#include "types.comp"

layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ4_0 {
   block_q4_0_packed16 block;
};

float16_t dequantFuncQ4_0(const in decodeBufQ4_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    const float16_t d = bl.block.d;
    const uint idx = coordInBlock[1];
    const uint shift = (idx & 0x10) >> 2;
    uint32_t qs = uint32_t(bl.block.qs[(idx & 0xE) >> 1]);
    qs >>= shift;
    qs &= 0x0F0F;
    qs = unpack8(qs)[idx & 1];
    float16_t ret = (float16_t(qs) - float16_t(8)) * d;
    return ret;
}

layout(buffer_reference, std430, buffer_reference_align = 4) buffer decodeBufQ4_1 {
   block_q4_1 block;
};

float16_t dequantFuncQ4_1(const in decodeBufQ4_1 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    const float16_t d = bl.block.d;
    const float16_t m = bl.block.m;
    const uint idx = coordInBlock[1];
    const uint iqs = idx & 0xF;
    const uint shift = (idx & 0x10) >> 2;
    uint32_t qs = bl.block.qs[iqs];
    qs >>= shift;
    qs &= 0xF;
    float16_t ret = float16_t(qs) * d + m;
    return ret;
}

layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ5_0 {
   block_q5_0 block;
};

float16_t dequantFuncQ5_0(const in decodeBufQ5_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    const float16_t d = bl.block.d;
    const uint idx = coordInBlock[1];
    const uint iqs = idx & 0xF;

    const uint uint_qh = uint(bl.block.qh[1]) << 16 | bl.block.qh[0];
    const uint qh = ((uint_qh >> idx) << 4) & 0x10;

    const uint shift = (idx & 0x10) >> 2;
    uint32_t qs = bl.block.qs[iqs];
    qs >>= shift;
    qs &= 0xF;

    float16_t ret = (float16_t(qs | qh) - float16_t(16)) * d;
    return ret;
}

layout(buffer_reference, std430, buffer_reference_align = 8) buffer decodeBufQ5_1 {
   block_q5_1 block;
};

float16_t dequantFuncQ5_1(const in decodeBufQ5_1 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    const float16_t d = bl.block.d;
    const float16_t m = bl.block.m;
    const uint idx = coordInBlock[1];
    const uint iqs = idx & 0xF;

    const uint uint_qh = bl.block.qh;
    const uint qh = ((uint_qh >> idx) << 4) & 0x10;

    const uint shift = (idx & 0x10) >> 2;
    uint32_t qs = bl.block.qs[iqs];
    qs >>= shift;
    qs &= 0xF;

    float16_t ret = float16_t(qs | qh) * d + m;
    return ret;
}

layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ8_0 {
   block_q8_0_packed16 block;
};

float16_t dequantFuncQ8_0(const in decodeBufQ8_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    const float16_t d = bl.block.d;
    const uint idx = coordInBlock[1];
    const uint iqs = idx;

    // Load 16b and select the byte for this element
    int32_t qs = unpack8(int32_t(bl.block.qs[(iqs & 0x1E) >> 1]))[iqs & 1];
    float16_t ret = float16_t(qs) * d;
    return ret;
}

layout(buffer_reference, std430, buffer_reference_align = 4) buffer decodeBufQ2_K {
   block_q2_K block;
};

layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ2_K_packed16 {
   block_q2_K_packed16 block;
};

float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    decodeBufQ2_K_packed16 bl16 = decodeBufQ2_K_packed16(bl);
    const f16vec2 d = bl.block.d;
    const uint idx = coordInBlock[1];

    const uint scalesi = (idx & 0xF0) >> 4;             // 0..15
    const uint qsshift = (idx & 0x60) >> 4;             // 0,2,4,6

    uint qs = uint32_t(bl16.block.qs[((idx & 0x80) >> 3) + ((idx & 0x1E) >> 1)]);
    qs = (qs >> qsshift) & 0x0303;
    qs = unpack8(qs)[idx & 1];

    const uint scales = bl.block.scales[scalesi];
    float16_t ret = d.x * float16_t(scales & 0xF) * float16_t(qs) - d.y * float16_t(scales >> 4);
    return ret;
}

layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ3_K {
   block_q3_K block;
};

float16_t dequantFuncQ3_K(const in decodeBufQ3_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    const uint idx = coordInBlock[1];
    const uint iqs = idx;

    const uint n = iqs / 128;                    // 0,1
    const uint qsi = n * 32 + (iqs % 32);        // 0..63
    const uint hmi =          (iqs % 32);        // 0..31
    const uint j = (iqs % 128) / 8;              // 0..15
    const uint is = iqs / 16;                    // 0..15
    const uint halfsplit = ((iqs % 128) / 32);   // 0,1,2,3
    const uint qsshift = halfsplit * 2;          // 0,2,4,6
    const uint m = 1 << (4 * n + halfsplit);     // 1,2,4,8,16,32,64,128

    uint32_t scaleidx0 = (is < 8) ? is : (is-8);
    uint32_t scaleidx0shift = (is < 8) ? 0 : 4;
    uint32_t scaleidx1 = is + 8 - (is/4)*4;
    uint32_t scaleidx1shift = (is/4)*2;

    const int8_t us = int8_t(((bl.block.scales[scaleidx0] >> scaleidx0shift) & 0xF) | (((bl.block.scales[scaleidx1] >> scaleidx1shift) & 3) << 4));

    const float16_t dl = bl.block.d * float16_t(us - 32);

    float16_t ret = dl * float16_t(int8_t((bl.block.qs[qsi    ] >> qsshift) & 3) - (((bl.block.hmask[hmi    ] & m) != 0) ? 0 : 4));

    return ret;
}

layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K {
   block_q4_K block;
};

layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K_packed16 {
   block_q4_K_packed16 block;
};

layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K_packed128 {
   block_q4_K_packed128 block;
};

float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    decodeBufQ4_K_packed16 bl16 = decodeBufQ4_K_packed16(bl);
    decodeBufQ4_K_packed128 bl128 = decodeBufQ4_K_packed128(bl);
    const uint idx = coordInBlock[1];

    const uint b = (idx & 0x20) >> 5;            // 0,1
    const uint is = (idx & 0xE0) >> 5;         // 0..7

    uvec4 v = bl128.block.q4k[0];

    const f16vec2 loadd = unpackFloat2x16(v.x);

    uint32_t sc;
    uint32_t mbyte;

    uint32_t scale0 = v.y;
    uint32_t scale4 = v.z;
    uint32_t scale8 = v.w;

    uint32_t sc_lo = scale0;
    uint32_t mb_lo = scale4;
    uint32_t sc_hi = (scale8 & 0x0F0F0F0F) | ((scale0 & 0xC0C0C0C0) >> 2);
    uint32_t mb_hi = ((scale8 & 0xF0F0F0F0) >> 4) | ((scale4 & 0xC0C0C0C0) >> 2);

    sc = is < 4 ? sc_lo : sc_hi;
    mbyte = is < 4 ? mb_lo : mb_hi;
    sc = sc >> (8 * (is & 3));
    mbyte = mbyte >> (8 * (is & 3));
    sc &= 0x3F;
    mbyte &= 0x3F;

    const float16_t d = loadd.x * float16_t(sc);
    const float16_t m = loadd.y * float16_t(mbyte);

    uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]);
    qs = (qs >> (b * 4 + 8 * (idx & 1))) & 0xF;

    float16_t ret = d * float16_t(qs) - m;

    return ret;
}

layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K {
   block_q5_K block;
};

layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K_packed16 {
   block_q5_K_packed16 block;
};

layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K_packed128 {
   block_q5_K_packed128 block;
};

float16_t dequantFuncQ5_K(const in decodeBufQ5_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    decodeBufQ5_K_packed16 bl16 = decodeBufQ5_K_packed16(bl);
    decodeBufQ5_K_packed128 bl128 = decodeBufQ5_K_packed128(bl);
    const uint idx = coordInBlock[1];

    const uint b = (idx & 0x20) >> 5;          // 0,1
    const uint is = (idx & 0xE0) >> 5;         // 0..7

    uvec4 v = bl128.block.q5k[0];

    const f16vec2 loadd = unpackFloat2x16(v.x);

    uint32_t sc;
    uint32_t mbyte;

    uint32_t scale0 = v.y;
    uint32_t scale4 = v.z;
    uint32_t scale8 = v.w;

    uint32_t sc_lo = scale0;
    uint32_t mb_lo = scale4;
    uint32_t sc_hi = (scale8 & 0x0F0F0F0F) | ((scale0 & 0xC0C0C0C0) >> 2);
    uint32_t mb_hi = ((scale8 & 0xF0F0F0F0) >> 4) | ((scale4 & 0xC0C0C0C0) >> 2);

    sc = is < 4 ? sc_lo : sc_hi;
    mbyte = is < 4 ? mb_lo : mb_hi;
    sc = sc >> (8 * (is & 3));
    mbyte = mbyte >> (8 * (is & 3));
    sc &= 0x3F;
    mbyte &= 0x3F;

    const float16_t d = loadd.x * float16_t(sc);
    const float16_t m = loadd.y * float16_t(mbyte);

    uint qh = uint32_t(bl16.block.qh[(idx & 0x1E) >> 1]);
    qh = ((qh >> is) & 0x101) << 4;

    uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]);
    qs = (qs >> (b * 4)) & 0x0F0F;
    qs = unpack8(qs | qh)[idx & 1];

    float16_t ret = d * (float16_t(qs)) - m;

    return ret;
}

layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ6_K {
   block_q6_K block;
};

layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ6_K_packed16 {
   block_q6_K_packed16 block;
};

float16_t dequantFuncQ6_K(const in decodeBufQ6_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    decodeBufQ6_K_packed16 bl16 = decodeBufQ6_K_packed16(bl);
    const uint idx = coordInBlock[1];

    const uint b = (idx & 0x40) >> 6;           // 0,1
    const uint qhshift = (idx & 0x60) >> 4;    // 0,2,4,6
    const uint is = (idx & 0xF0) >> 4;          // 0..15

    const float16_t dscale = bl.block.d * float16_t(bl.block.scales[is]);

    uint ql = uint32_t(bl16.block.ql[((idx & 0x80) >> 2) + ((idx & 0x3E) >> 1)]);
    ql = (ql >> (b * 4)) & 0x0F0F;

    uint qh = uint32_t(bl16.block.qh[((idx & 0x80) >> 3) + ((idx & 0x1E) >> 1)]);
    qh = ((qh >> qhshift) & 0x0303) << 4;

    int q = unpack8(ql | qh)[idx & 1];

    float16_t ret = dscale * float16_t(q - 32);

    return ret;
}

#if defined(DATA_A_IQ2_XXS)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XXS {
   block_iq2_xxs block;
};

layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XXS_packed16 {
   block_iq2_xxs_packed16 block;
};

float16_t dequantFuncIQ2_XXS(const in decodeBufIQ2_XXS bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    decodeBufIQ2_XXS_packed16 bl16 = decodeBufIQ2_XXS_packed16(bl);
    const float16_t d = bl.block.d;
    const uint idx = coordInBlock[1];

    const uint ib32 = (idx & 0xE0) >> 5; // 0..7
    const uint ib8 = (idx & 0x18) >> 3;  // 0..3
    const uint iqs = 8 * ib32 + ib8;

    const uint8_t qs = bl.block.qs[iqs];
    const uint signscale = pack32(u16vec2(bl16.block.qs[4*ib32+2], bl16.block.qs[4*ib32+3]));

    const float16_t dscale = bl.block.d * 0.25hf * (0.5hf + float16_t(signscale >> 28));
    uint sign = bitfieldExtract(signscale, 7 * int(ib8), 7);
    sign |= bitCount(sign) << 7;

    const uint8_t g = unpack8(iq2xxs_grid[qs][(idx & 4) >> 2])[idx & 3];

    float16_t ret = dscale * float16_t(g) * ((sign & (1 << (idx & 7))) != 0 ? -1.0hf : 1.0hf);

    return ret;
}
#endif

#if defined(DATA_A_IQ2_XS)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XS {
   block_iq2_xs block;
};

float16_t dequantFuncIQ2_XS(const in decodeBufIQ2_XS bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    const float16_t d = bl.block.d;
    const uint idx = coordInBlock[1];

    const uint is = (idx & 0xE0) >> 5;     // 0..8
    const uint sshift = (idx & 0x10) >> 2; // 0,4
    const uint iqs = (idx & 0xF8) >> 3;    // 0..63

    const uint16_t qs = bl.block.qs[iqs];
    const float16_t dscale = bl.block.d * 0.25hf * (0.5hf + float16_t((bl.block.scales[is] >> sshift) & 0xF));

    uint sign = uint(qs >> 9);
    sign |= bitCount(sign) << 7;
    const uint8_t g = unpack8(iq2xs_grid[qs & 0x1FF][(idx & 4) >> 2])[idx & 3];

    float16_t ret = dscale * float16_t(g) * ((sign & (1 << (idx & 7))) != 0 ? -1.0hf : 1.0hf);
    return ret;
}
#endif

#if defined(DATA_A_IQ2_S)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_S {
   block_iq2_s block;
};

float16_t dequantFuncIQ2_S(const in decodeBufIQ2_S bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    uint idx = coordInBlock[1];
    uint lsb = idx & 1;
    idx /= 2;

    const uint ib8 = (idx % 128) / 4; // 0..31
    const uint ib32 = ib8 / 4;        // 0..7

    const uint scale = (bl.block.scales[ib32] >> (2 * (ib8 & 2))) & 0xf;
    const uint qs = bl.block.qs[ib8];
    const uint qh = bl.block.qh[ib32];
    const uint qhshift = 2 * (ib8 % 4);
    const uint sign = bl.block.qs[QUANT_K / 8 + ib8] >> (2 * (idx % 4));

    const float d = float(bl.block.d);
    const float db = d * 0.25 * (0.5 + scale);
    const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
    const uint16_t grid = unpack16(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 2) >> 1])[idx & 1];
    const vec2 v = db * vec2(sign01) * vec2(unpack8(grid));
    return float16_t(v[lsb]);
}
#endif

#if defined(DATA_A_IQ3_XXS)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3_XXS {
   block_iq3_xxs block;
};

layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3_XXS_packed16 {
   block_iq3_xxs_packed16 block;
};

float16_t dequantFuncIQ3_XXS(const in decodeBufIQ3_XXS bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    uint idx = coordInBlock[1];
    uint lsb = idx & 1;
    idx /= 2;

    const uint iqs = (idx % 128) / 2;           // 0..63
    const uint is = QUANT_K / 4 + 4 * (iqs / 8); // 8 values

    const float d = float(bl.block.d);
    const uint qs = bl.block.qs[iqs];
    const uint signs = pack32(u8vec4(
        bl.block.qs[is+0],
        bl.block.qs[is+1],
        bl.block.qs[is+2],
        bl.block.qs[is+3]
    ));
    const float db = d * 0.5 * (0.5 + (signs >> 28));
    const uint32_t sign7 = bitfieldExtract(signs, 7 * (int(iqs / 2) % 4), 7);
    const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
    const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
    const uint grid = iq3xxs_grid[qs] >> (16 * (idx & 1));
    const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
    return float16_t(v[lsb]);
}
#endif

#if defined(DATA_A_IQ3_S)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3_S {
   block_iq3_s block;
};

float16_t dequantFuncIQ3_S(const in decodeBufIQ3_S bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    uint idx = coordInBlock[1];
    uint lsb = idx & 1;
    idx /= 2;

    const uint iqs = (idx % 128) / 2;           // 0..63
    const uint iqh = iqs / 8;

    const float d = float(bl.block.d);
    const uint qs = bl.block.qs[iqs];
    const uint qh = bl.block.qh[iqh];
    const int8_t sign = int8_t(bl.block.signs[iqs / 2] >> (2 * (idx % 4)));
    const uint scale = bl.block.scales[iqs / 16];
    const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(sign << 1, sign)));
    const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf));
    const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> (16 * (idx % 2));
    const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);

    return float16_t(v[lsb]);
}
#endif


#if defined(DATA_A_IQ4_NL)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ4_NL {
   block_iq4_nl block;
};

float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
    const float16_t d = bl.block.d;
    const uint idx = coordInBlock[1];
    const uint iqs = idx & 0xF;
    const uint shift = (idx & 0x10) >> 2;
    uint32_t qs = bl.block.qs[iqs];
    qs >>= shift;
    qs &= 0xF;
    float16_t ret = float16_t(kvalues_iq4nl[qs]) * d;
    return ret;
}
#endif

#if defined(DATA_A_Q4_0)
#define dequantFuncA dequantFuncQ4_0
#elif defined(DATA_A_Q4_1)
#define dequantFuncA dequantFuncQ4_1
#elif defined(DATA_A_Q5_0)
#define dequantFuncA dequantFuncQ5_0
#elif defined(DATA_A_Q5_1)
#define dequantFuncA dequantFuncQ5_1
#elif defined(DATA_A_Q8_0)
#define dequantFuncA dequantFuncQ8_0
#elif defined(DATA_A_Q2_K)
#define dequantFuncA dequantFuncQ2_K
#elif defined(DATA_A_Q3_K)
#define dequantFuncA dequantFuncQ3_K
#elif defined(DATA_A_Q4_K)
#define dequantFuncA dequantFuncQ4_K
#elif defined(DATA_A_Q5_K)
#define dequantFuncA dequantFuncQ5_K
#elif defined(DATA_A_Q6_K)
#define dequantFuncA dequantFuncQ6_K
#elif defined(DATA_A_IQ2_XXS)
#define dequantFuncA dequantFuncIQ2_XXS
#elif defined(DATA_A_IQ2_XS)
#define dequantFuncA dequantFuncIQ2_XS
#elif defined(DATA_A_IQ2_S)
#define dequantFuncA dequantFuncIQ2_S
#elif defined(DATA_A_IQ3_XXS)
#define dequantFuncA dequantFuncIQ3_XXS
#elif defined(DATA_A_IQ3_S)
#define dequantFuncA dequantFuncIQ3_S
#elif defined(DATA_A_IQ4_NL)
#define dequantFuncA dequantFuncIQ4_NL
#endif