			}
		}
		result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MulSub( float * float[%2d] ) %s", j, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
}

/*
============
TestDot
============
*/
void TestDot( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( float fdst0[COUNT] );
	ALIGN16( float fdst1[COUNT] );
	ALIGN16( float fsrc0[COUNT] );
	ALIGN16( float fsrc1[COUNT] );
	ALIGN16( idVec3 v3src0[COUNT] );
	ALIGN16( idVec3 v3src1[COUNT] );
	ALIGN16( idVec3 v3constant )( 1.0f, 2.0f, 3.0f );
	ALIGN16( idPlane v4src0[COUNT] );
	ALIGN16( idPlane v4constant )( 1.0f, 2.0f, 3.0f, 4.0f );
	ALIGN16( idDrawVert drawVerts[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		fsrc0[i] = srnd.CRandomFloat() * 10.0f;
		fsrc1[i] = srnd.CRandomFloat() * 10.0f;
		v3src0[i][0] = srnd.CRandomFloat() * 10.0f;
		v3src0[i][1] = srnd.CRandomFloat() * 10.0f;
		v3src0[i][2] = srnd.CRandomFloat() * 10.0f;
		v3src1[i][0] = srnd.CRandomFloat() * 10.0f;
		v3src1[i][1] = srnd.CRandomFloat() * 10.0f;
		v3src1[i][2] = srnd.CRandomFloat() * 10.0f;
		v4src0[i] = v3src0[i];
		v4src0[i][3] = srnd.CRandomFloat() * 10.0f;
		drawVerts[i].xyz = v3src0[i];
	}
	idLib::common->Printf( "====================================\n" );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Dot( fdst0, v3constant, v3src0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Dot( idVec3 * idVec3[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Dot( fdst1, v3constant, v3src0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Dot( idVec3 * idVec3[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Dot( fdst0, v3constant, v4src0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Dot( idVec3 * idPlane[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Dot( fdst1, v3constant, v4src0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Dot( idVec3 * idPlane[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );


	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Dot( fdst0, v3constant, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Dot( idVec3 * idDrawVert[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Dot( fdst1, v3constant, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Dot( idVec3 * idDrawVert[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );


	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Dot( fdst0, v4constant, v3src0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Dot( idPlane * idVec3[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Dot( fdst1, v4constant, v3src0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Dot( idPlane * idVec3[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );


	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Dot( fdst0, v4constant, v4src0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Dot( idPlane * idPlane[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Dot( fdst1, v4constant, v4src0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Dot( idPlane * idPlane[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );


	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Dot( fdst0, v4constant, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Dot( idPlane * idDrawVert[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Dot( fdst1, v4constant, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-5f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Dot( idPlane * idDrawVert[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );


	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Dot( fdst0, v3src0, v3src1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Dot( idVec3[] * idVec3[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Dot( fdst1, v3src0, v3src1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( idMath::Fabs( fdst0[i] - fdst1[i] ) > 1e-4f ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Dot( idVec3[] * idVec3[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	idLib::common->Printf( "====================================\n" );

	float dot1 = 0.0f, dot2 = 0.0f;
	for ( j = TEST_VECTOR_LENGTH_START; j < TEST_VECTOR_LENGTH_END && j < COUNT; j += TEST_VECTOR_LENGTH_STEP ) {

		bestClocksGeneric = 0;
		for ( i = 0; i < NUMTESTS; i++ ) {
			StartRecordTime( start );
			p_generic->Dot( dot1, fsrc0, fsrc1, j );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		PrintClocks( va( "generic->Dot( float[%2d] * float[%2d] )", j, j ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( i = 0; i < NUMTESTS; i++ ) {
			StartRecordTime( start );
			p_simd->Dot( dot2, fsrc0, fsrc1, j );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = idMath::Fabs( dot1 - dot2 ) < 1e-4f ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->Dot( float[%2d] * float[%2d] ) %s", j, j, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
}

/*
============
TestCompare
============
*/
void TestCompare( void ) {
	int i;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( float fsrc0[COUNT] );
	ALIGN16( byte bytedst[COUNT] );
	ALIGN16( byte bytedst2[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		fsrc0[i] = srnd.CRandomFloat() * 10.0f;
	}
	idLib::common->Printf( "====================================\n" );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->CmpGT( bytedst, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->CmpGT( float[] >= float )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->CmpGT( bytedst2, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( bytedst[i] != bytedst2[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->CmpGT( float[] >= float ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		memset( bytedst, 0, COUNT );
		StartRecordTime( start );
		p_generic->CmpGT( bytedst, 2, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->CmpGT( 2, float[] >= float )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		memset( bytedst2, 0, COUNT );
		StartRecordTime( start );
		p_simd->CmpGT( bytedst2, 2, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( bytedst[i] != bytedst2[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->CmpGT( 2, float[] >= float ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	// ======================

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->CmpGE( bytedst, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->CmpGE( float[] >= float )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->CmpGE( bytedst2, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( bytedst[i] != bytedst2[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->CmpGE( float[] >= float ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		memset( bytedst, 0, COUNT );
		StartRecordTime( start );
		p_generic->CmpGE( bytedst, 2, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->CmpGE( 2, float[] >= float )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		memset( bytedst2, 0, COUNT );
		StartRecordTime( start );
		p_simd->CmpGE( bytedst2, 2, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( bytedst[i] != bytedst2[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->CmpGE( 2, float[] >= float ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	// ======================

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->CmpLT( bytedst, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->CmpLT( float[] >= float )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->CmpLT( bytedst2, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( bytedst[i] != bytedst2[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->CmpLT( float[] >= float ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		memset( bytedst, 0, COUNT );
		StartRecordTime( start );
		p_generic->CmpLT( bytedst, 2, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->CmpLT( 2, float[] >= float )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		memset( bytedst2, 0, COUNT );
		StartRecordTime( start );
		p_simd->CmpLT( bytedst2, 2, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( bytedst[i] != bytedst2[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->CmpLT( 2, float[] >= float ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	// ======================

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->CmpLE( bytedst, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->CmpLE( float[] >= float )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->CmpLE( bytedst2, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( bytedst[i] != bytedst2[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->CmpLE( float[] >= float ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		memset( bytedst, 0, COUNT );
		StartRecordTime( start );
		p_generic->CmpLE( bytedst, 2, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->CmpLE( 2, float[] >= float )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		memset( bytedst2, 0, COUNT );
		StartRecordTime( start );
		p_simd->CmpLE( bytedst2, 2, fsrc0, 0.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( bytedst[i] != bytedst2[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->CmpLE( 2, float[] >= float ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestMinMax
============
*/
void TestMinMax( void ) {
	int i;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( float fsrc0[COUNT] );
	ALIGN16( idVec2 v2src0[COUNT] );
	ALIGN16( idVec3 v3src0[COUNT] );
	ALIGN16( idDrawVert drawVerts[COUNT] );
	ALIGN16( int indexes[COUNT] );
	float min = 0.0f, max = 0.0f, min2 = 0.0f, max2 = 0.0f;
	idVec2 v2min, v2max, v2min2, v2max2;
	idVec3 vmin, vmax, vmin2, vmax2;
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		fsrc0[i] = srnd.CRandomFloat() * 10.0f;
		v2src0[i][0] = srnd.CRandomFloat() * 10.0f;
		v2src0[i][1] = srnd.CRandomFloat() * 10.0f;
		v3src0[i][0] = srnd.CRandomFloat() * 10.0f;
		v3src0[i][1] = srnd.CRandomFloat() * 10.0f;
		v3src0[i][2] = srnd.CRandomFloat() * 10.0f;
		drawVerts[i].xyz = v3src0[i];
		indexes[i] = i;
	}
	idLib::common->Printf( "====================================\n" );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		min = idMath::INFINITY;
		max = -idMath::INFINITY;
		StartRecordTime( start );
		p_generic->MinMax( min, max, fsrc0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->MinMax( float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->MinMax( min2, max2, fsrc0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}
	result = ( min == min2 && max == max2 ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->MinMax( float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->MinMax( v2min, v2max, v2src0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->MinMax( idVec2[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->MinMax( v2min2, v2max2, v2src0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}
	result = ( v2min == v2min2 && v2max == v2max2 ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->MinMax( idVec2[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->MinMax( vmin, vmax, v3src0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->MinMax( idVec3[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->MinMax( vmin2, vmax2, v3src0, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}
	result = ( vmin == vmin2 && vmax == vmax2 ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->MinMax( idVec3[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->MinMax( vmin, vmax, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->MinMax( idDrawVert[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->MinMax( vmin2, vmax2, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}
	result = ( vmin == vmin2 && vmax == vmax2 ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->MinMax( idDrawVert[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->MinMax( vmin, vmax, drawVerts, indexes, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->MinMax( idDrawVert[], indexes[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->MinMax( vmin2, vmax2, drawVerts, indexes, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}
	result = ( vmin == vmin2 && vmax == vmax2 ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->MinMax( idDrawVert[], indexes[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestClamp
============
*/
void TestClamp( void ) {
	int i;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( float fdst0[COUNT] );
	ALIGN16( float fdst1[COUNT] );
	ALIGN16( float fsrc0[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		fsrc0[i] = srnd.CRandomFloat() * 10.0f;
	}
	idLib::common->Printf( "====================================\n" );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->Clamp( fdst0, fsrc0, -1.0f, 1.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Clamp( float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->Clamp( fdst1, fsrc0, -1.0f, 1.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( fdst0[i] != fdst1[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Clamp( float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );


	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->ClampMin( fdst0, fsrc0, -1.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->ClampMin( float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->ClampMin( fdst1, fsrc0, -1.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( fdst0[i] != fdst1[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->ClampMin( float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );


	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->ClampMax( fdst0, fsrc0, 1.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->ClampMax( float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->ClampMax( fdst1, fsrc0, 1.0f, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( fdst0[i] != fdst1[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->ClampMax( float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestMemcpy
============
*/
void TestMemcpy( void ) {
	int i, j;
	byte test0[8192];
	byte test1[8192];

	idRandom random( RANDOM_SEED );

	idLib::common->Printf( "====================================\n" );

	for ( i = 5; i < 8192; i += 31 ) {
		for ( j = 0; j < i; j++ ) {
			test0[j] = random.RandomInt( 255 );
		}
		p_simd->Memcpy( test1, test0, 8192 );
		for ( j = 0; j < i; j++ ) {
			if ( test1[j] != test0[j] ) {
				idLib::common->Printf( "   simd->Memcpy() " S_COLOR_RED "X\n" );
				return;
			}
		}
	}
	idLib::common->Printf( "   simd->Memcpy() ok\n" );
}

/*
============
TestMemset
============
*/
void TestMemset( void ) {
	int i, j, k;
	byte test[8192];

	for ( i = 0; i < 8192; i++ ) {
		test[i] = 0;
	}

	for ( i = 5; i < 8192; i += 31 ) {
		for ( j = -1; j <= 1; j++ ) {
			p_simd->Memset( test, j, i );
			for ( k = 0; k < i; k++ ) {
				if ( test[k] != ( byte )j ) {
					idLib::common->Printf( "   simd->Memset() " S_COLOR_RED "X\n" );
					return;
				}
			}
		}
	}
	idLib::common->Printf( "   simd->Memset() ok\n" );
}

#define	MATX_SIMD_EPSILON			1e-5f

/*
============
TestMatXMultiplyVecX
============
*/
void TestMatXMultiplyVecX( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	const char *result;
	idMatX mat;
	idVecX src( 6 );
	idVecX dst( 6 );
	idVecX tst( 6 );

	src[0] = 1.0f;
	src[1] = 2.0f;
	src[2] = 3.0f;
	src[3] = 4.0f;
	src[4] = 5.0f;
	src[5] = 6.0f;

	idLib::common->Printf( "================= NxN * Nx1 ===================\n" );

	for ( i = 1; i <= 6; i++ ) {
		mat.Random( i, i, RANDOM_SEED, -10.0f, 10.0f );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_generic->MatX_MultiplyVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;

		PrintClocks( va( "generic->MatX_MultiplyVecX %dx%d*%dx1", i, i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_simd->MatX_MultiplyVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_MultiplyVecX %dx%d*%dx1 %s", i, i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
	idLib::common->Printf( "================= Nx6 * 6x1 ===================\n" );

	for ( i = 1; i <= 6; i++ ) {
		mat.Random( i, 6, RANDOM_SEED, -10.0f, 10.0f );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_generic->MatX_MultiplyVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;

		PrintClocks( va( "generic->MatX_MultiplyVecX %dx6*6x1", i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_simd->MatX_MultiplyVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_MultiplyVecX %dx6*6x1 %s", i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
	idLib::common->Printf( "================= 6xN * Nx1 ===================\n" );

	for ( i = 1; i <= 6; i++ ) {
		mat.Random( 6, i, RANDOM_SEED, -10.0f, 10.0f );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_generic->MatX_MultiplyVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_MultiplyVecX 6x%d*%dx1", i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_simd->MatX_MultiplyVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_MultiplyVecX 6x%d*%dx1 %s", i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
}

/*
============
TestMatXMultiplyAddVecX
============
*/
void TestMatXMultiplyAddVecX( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	const char *result;
	idMatX mat;
	idVecX src( 6 );
	idVecX dst( 6 );
	idVecX tst( 6 );

	src[0] = 1.0f;
	src[1] = 2.0f;
	src[2] = 3.0f;
	src[3] = 4.0f;
	src[4] = 5.0f;
	src[5] = 6.0f;

	idLib::common->Printf( "================= NxN * Nx1 ===================\n" );

	for ( i = 1; i <= 6; i++ ) {
		mat.Random( i, i, RANDOM_SEED, -10.0f, 10.0f );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_generic->MatX_MultiplyAddVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_MultiplyAddVecX %dx%d*%dx1", i, i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_simd->MatX_MultiplyAddVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_MultiplyAddVecX %dx%d*%dx1 %s", i, i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
	idLib::common->Printf( "================= Nx6 * 6x1 ===================\n" );

	for ( i = 1; i <= 6; i++ ) {
		mat.Random( i, 6, RANDOM_SEED, -10.0f, 10.0f );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_generic->MatX_MultiplyAddVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_MultiplyAddVecX %dx6*6x1", i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_simd->MatX_MultiplyAddVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_MultiplyAddVecX %dx6*6x1 %s", i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
	idLib::common->Printf( "================= 6xN * Nx1 ===================\n" );

	for ( i = 1; i <= 6; i++ ) {
		mat.Random( 6, i, RANDOM_SEED, -10.0f, 10.0f );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_generic->MatX_MultiplyAddVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_MultiplyAddVecX 6x%d*%dx1", i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_simd->MatX_MultiplyAddVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_MultiplyAddVecX 6x%d*%dx1 %s", i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
}

/*
============
TestMatXTransposeMultiplyVecX
============
*/
void TestMatXTransposeMultiplyVecX( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	const char *result;
	idMatX mat;
	idVecX src( 6 );
	idVecX dst( 6 );
	idVecX tst( 6 );

	src[0] = 1.0f;
	src[1] = 2.0f;
	src[2] = 3.0f;
	src[3] = 4.0f;
	src[4] = 5.0f;
	src[5] = 6.0f;

	idLib::common->Printf( "================= Nx6 * Nx1 ===================\n" );

	for ( i = 1; i <= 6; i++ ) {
		mat.Random( i, 6, RANDOM_SEED, -10.0f, 10.0f );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_generic->MatX_TransposeMultiplyVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_TransposeMulVecX %dx6*%dx1", i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_simd->MatX_TransposeMultiplyVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_TransposeMulVecX %dx6*%dx1 %s", i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
	idLib::common->Printf( "================= 6xN * 6x1 ===================\n" );

	for ( i = 1; i <= 6; i++ ) {
		mat.Random( 6, i, RANDOM_SEED, -10.0f, 10.0f );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_generic->MatX_TransposeMultiplyVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_TransposeMulVecX 6x%d*6x1", i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_simd->MatX_TransposeMultiplyVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_TransposeMulVecX 6x%d*6x1 %s", i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
}

/*
============
TestMatXTransposeMultiplyAddVecX
============
*/
void TestMatXTransposeMultiplyAddVecX( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	const char *result;
	idMatX mat;
	idVecX src( 6 );
	idVecX dst( 6 );
	idVecX tst( 6 );

	src[0] = 1.0f;
	src[1] = 2.0f;
	src[2] = 3.0f;
	src[3] = 4.0f;
	src[4] = 5.0f;
	src[5] = 6.0f;

	idLib::common->Printf( "================= Nx6 * Nx1 ===================\n" );

	for ( i = 1; i <= 6; i++ ) {
		mat.Random( i, 6, RANDOM_SEED, -10.0f, 10.0f );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_generic->MatX_TransposeMultiplyAddVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_TransposeMulAddVecX %dx6*%dx1", i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_simd->MatX_TransposeMultiplyAddVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_TransposeMulAddVecX %dx6*%dx1 %s", i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}

	idLib::common->Printf( "================= 6xN * 6x1 ===================\n" );

	for ( i = 1; i <= 6; i++ ) {
		mat.Random( 6, i, RANDOM_SEED, -10.0f, 10.0f );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_generic->MatX_TransposeMultiplyAddVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_TransposeMulAddVecX 6x%d*6x1", i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			dst.Zero();
			StartRecordTime( start );
			p_simd->MatX_TransposeMultiplyAddVecX( dst, mat, src );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_TransposeMulAddVecX 6x%d*6x1 %s", i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
}

/*
============
TestMatXMultiplyMatX
============
*/
#define TEST_VALUE_RANGE			10.0f
#define	MATX_MATX_SIMD_EPSILON		1e-4f

void TestMatXMultiplyMatX( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	const char *result;
	idMatX m1, m2, dst, tst;

	idLib::common->Printf( "================= NxN * Nx6 ===================\n" );

	// NxN * Nx6
	for ( i = 1; i <= 5; i++ ) {
		m1.Random( i, i, RANDOM_SEED, -TEST_VALUE_RANGE, TEST_VALUE_RANGE );
		m2.Random( i, 6, RANDOM_SEED, -TEST_VALUE_RANGE, TEST_VALUE_RANGE );
		dst.SetSize( i, 6 );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_generic->MatX_MultiplyMatX( dst, m1, m2 );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_MultiplyMatX %dx%d*%dx6", i, i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_simd->MatX_MultiplyMatX( dst, m1, m2 );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_MultiplyMatX %dx%d*%dx6 %s", i, i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
	idLib::common->Printf( "================= 6xN * Nx6 ===================\n" );

	// 6xN * Nx6
	for ( i = 1; i <= 5; i++ ) {
		m1.Random( 6, i, RANDOM_SEED, -TEST_VALUE_RANGE, TEST_VALUE_RANGE );
		m2.Random( i, 6, RANDOM_SEED, -TEST_VALUE_RANGE, TEST_VALUE_RANGE );
		dst.SetSize( 6, 6 );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_generic->MatX_MultiplyMatX( dst, m1, m2 );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_MultiplyMatX 6x%d*%dx6", i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_simd->MatX_MultiplyMatX( dst, m1, m2 );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_MultiplyMatX 6x%d*%dx6 %s", i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
	idLib::common->Printf( "================= Nx6 * 6xN ===================\n" );

	// Nx6 * 6xN
	for ( i = 1; i <= 5; i++ ) {
		m1.Random( i, 6, RANDOM_SEED, -TEST_VALUE_RANGE, TEST_VALUE_RANGE );
		m2.Random( 6, i, RANDOM_SEED, -TEST_VALUE_RANGE, TEST_VALUE_RANGE );
		dst.SetSize( i, i );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_generic->MatX_MultiplyMatX( dst, m1, m2 );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_MultiplyMatX %dx6*6x%d", i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_simd->MatX_MultiplyMatX( dst, m1, m2 );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_MultiplyMatX %dx6*6x%d %s", i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
	idLib::common->Printf( "================= 6x6 * 6xN ===================\n" );

	// 6x6 * 6xN
	for ( i = 1; i <= 6; i++ ) {
		m1.Random( 6, 6, RANDOM_SEED, -TEST_VALUE_RANGE, TEST_VALUE_RANGE );
		m2.Random( 6, i, RANDOM_SEED, -TEST_VALUE_RANGE, TEST_VALUE_RANGE );
		dst.SetSize( 6, i );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_generic->MatX_MultiplyMatX( dst, m1, m2 );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_MultiplyMatX 6x6*6x%d", i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_simd->MatX_MultiplyMatX( dst, m1, m2 );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_MultiplyMatX 6x6*6x%d %s", i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
}

/*
============
TestMatXTransposeMultiplyMatX
============
*/
void TestMatXTransposeMultiplyMatX( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	const char *result;
	idMatX m1, m2, dst, tst;

	idLib::common->Printf( "================= Nx6 * NxN ===================\n" );

	// Nx6 * NxN
	for ( i = 1; i <= 5; i++ ) {
		m1.Random( i, 6, RANDOM_SEED, -TEST_VALUE_RANGE, TEST_VALUE_RANGE );
		m2.Random( i, i, RANDOM_SEED, -TEST_VALUE_RANGE, TEST_VALUE_RANGE );
		dst.SetSize( 6, i );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_generic->MatX_TransposeMultiplyMatX( dst, m1, m2 );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_TransMultiplyMatX %dx6*%dx%d", i, i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_simd->MatX_TransposeMultiplyMatX( dst, m1, m2 );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_TransMultiplyMatX %dx6*%dx%d %s", i, i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
	idLib::common->Printf( "================= 6xN * 6x6 ===================\n" );

	// 6xN * 6x6
	for ( i = 1; i <= 6; i++ ) {
		m1.Random( 6, i, RANDOM_SEED, -TEST_VALUE_RANGE, TEST_VALUE_RANGE );
		m2.Random( 6, 6, RANDOM_SEED, -TEST_VALUE_RANGE, TEST_VALUE_RANGE );
		dst.SetSize( i, 6 );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_generic->MatX_TransposeMultiplyMatX( dst, m1, m2 );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = dst;
		PrintClocks( va( "generic->MatX_TransMultiplyMatX 6x%d*6x6", i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_simd->MatX_TransposeMultiplyMatX( dst, m1, m2 );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = dst.Compare( tst, MATX_MATX_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_TransMultiplyMatX 6x%d*6x6 %s", i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
}

#define MATX_LTS_SIMD_EPSILON		1.0f
#define MATX_LTS_SOLVE_SIZE			25
#define MATX_LTS_SOLVE_START		1
#define MATX_LTS_SOLVE_STEP			7

/*
============
TestMatXLowerTriangularSolve
============
*/
void TestMatXLowerTriangularSolve( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	const char *result;
	idMatX L;
	idVecX x, b, tst;

	idLib::common->Printf( "====================================\n" );

	L.Random( MATX_LTS_SOLVE_SIZE, MATX_LTS_SOLVE_SIZE, 0, -1.0f, 1.0f );
	x.SetSize( MATX_LTS_SOLVE_SIZE );
	b.Random( MATX_LTS_SOLVE_SIZE, 0, -1.0f, 1.0f );

	for ( i = MATX_LTS_SOLVE_START; i < MATX_LTS_SOLVE_SIZE; i += MATX_LTS_SOLVE_STEP ) {

		x.Zero( i );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_generic->MatX_LowerTriangularSolve( L, x.ToFloatPtr(), b.ToFloatPtr(), i );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = x;
		x.Zero();

		PrintClocks( va( "generic->MatX_LowerTriangularSolve %dx%d", i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_simd->MatX_LowerTriangularSolve( L, x.ToFloatPtr(), b.ToFloatPtr(), i );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = x.Compare( tst, MATX_LTS_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_LowerTriangularSolve %dx%d %s", i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
}

/*
============
TestMatXLowerTriangularSolveTranspose
============
*/
void TestMatXLowerTriangularSolveTranspose( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	const char *result;
	idMatX L;
	idVecX x, b, tst;

	idLib::common->Printf( "====================================\n" );

	L.Random( MATX_LTS_SOLVE_SIZE, MATX_LTS_SOLVE_SIZE, 0, -1.0f, 1.0f );
	x.SetSize( MATX_LTS_SOLVE_SIZE );
	b.Random( MATX_LTS_SOLVE_SIZE, 0, -1.0f, 1.0f );

	for ( i = MATX_LTS_SOLVE_START; i < MATX_LTS_SOLVE_SIZE; i += MATX_LTS_SOLVE_STEP ) {

		x.Zero( i );

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_generic->MatX_LowerTriangularSolveTranspose( L, x.ToFloatPtr(), b.ToFloatPtr(), i );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		tst = x;
		x.Zero();

		PrintClocks( va( "generic->MatX_LowerTriangularSolveT %dx%d", i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			StartRecordTime( start );
			p_simd->MatX_LowerTriangularSolveTranspose( L, x.ToFloatPtr(), b.ToFloatPtr(), i );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = x.Compare( tst, MATX_LTS_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_LowerTriangularSolveT %dx%d %s", i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
}

#define MATX_LDLT_SIMD_EPSILON			0.1f
#define MATX_LDLT_FACTOR_SOLVE_SIZE		64

/*
============
TestMatXLDLTFactor
============
*/
void TestMatXLDLTFactor( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	const char *result;
	idMatX src, original, mat1, mat2;
	idVecX invDiag1, invDiag2;

	idLib::common->Printf( "====================================\n" );

	original.SetSize( MATX_LDLT_FACTOR_SOLVE_SIZE, MATX_LDLT_FACTOR_SOLVE_SIZE );
	src.Random( MATX_LDLT_FACTOR_SOLVE_SIZE, MATX_LDLT_FACTOR_SOLVE_SIZE, 0, -1.0f, 1.0f );
	src.TransposeMultiply( original, src );

	for ( i = MATX_LTS_SOLVE_START; i < MATX_LTS_SOLVE_SIZE; i += MATX_LTS_SOLVE_STEP ) {

		bestClocksGeneric = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			mat1 = original;
			invDiag1.Zero( MATX_LDLT_FACTOR_SOLVE_SIZE );
			StartRecordTime( start );
			p_generic->MatX_LDLTFactor( mat1, invDiag1, i );
			StopRecordTime( end );
			GetBest( start, end, bestClocksGeneric );
		}
		PrintClocks( va( "generic->MatX_LDLTFactor %dx%d", i, i ), 1, bestClocksGeneric );

		bestClocksSIMD = 0;
		for ( j = 0; j < NUMTESTS; j++ ) {
			mat2 = original;
			invDiag2.Zero( MATX_LDLT_FACTOR_SOLVE_SIZE );
			StartRecordTime( start );
			p_simd->MatX_LDLTFactor( mat2, invDiag2, i );
			StopRecordTime( end );
			GetBest( start, end, bestClocksSIMD );
		}
		result = mat1.Compare( mat2, MATX_LDLT_SIMD_EPSILON ) && invDiag1.Compare( invDiag2, MATX_LDLT_SIMD_EPSILON ) ? "ok" : S_COLOR_RED"X";
		PrintClocks( va( "   simd->MatX_LDLTFactor %dx%d %s", i, i, result ), 1, bestClocksSIMD, bestClocksGeneric );
	}
}

/*
============
TestBlendJoints
============
*/
void TestBlendJoints( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idJointQuat baseJoints[COUNT] );
	ALIGN16( idJointQuat joints1[COUNT] );
	ALIGN16( idJointQuat joints2[COUNT] );
	ALIGN16( idJointQuat blendJoints[COUNT] );
	ALIGN16( int index[COUNT] );
	float lerp = 0.3f;
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		idAngles angles;
		angles[0] = srnd.CRandomFloat() * 180.0f;
		angles[1] = srnd.CRandomFloat() * 180.0f;
		angles[2] = srnd.CRandomFloat() * 180.0f;
		baseJoints[i].q = angles.ToQuat();
		baseJoints[i].t[0] = srnd.CRandomFloat() * 10.0f;
		baseJoints[i].t[1] = srnd.CRandomFloat() * 10.0f;
		baseJoints[i].t[2] = srnd.CRandomFloat() * 10.0f;
		angles[0] = srnd.CRandomFloat() * 180.0f;
		angles[1] = srnd.CRandomFloat() * 180.0f;
		angles[2] = srnd.CRandomFloat() * 180.0f;
		blendJoints[i].q = angles.ToQuat();
		blendJoints[i].t[0] = srnd.CRandomFloat() * 10.0f;
		blendJoints[i].t[1] = srnd.CRandomFloat() * 10.0f;
		blendJoints[i].t[2] = srnd.CRandomFloat() * 10.0f;
		index[i] = i;
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < COUNT; j++ ) {
			joints1[j] = baseJoints[j];
		}
		StartRecordTime( start );
		p_generic->BlendJoints( joints1, blendJoints, lerp, index, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->BlendJoints()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < COUNT; j++ ) {
			joints2[j] = baseJoints[j];
		}
		StartRecordTime( start );
		p_simd->BlendJoints( joints2, blendJoints, lerp, index, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( !joints1[i].t.Compare( joints2[i].t, 1e-3f ) ) {
			break;
		}
		if ( !joints1[i].q.Compare( joints2[i].q, 1e-2f ) ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->BlendJoints() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestConvertJointQuatsToJointMats
============
*/
void TestConvertJointQuatsToJointMats( void ) {
	int i;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idJointQuat baseJoints[COUNT] );
	ALIGN16( idJointMat joints1[COUNT] );
	ALIGN16( idJointMat joints2[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		idAngles angles;
		angles[0] = srnd.CRandomFloat() * 180.0f;
		angles[1] = srnd.CRandomFloat() * 180.0f;
		angles[2] = srnd.CRandomFloat() * 180.0f;
		baseJoints[i].q = angles.ToQuat();
		baseJoints[i].t[0] = srnd.CRandomFloat() * 10.0f;
		baseJoints[i].t[1] = srnd.CRandomFloat() * 10.0f;
		baseJoints[i].t[2] = srnd.CRandomFloat() * 10.0f;
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->ConvertJointQuatsToJointMats( joints1, baseJoints, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->ConvertJointQuatsToJointMats()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->ConvertJointQuatsToJointMats( joints2, baseJoints, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( !joints1[i].Compare( joints2[i], 1e-4f ) ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->ConvertJointQuatsToJointMats() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestConvertJointMatsToJointQuats
============
*/
void TestConvertJointMatsToJointQuats( void ) {
	int i;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idJointMat baseJoints[COUNT] );
	ALIGN16( idJointQuat joints1[COUNT] );
	ALIGN16( idJointQuat joints2[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		idAngles angles;
		angles[0] = srnd.CRandomFloat() * 180.0f;
		angles[1] = srnd.CRandomFloat() * 180.0f;
		angles[2] = srnd.CRandomFloat() * 180.0f;
		baseJoints[i].SetRotation( angles.ToMat3() );
		idVec3 v;
		v[0] = srnd.CRandomFloat() * 10.0f;
		v[1] = srnd.CRandomFloat() * 10.0f;
		v[2] = srnd.CRandomFloat() * 10.0f;
		baseJoints[i].SetTranslation( v );
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->ConvertJointMatsToJointQuats( joints1, baseJoints, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->ConvertJointMatsToJointQuats()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->ConvertJointMatsToJointQuats( joints2, baseJoints, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( !joints1[i].q.Compare( joints2[i].q, 1e-4f ) ) {
			idLib::common->Printf( "ConvertJointMatsToJointQuats: broken q %i\n", i );
			break;
		}
		if ( !joints1[i].t.Compare( joints2[i].t, 1e-4f ) ) {
			idLib::common->Printf( "ConvertJointMatsToJointQuats: broken t %i\n", i );
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->ConvertJointMatsToJointQuats() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestTransformJoints
============
*/
void TestTransformJoints( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idJointMat joints[COUNT + 1] );
	ALIGN16( idJointMat joints1[COUNT + 1] );
	ALIGN16( idJointMat joints2[COUNT + 1] );
	ALIGN16( int parents[COUNT + 1] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i <= COUNT; i++ ) {
		idAngles angles;
		angles[0] = srnd.CRandomFloat() * 180.0f;
		angles[1] = srnd.CRandomFloat() * 180.0f;
		angles[2] = srnd.CRandomFloat() * 180.0f;
		joints[i].SetRotation( angles.ToMat3() );
		idVec3 v;
		v[0] = srnd.CRandomFloat() * 2.0f;
		v[1] = srnd.CRandomFloat() * 2.0f;
		v[2] = srnd.CRandomFloat() * 2.0f;
		joints[i].SetTranslation( v );
		parents[i] = i - 1;
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j <= COUNT; j++ ) {
			joints1[j] = joints[j];
		}
		StartRecordTime( start );
		p_generic->TransformJoints( joints1, parents, 1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->TransformJoints()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j <= COUNT; j++ ) {
			joints2[j] = joints[j];
		}
		StartRecordTime( start );
		p_simd->TransformJoints( joints2, parents, 1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( !joints1[i + 1].Compare( joints2[i + 1], 1e-4f ) ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->TransformJoints() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestUntransformJoints
============
*/
void TestUntransformJoints( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idJointMat joints[COUNT + 1] );
	ALIGN16( idJointMat joints1[COUNT + 1] );
	ALIGN16( idJointMat joints2[COUNT + 1] );
	ALIGN16( int parents[COUNT + 1] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i <= COUNT; i++ ) {
		idAngles angles;
		angles[0] = srnd.CRandomFloat() * 180.0f;
		angles[1] = srnd.CRandomFloat() * 180.0f;
		angles[2] = srnd.CRandomFloat() * 180.0f;
		joints[i].SetRotation( angles.ToMat3() );
		idVec3 v;
		v[0] = srnd.CRandomFloat() * 2.0f;
		v[1] = srnd.CRandomFloat() * 2.0f;
		v[2] = srnd.CRandomFloat() * 2.0f;
		joints[i].SetTranslation( v );
		parents[i] = i - 1;
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j <= COUNT; j++ ) {
			joints1[j] = joints[j];
		}
		StartRecordTime( start );
		p_generic->UntransformJoints( joints1, parents, 1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->UntransformJoints()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j <= COUNT; j++ ) {
			joints2[j] = joints[j];
		}
		StartRecordTime( start );
		p_simd->UntransformJoints( joints2, parents, 1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( !joints1[i + 1].Compare( joints2[i + 1], 1e-4f ) ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->UntransformJoints() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestTransformVerts
============
*/
#define NUMJOINTS	64
#define NUMVERTS	COUNT/2
void TestTransformVerts( void ) {
	int i;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idDrawVert drawVerts1[NUMVERTS] );
	ALIGN16( idDrawVert drawVerts2[NUMVERTS] );
	ALIGN16( idJointMat joints[NUMJOINTS] );
	ALIGN16( idVec4 weights[COUNT] );
	ALIGN16( int weightIndex[COUNT * 2] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < NUMJOINTS; i++ ) {
		idAngles angles;
		angles[0] = srnd.CRandomFloat() * 180.0f;
		angles[1] = srnd.CRandomFloat() * 180.0f;
		angles[2] = srnd.CRandomFloat() * 180.0f;
		joints[i].SetRotation( angles.ToMat3() );
		idVec3 v;
		v[0] = srnd.CRandomFloat() * 2.0f;
		v[1] = srnd.CRandomFloat() * 2.0f;
		v[2] = srnd.CRandomFloat() * 2.0f;
		joints[i].SetTranslation( v );
	}

	for ( i = 0; i < COUNT; i++ ) {
		weights[i][0] = srnd.CRandomFloat() * 2.0f;
		weights[i][1] = srnd.CRandomFloat() * 2.0f;
		weights[i][2] = srnd.CRandomFloat() * 2.0f;
		weights[i][3] = srnd.CRandomFloat();
		weightIndex[i * 2 + 0] = ( i * NUMJOINTS / COUNT ) * sizeof( idJointMat );
		weightIndex[i * 2 + 1] = i & 1;
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->TransformVerts( drawVerts1, NUMVERTS, joints, weights, weightIndex, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->TransformVerts()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->TransformVerts( drawVerts2, NUMVERTS, joints, weights, weightIndex, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < NUMVERTS; i++ ) {
		if ( !drawVerts1[i].xyz.Compare( drawVerts2[i].xyz, 0.5f ) ) {
			break;
		}
	}
	result = ( i >= NUMVERTS ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->TransformVerts() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestTracePointCull
============
*/
void TestTracePointCull( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idPlane planes[4] );
	ALIGN16( idDrawVert drawVerts[COUNT] );
	ALIGN16( byte cullBits1[COUNT] );
	ALIGN16( byte cullBits2[COUNT] );
	byte totalOr1 = 0, totalOr2 = 0;
	const char *result;

	idRandom srnd( RANDOM_SEED );

	planes[0].SetNormal( idVec3( 1,  0, 0 ) );
	planes[1].SetNormal( idVec3( -1,  0, 0 ) );
	planes[2].SetNormal( idVec3( 0,  1, 0 ) );
	planes[3].SetNormal( idVec3( 0, -1, 0 ) );
	planes[0][3] = -5.3f;
	planes[1][3] = 5.3f;
	planes[2][3] = -3.4f;
	planes[3][3] = 3.4f;

	for ( i = 0; i < COUNT; i++ ) {
		for ( j = 0; j < 3; j++ ) {
			drawVerts[i].xyz[j] = srnd.CRandomFloat() * 10.0f;
		}
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->TracePointCull( cullBits1, totalOr1, 0.0f, planes, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->TracePointCull()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->TracePointCull( cullBits2, totalOr2, 0.0f, planes, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( cullBits1[i] != cullBits2[i] ) {
			break;
		}
	}
	result = ( i >= COUNT && totalOr1 == totalOr2 ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->TracePointCull() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestDecalPointCull
============
*/
void TestDecalPointCull( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idPlane planes[6] );
	ALIGN16( idDrawVert drawVerts[COUNT] );
	ALIGN16( byte cullBits1[COUNT] );
	ALIGN16( byte cullBits2[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	planes[0].SetNormal( idVec3( 1,  0,  0 ) );
	planes[1].SetNormal( idVec3( -1,  0,  0 ) );
	planes[2].SetNormal( idVec3( 0,  1,  0 ) );
	planes[3].SetNormal( idVec3( 0, -1,  0 ) );
	planes[4].SetNormal( idVec3( 0,  0,  1 ) );
	planes[5].SetNormal( idVec3( 0,  0, -1 ) );
	planes[0][3] = -5.3f;
	planes[1][3] = 5.3f;
	planes[2][3] = -4.4f;
	planes[3][3] = 4.4f;
	planes[4][3] = -3.5f;
	planes[5][3] = 3.5f;

	for ( i = 0; i < COUNT; i++ ) {
		for ( j = 0; j < 3; j++ ) {
			drawVerts[i].xyz[j] = srnd.CRandomFloat() * 10.0f;
		}
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->DecalPointCull( cullBits1, planes, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->DecalPointCull()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->DecalPointCull( cullBits2, planes, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( cullBits1[i] != cullBits2[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->DecalPointCull() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestOverlayPointCull
============
*/
void TestOverlayPointCull( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idPlane planes[2] );
	ALIGN16( idDrawVert drawVerts[COUNT] );
	ALIGN16( byte cullBits1[COUNT] );
	ALIGN16( byte cullBits2[COUNT] );
	ALIGN16( idVec2 texCoords1[COUNT] );
	ALIGN16( idVec2 texCoords2[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	planes[0].SetNormal( idVec3( 0.3f, 0.2f, 0.9f ) );
	planes[1].SetNormal( idVec3( 0.9f, 0.2f, 0.3f ) );
	planes[0][3] = -5.3f;
	planes[1][3] = -4.3f;

	for ( i = 0; i < COUNT; i++ ) {
		for ( j = 0; j < 3; j++ ) {
			drawVerts[i].xyz[j] = srnd.CRandomFloat() * 10.0f;
		}
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->OverlayPointCull( cullBits1, texCoords1, planes, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->OverlayPointCull()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->OverlayPointCull( cullBits2, texCoords2, planes, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( cullBits1[i] != cullBits2[i] ) {
			break;
		}
		if ( !texCoords1[i].Compare( texCoords2[i], 1e-4f ) ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->OverlayPointCull() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestDeriveTriPlanes
============
*/
void TestDeriveTriPlanes( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idDrawVert drawVerts1[COUNT] );
	ALIGN16( idDrawVert drawVerts2[COUNT] );
	ALIGN16( idPlane planes1[COUNT] );
	ALIGN16( idPlane planes2[COUNT] );
	ALIGN16( int indexes[COUNT * 3] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		for ( j = 0; j < 3; j++ ) {
			drawVerts1[i].xyz[j] = srnd.CRandomFloat() * 10.0f;
		}
		for ( j = 0; j < 2; j++ ) {
			drawVerts1[i].st[j] = srnd.CRandomFloat();
		}
		drawVerts2[i] = drawVerts1[i];
	}

	for ( i = 0; i < COUNT; i++ ) {
		indexes[i * 3 + 0] = ( i + 0 ) % COUNT;
		indexes[i * 3 + 1] = ( i + 1 ) % COUNT;
		indexes[i * 3 + 2] = ( i + 2 ) % COUNT;
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->DeriveTriPlanes( planes1, drawVerts1, COUNT, indexes, COUNT * 3 );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->DeriveTriPlanes()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->DeriveTriPlanes( planes2, drawVerts2, COUNT, indexes, COUNT * 3 );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( !planes1[i].Compare( planes2[i], 1e-1f, 1e-1f ) ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->DeriveTriPlanes() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestDeriveTangents
============
*/
void TestDeriveTangents( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idDrawVert drawVerts1[COUNT] );
	ALIGN16( idDrawVert drawVerts2[COUNT] );
	ALIGN16( idPlane planes1[COUNT] );
	ALIGN16( idPlane planes2[COUNT] );
	ALIGN16( int indexes[COUNT * 3] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		for ( j = 0; j < 3; j++ ) {
			drawVerts1[i].xyz[j] = srnd.CRandomFloat() * 10.0f;
		}
		for ( j = 0; j < 2; j++ ) {
			drawVerts1[i].st[j] = srnd.CRandomFloat();
		}
		drawVerts2[i] = drawVerts1[i];
	}

	for ( i = 0; i < COUNT; i++ ) {
		indexes[i * 3 + 0] = ( i + 0 ) % COUNT;
		indexes[i * 3 + 1] = ( i + 1 ) % COUNT;
		indexes[i * 3 + 2] = ( i + 2 ) % COUNT;
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->DeriveTangents( planes1, drawVerts1, COUNT, indexes, COUNT * 3 );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->DeriveTangents()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->DeriveTangents( planes2, drawVerts2, COUNT, indexes, COUNT * 3 );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		idVec3 v1, v2;

		v1 = drawVerts1[i].normal;
		v1.Normalize();
		v2 = drawVerts2[i].normal;
		v2.Normalize();
		if ( !v1.Compare( v2, 1e-1f ) ) {
			idLib::common->Printf( "DeriveTangents: broken at normal %i\n -- expecting %s got %s", i, v1.ToString(), v2.ToString() );
			break;
		}
		v1 = drawVerts1[i].tangents[0];
		v1.Normalize();
		v2 = drawVerts2[i].tangents[0];
		v2.Normalize();
		if ( !v1.Compare( v2, 1e-1f ) ) {
			idLib::common->Printf( "DeriveTangents: broken at tangent0 %i -- expecting %s got %s\n", i, v1.ToString(), v2.ToString() );
			break;
		}
		v1 = drawVerts1[i].tangents[1];
		v1.Normalize();
		v2 = drawVerts2[i].tangents[1];
		v2.Normalize();
		if ( !v1.Compare( v2, 1e-1f ) ) {
			idLib::common->Printf( "DeriveTangents: broken at tangent1 %i -- expecting %s got %s\n", i, v1.ToString(), v2.ToString() );
			break;
		}
		if ( !planes1[i].Compare( planes2[i], 1e-1f, 1e-1f ) ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->DeriveTangents() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestDeriveUnsmoothedTangents
============
*/
void TestDeriveUnsmoothedTangents( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idDrawVert drawVerts1[COUNT] );
	ALIGN16( idDrawVert drawVerts2[COUNT] );
	ALIGN16( dominantTri_s dominantTris[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		for ( j = 0; j < 3; j++ ) {
			drawVerts1[i].xyz[j] = srnd.CRandomFloat() * 10.0f;
		}
		for ( j = 0; j < 2; j++ ) {
			drawVerts1[i].st[j] = srnd.CRandomFloat();
		}
		drawVerts2[i] = drawVerts1[i];

		dominantTris[i].v2 = ( i + 1 + srnd.RandomInt( 8 ) ) % COUNT;
		dominantTris[i].v3 = ( i + 9 + srnd.RandomInt( 8 ) ) % COUNT;
		dominantTris[i].normalizationScale[0] = srnd.CRandomFloat();
		dominantTris[i].normalizationScale[1] = srnd.CRandomFloat();
		dominantTris[i].normalizationScale[2] = srnd.CRandomFloat();
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->DeriveUnsmoothedTangents( drawVerts1, dominantTris, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->DeriveUnsmoothedTangents()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->DeriveUnsmoothedTangents( drawVerts2, dominantTris, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		idVec3 v1, v2;

		v1 = drawVerts1[i].normal;
		v1.Normalize();
		v2 = drawVerts2[i].normal;
		v2.Normalize();
		if ( !v1.Compare( v2, 1e-1f ) ) {
			break;
		}
		v1 = drawVerts1[i].tangents[0];
		v1.Normalize();
		v2 = drawVerts2[i].tangents[0];
		v2.Normalize();
		if ( !v1.Compare( v2, 1e-1f ) ) {
			break;
		}
		v1 = drawVerts1[i].tangents[1];
		v1.Normalize();
		v2 = drawVerts2[i].tangents[1];
		v2.Normalize();
		if ( !v1.Compare( v2, 1e-1f ) ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->DeriveUnsmoothedTangents() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestNormalizeTangents
============
*/
void TestNormalizeTangents( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idDrawVert drawVerts1[COUNT] );
	ALIGN16( idDrawVert drawVerts2[COUNT] );
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		for ( j = 0; j < 3; j++ ) {
			drawVerts1[i].normal[j] = srnd.CRandomFloat() * 10.0f;
			drawVerts1[i].tangents[0][j] = srnd.CRandomFloat() * 10.0f;
			drawVerts1[i].tangents[1][j] = srnd.CRandomFloat() * 10.0f;
		}
		drawVerts2[i] = drawVerts1[i];
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->NormalizeTangents( drawVerts1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->NormalizeTangents()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->NormalizeTangents( drawVerts2, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( !drawVerts1[i].normal.Compare( drawVerts2[i].normal, 1e-2f ) ) {
			break;
		}
		if ( !drawVerts1[i].tangents[0].Compare( drawVerts2[i].tangents[0], 1e-2f ) ) {
			break;
		}
		if ( !drawVerts1[i].tangents[1].Compare( drawVerts2[i].tangents[1], 1e-2f ) ) {
			break;
		}

		// since we're doing a lot of unaligned work, added this check to
		// make sure xyz wasn't getting overwritten
		if ( !drawVerts1[i].xyz.Compare( drawVerts2[i].xyz, 1e-2f ) ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->NormalizeTangents() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestCreateShadowCache
============
*/
void TestCreateShadowCache( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( idDrawVert drawVerts[COUNT] );
	ALIGN16( idVec4 vertexCache1[COUNT * 2] );
	ALIGN16( idVec4 vertexCache2[COUNT * 2] );
	ALIGN16( int originalVertRemap[COUNT] );
	ALIGN16( int vertRemap1[COUNT] );
	ALIGN16( int vertRemap2[COUNT] );
	ALIGN16( idVec3 lightOrigin );
	int numVerts1 = 0, numVerts2 = 0;
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		drawVerts[i].xyz[0] = srnd.CRandomFloat() * 100.0f;
		drawVerts[i].xyz[1] = srnd.CRandomFloat() * 100.0f;
		drawVerts[i].xyz[2] = srnd.CRandomFloat() * 100.0f;
		originalVertRemap[i] = ( srnd.CRandomFloat() > 0.0f ) ? -1 : 0;
	}
	lightOrigin[0] = srnd.CRandomFloat() * 100.0f;
	lightOrigin[1] = srnd.CRandomFloat() * 100.0f;
	lightOrigin[2] = srnd.CRandomFloat() * 100.0f;

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < COUNT; j++ ) {
			vertRemap1[j] = originalVertRemap[j];
		}
		StartRecordTime( start );
		numVerts1 = p_generic->CreateShadowCache( vertexCache1, vertRemap1, lightOrigin, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->CreateShadowCache()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < COUNT; j++ ) {
			vertRemap2[j] = originalVertRemap[j];
		}
		StartRecordTime( start );
		numVerts2 = p_simd->CreateShadowCache( vertexCache2, vertRemap2, lightOrigin, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( i < ( numVerts1 / 2 ) ) {
			if ( !vertexCache1[i * 2 + 0].Compare( vertexCache2[i * 2 + 0], 1e-2f ) ) {
				break;
			}
			if ( !vertexCache1[i * 2 + 1].Compare( vertexCache2[i * 2 + 1], 1e-2f ) ) {
				break;
			}
		}
		if ( vertRemap1[i] != vertRemap2[i] ) {
			break;
		}
	}
	result = ( i >= COUNT && numVerts1 == numVerts2 ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->CreateShadowCache() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->CreateVertexProgramShadowCache( vertexCache1, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->CreateVertexProgramShadowCache()", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->CreateVertexProgramShadowCache( vertexCache2, drawVerts, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( !vertexCache1[i * 2 + 0].Compare( vertexCache2[i * 2 + 0], 1e-2f ) ) {
			break;
		}
		if ( !vertexCache1[i * 2 + 1].Compare( vertexCache2[i * 2 + 1], 1e-2f ) ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->CreateVertexProgramShadowCache() %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestConvertRGTCFromRGBA8
============
*/
void TestConvertRGTCFromRGBA8() {
	idRandom rnd(RANDOM_SEED);

	static const int SIZE = 64;
	idList<byte> inputData, outputGeneric, outputSIMD;
	inputData.SetNum(SIZE * SIZE * 4);
	outputGeneric.SetNum(SIZE * SIZE);
	outputSIMD.SetNum(SIZE * SIZE);
	for (int i = 0; i < inputData.Num(); i++)
		inputData[i] = rnd.RandomInt(256);

	TIME_TYPE start, end;
	TIME_TYPE bestClocksGeneric = 0;
	for ( int i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_generic->CompressRGTCFromRGBA8((const byte*)inputData.Ptr(), SIZE, SIZE, 4 * SIZE, outputGeneric.Ptr());
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( va( "generic->CompressRGTCFromRGBA8( %d )", SIZE ), SIZE * SIZE, bestClocksGeneric );

	TIME_TYPE bestClocksSIMD = 0;
	for ( int i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		p_simd->CompressRGTCFromRGBA8((const byte*)inputData.Ptr(), SIZE, SIZE, 4 * SIZE, outputSIMD.Ptr());
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	int diffPos = -1;
	for (int i = 0; i < outputGeneric.Num(); i++)
		if (outputGeneric[i] != outputSIMD[i]) {
			diffPos = i;
			break;
		}
	const char *result = diffPos < 0 ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->CompressRGTCFromRGBA8( %d ) %s", SIZE, result ), SIZE * SIZE, bestClocksSIMD, bestClocksGeneric );

	// Additional tests: various image sizes, better value distributions

	for (int h = 1; h <= 43; h++) {
		for (int w = 1; w <= 43; w++) {
			int stride = w + rnd.RandomInt(10);
			idList<dword> rgbaPixels;
			rgbaPixels.SetNum(h * stride);

			idVec2 avg;
			avg.x = rnd.RandomFloat();
			avg.y = rnd.RandomFloat();
			for (int i = 0; i < h; i++)
				for (int j = 0; j < w; j++) {
					float radius = float(i + j) / (h + w);
					idVec2 val;
					val.x = idMath::ClampFloat(0, 1, avg.x + radius * rnd.CRandomFloat());
					val.y = idMath::ClampFloat(0, 1, avg.y + radius * rnd.CRandomFloat());
					dword red = idMath::Round(val.x * 255.0f);
					dword green = idMath::Round(val.y * 255.0f);
					dword blue = rnd.RandomInt(256);
					dword alpha = rnd.RandomInt(256);
					rgbaPixels[i * stride + j] = red + (green << 8) + (blue << 16) + (alpha << 24);
				}

			int outSize = ((h + 3) / 4) * ((w + 3) / 4) * 16;
			idList<byte> outputGen, outputFast;
			outputGen.SetNum(outSize);
			outputFast.SetNum(outSize);

			p_generic->CompressRGTCFromRGBA8((const byte*)rgbaPixels.Ptr(), w, h, 4 * stride, outputGen.Ptr());
			p_simd->CompressRGTCFromRGBA8((const byte*)rgbaPixels.Ptr(), w, h, 4 * stride, outputFast.Ptr());

			if (memcmp(outputGen.Ptr(), outputFast.Ptr(), outSize) != 0)
				common->Error("TestConvertRGTCFromRGBA8 output mismatch");
		}
	}
}

/*
============
TestSoundUpSampling
============
*/
#define SOUND_UPSAMPLE_EPSILON		1.0f

void TestSoundUpSampling( void ) {
	int i;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( short pcm[MIXBUFFER_SAMPLES * 2] );
	ALIGN16( float ogg0[MIXBUFFER_SAMPLES * 2] );
	ALIGN16( float ogg1[MIXBUFFER_SAMPLES * 2] );
	ALIGN16( float samples1[MIXBUFFER_SAMPLES * 2] );
	ALIGN16( float samples2[MIXBUFFER_SAMPLES * 2] );
	float *ogg[2];
	int kHz, numSpeakers;
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < MIXBUFFER_SAMPLES * 2; i++ ) {
		pcm[i] = srnd.RandomInt( ( 1 << 16 ) ) - ( 1 << 15 );
		ogg0[i] = srnd.RandomFloat();
		ogg1[i] = srnd.RandomFloat();
	}
	ogg[0] = ogg0;
	ogg[1] = ogg1;

	for ( numSpeakers = 1; numSpeakers <= 2; numSpeakers++ ) {

		for ( kHz = 11025; kHz <= 44100; kHz *= 2 ) {
			bestClocksGeneric = 0;
			for ( i = 0; i < NUMTESTS; i++ ) {
				StartRecordTime( start );
				p_generic->UpSamplePCMTo44kHz( samples1, pcm, MIXBUFFER_SAMPLES * numSpeakers * kHz / 44100, kHz, numSpeakers );
				StopRecordTime( end );
				GetBest( start, end, bestClocksGeneric );
			}
			PrintClocks( va( "generic->UpSamplePCMTo44kHz( %d, %d )", kHz, numSpeakers ), MIXBUFFER_SAMPLES * numSpeakers * kHz / 44100, bestClocksGeneric );

			bestClocksSIMD = 0;
			for ( i = 0; i < NUMTESTS; i++ ) {
				StartRecordTime( start );
				p_simd->UpSamplePCMTo44kHz( samples2, pcm, MIXBUFFER_SAMPLES * numSpeakers * kHz / 44100, kHz, numSpeakers );
				StopRecordTime( end );
				GetBest( start, end, bestClocksSIMD );
			}

			for ( i = 0; i < MIXBUFFER_SAMPLES * numSpeakers; i++ ) {
				if ( idMath::Fabs( samples1[i] - samples2[i] ) > SOUND_UPSAMPLE_EPSILON ) {
					break;
				}
			}
			result = ( i >= MIXBUFFER_SAMPLES * numSpeakers ) ? "ok" : S_COLOR_RED"X";
			PrintClocks( va( "   simd->UpSamplePCMTo44kHz( %d, %d ) %s", kHz, numSpeakers, result ), MIXBUFFER_SAMPLES * numSpeakers * kHz / 44100, bestClocksSIMD, bestClocksGeneric );
		}
	}

	for ( numSpeakers = 1; numSpeakers <= 2; numSpeakers++ ) {

		for ( kHz = 11025; kHz <= 44100; kHz *= 2 ) {
			bestClocksGeneric = 0;
			for ( i = 0; i < NUMTESTS; i++ ) {
				StartRecordTime( start );
				p_generic->UpSampleOGGTo44kHz( samples1, ogg, MIXBUFFER_SAMPLES * numSpeakers * kHz / 44100, kHz, numSpeakers );
				StopRecordTime( end );
				GetBest( start, end, bestClocksGeneric );
			}
			PrintClocks( va( "generic->UpSampleOGGTo44kHz( %d, %d )", kHz, numSpeakers ), MIXBUFFER_SAMPLES * numSpeakers * kHz / 44100, bestClocksGeneric );

			bestClocksSIMD = 0;
			for ( i = 0; i < NUMTESTS; i++ ) {
				StartRecordTime( start );
				p_simd->UpSampleOGGTo44kHz( samples2, ogg, MIXBUFFER_SAMPLES * numSpeakers * kHz / 44100, kHz, numSpeakers );
				StopRecordTime( end );
				GetBest( start, end, bestClocksSIMD );
			}

			for ( i = 0; i < MIXBUFFER_SAMPLES * numSpeakers; i++ ) {
				if ( idMath::Fabs( samples1[i] - samples2[i] ) > SOUND_UPSAMPLE_EPSILON ) {
					break;
				}
			}
			result = ( i >= MIXBUFFER_SAMPLES ) ? "ok" : S_COLOR_RED"X";
			PrintClocks( va( "   simd->UpSampleOGGTo44kHz( %d, %d ) %s", kHz, numSpeakers, result ), MIXBUFFER_SAMPLES * numSpeakers * kHz / 44100, bestClocksSIMD, bestClocksGeneric );
		}
	}
}

/*
============
TestSoundMixing
============
*/
#define SOUND_MIX_EPSILON		2.0f

void TestSoundMixing( void ) {
	int i, j;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( float origMixBuffer[MIXBUFFER_SAMPLES * 6] );
	ALIGN16( float mixBuffer1[MIXBUFFER_SAMPLES * 6] );
	ALIGN16( float mixBuffer2[MIXBUFFER_SAMPLES * 6] );
	ALIGN16( float samples[MIXBUFFER_SAMPLES * 6] );
	ALIGN16( short outSamples1[MIXBUFFER_SAMPLES * 6] );
	ALIGN16( short outSamples2[MIXBUFFER_SAMPLES * 6] );
	float lastV[6];
	float currentV[6];
	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < 6; i++ ) {
		lastV[i] = srnd.CRandomFloat();
		currentV[i] = srnd.CRandomFloat();
	}

	for ( i = 0; i < MIXBUFFER_SAMPLES * 6; i++ ) {
		origMixBuffer[i] = srnd.CRandomFloat();
		samples[i] = srnd.RandomInt( ( 1 << 16 ) ) - ( 1 << 15 );
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < MIXBUFFER_SAMPLES * 6; j++ ) {
			mixBuffer1[j] = origMixBuffer[j];
		}
		StartRecordTime( start );
		p_generic->MixSoundTwoSpeakerMono( mixBuffer1, samples, MIXBUFFER_SAMPLES, lastV, currentV );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->MixSoundTwoSpeakerMono()", MIXBUFFER_SAMPLES, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < MIXBUFFER_SAMPLES * 6; j++ ) {
			mixBuffer2[j] = origMixBuffer[j];
		}
		StartRecordTime( start );
		p_simd->MixSoundTwoSpeakerMono( mixBuffer2, samples, MIXBUFFER_SAMPLES, lastV, currentV );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < MIXBUFFER_SAMPLES * 6; i++ ) {
		if ( idMath::Fabs( mixBuffer1[i] - mixBuffer2[i] ) > SOUND_MIX_EPSILON ) {
			break;
		}
	}
	result = ( i >= MIXBUFFER_SAMPLES * 6 ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->MixSoundTwoSpeakerMono() %s", result ), MIXBUFFER_SAMPLES, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < MIXBUFFER_SAMPLES * 6; j++ ) {
			mixBuffer1[j] = origMixBuffer[j];
		}
		StartRecordTime( start );
		p_generic->MixSoundTwoSpeakerStereo( mixBuffer1, samples, MIXBUFFER_SAMPLES, lastV, currentV );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->MixSoundTwoSpeakerStereo()", MIXBUFFER_SAMPLES, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < MIXBUFFER_SAMPLES * 6; j++ ) {
			mixBuffer2[j] = origMixBuffer[j];
		}
		StartRecordTime( start );
		p_simd->MixSoundTwoSpeakerStereo( mixBuffer2, samples, MIXBUFFER_SAMPLES, lastV, currentV );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < MIXBUFFER_SAMPLES * 6; i++ ) {
		if ( idMath::Fabs( mixBuffer1[i] - mixBuffer2[i] ) > SOUND_MIX_EPSILON ) {
			break;
		}
	}
	result = ( i >= MIXBUFFER_SAMPLES * 6 ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->MixSoundTwoSpeakerStereo() %s", result ), MIXBUFFER_SAMPLES, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < MIXBUFFER_SAMPLES * 6; j++ ) {
			mixBuffer1[j] = origMixBuffer[j];
		}
		StartRecordTime( start );
		p_generic->MixSoundSixSpeakerMono( mixBuffer1, samples, MIXBUFFER_SAMPLES, lastV, currentV );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->MixSoundSixSpeakerMono()", MIXBUFFER_SAMPLES, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < MIXBUFFER_SAMPLES * 6; j++ ) {
			mixBuffer2[j] = origMixBuffer[j];
		}
		StartRecordTime( start );
		p_simd->MixSoundSixSpeakerMono( mixBuffer2, samples, MIXBUFFER_SAMPLES, lastV, currentV );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < MIXBUFFER_SAMPLES * 6; i++ ) {
		if ( idMath::Fabs( mixBuffer1[i] - mixBuffer2[i] ) > SOUND_MIX_EPSILON ) {
			break;
		}
	}
	result = ( i >= MIXBUFFER_SAMPLES * 6 ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->MixSoundSixSpeakerMono() %s", result ), MIXBUFFER_SAMPLES, bestClocksSIMD, bestClocksGeneric );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < MIXBUFFER_SAMPLES * 6; j++ ) {
			mixBuffer1[j] = origMixBuffer[j];
		}
		StartRecordTime( start );
		p_generic->MixSoundSixSpeakerStereo( mixBuffer1, samples, MIXBUFFER_SAMPLES, lastV, currentV );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->MixSoundSixSpeakerStereo()", MIXBUFFER_SAMPLES, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < MIXBUFFER_SAMPLES * 6; j++ ) {
			mixBuffer2[j] = origMixBuffer[j];
		}
		StartRecordTime( start );
		p_simd->MixSoundSixSpeakerStereo( mixBuffer2, samples, MIXBUFFER_SAMPLES, lastV, currentV );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < MIXBUFFER_SAMPLES * 6; i++ ) {
		if ( idMath::Fabs( mixBuffer1[i] - mixBuffer2[i] ) > SOUND_MIX_EPSILON ) {
			break;
		}
	}
	result = ( i >= MIXBUFFER_SAMPLES * 6 ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->MixSoundSixSpeakerStereo() %s", result ), MIXBUFFER_SAMPLES, bestClocksSIMD, bestClocksGeneric );

	for ( i = 0; i < MIXBUFFER_SAMPLES * 6; i++ ) {
		origMixBuffer[i] = srnd.RandomInt( ( 1 << 17 ) ) - ( 1 << 16 );
	}

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < MIXBUFFER_SAMPLES * 6; j++ ) {
			mixBuffer1[j] = origMixBuffer[j];
		}
		StartRecordTime( start );
		p_generic->MixedSoundToSamples( outSamples1, mixBuffer1, MIXBUFFER_SAMPLES * 6 );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->MixedSoundToSamples()", MIXBUFFER_SAMPLES, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		for ( j = 0; j < MIXBUFFER_SAMPLES * 6; j++ ) {
			mixBuffer2[j] = origMixBuffer[j];
		}
		StartRecordTime( start );
		p_simd->MixedSoundToSamples( outSamples2, mixBuffer2, MIXBUFFER_SAMPLES * 6 );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < MIXBUFFER_SAMPLES * 6; i++ ) {
		if ( outSamples1[i] != outSamples2[i] ) {
			break;
		}
	}
	result = ( i >= MIXBUFFER_SAMPLES * 6 ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->MixedSoundToSamples() %s", result ), MIXBUFFER_SAMPLES, bestClocksSIMD, bestClocksGeneric );
}

/*
============
TestMath
============
*/
void TestMath( void ) {
	int i;
	TIME_TYPE start, end, bestClocks;

	idLib::common->Printf( "====================================\n" );

	float tst = -1.0f;
	float tst2 = 1.0f;
	float testvar = 1.0f;
	idRandom rnd;

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = fabs( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "            fabs( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		int tmp = * ( int * ) &tst;
		tmp &= 0x7FFFFFFF;
		tst = * ( float * ) &tmp;
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "    idMath::Fabs( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = 10.0f + 100.0f * rnd.RandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = sqrt( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst * 0.01f;
		tst = 10.0f + 100.0f * rnd.RandomFloat();
	}
	PrintClocks( "            sqrt( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.RandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Sqrt( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.RandomFloat();
	}
	PrintClocks( "    idMath::Sqrt( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.RandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Sqrt16( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.RandomFloat();
	}
	PrintClocks( "  idMath::Sqrt16( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.RandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Sqrt64( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.RandomFloat();
	}
	PrintClocks( "  idMath::Sqrt64( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.RandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = tst * idMath::RSqrt( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.RandomFloat();
	}
	PrintClocks( "   idMath::RSqrt( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Sin( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "     idMath::Sin( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Sin16( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "   idMath::Sin16( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Cos( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "     idMath::Cos( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Cos16( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "   idMath::Cos16( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		idMath::SinCos( tst, tst, tst2 );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "  idMath::SinCos( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		idMath::SinCos16( tst, tst, tst2 );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "idMath::SinCos16( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Tan( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "     idMath::Tan( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Tan16( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "   idMath::Tan16( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::ASin( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst * ( 1.0f / idMath::PI );
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "    idMath::ASin( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::ASin16( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst * ( 1.0f / idMath::PI );
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "  idMath::ASin16( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::ACos( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst * ( 1.0f / idMath::PI );
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "    idMath::ACos( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::ACos16( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst * ( 1.0f / idMath::PI );
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "  idMath::ACos16( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::ATan( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "    idMath::ATan( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::ATan16( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "  idMath::ATan16( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Pow( 2.7f, tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst * 0.1f;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "    idMath::Pow( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Pow16( 2.7f, tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst * 0.1f;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "  idMath::Pow16( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Exp( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst * 0.1f;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "    idMath::Exp( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		tst = idMath::Exp16( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst * 0.1f;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "  idMath::Exp16( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		tst = fabs( tst ) + 1.0f;
		StartRecordTime( start );
		tst = idMath::Log( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "    idMath::Log( tst )", 1, bestClocks );

	bestClocks = 0;
	tst = rnd.CRandomFloat();
	for ( i = 0; i < NUMTESTS; i++ ) {
		tst = fabs( tst ) + 1.0f;
		StartRecordTime( start );
		tst = idMath::Log16( tst );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
		testvar = ( testvar + tst ) * tst;
		tst = rnd.CRandomFloat();
	}
	PrintClocks( "  idMath::Log16( tst )", 1, bestClocks );

	//idLib::common->Printf( "testvar = %f\n", testvar );

	idMat3 resultMat3;
	idQuat fromQuat, toQuat, resultQuat;
	idCQuat cq;
	idAngles ang;

	fromQuat = idAngles( 30, 45, 0 ).ToQuat();
	toQuat = idAngles( 45, 0, 0 ).ToQuat();
	cq = idAngles( 30, 45, 0 ).ToQuat().ToCQuat();
	ang = idAngles( 30, 40, 50 );

	bestClocks = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		resultMat3 = fromQuat.ToMat3();
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
	}
	PrintClocks( "       idQuat::ToMat3()", 1, bestClocks );

	bestClocks = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		resultQuat.Slerp( fromQuat, toQuat, 0.3f );
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
	}
	PrintClocks( "        idQuat::Slerp()", 1, bestClocks );

	bestClocks = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		resultQuat = cq.ToQuat();
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
	}
	PrintClocks( "      idCQuat::ToQuat()", 1, bestClocks );

	bestClocks = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		resultQuat = ang.ToQuat();
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
	}
	PrintClocks( "     idAngles::ToQuat()", 1, bestClocks );

	bestClocks = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {
		StartRecordTime( start );
		resultMat3 = ang.ToMat3();
		StopRecordTime( end );
		GetBest( start, end, bestClocks );
	}
	PrintClocks( "     idAngles::ToMat3()", 1, bestClocks );
}

/*
============
TestNegate
============
*/

// this wasn't previously in the test
void TestNegate( void ) {
	int i;
	TIME_TYPE start, end, bestClocksGeneric, bestClocksSIMD;
	ALIGN16( float fsrc0[COUNT] );
	ALIGN16( float fsrc1[COUNT] );
	ALIGN16( float fsrc2[COUNT] );

	const char *result;

	idRandom srnd( RANDOM_SEED );

	for ( i = 0; i < COUNT; i++ ) {
		fsrc0[i] = fsrc1[i] = fsrc2[i] = srnd.CRandomFloat() * 10.0f;
	}
	idLib::common->Printf( "====================================\n" );

	bestClocksGeneric = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {

		memcpy( &fsrc1[0], &fsrc0[0], COUNT * sizeof( float ) );

		StartRecordTime( start );
		p_generic->Negate16( fsrc1, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksGeneric );
	}
	PrintClocks( "generic->Negate16( float[] )", COUNT, bestClocksGeneric );

	bestClocksSIMD = 0;
	for ( i = 0; i < NUMTESTS; i++ ) {

		memcpy( &fsrc2[0], &fsrc0[0], COUNT * sizeof( float ) );

		StartRecordTime( start );
		p_simd->Negate16( fsrc2, COUNT );
		StopRecordTime( end );
		GetBest( start, end, bestClocksSIMD );
	}

	for ( i = 0; i < COUNT; i++ ) {
		if ( fsrc1[i] != fsrc2[i] ) {
			break;
		}
	}
	result = ( i >= COUNT ) ? "ok" : S_COLOR_RED"X";
	PrintClocks( va( "   simd->Negate16( float[] ) %s", result ), COUNT, bestClocksSIMD, bestClocksGeneric );
}


/*
============
idSIMD::Test_f
============
*/
void idSIMD::Test_f( const idCmdArgs &args ) {

#ifdef _WIN32
	SetThreadPriority( GetCurrentThread(), THREAD_PRIORITY_TIME_CRITICAL );
#endif /* _WIN32 */

	p_simd = SIMDProcessor;
	p_generic = CreateProcessor( "generic" );

	int testBits = -1;
	if ( idStr::Length( args.Argv( 1 ) ) != 0 ) {
		idStr argString = args.Argv( 1 );
		if ( argString.IsNumeric() )
			testBits = atoi( argString.c_str() );
	}

	idLib::common->SetRefreshOnPrint( true );

	idLib::common->Printf( "using %s for SIMD processing\n", p_simd->GetName() );

	GetBaseClocks();

	if ( testBits & 1 ) {
		TestMath();
		TestAdd();
		TestSub();
		TestMul();
		TestDiv();
		TestMulAdd();
		TestMulSub();
		TestDot();
		TestCompare();
		TestMinMax();
		TestClamp();
		TestMemcpy();
		TestMemset();
		TestNegate();
	}

	if ( testBits & 2 ) {
		TestMatXMultiplyVecX();
		TestMatXMultiplyAddVecX();
		TestMatXTransposeMultiplyVecX();
		TestMatXTransposeMultiplyAddVecX();
		TestMatXMultiplyMatX();
		TestMatXTransposeMultiplyMatX();
		TestMatXLowerTriangularSolve();
		TestMatXLowerTriangularSolveTranspose();
		TestMatXLDLTFactor();
	}

	if ( testBits & 4 ) {
		idLib::common->Printf( "====================================\n" );

		TestBlendJoints();
		TestConvertJointQuatsToJointMats();
		TestConvertJointMatsToJointQuats();
		TestTransformJoints();
		TestUntransformJoints();
		TestTransformVerts();
		TestTracePointCull();
		TestDecalPointCull();
		TestOverlayPointCull();
		TestDeriveTriPlanes();
		TestDeriveTangents();
		TestDeriveUnsmoothedTangents();
		TestNormalizeTangents();
		TestCreateShadowCache();
		TestConvertRGTCFromRGBA8();
	}

	if ( testBits & 8 ) {
		idLib::common->Printf( "====================================\n" );

		TestSoundUpSampling();
		TestSoundMixing();
	}

	idLib::common->SetRefreshOnPrint( false );

	delete p_generic;
	p_simd = NULL;
	p_generic = NULL;

#ifdef _WIN32
	SetThreadPriority( GetCurrentThread(), THREAD_PRIORITY_NORMAL );
#endif /* _WIN32 */
}