PtexSeparableKernel.h

Go to the documentation of this file.

#ifndef PtexSeparableKernel_h
#define PtexSeparableKernel_h
 
/*
PTEX SOFTWARE
Copyright 2014 Disney Enterprises, Inc.  All rights reserved
 
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
 
  * Redistributions of source code must retain the above copyright
    notice, this list of conditions and the following disclaimer.
 
  * Redistributions in binary form must reproduce the above copyright
    notice, this list of conditions and the following disclaimer in
    the documentation and/or other materials provided with the
    distribution.
 
  * The names "Disney", "Walt Disney Pictures", "Walt Disney Animation
    Studios" or the names of its contributors may NOT be used to
    endorse or promote products derived from this software without
    specific prior written permission from Walt Disney Pictures.
 
Disclaimer: THIS SOFTWARE IS PROVIDED BY WALT DISNEY PICTURES AND
CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING,
BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY, FITNESS
FOR A PARTICULAR PURPOSE, NONINFRINGEMENT AND TITLE ARE DISCLAIMED.
IN NO EVENT SHALL WALT DISNEY PICTURES, THE COPYRIGHT HOLDER OR
CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND BASED ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGES.
*/
 
#include <assert.h>
#include <algorithm>
#include <numeric>
#include "Ptexture.h"
#include "PtexUtils.h"
 
PTEX_NAMESPACE_BEGIN
 
// Separable convolution kernel
class PtexSeparableKernel {
 public:
    Res res;                    // resolution that kernel was built for
    int u, v;                   // uv offset within face data
    int uw, vw;                 // kernel width
    float* ku;                  // kernel weights in u
    float* kv;                  // kernel weights in v
    static const int kmax = 10; // max kernel width
    float kubuff[kmax];
    float kvbuff[kmax];
    int rot;
 
    PtexSeparableKernel()
        : res(0), u(0), v(0), uw(0), vw(0), ku(kubuff), kv(kvbuff), rot(0)
    {
        kubuff[0] = 0; // keep cppcheck happy
        kvbuff[0] = 0;
    }
 
    PtexSeparableKernel(const PtexSeparableKernel& k)
    {
        set(k.res, k.u, k.v, k.uw, k.vw, k.ku, k.kv, k.rot);
    }
 
    PtexSeparableKernel& operator= (const PtexSeparableKernel& k)
    {
        set(k.res, k.u, k.v, k.uw, k.vw, k.ku, k.kv, k.rot);
        return *this;
    }
 
    void set(Res resVal,
             int uVal, int vVal,
             int uwVal, int vwVal,
             const float* kuVal, const float* kvVal, int rotVal=0)
    {
        assert(uwVal <= kmax && vwVal <= kmax);
        res = resVal;
        u = uVal;
        v = vVal;
        uw = uwVal;
        vw = vwVal;
        memcpy(kubuff, kuVal, sizeof(*ku)*uw);
        memcpy(kvbuff, kvVal, sizeof(*kv)*vw);
        ku = kubuff;
        kv = kvbuff;
        rot = rotVal;
    }
 
    void stripZeros()
    {
        while (ku[0] == 0) { ku++; u++; uw--; }
        while (ku[uw-1] == 0) { uw--; }
        while (kv[0] == 0) { kv++; v++; vw--; }
        while (kv[vw-1] == 0) { vw--; }
        assert(uw > 0 && vw > 0);
    }
 
    float weight() const
    {
        return accumulate(ku, uw) * accumulate(kv, vw);
    }
 
    void mergeL(BorderMode mode)
    {
        int w = -u;
        if (mode != m_black)
            ku[w] += accumulate(ku, w);
        ku += w;
        uw -= w;
        u = 0;
    }
 
    void mergeR(BorderMode mode)
    {
        int w = uw + u - res.u();
        float* kp = ku + uw - w;
        if (mode != m_black)
            kp[-1] += accumulate(kp, w);
        uw -= w;
    }
 
    void mergeB(BorderMode mode)
    {
        int w = -v;
        if (mode != m_black)
            kv[w] += accumulate(kv, w);
        kv += w;
        vw -= w;
        v = 0;
    }
 
    void mergeT(BorderMode mode)
    {
        int w = vw + v - res.v();
        float* kp = kv + vw - w;
        if (mode != m_black)
            kp[-1] += accumulate(kp, w);
        vw -= w;
    }
 
    void splitL(PtexSeparableKernel& k)
    {
        // split off left piece of width w into k
        int w = -u;
 
        if (w < uw) {
            // normal case - split off a portion
            //    res  u          v  uw vw  ku  kv
            k.set(res, res.u()-w, v, w, vw, ku, kv);
 
            // update local
            u = 0;
            uw -= w;
            ku += w;
        }
        else {
            // entire kernel is split off
            k = *this;
            k.u += res.u();
            u = 0; uw = 0;
        }
    }
 
    void splitR(PtexSeparableKernel& k)
    {
        // split off right piece of width w into k
        int w = u + uw - res.u();
 
        if (w < uw) {
            // normal case - split off a portion
            //    res  u  v  uw vw  ku           kv
            k.set(res, 0, v, w, vw, ku + uw - w, kv);
 
            // update local
            uw -= w;
        }
        else {
            // entire kernel is split off
            k = *this;
            k.u -= res.u();
            u = 0; uw = 0;
        }
    }
 
    void splitB(PtexSeparableKernel& k)
    {
        // split off bottom piece of width w into k
        int w = -v;
        if (w < vw) {
            // normal case - split off a portion
            //    res  u  v          uw vw  ku  kv
            k.set(res, u, res.v()-w, uw, w, ku, kv);
 
            // update local
            v = 0;
            vw -= w;
            kv += w;
        }
        else {
            // entire kernel is split off
            k = *this;
            k.v += res.v();
            v = 0; vw = 0;
        }
    }
 
    void splitT(PtexSeparableKernel& k)
    {
        // split off top piece of width w into k
        int w = v + vw - res.v();
        if (w < vw) {
            // normal case - split off a portion
            //    res  u  v  uw vw  ku  kv
            k.set(res, u, 0, uw, w, ku, kv + vw - w);
 
            // update local
            vw -= w;
        }
        else {
            // entire kernel is split off
            k = *this;
            k.v -= res.v();
            v = 0; vw = 0;
        }
    }
 
    void flipU()
    {
        u = res.u() - u - uw;
        std::reverse(ku, ku+uw);
    }
 
    void flipV()
    {
        v = res.v() - v - vw;
        std::reverse(kv, kv+vw);
    }
 
    void swapUV()
    {
        res.swapuv();
        std::swap(u, v);
        std::swap(uw, vw);
        std::swap(ku, kv);
    }
 
    void rotate(int rotVal)
    {
        // rotate kernel 'rot' steps ccw
        switch (rotVal & 3) {
        default: return;
        case 1: flipU(); swapUV(); break;
        case 2: flipU(); flipV(); break;
        case 3: flipV(); swapUV(); break;
        }
        rot = (rot + rotVal)&3;
    }
 
    bool adjustMainToSubface(int eid)
    {
        // to adjust the kernel for the subface, we must adjust the res down and offset the uv coords
        // however, if the res is already zero, we must upres the kernel first
        if (res.ulog2 == 0) upresU();
        if (res.vlog2 == 0) upresV();
 
        if (res.ulog2 > 0) res.ulog2--;
        if (res.vlog2 > 0) res.vlog2--;
 
        // offset uv coords and determine whether target subface is the primary one
        bool primary = 0;
        int resu = res.u(), resv = res.v();
        switch (eid&3) {
        case e_bottom:
            primary = (u < resu);
            v -= resv;
            if (!primary) u -= resu;
            break;
        case e_right:
            primary = (v < resv);
            if (!primary) v -= resv;
            break;
        case e_top:
            primary = (u >= resu);
            if (primary) u -= resu;
            break;
        case e_left:
            primary = (v >= resv);
            u -= resu;
            if (primary) v -= resv;
            break;
        }
        return primary;
    }
 
    void adjustSubfaceToMain(int eid)
    {
        switch (eid&3) {
        case e_bottom: v += res.v(); break;
        case e_right:  break;
        case e_top:    u += res.u(); break;
        case e_left:   u += res.u(); v += res.v(); break;
        }
        res.ulog2++; res.vlog2++;
    }
 
    void downresU()
    {
        float* src = ku;
        float* dst = ku;
 
        // skip odd leading sample (if any)
        if (u & 1) {
            dst++;
            src++;
            uw--;
        }
 
        // combine even pairs
        for (int i = uw/2; i > 0; i--) {
            *dst++ = src[0] + src[1];
            src += 2;
        }
 
        // copy odd trailing sample (if any)
        if (uw & 1) {
            *dst++ = *src++;
        }
 
        // update state
        u /= 2;
        uw = int(dst - ku);
        res.ulog2--;
    }
 
    void downresV()
    {
        float* src = kv;
        float* dst = kv;
 
        // skip odd leading sample (if any)
        if (v & 1) {
            dst++;
            src++;
            vw--;
        }
 
        // combine even pairs
        for (int i = vw/2; i > 0; i--) {
            *dst++ = src[0] + src[1];
            src += 2;
        }
 
        // copy odd trailing sample (if any)
        if (vw & 1) {
            *dst++ = *src++;
        }
 
        // update state
        v /= 2;
        vw = int(dst - kv);
        res.vlog2--;
    }
 
    void upresU()
    {
        float* src = ku + uw-1;
        float* dst = ku + uw*2-2;
        for (int i = uw; i > 0; i--) {
            dst[0] = dst[1] = *src-- / 2;
            dst -=2;
        }
        uw *= 2;
        u *= 2;
        res.ulog2++;
    }
 
    void upresV()
    {
        float* src = kv + vw-1;
        float* dst = kv + vw*2-2;
        for (int i = vw; i > 0; i--) {
            dst[0] = dst[1] = *src-- / 2;
            dst -=2;
        }
        vw *= 2;
        v *= 2;
        res.vlog2++;
    }
 
    float makeSymmetric(float initialWeight)
    {
        assert(u == 0 && v == 0);
 
        // downres higher-res dimension until equal
        if (res.ulog2 > res.vlog2) {
            do { downresU(); } while(res.ulog2 > res.vlog2);
        }
        else if (res.vlog2 > res.ulog2) {
            do { downresV(); } while (res.vlog2 > res.ulog2);
        }
 
        // truncate excess samples in longer dimension
        uw = vw = PtexUtils::min(uw, vw);
 
        // combine corresponding u and v samples and compute new kernel weight
        float newWeight = 0;
        for (int i = 0; i < uw; i++) {
            float sum = ku[i] + kv[i];
            ku[i] = kv[i] = sum;
            newWeight += sum;
        }
        newWeight *= newWeight; // equivalent to k.weight() ( = sum(ku)*sum(kv) )
 
        // compute scale factor to compensate for weight change
        float scale = newWeight == 0 ? 1.f : initialWeight / newWeight;
 
        // Note: a sharpening kernel (like Mitchell) can produce
        // negative weights which may cancel out when adding the two
        // kernel axes together, and this can cause the compensation
        // scale factor to spike up.  We expect the scale factor to be
        // less than one in "normal" cases (i.e. ku*kv <= (ku+kv)^2 if ku
        // and kv are both positive), so clamping to -1..1 will have
        // no effect on positive kernels.  If there are negative
        // weights, the clamping will just limit the amount of
        // sharpening happening at the corners, and the result will
        // still be smooth.
 
        // clamp scale factor to -1..1 range
        if (scale >= 1) {
            // scale by 1 (i.e. do nothing)
        }
        else {
            if (scale < -1) {
                // a negative scale means the original kernel had an overall negative weight
                // after making symmetric, the kernel will always be positive
                // scale ku by -1
                // note: choice of u is arbitrary; we could have scaled u or v (but not both)
                for (int i = 0; i < uw; i++) ku[i] *= -1;
                newWeight = -newWeight;
            }
            else {
                // scale ku to restore initialWeight (again, choice of u instead of v is arbitrary)
                for (int i = 0; i < uw; i++) ku[i] *= scale;
                newWeight = initialWeight;
            }
        }
        return newWeight;
    }
 
    void apply(float* dst, void* data, DataType dt, int nChan, int nTxChan)
    {
        // dispatch specialized apply function
        ApplyFn fn = applyFunctions[(nChan!=nTxChan)*20 + ((unsigned)nChan<=4)*nChan*4 + dt];
        fn(*this, dst, data, nChan, nTxChan);
    }
 
    void applyConst(float* dst, void* data, DataType dt, int nChan)
    {
        PtexUtils::applyConst(weight(), dst, data, dt, nChan);
    }
 
 private:
    typedef void (*ApplyFn)(PtexSeparableKernel& k, float* dst, void* data, int nChan, int nTxChan);
    typedef void (*ApplyConstFn)(float weight, float* dst, void* data, int nChan);
    static ApplyFn applyFunctions[40];
    static ApplyConstFn applyConstFunctions[20];
    static inline float accumulate(const float* p, int n)
    {
        float result = 0;
        for (const float* e = p + n; p != e; p++) result += *p;
        return result;
    }
};
 
PTEX_NAMESPACE_END
 
#endif