Merge "Fix color cube tests."

            for (int y = 0; y < sy; y++) {

                for (int x = 0; x < sx; x++ ) {

                    dat[z*sy*sx + y*sx + x] = 0xff000000 |

                        ((x | (x<<2)) << 16) |

                        ((y | (y<<2)) << 8) |

                        ((z | (z<<2)) << 0);

                    dat[z*sy*sx + y*sx + x] = //0xff000000 |

                        (((x >> 2) | (x<<3)) << 0) |

                        (((y >> 2) | (y<<3)) << 8) |

                        ((z | (z<<4)) << 16);

                }

static rs_allocation gCube;

static int4 gDims;

static int4 gFracMask;

static int4 gFracBits;

static short4 gDims;

static short4 gFracMask;

static short4 gFracBits;

static short4 gFracShift;

static int4 gFinalShift;

static int4 gFinalAdd;

void setCube(rs_allocation c) {

    gCube = c;

    gDims.x = rsAllocationGetDimX(gCube);

    gDims.y = rsAllocationGetDimY(gCube);

    gDims.z = rsAllocationGetDimZ(gCube);

    gDims.x = rsAllocationGetDimX(gCube) - 1;

    gDims.y = rsAllocationGetDimY(gCube) - 1;

    gDims.z = rsAllocationGetDimZ(gCube) - 1;

    gDims.w = 0;

    gFracMask = gDims - 1;

    gFracBits = (int4)32 - clz(gFracMask);

    gFracMask = gDims;

    gFracBits = (short4)32 - clz(gFracMask);

    gFracShift = (short4)8 - gFracBits;

    rsDebug("dims", gDims);

    rsDebug("gFracMask", gFracMask);

    rsDebug("gFracBits", gFracBits);

    gFinalShift = gFracShift.x + gFracShift.y + gFracShift.z;

    gFinalAdd = (((int4)1 << gFinalShift) - (int4)1) >> (int4)1;

    rsDebug("gFinalShift", gFinalShift);

    rsDebug("gFinalAdd", gFinalAdd);

}

void root(const uchar4 *in, uchar4 *out) {

void root(const uchar4 *in, uchar4 *out, uint32_t x, uint32_t y) {

    //rsDebug("root", in);

    int4 coord1 = convert_int4(*in);

    int4 coord2 = min(coord1 + 1, gDims);

    short4 baseCoord = convert_short4(*in);

    short4 coord1 = baseCoord >> gFracShift;

    short4 coord2 = min(coord1 + (short4)1, gDims);

    short4 weight2 = baseCoord - (coord1 << gFracShift);

    short4 weight1 = ((short4)1 << gFracShift) - weight2;

    uchar4 v1 = rsGetElementAt_uchar4(gCube, coord1.x >> 3, coord1.y >> 3, coord1.z >> 4);

    ushort4 v000 = convert_ushort4(rsGetElementAt_uchar4(gCube, coord1.x, coord1.y, coord1.z));

    ushort4 v100 = convert_ushort4(rsGetElementAt_uchar4(gCube, coord2.x, coord1.y, coord1.z));

    ushort4 v010 = convert_ushort4(rsGetElementAt_uchar4(gCube, coord1.x, coord2.y, coord1.z));

    ushort4 v110 = convert_ushort4(rsGetElementAt_uchar4(gCube, coord2.x, coord2.y, coord1.z));

    ushort4 v001 = convert_ushort4(rsGetElementAt_uchar4(gCube, coord1.x, coord1.y, coord2.z));

    ushort4 v101 = convert_ushort4(rsGetElementAt_uchar4(gCube, coord2.x, coord1.y, coord2.z));

    ushort4 v011 = convert_ushort4(rsGetElementAt_uchar4(gCube, coord1.x, coord2.y, coord2.z));

    ushort4 v111 = convert_ushort4(rsGetElementAt_uchar4(gCube, coord2.x, coord2.y, coord2.z));

    //rsDebug("coord1", coord1);

    //rsDebug("coord2", coord2);

    uint4 yz00 = convert_uint4((v000 * weight1.x) + (v100 * weight2.x));

    uint4 yz10 = convert_uint4((v010 * weight1.x) + (v110 * weight2.x));

    uint4 yz01 = convert_uint4((v001 * weight1.x) + (v101 * weight2.x));

    uint4 yz11 = convert_uint4((v011 * weight1.x) + (v111 * weight2.x));

    uint4 z0 = (yz00 * weight1.y) + (yz10 * weight2.y);

    uint4 z1 = (yz01 * weight1.y) + (yz11 * weight2.y);

    uint4 v = (z0 * weight1.z) + (z1 * weight2.z);

    #if 0

    if (x + y < 100) {

        rsDebug("coord1", coord1);

        rsDebug("coord2", coord2);

        rsDebug("weight1", weight1);

        rsDebug("weight2", weight2);

        rsDebug("yz00", yz00);

        rsDebug("z0", z0);

        rsDebug("v", v);

    }

    #endif

    *out = v1;

    *out = convert_uchar4((v + gFinalAdd) >> gFinalShift);

    out->a = 0xff;

}