castano

// Returns p bit that results in the lowest RGB quantization error.
inline int quantize_rgb_bc7_mode6(float r, float g, float b, float * out_qr, float * out_qg, float * out_qb) {
    float qr = 2 * trunc(r * 127.5f);
    float qg = 2 * trunc(g * 127.5f);
    float qb = 2 * trunc(b * 127.5f);

    float dr = 255 * r - qr;
    float dg = 255 * g - qg;
    float db = 255 * b - qb;

castano

template <typename T>
struct Compare {
    T lambda;

#if _MSC_VER || __APPLE__
    static int compare(void * cmp, const void * a, const void * b)
#else
    static int compare(const void * a, const void * b, void * cmp)
#endif
    {

castano

// This code is in the public domain -- Ignacio Castaño <[email protected]>

#include "Sphere.h"
#include "Vector.inl"
#include "Box.inl"

#include <float.h> // FLT_MAX

const float radiusEpsilon = 1e-4f;

castano

// Emulating gathers using loads and permutevar8. This made the entire compressor about 15% faster. Both methods require AVX2.

// Load 4 uint8 per lane.
__m256i packedClusterIndex = _mm256_load_si256((__m256i *)&s_fourCluster[i]);

if (count <= 8) {
    // Load r_sat in one register:
    Wide8 r07 = load8(r_sat);
    Wide8 g07 = load8(g_sat);
    Wide8 b07 = load8(b_sat);

castano

    Wide8 besterror8 = broadcast8(FLT_MAX);
    Vector3_Wide8 beststart8;
    Vector3_Wide8 bestend8;

    // check all possible clusters for this total order
    for (int i = 0; i < s_fourClusterTotal[count - 1]; i += 8)
    {
        //uint c0 = s_fourCluster[i].c0;
        //uint c1 = s_fourCluster[i].c1;
        //uint c2 = s_fourCluster[i].c2;

castano

rgbcx v1.12
icbc v1.0
                 RMSE           PSNR            Time

Kodak/Waterloo Image Set:

rgbcx-0          8.128142       29.930977       0.163412
rgbcx-1          8.041399       30.024172       0.182219
rgbcx-2          8.198924       29.855667       0.213994
rgbcx-3          8.005721       30.062794       0.242393

castano

Perfect Quantization of DXT endpoints
-------------------------------------

One of the issues that affect the quality of most DXT compressors is the way floating point colors are rounded.

For example, stb_dxt does:

    max16 =  (unsigned short)(stb__sclamp((At1_r*yy - At2_r*xy)*frb+0.5f,0,31) << 11);
    max16 |= (unsigned short)(stb__sclamp((At1_g*yy - At2_g*xy)*fg +0.5f,0,63) << 5);
    max16 |= (unsigned short)(stb__sclamp((At1_b*yy - At2_b*xy)*frb+0.5f,0,31) << 0);

castano

                 RMSE           PSNR            Time

Kodak/Waterloo Image Set:

stb              8.202766       29.851597       0.258041
stb-hq           8.009301       30.058910       0.284019
nvtt-fast        8.089954       29.971882       0.445670

nvtt             7.616215       30.496019       6.806233
nvtt-hq          7.562366       30.557650       13.081200

castano

// This code is in the public domain -- Ignacio Castaño <[email protected]>

#include "Sphere.h"
#include "Vector.inl"
#include "Box.inl"

#include <float.h> // FLT_MAX

using namespace nv;

castano

I didn't think there was much to gain from optimizing quantization intervals for vertex compression, but I thought it would give it a try and see it for myself.

I have a set of vertices X that I want to quantize. The trivial way to do that is to transform them to the [0,1] range and store them using an integer UNORM format.

At runtime the vertices are reconstructed as follows:

X = Q * m + a

Where: