#version 450

/*
 * Iris Vulkan GEMM (Q8_0 weights)
 *
 * Computes C[M,N] = alpha * op(A)[M,K] @ op(B)[K,N] + beta * C[M,N]
 * where A and C are f32 and B is a GGML Q8_0-quantized weight matrix. Used for
 * weight-bound linear layers: A is the f32 activation, B is the Q8_0 weight.
 *
 * GGML Q8_0 packs the weight into blocks of 32 elements. Each block is 34
 * bytes laid out as: an fp16 scale (2 bytes) followed by 32 signed int8
 * quants. A weight element at flat index e dequantizes to
 *   value = float(qs[e % 32]) * fp16_to_f32(scale[e / 32])
 * Blocks never cross a matrix row because every weight's contraction
 * dimension (K) is a multiple of 32, so the flat index e == row*K + col maps
 * cleanly onto the block stream.
 *
 * The block buffer is read as a uint[] (no 16-bit storage extension needed):
 * the fp16 scale at the block start is always 16-bit aligned within a word
 * (block byte offset 34*b is congruent to 0 or 2 mod 4), and the int8 quants
 * are extracted with byte-granular shifts.
 */

layout(local_size_x = 16, local_size_y = 16) in;

layout(std430, binding = 0) readonly buffer ABuf { float A[]; };
layout(std430, binding = 1) readonly buffer BBuf { uint Bw[]; };
layout(std430, binding = 2) buffer CBuf { float C[]; };

layout(push_constant) uniform PC {
    uint M, N, K;
    uint lda, ldb, ldc;
    uint ta, tb;
    float alpha, beta;
} pc;

shared float As[16][16];
shared float Bs[16][16];

/* Dequantize the Q8_0 weight element at flat index e. */
float load_q8(uint e) {
    uint block  = e >> 5u;          /* e / 32 */
    uint within = e & 31u;          /* e % 32 */
    uint base   = block * 34u;      /* byte offset of this block */

    /* fp16 scale at byte `base` (16-bit aligned within its word). */
    uint sword = Bw[base >> 2u];
    vec2 shalf = unpackHalf2x16(sword);
    float scale = ((base & 2u) == 0u) ? shalf.x : shalf.y;

    /* int8 quant at byte base + 2 + within. */
    uint qoff  = base + 2u + within;
    uint qword = Bw[qoff >> 2u];
    uint b     = (qword >> ((qoff & 3u) * 8u)) & 0xffu;
    int  q     = (b < 128u) ? int(b) : int(b) - 256;   /* sign extend */

    return float(q) * scale;
}

void main() {
    uint row = gl_GlobalInvocationID.y;
    uint col = gl_GlobalInvocationID.x;
    uint tx = gl_LocalInvocationID.x;
    uint ty = gl_LocalInvocationID.y;

    float acc = 0.0;
    uint numTiles = (pc.K + 15u) / 16u;

    for (uint t = 0u; t < numTiles; t++) {
        uint kA = t * 16u + tx;
        uint kB = t * 16u + ty;

        float av = 0.0;
        if (row < pc.M && kA < pc.K) {
            av = (pc.ta == 0u) ? A[row * pc.lda + kA]
                               : A[kA * pc.lda + row];
        }
        As[ty][tx] = av;

        float bv = 0.0;
        if (col < pc.N && kB < pc.K) {
            uint e = (pc.tb == 0u) ? (kB * pc.ldb + col)
                                   : (col * pc.ldb + kB);
            bv = load_q8(e);
        }
        Bs[ty][tx] = bv;

        barrier();

        for (uint k = 0u; k < 16u; k++) {
            acc += As[ty][k] * Bs[k][tx];
        }

        barrier();
    }

    if (row < pc.M && col < pc.N) {
        uint ci = row * pc.ldc + col;
        float prev = (pc.beta != 0.0) ? C[ci] : 0.0;
        C[ci] = pc.alpha * acc + pc.beta * prev;
    }
}