Fast O(N) median filter for GDI+ bitmaps impl w/ SSE2 thunks based on http://nomis80.org/ctmf.c

mdMedianBlur.bas

Option Explicit
DefObj A-Z

'--- for VirtualProtect
Private Const PAGE_EXECUTE_READWRITE        As Long = &H40
Private Const MEM_COMMIT                    As Long = &H1000
'--- for CryptStringToBinary
Private Const CRYPT_STRING_BASE64           As Long = 1
'--- for GdipBitmapLockBits
Private Const ImageLockModeRead             As Long = &H1
Private Const ImageLockModeWrite            As Long = &H2
Private Const PixelFormat32bppARGB          As Long = &H26200A
Private Const PixelFormat24bppRGB           As Long = &H21808
Private Const PixelFormatAlpha              As Long = &H40000

Private Declare Sub CopyMemory Lib "kernel32" Alias "RtlMoveMemory" (Destination As Any, Source As Any, ByVal Length As Long)
Private Declare Function VirtualProtect Lib "kernel32" (ByVal lpAddress As Long, ByVal dwSize As Long, ByVal flNewProtect As Long, ByRef lpflOldProtect As Long) As Long
Private Declare Function VirtualAlloc Lib "kernel32" (ByVal lpAddress As Long, ByVal dwSize As Long, ByVal flAllocationType As Long, ByVal flProtect As Long) As Long
Private Declare Function CryptStringToBinary Lib "crypt32" Alias "CryptStringToBinaryW" (ByVal pszString As Long, ByVal cchString As Long, ByVal dwFlags As Long, ByVal pbBinary As Long, ByRef pcbBinary As Long, ByRef pdwSkip As Long, ByRef pdwFlags As Long) As Long
Private Declare Function GetModuleHandle Lib "kernel32" Alias "GetModuleHandleA" (ByVal lpModuleName As String) As Long
Private Declare Function GetProcAddress Lib "kernel32" (ByVal hModule As Long, ByVal lpProcName As String) As Long
'--- gdi+
Private Declare Function GdipGetImagePixelFormat Lib "gdiplus" (ByVal Image As Long, nFormat As Long) As Long
Private Declare Function GdipBitmapLockBits Lib "gdiplus" (ByVal hBitmap As Long, lpRect As Any, ByVal lFlags As Long, ByVal lPixelFormat As Long, uLockedBitmapData As BitmapData) As Long
Private Declare Function GdipBitmapUnlockBits Lib "gdiplus" (ByVal hBitmap As Long, uLockedBitmapData As BitmapData) As Long

Private Type BitmapData
    Width               As Long
    Height              As Long
    Stride              As Long
    PixelFormat         As Long
    Scan0               As Long
    Reserved            As Long
End Type

Public Function MedianBlur( _
            ByVal hBitmap As Long, _
            ByVal lRadius As Long) As Boolean
    Const L2_CACHE_SIZE As Long = 512 * 1024&
    Dim lPixelFormat    As Long
    Dim uData           As BitmapData
    Dim aBuffer()       As Byte
    Dim aImport(0 To 1) As Long

    On Error GoTo EH
    If GdipGetImagePixelFormat(hBitmap, lPixelFormat) <> 0 Then
        GoTo QH
    End If
    If GdipBitmapLockBits(hBitmap, ByVal 0, ImageLockModeRead Or ImageLockModeWrite, IIf((lPixelFormat And PixelFormatAlpha) <> 0, _
            PixelFormat32bppARGB, PixelFormat24bppRGB), uData) <> 0 Then
        GoTo QH
    End If
    ReDim aBuffer(0 To uData.Stride - 1, 0 To uData.Height - 1) As Byte
    aImport(0) = GetProcAddress(GetModuleHandle("ole32"), "CoTaskMemAlloc")
    aImport(1) = GetProcAddress(GetModuleHandle("ole32"), "CoTaskMemFree")
    Call CTimeMedianFilter(uData.Scan0, VarPtr(aBuffer(0, 0)), uData.Width, uData.Height, uData.Stride, uData.Stride, lRadius, _
        IIf((lPixelFormat And PixelFormatAlpha) <> 0, 4, 3), L2_CACHE_SIZE, VarPtr(aImport(0)))
    Call CopyMemory(ByVal uData.Scan0, aBuffer(0, 0), uData.Stride * uData.Height)
    '--- success
    MedianBlur = True
QH:
    On Error Resume Next
    If uData.Scan0 <> 0 Then
        Call GdipBitmapUnlockBits(hBitmap, uData)
    End If
    Exit Function
EH:
    Debug.Print "Critical error: " & Err.Description & " [MedianBlur]"
    Resume QH
End Function

'--- Based on http://nomis80.org/ctmf.c - Constant-time median filtering
'--- Copyright (C) 2006  Simon Perreault, licensed under the GPL v3+
Private Sub CTimeMedianFilter(ByVal lSrcPtr As Long, ByVal lDstPtr As Long, ByVal lWidth As Long, ByVal lHeight As Long, ByVal lSrcStep As Long, ByVal lDstStep As Long, ByVal lRadius As Long, ByVal lNumChannels As Long, ByVal lMaxMemUsage As Long, ByVal lImportPtr As Long)
    Const STR_THUNK     As String = "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" & _
                                    "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" & _
                                    "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" & _
                                    "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" & _
                                    "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" ' 11.4.2019 11:16:59
    pvPatchThunk AddressOf mdMedianBlur.CTimeMedianFilter, STR_THUNK
    CTimeMedianFilter lSrcPtr, lDstPtr, lWidth, lHeight, lSrcStep, lDstStep, lRadius, lNumChannels, lMaxMemUsage, lImportPtr
End Sub

Private Sub pvPatchThunk(ByVal pfn As Long, sThunkStr As String)
    Dim lThunkSize      As Long
    Dim lThunkPtr       As Long
    Dim bInIDE          As Boolean

    '--- decode thunk
    Call CryptStringToBinary(StrPtr(sThunkStr), Len(sThunkStr), CRYPT_STRING_BASE64, 0, lThunkSize, 0, 0)
    lThunkPtr = VirtualAlloc(0, lThunkSize, MEM_COMMIT, PAGE_EXECUTE_READWRITE)
    Call CryptStringToBinary(StrPtr(sThunkStr), Len(sThunkStr), CRYPT_STRING_BASE64, lThunkPtr, lThunkSize, 0, 0)
    '--- patch func
    Debug.Assert pvSetTrue(bInIDE)
    If bInIDE Then
        Call CopyMemory(pfn, ByVal pfn + &H16, 4)
    Else
        Call VirtualProtect(pfn, 8, PAGE_EXECUTE_READWRITE, 0)
    End If
    ' B8 00 00 00 00       mov         eax,00000000h
    ' FF E0                jmp         eax
    Call CopyMemory(ByVal pfn, 6333077358968.8504@, 8)
    Call CopyMemory(ByVal (pfn Xor &H80000000) + 1 Xor &H80000000, lThunkPtr, 4)
End Sub

Private Function pvSetTrue(bValue As Boolean) As Boolean
    bValue = True
    pvSetTrue = True
End Function

thunk.cpp

#include <stdio.h>
#include <windows.h>
#include <intrin.h>

#pragma comment(lib, "crypt32")
extern "C" void * __cdecl memset(void *, int, size_t);
#pragma intrinsic(memset)

#define MY_SSE2

/*
 * ctmf.c - Constant-time median filtering
 * Copyright (C) 2006  Simon Perreault
 *
 * Reference: S. Perreault and P. Hébert, "Median Filtering in Constant Time",
 * IEEE Transactions on Image Processing, September 2007.
 *
 * This program has been obtained from http://nomis80.org/ctmf.html. No patent
 * covers this program, although it is subject to the following license:
 *
 *  This program 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.
 *
 *  This program is distributed in the hope that it will be useful,
 *  but WITHOUT ANY WARRANTY; without even the implied warranty of
 *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *  GNU General Public License for more details.
 *
 *  You should have received a copy of the GNU General Public License
 *  along with this program.  If not, see <http://www.gnu.org/licenses/>.
 *
 * Contact:
 *  Laboratoire de vision et systèmes numériques
 *  Pavillon Adrien-Pouliot
 *  Université Laval
 *  Sainte-Foy, Québec, Canada
 *  G1K 7P4
 *
 *  perreaul@gel.ulaval.ca
 */

/* Standard C includes */
#include <assert.h>
#include <math.h>
#include <stdlib.h>
#include <string.h>

/* Type declarations */
#ifdef _MSC_VER
#include <basetsd.h>
typedef UINT8 uint8_t;
typedef UINT16 uint16_t;
typedef UINT32 uint32_t;
#pragma warning( disable: 4799 )
#else
#include <stdint.h>
#endif

/* Intrinsic declarations */
#ifdef MY_SSE2
#include <emmintrin.h>
#elif defined(__SSE2__) || defined(__MMX__)
#if defined(__SSE2__)
#include <emmintrin.h>
#elif defined(__MMX__)
#include <mmintrin.h>
#endif
#if defined(__GNUC__)
#include <mm_malloc.h>
#elif defined(_MSC_VER)
#include <malloc.h>
#endif
#elif defined(__ALTIVEC__)
#include <altivec.h>
#endif

/* Compiler peculiarities */
#if defined(__GNUC__)
#include <stdint.h>
#define inline __inline__
#define align(x) __attribute__ ((aligned (x)))
#elif defined(_MSC_VER)
#define inline __inline
#define align(x) __declspec(align(x))
#else
#define inline
#define align(x)
#endif

#ifndef MIN
#define MIN(a,b) ((a) > (b) ? (b) : (a))
#endif

#ifndef MAX
#define MAX(a,b) ((a) < (b) ? (b) : (a))
#endif

/**
 * This structure represents a two-tier histogram. The first tier (known as the
 * "coarse" level) is 4 bit wide and the second tier (known as the "fine" level)
 * is 8 bit wide. Pixels inserted in the fine level also get inserted into the
 * coarse bucket designated by the 4 MSBs of the fine bucket value.
 *
 * The structure is aligned on 16 bytes, which is a prerequisite for SIMD
 * instructions. Each bucket is 16 bit wide, which means that extra care must be
 * taken to prevent overflow.
 */
typedef struct align(16)
{
    uint16_t coarse[16];
    uint16_t fine[16][16];
} Histogram;

typedef struct 
{
    void *(__stdcall *malloc)(size_t);
    void (__stdcall *free)(void *);
} import_t;

static void ctmf_helper(
        const unsigned char* const src, unsigned char* const dst,
        const int width, const int height,
        const int src_step, const int dst_step,
        const int r, const int cn,
        const int pad_left, const int pad_right,
        const import_t *
        );

/**
 * \brief Constant-time median filtering
 *
 * This function does a median filtering of an 8-bit image. The source image is
 * processed as if it was padded with zeros. The median kernel is square with
 * odd dimensions. Images of arbitrary size may be processed.
 *
 * To process multi-channel images, you must call this function multiple times,
 * changing the source and destination adresses and steps such that each channel
 * is processed as an independent single-channel image.
 *
 * Processing images of arbitrary bit depth is not supported.
 *
 * The computing time is O(1) per pixel, independent of the radius of the
 * filter. The algorithm's initialization is O(r*width), but it is negligible.
 * Memory usage is simple: it will be as big as the cache size, or smaller if
 * the image is small. For efficiency, the histograms' bins are 16-bit wide.
 * This may become too small and lead to overflow as \a r increases.
 *
 * \param src           Source image data.
 * \param dst           Destination image data. Must be preallocated.
 * \param width         Image width, in pixels.
 * \param height        Image height, in pixels.
 * \param src_step      Distance between adjacent pixels on the same column in
 *                      the source image, in bytes.
 * \param dst_step      Distance between adjacent pixels on the same column in
 *                      the destination image, in bytes.
 * \param r             Median filter radius. The kernel will be a 2*r+1 by
 *                      2*r+1 square.
 * \param cn            Number of channels. For example, a grayscale image would
 *                      have cn=1 while an RGB image would have cn=3.
 * \param memsize       Maximum amount of memory to use, in bytes. Set this to
 *                      the size of the L2 cache, then vary it slightly and
 *                      measure the processing time to find the optimal value.
 *                      For example, a 512 kB L2 cache would have
 *                      memsize=512*1024 initially.
 */
void __stdcall ctmf(
        const unsigned char* const src, unsigned char* const dst,
        const int width, const int height,
        const int src_step, const int dst_step,
        const int r, const int cn, const long unsigned int memsize, 
        const import_t *pfn
        )
{
    /*
     * Processing the image in vertical stripes is an optimization made
     * necessary by the limited size of the CPU cache. Each histogram is 544
     * bytes big and therefore I can fit a limited number of them in the cache.
     * That number may sometimes be smaller than the image width, which would be
     * the number of histograms I would need without stripes.
     *
     * I need to keep histograms in the cache so that they are available
     * quickly when processing a new row. Each row needs access to the previous
     * row's histograms. If there are too many histograms to fit in the cache,
     * thrashing to RAM happens.
     *
     * To solve this problem, I figure out the maximum number of histograms
     * that can fit in cache. From this is determined the number of stripes in
     * an image. The formulas below make the stripes all the same size and use
     * as few stripes as possible.
     *
     * Note that each stripe causes an overlap on the neighboring stripes, as
     * when mowing the lawn. That overlap is proportional to r. When the overlap
     * is a significant size in comparison with the stripe size, then we are not
     * O(1) anymore, but O(r). In fact, we have been O(r) all along, but the
     * initialization term was neglected, as it has been (and rightly so) in B.
     * Weiss, "Fast Median and Bilateral Filtering", SIGGRAPH, 2006. Processing
     * by stripes only makes that initialization term bigger.
     *
     * Also, note that the leftmost and rightmost stripes don't need overlap.
     * A flag is passed to ctmf_helper() so that it treats these cases as if the
     * image was zero-padded.
     */
    int denom = (memsize / sizeof(Histogram) - 2*r);
    int stripes = (width - 2*r + denom-1) / denom;
    int stripe_size = ( width + stripes*2*r - 2*r + stripes-1) / stripes;

    int i;

    for ( i = 0; i < width; i += stripe_size - 2*r ) {
        int stripe = stripe_size;
        /* Make sure that the filter kernel fits into one stripe. */
        if ( i + stripe_size - 2*r >= width || width - (i + stripe_size - 2*r) < 2*r+1 ) {
            stripe = width - i;
        }

        ctmf_helper( src + cn*i, dst + cn*i, stripe, height, src_step, dst_step, r, cn,
                i == 0, stripe == width - i, pfn );

        if ( stripe == width - i ) {
            break;
        }
    }
}

/**
 * HOP is short for Histogram OPeration. This macro makes an operation \a op on
 * histogram \a h for pixel value \a x. It takes care of handling both levels.
 */
#define HOP(h,x,op) \
    h.coarse[x>>4] op; \
    *((uint16_t*) h.fine + x) op;

#define COP(c,j,x,op) \
    h_coarse[ 16*(n*c+j) + (x>>4) ] op; \
    h_fine[ 16 * (n*(16*c+(x>>4)) + j) + (x & 0xF) ] op;

/**
 * Adds histograms \a x and \a y and stores the result in \a y. Makes use of
 * SSE2, MMX or Altivec, if available.
 */
#ifdef MY_SSE2
static inline void histogram_add( const uint16_t x[16], uint16_t y[16] )
{
    *(__m128i*) &y[0] = _mm_add_epi16( *(__m128i*) &y[0], *(__m128i*) &x[0] );
    *(__m128i*) &y[8] = _mm_add_epi16( *(__m128i*) &y[8], *(__m128i*) &x[8] );
}
#elif defined(__SSE2__)
static inline void histogram_add( const uint16_t x[16], uint16_t y[16] )
{
    *(__m128i*) &y[0] = _mm_add_epi16( *(__m128i*) &y[0], *(__m128i*) &x[0] );
    *(__m128i*) &y[8] = _mm_add_epi16( *(__m128i*) &y[8], *(__m128i*) &x[8] );
}
#elif defined(__MMX__)
static inline void histogram_add( const uint16_t x[16], uint16_t y[16] )
{
    *(__m64*) &y[0]  = _mm_add_pi16( *(__m64*) &y[0],  *(__m64*) &x[0]  );
    *(__m64*) &y[4]  = _mm_add_pi16( *(__m64*) &y[4],  *(__m64*) &x[4]  );
    *(__m64*) &y[8]  = _mm_add_pi16( *(__m64*) &y[8],  *(__m64*) &x[8]  );
    *(__m64*) &y[12] = _mm_add_pi16( *(__m64*) &y[12], *(__m64*) &x[12] );
}
#elif defined(__ALTIVEC__)
static inline void histogram_add( const uint16_t x[16], uint16_t y[16] )
{
    *(vector unsigned short*) &y[0] = vec_add( *(vector unsigned short*) &y[0], *(vector unsigned short*) &x[0] );
    *(vector unsigned short*) &y[8] = vec_add( *(vector unsigned short*) &y[8], *(vector unsigned short*) &x[8] );
}
#else
static inline void histogram_add( const uint16_t x[16], uint16_t y[16] )
{
    int i;
    for ( i = 0; i < 16; ++i ) {
        y[i] += x[i];
    }
}
#endif

/**
 * Subtracts histogram \a x from \a y and stores the result in \a y. Makes use
 * of SSE2, MMX or Altivec, if available.
 */
#ifdef MY_SSE2
static inline void histogram_sub( const uint16_t x[16], uint16_t y[16] )
{
    *(__m128i*) &y[0] = _mm_sub_epi16( *(__m128i*) &y[0], *(__m128i*) &x[0] );
    *(__m128i*) &y[8] = _mm_sub_epi16( *(__m128i*) &y[8], *(__m128i*) &x[8] );
}
#elif defined(__SSE2__)
static inline void histogram_sub( const uint16_t x[16], uint16_t y[16] )
{
    *(__m128i*) &y[0] = _mm_sub_epi16( *(__m128i*) &y[0], *(__m128i*) &x[0] );
    *(__m128i*) &y[8] = _mm_sub_epi16( *(__m128i*) &y[8], *(__m128i*) &x[8] );
}
#elif defined(__MMX__)
static inline void histogram_sub( const uint16_t x[16], uint16_t y[16] )
{
    *(__m64*) &y[0]  = _mm_sub_pi16( *(__m64*) &y[0],  *(__m64*) &x[0]  );
    *(__m64*) &y[4]  = _mm_sub_pi16( *(__m64*) &y[4],  *(__m64*) &x[4]  );
    *(__m64*) &y[8]  = _mm_sub_pi16( *(__m64*) &y[8],  *(__m64*) &x[8]  );
    *(__m64*) &y[12] = _mm_sub_pi16( *(__m64*) &y[12], *(__m64*) &x[12] );
}
#elif defined(__ALTIVEC__)
static inline void histogram_sub( const uint16_t x[16], uint16_t y[16] )
{
    *(vector unsigned short*) &y[0] = vec_sub( *(vector unsigned short*) &y[0], *(vector unsigned short*) &x[0] );
    *(vector unsigned short*) &y[8] = vec_sub( *(vector unsigned short*) &y[8], *(vector unsigned short*) &x[8] );
}
#else
static inline void histogram_sub( const uint16_t x[16], uint16_t y[16] )
{
    int i;
    for ( i = 0; i < 16; ++i ) {
        y[i] -= x[i];
    }
}
#endif

static inline void histogram_muladd( const uint16_t a, const uint16_t x[16],
        uint16_t y[16] )
{
    int i;
    for ( i = 0; i < 16; ++i ) {
        y[i] += a * x[i];
    }
}

#pragma function(memset)
static inline void * __cdecl memset(void *pTarget, int value, size_t cbTarget)
{
    __stosb((unsigned char *)pTarget, value, cbTarget);
    return pTarget;
}

static void ctmf_helper(
        const unsigned char* const src, unsigned char* const dst,
        const int width, const int height,
        const int src_step, const int dst_step,
        const int r, const int cn,
        const int pad_left, const int pad_right, 
        const import_t *pfn
        )
{
    const int m = height, n = width;
    int i, j, k, c;
    const unsigned char *p, *q;

    Histogram H[4];
    uint16_t *h_coarse, *h_fine, luc[4][16];
    void *m_coarse, *m_fine;

    assert( src );
    assert( dst );
    assert( r >= 0 );
    assert( width >= 2*r+1 );
    assert( height >= 2*r+1 );
    assert( src_step != 0 );
    assert( dst_step != 0 );

    /* SSE2 and MMX need aligned memory, provided by _mm_malloc(). */
#if defined(__SSE2__) || defined(__MMX__)
    h_coarse = (uint16_t*) _mm_malloc(  1 * 16 * n * cn * sizeof(uint16_t), 16 );
    h_fine   = (uint16_t*) _mm_malloc( 16 * 16 * n * cn * sizeof(uint16_t), 16 );
#else
    m_coarse = pfn->malloc(  1 * 16 * n * cn * sizeof(uint16_t) + 16);
    m_fine   = pfn->malloc( 16 * 16 * n * cn * sizeof(uint16_t) + 16);
    h_coarse = (uint16_t*)(((intptr_t)m_coarse + 15) & -16);
    h_fine   = (uint16_t*)(((intptr_t)m_fine + 15) & -16);
#endif
    memset( h_coarse, 0,  1 * 16 * n * cn * sizeof(uint16_t) );
    memset( h_fine,   0, 16 * 16 * n * cn * sizeof(uint16_t) );

    /* First row initialization */
    for ( j = 0; j < n; ++j ) {
        for ( c = 0; c < cn; ++c ) {
            COP( c, j, src[cn*j+c], += r+1 );
        }
    }
    for ( i = 0; i < r; ++i ) {
        for ( j = 0; j < n; ++j ) {
            for ( c = 0; c < cn; ++c ) {
                COP( c, j, src[src_step*i+cn*j+c], ++ );
            }
        }
    }

    for ( i = 0; i < m; ++i ) {

        /* Update column histograms for entire row. */
        p = src + src_step * MAX( 0, i-r-1 );
        q = p + cn * n;
        for ( j = 0; p != q; ++j ) {
            for ( c = 0; c < cn; ++c, ++p ) {
                COP( c, j, *p, -- );
            }
        }

        p = src + src_step * MIN( m-1, i+r );
        q = p + cn * n;
        for ( j = 0; p != q; ++j ) {
            for ( c = 0; c < cn; ++c, ++p ) {
                COP( c, j, *p, ++ );
            }
        }

        /* First column initialization */
        memset( H, 0, cn*sizeof(H[0]) );
        memset( luc, 0, cn*sizeof(luc[0]) );
        if ( pad_left ) {
            for ( c = 0; c < cn; ++c ) {
                histogram_muladd( r, &h_coarse[16*n*c], H[c].coarse );
            }
        }
        for ( j = 0; j < (pad_left ? r : 2*r); ++j ) {
            for ( c = 0; c < cn; ++c ) {
                histogram_add( &h_coarse[16*(n*c+j)], H[c].coarse );
            }
        }
        for ( c = 0; c < cn; ++c ) {
            for ( k = 0; k < 16; ++k ) {
                histogram_muladd( 2*r+1, &h_fine[16*n*(16*c+k)], &H[c].fine[k][0] );
            }
        }

        for ( j = pad_left ? 0 : r; j < (pad_right ? n : n-r); ++j ) {
            for ( c = 0; c < cn; ++c ) {
                const uint16_t t = 2*r*r + 2*r;
                uint16_t sum = 0, *segment;
                int b;

                histogram_add( &h_coarse[16*(n*c + MIN(j+r,n-1))], H[c].coarse );

                /* Find median at coarse level */
                for ( k = 0; k < 16 ; ++k ) {
                    sum += H[c].coarse[k];
                    if ( sum > t ) {
                        sum -= H[c].coarse[k];
                        break;
                    }
                }
                assert( k < 16 );

                /* Update corresponding histogram segment */
                if ( luc[c][k] <= j-r ) {
                    memset( &H[c].fine[k], 0, 16 * sizeof(uint16_t) );
                    for ( luc[c][k] = j-r; luc[c][k] < MIN(j+r+1,n); ++luc[c][k] ) {
                        histogram_add( &h_fine[16*(n*(16*c+k)+luc[c][k])], H[c].fine[k] );
                    }
                    if ( luc[c][k] < j+r+1 ) {
                        histogram_muladd( j+r+1 - n, &h_fine[16*(n*(16*c+k)+(n-1))], &H[c].fine[k][0] );
                        luc[c][k] = j+r+1;
                    }
                }
                else {
                    for ( ; luc[c][k] < j+r+1; ++luc[c][k] ) {
                        histogram_sub( &h_fine[16*(n*(16*c+k)+MAX(luc[c][k]-2*r-1,0))], H[c].fine[k] );
                        histogram_add( &h_fine[16*(n*(16*c+k)+MIN(luc[c][k],n-1))], H[c].fine[k] );
                    }
                }

                histogram_sub( &h_coarse[16*(n*c+MAX(j-r,0))], H[c].coarse );

                /* Find median in segment */
                segment = H[c].fine[k];
                for ( b = 0; b < 16 ; ++b ) {
                    sum += segment[b];
                    if ( sum > t ) {
                        dst[dst_step*i+cn*j+c] = 16*k + b;
                        break;
                    }
                }
                assert( b < 16 );
            }
        }
    }

#if defined(__SSE2__) || defined(__MMX__)
    _mm_empty();
    _mm_free(h_coarse);
    _mm_free(h_fine);
#else
    pfn->free(m_coarse);
    pfn->free(m_fine);
#endif
}

#define THUNK_SIZE ((char *)main - (char *)ctmf)

BSTR GetCurrentDateTime();

void main()
{
    //ctmf(0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
    WCHAR szBuffer[10000];
    DWORD dwBufSize = _countof(szBuffer);
    CryptBinaryToString((BYTE *)ctmf, THUNK_SIZE, CRYPT_STRING_BASE64, szBuffer, &dwBufSize);
    for (int i = 0, j = 0; (szBuffer[j] = szBuffer[i]) != 0; ++i) {
        j += (szBuffer[j] != '\r' && szBuffer[j] != '\n');
        if (j % 768 == 0) wcscpy(szBuffer + j, L"\" & _\r\n\""), j += 8;
    }
    printf("Const STR_THUNK     As String = \"%S\" ' %S\n", szBuffer, GetCurrentDateTime());
}

BSTR GetCurrentDateTime()
{
    SYSTEMTIME  st;
    DATE        dt;
    VARIANT     vdt = { VT_DATE, };
    VARIANT     vstr = { VT_EMPTY };

    GetLocalTime(&st);
    SystemTimeToVariantTime(&st, &dt);
    vdt.date = dt;
    VariantChangeType(&vstr, &vdt, 0, VT_BSTR);
    return vstr.bstrVal;
}