layout(set=0, binding=0) writeonly buffer s0 {
    FLOAT data[];
} uOutput;

layout(set=0, binding=1) readonly buffer s1 {
    FLOAT data[];
} uQuery;

layout(set=0, binding=2) readonly buffer s2 {
    FLOAT data[];
} uKey;

layout(set=0, binding=3) readonly buffer s3 {
    FLOAT data[];
} uValue;

layout(set=0, binding=4) readonly buffer s4 {
    FLOAT4 data[];
} uCacheKey;

layout(set=0, binding=5) readonly buffer s5 {
    FLOAT4 data[];
} uCacheValue;

layout(set=0, binding=6) readonly buffer s6 {
    FLOAT data[];
} uMask;

layout(set=0, binding=7) readonly uniform constBuffer {
    ivec4 s0; // qLen, kLen, headNum, kvHeadNum
    ivec4 s1; // headDim, group, pastLen, totalLen
    ivec4 s2; // maskQlen, maskKvlen, maskMode, cacheMaxLen
    vec4 f0;  // scale, 0, 0, 0
} uConst;

layout(local_size_x = 8, local_size_y = 8, local_size_z = 1) in;

const int MAX_D4 = 64; // headDim <= 256
const float NEG_INF = -3.402823466e38;

void main() {
    const int h = int(gl_GlobalInvocationID.x);
    const int q = int(gl_GlobalInvocationID.y);

    const int qLen = uConst.s0.x;
    const int kLen = uConst.s0.y;
    const int headNum = uConst.s0.z;
    const int kvHeadNum = uConst.s0.w;

    if (h >= headNum || q >= qLen) {
        return;
    }

    const int headDim = uConst.s1.x;
    const int group = uConst.s1.y;
    const int pastLen = uConst.s1.z;
    const int totalLen = uConst.s1.w;

    const int maskMode = uConst.s2.z;
    const int maskQlen = uConst.s2.x;
    const int maskKvlen = uConst.s2.y;
    const int maxLen = uConst.s2.w;

    const float scale = uConst.f0.x;

    const int kvh = h / group;
    if (kvh >= kvHeadNum) {
        return;
    }
    const int d4Size = headDim / 4;
    if (d4Size > MAX_D4) {
        return;
    }
    if (maxLen <= 0) {
        return;
    }

    // query: [1, qLen, headNum, headDim]
    const int qBase = (q * headNum + h) * headDim;
    vec4 outVec[MAX_D4];
    for (int i = 0; i < d4Size; ++i) {
        outVec[i] = vec4(0.0);
    }

    float m = -3.402823466e38;
    float s = 0.0;

    const int kStart = (maskMode == 1 && maskKvlen > 0 && totalLen > maskKvlen) ? (totalLen - maskKvlen) : 0;

    for (int k = 0; k < totalLen; ++k) {
        float dotSum = 0.0;
        for (int i = 0; i < d4Size; ++i) {
            const int d = i * 4;
            const int qOff = qBase + d;
            // cacheK (packed): [kvHeadNum, headDim/4, maxLen, 4]
            const int kOff = (kvh * d4Size + i) * maxLen + k;
            vec4 qv = vec4(float(uQuery.data[qOff + 0]), float(uQuery.data[qOff + 1]), float(uQuery.data[qOff + 2]),
                           float(uQuery.data[qOff + 3]));
            vec4 kv = vec4(uCacheKey.data[kOff]);
            dotSum += dot(qv, kv);
        }
        float score = dotSum * scale;
        if (maskMode == 1) {
            if (q < maskQlen && maskKvlen > 0 && k >= kStart) {
                const int mk = k - kStart;
                if (mk >= 0 && mk < maskKvlen) {
                    score += float(uMask.data[q * maskKvlen + mk]);
                }
            }
        } else if (maskMode == 2 && k > (pastLen + q)) {
            score = NEG_INF;
        }

        float w = 0.0;
        if (score > m) {
            float factor = exp(m - score);
            for (int i = 0; i < d4Size; ++i) {
                outVec[i] *= factor;
            }
            s = s * factor + 1.0;
            m = score;
            w = 1.0;
        } else {
            w = exp(score - m);
            s += w;
        }

        // cacheV (kvh-major, packed-D): [kvHeadNum, maxLen, headDim/4, 4]
        const int vOff4Base = ((kvh * maxLen + k) * d4Size);
        for (int i = 0; i < d4Size; ++i) {
            vec4 vv = vec4(uCacheValue.data[vOff4Base + i]);
            outVec[i] += w * vv;
        }
    }

    float invS = 1.0 / s;
    // output: [1, qLen, headNum * headDim]
    const int outBase = (q * headNum + h) * headDim;
    for (int i = 0; i < d4Size; ++i) {
        vec4 v = outVec[i] * invS;
        const int d = i * 4;
        uOutput.data[outBase + d + 0] = FLOAT(v.x);
        uOutput.data[outBase + d + 1] = FLOAT(v.y);
        uOutput.data[outBase + d + 2] = FLOAT(v.z);
        uOutput.data[outBase + d + 3] = FLOAT(v.w);
    }
}