drewcassidy/nvidia-texture-tools
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

More cleanup. Remove files that are not strictly required.

Browse Source
This commit is contained in:
castano
2009-03-01 02:38:24 +00:00
parent 88fc5ca18e
commit 03c9ec0f62
26 changed files with 0 additions and 3939 deletions
Download Patch File Download Diff File Expand all files Collapse all files

168
src/nvcore/BitArray.h
View File

@ -1,168 +0,0 @@
// This code is in the public domain -- castanyo@yahoo.es
#ifndef NV_CORE_BITARRAY_H
#define NV_CORE_BITARRAY_H
#include <nvcore/nvcore.h>
#include <nvcore/Containers.h>
namespace nv
{
/// Count the bits of @a x.
inline uint bitsSet(uint8 x) {
uint count = 0;
for(; x != 0; x >>= 1) {
count += (x & 1);
}
return count;
}
/// Count the bits of @a x.
inline uint bitsSet(uint32 x, int bits) {
uint count = 0;
for(; x != 0 && bits != 0; x >>= 1, bits--) {
count += (x & 1);
}
return count;
}
/// Simple bit array.
class BitArray
{
public:
/// Default ctor.
BitArray() {}
/// Ctor with initial m_size.
BitArray(uint sz)
{
resize(sz);
}
/// Get array m_size.
uint size() const { return m_size; }
/// Clear array m_size.
void clear() { resize(0); }
/// Set array m_size.
void resize(uint sz)
{
m_size = sz;
m_bitArray.resize( (m_size + 7) >> 3 );
}
/// Get bit.
bool bitAt(uint b) const
{
nvDebugCheck( b < m_size );
return (m_bitArray[b >> 3] & (1 << (b & 7))) != 0;
}
/// Set a bit.
void setBitAt(uint b)
{
nvDebugCheck( b < m_size );
m_bitArray[b >> 3] |= (1 << (b & 7));
}
/// Clear a bit.
void clearBitAt( uint b )
{
nvDebugCheck( b < m_size );
m_bitArray[b >> 3] &= ~(1 << (b & 7));
}
/// Clear all the bits.
void clearAll()
{
memset(m_bitArray.mutableBuffer(), 0, m_bitArray.size());
}
/// Set all the bits.
void setAll()
{
memset(m_bitArray.mutableBuffer(), 0xFF, m_bitArray.size());
}
/// Toggle all the bits.
void toggleAll()
{
const uint byte_num = m_bitArray.size();
for(uint b = 0; b < byte_num; b++) {
m_bitArray[b] ^= 0xFF;
}
}
/// Get a byte of the bit array.
const uint8 & byteAt(uint index) const
{
return m_bitArray[index];
}
/// Set the given byte of the byte array.
void setByteAt(uint index, uint8 b)
{
m_bitArray[index] = b;
}
/// Count the number of bits set.
uint countSetBits() const
{
const uint num = m_bitArray.size();
if( num == 0 ) {
return 0;
}
uint count = 0;
for(uint i = 0; i < num - 1; i++) {
count += bitsSet(m_bitArray[i]);
}
count += bitsSet(m_bitArray[num-1], m_size & 0x7);
//piDebugCheck(count + countClearBits() == m_size);
return count;
}
/// Count the number of bits clear.
uint countClearBits() const {
const uint num = m_bitArray.size();
if( num == 0 ) {
return 0;
}
uint count = 0;
for(uint i = 0; i < num - 1; i++) {
count += bitsSet(~m_bitArray[i]);
}
count += bitsSet(~m_bitArray[num-1], m_size & 0x7);
//piDebugCheck(count + countSetBits() == m_size);
return count;
}
friend void swap(BitArray & a, BitArray & b)
{
swap(a.m_size, b.m_size);
swap(a.m_bitArray, b.m_bitArray);
}
private:
/// Number of bits stored.
uint m_size;
/// Array of bits.
Array<uint8> m_bitArray;
};
} // nv namespace
#endif // _PI_CORE_BITARRAY_H_

37
src/nvcore/CMakeLists.txt
View File

@ -7,9 +7,6 @@ SET(CORE_SRCS
DefsGnucWin32.h
DefsVcWin32.h
Ptr.h
RefCounted.h
RefCounted.cpp
BitArray.h
Memory.h
Memory.cpp
Debug.h
@ -17,10 +14,6 @@ SET(CORE_SRCS
Containers.h
StrLib.h
StrLib.cpp
Radix.h
Radix.cpp
CpuInfo.h
CpuInfo.cpp
Algorithms.h
Timer.h
Library.h
@ -31,41 +24,11 @@ SET(CORE_SRCS
TextReader.cpp
TextWriter.h
TextWriter.cpp
Tokenizer.h
Tokenizer.cpp
FileSystem.h
FileSystem.cpp)
INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
# For Windows64 in MSVC we need to add the assembly version of vsscanf
IF(MSVC AND NV_SYSTEM_PROCESSOR STREQUAL "AMD64")
SET(VSSCANF_ASM_NAME "vsscanf_proxy_win64")
IF(MSVC_IDE)
# $(IntDir) is a macro expanded to the intermediate directory of the selected solution configuration
SET(VSSCANF_ASM_INTDIR "$(IntDir)")
ELSE(MSVC_IDE)
# For some reason the NMake generator doesn't work properly with the generated .obj source:
# it requires the absolute path. So this is a hack which worked as of cmake 2.6.0 patch 0
GET_FILENAME_COMPONENT(VSSCANF_ASM_INTDIR
"${nvcore_BINARY_DIR}/CMakeFiles/nvcore.dir" ABSOLUTE)
ENDIF(MSVC_IDE)
SET(VSSCANF_ASM_SRC "${CMAKE_CURRENT_SOURCE_DIR}/${VSSCANF_ASM_NAME}.masm")
SET(VSSCANF_ASM_OBJ "${VSSCANF_ASM_INTDIR}/${VSSCANF_ASM_NAME}.obj")
# Adds the assembly output to the sources and adds the custom command to generate it
SET(CORE_SRCS
${CORE_SRCS}
${VSSCANF_ASM_OBJ}
)
ADD_CUSTOM_COMMAND(OUTPUT ${VSSCANF_ASM_OBJ}
MAIN_DEPENDENCY ${VSSCANF_ASM_SRC}
COMMAND ml64
ARGS /nologo /Fo ${VSSCANF_ASM_OBJ} /c /Cx ${VSSCANF_ASM_SRC}
)
ENDIF(MSVC AND NV_SYSTEM_PROCESSOR STREQUAL "AMD64")
# targets
ADD_DEFINITIONS(-DNVCORE_EXPORTS)

59
src/nvcore/Constraints.h
View File

@ -1,59 +0,0 @@
// This code is in the public domain -- castanyo@yahoo.es
#ifndef NV_CORE_ALGORITHMS_H
#define NV_CORE_ALGORITHMS_H
#include <nvcore/nvcore.h>
namespace nv
{
// Cool constraints from "Imperfect C++"
// must_be_pod
template <typename T>
struct must_be_pod
{
static void constraints()
{
union { T T_is_not_POD_type; };
}
};
// must_be_pod_or_void
template <typename T>
struct must_be_pod_or_void
{
static void constraints()
{
union { T T_is_not_POD_type; };
}
};
template <> struct must_be_pod_or_void<void> {};
// size_of
template <typename T>
struct size_of
{
enum { value = sizeof(T) };
};
template <>
struct size_of<void>
{
enum { value = 0 };
};
// must_be_same_size
template <typename T1, typename T2>
struct must_be_same_size
{
static void constraints()
{
const int T1_not_same_size_as_T2 = size_of<T1>::value == size_of<T2>::value;
int i[T1_not_same_size_as_T2];
}
};
} // nv namespace
#endif // NV_CORE_ALGORITHMS_H

162
src/nvcore/CpuInfo.cpp
View File

@ -1,162 +0,0 @@
// This code is in the public domain -- castanyo@yahoo.es
#include <nvcore/CpuInfo.h>
#include <nvcore/Debug.h>
using namespace nv;
#if NV_OS_WIN32
#define _WIN32_WINNT 0x0501
#define WIN32_LEAN_AND_MEAN
#include <windows.h>
typedef BOOL (WINAPI *LPFN_ISWOW64PROCESS) (HANDLE, PBOOL);
static bool isWow64()
{
LPFN_ISWOW64PROCESS fnIsWow64Process = (LPFN_ISWOW64PROCESS)GetProcAddress(GetModuleHandle(TEXT("kernel32")), "IsWow64Process");
BOOL bIsWow64 = FALSE;
if (NULL != fnIsWow64Process)
{
if (!fnIsWow64Process(GetCurrentProcess(), &bIsWow64))
{
return false;
}
}
return bIsWow64 == TRUE;
}
#endif // NV_OS_WIN32
#if NV_OS_LINUX
#include <string.h>
#include <sched.h>
#endif // NV_OS_LINUX
#if NV_OS_DARWIN
#include <sys/types.h>
#include <sys/sysctl.h>
#endif // NV_OS_DARWIN
// Initialize the data and the local defines, which are designed
// to match the positions in cpuid
uint CpuInfo::m_cpu = ~0x0;
uint CpuInfo::m_procCount = 0;
#define NV_CPUINFO_MMX_MASK (1<<23)
#define NV_CPUINFO_SSE_MASK (1<<25)
#define NV_CPUINFO_SSE2_MASK (1<<26)
#define NV_CPUINFO_SSE3_MASK (1)
uint CpuInfo::processorCount()
{
if (m_procCount == 0) {
#if NV_OS_WIN32
SYSTEM_INFO sysInfo;
typedef BOOL (WINAPI *LPFN_ISWOW64PROCESS) (HANDLE, PBOOL);
if (isWow64())
{
GetNativeSystemInfo(&sysInfo);
}
else
{
GetSystemInfo(&sysInfo);
}
uint count = (uint)sysInfo.dwNumberOfProcessors;
m_procCount = count;
#elif NV_OS_LINUX
// Code from x264 (July 6 snapshot) cpu.c:271
uint bit;
uint np;
cpu_set_t p_aff;
memset( &p_aff, 0, sizeof(p_aff) );
sched_getaffinity( 0, sizeof(p_aff), &p_aff );
for( np = 0, bit = 0; bit < sizeof(p_aff); bit++ )
np += (((uint8 *)&p_aff)[bit / 8] >> (bit % 8)) & 1;
m_procCount = np;
#elif NV_OS_DARWIN
// Code from x264 (July 6 snapshot) cpu.c:286
uint numberOfCPUs;
size_t length = sizeof( numberOfCPUs );
if( sysctlbyname("hw.ncpu", &numberOfCPUs, &length, NULL, 0) )
{
numberOfCPUs = 1;
}
m_procCount = numberOfCPUs;
#else
m_procCount = 1;
#endif
}
nvDebugCheck(m_procCount > 0);
return m_procCount;
}
uint CpuInfo::coreCount()
{
return 1;
}
bool CpuInfo::hasMMX()
{
return (cpu() & NV_CPUINFO_MMX_MASK) != 0;
}
bool CpuInfo::hasSSE()
{
return (cpu() & NV_CPUINFO_SSE_MASK) != 0;
}
bool CpuInfo::hasSSE2()
{
return (cpu() & NV_CPUINFO_SSE2_MASK) != 0;
}
bool CpuInfo::hasSSE3()
{
return (cpu() & NV_CPUINFO_SSE3_MASK) != 0;
}
inline int CpuInfo::cpu() {
if (m_cpu == ~0x0) {
m_cpu = 0;
#if NV_CC_MSVC
int CPUInfo[4] = {-1};
__cpuid(CPUInfo, /*InfoType*/ 1);
if (CPUInfo[2] & NV_CPUINFO_SSE3_MASK) {
m_cpu |= NV_CPUINFO_SSE3_MASK;
}
if (CPUInfo[3] & NV_CPUINFO_MMX_MASK) {
m_cpu |= NV_CPUINFO_MMX_MASK;
}
if (CPUInfo[3] & NV_CPUINFO_SSE_MASK) {
m_cpu |= NV_CPUINFO_SSE_MASK;
}
if (CPUInfo[3] & NV_CPUINFO_SSE2_MASK) {
m_cpu |= NV_CPUINFO_SSE2_MASK;
}
#elif NV_CC_GNUC
// TODO: add the proper inline assembly
#if NV_CPU_X86
#elif NV_CPU_X86_64
#endif // NV_CPU_X86_64
#endif // NV_CC_GNUC
}
return m_cpu;
}

109
src/nvcore/CpuInfo.h
View File

@ -1,109 +0,0 @@
// This code is in the public domain -- castanyo@yahoo.es
#ifndef NV_CORE_CPUINFO_H
#define NV_CORE_CPUINFO_H
#include <nvcore/nvcore.h>
#if NV_CC_MSVC
#if _MSC_VER >= 1400
# include <intrin.h> // __rdtsc
#endif
#endif
namespace nv
{
// CPU Information.
class CpuInfo
{
protected:
static int cpu();
private:
// Cache of the CPU data
static uint m_cpu;
static uint m_procCount;
public:
static uint processorCount();
static uint coreCount();
static bool hasMMX();
static bool hasSSE();
static bool hasSSE2();
static bool hasSSE3();
};
#if NV_CC_MSVC
#if _MSC_VER < 1400
inline uint64 rdtsc()
{
uint64 t;
__asm rdtsc
__asm mov DWORD PTR [t], eax
__asm mov DWORD PTR [t+4], edx
return t;
}
#else
#pragma intrinsic(__rdtsc)
inline uint64 rdtsc()
{
return __rdtsc();
}
#endif
#endif
#if NV_CC_GNUC
#if defined(__i386__)
inline /*volatile*/ uint64 rdtsc()
{
uint64 x;
//__asm__ volatile ("rdtsc" : "=A" (x));
__asm__ volatile (".byte 0x0f, 0x31" : "=A" (x));
return x;
}
#elif defined(__x86_64__)
static __inline__ uint64 rdtsc(void)
{
unsigned int hi, lo;
__asm__ __volatile__ ("rdtsc" : "=a"(lo), "=d"(hi));
return ( (unsigned long long)lo)|( ((unsigned long long)hi)<<32 );
}
#elif defined(__powerpc__)
static __inline__ uint64 rdtsc(void)
{
uint64 result=0;
unsigned long int upper, lower, tmp;
__asm__ volatile(
"0: \n"
"\tmftbu %0 \n"
"\tmftb %1 \n"
"\tmftbu %2 \n"
"\tcmpw %2,%0 \n"
"\tbne 0b \n"
: "=r"(upper),"=r"(lower),"=r"(tmp)
);
result = upper;
result = result<<32;
result = result|lower;
return(result);
}
#endif
#endif // NV_CC_GNUC
} // nv namespace
#endif // NV_CORE_CPUINFO_H

30
src/nvcore/Prefetch.h
View File

@ -1,30 +0,0 @@
// This code is in the public domain -- castanyo@yahoo.es
#ifndef NV_CORE_PREFETCH_H
#define NV_CORE_PREFETCH_H
#include <nvcore/nvcore.h>
// nvPrefetch
#if NV_CC_GNUC
#define nvPrefetch(ptr) __builtin_prefetch(ptr)
#elif NV_CC_MSVC
// Uses SSE Intrinsics for both x86 and x86_64
#include <xmmintrin.h>
__forceinline void nvPrefetch(const void * mem)
{
_mm_prefetch(static_cast<const char*>(mem), _MM_HINT_T0); /* prefetcht0 */
// _mm_prefetch(static_cast<const char*>(mem), _MM_HINT_NTA); /* prefetchnta */
}
#else
// do nothing in other case.
#define nvPrefetch(ptr)
#endif // NV_CC_MSVC
#endif // NV_CORE_PREFETCH_H

484
src/nvcore/Radix.cpp
View File

@ -1,484 +0,0 @@
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/**
* Contains source code from the article "Radix Sort Revisited".
* \file Radix.cpp
* \author Pierre Terdiman
* \date April, 4, 2000
*/
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// References:
// http://www.codercorner.com/RadixSortRevisited.htm
// http://www.stereopsis.com/radix.html
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/**
* Revisited Radix Sort.
* This is my new radix routine:
* - it uses indices and doesn't recopy the values anymore, hence wasting less ram
* - it creates all the histograms in one run instead of four
* - it sorts words faster than dwords and bytes faster than words
* - it correctly sorts negative floating-point values by patching the offsets
* - it automatically takes advantage of temporal coherence
* - multiple keys support is a side effect of temporal coherence
* - it may be worth recoding in asm... (mainly to use FCOMI, FCMOV, etc) [it's probably memory-bound anyway]
*
* History:
* - 08.15.98: very first version
* - 04.04.00: recoded for the radix article
* - 12.xx.00: code lifting
* - 09.18.01: faster CHECK_PASS_VALIDITY thanks to Mark D. Shattuck (who provided other tips, not included here)
* - 10.11.01: added local ram support
* - 01.20.02: bugfix! In very particular cases the last pass was skipped in the float code-path, leading to incorrect sorting......
* - 01.02.02: - "mIndices" renamed => "mRanks". That's a rank sorter after all.
* - ranks are not "reset" anymore, but implicit on first calls
* - 07.05.02: offsets rewritten with one less indirection.
* - 11.03.02: "bool" replaced with RadixHint enum
* - 07.15.04: stack-based radix added
* - we want to use the radix sort but without making it static, and without allocating anything.
* - we internally allocate two arrays of ranks. Each of them has N uint32s to sort N values.
* - 1Mb/2/sizeof(uint32) = 131072 values max, at the same time.
* - 09.22.04: - adapted to MacOS by Chris Lamb
* - 01.12.06: - added optimizations suggested by Kyle Hubert
* - 04.06.08: - Fix bug negative zero sorting bug by Ignacio Castaño
*
* \class RadixSort
* \author Pierre Terdiman
* \version 1.5
* \date August, 15, 1998
*/
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Header
#include <nvcore/Radix.h>
#include <string.h> // memset
//using namespace IceCore;
#define INVALIDATE_RANKS mCurrentSize|=0x80000000
#define VALIDATE_RANKS mCurrentSize&=0x7fffffff
#define CURRENT_SIZE (mCurrentSize&0x7fffffff)
#define INVALID_RANKS (mCurrentSize&0x80000000)
#if NV_BIG_ENDIAN
#define H0_OFFSET 768
#define H1_OFFSET 512
#define H2_OFFSET 256
#define H3_OFFSET 0
#define BYTES_INC (3-j)
#else
#define H0_OFFSET 0
#define H1_OFFSET 256
#define H2_OFFSET 512
#define H3_OFFSET 768
#define BYTES_INC j
#endif
#define CREATE_HISTOGRAMS(type, buffer) \
/* Clear counters/histograms */ \
memset(mHistogram, 0, 256*4*sizeof(uint32)); \
\
/* Prepare to count */ \
const uint8* p = (const uint8*)input; \
const uint8* pe = &p[nb*4]; \
uint32* h0= &mHistogram[H0_OFFSET]; /* Histogram for first pass (LSB) */ \
uint32* h1= &mHistogram[H1_OFFSET]; /* Histogram for second pass */ \
uint32* h2= &mHistogram[H2_OFFSET]; /* Histogram for third pass */ \
uint32* h3= &mHistogram[H3_OFFSET]; /* Histogram for last pass (MSB) */ \
\
bool AlreadySorted = true; /* Optimism... */ \
\
if(INVALID_RANKS) \
{ \
/* Prepare for temporal coherence */ \
type* Running = (type*)buffer; \
type PrevVal = *Running; \
\
while(p!=pe) \
{ \
/* Read input buffer in previous sorted order */ \
type Val = *Running++; \
/* Check whether already sorted or not */ \
if(Val<PrevVal) { AlreadySorted = false; break; } /* Early out */ \
/* Update for next iteration */ \
PrevVal = Val; \
\
/* Create histograms */ \
h0[*p++]++; h1[*p++]++; h2[*p++]++; h3[*p++]++; \
} \
\
/* If all input values are already sorted, we just have to return and leave the */ \
/* previous list unchanged. That way the routine may take advantage of temporal */ \
/* coherence, for example when used to sort transparent faces. */ \
if(AlreadySorted) \
{ \
mNbHits++; \
for(uint32 i=0;i<nb;i++) mRanks[i] = i; \
return *this; \
} \
} \
else \
{ \
/* Prepare for temporal coherence */ \
const uint32* Indices = mRanks; \
type PrevVal = (type)buffer[*Indices]; \
\
while(p!=pe) \
{ \
/* Read input buffer in previous sorted order */ \
type Val = (type)buffer[*Indices++]; \
/* Check whether already sorted or not */ \
if(Val<PrevVal) { AlreadySorted = false; break; } /* Early out */ \
/* Update for next iteration */ \
PrevVal = Val; \
\
/* Create histograms */ \
h0[*p++]++; h1[*p++]++; h2[*p++]++; h3[*p++]++; \
} \
\
/* If all input values are already sorted, we just have to return and leave the */ \
/* previous list unchanged. That way the routine may take advantage of temporal */ \
/* coherence, for example when used to sort transparent faces. */ \
if(AlreadySorted) { mNbHits++; return *this; } \
} \
\
/* Else there has been an early out and we must finish computing the histograms */ \
while(p!=pe) \
{ \
/* Create histograms without the previous overhead */ \
h0[*p++]++; h1[*p++]++; h2[*p++]++; h3[*p++]++; \
}
#define CHECK_PASS_VALIDITY(pass) \
/* Shortcut to current counters */ \
const uint32* CurCount = &mHistogram[pass<<8]; \
\
/* Reset flag. The sorting pass is supposed to be performed. (default) */ \
bool PerformPass = true; \
\
/* Check pass validity */ \
\
/* If all values have the same byte, sorting is useless. */ \
/* It may happen when sorting bytes or words instead of dwords. */ \
/* This routine actually sorts words faster than dwords, and bytes */ \
/* faster than words. Standard running time (O(4*n))is reduced to O(2*n) */ \
/* for words and O(n) for bytes. Running time for floats depends on actual values... */ \
\
/* Get first byte */ \
uint8 UniqueVal = *(((uint8*)input)+pass); \
\
/* Check that byte's counter */ \
if(CurCount[UniqueVal]==nb) PerformPass=false;
using namespace nv;
/// Constructor.
RadixSort::RadixSort() : mRanks(NULL), mRanks2(NULL), mCurrentSize(0), mTotalCalls(0), mNbHits(0), mDeleteRanks(true)
{
// Initialize indices
INVALIDATE_RANKS;
}
/// Destructor.
RadixSort::~RadixSort()
{
// Release everything
if(mDeleteRanks)
{
delete [] mRanks2;
delete [] mRanks;
}
}
/// Resizes the inner lists.
/// \param nb [in] new size (number of dwords)
/// \return true if success
bool RadixSort::resize(uint32 nb)
{
if(mDeleteRanks)
{
// Free previously used ram
delete [] mRanks2;
delete [] mRanks;
// Get some fresh one
mRanks = new uint32[nb];
mRanks2 = new uint32[nb];
}
return true;
}
inline void RadixSort::checkResize(uint32 nb)
{
uint32 CurSize = CURRENT_SIZE;
if(nb!=CurSize)
{
if(nb>CurSize) resize(nb);
mCurrentSize = nb;
INVALIDATE_RANKS;
}
}
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/**
* Main sort routine.
* This one is for integer values. After the call, mIndices contains a list of indices in sorted order, i.e. in the order you may process your data.
* \param input [in] a list of integer values to sort
* \param nb [in] number of values to sort
* \param signedvalues [in] true to handle negative values, false if you know your input buffer only contains positive values
* \return Self-Reference
*/
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
RadixSort& RadixSort::sort(const uint32* input, uint32 nb, bool signedValues/*=true*/)
{
// Checkings
if(!input || !nb || nb&0x80000000) return *this;
// Stats
mTotalCalls++;
// Resize lists if needed
checkResize(nb);
// Allocate histograms & offsets on the stack
uint32 mHistogram[256*4];
uint32* mLink[256];
// Create histograms (counters). Counters for all passes are created in one run.
// Pros: read input buffer once instead of four times
// Cons: mHistogram is 4Kb instead of 1Kb
// We must take care of signed/unsigned values for temporal coherence.... I just
// have 2 code paths even if just a single opcode changes. Self-modifying code, someone?
if(!signedValues) { CREATE_HISTOGRAMS(uint32, input); }
else { CREATE_HISTOGRAMS(int32, input); }
// Radix sort, j is the pass number (0=LSB, 3=MSB)
for(uint32 j=0;j<4;j++)
{
CHECK_PASS_VALIDITY(j);
// Sometimes the fourth (negative) pass is skipped because all numbers are negative and the MSB is 0xFF (for example). This is
// not a problem, numbers are correctly sorted anyway.
if(PerformPass)
{
// Should we care about negative values?
if(j!=3 || !signedValues)
{
// Here we deal with positive values only
// Create offsets
mLink[0] = mRanks2;
for(uint32 i=1;i<256;i++) mLink[i] = mLink[i-1] + CurCount[i-1];
}
else
{
// This is a special case to correctly handle negative integers. They're sorted in the right order but at the wrong place.
mLink[128] = mRanks2;
for(uint32 i=129;i<256;i++) mLink[i] = mLink[i-1] + CurCount[i-1];
mLink[0] = mLink[255] + CurCount[255];
for(uint32 i=1;i<128;i++) mLink[i] = mLink[i-1] + CurCount[i-1];
}
// Perform Radix Sort
const uint8* InputBytes = (const uint8*)input;
InputBytes += BYTES_INC;
if(INVALID_RANKS)
{
for(uint32 i=0;i<nb;i++) *mLink[InputBytes[i<<2]]++ = i;
VALIDATE_RANKS;
}
else
{
const uint32* Indices = mRanks;
const uint32* IndicesEnd = &mRanks[nb];
while(Indices!=IndicesEnd)
{
uint32 id = *Indices++;
*mLink[InputBytes[id<<2]]++ = id;
}
}
// Swap pointers for next pass. Valid indices - the most recent ones - are in mRanks after the swap.
uint32* Tmp = mRanks;
mRanks = mRanks2;
mRanks2 = Tmp;
}
}
return *this;
}
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/**
* Main sort routine.
* This one is for floating-point values. After the call, mIndices contains a list of indices in sorted order, i.e. in the order you may process your data.
* \param input [in] a list of floating-point values to sort
* \param nb [in] number of values to sort
* \return Self-Reference
* \warning only sorts IEEE floating-point values
*/
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
RadixSort& RadixSort::sort(const float* input2, uint32 nb)
{
// Checkings
if(!input2 || !nb || nb&0x80000000) return *this;
// Stats
mTotalCalls++;
const uint32* input = (const uint32*)input2;
// Resize lists if needed
checkResize(nb);
// Allocate histograms & offsets on the stack
uint32 mHistogram[256*4];
uint32* mLink[256];
// Create histograms (counters). Counters for all passes are created in one run.
// Pros: read input buffer once instead of four times
// Cons: mHistogram is 4Kb instead of 1Kb
// Floating-point values are always supposed to be signed values, so there's only one code path there.
// Please note the floating point comparison needed for temporal coherence! Although the resulting asm code
// is dreadful, this is surprisingly not such a performance hit - well, I suppose that's a big one on first
// generation Pentiums....We can't make comparison on integer representations because, as Chris said, it just
// wouldn't work with mixed positive/negative values....
{ CREATE_HISTOGRAMS(float, input2); }
// Radix sort, j is the pass number (0=LSB, 3=MSB)
for(uint32 j=0;j<4;j++)
{
// Should we care about negative values?
if(j!=3)
{
// Here we deal with positive values only
CHECK_PASS_VALIDITY(j);
if(PerformPass)
{
// Create offsets
mLink[0] = mRanks2;
for(uint32 i=1;i<256;i++) mLink[i] = mLink[i-1] + CurCount[i-1];
// Perform Radix Sort
const uint8* InputBytes = (const uint8*)input;
InputBytes += BYTES_INC;
if(INVALID_RANKS)
{
for(uint32 i=0;i<nb;i++) *mLink[InputBytes[i<<2]]++ = i;
VALIDATE_RANKS;
}
else
{
const uint32* Indices = mRanks;
const uint32* IndicesEnd = &mRanks[nb];
while(Indices!=IndicesEnd)
{
uint32 id = *Indices++;
*mLink[InputBytes[id<<2]]++ = id;
}
}
// Swap pointers for next pass. Valid indices - the most recent ones - are in mRanks after the swap.
uint32* Tmp = mRanks;
mRanks = mRanks2;
mRanks2 = Tmp;
}
}
else
{
// This is a special case to correctly handle negative values
CHECK_PASS_VALIDITY(j);
if(PerformPass)
{
mLink[255] = mRanks2 + CurCount[255];
for(uint32 i = 254; i > 126; i--) mLink[i] = mLink[i+1] + CurCount[i];
mLink[0] = mLink[127] + CurCount[127];
for(uint32 i = 1; i < 127; i++) mLink[i] = mLink[i-1] + CurCount[i-1];
// Perform Radix Sort
if(INVALID_RANKS)
{
for(uint32 i=0;i<nb;i++)
{
uint32 Radix = input[i]>>24; // Radix byte, same as above. AND is useless here (uint32).
// ### cmp to be killed. Not good. Later.
if(Radix<128) *mLink[Radix]++ = i; // Number is positive, same as above
else *(--mLink[Radix]) = i; // Number is negative, flip the sorting order
}
VALIDATE_RANKS;
}
else
{
for(uint32 i=0;i<nb;i++)
{
uint32 Radix = input[mRanks[i]]>>24; // Radix byte, same as above. AND is useless here (uint32).
// ### cmp to be killed. Not good. Later.
if(Radix<128) *mLink[Radix]++ = mRanks[i]; // Number is positive, same as above
else *(--mLink[Radix]) = mRanks[i]; // Number is negative, flip the sorting order
}
}
// Swap pointers for next pass. Valid indices - the most recent ones - are in mRanks after the swap.
uint32* Tmp = mRanks;
mRanks = mRanks2;
mRanks2 = Tmp;
}
else
{
// The pass is useless, yet we still have to reverse the order of current list if all values are negative.
if(UniqueVal>=128)
{
if(INVALID_RANKS)
{
// ###Possible?
for(uint32 i=0;i<nb;i++) mRanks2[i] = nb-i-1;
VALIDATE_RANKS;
}
else
{
for(uint32 i=0;i<nb;i++) mRanks2[i] = mRanks[nb-i-1];
}
// Swap pointers for next pass. Valid indices - the most recent ones - are in mRanks after the swap.
uint32* Tmp = mRanks;
mRanks = mRanks2;
mRanks2 = Tmp;
}
}
}
}
return *this;
}
bool RadixSort::setRankBuffers(uint32* ranks0, uint32* ranks1)
{
if(!ranks0 || !ranks1) return false;
mRanks = ranks0;
mRanks2 = ranks1;
mDeleteRanks = false;
return true;
}
RadixSort & RadixSort::sort(const Array<int> & input)
{
return sort((const uint32 *)input.buffer(), input.count(), true);
}
RadixSort & RadixSort::sort(const Array<uint> & input)
{
return sort(input.buffer(), input.count(), false);
}
RadixSort & RadixSort::sort(const Array<float> & input)
{
return sort(input.buffer(), input.count());
}

73
src/nvcore/Radix.h
View File

@ -1,73 +0,0 @@
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
/**
* Contains source code from the article "Radix Sort Revisited".
* \file Radix.h
* \author Pierre Terdiman
* \date April, 4, 2000
*/
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Include Guard
#ifndef NV_CORE_RADIXSORT_H
#define NV_CORE_RADIXSORT_H
#include <nvcore/nvcore.h>
#include <nvcore/Containers.h>
namespace nv
{
class NVCORE_CLASS RadixSort
{
NV_FORBID_COPY(RadixSort);
public:
// Constructor/Destructor
RadixSort();
~RadixSort();
// Sorting methods
RadixSort & sort(const uint32* input, uint32 nb, bool signedValues=true);
RadixSort & sort(const float* input, uint32 nb);
// Helpers
RadixSort & sort(const Array<int> & input);
RadixSort & sort(const Array<uint> & input);
RadixSort & sort(const Array<float> & input);
//! Access to results. mRanks is a list of indices in sorted order, i.e. in the order you may further process your data
inline /*const*/ uint32 * ranks() /*const*/ { return mRanks; }
//! mIndices2 gets trashed on calling the sort routine, but otherwise you can recycle it the way you want.
inline uint32 * recyclable() const { return mRanks2; }
// Stats
//! Returns the total number of calls to the radix sorter.
inline uint32 totalCalls() const { return mTotalCalls; }
//! Returns the number of early exits due to temporal coherence.
inline uint32 hits() const { return mNbHits; }
bool setRankBuffers(uint32* ranks0, uint32* ranks1);
private:
uint32 mCurrentSize; //!< Current size of the indices list
uint32 * mRanks; //!< Two lists, swapped each pass
uint32 * mRanks2;
// Stats
uint32 mTotalCalls; //!< Total number of calls to the sort routine
uint32 mNbHits; //!< Number of early exits due to coherence
// Stack-radix
bool mDeleteRanks; //!<
// Internal methods
void checkResize(uint32 nb);
bool resize(uint32 nb);
};
} // nv namespace
#endif // NV_CORE_RADIXSORT_H

9
src/nvcore/RefCounted.cpp
View File

@ -1,9 +0,0 @@
// This code is in the public domain -- castanyo@yahoo.es
#include "RefCounted.h"
using namespace nv;
int nv::RefCounted::s_total_ref_count = 0;
int nv::RefCounted::s_total_obj_count = 0;

119
src/nvcore/RefCounted.h
View File

@ -1,119 +0,0 @@
// This code is in the public domain -- castanyo@yahoo.es
#ifndef NV_CORE_REFCOUNTED_H
#define NV_CORE_REFCOUNTED_H
#include <nvcore/nvcore.h>
#include <nvcore/Debug.h>
#define NV_DECLARE_PTR(Class) \
template <class T> class SmartPtr; \
typedef SmartPtr<class Class> Class ## Ptr; \
typedef SmartPtr<const class Class> Class ## ConstPtr
namespace nv
{
/// Reference counted base class to be used with SmartPtr and WeakPtr.
class RefCounted
{
NV_FORBID_COPY(RefCounted);
public:
/// Ctor.
RefCounted() : m_count(0)/*, m_weak_proxy(NULL)*/
{
s_total_obj_count++;
}
/// Virtual dtor.
virtual ~RefCounted()
{
nvCheck( m_count == 0 );
nvCheck( s_total_obj_count > 0 );
s_total_obj_count--;
}
/// Increase reference count.
uint addRef() const
{
s_total_ref_count++;
m_count++;
return m_count;
}
/// Decrease reference count and remove when 0.
uint release() const
{
nvCheck( m_count > 0 );
s_total_ref_count--;
m_count--;
if( m_count == 0 ) {
// releaseWeakProxy();
delete this;
return 0;
}
return m_count;
}
/*
/// Get weak proxy.
WeakProxy * getWeakProxy() const
{
if (m_weak_proxy == NULL) {
m_weak_proxy = new WeakProxy;
m_weak_proxy->AddRef();
}
return m_weak_proxy;
}
/// Release the weak proxy.
void releaseWeakProxy() const
{
if (m_weak_proxy != NULL) {
m_weak_proxy->NotifyObjectDied();
m_weak_proxy->Release();
m_weak_proxy = NULL;
}
}
*/
/** @name Debug methods: */
//@{
/// Get reference count.
int refCount() const
{
return m_count;
}
/// Get total number of objects.
static int totalObjectCount()
{
return s_total_obj_count;
}
/// Get total number of references.
static int totalReferenceCount()
{
return s_total_ref_count;
}
//@}
private:
NVCORE_API static int s_total_ref_count;
NVCORE_API static int s_total_obj_count;
mutable int m_count;
// mutable WeakProxy * weak_proxy;
};
} // nv namespace
#endif // NV_CORE_REFCOUNTED_H

259
src/nvcore/Tokenizer.cpp
View File

@ -1,259 +0,0 @@
// This code is in the public domain -- castano@gmail.com
#include "Tokenizer.h"
#include <nvcore/StrLib.h>
#include <stdio.h> // vsscanf
#include <stdarg.h> // va_list
#include <stdlib.h> // atof, atoi
#if NV_CC_MSVC
#if defined NV_CPU_X86
/* vsscanf for Win32
* Written 5/2003 by <mgix@mgix.com>
* This code is in the Public Domain
*/
#include <malloc.h> // alloca
//#include <string.h>
static int vsscanf(const char * buffer, const char * format, va_list argPtr)
{
// Get an upper bound for the # of args
size_t count = 0;
const char *p = format;
while(1) {
char c = *(p++);
if(c==0) break;
if(c=='%' && (p[0]!='*' && p[0]!='%')) ++count;
}
// Make a local stack
size_t stackSize = (2+count)*sizeof(void*);
void **newStack = (void**)alloca(stackSize);
// Fill local stack the way sscanf likes it
newStack[0] = (void*)buffer;
newStack[1] = (void*)format;
memcpy(newStack+2, argPtr, count*sizeof(void*));
// @@ Use: CALL DWORD PTR [sscanf]
// Warp into system sscanf with new stack
int result;
void *savedESP;
__asm
{
mov savedESP, esp
mov esp, newStack
#if _MSC_VER >= 1400
call DWORD PTR [sscanf_s]
#else
call DWORD PTR [sscanf]
#endif
mov esp, savedESP
mov result, eax
}
return result;
}
#elif defined NV_CPU_X86_64
/* Prototype of the helper assembly function */
#ifdef __cplusplus
extern "C" {
#endif
int vsscanf_proxy_win64(const char * buffer, const char * format, va_list argPtr, __int64 count);
#ifdef __cplusplus
}
#endif
/* MASM64 version of the above vsscanf */
static int vsscanf(const char * buffer, const char * format, va_list argPtr)
{
// Get an upper bound for the # of args
__int64 count = 0;
const char *p = format;
while(1) {
char c = *(p++);
if(c==0) break;
if(c=='%' && (p[0]!='*' && p[0]!='%')) ++count;
}
return vsscanf_proxy_win64(buffer, format, argPtr, count);
}
/*#error vsscanf doesn't work on MSVC for x64*/
#else
#error Unknown processor for MSVC
#endif
#endif // NV_CC_MSVC
using namespace nv;
Token::Token() :
m_str(""), m_len(0)
{
}
Token::Token(const Token & token) :
m_str(token.m_str), m_len(token.m_len)
{
}
Token::Token(const char * str, int len) :
m_str(str), m_len(len)
{
}
bool Token::operator==(const char * str) const
{
return strncmp(m_str, str, m_len) == 0;
}
bool Token::operator!=(const char * str) const
{
return strncmp(m_str, str, m_len) != 0;
}
bool Token::isNull()
{
return m_len != 0;
}
float Token::toFloat() const
{
return float(atof(m_str));
}
int Token::toInt() const
{
return atoi(m_str);
}
uint Token::toUnsignedInt() const
{
// @@ TBD
return uint(atoi(m_str));
}
String Token::toString() const
{
return String(m_str, m_len);
}
bool Token::parse(const char * format, int count, ...) const
{
va_list arg;
va_start(arg, count);
int readCount = vsscanf(m_str, format, arg);
va_end(arg);
return readCount == count;
}
Tokenizer::Tokenizer(Stream * stream) :
m_reader(stream), m_lineNumber(0), m_columnNumber(0), m_delimiters("{}()="), m_spaces(" \t")
{
}
bool Tokenizer::nextLine(bool skipEmptyLines /*= true*/)
{
do {
if (!readLine()) {
return false;
}
}
while (!readToken() && skipEmptyLines);
return true;
}
bool Tokenizer::nextToken(bool skipEndOfLine /*= false*/)
{
if (!readToken()) {
if (!skipEndOfLine) {
return false;
}
else {
return nextLine(true);
}
}
return true;
}
bool Tokenizer::readToken()
{
skipSpaces();
const char * begin = m_line + m_columnNumber;
if (*begin == '\0') {
return false;
}
char c = readChar();
if (isDelimiter(c)) {
m_token = Token(begin, 1);
return true;
}
// @@ Add support for quoted tokens "", ''
int len = 0;
while (!isDelimiter(c) && !isSpace(c) && c != '\0') {
c = readChar();
len++;
}
m_columnNumber--;
m_token = Token(begin, len);
return true;
}
char Tokenizer::readChar()
{
return m_line[m_columnNumber++];
}
bool Tokenizer::readLine()
{
m_lineNumber++;
m_columnNumber = 0;
m_line = m_reader.readLine();
return m_line != NULL;
}
void Tokenizer::skipSpaces()
{
while (isSpace(readChar())) {}
m_columnNumber--;
}
bool Tokenizer::isSpace(char c)
{
uint i = 0;
while (m_spaces[i] != '\0') {
if (c == m_spaces[i]) {
return true;
}
i++;
}
return false;
}
bool Tokenizer::isDelimiter(char c)
{
uint i = 0;
while (m_delimiters[i] != '\0') {
if (c == m_delimiters[i]) {
return true;
}
i++;
}
return false;
}

98
src/nvcore/Tokenizer.h
View File

@ -1,98 +0,0 @@
// This code is in the public domain -- castano@gmail.com
#ifndef NV_CORE_TOKENIZER_H
#define NV_CORE_TOKENIZER_H
#include <nvcore/StrLib.h>
#include <nvcore/Stream.h>
#include <nvcore/TextReader.h>
namespace nv
{
/// A token produced by the Tokenizer.
class NVCORE_CLASS Token
{
public:
Token();
Token(const Token & token);
Token(const char * str, int len);
bool operator==(const char * str) const;
bool operator!=(const char * str) const;
bool isNull();
float toFloat() const;
int toInt() const;
uint toUnsignedInt() const;
String toString() const;
bool parse(const char * format, int count, ...) const __attribute__((format (scanf, 2, 4)));
private:
const char * m_str;
int m_len;
};
/// Exception thrown by the tokenizer.
class TokenizerException
{
public:
TokenizerException(int line, int column) : m_line(line), m_column(column) {}
int line() const { return m_line; }
int column() const { return m_column; }
private:
int m_line;
int m_column;
};
// @@ Use enums instead of bools for clarity!
//enum SkipEmptyLines { skipEmptyLines, noSkipEmptyLines };
//enum SkipEndOfLine { skipEndOfLine, noSkipEndOfLine };
/// A simple stream tokenizer.
class NVCORE_CLASS Tokenizer
{
public:
Tokenizer(Stream * stream);
bool nextLine(bool skipEmptyLines = true);
bool nextToken(bool skipEndOfLine = false);
const Token & token() const { return m_token; }
int lineNumber() const { return m_lineNumber; }
int columnNumber() const { return m_columnNumber; }
void setDelimiters(const char * str) { m_delimiters = str; }
const char * delimiters() const { return m_delimiters; }
void setSpaces(const char * str) { m_spaces = str; }
const char * spaces() const { return m_spaces; }
private:
char readChar();
bool readLine();
bool readToken();
void skipSpaces();
bool isSpace(char c);
bool isDelimiter(char c);
private:
TextReader m_reader;
const char * m_line;
Token m_token;
int m_lineNumber;
int m_columnNumber;
const char * m_delimiters;
const char * m_spaces;
};
} // nv namespace
#endif // NV_CORE_TOKENIZER_H

124
src/nvcore/vsscanf_proxy_win64.masm
View File

@ -1,124 +0,0 @@
; MASM x64 version of
; vsscanf for Win32
; originally written 5/2003 by <mgix@mgix.com>
;
; This was done because MSVC does not accept inline assembly code
; for the x64 platform, so this file implements almost the whole
; module in assembly using the amd64 ABI
;
; 06/17/2008 by edgarv [at] nvidia com
; Definition of memcpy
memcpy PROTO dest:Ptr, src:Ptr, numbytes:QWORD
; Definition of sscanf
sscanf PROTO buffer:Ptr Byte, format:Ptr Byte, args:VARARG
; Start a code segment named "_TEXT" by default
.CODE
; Entry point of our function: at this point we can use
; named parameters
ALIGN 16
PUBLIC vsscanf_proxy_win64
; Because the x64 code uses the fast call convention, only
; the arguments beyond the 4th one are available from the stack.
; The first four parameters are in RCX, RDX, R8 and R9
; Parameters:
; const char* buffer
; const char* format
; va_list argPtr
; size_t count
vsscanf_proxy_win64 PROC, \
buffer:PTR Byte, format:PTR Byte, argPtr:PTR, count:QWORD
; Allocates space for our local variable, savedRDP
sub rsp, 08h
; Copies the parameters from the registers to the memory: before warping to
; sscanf we will call memcpy, and those registers can just dissapear!
mov buffer, rcx
mov format, rdx
mov argPtr, r8
mov count, r9
; Allocate extra space in the stack for (2+count)*sizeof(void*),
; this is (2+count)*(8)
mov r10, r9
add r10, 2 ; count += 2
sal r10, 3 ; count *= 8
add r10, 0fh ; To force alignment to 16bytes
and r10, 0fffffffffffffff0h
sub rsp, r10 ; Actual stack allocation
; Continues by copying all the arguments in the "alloca" space
mov [rsp], rcx ; newStack[0] = (void*)buffer;
mov [rsp + 08h], rdx ; newStack[1] = (void*)format;
; Calls memcpy(newStack+2, argPtr, count*sizeof(void*));
mov rcx, rsp
add rcx, 010h ; newStack+2
mov rdx, r8 ; argPtr
mov r8, r9
sal r8, 3 ; count*sizeof(void*)
; Prepares extra stack space as required by the ABI for 4 arguments, and calls memcpy
sub rsp, 020h
call memcpy
; Restore the stack
add rsp, 020h
; Saves rsp in memory
mov qword ptr [rbp - 8], rsp
; Does exactly the same trick as before: warp into system sscanf with the new stack,
; but this time we also setup the arguments in the registers according to the amd64 ABI
; If there was at least one argument (after buffer and format), we need to copy that
; to r8, and if there was a second one we must copy that to r9
; (the first arguments to sscanf are always the buffer and the format)
mov r10, count
; Copy the first argument to r8 (if it exists)
cmp r10, 0
je args_memcpy
mov r8, [rsp + 10h]
; Copy the second argument to r9 (if it exists)
cmp r10, 1
je args_memcpy
mov r9, [rsp + 18h]
args_memcpy:
; Copies the buffer and format to rcx and rdx
mov rdx, [rsp + 08h]
mov rcx, [rsp]
; Finally, calls sscanf using the current stack
call sscanf
; At this point the return value is alreay in rax
; Restores rsp
mov rsp, qword ptr [rbp - 8]
; Undoes the alloca
add rsp, r10
; Restores the space for local variables
add rsp, 08h
; Remember, the return value is already in rax since the sscanf call
ret
vsscanf_proxy_win64 ENDP
END