transposition.h

Go to the documentation of this file.

#pragma once

 
#include "fl/stl/compiler_control.h"
#include "fl/stl/int.h"
#include "fl/stl/span.h"
#include "fl/stl/optional.h"
#include "fl/stl/cstring.h"
#include "fl/stl/noexcept.h"
 
FL_OPTIMIZATION_LEVEL_O3_BEGIN
 
namespace fl {
 
// ============================================================================
// Core 8x1 Bit Transpose Functions
// ============================================================================
 
// Note: These transpose functions are used across multiple platforms,
// so they are defined for all targets

typedef union {
    fl::u8 raw;   
    struct {
        fl::u32 a0:1;  
        fl::u32 a1:1;  
        fl::u32 a2:1;  
        fl::u32 a3:1;  
        fl::u32 a4:1;  
        fl::u32 a5:1;  
        fl::u32 a6:1;  
        fl::u32 a7:1;  
    };
} just8bits;

typedef struct {
    fl::u32 a0:1;  
    fl::u32 a1:1;  
    fl::u32 a2:1;  
    fl::u32 a3:1;  
    fl::u32 a4:1;  
    fl::u32 a5:1;  
    fl::u32 a6:1;  
    fl::u32 a7:1;  
    fl::u32 b0:1;  
    fl::u32 b1:1;  
    fl::u32 b2:1;  
    fl::u32 b3:1;  
    fl::u32 b4:1;  
    fl::u32 b5:1;  
    fl::u32 b6:1;  
    fl::u32 b7:1;  
    fl::u32 c0:1;  
    fl::u32 c1:1;  
    fl::u32 c2:1;  
    fl::u32 c3:1;  
    fl::u32 c4:1;  
    fl::u32 c5:1;  
    fl::u32 c6:1;  
    fl::u32 c7:1;  
    fl::u32 d0:1;  
    fl::u32 d1:1;  
    fl::u32 d2:1;  
    fl::u32 d3:1;  
    fl::u32 d4:1;  
    fl::u32 d5:1;  
    fl::u32 d6:1;  
    fl::u32 d7:1;  
} sub4;

typedef union {
    fl::u32 word[2];  
    fl::u8 bytes[8];  
    struct {
        sub4 a;  
        sub4 b;  
    };
} bitswap_type;

void transpose8x1_noinline(unsigned char *A, unsigned char *B) FL_NOEXCEPT;

FASTLED_FORCE_INLINE void transpose8x1(unsigned char *A, unsigned char *B) FL_NOEXCEPT {
    fl::u32 x, y, t;
 
    // Load the array and pack it into x and y.
    y = *(fl::u32*)(A);
    x = *(fl::u32*)(A+4);
 
    // pre-transform x
    t = (x ^ (x >> 7)) & 0x00AA00AA;  x = x ^ t ^ (t << 7);
    t = (x ^ (x >>14)) & 0x0000CCCC;  x = x ^ t ^ (t <<14);
 
    // pre-transform y
    t = (y ^ (y >> 7)) & 0x00AA00AA;  y = y ^ t ^ (t << 7);
    t = (y ^ (y >>14)) & 0x0000CCCC;  y = y ^ t ^ (t <<14);
 
    // final transform
    t = (x & 0xF0F0F0F0) | ((y >> 4) & 0x0F0F0F0F);
    y = ((x << 4) & 0xF0F0F0F0) | (y & 0x0F0F0F0F);
    x = t;
 
    *((u32*)B) = y;
    *((u32*)(B+4)) = x;
}

FASTLED_FORCE_INLINE void transpose8x1_MSB(unsigned char *A, unsigned char *B) FL_NOEXCEPT {
    fl::u32 x, y, t;
 
    // Load the array and pack it into x and y.
    y = *(fl::u32*)(A);
    x = *(fl::u32*)(A+4);
 
    // pre-transform x
    t = (x ^ (x >> 7)) & 0x00AA00AA;  x = x ^ t ^ (t << 7);
    t = (x ^ (x >>14)) & 0x0000CCCC;  x = x ^ t ^ (t <<14);
 
    // pre-transform y
    t = (y ^ (y >> 7)) & 0x00AA00AA;  y = y ^ t ^ (t << 7);
    t = (y ^ (y >>14)) & 0x0000CCCC;  y = y ^ t ^ (t <<14);
 
    // final transform
    t = (x & 0xF0F0F0F0) | ((y >> 4) & 0x0F0F0F0F);
    y = ((x << 4) & 0xF0F0F0F0) | (y & 0x0F0F0F0F);
    x = t;
 
    B[7] = y; y >>= 8;
    B[6] = y; y >>= 8;
    B[5] = y; y >>= 8;
    B[4] = y;
 
    B[3] = x; x >>= 8;
    B[2] = x; x >>= 8;
    B[1] = x; x >>= 8;
    B[0] = x;
}

template<int m, int n>
FASTLED_FORCE_INLINE void transpose8(unsigned char *A, unsigned char *B) FL_NOEXCEPT {
    fl::u32 x, y, t;
 
    // Load the array and pack it into x and y.
    if(m == 1) {
        y = *(fl::u32*)(A);
        x = *(fl::u32*)(A+4);
    } else {
        x = (fl::u32(A[0])<<24)   | (fl::u32(A[m])<<16)   | (fl::u32(A[2*m])<<8) | A[3*m];
        y = (fl::u32(A[4*m])<<24) | (fl::u32(A[5*m])<<16) | (fl::u32(A[6*m])<<8) | A[7*m];
    }
 
    // pre-transform x
    t = (x ^ (x >> 7)) & 0x00AA00AA;  x = x ^ t ^ (t << 7);
    t = (x ^ (x >>14)) & 0x0000CCCC;  x = x ^ t ^ (t <<14);
 
    // pre-transform y
    t = (y ^ (y >> 7)) & 0x00AA00AA;  y = y ^ t ^ (t << 7);
    t = (y ^ (y >>14)) & 0x0000CCCC;  y = y ^ t ^ (t <<14);
 
    // final transform
    t = (x & 0xF0F0F0F0) | ((y >> 4) & 0x0F0F0F0F);
    y = ((x << 4) & 0xF0F0F0F0) | (y & 0x0F0F0F0F);
    x = t;
 
    B[7*n] = y; y >>= 8;
    B[6*n] = y; y >>= 8;
    B[5*n] = y; y >>= 8;
    B[4*n] = y;
 
    B[3*n] = x; x >>= 8;
    B[2*n] = x; x >>= 8;
    B[n] = x; x >>= 8;
    B[0] = x;
}
 
// ============================================================================
// Low-Level ISR-Safe Transposition Primitives
// ============================================================================

inline void transpose_2lane_inline(
    const u8* lane0_byte,
    const u8* lane1_byte,
    u8* output,
    size_t num_bytes
) FL_NOEXCEPT;

inline void transpose_4lane_inline(
    const u8* const lanes[4],
    u8* output,
    size_t num_bytes
) FL_NOEXCEPT;

inline void transpose_8lane_inline(
    const u8* const lanes[8],
    u8* output,
    size_t num_bytes
) FL_NOEXCEPT;

inline void transpose_16lane_inline(
    const u8* const lanes[16],
    u8* output,
    size_t num_bytes
) FL_NOEXCEPT;

template<typename TSource>
inline void transpose_generic_inline(
    const TSource* const lanes[],
    size_t num_lanes,
    u8* output,
    size_t num_items
) FL_NOEXCEPT;
 
// Implementation of inline ISR-safe primitives
 
inline void transpose_2lane_inline(
    const u8* lane0_byte,
    const u8* lane1_byte,
    u8* output,
    size_t num_bytes
) FL_NOEXCEPT {
    for (size_t byte_idx = 0; byte_idx < num_bytes; byte_idx++) {
        u8 a = lane0_byte[byte_idx];
        u8 b = lane1_byte[byte_idx];
 
        // dest[0] contains bit pairs for positions 7,6,5,4 (MSB first)
        output[byte_idx * 2 + 0] =
            ((a >> 7) & 0x01) << 0 | ((b >> 7) & 0x01) << 1 |
            ((a >> 6) & 0x01) << 2 | ((b >> 6) & 0x01) << 3 |
            ((a >> 5) & 0x01) << 4 | ((b >> 5) & 0x01) << 5 |
            ((a >> 4) & 0x01) << 6 | ((b >> 4) & 0x01) << 7;
 
        // dest[1] contains bit pairs for positions 3,2,1,0 (LSB)
        output[byte_idx * 2 + 1] =
            ((a >> 3) & 0x01) << 0 | ((b >> 3) & 0x01) << 1 |
            ((a >> 2) & 0x01) << 2 | ((b >> 2) & 0x01) << 3 |
            ((a >> 1) & 0x01) << 4 | ((b >> 1) & 0x01) << 5 |
            ((a >> 0) & 0x01) << 6 | ((b >> 0) & 0x01) << 7;
    }
}
 
inline void transpose_4lane_inline(
    const u8* const lanes[4],
    u8* output,
    size_t num_bytes
) FL_NOEXCEPT {
    for (size_t byte_idx = 0; byte_idx < num_bytes; byte_idx++) {
        u8 a = lanes[0][byte_idx];
        u8 b = lanes[1][byte_idx];
        u8 c = lanes[2][byte_idx];
        u8 d = lanes[3][byte_idx];
 
        u8* dest = &output[byte_idx * 4];
 
        dest[0] = ((a >> 7) & 0x01) << 0 | ((b >> 7) & 0x01) << 1 | ((c >> 7) & 0x01) << 2 | ((d >> 7) & 0x01) << 3 |
                  ((a >> 6) & 0x01) << 4 | ((b >> 6) & 0x01) << 5 | ((c >> 6) & 0x01) << 6 | ((d >> 6) & 0x01) << 7;
 
        dest[1] = ((a >> 5) & 0x01) << 0 | ((b >> 5) & 0x01) << 1 | ((c >> 5) & 0x01) << 2 | ((d >> 5) & 0x01) << 3 |
                  ((a >> 4) & 0x01) << 4 | ((b >> 4) & 0x01) << 5 | ((c >> 4) & 0x01) << 6 | ((d >> 4) & 0x01) << 7;
 
        dest[2] = ((a >> 3) & 0x01) << 0 | ((b >> 3) & 0x01) << 1 | ((c >> 3) & 0x01) << 2 | ((d >> 3) & 0x01) << 3 |
                  ((a >> 2) & 0x01) << 4 | ((b >> 2) & 0x01) << 5 | ((c >> 2) & 0x01) << 6 | ((d >> 2) & 0x01) << 7;
 
        dest[3] = ((a >> 1) & 0x01) << 0 | ((b >> 1) & 0x01) << 1 | ((c >> 1) & 0x01) << 2 | ((d >> 1) & 0x01) << 3 |
                  ((a >> 0) & 0x01) << 4 | ((b >> 0) & 0x01) << 5 | ((c >> 0) & 0x01) << 6 | ((d >> 0) & 0x01) << 7;
    }
}
 
inline void transpose_8lane_inline(
    const u8* const lanes[8],
    u8* output,
    size_t num_bytes
) FL_NOEXCEPT {
    for (size_t byte_idx = 0; byte_idx < num_bytes; byte_idx++) {
        // Pack 8 bytes into a single 64-bit register
        // This reduces register pressure and enables parallel bit extraction
        u64 packed =
            ((u64)lanes[0][byte_idx] << 0)  |
            ((u64)lanes[1][byte_idx] << 8)  |
            ((u64)lanes[2][byte_idx] << 16) |
            ((u64)lanes[3][byte_idx] << 24) |
            ((u64)lanes[4][byte_idx] << 32) |
            ((u64)lanes[5][byte_idx] << 40) |
            ((u64)lanes[6][byte_idx] << 48) |
            ((u64)lanes[7][byte_idx] << 56);
 
        u8* dest = &output[byte_idx * 8];
 
        // Extract bits in parallel (compiler can optimize independent shifts)
        for (int bit = 7; bit >= 0; bit--) {
            dest[7 - bit] =
                ((packed >> (bit + 0))  & 0x01) << 0 |
                ((packed >> (bit + 8))  & 0x01) << 1 |
                ((packed >> (bit + 16)) & 0x01) << 2 |
                ((packed >> (bit + 24)) & 0x01) << 3 |
                ((packed >> (bit + 32)) & 0x01) << 4 |
                ((packed >> (bit + 40)) & 0x01) << 5 |
                ((packed >> (bit + 48)) & 0x01) << 6 |
                ((packed >> (bit + 56)) & 0x01) << 7;
        }
    }
}
 
inline void transpose_16lane_inline(
    const u8* const lanes[16],
    u8* output,
    size_t num_bytes
) FL_NOEXCEPT {
    for (size_t byte_idx = 0; byte_idx < num_bytes; byte_idx++) {
        // Pack lanes 0-7 into first 64-bit register
        u64 packed_lo =
            ((u64)lanes[0][byte_idx] << 0)  |
            ((u64)lanes[1][byte_idx] << 8)  |
            ((u64)lanes[2][byte_idx] << 16) |
            ((u64)lanes[3][byte_idx] << 24) |
            ((u64)lanes[4][byte_idx] << 32) |
            ((u64)lanes[5][byte_idx] << 40) |
            ((u64)lanes[6][byte_idx] << 48) |
            ((u64)lanes[7][byte_idx] << 56);
 
        // Pack lanes 8-15 into second 64-bit register
        u64 packed_hi =
            ((u64)lanes[8][byte_idx]  << 0)  |
            ((u64)lanes[9][byte_idx]  << 8)  |
            ((u64)lanes[10][byte_idx] << 16) |
            ((u64)lanes[11][byte_idx] << 24) |
            ((u64)lanes[12][byte_idx] << 32) |
            ((u64)lanes[13][byte_idx] << 40) |
            ((u64)lanes[14][byte_idx] << 48) |
            ((u64)lanes[15][byte_idx] << 56);
 
        u8* dest = &output[byte_idx * 16];
 
        // Extract bits in parallel from both packed registers
        for (int bit = 7; bit >= 0; bit--) {
            dest[7 - bit] =
                ((packed_lo >> (bit + 0))  & 0x01) << 0 |
                ((packed_lo >> (bit + 8))  & 0x01) << 1 |
                ((packed_lo >> (bit + 16)) & 0x01) << 2 |
                ((packed_lo >> (bit + 24)) & 0x01) << 3 |
                ((packed_lo >> (bit + 32)) & 0x01) << 4 |
                ((packed_lo >> (bit + 40)) & 0x01) << 5 |
                ((packed_lo >> (bit + 48)) & 0x01) << 6 |
                ((packed_lo >> (bit + 56)) & 0x01) << 7;
 
            dest[15 - bit] =
                ((packed_hi >> (bit + 0))  & 0x01) << 0 |
                ((packed_hi >> (bit + 8))  & 0x01) << 1 |
                ((packed_hi >> (bit + 16)) & 0x01) << 2 |
                ((packed_hi >> (bit + 24)) & 0x01) << 3 |
                ((packed_hi >> (bit + 32)) & 0x01) << 4 |
                ((packed_hi >> (bit + 40)) & 0x01) << 5 |
                ((packed_hi >> (bit + 48)) & 0x01) << 6 |
                ((packed_hi >> (bit + 56)) & 0x01) << 7;
        }
    }
}
 
template<typename TSource>
inline void transpose_generic_inline(
    const TSource* const lanes[],
    size_t num_lanes,
    u8* output,
    size_t num_items
) FL_NOEXCEPT {
    constexpr size_t bits_per_item = sizeof(TSource) * 8;
 
    for (size_t item_idx = 0; item_idx < num_items; item_idx++) {
        u8* dest = &output[item_idx * bits_per_item];
 
        // Process each bit position in the source data (MSB to LSB)
        for (size_t bit_pos = 0; bit_pos < bits_per_item; bit_pos++) {
            size_t src_bit = (bits_per_item - 1) - bit_pos;
            u8 output_byte = 0;
 
            // Extract bit from each lane (up to 8 lanes per output byte)
            for (size_t lane = 0; lane < num_lanes && lane < 8; lane++) {
                TSource src_value = lanes[lane][item_idx];
                u8 bit = (src_value >> src_bit) & 0x01;
                output_byte |= (bit << (7 - lane));
            }
 
            dest[bit_pos] = output_byte;
        }
    }
}
 
// ============================================================================
// SPI Multi-Lane Transposer
// ============================================================================

class SPITransposer {
public:
    struct LaneData {
        fl::span<const u8> payload;        
        fl::span<const u8> padding_frame;  
    };

    static bool transpose2(const fl::optional<LaneData>& lane0,
                          const fl::optional<LaneData>& lane1,
                          fl::span<u8> output,
                          const char** error = nullptr) FL_NOEXCEPT;

    static bool transpose4(const fl::optional<LaneData>& lane0,
                          const fl::optional<LaneData>& lane1,
                          const fl::optional<LaneData>& lane2,
                          const fl::optional<LaneData>& lane3,
                          fl::span<u8> output,
                          const char** error = nullptr) FL_NOEXCEPT;

    static bool transpose8(const fl::optional<LaneData> lanes[8],
                          fl::span<u8> output,
                          const char** error = nullptr) FL_NOEXCEPT;

    static bool transpose16(const fl::optional<LaneData> lanes[16],
                           fl::span<u8> output,
                           const char** error = nullptr) FL_NOEXCEPT;
 
private:
    static u8 getLaneByte(const LaneData& lane, size_t byte_idx, size_t max_size) FL_NOEXCEPT;
};
 
// ============================================================================
// Parallel Strip Transposer (RP2040/RP2350 PIO)
// ============================================================================

FASTLED_FORCE_INLINE void transpose_8strips(
    const u8* const input[8],
    u8* output,
    u16 num_leds,
    u8 bytes_per_led
) FL_NOEXCEPT {
    // Process each LED
    for (u16 led = 0; led < num_leds; led++) {
        u8 temp_input[8];
 
        // Process each byte in the LED
        for (u8 byte_idx = 0; byte_idx < bytes_per_led; byte_idx++) {
            // Collect one byte from each strip for this byte position
            for (int strip = 0; strip < 8; strip++) {
                temp_input[strip] = input[strip][led * bytes_per_led + byte_idx];
            }
 
            // Transpose 8 bytes → 8 bytes (1 bit from each strip per output byte)
            transpose8x1_MSB(temp_input, output);
 
            // Advance output pointer by 8 bytes
            output += 8;
        }
    }
}

FASTLED_FORCE_INLINE void transpose_4strips(
    const u8* const input[4],
    u8* output,
    u16 num_leds,
    u8 bytes_per_led
) FL_NOEXCEPT {
    // Process each LED
    for (u16 led = 0; led < num_leds; led++) {
        // Process each byte in the LED
        for (u8 byte_idx = 0; byte_idx < bytes_per_led; byte_idx++) {
            // Collect one byte from each strip for this byte position
            u8 strip_bytes[4];
            for (int strip = 0; strip < 4; strip++) {
                strip_bytes[strip] = input[strip][led * bytes_per_led + byte_idx];
            }
 
            // Transpose: extract each bit position from all 4 strips
            for (int bit = 7; bit >= 0; bit--) {
                u8 output_byte = 0;
                // Pack bits from all 4 strips into lower 4 bits
                for (int strip = 0; strip < 4; strip++) {
                    output_byte |= ((strip_bytes[strip] >> bit) & 1) << strip;
                }
                *output++ = output_byte;
            }
        }
    }
}

FASTLED_FORCE_INLINE void transpose_2strips(
    const u8* const input[2],
    u8* output,
    u16 num_leds,
    u8 bytes_per_led
) FL_NOEXCEPT {
    // Process each LED
    for (u16 led = 0; led < num_leds; led++) {
        // Process each byte in the LED
        for (u8 byte_idx = 0; byte_idx < bytes_per_led; byte_idx++) {
            // Collect one byte from each strip for this byte position
            u8 strip_bytes[2];
            strip_bytes[0] = input[0][led * bytes_per_led + byte_idx];
            strip_bytes[1] = input[1][led * bytes_per_led + byte_idx];
 
            // Transpose: extract each bit position from both strips
            for (int bit = 7; bit >= 0; bit--) {
                u8 output_byte =
                    ((strip_bytes[0] >> bit) & 1) |
                    (((strip_bytes[1] >> bit) & 1) << 1);
                *output++ = output_byte;
            }
        }
    }
}

FASTLED_FORCE_INLINE u32 calculate_transpose_buffer_size(u16 num_leds, u8 bytes_per_led) FL_NOEXCEPT {
    return num_leds * bytes_per_led * 8;
}

inline bool transpose_strips(
    u8 num_strips,
    const u8* const* input,
    u8* output,
    u16 num_leds,
    u8 bytes_per_led
) FL_NOEXCEPT {
    switch (num_strips) {
        case 8:
            transpose_8strips(input, output, num_leds, bytes_per_led);
            return true;
        case 4:
            transpose_4strips(input, output, num_leds, bytes_per_led);
            return true;
        case 2:
            transpose_2strips(input, output, num_leds, bytes_per_led);
            return true;
        default:
            return false;  // Invalid strip count
    }
}
 
// ============================================================================
// PARLIO Wave8 Transposer (ESP32-S3 Parallel I/O)
// ============================================================================

template<size_t DATA_WIDTH>
FASTLED_FORCE_INLINE FL_IRAM FL_OPTIMIZE_FUNCTION size_t transpose_wave8byte_parlio_template(
    const u8* FL_RESTRICT_PARAM laneWaveforms,
    u8* FL_RESTRICT_PARAM outputBuffer
) FL_NOEXCEPT {
    constexpr size_t bytes_per_lane = 8;   // sizeof(Wave8Byte)
    constexpr size_t pulsesPerByte = 64;   // 8 bits × 8 pulses per bit
    size_t outputIdx = 0;
 
    // Note: Using regular if statements (C++11 compatible)
    // Compiler optimizes away dead branches for constant template parameters
    if (DATA_WIDTH == 8) {
        // Special optimized case for 8 lanes with bit packing
        // Optimized: Hoist packing outside inner loop to reduce redundant operations
        for (size_t bit_pos = 0; bit_pos < 8; bit_pos++) {
            // Pack 8 wave8_byte values into a single 64-bit register for parallel extraction
            // This packing is done once per bit_pos (8 times) instead of 64 times
            u64 packed =
                ((u64)laneWaveforms[0 * bytes_per_lane + bit_pos] << 0)  |
                ((u64)laneWaveforms[1 * bytes_per_lane + bit_pos] << 8)  |
                ((u64)laneWaveforms[2 * bytes_per_lane + bit_pos] << 16) |
                ((u64)laneWaveforms[3 * bytes_per_lane + bit_pos] << 24) |
                ((u64)laneWaveforms[4 * bytes_per_lane + bit_pos] << 32) |
                ((u64)laneWaveforms[5 * bytes_per_lane + bit_pos] << 40) |
                ((u64)laneWaveforms[6 * bytes_per_lane + bit_pos] << 48) |
                ((u64)laneWaveforms[7 * bytes_per_lane + bit_pos] << 56);
 
            // Inner loop: extract 8 pulses from the packed data
            for (size_t pulse_bit = 0; pulse_bit < 8; pulse_bit++) {
                // Extract pulse bits in parallel (compiler can optimize independent shifts)
                outputBuffer[outputIdx++] =
                    ((packed >> (7 - pulse_bit + 0))  & 0x01) << 0 |
                    ((packed >> (7 - pulse_bit + 8))  & 0x01) << 1 |
                    ((packed >> (7 - pulse_bit + 16)) & 0x01) << 2 |
                    ((packed >> (7 - pulse_bit + 24)) & 0x01) << 3 |
                    ((packed >> (7 - pulse_bit + 32)) & 0x01) << 4 |
                    ((packed >> (7 - pulse_bit + 40)) & 0x01) << 5 |
                    ((packed >> (7 - pulse_bit + 48)) & 0x01) << 6 |
                    ((packed >> (7 - pulse_bit + 56)) & 0x01) << 7;
            }
        }
    } else if (DATA_WIDTH <= 8) {
        // Pack into single bytes (compile-time branch elimination via template instantiation)
        // Guard against division by zero when DATA_WIDTH > 8 (shouldn't execute this branch, but compiler still evaluates it)
        const size_t ticksPerByte = (DATA_WIDTH > 8) ? 1 : (8 / DATA_WIDTH);
        const size_t numOutputBytes = (pulsesPerByte + ticksPerByte - 1) / ticksPerByte;
 
        for (size_t outputByteIdx = 0; outputByteIdx < numOutputBytes; outputByteIdx++) {
            u8 outputByte = 0;
 
            FL_UNROLL(8)
            for (size_t t = 0; t < ticksPerByte; t++) {
                size_t pulse_idx = outputByteIdx * ticksPerByte + t;
                if (pulse_idx >= pulsesPerByte)
                    break;
 
                size_t bit_pos = pulse_idx / 8;
                size_t pulse_bit = pulse_idx % 8;
 
                FL_UNROLL(8)
                for (size_t lane = 0; lane < DATA_WIDTH; lane++) {
                    const u8* laneWaveform = laneWaveforms + (lane * bytes_per_lane);
                    u8 wave8_byte = laneWaveform[bit_pos];
                    u8 pulse = (wave8_byte >> (7 - pulse_bit)) & 1;
 
                    size_t bitPos = t * DATA_WIDTH + lane;
                    outputByte |= (pulse << bitPos);
                }
            }
 
            outputBuffer[outputIdx++] = outputByte;
        }
    } else if (DATA_WIDTH == 16) {
        // Pack into 16-bit words (compile-time branch)
        // Optimized: Software pipelining + output buffering
        // Process 2 bit positions in parallel for better ILP, and batch writes for better cache efficiency
 
        // Output buffer: accumulate 16 words (32 bytes) before writing
        // This aligns with typical 32-byte cache lines and reduces memory write overhead
        u8 writeBuffer[32];
        size_t writeIdx = 0;
 
        for (size_t bit_pos = 0; bit_pos < 8; bit_pos += 2) {
            // Pack 16 wave8_byte values for TWO bit positions simultaneously
            // This enables instruction-level parallelism and better register utilization
            u64 packed_lo_0 =
                ((u64)laneWaveforms[0 * bytes_per_lane + bit_pos + 0] << 0)  |
                ((u64)laneWaveforms[1 * bytes_per_lane + bit_pos + 0] << 8)  |
                ((u64)laneWaveforms[2 * bytes_per_lane + bit_pos + 0] << 16) |
                ((u64)laneWaveforms[3 * bytes_per_lane + bit_pos + 0] << 24) |
                ((u64)laneWaveforms[4 * bytes_per_lane + bit_pos + 0] << 32) |
                ((u64)laneWaveforms[5 * bytes_per_lane + bit_pos + 0] << 40) |
                ((u64)laneWaveforms[6 * bytes_per_lane + bit_pos + 0] << 48) |
                ((u64)laneWaveforms[7 * bytes_per_lane + bit_pos + 0] << 56);
 
            u64 packed_hi_0 =
                ((u64)laneWaveforms[8  * bytes_per_lane + bit_pos + 0] << 0)  |
                ((u64)laneWaveforms[9  * bytes_per_lane + bit_pos + 0] << 8)  |
                ((u64)laneWaveforms[10 * bytes_per_lane + bit_pos + 0] << 16) |
                ((u64)laneWaveforms[11 * bytes_per_lane + bit_pos + 0] << 24) |
                ((u64)laneWaveforms[12 * bytes_per_lane + bit_pos + 0] << 32) |
                ((u64)laneWaveforms[13 * bytes_per_lane + bit_pos + 0] << 40) |
                ((u64)laneWaveforms[14 * bytes_per_lane + bit_pos + 0] << 48) |
                ((u64)laneWaveforms[15 * bytes_per_lane + bit_pos + 0] << 56);
 
            u64 packed_lo_1 =
                ((u64)laneWaveforms[0 * bytes_per_lane + bit_pos + 1] << 0)  |
                ((u64)laneWaveforms[1 * bytes_per_lane + bit_pos + 1] << 8)  |
                ((u64)laneWaveforms[2 * bytes_per_lane + bit_pos + 1] << 16) |
                ((u64)laneWaveforms[3 * bytes_per_lane + bit_pos + 1] << 24) |
                ((u64)laneWaveforms[4 * bytes_per_lane + bit_pos + 1] << 32) |
                ((u64)laneWaveforms[5 * bytes_per_lane + bit_pos + 1] << 40) |
                ((u64)laneWaveforms[6 * bytes_per_lane + bit_pos + 1] << 48) |
                ((u64)laneWaveforms[7 * bytes_per_lane + bit_pos + 1] << 56);
 
            u64 packed_hi_1 =
                ((u64)laneWaveforms[8  * bytes_per_lane + bit_pos + 1] << 0)  |
                ((u64)laneWaveforms[9  * bytes_per_lane + bit_pos + 1] << 8)  |
                ((u64)laneWaveforms[10 * bytes_per_lane + bit_pos + 1] << 16) |
                ((u64)laneWaveforms[11 * bytes_per_lane + bit_pos + 1] << 24) |
                ((u64)laneWaveforms[12 * bytes_per_lane + bit_pos + 1] << 32) |
                ((u64)laneWaveforms[13 * bytes_per_lane + bit_pos + 1] << 40) |
                ((u64)laneWaveforms[14 * bytes_per_lane + bit_pos + 1] << 48) |
                ((u64)laneWaveforms[15 * bytes_per_lane + bit_pos + 1] << 56);
 
            // Inner loop: interleave extraction from both bit positions
            // This allows CPU to execute independent operations in parallel
            for (size_t pulse_bit = 0; pulse_bit < 8; pulse_bit++) {
                // Extract pulse bits for first bit position
                u16 outputWord_0 =
                    ((packed_lo_0 >> (7 - pulse_bit + 0))  & 0x01) << 0  |
                    ((packed_lo_0 >> (7 - pulse_bit + 8))  & 0x01) << 1  |
                    ((packed_lo_0 >> (7 - pulse_bit + 16)) & 0x01) << 2  |
                    ((packed_lo_0 >> (7 - pulse_bit + 24)) & 0x01) << 3  |
                    ((packed_lo_0 >> (7 - pulse_bit + 32)) & 0x01) << 4  |
                    ((packed_lo_0 >> (7 - pulse_bit + 40)) & 0x01) << 5  |
                    ((packed_lo_0 >> (7 - pulse_bit + 48)) & 0x01) << 6  |
                    ((packed_lo_0 >> (7 - pulse_bit + 56)) & 0x01) << 7  |
                    ((packed_hi_0 >> (7 - pulse_bit + 0))  & 0x01) << 8  |
                    ((packed_hi_0 >> (7 - pulse_bit + 8))  & 0x01) << 9  |
                    ((packed_hi_0 >> (7 - pulse_bit + 16)) & 0x01) << 10 |
                    ((packed_hi_0 >> (7 - pulse_bit + 24)) & 0x01) << 11 |
                    ((packed_hi_0 >> (7 - pulse_bit + 32)) & 0x01) << 12 |
                    ((packed_hi_0 >> (7 - pulse_bit + 40)) & 0x01) << 13 |
                    ((packed_hi_0 >> (7 - pulse_bit + 48)) & 0x01) << 14 |
                    ((packed_hi_0 >> (7 - pulse_bit + 56)) & 0x01) << 15;
 
                // Extract pulse bits for second bit position
                u16 outputWord_1 =
                    ((packed_lo_1 >> (7 - pulse_bit + 0))  & 0x01) << 0  |
                    ((packed_lo_1 >> (7 - pulse_bit + 8))  & 0x01) << 1  |
                    ((packed_lo_1 >> (7 - pulse_bit + 16)) & 0x01) << 2  |
                    ((packed_lo_1 >> (7 - pulse_bit + 24)) & 0x01) << 3  |
                    ((packed_lo_1 >> (7 - pulse_bit + 32)) & 0x01) << 4  |
                    ((packed_lo_1 >> (7 - pulse_bit + 40)) & 0x01) << 5  |
                    ((packed_lo_1 >> (7 - pulse_bit + 48)) & 0x01) << 6  |
                    ((packed_lo_1 >> (7 - pulse_bit + 56)) & 0x01) << 7  |
                    ((packed_hi_1 >> (7 - pulse_bit + 0))  & 0x01) << 8  |
                    ((packed_hi_1 >> (7 - pulse_bit + 8))  & 0x01) << 9  |
                    ((packed_hi_1 >> (7 - pulse_bit + 16)) & 0x01) << 10 |
                    ((packed_hi_1 >> (7 - pulse_bit + 24)) & 0x01) << 11 |
                    ((packed_hi_1 >> (7 - pulse_bit + 32)) & 0x01) << 12 |
                    ((packed_hi_1 >> (7 - pulse_bit + 40)) & 0x01) << 13 |
                    ((packed_hi_1 >> (7 - pulse_bit + 48)) & 0x01) << 14 |
                    ((packed_hi_1 >> (7 - pulse_bit + 56)) & 0x01) << 15;
 
                // Write to buffer instead of directly to output
                writeBuffer[writeIdx++] = outputWord_0 & 0xFF;
                writeBuffer[writeIdx++] = (outputWord_0 >> 8) & 0xFF;
                writeBuffer[writeIdx++] = outputWord_1 & 0xFF;
                writeBuffer[writeIdx++] = (outputWord_1 >> 8) & 0xFF;
            }
 
            // Flush buffer when full (16 words = 32 bytes)
            // This triggers efficient burst writes that align with cache lines
            if (writeIdx == 32) {
                fl::memcpy(&outputBuffer[outputIdx], writeBuffer, 32);
                outputIdx += 32;
                writeIdx = 0;
            }
        }
    } else {
        // Invalid DATA_WIDTH (compile-time error if template instantiated with wrong value)
        return 0;
    }
 
    return outputIdx;
}

FASTLED_FORCE_INLINE FL_IRAM size_t transpose_wave8byte_parlio(
    const u8* FL_RESTRICT_PARAM laneWaveforms,
    size_t data_width,
    u8* FL_RESTRICT_PARAM outputBuffer
) FL_NOEXCEPT {
    // Dispatch to template specialization based on runtime data_width
    // Compiler generates optimized code for each specialization (no runtime branching)
    switch (data_width) {
        case 1:
            return transpose_wave8byte_parlio_template<1>(laneWaveforms, outputBuffer);
        case 2:
            return transpose_wave8byte_parlio_template<2>(laneWaveforms, outputBuffer);
        case 4:
            return transpose_wave8byte_parlio_template<4>(laneWaveforms, outputBuffer);
        case 8:
            return transpose_wave8byte_parlio_template<8>(laneWaveforms, outputBuffer);
        case 16:
            return transpose_wave8byte_parlio_template<16>(laneWaveforms, outputBuffer);
        default:
            // Invalid data_width
            return 0;
    }
}
 
}  // namespace fl
 
FL_OPTIMIZATION_LEVEL_O3_END