crossxtex/DirectXTex/filters.h

//-------------------------------------------------------------------------------------
// filters.h
//  
// Utility header with helpers for implementing image filters
//
// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
//-------------------------------------------------------------------------------------

#pragma once

#include <DirectXMath.h>
#include <DirectXPackedVector.h>

#include <memory>

#include "scoped.h"

namespace DirectX
{

//-------------------------------------------------------------------------------------
// Box filtering helpers
//-------------------------------------------------------------------------------------

XMGLOBALCONST XMVECTORF32 g_boxScale   = { { { 0.25f, 0.25f, 0.25f, 0.25f } } };
XMGLOBALCONST XMVECTORF32 g_boxScale3D = { { { 0.125f, 0.125f, 0.125f, 0.125f } } };

#define AVERAGE4( res, p0, p1, p2, p3 ) \
{ \
    XMVECTOR v = XMVectorAdd((p0), (p1)); \
    v = XMVectorAdd(v, (p2)); \
    v = XMVectorAdd(v, (p3)); \
    res = XMVectorMultiply(v, g_boxScale); \
}

#define AVERAGE8( res, p0, p1, p2, p3, p4, p5, p6, p7) \
{ \
    XMVECTOR v = XMVectorAdd((p0), (p1)); \
    v = XMVectorAdd(v, (p2)); \
    v = XMVectorAdd(v, (p3)); \
    v = XMVectorAdd(v, (p4)); \
    v = XMVectorAdd(v, (p5)); \
    v = XMVectorAdd(v, (p6)); \
    v = XMVectorAdd(v, (p7)); \
    res = XMVectorMultiply(v, g_boxScale3D); \
}


//-------------------------------------------------------------------------------------
// Linear filtering helpers
//-------------------------------------------------------------------------------------

struct LinearFilter
{
    size_t  u0;
    float   weight0;
    size_t  u1;
    float   weight1;
};

inline void _CreateLinearFilter(_In_ size_t source, _In_ size_t dest, _In_ bool wrap, _Out_writes_(dest) LinearFilter* lf)
{
    assert(source > 0);
    assert(dest > 0);
    assert(lf != nullptr);

    float scale = float(source) / float(dest);

    // Mirror is the same case as clamp for linear

    for (size_t u = 0; u < dest; ++u)
    {
        float srcB = (float(u) + 0.5f) * scale + 0.5f;

        ptrdiff_t isrcB = ptrdiff_t(srcB);
        ptrdiff_t isrcA = isrcB - 1;

        if (isrcA < 0)
        {
            isrcA = (wrap) ? (ptrdiff_t(source) - 1) : 0;
        }

        if (size_t(isrcB) >= source)
        {
            isrcB = (wrap) ? 0 : (ptrdiff_t(source) - 1);
        }

        float weight = 1.0f + float(isrcB) - srcB;

        auto& entry = lf[u];
        entry.u0 = size_t(isrcA);
        entry.weight0 = weight;

        entry.u1 = size_t(isrcB);
        entry.weight1 = 1.0f - weight;
    }
}

#define BILINEAR_INTERPOLATE( res, x, y, r0, r1 ) \
    res = XMVectorAdd(XMVectorScale(XMVectorAdd(XMVectorScale((r0)[ x.u0 ], x.weight0), XMVectorScale((r0)[ x.u1 ], x.weight1)), y.weight0), \
                      XMVectorScale(XMVectorAdd(XMVectorScale((r1)[ x.u0 ], x.weight0), XMVectorScale((r1)[ x.u1 ], x.weight1)), y.weight1) )

#define TRILINEAR_INTERPOLATE( res, x, y, z, r0, r1, r2, r3 ) \
{\
    XMVECTOR a0 = XMVectorScale(XMVectorAdd(XMVectorScale((r0)[ x.u0 ], x.weight0 ), XMVectorScale((r0)[ x.u1 ], x.weight1)), y.weight0); \
    XMVECTOR a1 = XMVectorScale(XMVectorAdd(XMVectorScale((r1)[ x.u0 ], x.weight0 ), XMVectorScale((r1)[ x.u1 ], x.weight1)), y.weight1); \
    XMVECTOR a2 = XMVectorScale(XMVectorAdd(XMVectorScale((r2)[ x.u0 ], x.weight0 ), XMVectorScale((r2)[ x.u1 ], x.weight1)), y.weight0); \
    XMVECTOR a3 = XMVectorScale(XMVectorAdd(XMVectorScale((r3)[ x.u0 ], x.weight0 ), XMVectorScale((r3)[ x.u1 ], x.weight1)), y.weight1); \
    res = XMVectorAdd(XMVectorScale(XMVectorAdd(a0, a1), z.weight0), XMVectorScale(XMVectorAdd(a2, a3), z.weight1)); \
}

//-------------------------------------------------------------------------------------
// Cubic filtering helpers
//-------------------------------------------------------------------------------------

XMGLOBALCONST XMVECTORF32 g_cubicThird = { { { 1.f / 3.f, 1.f / 3.f, 1.f / 3.f, 1.f / 3.f } } };
XMGLOBALCONST XMVECTORF32 g_cubicSixth = { { { 1.f / 6.f, 1.f / 6.f, 1.f / 6.f, 1.f / 6.f } } };
XMGLOBALCONST XMVECTORF32 g_cubicHalf  = { { { 1.f / 2.f, 1.f / 2.f, 1.f / 2.f, 1.f / 2.f } } };

inline ptrdiff_t bounduvw(ptrdiff_t u, ptrdiff_t maxu, bool wrap, bool mirror)
{
    if (wrap)
    {
        if (u < 0)
        {
            u = maxu + u + 1;
        }
        else if (u > maxu)
        {
            u = u - maxu - 1;
        }
    }
    else if (mirror)
    {
        if (u < 0)
        {
            u = (-u) - 1;
        }
        else if (u > maxu)
        {
            u = maxu - (u - maxu - 1);
        }
    }

    // Handles clamp, but also a safety factor for degenerate images for wrap/mirror
    u = std::min<ptrdiff_t>(u, maxu);
    u = std::max<ptrdiff_t>(u, 0);

    return u;
}

struct CubicFilter
{
    size_t  u0;
    size_t  u1;
    size_t  u2;
    size_t  u3;
    float   x;
};

inline void _CreateCubicFilter(_In_ size_t source, _In_ size_t dest, _In_ bool wrap, _In_ bool mirror, _Out_writes_(dest) CubicFilter* cf)
{
    assert(source > 0);
    assert(dest > 0);
    assert(cf != nullptr);

    float scale = float(source) / float(dest);

    for (size_t u = 0; u < dest; ++u)
    {
        float srcB = (float(u) + 0.5f) * scale - 0.5f;

        ptrdiff_t isrcB = bounduvw(ptrdiff_t(srcB), ptrdiff_t(source) - 1, wrap, mirror);
        ptrdiff_t isrcA = bounduvw(isrcB - 1, ptrdiff_t(source) - 1, wrap, mirror);
        ptrdiff_t isrcC = bounduvw(isrcB + 1, ptrdiff_t(source) - 1, wrap, mirror);
        ptrdiff_t isrcD = bounduvw(isrcB + 2, ptrdiff_t(source) - 1, wrap, mirror);

        auto& entry = cf[u];
        entry.u0 = size_t(isrcA);
        entry.u1 = size_t(isrcB);
        entry.u2 = size_t(isrcC);
        entry.u3 = size_t(isrcD);

        float x = srcB - float(isrcB);
        entry.x = x;
    }
}

#define CUBIC_INTERPOLATE( res, dx, p0, p1, p2, p3 ) \
{ \
    XMVECTOR a0 = (p1); \
    XMVECTOR d0 = XMVectorSubtract(p0, a0); \
    XMVECTOR d2 = XMVectorSubtract(p2, a0); \
    XMVECTOR d3 = XMVectorSubtract(p3, a0); \
    XMVECTOR a1 = XMVectorSubtract(d2, XMVectorMultiply(g_cubicThird, d0)); \
    a1 = XMVectorSubtract(a1, XMVectorMultiply(g_cubicSixth, d3)); \
    XMVECTOR a2 = XMVectorAdd(XMVectorMultiply(g_cubicHalf, d0), XMVectorMultiply(g_cubicHalf, d2)); \
    XMVECTOR a3 = XMVectorSubtract(XMVectorMultiply(g_cubicSixth, d3), XMVectorMultiply(g_cubicSixth, d0)); \
    a3 = XMVectorSubtract(a3, XMVectorMultiply(g_cubicHalf, d2)); \
    XMVECTOR vdx = XMVectorReplicate(dx); \
    XMVECTOR vdx2 = XMVectorMultiply(vdx, vdx); \
    XMVECTOR vdx3 = XMVectorMultiply(vdx2, vdx); \
    res = XMVectorAdd(XMVectorAdd(XMVectorAdd(a0, XMVectorMultiply(a1, vdx)), XMVectorMultiply(a2, vdx2)), XMVectorMultiply(a3, vdx3)); \
}


//-------------------------------------------------------------------------------------
// Triangle filtering helpers
//-------------------------------------------------------------------------------------

namespace TriangleFilter
{
    struct FilterTo
    {
        size_t      u;
        float       weight;
    };

    struct FilterFrom
    {
        size_t      count;
        size_t      sizeInBytes;
        FilterTo    to[1]; // variable-sized array
    };

    struct Filter
    {
        size_t      sizeInBytes;
        size_t      totalSize;
        FilterFrom  from[1]; // variable-sized array
    };

    struct TriangleRow
    {
        size_t                      remaining;
        TriangleRow*                next;
        ScopedAlignedArrayXMVECTOR  scanline;

        TriangleRow() noexcept : remaining(0), next(nullptr) {}
    };

    static const size_t TF_FILTER_SIZE = sizeof(Filter) - sizeof(FilterFrom);
    static const size_t TF_FROM_SIZE = sizeof(FilterFrom) - sizeof(FilterTo);
    static const size_t TF_TO_SIZE = sizeof(FilterTo);

    static const float TF_EPSILON = 0.00001f;

    inline HRESULT _Create(_In_ size_t source, _In_ size_t dest, _In_ bool wrap, _Inout_ std::unique_ptr<Filter>& tf)
    {
        assert(source > 0);
        assert(dest > 0);

        float scale = float(dest) / float(source);
        float scaleInv = 0.5f / scale;

        // Determine storage required for filter and allocate memory if needed
        size_t totalSize = TF_FILTER_SIZE + TF_FROM_SIZE + TF_TO_SIZE;
        float repeat = (wrap) ? 1.f : 0.f;

        for (size_t u = 0; u < source; ++u)
        {
            float src = float(u) - 0.5f;
            float destMin = src * scale;
            float destMax = destMin + scale;
            float t = destMax - destMin + repeat + 1.f;
            totalSize += TF_FROM_SIZE + TF_TO_SIZE + size_t(t) * TF_TO_SIZE * 2;
        }

        uint8_t* pFilter = nullptr;

        if (tf)
        {
            // See if existing filter memory block is large enough to reuse
            if (tf->totalSize >= totalSize)
            {
                pFilter = reinterpret_cast<uint8_t*>(tf.get());
            }
            else
            {
                // Need to reallocate filter memory block
                tf.reset(nullptr);
            }
        }

        if (!tf)
        {
            // Allocate filter memory block
            pFilter = new (std::nothrow) uint8_t[totalSize];
            if (!pFilter)
                return E_OUTOFMEMORY;

            tf.reset(reinterpret_cast<Filter*>(pFilter));
            tf->totalSize = totalSize;
        }

        assert(pFilter != nullptr);
        _Analysis_assume_(pFilter != 0);

        // Filter setup
        size_t sizeInBytes = TF_FILTER_SIZE;
        size_t accumU = 0;
        float accumWeight = 0.f;

        for (size_t u = 0; u < source; ++u)
        {
            // Setup from entry
            size_t sizeFrom = sizeInBytes;
            auto pFrom = reinterpret_cast<FilterFrom*>(pFilter + sizeInBytes);
            sizeInBytes += TF_FROM_SIZE;

            if (sizeInBytes > totalSize)
                return E_FAIL;

            size_t toCount = 0;

            // Perform two passes to capture the influences from both sides
            for (size_t j = 0; j < 2; ++j)
            {
                float src = float(u + j) - 0.5f;

                float destMin = src * scale;
                float destMax = destMin + scale;

                if (!wrap)
                {
                    // Clamp
                    if (destMin < 0.f)
                        destMin = 0.f;
                    if (destMax > float(dest))
                        destMax = float(dest);
                }

                for (auto k = static_cast<ptrdiff_t>(floorf(destMin)); float(k) < destMax; ++k)
                {
                    float d0 = float(k);
                    float d1 = d0 + 1.f;

                    size_t u0;
                    if (k < 0)
                    {
                        // Handle wrap
                        u0 = size_t(k + ptrdiff_t(dest));
                    }
                    else if (k >= ptrdiff_t(dest))
                    {
                        // Handle wrap
                        u0 = size_t(k - ptrdiff_t(dest));
                    }
                    else
                    {
                        u0 = size_t(k);
                    }

                    // Save previous accumulated weight (if any)
                    if (u0 != accumU)
                    {
                        if (accumWeight > TF_EPSILON)
                        {
                            auto pTo = reinterpret_cast<FilterTo*>(pFilter + sizeInBytes);
                            sizeInBytes += TF_TO_SIZE;
                            ++toCount;

                            if (sizeInBytes > totalSize)
                                return E_FAIL;

                            pTo->u = accumU;
                            pTo->weight = accumWeight;
                        }

                        accumWeight = 0.f;
                        accumU = u0;
                    }

                    // Clip destination
                    if (d0 < destMin)
                        d0 = destMin;
                    if (d1 > destMax)
                        d1 = destMax;

                    // Calculate average weight over destination pixel

                    float weight;
                    if (!wrap && src < 0.f)
                        weight = 1.f;
                    else if (!wrap && ((src + 1.f) >= float(source)))
                        weight = 0.f;
                    else
                        weight = (d0 + d1) * scaleInv - src;

                    accumWeight += (d1 - d0) * (j ? (1.f - weight) : weight);
                }
            }

            // Store accumulated weight
            if (accumWeight > TF_EPSILON)
            {
                auto pTo = reinterpret_cast<FilterTo*>(pFilter + sizeInBytes);
                sizeInBytes += TF_TO_SIZE;
                ++toCount;

                if (sizeInBytes > totalSize)
                    return E_FAIL;

                pTo->u = accumU;
                pTo->weight = accumWeight;
            }

            accumWeight = 0.f;

            // Finalize from entry
            pFrom->count = toCount;
            pFrom->sizeInBytes = sizeInBytes - sizeFrom;
        }

        tf->sizeInBytes = sizeInBytes;

        return S_OK;
    }

} // namespace TriangleFilter

} // namespace DirectX