Function wave8Untranspose_4

src/fl/channels/wave8.cpp.hpp:371–415

}

FL_OPTIMIZE_FUNCTION
void wave8Untranspose_4(const u8 (&FL_RESTRICT_PARAM transposed)[4 * sizeof(Wave8Byte)],
                        u8 (&FL_RESTRICT_PARAM output)[4 * sizeof(Wave8Byte)]) {
    // Reverse the 4-lane transposition
    // Input: 32 bytes of interleaved data (4 bytes per symbol, 8 symbols)
    // Output: 4 Wave8Byte structures (32 bytes total, de-interleaved)

    Wave8Byte lane_waves[4];

    // Process each of the 8 symbols
    for (int symbol_idx = 0; symbol_idx < 8; symbol_idx++) {
        u8 lane_bytes[4] = {0, 0, 0, 0};

        // Process 4 input bytes (2 pulses per byte)
        for (int byte_idx = 0; byte_idx < 4; byte_idx++) {
            u8 input_byte = transposed[symbol_idx * 4 + byte_idx];

            // Calculate which pulse bits these correspond to
            int pulse_bit_hi = 7 - (byte_idx * 2);
            int pulse_bit_lo = pulse_bit_hi - 1;

            // De-interleave 4 lanes from this byte
            // Bit layout: [L3_hi, L2_hi, L1_hi, L0_hi, L3_lo, L2_lo, L1_lo, L0_lo]
            for (int lane = 0; lane < 4; lane++) {
                // Extract bits for this lane
                u8 pulse_hi = (input_byte >> (4 + lane)) & 1;
                u8 pulse_lo = (input_byte >> lane) & 1;

                // Reconstruct lane byte
                lane_bytes[lane] |= (pulse_hi << pulse_bit_hi);
                lane_bytes[lane] |= (pulse_lo << pulse_bit_lo);
            }
        }

        // Store de-interleaved bytes
        for (int lane = 0; lane < 4; lane++) {
            lane_waves[lane].symbols[symbol_idx].data = lane_bytes[lane];
        }
    }

    // Copy de-interleaved data to output
    for (int lane = 0; lane < 4; lane++) {
        fl::isr::memcpy(output + lane * sizeof(Wave8Byte), &lane_waves[lane], sizeof(Wave8Byte));
    }
}

FL_OPTIMIZE_FUNCTION
void wave8Untranspose_8(const u8 (&FL_RESTRICT_PARAM transposed)[8 * sizeof(Wave8Byte)],