wave8_transpose_4_bf1()

void fl::detail::wave8_transpose_4_bf1	(	const u8	lanes[4],
		u8	W0,
		u8	W1,
		u8	output[4 *sizeof(Wave8Byte)] )

Definition at line 227 of file wave8.cpp.hpp.

                                                                          {
    u8 d_mask[8];
    u8 m0_mask[8];
    const u8 D_byte = W0 ^ W1;
    for (int p = 0; p < 8; ++p) {
        const int shift = 7 - p;
        d_mask[p] = ((D_byte >> shift) & 1) ? 0xFFu : 0x00u;
        m0_mask[p] = ((W0 >> shift) & 1) ? 0xFFu : 0x00u;
    }
    const u32 aLo = spreadA(lanes[0]) | (spreadA(lanes[1]) << 1)
                  | (spreadA(lanes[2]) << 2) | (spreadA(lanes[3]) << 3);
    const u32 bLo = spreadB(lanes[0]) | (spreadB(lanes[1]) << 1)
                  | (spreadB(lanes[2]) << 2) | (spreadB(lanes[3]) << 3);
    u8 cols[8];
    cols[0] = static_cast<u8>(aLo);
    cols[1] = static_cast<u8>(aLo >> 8);
    cols[2] = static_cast<u8>(aLo >> 16);
    cols[3] = static_cast<u8>(aLo >> 24);
    cols[4] = static_cast<u8>(bLo);
    cols[5] = static_cast<u8>(bLo >> 8);
    cols[6] = static_cast<u8>(bLo >> 16);
    cols[7] = static_cast<u8>(bLo >> 24);
    for (int s = 0; s < 8; ++s) {
        const u8 col = cols[s];
        for (int k = 0; k < 4; ++k) {
            const int p_hi = 2 * k;
            const int p_lo = 2 * k + 1;
            const u8 hi = static_cast<u8>((m0_mask[p_hi] & 0xF0u) ^ ((col << 4) & d_mask[p_hi]));
            const u8 lo = static_cast<u8>((m0_mask[p_lo] & 0x0Fu) ^ (col & d_mask[p_lo]));
            output[s * 4 + k] = static_cast<u8>(hi | lo);
        }
    }
}

References fl::FL_NO_EXCEPT, spreadA(), spreadB(), fl::W0, and fl::W1.

Referenced by fl::wave8Transpose_4_bf1().

Here is the call graph for this function:

Here is the caller graph for this function: