#version 450

#extension GL_EXT_control_flow_attributes : enable
#extension GL_EXT_shader_16bit_storage : require

#ifdef FLOAT16
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
#endif

#ifdef MUL_MAT_ID
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#endif

#include "types.comp"

#ifndef LOAD_VEC_A
#define LOAD_VEC_A 1
#endif
#ifndef LOAD_VEC_B
#define LOAD_VEC_B 1
#endif

layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;

layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};

#ifdef MUL_MAT_ID
layout (binding = 3) readonly buffer IDS {int data_ids[];};
#endif

layout (push_constant) uniform parameter
{
    uint M;
    uint N;
    uint K;
    uint stride_a;
    uint stride_b;
    uint stride_d;

    uint batch_stride_a;
    uint batch_stride_b;
    uint batch_stride_d;

#ifdef MUL_MAT_ID
    uint nei0;
    uint nei1;
    uint nbi1;
    uint ne11;
#else
    uint k_split;
    uint ne02;
    uint ne12;
    uint broadcast2;
    uint broadcast3;
#endif
} p;

layout (constant_id = 1) const uint BM = 64;
layout (constant_id = 2) const uint BN = 64;
layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working with a quant
layout (constant_id = 4) const uint WM = 32;
layout (constant_id = 5) const uint WN = 32;
layout (constant_id = 6) const uint WMITER = 2;
layout (constant_id = 7) const uint TM = 4;
layout (constant_id = 8) const uint TN = 2;
layout (constant_id = 9) const uint WARP = 32;

shared FLOAT_TYPE buf_a[BM * (BK+1)];
shared FLOAT_TYPE buf_b[BN * (BK+1)];

#ifdef MUL_MAT_ID
shared u16vec2 row_ids[2048];
#endif

void main() {
#ifdef MUL_MAT_ID
    const uint expert_idx = gl_GlobalInvocationID.z;
#else
    const uint batch_idx = gl_GlobalInvocationID.z;

    const uint i13 = batch_idx / p.ne12;
    const uint i12 = batch_idx % p.ne12;

    const uint i03 = i13 / p.broadcast3;
    const uint i02 = i12 / p.broadcast2;

    const uint batch_idx_a = i03 * p.ne02 + i02;
#endif

    const uint blocks_m = (p.M + BM - 1) / BM;
    const uint ir = gl_WorkGroupID.x % blocks_m;
    const uint ik = gl_WorkGroupID.x / blocks_m;
    const uint ic = gl_WorkGroupID.y;

    const uint warp_i = gl_LocalInvocationID.x / WARP;
    const uint warp_r = warp_i % (BM / WM);
    const uint warp_c = warp_i / (BM / WM);

    const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
    const uint WSUBM = WM / WMITER;
    const uint WSUBN = WN / WNITER;

    const uint tiw = gl_LocalInvocationID.x % WARP;
    const uint tiwr = tiw % (WSUBM / TM);
    const uint tiwc = tiw / (WSUBM / TM);

    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A);
    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A);
    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B);
    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B);

    const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A / BK;
    const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B / BK;

#ifdef MUL_MAT_ID
    uint _ne1 = 0;
    for (uint ii1 = 0; ii1 < p.nei1; ii1++) {
        for (uint ii0 = 0; ii0 < p.nei0; ii0++) {
            if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) {
                row_ids[_ne1] = u16vec2(ii0, ii1);
                _ne1++;
            }
        }
    }

    barrier();

    // Workgroup has no work
    if (ic * BN >= _ne1) return;
#endif

#ifdef MUL_MAT_ID
    const uint start_k = 0;
    const uint end_k = p.K;
#else
    const uint start_k = ik * p.k_split;
    const uint end_k = min(p.K, (ik + 1) * p.k_split);
#endif

    uint pos_a = (
#ifdef MUL_MAT_ID
        expert_idx * p.batch_stride_a +
#else
        batch_idx_a * p.batch_stride_a +
#endif
        ir * BM * p.stride_a + start_k) / LOAD_VEC_A;
#ifdef MUL_MAT_ID
    uint pos_b = 0;
#else
    uint pos_b = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / LOAD_VEC_B;
#endif

    float sums[WMITER * TM * WNITER * TN];
    FLOAT_TYPE cache_a[WMITER * TM];
    FLOAT_TYPE cache_b[WNITER * TN];

    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
        sums[i] = 0.0f;
    }

    [[unroll]] for (uint block = start_k; block < end_k; block += BK) {
        [[unroll]] for (uint l = 0; l < BM; l += loadstride_a) {

#if defined(DATA_A_F32) || defined(DATA_A_F16)
#if LOAD_VEC_A == 8
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
            buf_a[buf_idx    ] = FLOAT_TYPE(data_a[idx][0].x);
            buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx][0].y);
            buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx][0].z);
            buf_a[buf_idx + 3] = FLOAT_TYPE(data_a[idx][0].w);
            buf_a[buf_idx + 4] = FLOAT_TYPE(data_a[idx][1].x);
            buf_a[buf_idx + 5] = FLOAT_TYPE(data_a[idx][1].y);
            buf_a[buf_idx + 6] = FLOAT_TYPE(data_a[idx][1].z);
            buf_a[buf_idx + 7] = FLOAT_TYPE(data_a[idx][1].w);
#elif LOAD_VEC_A == 4
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
            buf_a[buf_idx    ] = FLOAT_TYPE(data_a[idx].x);
            buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx].y);
            buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx].z);
            buf_a[buf_idx + 3] = FLOAT_TYPE(data_a[idx].w);
#else
            if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) {
                buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]);
            } else {
                buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(0.0f);
            }
#endif
#elif defined(DATA_A_Q4_0)
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;

            const uint ib = idx / 16;
            const uint iqs = idx & 0xF;

            const float d = float(data_a[ib].d);
            const uint vui = uint(data_a[ib].qs[iqs]);
            const vec2 v = (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;

            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_Q4_1)
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;

            const uint ib = idx / 16;
            const uint iqs = idx & 0xF;

            const float d = float(data_a[ib].d);
            const float m = float(data_a[ib].m);
            const uint vui = uint(data_a[ib].qs[iqs]);
            const vec2 v = vec2(vui & 0xF, vui >> 4) * d + m;

            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_Q5_0)
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;

            const uint ib = idx / 16;
            const uint iqs = idx & 0xF;

            const float d = float(data_a[ib].d);
            const uint uint_qh = uint(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0];
            const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
            const uint vui = uint(data_a[ib].qs[iqs]);
            const vec2 v = (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;

            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_Q5_1)
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;

            const uint ib = idx / 16;
            const uint iqs = idx & 0xF;

            const float d = float(data_a[ib].d);
            const float m = float(data_a[ib].m);
            const uint uint_qh = data_a[ib].qh;
            const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
            const uint vui = uint(data_a[ib].qs[iqs]);
            const vec2 v = vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;

            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_Q8_0)
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;

            const uint ib = idx / 16;
            const uint iqs = (idx & 0xF) * 2;

            const float d = float(data_a[ib].d);
            const vec2 v = vec2(int(data_a[ib].qs[iqs]), int(data_a[ib].qs[iqs + 1])) * d;

            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_Q2_K)
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;

            const uint ib = idx / 128;                         // 2 values per idx
            const uint iqs = idx % 128;                        // 0..127

            const uint qsi = (iqs / 64) * 32 + (iqs % 16) * 2; // 0,2,4..30
            const uint scalesi = iqs / 8;                      // 0..15
            const uint qsshift = ((iqs % 64) / 16) * 2;        // 0,2,4,6

            const uvec2 qs = uvec2(data_a[ib].qs[qsi], data_a[ib].qs[qsi + 1]);
            const uint scales = data_a[ib].scales[scalesi];
            const vec2 d = vec2(data_a[ib].d);

            const vec2 v = d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);

            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_Q3_K)
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;

            const uint ib = idx / 128;                   // 2 values per idx
            const uint iqs = idx % 128;                  // 0..127

            const uint n = iqs / 64;                     // 0,1
            const uint qsi = n * 32 + (iqs % 16) * 2;    // 0,2,4..62
            const uint hmi =          (iqs % 16) * 2;    // 0,2,4..30
            const uint j = (iqs % 64) / 4;               // 0..3
            const uint is = iqs / 8;                     // 0..15
            const uint halfsplit = ((iqs % 64) / 16);    // 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

            const int8_t us = int8_t(is <  4 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+8] >> 0) & 3) << 4) :
                                    is <  8 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+4] >> 2) & 3) << 4) :
                                    is < 12 ? (data_a[ib].scales[is-8] >>  4) | (((data_a[ib].scales[is+0] >> 4) & 3) << 4) :
                                            (data_a[ib].scales[is-8] >>  4) | (((data_a[ib].scales[is-4] >> 6) & 3) << 4));
            const float dl = float(data_a[ib].d) * float(us - 32);

            buf_a[buf_idx    ] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi    ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi    ] & m) != 0) ? 0 : 4)));
            buf_a[buf_idx + 1] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));
#elif defined(DATA_A_Q4_K)
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;

            const uint ib = idx / 128;                 // 2 values per idx
            const uint iqs = idx % 128;                // 0..127

            const uint n = iqs / 32;                   // 0,1,2,3
            const uint b = (iqs % 32) / 16;            // 0,1
            const uint is = 2 * n + b;                 // 0..7
            const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126

            const vec2 loadd = vec2(data_a[ib].d);

            uint8_t sc;
            uint8_t mbyte;
            if (is < 4) {
                sc    = uint8_t(data_a[ib].scales[is    ] & 63);
                mbyte = uint8_t(data_a[ib].scales[is + 4] & 63);
            } else {
                sc    = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4));
                mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
            }
            const float d = loadd.x * sc;
            const float m = loadd.y * mbyte;

            buf_a[buf_idx    ] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) - m);
            buf_a[buf_idx + 1] = FLOAT_TYPE(d * float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) - m);
#elif defined(DATA_A_Q5_K)
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;

            const uint ib = idx / 128;                 // 2 values per idx
            const uint iqs = idx % 128;                // 0..127

            const uint n = iqs / 32;                   // 0,1,2,3
            const uint b = (iqs % 32) / 16;            // 0,1
            const uint is = 2 * n + b;                 // 0..7
            const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
            const uint qhi = (iqs % 16) * 2;           // 0,2,4..30

            const uint8_t hm = uint8_t(1 << (iqs / 16));

            const vec2 loadd = vec2(data_a[ib].d);

            uint8_t sc;
            uint8_t mbyte;
            if (is < 4) {
                sc    = uint8_t(data_a[ib].scales[is    ] & 63);
                mbyte = uint8_t(data_a[ib].scales[is + 4] & 63);
            } else {
                sc    = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4));
                mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
            }
            const float d = loadd.x * sc;
            const float m = loadd.y * mbyte;

            buf_a[buf_idx    ] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi    ] & hm) != 0 ? 16 : 0)) - m);
            buf_a[buf_idx + 1] = FLOAT_TYPE(d * (float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0)) - m);
#elif defined(DATA_A_Q6_K)
            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;

            const uint ib = idx / 128;                  // 2 values per idx
            const uint iqs = idx % 128;                 // 0..127

            const uint n = iqs / 64;                    // 0,1
            const uint b = (iqs % 64) / 32;             // 0,1
            const uint is_b = (iqs % 16) / 8;           // 0,1
            const uint qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
            const uint is = 8 * n + qhshift + is_b;     // 0..15
            const uint qsi = n * 64 + (iqs % 32) * 2;   // 0,2,4..126
            const uint qhi = n * 32 + (iqs % 16) * 2;   // 0,2,4..62

            const float dscale = float(data_a[ib].d) * float(data_a[ib].scales[is]);

            buf_a[buf_idx    ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32));
            buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
#endif
        }
        [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {
#if LOAD_VEC_B == 8
#ifdef MUL_MAT_ID
            const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l];
            const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b;
#else
            const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
#endif
            const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B;
            buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx][0].x);
            buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx][0].y);
            buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx][0].z);
            buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx][0].w);
            buf_b[buf_idx + 4] = FLOAT_TYPE(data_b[idx][1].x);
            buf_b[buf_idx + 5] = FLOAT_TYPE(data_b[idx][1].y);
            buf_b[buf_idx + 6] = FLOAT_TYPE(data_b[idx][1].z);
            buf_b[buf_idx + 7] = FLOAT_TYPE(data_b[idx][1].w);
#elif LOAD_VEC_B == 4
#ifdef MUL_MAT_ID
            const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l];
            const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b;
#else
            const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
#endif
            const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B;
            buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx].x);
            buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx].y);
            buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx].z);
            buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx].w);
#elif !MUL_MAT_ID
            if (ic * BN + loadc_b + l < p.N && block + loadr_b < end_k) {
                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]);
            } else {
                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f);
            }
#else
            const uint row_i = ic * BN + loadc_b + l;
            if (row_i < _ne1) {
                const u16vec2 row_idx = row_ids[row_i];
                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
            } else {
                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f);
            }
#endif
        }

        barrier();

        pos_a += BK / LOAD_VEC_A;
        pos_b += BK / LOAD_VEC_B;

        for (uint i = 0; i < BK; i++) {
            // Load from shared into cache
            [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                [[unroll]] for (uint j = 0; j < TM; j++) {
                    cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * (BK+1) + i];
                }
            }
            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
                [[unroll]] for (uint j = 0; j < TN; j++) {
                    cache_b[wsic * TN + j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * (BK+1) + i];
                }
            }

            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
                [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                    [[unroll]] for (uint cc = 0; cc < TN; cc++) {
                        [[unroll]] for (uint cr = 0; cr < TM; cr++) {
                            sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr] += float(cache_a[wsir * TM + cr]) * float(cache_b[wsic * TN + cc]);
                        }
                    }
                }
            }
        }

        barrier();
    }

    const uint dr = ir * BM + warp_r * WM;
    const uint dc = ic * BN + warp_c * WN;

#ifndef MUL_MAT_ID
    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
#endif

    [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
        [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {

            const uint dr_warp = dr + wsir * WSUBM + tiwr * TM;
            const uint dc_warp = dc + wsic * WSUBN + tiwc * TN;
            [[unroll]] for (uint cc = 0; cc < TN; cc++) {
#ifdef MUL_MAT_ID
                const uint row_i = dc_warp + cc;
                if (row_i >= _ne1) break;

                const u16vec2 row_idx = row_ids[row_i];
#endif
                [[unroll]] for (uint cr = 0; cr < TM; cr++) {
#ifdef MUL_MAT_ID
                    data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
#else
                    if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
                    }
#endif
                }
            }
        }
    }
}