FasTC/BPTCEncoder/src/RGBAEndpointsSIMD.cpp

/* FasTC
 * Copyright (c) 2012 University of North Carolina at Chapel Hill. All rights reserved.
 *
 * Permission to use, copy, modify, and distribute this software and its documentation for educational, 
 * research, and non-profit purposes, without fee, and without a written agreement is hereby granted, 
 * provided that the above copyright notice, this paragraph, and the following four paragraphs appear 
 * in all copies.
 *
 * Permission to incorporate this software into commercial products may be obtained by contacting the 
 * authors or the Office of Technology Development at the University of North Carolina at Chapel Hill <otd@unc.edu>.
 *
 * This software program and documentation are copyrighted by the University of North Carolina at Chapel Hill. 
 * The software program and documentation are supplied "as is," without any accompanying services from the 
 * University of North Carolina at Chapel Hill or the authors. The University of North Carolina at Chapel Hill 
 * and the authors do not warrant that the operation of the program will be uninterrupted or error-free. The 
 * end-user understands that the program was developed for research purposes and is advised not to rely 
 * exclusively on the program for any reason.
 *
 * IN NO EVENT SHALL THE UNIVERSITY OF NORTH CAROLINA AT CHAPEL HILL OR THE AUTHORS BE LIABLE TO ANY PARTY FOR 
 * DIRECT, INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES, INCLUDING LOST PROFITS, ARISING OUT OF THE 
 * USE OF THIS SOFTWARE AND ITS DOCUMENTATION, EVEN IF THE UNIVERSITY OF NORTH CAROLINA AT CHAPEL HILL OR THE 
 * AUTHORS HAVE BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 * THE UNIVERSITY OF NORTH CAROLINA AT CHAPEL HILL AND THE AUTHORS SPECIFICALLY DISCLAIM ANY WARRANTIES, INCLUDING, 
 * BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE AND ANY 
 * STATUTORY WARRANTY OF NON-INFRINGEMENT. THE SOFTWARE PROVIDED HEREUNDER IS ON AN "AS IS" BASIS, AND THE UNIVERSITY 
 * OF NORTH CAROLINA AT CHAPEL HILL AND THE AUTHORS HAVE NO OBLIGATIONS TO PROVIDE MAINTENANCE, SUPPORT, UPDATES, 
 * ENHANCEMENTS, OR MODIFICATIONS.
 *
 * Please send all BUG REPORTS to <pavel@cs.unc.edu>.
 *
 * The authors may be contacted via:
 *
 * Pavel Krajcevski
 * Dept of Computer Science
 * 201 S Columbia St
 * Frederick P. Brooks, Jr. Computer Science Bldg
 * Chapel Hill, NC 27599-3175
 * USA
 * 
 * <http://gamma.cs.unc.edu/FasTC/>
 */

// The original lisence from the code available at the following location:
// http://software.intel.com/en-us/vcsource/samples/fast-texture-compression
//
// This code has been modified significantly from the original.

//--------------------------------------------------------------------------------------
// Copyright 2011 Intel Corporation
// All Rights Reserved
//
// Permission is granted to use, copy, distribute and prepare derivative works of this
// software for any purpose and without fee, provided, that the above copyright notice
// and this statement appear in all copies.  Intel makes no representations about the
// suitability of this software for any purpose.  THIS SOFTWARE IS PROVIDED "AS IS."
// INTEL SPECIFICALLY DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, AND ALL LIABILITY,
// INCLUDING CONSEQUENTIAL AND OTHER INDIRECT DAMAGES, FOR THE USE OF THIS SOFTWARE,
// INCLUDING LIABILITY FOR INFRINGEMENT OF ANY PROPRIETARY RIGHTS, AND INCLUDING THE
// WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.  Intel does not
// assume any responsibility for any errors which may appear in this software nor any
// responsibility to update it.
//
//--------------------------------------------------------------------------------------

#include "BC7Config.h"
#include "RGBAEndpointsSIMD.h"
#include "BC7Compressor.h"
#include "BC7CompressionModeSIMD.h"

#include <cassert>
#include <cfloat>

#ifndef HAS_SSE_POPCNT 
static inline uint32 popcnt32(uint32 x) {
  uint32 m1 = 0x55555555;
  uint32 m2 = 0x33333333;
  uint32 m3 = 0x0f0f0f0f;
  x -= (x>>1) & 1;
  x = (x&m2) + ((x>>2)&m2);
  x = (x+(x>>4))&m3;
  x += x>>8;
  return (x+(x>>16)) & 0x3f;
}
#endif

///////////////////////////////////////////////////////////////////////////////
//
// RGBAVectorSIMD implementation
//
///////////////////////////////////////////////////////////////////////////////

/* Original scalar implementation:

  // If the mask is all the bits, then we can just return the value.
  if(mask == 0xFF) {
    return val;
  }

  uint32 prec = CountBitsInMask(mask);
  const uint32 step = 1 << (8 - prec);

  assert(step-1 == uint8(~mask));

  uint32 lval = val & mask;
  uint32 hval = lval + step;

  if(pBit >= 0) {
    prec++;
    lval |= !!(pBit) << (8 - prec);
    hval |= !!(pBit) << (8 - prec);
  }

  if(lval > val) {
    lval -= step;
    hval -= step;
  }

  lval |= lval >> prec;
  hval |= hval >> prec;

  if(sad(val, lval) < sad(val, hval))
    return lval;
  else
    return hval;
*/

// !TODO! AVX2 supports an instruction known as vsllv, which shifts a vector
// by the values stored in another vector. I.e. you can do something like this:
//
// __m128i shiftVals = _mm_set_epi32(1, 2, 3, 4);
// __m128i someVector = _mm_set1_epi32(1) ;
// __m128i shifted = _mm_srav_epi32 (someVector, shiftVals);
//
// and the result will be the same as __mm_Set_epi32(1, 4, 8, 16);
//
// This is useful because our color channels may have different precisions
// when we're quantizing them, such as for BC7 modes 4 and 5. Hence, we would
// want to do our quantization as accurately as possible, but currently it would
// be very hard to vectorize.

#ifdef _MSC_VER
#define ALIGN_SSE __declspec ( align(16) )
#else
#define ALIGN_SSE __attribute__((aligned(16)))
#endif

// Constants. There are two ways to specify them: either by using the _mm_set* 
// intrinsics, or by defining them as aligned arrays. You want to do the former 
// when you use them infrequently, and the latter when you use them multiple times
// in a short time frame (like in an inner loop)
static const __m128 kZero = _mm_set1_ps(0.0f);
static const __m128 kByteMax = _mm_set1_ps(255.0f);
static const __m128 kHalfVector = _mm_set1_ps(0.5f);
static const __m128i kOneVector = _mm_set1_epi32(1);
static const __m128i kZeroVector = _mm_set1_epi32(0);
static const ALIGN_SSE uint32 kThirtyTwoVector[4] = { 32, 32, 32, 32 };
static const __m128i kByteValMask = _mm_set_epi32(0xFF, 0xFF, 0xFF, 0xFF);

static inline __m128i sad(const __m128i &a, const __m128i &b) {
  const __m128i maxab = _mm_max_epu8(a, b);
  const __m128i minab = _mm_min_epu8(a, b);
  return _mm_and_si128( kByteValMask, _mm_subs_epu8( maxab, minab ) );
}

__m128i RGBAVectorSIMD::ToPixel(const __m128i &qmask) const {

  // !SPEED! We should figure out a way to get rid of these scalar operations.
#ifdef HAS_SSE_POPCNT
  const uint32 prec = _mm_popcnt_u32(((uint32 *)(&qmask))[0]);
#else
  const uint32 prec = popcnt32(((uint32 *)(&qmask))[0]);
#endif
  
  assert(r >= 0.0f && r <= 255.0f);
  assert(g >= 0.0f && g <= 255.0f);
  assert(b >= 0.0f && b <= 255.0f);
  assert(a >= 0.0f && a <= 255.0f);
  assert(((uint32 *)(&qmask))[3] == 0xFF || ((uint32 *)(&qmask))[3] == ((uint32 *)(&qmask))[0]);
  assert(((uint32 *)(&qmask))[2] == ((uint32 *)(&qmask))[1] && ((uint32 *)(&qmask))[0] == ((uint32 *)(&qmask))[1]);

  const __m128i val = _mm_cvtps_epi32( _mm_add_ps(kHalfVector, vec) );

  const __m128i step = _mm_slli_epi32( kOneVector, 8 - prec );
  const __m128i &mask = qmask;

  __m128i lval = _mm_and_si128(val, mask);
  __m128i hval = _mm_add_epi32(lval, step);

  const __m128i lvalShift = _mm_srli_epi32(lval, prec);
  const __m128i hvalShift = _mm_srli_epi32(hval, prec);

  lval = _mm_or_si128(lval, lvalShift);
  hval = _mm_or_si128(hval, hvalShift);

  const __m128i lvald = _mm_sub_epi32( val, lval );
  const __m128i hvald = _mm_sub_epi32( hval, val );

  const __m128i vd = _mm_cmplt_epi32(lvald, hvald);
  __m128i ans = _mm_blendv_epi8(hval, lval, vd);

  const __m128i chanExact = _mm_cmpeq_epi32(mask, kByteValMask);
  ans = _mm_blendv_epi8( ans, val, chanExact );
  return ans;
}

__m128i RGBAVectorSIMD::ToPixel(const __m128i &qmask, const int pBit) const {
  
  // !SPEED! We should figure out a way to get rid of these scalar operations.
#ifdef HAS_SSE_POPCNT
  const uint32 prec = _mm_popcnt_u32(((uint32 *)(&qmask))[0]);
#else
  const uint32 prec = popcnt32(((uint32 *)(&qmask))[0]);
#endif
  
  assert(r >= 0.0f && r <= 255.0f);
  assert(g >= 0.0f && g <= 255.0f);
  assert(b >= 0.0f && b <= 255.0f);
  assert(a >= 0.0f && a <= 255.0f);
  assert(((uint32 *)(&qmask))[3] == 0xFF || ((uint32 *)(&qmask))[3] == ((uint32 *)(&qmask))[0]);
  assert(((uint32 *)(&qmask))[2] == ((uint32 *)(&qmask))[1] && ((uint32 *)(&qmask))[0] == ((uint32 *)(&qmask))[1]);

  const __m128i val = _mm_cvtps_epi32( _mm_add_ps(kHalfVector, vec) );
  const __m128i pbit = _mm_set1_epi32(!!pBit);

  const __m128i &mask = qmask; // _mm_set_epi32(alphaMask, channelMask, channelMask, channelMask);
  const __m128i step = _mm_slli_epi32( kOneVector, 8 - prec );

  __m128i lval = _mm_and_si128( val, mask );
  __m128i hval = _mm_add_epi32( lval, step );

  const __m128i pBitShifted = _mm_slli_epi32(pbit, 7 - prec);
  lval = _mm_or_si128(lval, pBitShifted );
  hval = _mm_or_si128(hval, pBitShifted);

  // These next three lines we make sure that after adding the pbit that val is
  // still in between lval and hval. If it isn't, then we subtract a
  // step from both. Now, val should be larger than lval and less than
  // hval, but certain situations make this not always the case (e.g. val
  // is 0, precision is 4 bits, and pbit is 1). Hence, we add back the
  // step if it goes below zero, making it equivalent to hval and so it
  // doesn't matter which we choose.
  {
    __m128i cmp = _mm_cmpgt_epi32(lval, val);
    cmp = _mm_mullo_epi32(cmp, step);
    lval = _mm_add_epi32(lval, cmp);
    hval = _mm_add_epi32(hval, cmp);

    cmp = _mm_cmplt_epi32(lval, kZeroVector);
    cmp = _mm_mullo_epi32(cmp, step);
    lval = _mm_sub_epi32(lval, cmp);
  }

  const __m128i lvalShift = _mm_srli_epi32(lval, prec + 1);
  const __m128i hvalShift = _mm_srli_epi32(hval, prec + 1);

  lval = _mm_or_si128(lval, lvalShift);
  hval = _mm_or_si128(hval, hvalShift);

  const __m128i lvald = _mm_sub_epi32( val, lval );
  const __m128i hvald = _mm_sub_epi32( hval, val );

  const __m128i vd = _mm_cmplt_epi32(lvald, hvald);
  __m128i ans = _mm_blendv_epi8(hval, lval, vd);

  const __m128i chanExact = _mm_cmpeq_epi32(mask, kByteValMask);
  ans = _mm_blendv_epi8( ans, val, chanExact );
  return ans;
}

///////////////////////////////////////////////////////////////////////////////
//
// RGBAMatrixSIMD implementation
//
///////////////////////////////////////////////////////////////////////////////

RGBAVectorSIMD RGBAMatrixSIMD::operator *(const RGBAVectorSIMD &p) const {
  
  __m128 xVec = _mm_set1_ps( p.x );
  __m128 yVec = _mm_set1_ps( p.y );
  __m128 zVec = _mm_set1_ps( p.z );
  __m128 wVec = _mm_set1_ps( p.w );

  __m128 vec1 = _mm_mul_ps( xVec, col[0] );
  __m128 vec2 = _mm_mul_ps( yVec, col[1] );
  __m128 vec3 = _mm_mul_ps( zVec, col[2] );
  __m128 vec4 = _mm_mul_ps( wVec, col[3] );

  return RGBAVectorSIMD( _mm_add_ps( _mm_add_ps( vec1, vec2 ), _mm_add_ps( vec3, vec4 ) ) );
}

///////////////////////////////////////////////////////////////////////////////
//
// Cluster implementation
//
///////////////////////////////////////////////////////////////////////////////

RGBAClusterSIMD::RGBAClusterSIMD(const RGBAClusterSIMD &left, const RGBAClusterSIMD &right) {

  assert(!(left.m_PointBitString & right.m_PointBitString));

  *this = left;
  for(int i = 0; i < right.m_NumPoints; i++) {

    const RGBAVectorSIMD &p = right.m_DataPoints[i];

    assert(m_NumPoints < kMaxNumDataPoints);
    m_Total += p;
    m_DataPoints[m_NumPoints++] = p;

    m_Min.vec = _mm_min_ps(m_Min.vec, p.vec);
    m_Max.vec = _mm_max_ps(m_Max.vec, p.vec);
  }

  m_PointBitString = left.m_PointBitString | right.m_PointBitString;
  m_PrincipalAxisCached = false;
}  

void RGBAClusterSIMD::AddPoint(const RGBAVectorSIMD &p, int idx) {
  assert(m_NumPoints < kMaxNumDataPoints);
  m_Total += p;
  m_DataPoints[m_NumPoints++] = p;
  m_PointBitString |= 1 << idx;

  m_Min.vec = _mm_min_ps(m_Min.vec, p.vec);
  m_Max.vec = _mm_max_ps(m_Max.vec, p.vec);
}

float RGBAClusterSIMD::QuantizedError(const RGBAVectorSIMD &p1, const RGBAVectorSIMD &p2, const uint8 nBuckets, const __m128i &bitMask, const int pbits[2], __m128i *indices) const {

  // nBuckets should be a power of two.
  assert(!(nBuckets & (nBuckets - 1)));

#ifdef HAS_SSE_POPCNT
  const uint8 indexPrec = 8-_mm_popcnt_u32(~(nBuckets - 1) & 0xFF);
#else
  const uint8 indexPrec = 8-popcnt32(~(nBuckets - 1) & 0xFF);
#endif
  assert(indexPrec >= 2 && indexPrec <= 4);

  typedef __m128i tInterpPair[2];
  typedef tInterpPair tInterpLevel[16];
  const tInterpLevel *interpVals = kBC7InterpolationValuesSIMD + (indexPrec - 1);

  __m128i qp1, qp2;
  if(pbits) {
    qp1 = p1.ToPixel(bitMask, pbits[0]);
    qp2 = p2.ToPixel(bitMask, pbits[1]);
  }
  else {
    qp1 = p1.ToPixel(bitMask);
    qp2 = p2.ToPixel(bitMask);
  }

  __m128 errorMetricVec = _mm_load_ps( BC7C::GetErrorMetric() );

  __m128 totalError = kZero;
  for(int i = 0; i < m_NumPoints; i++) {

    const __m128i pixel = m_DataPoints[i].ToPixel( kByteValMask );

    __m128 minError = _mm_set1_ps(FLT_MAX);
    __m128i bestBucket = _mm_set1_epi32(-1);
    for(int j = 0; j < nBuckets; j++) {

      const __m128i jVec = _mm_set1_epi32(j);
      const __m128i interp0 = (*interpVals)[j][0];
      const __m128i interp1 = (*interpVals)[j][1];

      const __m128i ip0 = _mm_mullo_epi32( qp1, interp0 );
      const __m128i ip1 = _mm_mullo_epi32( qp2, interp1 );
      const __m128i ip = _mm_add_epi32( *((const __m128i *)kThirtyTwoVector), _mm_add_epi32( ip0, ip1 ) );
      const __m128i dist = sad( _mm_and_si128( _mm_srli_epi32( ip, 6 ), kByteValMask ), pixel );
      __m128 errorVec = _mm_cvtepi32_ps( dist );
      
      errorVec = _mm_mul_ps( errorVec, errorMetricVec );
      errorVec = _mm_mul_ps( errorVec, errorVec );
      errorVec = _mm_hadd_ps( errorVec, errorVec );
      errorVec = _mm_hadd_ps( errorVec, errorVec );

      const __m128 cmp = _mm_cmple_ps( errorVec, minError );
      minError = _mm_blendv_ps( minError, errorVec, cmp );
      bestBucket = _mm_blendv_epi8( bestBucket, jVec, _mm_castps_si128( cmp ) );

      // Conceptually, once the error starts growing, it doesn't stop growing (we're moving
      // farther away from the reference point along the line). Hence we can early out here.
      // However, quanitzation artifacts mean that this is not ALWAYS the case, so we do suffer
      // about 0.01 RMS error. 
      if(!((uint8 *)(&cmp))[0])
        break;
    }

    totalError = _mm_add_ps(totalError, minError);
    if(indices) ((uint32 *)indices)[i] = ((uint32 *)(&bestBucket))[0];
  }

  return ((float *)(&totalError))[0];
}

///////////////////////////////////////////////////////////////////////////////
//
// Utility function implementation
//
///////////////////////////////////////////////////////////////////////////////

void ClampEndpoints(RGBAVectorSIMD &p1, RGBAVectorSIMD &p2) {
  p1.vec = _mm_min_ps( kByteMax, _mm_max_ps( p1.vec, kZero ) );
  p2.vec = _mm_min_ps( kByteMax, _mm_max_ps( p2.vec, kZero ) );
}

void GetPrincipalAxis(const RGBAClusterSIMD &c, RGBADirSIMD &axis) {

  if(c.GetNumPoints() == 2) {
    axis = c.GetPoint(1) - c.GetPoint(0);
    return;
  }

  RGBAVectorSIMD avg = c.GetTotal();
  avg /= float(c.GetNumPoints());

  // We use these vectors for calculating the covariance matrix...
  RGBAVectorSIMD toPts[kMaxNumDataPoints];
  RGBAVectorSIMD toPtsMax(-FLT_MAX);
  for(int i = 0; i < c.GetNumPoints(); i++) {
    toPts[i] = c.GetPoint(i) - avg;
    toPtsMax.vec = _mm_max_ps(toPtsMax.vec, toPts[i].vec);
  }

  RGBAMatrixSIMD covMatrix;

  // Compute covariance.
  const float fNumPoints = float(c.GetNumPoints());
  for(int i = 0; i < kNumColorChannels; i++) {
    for(int j = 0; j <= i; j++) {

      float sum = 0.0;
      for(int k = 0; k < c.GetNumPoints(); k++) {
        sum += toPts[k].c[i] * toPts[k].c[j];
      }

      covMatrix(i, j) = sum / fNumPoints;
      covMatrix(j, i) = covMatrix(i, j);
    }
  }

  // !SPEED! Find eigenvectors by using the power method. This is good because the
  // matrix is only 4x4, which allows us to use SIMD...
  RGBAVectorSIMD b = toPtsMax;
  assert(b.Length() > 0);
  b /= b.Length();

  RGBAVectorSIMD newB = covMatrix * b;

  // !HACK! If the principal eigenvector of the covariance matrix
  // converges to zero, that means that the points lie equally 
  // spaced on a sphere in this space. In this (extremely rare)
  // situation, just choose a point and use it as the principal 
  // direction.
  const float newBlen = newB.Length();
  if(newBlen < 1e-10) {
    axis = toPts[0];
    return;
  }

  for(int i = 0; i < 8; i++) {
    newB = covMatrix * b;
    newB.Normalize();
    b = newB;
  }

  axis = b;
}