Skip to content

Commit

Permalink
More cube map filtering optimizations
Browse files Browse the repository at this point in the history
  • Loading branch information
@fo76utils
fo76utils committed Feb 4, 2025
1 parent 6574c9d commit c3f929e
Show file tree
Hide file tree
Showing 4 changed files with 119 additions and 62 deletions.
154 changes: 100 additions & 54 deletions lib/libfo76utils/src/ddstxt16.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1174,73 +1174,119 @@ FloatVector4 DDSTexture16::cubeMap(float x, float y, float z,
return c0;
}

#if ENABLE_GCC_SIMD_32
using FloatVecType = FloatVector8;
#else
using FloatVecType = FloatVector4;
#endif

FloatVector4 DDSTexture16::cubeMapImportanceSample(
FloatVector4 t, FloatVector4 b, FloatVector4 n,
const FloatVector4 *sampleBuf, size_t sampleCnt) const
{
if (!(channelCntFlags & 0x80)) [[unlikely]]
return cubeMap(n[0], n[1], n[2], 0.0f);
FloatVector4 c(0.0f);
do
#if ENABLE_GCC_SIMD_32
constexpr size_t k = 8;
#else
constexpr size_t k = 4;
#endif
FloatVecType xf0, yf0, xf1, yf1;
std::int32_t xi0[k], yi0[k], xi1[k], yi1[k], i[k];
std::int32_t m0[k], xMask[k];
FloatVecType mf, w;
for (size_t j = 0; true; j = (j + 1) & (k - 1))
{
FloatVector4 v = *(sampleBuf++);
float mipLevel = v[3];
int m0 = int(mipLevel);
float mf = float(m0);
unsigned int xMask = xMaskMip0 >> (unsigned char) m0;
FloatVector4 w = FloatVector4(float(int(xMask + 1U)));
FloatVector4 xyz = (t * v[0]) + (b * v[1]) + (n * v[2]);
FloatVector4 xyzm(xyz);
xyzm.absValues();
unsigned int m = xyz.getSignMask();
FloatVector4 tmp(xyzm);
tmp.maxValues(FloatVector4(xyzm).shuffleValues(0xC9));
tmp.maxValues(FloatVector4(xyzm).shuffleValues(0xD2));
unsigned int m2 = (xyzm - tmp).getSignMask();
float d = w[0] / tmp[0];
size_t i = m2 & 1U;
w = w * FloatVector4(0.5f, 0.5f, 0.25f, 0.25f) - 0.5f;
FloatVector4 s(0.5f, -0.5f, 0.25f, -0.25f);
if (!i) // +X (0), -X (1)
{
if (!(m & 1U))
s.shuffleValues(0xF5);
xyz.shuffleValues(0x66); // ZYZY
}
else if (!(m2 & 2U)) // +Y (2), -Y (3)
if (!j)
{
if (!(m & 2U))
s.shuffleValues(0xA0);
xyz.shuffleValues(0x88); // XZXZ
}
else // +Z (4), -Z (5)
{
i = 2;
if (m & 4U)
s.shuffleValues(0xF5);
xyz.shuffleValues(0x44); // XYXY
if (sampleCnt < k)
break;
sampleCnt = sampleCnt - k;
#if ENABLE_GCC_SIMD_32
FloatVecType xTmp(sampleBuf[0], sampleBuf[4]);
FloatVecType yTmp(sampleBuf[1], sampleBuf[5]);
FloatVecType zTmp(sampleBuf[2], sampleBuf[6]);
FloatVecType mipLevel(sampleBuf[3], sampleBuf[7]);
#else
FloatVecType xTmp(sampleBuf[0]);
FloatVecType yTmp(sampleBuf[1]);
FloatVecType zTmp(sampleBuf[2]);
FloatVecType mipLevel(sampleBuf[3]);
#endif
sampleBuf = sampleBuf + k;
w = zTmp;
mf = mipLevel;
mipLevel.floorValues().convertToInt32(m0);
mf -= mipLevel;
xMask[0] = std::int32_t(xMaskMip0 >> (unsigned char) m0[0]);
xMask[1] = std::int32_t(xMaskMip0 >> (unsigned char) m0[1]);
xMask[2] = std::int32_t(xMaskMip0 >> (unsigned char) m0[2]);
xMask[3] = std::int32_t(xMaskMip0 >> (unsigned char) m0[3]);
#if ENABLE_GCC_SIMD_32
xMask[4] = std::int32_t(xMaskMip0 >> (unsigned char) m0[4]);
xMask[5] = std::int32_t(xMaskMip0 >> (unsigned char) m0[5]);
xMask[6] = std::int32_t(xMaskMip0 >> (unsigned char) m0[6]);
xMask[7] = std::int32_t(xMaskMip0 >> (unsigned char) m0[7]);
FloatVecType wTmp = FloatVecType(xMask) * 0.5f;
#else
FloatVecType wTmp = FloatVecType::convertInt32(xMask) * 0.5f;
#endif
FloatVecType x = (xTmp * t[0]) + (yTmp * b[0]) + (zTmp * n[0]);
FloatVecType y = (xTmp * t[1]) + (yTmp * b[1]) + (zTmp * n[1]);
FloatVecType z = (xTmp * t[2]) + (yTmp * b[2]) + (zTmp * n[2]);
xTmp = FloatVecType(x).absValues();
yTmp = FloatVecType(y).absValues();
zTmp = FloatVecType(z).absValues();
FloatVecType mTmp = FloatVecType(xTmp).maxValues(yTmp).maxValues(zTmp);
FloatVecType d = (wTmp + 0.5f) / mTmp;
// -1 if face >= 2
xTmp = FloatVecType(1.0f).blendValues(FloatVecType(-1.0f), xTmp - mTmp);
// -1 if face 2 or 3
yTmp = FloatVecType(xTmp).blendValues(FloatVecType(1.0f), yTmp - mTmp);
// face 0, 1: X = -z / x, Y = -y / abs(x)
// face 2, 3: X = x / abs(y), Y = z / y
// face 4, 5: X = x / z, Y = -y / abs(z)
FloatVecType f0(0.0f);
f0.blendValues(FloatVecType(4.0f), xTmp);
f0.blendValues(FloatVecType(2.0f), yTmp);
// f0 = face & ~1, f1 = 1 - (face & 1) * 2
FloatVecType f1(x);
f1.blendValues(z, xTmp).blendValues(y, yTmp);
f1 = FloatVecType(1.0f).blendValues(FloatVecType(-1.0f), f1);
xf0 = (z * -1.0f).blendValues(x, xTmp) * f1;
yf0 = (y * -1.0f).blendValues(z * f1, yTmp);
xf0.blendValues(x, yTmp);
f0.blendValues(f0 + 1.0f, f1).convertToInt32(i);
xf0 = xf0 * d + wTmp;
yf0 = yf0 * d + wTmp;
xf1 = xf0 * 0.5f - 0.25f;
yf1 = yf0 * 0.5f - 0.25f;
xTmp = FloatVecType(xf0).floorValues();
yTmp = FloatVecType(yf0).floorValues();
xTmp.convertToInt32(xi0);
yTmp.convertToInt32(yi0);
xf0 -= xTmp;
yf0 -= yTmp;
xTmp = FloatVecType(xf1).floorValues();
yTmp = FloatVecType(yf1).floorValues();
xTmp.convertToInt32(xi1);
yTmp.convertToInt32(yi1);
xf1 -= xTmp;
yf1 -= yTmp;
}
xyz = xyz * s * d + w;
i = (i << 1) | ((m >> i) & 1U);
FloatVector4 xy_f(xyz);
xy_f.floorValues();
std::int32_t xy_i[4];
xy_f.convertToInt32(xy_i);
xy_f = xyz - xy_f;
FloatVector4 c0(getPixelB_Cube(textureData[m0], xy_i[0], xy_i[1], i,
textureDataSize, xy_f[0], xy_f[1], xMask));
if (xMask && mf != mipLevel) [[likely]]
FloatVector4 c0(getPixelB_Cube(textureData[m0[j]], xi0[j], yi0[j], i[j],
textureDataSize, xf0[j], yf0[j], xMask[j]));
if (xMask[j] && mf[j] > 0.0f) [[likely]]
{
mf = mipLevel - mf;
c0 -= (c0 * mf);
FloatVector4 c1(getPixelB_Cube(textureData[m0 + 1], xy_i[2], xy_i[3], i,
textureDataSize, xy_f[2], xy_f[3],
xMask >> 1));
c0 += (c1 * mf);
c0 -= (c0 * mf[j]);
FloatVector4 c1(getPixelB_Cube(textureData[m0[j] + 1], xi1[j], yi1[j],
i[j], textureDataSize, xf1[j], yf1[j],
xMask[j] >> 1));
c0 += (c1 * mf[j]);
}
c += (c0 * v[2]);
c += (c0 * w[j]);
}
while (--sampleCnt);
return c;
}

8 changes: 6 additions & 2 deletions lib/libfo76utils/src/ddstxt16.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -227,8 +227,12 @@ class DDSTexture16
// y = -1.0 to 1.0: S to N
// z = -1.0 to 1.0: bottom to top
FloatVector4 cubeMap(float x, float y, float z, float mipLevel) const;
// Calculate a weighted sum of cube map samples, each element of sampleBuf
// is a vector (X, Y, Z) and a mip level (W) in the range 0.0 to 16.0:
// Calculate a weighted sum of cube map samples, sampleBuf is expected to
// contain X, Y, Z vectors and mip levels (W) in the range 0.0 to 16.0 in
// this format:
// (X0, X1, X2, X3), (Y0, Y1, Y2, Y3), (Z0, Z1, Z2, Z3), (W0, W1, W2, W3)
// (X4, X5, X6, X7), (Y4, Y5, Y6, Y7), ...
// sampleCnt must be a multiple of 8. Each sample is calculated as follows:
// v = (t * X) + (b * Y) + (n * Z)
// sample = cubeMap(v[0], v[1], v[2], W) * Z
FloatVector4 cubeMapImportanceSample(
Expand Down
16 changes: 11 additions & 5 deletions lib/libfo76utils/src/sfcube2.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -329,8 +329,7 @@ size_t SFCubeMapFilter::convertImage(
{
int n = int(importanceSampleCnt);
importanceSampleTable = &importanceSampleBuf;
importanceSampleBuf.clear();
importanceSampleBuf.reserve(size_t(n));
importanceSampleBuf.resize(size_t(n));
float a = roughness * roughness;
float a2 = a * a;
filterParam = 0.0f;
Expand All @@ -341,8 +340,16 @@ size_t SFCubeMapFilter::convertImage(
float nDotH = h[2];
FloatVector4 l(h * (nDotH * 2.0f) // L = reflect(-N, H)
- FloatVector4(0.0f, 0.0f, 1.0f, 0.0f));
if (!(l[2] > 0.0f))
float *bufp = &(importanceSampleBuf.data()[i & ~3][i & 3]);
bufp[0] = l[0];
bufp[4] = l[1];
if (!(l[2] > 0.0f)) [[unlikely]]
{
bufp[8] = 0.0f;
bufp[12] = 16.0f;
continue;
}
bufp[8] = l[2];
filterParam += l[2];
// calculate mip level, based on formula from
// https://chetanjags.wordpress.com/2015/08/26/image-based-lighting/
Expand All @@ -351,8 +358,7 @@ size_t SFCubeMapFilter::convertImage(
float mipLevel = // mip bias = +1.0
float(std::log2(float(t.getWidth()) * float(t.getWidth())
/ (float(n) * d))) * 0.5f + 2.29248125f;
l[3] = std::min(std::max(mipLevel, 0.0f), 16.0f);
importanceSampleBuf.push_back(l);
bufp[12] = std::min(std::max(mipLevel, 0.0f), 16.0f);
}
filterParam = normalizeScale / filterParam;
}
Expand Down
3 changes: 2 additions & 1 deletion lib/libfo76utils/src/sfcube2.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -71,7 +71,8 @@ class SFCubeMapFilter
errorMessage("SFCubeMapFilter: invalid output dimensions");
width = std::uint32_t(w);
}
// set the number of samples to use for importance sampling (-1: disable)
// Set the number of samples to use for importance sampling.
// 'n' should be a multiple of 8, or -1 to use maximum quality
inline void setImportanceSamplingQuality(std::int32_t n)
{
importanceSampleCnt = std::uint32_t(n);
Expand Down

0 comments on commit c3f929e

Please sign in to comment.