suyu/src/video_core/host_shaders/astc_decoder.comp

// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.

#version 450

#ifdef VULKAN

#define BEGIN_PUSH_CONSTANTS layout(push_constant) uniform PushConstants {
#define END_PUSH_CONSTANTS };
#define UNIFORM(n)
#define BINDING_INPUT_BUFFER 0
#define BINDING_SWIZZLE_BUFFER 1
#define BINDING_OUTPUT_IMAGE 2

#else // ^^^ Vulkan ^^^ // vvv OpenGL vvv

#define BEGIN_PUSH_CONSTANTS
#define END_PUSH_CONSTANTS
#define UNIFORM(n) layout(location = n) uniform
#define BINDING_INPUT_BUFFER 0
#define BINDING_SWIZZLE_BUFFER 1
#define BINDING_OUTPUT_IMAGE 0

#endif

layout(local_size_x = 32, local_size_y = 32, local_size_z = 1) in;

BEGIN_PUSH_CONSTANTS
UNIFORM(1) uvec2 block_dims;

UNIFORM(2) uint bytes_per_block_log2;
UNIFORM(3) uint layer_stride;
UNIFORM(4) uint block_size;
UNIFORM(5) uint x_shift;
UNIFORM(6) uint block_height;
UNIFORM(7) uint block_height_mask;
END_PUSH_CONSTANTS

struct EncodingData {
    uint encoding;
    uint num_bits;
    uint bit_value;
    uint quint_trit_value;
};

struct TexelWeightParams {
    uvec2 size;
    uint max_weight;
    bool dual_plane;
    bool error_state;
    bool void_extent_ldr;
    bool void_extent_hdr;
};

// Swizzle data
layout(binding = BINDING_SWIZZLE_BUFFER, std430) readonly buffer SwizzleTable {
    uint swizzle_table[];
};

layout(binding = BINDING_INPUT_BUFFER, std430) readonly buffer InputBufferU32 {
    uint astc_data[];
};

layout(binding = BINDING_OUTPUT_IMAGE, rgba8) uniform writeonly image2DArray dest_image;

const uint GOB_SIZE_X = 64;
const uint GOB_SIZE_Y = 8;
const uint GOB_SIZE_Z = 1;
const uint GOB_SIZE = GOB_SIZE_X * GOB_SIZE_Y * GOB_SIZE_Z;

const uint GOB_SIZE_X_SHIFT = 6;
const uint GOB_SIZE_Y_SHIFT = 3;
const uint GOB_SIZE_Z_SHIFT = 0;
const uint GOB_SIZE_SHIFT = GOB_SIZE_X_SHIFT + GOB_SIZE_Y_SHIFT + GOB_SIZE_Z_SHIFT;

const uvec2 SWIZZLE_MASK = uvec2(GOB_SIZE_X - 1, GOB_SIZE_Y - 1);

const int BLOCK_SIZE_IN_BYTES = 16;

const int BLOCK_INFO_ERROR = 0;
const int BLOCK_INFO_VOID_EXTENT_HDR = 1;
const int BLOCK_INFO_VOID_EXTENT_LDR = 2;
const int BLOCK_INFO_NORMAL = 3;

const int JUST_BITS = 0;
const int QUINT = 1;
const int TRIT = 2;

// ASTC Encodings data, sorted in ascending order based on their BitLength value
// (see GetBitLength() function)
EncodingData encoding_values[22] = EncodingData[](
    EncodingData(JUST_BITS, 0, 0, 0), EncodingData(JUST_BITS, 1, 0, 0), EncodingData(TRIT, 0, 0, 0),
    EncodingData(JUST_BITS, 2, 0, 0), EncodingData(QUINT, 0, 0, 0), EncodingData(TRIT, 1, 0, 0),
    EncodingData(JUST_BITS, 3, 0, 0), EncodingData(QUINT, 1, 0, 0), EncodingData(TRIT, 2, 0, 0),
    EncodingData(JUST_BITS, 4, 0, 0), EncodingData(QUINT, 2, 0, 0), EncodingData(TRIT, 3, 0, 0),
    EncodingData(JUST_BITS, 5, 0, 0), EncodingData(QUINT, 3, 0, 0), EncodingData(TRIT, 4, 0, 0),
    EncodingData(JUST_BITS, 6, 0, 0), EncodingData(QUINT, 4, 0, 0), EncodingData(TRIT, 5, 0, 0),
    EncodingData(JUST_BITS, 7, 0, 0), EncodingData(QUINT, 5, 0, 0), EncodingData(TRIT, 6, 0, 0),
    EncodingData(JUST_BITS, 8, 0, 0)
);

// The following constants are expanded variants of the Replicate()
// function calls corresponding to the following arguments:
// value: index into the generated table
// num_bits: the after "REPLICATE" in the table name. i.e. 4 is num_bits in REPLICATE_4.
// to_bit: the integer after "TO_"
const uint REPLICATE_BIT_TO_7_TABLE[2] = uint[](0, 127);
const uint REPLICATE_1_BIT_TO_9_TABLE[2] = uint[](0, 511);

const uint REPLICATE_1_BIT_TO_8_TABLE[2] = uint[](0, 255);
const uint REPLICATE_2_BIT_TO_8_TABLE[4] = uint[](0, 85, 170, 255);
const uint REPLICATE_3_BIT_TO_8_TABLE[8] = uint[](0, 36, 73, 109, 146, 182, 219, 255);
const uint REPLICATE_4_BIT_TO_8_TABLE[16] =
    uint[](0, 17, 34, 51, 68, 85, 102, 119, 136, 153, 170, 187, 204, 221, 238, 255);
const uint REPLICATE_5_BIT_TO_8_TABLE[32] =
    uint[](0, 8, 16, 24, 33, 41, 49, 57, 66, 74, 82, 90, 99, 107, 115, 123, 132, 140, 148, 156, 165,
           173, 181, 189, 198, 206, 214, 222, 231, 239, 247, 255);
const uint REPLICATE_1_BIT_TO_6_TABLE[2] = uint[](0, 63);
const uint REPLICATE_2_BIT_TO_6_TABLE[4] = uint[](0, 21, 42, 63);
const uint REPLICATE_3_BIT_TO_6_TABLE[8] = uint[](0, 9, 18, 27, 36, 45, 54, 63);
const uint REPLICATE_4_BIT_TO_6_TABLE[16] =
    uint[](0, 4, 8, 12, 17, 21, 25, 29, 34, 38, 42, 46, 51, 55, 59, 63);
const uint REPLICATE_5_BIT_TO_6_TABLE[32] =
    uint[](0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30, 33, 35, 37, 39, 41, 43, 45,
           47, 49, 51, 53, 55, 57, 59, 61, 63);
const uint REPLICATE_6_BIT_TO_8_TABLE[64] =
    uint[](0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 52, 56, 60, 65, 69, 73, 77, 81, 85, 89,
           93, 97, 101, 105, 109, 113, 117, 121, 125, 130, 134, 138, 142, 146, 150, 154, 158, 162,
           166, 170, 174, 178, 182, 186, 190, 195, 199, 203, 207, 211, 215, 219, 223, 227, 231, 235,
           239, 243, 247, 251, 255);
const uint REPLICATE_7_BIT_TO_8_TABLE[128] =
    uint[](0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30, 32, 34, 36, 38, 40, 42, 44,
           46, 48, 50, 52, 54, 56, 58, 60, 62, 64, 66, 68, 70, 72, 74, 76, 78, 80, 82, 84, 86, 88,
           90, 92, 94, 96, 98, 100, 102, 104, 106, 108, 110, 112, 114, 116, 118, 120, 122, 124, 126,
           129, 131, 133, 135, 137, 139, 141, 143, 145, 147, 149, 151, 153, 155, 157, 159, 161, 163,
           165, 167, 169, 171, 173, 175, 177, 179, 181, 183, 185, 187, 189, 191, 193, 195, 197, 199,
           201, 203, 205, 207, 209, 211, 213, 215, 217, 219, 221, 223, 225, 227, 229, 231, 233, 235,
           237, 239, 241, 243, 245, 247, 249, 251, 253, 255);

// Input ASTC texture globals
uint current_index = 0;
int bitsread = 0;
uint total_bitsread = 0;
uint local_buff[16];

// Color data globals
uint color_endpoint_data[16];
int color_bitsread = 0;
uint total_color_bitsread = 0;
int color_index = 0;

// Four values, two endpoints, four maximum paritions
uint color_values[32];
int colvals_index = 0;

// Weight data globals
uint texel_weight_data[16];
int texel_bitsread = 0;
uint total_texel_bitsread = 0;
int texel_index = 0;

bool texel_flag = false;

// Global "vectors" to be pushed into when decoding
EncodingData result_vector[100];
int result_index = 0;

EncodingData texel_vector[100];
int texel_vector_index = 0;

uint unquantized_texel_weights[2][144];

uint SwizzleOffset(uvec2 pos) {
    pos = pos & SWIZZLE_MASK;
    return swizzle_table[pos.y * 64 + pos.x];
}

uint ReadTexel(uint offset) {
    // extract the 8-bit value from the 32-bit packed data.
    return bitfieldExtract(astc_data[offset / 4], int((offset * 8) & 24), 8);
}

// Replicates low num_bits such that [(to_bit - 1):(to_bit - 1 - from_bit)]
// is the same as [(num_bits - 1):0] and repeats all the way down.
uint Replicate(uint val, uint num_bits, uint to_bit) {
    if (num_bits == 0 || to_bit == 0) {
        return 0;
    }
    const uint v = val & uint((1 << num_bits) - 1);
    uint res = v;
    uint reslen = num_bits;
    while (reslen < to_bit) {
        uint comp = 0;
        if (num_bits > to_bit - reslen) {
            uint newshift = to_bit - reslen;
            comp = num_bits - newshift;
            num_bits = newshift;
        }
        res = uint(res << num_bits);
        res = uint(res | (v >> comp));
        reslen += num_bits;
    }
    return res;
}

uvec4 ReplicateByteTo16(uvec4 value) {
    return value * 0x101;
}

uint ReplicateBitTo7(uint value) {
    return REPLICATE_BIT_TO_7_TABLE[value];
}

uint ReplicateBitTo9(uint value) {
    return REPLICATE_1_BIT_TO_9_TABLE[value];
}

uint FastReplicateTo8(uint value, uint num_bits) {
    switch (num_bits) {
    case 1:
        return REPLICATE_1_BIT_TO_8_TABLE[value];
    case 2:
        return REPLICATE_2_BIT_TO_8_TABLE[value];
    case 3:
        return REPLICATE_3_BIT_TO_8_TABLE[value];
    case 4:
        return REPLICATE_4_BIT_TO_8_TABLE[value];
    case 5:
        return REPLICATE_5_BIT_TO_8_TABLE[value];
    case 6:
        return REPLICATE_6_BIT_TO_8_TABLE[value];
    case 7:
        return REPLICATE_7_BIT_TO_8_TABLE[value];
    case 8:
        return value;
    }
    return Replicate(value, num_bits, 8);
}

uint FastReplicateTo6(uint value, uint num_bits) {
    switch (num_bits) {
    case 1:
        return REPLICATE_1_BIT_TO_6_TABLE[value];
    case 2:
        return REPLICATE_2_BIT_TO_6_TABLE[value];
    case 3:
        return REPLICATE_3_BIT_TO_6_TABLE[value];
    case 4:
        return REPLICATE_4_BIT_TO_6_TABLE[value];
    case 5:
        return REPLICATE_5_BIT_TO_6_TABLE[value];
    }
    return Replicate(value, num_bits, 6);
}

uint Div3Floor(uint v) {
    return (v * 0x5556) >> 16;
}

uint Div3Ceil(uint v) {
    return Div3Floor(v + 2);
}

uint Div5Floor(uint v) {
    return (v * 0x3334) >> 16;
}

uint Div5Ceil(uint v) {
    return Div5Floor(v + 4);
}

uint Hash52(uint p) {
    p ^= p >> 15;
    p -= p << 17;
    p += p << 7;
    p += p << 4;
    p ^= p >> 5;
    p += p << 16;
    p ^= p >> 7;
    p ^= p >> 3;
    p ^= p << 6;
    p ^= p >> 17;
    return p;
}

uint Select2DPartition(uint seed, uint x, uint y, uint partition_count, bool small_block) {
    if (small_block) {
        x <<= 1;
        y <<= 1;
    }

    seed += (partition_count - 1) * 1024;

    uint rnum = Hash52(uint(seed));
    uint seed1 = uint(rnum & 0xF);
    uint seed2 = uint((rnum >> 4) & 0xF);
    uint seed3 = uint((rnum >> 8) & 0xF);
    uint seed4 = uint((rnum >> 12) & 0xF);
    uint seed5 = uint((rnum >> 16) & 0xF);
    uint seed6 = uint((rnum >> 20) & 0xF);
    uint seed7 = uint((rnum >> 24) & 0xF);
    uint seed8 = uint((rnum >> 28) & 0xF);

    seed1 = (seed1 * seed1);
    seed2 = (seed2 * seed2);
    seed3 = (seed3 * seed3);
    seed4 = (seed4 * seed4);
    seed5 = (seed5 * seed5);
    seed6 = (seed6 * seed6);
    seed7 = (seed7 * seed7);
    seed8 = (seed8 * seed8);

    uint sh1, sh2;
    if ((seed & 1) > 0) {
        sh1 = (seed & 2) > 0 ? 4 : 5;
        sh2 = (partition_count == 3) ? 6 : 5;
    } else {
        sh1 = (partition_count == 3) ? 6 : 5;
        sh2 = (seed & 2) > 0 ? 4 : 5;
    }
    seed1 >>= sh1;
    seed2 >>= sh2;
    seed3 >>= sh1;
    seed4 >>= sh2;
    seed5 >>= sh1;
    seed6 >>= sh2;
    seed7 >>= sh1;
    seed8 >>= sh2;

    uint a = seed1 * x + seed2 * y + (rnum >> 14);
    uint b = seed3 * x + seed4 * y + (rnum >> 10);
    uint c = seed5 * x + seed6 * y + (rnum >> 6);
    uint d = seed7 * x + seed8 * y + (rnum >> 2);

    a &= 0x3F;
    b &= 0x3F;
    c &= 0x3F;
    d &= 0x3F;

    if (partition_count < 4) {
        d = 0;
    }
    if (partition_count < 3) {
        c = 0;
    }

    if (a >= b && a >= c && a >= d) {
        return 0;
    } else if (b >= c && b >= d) {
        return 1;
    } else if (c >= d) {
        return 2;
    } else {
        return 3;
    }
}

uint ReadBit() {
    if (current_index >= local_buff.length()) {
        return 0;
    }
    uint bit = bitfieldExtract(local_buff[current_index], bitsread, 1);
    ++bitsread;
    ++total_bitsread;
    if (bitsread == 8) {
        ++current_index;
        bitsread = 0;
    }
    return bit;
}

uint StreamBits(uint num_bits) {
    uint ret = 0;
    for (uint i = 0; i < num_bits; i++) {
        ret |= ((ReadBit() & 1) << i);
    }
    return ret;
}

uint ReadColorBit() {
    uint bit = 0;
    if (texel_flag) {
        bit = bitfieldExtract(texel_weight_data[texel_index], texel_bitsread, 1);
        ++texel_bitsread;
        ++total_texel_bitsread;
        if (texel_bitsread == 8) {
            ++texel_index;
            texel_bitsread = 0;
        }
    } else {
        bit = bitfieldExtract(color_endpoint_data[color_index], color_bitsread, 1);
        ++color_bitsread;
        ++total_color_bitsread;
        if (color_bitsread == 8) {
            ++color_index;
            color_bitsread = 0;
        }
    }
    return bit;
}

uint StreamColorBits(uint num_bits) {
    uint ret = 0;
    for (uint i = 0; i < num_bits; i++) {
        ret |= ((ReadColorBit() & 1) << i);
    }
    return ret;
}

void ResultEmplaceBack(EncodingData val) {
    if (texel_flag) {
        texel_vector[texel_vector_index] = val;
        ++texel_vector_index;
    } else {
        result_vector[result_index] = val;
        ++result_index;
    }
}

// Returns the number of bits required to encode n_vals values.
uint GetBitLength(uint n_vals, uint encoding_index) {
    uint total_bits = encoding_values[encoding_index].num_bits * n_vals;
    if (encoding_values[encoding_index].encoding == TRIT) {
        total_bits += Div5Ceil(n_vals * 8);
    } else if (encoding_values[encoding_index].encoding == QUINT) {
        total_bits += Div3Ceil(n_vals * 7);
    }
    return total_bits;
}

uint GetNumWeightValues(uvec2 size, bool dual_plane) {
    uint n_vals = size.x * size.y;
    if (dual_plane) {
        n_vals *= 2;
    }
    return n_vals;
}

uint GetPackedBitSize(uvec2 size, bool dual_plane, uint max_weight) {
    uint n_vals = GetNumWeightValues(size, dual_plane);
    return GetBitLength(n_vals, max_weight);
}

uint BitsBracket(uint bits, uint pos) {
    return ((bits >> pos) & 1);
}

uint BitsOp(uint bits, uint start, uint end) {
    if (start == end) {
        return BitsBracket(bits, start);
    } else if (start > end) {
        uint t = start;
        start = end;
        end = t;
    }

    uint mask = (1 << (end - start + 1)) - 1;
    return ((bits >> start) & mask);
}

void DecodeQuintBlock(uint num_bits) {
    uint m[3];
    uint q[3];
    uint Q;
    m[0] = StreamColorBits(num_bits);
    Q = StreamColorBits(3);
    m[1] = StreamColorBits(num_bits);
    Q |= StreamColorBits(2) << 3;
    m[2] = StreamColorBits(num_bits);
    Q |= StreamColorBits(2) << 5;
    if (BitsOp(Q, 1, 2) == 3 && BitsOp(Q, 5, 6) == 0) {
        q[0] = 4;
        q[1] = 4;
        q[2] = (BitsBracket(Q, 0) << 2) | ((BitsBracket(Q, 4) & ~BitsBracket(Q, 0)) << 1) |
               (BitsBracket(Q, 3) & ~BitsBracket(Q, 0));
    } else {
        uint C = 0;
        if (BitsOp(Q, 1, 2) == 3) {
            q[2] = 4;
            C = (BitsOp(Q, 3, 4) << 3) | ((~BitsOp(Q, 5, 6) & 3) << 1) | BitsBracket(Q, 0);
        } else {
            q[2] = BitsOp(Q, 5, 6);
            C = BitsOp(Q, 0, 4);
        }
        if (BitsOp(C, 0, 2) == 5) {
            q[1] = 4;
            q[0] = BitsOp(C, 3, 4);
        } else {
            q[1] = BitsOp(C, 3, 4);
            q[0] = BitsOp(C, 0, 2);
        }
    }
    for (uint i = 0; i < 3; i++) {
        EncodingData val;
        val.encoding = QUINT;
        val.num_bits = num_bits;
        val.bit_value = m[i];
        val.quint_trit_value = q[i];
        ResultEmplaceBack(val);
    }
}

void DecodeTritBlock(uint num_bits) {
    uint m[5];
    uint t[5];
    uint T;
    m[0] = StreamColorBits(num_bits);
    T = StreamColorBits(2);
    m[1] = StreamColorBits(num_bits);
    T |= StreamColorBits(2) << 2;
    m[2] = StreamColorBits(num_bits);
    T |= StreamColorBits(1) << 4;
    m[3] = StreamColorBits(num_bits);
    T |= StreamColorBits(2) << 5;
    m[4] = StreamColorBits(num_bits);
    T |= StreamColorBits(1) << 7;
    uint C = 0;
    if (BitsOp(T, 2, 4) == 7) {
        C = (BitsOp(T, 5, 7) << 2) | BitsOp(T, 0, 1);
        t[4] = 2;
        t[3] = 2;
    } else {
        C = BitsOp(T, 0, 4);
        if (BitsOp(T, 5, 6) == 3) {
            t[4] = 2;
            t[3] = BitsBracket(T, 7);
        } else {
            t[4] = BitsBracket(T, 7);
            t[3] = BitsOp(T, 5, 6);
        }
    }
    if (BitsOp(C, 0, 1) == 3) {
        t[2] = 2;
        t[1] = BitsBracket(C, 4);
        t[0] = (BitsBracket(C, 3) << 1) | (BitsBracket(C, 2) & ~BitsBracket(C, 3));
    } else if (BitsOp(C, 2, 3) == 3) {
        t[2] = 2;
        t[1] = 2;
        t[0] = BitsOp(C, 0, 1);
    } else {
        t[2] = BitsBracket(C, 4);
        t[1] = BitsOp(C, 2, 3);
        t[0] = (BitsBracket(C, 1) << 1) | (BitsBracket(C, 0) & ~BitsBracket(C, 1));
    }
    for (uint i = 0; i < 5; i++) {
        EncodingData val;
        val.encoding = TRIT;
        val.num_bits = num_bits;
        val.bit_value = m[i];
        val.quint_trit_value = t[i];
        ResultEmplaceBack(val);
    }
}

void DecodeIntegerSequence(uint max_range, uint num_values) {
    EncodingData val = encoding_values[max_range];
    uint vals_decoded = 0;
    while (vals_decoded < num_values) {
        switch (val.encoding) {
        case QUINT:
            DecodeQuintBlock(val.num_bits);
            vals_decoded += 3;
            break;
        case TRIT:
            DecodeTritBlock(val.num_bits);
            vals_decoded += 5;
            break;
        case JUST_BITS:
            val.bit_value = StreamColorBits(val.num_bits);
            ResultEmplaceBack(val);
            vals_decoded++;
            break;
        }
    }
}

void DecodeColorValues(uvec4 modes, uint num_partitions, uint color_data_bits) {
    uint num_values = 0;
    for (uint i = 0; i < num_partitions; i++) {
        num_values += ((modes[i] >> 2) + 1) << 1;
    }
    // Find the largest encoding that's within color_data_bits
    // TODO(ameerj): profile with binary search
    int range = 0;
    while (++range < encoding_values.length()) {
        uint bit_length = GetBitLength(num_values, range);
        if (bit_length > color_data_bits) {
            break;
        }
    }
    DecodeIntegerSequence(range - 1, num_values);
    uint out_index = 0;
    for (int itr = 0; itr < result_index; ++itr) {
        if (out_index >= num_values) {
            break;
        }
        EncodingData val = result_vector[itr];
        uint bitlen = val.num_bits;
        uint bitval = val.bit_value;
        uint A = 0, B = 0, C = 0, D = 0;
        A = ReplicateBitTo9((bitval & 1));
        switch (val.encoding) {
        case JUST_BITS:
            color_values[out_index++] = FastReplicateTo8(bitval, bitlen);
            break;
        case TRIT: {
            D = val.quint_trit_value;
            switch (bitlen) {
            case 1:
                C = 204;
                break;
            case 2: {
                C = 93;
                uint b = (bitval >> 1) & 1;
                B = (b << 8) | (b << 4) | (b << 2) | (b << 1);
                break;
            }
            case 3: {
                C = 44;
                uint cb = (bitval >> 1) & 3;
                B = (cb << 7) | (cb << 2) | cb;
                break;
            }
            case 4: {
                C = 22;
                uint dcb = (bitval >> 1) & 7;
                B = (dcb << 6) | dcb;
                break;
            }
            case 5: {
                C = 11;
                uint edcb = (bitval >> 1) & 0xF;
                B = (edcb << 5) | (edcb >> 2);
                break;
            }
            case 6: {
                C = 5;
                uint fedcb = (bitval >> 1) & 0x1F;
                B = (fedcb << 4) | (fedcb >> 4);
                break;
            }
            }
            break;
        }
        case QUINT: {
            D = val.quint_trit_value;
            switch (bitlen) {
            case 1:
                C = 113;
                break;
            case 2: {
                C = 54;
                uint b = (bitval >> 1) & 1;
                B = (b << 8) | (b << 3) | (b << 2);
                break;
            }
            case 3: {
                C = 26;
                uint cb = (bitval >> 1) & 3;
                B = (cb << 7) | (cb << 1) | (cb >> 1);
                break;
            }
            case 4: {
                C = 13;
                uint dcb = (bitval >> 1) & 7;
                B = (dcb << 6) | (dcb >> 1);
                break;
            }
            case 5: {
                C = 6;
                uint edcb = (bitval >> 1) & 0xF;
                B = (edcb << 5) | (edcb >> 3);
                break;
            }
            }
            break;
        }
        }
        if (val.encoding != JUST_BITS) {
            uint T = (D * C) + B;
            T ^= A;
            T = (A & 0x80) | (T >> 2);
            color_values[out_index++] = T;
        }
    }
}

ivec2 BitTransferSigned(int a, int b) {
    ivec2 transferred;
    transferred.y = b >> 1;
    transferred.y |= a & 0x80;
    transferred.x = a >> 1;
    transferred.x &= 0x3F;
    if ((transferred.x & 0x20) > 0) {
        transferred.x -= 0x40;
    }
    return transferred;
}

uvec4 ClampByte(ivec4 color) {
    for (uint i = 0; i < 4; ++i) {
        color[i] = (color[i] < 0) ? 0 : ((color[i] > 255) ? 255 : color[i]);
    }
    return uvec4(color);
}

ivec4 BlueContract(int a, int r, int g, int b) {
    return ivec4(a, (r + b) >> 1, (g + b) >> 1, b);
}

void ComputeEndpoints(out uvec4 ep1, out uvec4 ep2, uint color_endpoint_mode) {
#define READ_UINT_VALUES(N)                                                                        \
    uint v[N];                                                                                     \
    for (uint i = 0; i < N; i++) {                                                                 \
        v[i] = color_values[colvals_index++];                                                      \
    }

#define READ_INT_VALUES(N)                                                                         \
    int v[N];                                                                                      \
    for (uint i = 0; i < N; i++) {                                                                 \
        v[i] = int(color_values[colvals_index++]);                                                 \
    }

    switch (color_endpoint_mode) {
    case 0: {
        READ_UINT_VALUES(2)
        ep1 = uvec4(0xFF, v[0], v[0], v[0]);
        ep2 = uvec4(0xFF, v[1], v[1], v[1]);
        break;
    }
    case 1: {
        READ_UINT_VALUES(2)
        uint L0 = (v[0] >> 2) | (v[1] & 0xC0);
        uint L1 = min(L0 + (v[1] & 0x3F), 0xFFU);
        ep1 = uvec4(0xFF, L0, L0, L0);
        ep2 = uvec4(0xFF, L1, L1, L1);
        break;
    }
    case 4: {
        READ_UINT_VALUES(4)
        ep1 = uvec4(v[2], v[0], v[0], v[0]);
        ep2 = uvec4(v[3], v[1], v[1], v[1]);
        break;
    }
    case 5: {
        READ_INT_VALUES(4)
        ivec2 transferred = BitTransferSigned(v[1], v[0]);
        v[1] = transferred.x;
        v[0] = transferred.y;
        transferred = BitTransferSigned(v[3], v[2]);
        v[3] = transferred.x;
        v[2] = transferred.y;
        ep1 = ClampByte(ivec4(v[2], v[0], v[0], v[0]));
        ep2 = ClampByte(ivec4(v[2] + v[3], v[0] + v[1], v[0] + v[1], v[0] + v[1]));
        break;
    }
    case 6: {
        READ_UINT_VALUES(4)
        ep1 = uvec4(0xFF, (v[0] * v[3]) >> 8, (v[1] * v[3]) >> 8, (v[2] * v[3]) >> 8);
        ep2 = uvec4(0xFF, v[0], v[1], v[2]);
        break;
    }
    case 8: {
        READ_UINT_VALUES(6)
        if ((v[1] + v[3] + v[5]) >= (v[0] + v[2] + v[4])) {
            ep1 = uvec4(0xFF, v[0], v[2], v[4]);
            ep2 = uvec4(0xFF, v[1], v[3], v[5]);
        } else {
            ep1 = uvec4(BlueContract(0xFF, int(v[1]), int(v[3]), int(v[5])));
            ep2 = uvec4(BlueContract(0xFF, int(v[0]), int(v[2]), int(v[4])));
        }
        break;
    }
    case 9: {
        READ_INT_VALUES(6)
        ivec2 transferred = BitTransferSigned(v[1], v[0]);
        v[1] = transferred.x;
        v[0] = transferred.y;
        transferred = BitTransferSigned(v[3], v[2]);
        v[3] = transferred.x;
        v[2] = transferred.y;
        transferred = BitTransferSigned(v[5], v[4]);
        v[5] = transferred.x;
        v[4] = transferred.y;
        if ((v[1] + v[3] + v[5]) >= 0) {
            ep1 = ClampByte(ivec4(0xFF, v[0], v[2], v[4]));
            ep2 = ClampByte(ivec4(0xFF, v[0] + v[1], v[2] + v[3], v[4] + v[5]));
        } else {
            ep1 = ClampByte(BlueContract(0xFF, v[0] + v[1], v[2] + v[3], v[4] + v[5]));
            ep2 = ClampByte(BlueContract(0xFF, v[0], v[2], v[4]));
        }
        break;
    }
    case 10: {
        READ_UINT_VALUES(6)
        ep1 = uvec4(v[4], (v[0] * v[3]) >> 8, (v[1] * v[3]) >> 8, (v[2] * v[3]) >> 8);
        ep2 = uvec4(v[5], v[0], v[1], v[2]);
        break;
    }
    case 12: {
        READ_UINT_VALUES(8)
        if ((v[1] + v[3] + v[5]) >= (v[0] + v[2] + v[4])) {
            ep1 = uvec4(v[6], v[0], v[2], v[4]);
            ep2 = uvec4(v[7], v[1], v[3], v[5]);
        } else {
            ep1 = uvec4(BlueContract(int(v[7]), int(v[1]), int(v[3]), int(v[5])));
            ep2 = uvec4(BlueContract(int(v[6]), int(v[0]), int(v[2]), int(v[4])));
        }
        break;
    }
    case 13: {
        READ_INT_VALUES(8)
        ivec2 transferred = BitTransferSigned(v[1], v[0]);
        v[1] = transferred.x;
        v[0] = transferred.y;
        transferred = BitTransferSigned(v[3], v[2]);
        v[3] = transferred.x;
        v[2] = transferred.y;

        transferred = BitTransferSigned(v[5], v[4]);
        v[5] = transferred.x;
        v[4] = transferred.y;

        transferred = BitTransferSigned(v[7], v[6]);
        v[7] = transferred.x;
        v[6] = transferred.y;

        if ((v[1] + v[3] + v[5]) >= 0) {
            ep1 = ClampByte(ivec4(v[6], v[0], v[2], v[4]));
            ep2 = ClampByte(ivec4(v[7] + v[6], v[0] + v[1], v[2] + v[3], v[4] + v[5]));
        } else {
            ep1 = ClampByte(BlueContract(v[6] + v[7], v[0] + v[1], v[2] + v[3], v[4] + v[5]));
            ep2 = ClampByte(BlueContract(v[6], v[0], v[2], v[4]));
        }
        break;
    }
    default: {
        // HDR mode, or more likely a bug computing the color_endpoint_mode
        ep1 = uvec4(0xFF, 0xFF, 0, 0);
        ep2 = uvec4(0xFF, 0xFF, 0, 0);
        break;
    }
    }
#undef READ_UINT_VALUES
#undef READ_INT_VALUES
}

uint UnquantizeTexelWeight(EncodingData val) {
    uint bitval = val.bit_value;
    uint bitlen = val.num_bits;
    uint A = ReplicateBitTo7((bitval & 1));
    uint B = 0, C = 0, D = 0;
    uint result = 0;
    switch (val.encoding) {
    case JUST_BITS:
        result = FastReplicateTo6(bitval, bitlen);
        break;
    case TRIT: {
        D = val.quint_trit_value;
        switch (bitlen) {
        case 0: {
            uint results[3] = {0, 32, 63};
            result = results[D];
            break;
        }
        case 1: {
            C = 50;
            break;
        }
        case 2: {
            C = 23;
            uint b = (bitval >> 1) & 1;
            B = (b << 6) | (b << 2) | b;
            break;
        }
        case 3: {
            C = 11;
            uint cb = (bitval >> 1) & 3;
            B = (cb << 5) | cb;
            break;
        }
        default:
            break;
        }
        break;
    }
    case QUINT: {
        D = val.quint_trit_value;
        switch (bitlen) {
        case 0: {
            uint results[5] = {0, 16, 32, 47, 63};
            result = results[D];
            break;
        }
        case 1: {
            C = 28;
            break;
        }
        case 2: {
            C = 13;
            uint b = (bitval >> 1) & 1;
            B = (b << 6) | (b << 1);
            break;
        }
        }
        break;
    }
    }
    if (val.encoding != JUST_BITS && bitlen > 0) {
        result = D * C + B;
        result ^= A;
        result = (A & 0x20) | (result >> 2);
    }
    if (result > 32) {
        result += 1;
    }
    return result;
}

void UnquantizeTexelWeights(bool dual_plane, uvec2 size) {
    uint weight_idx = 0;
    uint unquantized[2][144];
    uint area = size.x * size.y;
    for (uint itr = 0; itr < texel_vector_index; itr++) {
        unquantized[0][weight_idx] = UnquantizeTexelWeight(texel_vector[itr]);
        if (dual_plane) {
            ++itr;
            unquantized[1][weight_idx] = UnquantizeTexelWeight(texel_vector[itr]);
            if (itr == texel_vector_index) {
                break;
            }
        }
        if (++weight_idx >= (area))
            break;
    }

    const uint Ds = uint((block_dims.x * 0.5f + 1024) / (block_dims.x - 1));
    const uint Dt = uint((block_dims.y * 0.5f + 1024) / (block_dims.y - 1));
    const uint k_plane_scale = dual_plane ? 2 : 1;
    for (uint plane = 0; plane < k_plane_scale; plane++) {
        for (uint t = 0; t < block_dims.y; t++) {
            for (uint s = 0; s < block_dims.x; s++) {
                uint cs = Ds * s;
                uint ct = Dt * t;
                uint gs = (cs * (size.x - 1) + 32) >> 6;
                uint gt = (ct * (size.y - 1) + 32) >> 6;
                uint js = gs >> 4;
                uint fs = gs & 0xF;
                uint jt = gt >> 4;
                uint ft = gt & 0x0F;
                uint w11 = (fs * ft + 8) >> 4;
                uint w10 = ft - w11;
                uint w01 = fs - w11;
                uint w00 = 16 - fs - ft + w11;
                uvec4 w = uvec4(w00, w01, w10, w11);
                uint v0 = jt * size.x + js;

                uvec4 p = uvec4(0);
                if (v0 < area) {
                    p.x = unquantized[plane][v0];
                }
                if ((v0 + 1) < (area)) {
                    p.y = unquantized[plane][v0 + 1];
                }
                if ((v0 + size.x) < (area)) {
                    p.z = unquantized[plane][(v0 + size.x)];
                }
                if ((v0 + size.x + 1) < (area)) {
                    p.w = unquantized[plane][(v0 + size.x + 1)];
                }
                unquantized_texel_weights[plane][t * block_dims.x + s] = (uint(dot(p, w)) + 8) >> 4;
            }
        }
    }
}

int FindLayout(uint mode) {
    if ((mode & 3) != 0) {
        if ((mode & 8) != 0) {
            if ((mode & 4) != 0) {
                if ((mode & 0x100) != 0) {
                    return 4;
                }
                return 3;
            }
            return 2;
        }
        if ((mode & 4) != 0) {
            return 1;
        }
        return 0;
    }
    if ((mode & 0x100) != 0) {
        if ((mode & 0x80) != 0) {
            if ((mode & 0x20) != 0) {
                return 8;
            }
            return 7;
        }
        return 9;
    }
    if ((mode & 0x80) != 0) {
        return 6;
    }
    return 5;
}

TexelWeightParams DecodeBlockInfo(uint block_index) {
    TexelWeightParams params = TexelWeightParams(uvec2(0), 0, false, false, false, false);
    uint mode = StreamBits(11);
    if ((mode & 0x1ff) == 0x1fc) {
        if ((mode & 0x200) != 0) {
            params.void_extent_hdr = true;
        } else {
            params.void_extent_ldr = true;
        }
        if ((mode & 0x400) == 0 || StreamBits(1) == 0) {
            params.error_state = true;
        }
        return params;
    }
    if ((mode & 0xf) == 0) {
        params.error_state = true;
        return params;
    }
    if ((mode & 3) == 0 && (mode & 0x1c0) == 0x1c0) {
        params.error_state = true;
        return params;
    }
    uint A, B;
    uint mode_layout = FindLayout(mode);
    switch (mode_layout) {
    case 0:
        A = (mode >> 5) & 0x3;
        B = (mode >> 7) & 0x3;
        params.size = uvec2(B + 4, A + 2);
        break;
    case 1:
        A = (mode >> 5) & 0x3;
        B = (mode >> 7) & 0x3;
        params.size = uvec2(B + 8, A + 2);
        break;
    case 2:
        A = (mode >> 5) & 0x3;
        B = (mode >> 7) & 0x3;
        params.size = uvec2(A + 2, B + 8);
        break;
    case 3:
        A = (mode >> 5) & 0x3;
        B = (mode >> 7) & 0x1;
        params.size = uvec2(A + 2, B + 6);
        break;
    case 4:
        A = (mode >> 5) & 0x3;
        B = (mode >> 7) & 0x1;
        params.size = uvec2(B + 2, A + 2);
        break;
    case 5:
        A = (mode >> 5) & 0x3;
        params.size = uvec2(12, A + 2);
        break;
    case 6:
        A = (mode >> 5) & 0x3;
        params.size = uvec2(A + 2, 12);
        break;
    case 7:
        params.size = uvec2(6, 10);
        break;
    case 8:
        params.size = uvec2(10, 6);
        break;
    case 9:
        A = (mode >> 5) & 0x3;
        B = (mode >> 9) & 0x3;
        params.size = uvec2(A + 6, B + 6);
        break;
    default:
        params.error_state = true;
        break;
    }
    params.dual_plane = (mode_layout != 9) && ((mode & 0x400) != 0);
    uint weight_index = (mode & 0x10) != 0 ? 1 : 0;
    if (mode_layout < 5) {
        weight_index |= (mode & 0x3) << 1;
    } else {
        weight_index |= (mode & 0xc) >> 1;
    }
    weight_index -= 2;
    if ((mode_layout != 9) && ((mode & 0x200) != 0)) {
        const int max_weights[6] = int[6](7, 8, 9, 10, 11, 12);
        params.max_weight = max_weights[weight_index];
    } else {
        const int max_weights[6] = int[6](1, 2, 3, 4, 5, 6);
        params.max_weight = max_weights[weight_index];
    }
    return params;
}

void FillError(ivec3 coord) {
    for (uint j = 0; j < block_dims.y; j++) {
        for (uint i = 0; i < block_dims.x; i++) {
            imageStore(dest_image, coord + ivec3(i, j, 0), vec4(1.0, 1.0, 0.0, 1.0));
        }
    }
}

void FillVoidExtentLDR(ivec3 coord) {
    StreamBits(52);
    uint r_u = StreamBits(16);
    uint g_u = StreamBits(16);
    uint b_u = StreamBits(16);
    uint a_u = StreamBits(16);
    float a = float(a_u) / 65535.0f;
    float r = float(r_u) / 65535.0f;
    float g = float(g_u) / 65535.0f;
    float b = float(b_u) / 65535.0f;
    for (uint j = 0; j < block_dims.y; j++) {
        for (uint i = 0; i < block_dims.x; i++) {
            imageStore(dest_image, coord + ivec3(i, j, 0), vec4(r, g, b, a));
        }
    }
}

void DecompressBlock(ivec3 coord, uint block_index) {
    TexelWeightParams params = DecodeBlockInfo(block_index);
    if (params.error_state) {
        FillError(coord);
        return;
    }
    if (params.void_extent_hdr) {
        FillError(coord);
        return;
    }
    if (params.void_extent_ldr) {
        FillVoidExtentLDR(coord);
        return;
    }
    if ((params.size.x > block_dims.x) || (params.size.y > block_dims.y)) {
        FillError(coord);
        return;
    }
    uint num_partitions = StreamBits(2) + 1;
    if (num_partitions > 4 || (num_partitions == 4 && params.dual_plane)) {
        FillError(coord);
        return;
    }
    int plane_index = -1;
    uint partition_index = 1;
    uvec4 color_endpoint_mode = uvec4(0);
    uint ced_pointer = 0;
    uint base_cem = 0;
    if (num_partitions == 1) {
        color_endpoint_mode.x = StreamBits(4);
        partition_index = 0;
    } else {
        partition_index = StreamBits(10);
        base_cem = StreamBits(6);
    }
    uint base_mode = base_cem & 3;
    uint weight_bits = GetPackedBitSize(params.size, params.dual_plane, params.max_weight);
    uint remaining_bits = 128 - weight_bits - total_bitsread;
    uint extra_cem_bits = 0;
    if (base_mode > 0) {
        switch (num_partitions) {
        case 2:
            extra_cem_bits += 2;
            break;
        case 3:
            extra_cem_bits += 5;
            break;
        case 4:
            extra_cem_bits += 8;
            break;
        default:
            return;
        }
    }
    remaining_bits -= extra_cem_bits;
    uint plane_selector_bits = 0;
    if (params.dual_plane) {
        plane_selector_bits = 2;
    }
    remaining_bits -= plane_selector_bits;
    if (remaining_bits > 128) {
        // Bad data, more remaining bits than 4 bytes
        // return early
        return;
    }
    // Read color data...
    uint color_data_bits = remaining_bits;
    while (remaining_bits > 0) {
        int nb = int(min(remaining_bits, 8U));
        uint b = StreamBits(nb);
        color_endpoint_data[ced_pointer] = uint(bitfieldExtract(b, 0, nb));
        ++ced_pointer;
        remaining_bits -= nb;
    }
    plane_index = int(StreamBits(plane_selector_bits));
    if (base_mode > 0) {
        uint extra_cem = StreamBits(extra_cem_bits);
        uint cem = (extra_cem << 6) | base_cem;
        cem >>= 2;
        uvec4 C = uvec4(0);
        for (uint i = 0; i < num_partitions; i++) {
            C[i] = (cem & 1);
            cem >>= 1;
        }
        uvec4 M = uvec4(0);
        for (uint i = 0; i < num_partitions; i++) {
            M[i] = cem & 3;
            cem >>= 2;
        }
        for (uint i = 0; i < num_partitions; i++) {
            color_endpoint_mode[i] = base_mode;
            if (C[i] == 0) {
                --color_endpoint_mode[i];
            }
            color_endpoint_mode[i] <<= 2;
            color_endpoint_mode[i] |= M[i];
        }
    } else if (num_partitions > 1) {
        uint cem = base_cem >> 2;
        for (uint i = 0; i < num_partitions; i++) {
            color_endpoint_mode[i] = cem;
        }
    }
    DecodeColorValues(color_endpoint_mode, num_partitions, color_data_bits);

    uvec4 endpoints[4][2];
    for (uint i = 0; i < num_partitions; i++) {
        ComputeEndpoints(endpoints[i][0], endpoints[i][1], color_endpoint_mode[i]);
    }

    for (uint i = 0; i < 16; i++) {
        texel_weight_data[i] = local_buff[i];
    }
    for (uint i = 0; i < 8; i++) {
#define REVERSE_BYTE(b) ((b * 0x0802U & 0x22110U) | (b * 0x8020U & 0x88440U)) * 0x10101U >> 16
        uint a = REVERSE_BYTE(texel_weight_data[i]);
        uint b = REVERSE_BYTE(texel_weight_data[15 - i]);
#undef REVERSE_BYTE
        texel_weight_data[i] = uint(bitfieldExtract(b, 0, 8));
        texel_weight_data[15 - i] = uint(bitfieldExtract(a, 0, 8));
    }
    uint clear_byte_start =
        (GetPackedBitSize(params.size, params.dual_plane, params.max_weight) >> 3) + 1;
    texel_weight_data[clear_byte_start - 1] =
        texel_weight_data[clear_byte_start - 1] &
        uint(
            ((1 << (GetPackedBitSize(params.size, params.dual_plane, params.max_weight) % 8)) - 1));
    for (uint i = 0; i < 16 - clear_byte_start; i++) {
        texel_weight_data[clear_byte_start + i] = 0U;
    }
    texel_flag = true; // use texel "vector" and bit stream in integer decoding
    DecodeIntegerSequence(params.max_weight, GetNumWeightValues(params.size, params.dual_plane));

    UnquantizeTexelWeights(params.dual_plane, params.size);

    for (uint j = 0; j < block_dims.y; j++) {
        for (uint i = 0; i < block_dims.x; i++) {
            uint local_partition = 0;
            if (num_partitions > 1) {
                local_partition = Select2DPartition(partition_index, i, j, num_partitions,
                                                     (block_dims.y * block_dims.x) < 32);
            }
            vec4 p;
            uvec4 C0 = ReplicateByteTo16(endpoints[local_partition][0]);
            uvec4 C1 = ReplicateByteTo16(endpoints[local_partition][1]);
            uvec4 plane_vec = uvec4(0);
            uvec4 weight_vec = uvec4(0);
            for (uint c = 0; c < 4; c++) {
                if (params.dual_plane && (((plane_index + 1) & 3) == c)) {
                    plane_vec[c] = 1;
                }
                weight_vec[c] = unquantized_texel_weights[plane_vec[c]][j * block_dims.x + i];
            }
            vec4 Cf = vec4((C0 * (uvec4(64) - weight_vec) + C1 * weight_vec + uvec4(32)) / 64);
            p = (Cf / 65535.0);
            imageStore(dest_image, coord + ivec3(i, j, 0), p.gbar);
        }
    }
}

void main() {
    uvec3 pos = gl_GlobalInvocationID;
    pos.x <<= bytes_per_block_log2;

    // Read as soon as possible due to its latency
    const uint swizzle = SwizzleOffset(pos.xy);

    const uint block_y = pos.y >> GOB_SIZE_Y_SHIFT;

    uint offset = 0;
    offset += pos.z * layer_stride;
    offset += (block_y >> block_height) * block_size;
    offset += (block_y & block_height_mask) << GOB_SIZE_SHIFT;
    offset += (pos.x >> GOB_SIZE_X_SHIFT) << x_shift;
    offset += swizzle;

    const ivec3 coord = ivec3(gl_GlobalInvocationID * uvec3(block_dims, 1));
    if (any(greaterThanEqual(coord, imageSize(dest_image)))) {
        return;
    }
    uint block_index =
        pos.z * gl_WorkGroupSize.x * gl_WorkGroupSize.y + pos.y * gl_WorkGroupSize.x + pos.x;

    current_index = 0;
    bitsread = 0;
    for (int i = 0; i < 16; i++) {
        local_buff[i] = ReadTexel(offset + i);
    }
    DecompressBlock(coord, block_index);
}