#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
#if defined(DATA_A_IQ1_M)
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#endif
#if !defined(DATA_A_F32) && !defined(DATA_A_F16) && !defined(DATA_A_BF16)
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#endif

#if defined(DATA_A_BF16) && defined(COOPMAT)
#extension GL_EXT_bfloat16 : enable
#endif

#ifdef COOPMAT
#extension GL_KHR_cooperative_matrix : enable
#extension GL_KHR_memory_scope_semantics : enable
#endif

#if defined(COOPMAT) || defined(MUL_MAT_ID_USE_SUBGROUPS)
#extension GL_KHR_shader_subgroup_basic : enable
#extension GL_KHR_shader_subgroup_ballot : enable
#endif

#ifdef MUL_MAT_ID
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#endif

#ifdef MULMAT_QUANT
#include "ggml_type_ids.glsl"
layout (constant_id = 12) const uint MmTypeA = 0;
#endif

#include "types.glsl"
#include "dot_product_funcs.glsl"

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

layout (constant_id = 11) const uint ALIGNED = 0;

#ifdef MULMAT_QUANT

uint mm_load_vec_a() {
    switch (MmTypeA) {
    case GGML_TYPE_Q1_0:
    case GGML_TYPE_Q4_0:
    case GGML_TYPE_Q4_1:
    case GGML_TYPE_Q5_1:
        return 8u;
    case GGML_TYPE_Q2_0:
    case GGML_TYPE_Q5_0:
    case GGML_TYPE_Q8_0:
    case GGML_TYPE_Q2_K:
    case GGML_TYPE_Q4_K:
    case GGML_TYPE_Q5_K:
        return 4u;
    default:
        return 2u;
    }
}
#endif

#if !defined(TO_FLOAT_TYPE)
#define TO_FLOAT_TYPE FLOAT_TYPE
#endif

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

#ifndef MULMAT_QUANT
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
#if defined(DATA_A_F32)
layout (binding = 0) readonly buffer A_SCALAR {float data_a_scalar[];};
#elif defined(DATA_A_F16)
layout (binding = 0) readonly buffer A_SCALAR {float16_t data_a_scalar[];};
#elif defined(DATA_A_BF16)
layout (binding = 0) readonly buffer A_SCALAR {uint16_t data_a_scalar[];};
#endif
#if defined(A_TYPE_PACKED16)
layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
#endif
#if defined(A_TYPE_PACKED32)
layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
#endif
#else
// Unpacked struct aliases
layout (binding = 0) readonly buffer BUF_Q1_0     { block_q1_0 data[];     } a_q1_0;
layout (binding = 0) readonly buffer BUF_Q2_0     { block_q2_0 data[];     } a_q2_0;
layout (binding = 0) readonly buffer BUF_Q2_K     { block_q2_K data[];     } a_q2_k;
layout (binding = 0) readonly buffer BUF_Q3_K     { block_q3_K data[];     } a_q3_k;
layout (binding = 0) readonly buffer BUF_Q4_K     { block_q4_K data[];     } a_q4_k;
layout (binding = 0) readonly buffer BUF_Q5_K     { block_q5_K data[];     } a_q5_k;
layout (binding = 0) readonly buffer BUF_Q6_K     { block_q6_K data[];     } a_q6_k;
layout (binding = 0) readonly buffer BUF_TQ1_0    { block_tq1_0 data[];    } a_tq1_0;
layout (binding = 0) readonly buffer BUF_TQ2_0    { block_tq2_0 data[];    } a_tq2_0;
// Packed16 aliases
layout (binding = 0) readonly buffer BUF_Q4_0_P16   { block_q4_0_packed16 data[];   } a_q4_0_p16;
layout (binding = 0) readonly buffer BUF_Q5_0_P16   { block_q5_0_packed16 data[];   } a_q5_0_p16;
layout (binding = 0) readonly buffer BUF_Q8_0_P16   { block_q8_0_packed16 data[];   } a_q8_0_p16;
layout (binding = 0) readonly buffer BUF_Q3_K_P16   { block_q3_K_packed16 data[];   } a_q3_k_p16;
layout (binding = 0) readonly buffer BUF_Q6_K_P16   { block_q6_K_packed16 data[];   } a_q6_k_p16;
// Packed32 aliases
layout (binding = 0) readonly buffer BUF_Q4_1_P32   { block_q4_1_packed32 data[];   } a_q4_1_p32;
layout (binding = 0) readonly buffer BUF_Q5_1_P32   { block_q5_1_packed32 data[];   } a_q5_1_p32;
layout (binding = 0) readonly buffer BUF_Q2_K_P32   { block_q2_K_packed32 data[];   } a_q2_k_p32;
layout (binding = 0) readonly buffer BUF_Q4_K_P32   { block_q4_K_packed32 data[];   } a_q4_k_p32;
layout (binding = 0) readonly buffer BUF_Q5_K_P32   { block_q5_K_packed32 data[];   } a_q5_k_p32;
#endif

layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
layout (binding = 1) readonly buffer B_SCALAR {B_TYPE_SCALAR data_b_scalar[];};
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};

#ifdef MUL_MAT_ID
layout (binding = 3) readonly buffer IDS {int data_ids[];};
layout (binding = 4) readonly buffer Counts {int data_expert_count[];};
#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;
    uint n_experts;
    uint hoist_row_ids;
#else
    uint base_work_group_z;
    uint num_batches;
    uint k_split;
    uint ne02;
    uint ne12;
    uint broadcast2;
    uint broadcast3;
#endif
} p;

layout (constant_id = 0) const uint BLOCK_SIZE = 64;
layout (constant_id = 1) const uint BM = 64;
layout (constant_id = 2) const uint BN = 64;
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 TK = 1;  // Only needed for coopmat
layout (constant_id = 10) const uint WARP = 32;

#if defined(DATA_A_F32) || defined(DATA_A_F16)
#define BK 32
#define BK_STEP 4
#else
layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working with a quant
#define BK_STEP 2
#endif

#ifdef COOPMAT
#ifdef MULMAT_QUANT
layout(constant_id = 13) const uint SHMEM_STRIDE_PAD = 4;
layout(constant_id = 14) const bool APPLY_SLM_A_RESHAPE = false;
#else
layout(constant_id = 12) const uint SHMEM_STRIDE_PAD = 4;
layout(constant_id = 13) const bool APPLY_SLM_A_RESHAPE = false;
#endif
#else
const uint SHMEM_STRIDE_PAD = 1;
const bool APPLY_SLM_A_RESHAPE = false;
#endif
#define SHMEM_STRIDE (BK / 2 + SHMEM_STRIDE_PAD)

shared FLOAT_TYPEV2 buf_a[BM * SHMEM_STRIDE];
shared FLOAT_TYPEV2 buf_b[BN * SHMEM_STRIDE];

#define NUM_WARPS (BLOCK_SIZE / WARP)

#ifdef COOPMAT
shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
#endif

#include "mul_mm_id_funcs.glsl"
#include "mul_mm_funcs.glsl"

#ifdef MULMAT_QUANT
#include "iq_shmem_init.glsl"
#endif

void main() {
    const uint ic = gl_WorkGroupID.y;

#ifdef MUL_MAT_ID
    const uint expert_idx = gl_WorkGroupID.z;
    if (ic * BN >= data_expert_count[expert_idx]) {
        return;
    }
#endif
#if defined(NEEDS_INIT_IQ_SHMEM) || defined(MULMAT_QUANT)
    init_iq_shmem(gl_WorkGroupSize);
#endif

#ifndef MUL_MAT_ID
    const uint batch_idx = gl_WorkGroupID.z + p.base_work_group_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 WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
    const uint WSUBM = WM / WMITER;
    const uint WSUBN = WN / WNITER;

#ifdef COOPMAT
    const uint warp_i = gl_SubgroupID;

    const uint tiw = gl_SubgroupInvocationID;

    const uint cms_per_row = WM / TM;
    const uint cms_per_col = WN / TN;

    const uint storestride = WARP / TM;
    const uint store_r = tiw % TM;
    const uint store_c = tiw / TM;
#else
    const uint warp_i = gl_LocalInvocationID.x / WARP;

    const uint tiw = gl_LocalInvocationID.x % WARP;

    const uint tiwr = tiw % (WSUBM / TM);
    const uint tiwc = tiw / (WSUBM / TM);
#endif

    const uint warp_r = warp_i % (BM / WM);
    const uint warp_c = warp_i / (BM / WM);

#if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)
    const uint LOAD_VEC_A_EFF = (ALIGNED != 0) ? LOAD_VEC_A : 1;
    const uint LOAD_VEC_BATCH_A = (ALIGNED != 0) ? 1 : 2;
#elif !defined(MULMAT_QUANT)
    const uint LOAD_VEC_A_EFF = LOAD_VEC_A;
    const uint LOAD_VEC_BATCH_A = 1;
#else
    const uint LOAD_VEC_A_EFF = mm_load_vec_a();
    const uint LOAD_VEC_BATCH_A = 1;
#endif
    const uint LOAD_VEC_B_EFF = (ALIGNED != 0) ? LOAD_VEC_B : 1;
    const uint LOAD_VEC_BATCH_B = (ALIGNED != 0) ? 1 : 2;

    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A_EFF / LOAD_VEC_BATCH_A);
    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A_EFF / LOAD_VEC_BATCH_A);
    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B_EFF / LOAD_VEC_BATCH_B);
    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B_EFF / LOAD_VEC_BATCH_B);

    const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A_EFF * LOAD_VEC_BATCH_A / BK;
    const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B_EFF * LOAD_VEC_BATCH_B / BK;

#ifdef MUL_MAT_ID
    if (p.hoist_row_ids != 0) {
        load_row_ids_hoisted(expert_idx, ic);
    } else {
#ifdef MUL_MAT_ID_USE_SUBGROUPS
        if (bitCount(p.nei0) == 1) {
            load_row_ids(expert_idx, true, ic);
        } else {
            load_row_ids(expert_idx, false, ic);
        }
#else
        _ne1 = 0;
        for (uint ii1 = 0; ii1 < p.nei1 && _ne1 < (ic + 1) * BN; ii1++) {
            for (uint ii0 = 0; ii0 < p.nei0 && _ne1 < (ic + 1) * BN; ii0++) {
                if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) {
                    if (_ne1 >= ic * BN) {
                        row_ids[_ne1 - ic * BN] = u16vec2(ii0, ii1);
                    }
                    _ne1++;
                }
            }
        }

        barrier();
#endif
    }

    // 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 / LOAD_VEC_A_EFF) +
#else
        batch_idx_a * (p.batch_stride_a / LOAD_VEC_A_EFF) +
#endif
        (ir * BM * p.stride_a + start_k) / LOAD_VEC_A_EFF;
#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_EFF;
#endif

#ifdef COOPMAT
    coopmat<FLOAT_TYPE, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a;
    coopmat<FLOAT_TYPE, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
    coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col];

    [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) {
        sums[i] = coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0f);
    }
#else
    ACC_TYPEV2 sums[WMITER * TM * WNITER * TN/2];
#if defined(DATA_A_F32) || defined(DATA_A_F16)
    FLOAT_TYPEV4 cache_a[WMITER * TM];
    FLOAT_TYPEV4 cache_b;
#else
    FLOAT_TYPEV2 cache_a[WMITER * TM];
    FLOAT_TYPEV2 cache_b;
#endif

    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN/2; i++) {
        sums[i] = ACC_TYPEV2(0.0f, 0.0f);
    }
#endif

    for (uint block = start_k; block < end_k; block += BK) {
        [[unroll]] for (uint l = 0; l < BM; l += loadstride_a) {
            load_a_to_shmem(pos_a, loadr_a, loadc_a + l, ir * BM + loadc_a + l, block, end_k);
        }
        [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {
#if !defined(MUL_MAT_ID)
            load_b_to_shmem(pos_b, loadr_b, loadc_b + l, ic * BN + loadc_b + l, block, end_k);
#else
            load_b_to_shmem(pos_b, loadr_b, loadc_b + l, ic, _ne1, block, end_k);
#endif
        }

        barrier();

        pos_a += BK / LOAD_VEC_A_EFF;
        pos_b += BK / LOAD_VEC_B_EFF;

#ifdef COOPMAT
        [[unroll]] for (uint i = 0; i < BK; i += TK) {
            [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
                // Load from shared into cache
                coopMatLoad(cache_a, buf_a, a_shmem_index(warp_r * WM + cm_row * TM, i / 2), a_shmem_stride(), gl_CooperativeMatrixLayoutRowMajor);

                [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
                    coopMatLoad(cache_b, buf_b, (warp_c * WN + cm_col * TN) * SHMEM_STRIDE + i / 2, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor);

                    sums[cm_col * cms_per_row + cm_row] = coopMatMulAdd(cache_a, cache_b, sums[cm_col * cms_per_row + cm_row]);
                }
            }
        }
#else
        [[unroll]] for (uint i = 0; i < BK / BK_STEP; i++) {
            // Load from shared into cache
            [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                [[unroll]] for (uint j = 0; j < TM; j++) {
                #if defined(DATA_A_F32) || defined(DATA_A_F16)
                    cache_a[wsir * TM + j].xy = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + 2 * i    ];
                    cache_a[wsir * TM + j].zw = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + 2 * i + 1];
                #else
                    cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + i];
                #endif
                }
            }

            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
                [[unroll]] for (uint cc = 0; cc < TN; cc++) {
                #if defined(DATA_A_F32) || defined(DATA_A_F16)
                    cache_b.xy = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + 2 * i    ];
                    cache_b.zw = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + 2 * i + 1];
                #else
                    cache_b = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + i];
                #endif

                    [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                        [[unroll]] for (uint cr = 0; cr < TM / 2; cr++) {
                            // [WNITER][TN][WMITER][TM / 2] -> [wsic][cc][wsir][cr]
                            const uint sums_idx = (wsic * TN + cc) * WMITER * (TM / 2) + wsir * (TM / 2) + cr;
                            sums[sums_idx].x = dot_product(cache_a[wsir * TM + 2 * cr    ], cache_b, sums[sums_idx].x);
                            sums[sums_idx].y = dot_product(cache_a[wsir * TM + 2 * cr + 1], cache_b, sums[sums_idx].y);
                        }
                    }
                }
            }

        }
#endif

        barrier();
    }

#if defined(ACC_TYPE_MAX)
#ifdef COOPMAT
    [[unroll]] for (uint j = 0; j < cms_per_row * cms_per_col; j++) {
        [[unroll]] for (uint i = 0; i < sums[j].length(); ++i) {
            sums[j][i] = clamp(sums[j][i], -ACC_TYPE_MAX, ACC_TYPE_MAX);
        }
    }
#else
    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN/2; i++) {
        sums[i].x = clamp(sums[i].x, -ACC_TYPE_MAX, ACC_TYPE_MAX);
        sums[i].y = clamp(sums[i].y, -ACC_TYPE_MAX, ACC_TYPE_MAX);
    }
#endif
#endif

    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 * p.num_batches;
#endif

#ifdef COOPMAT
#ifdef MUL_MAT_ID
    [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
        [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
            coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);

            barrier();
            [[unroll]] for (uint col = 0; col < TN; col += storestride) {
                const uint row_i = dc + cm_col * TN + col + store_c;
                if (row_i >= _ne1) break;

                const u16vec2 row_idx = row_ids[row_i - ic * BN];

                if (dr + cm_row * TM + store_r < p.M) {
                    data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
                }
            }
            barrier();
        }
    }
#else
    const bool is_aligned = p.stride_d % 4 == 0;  // Assumption: D_TYPE == float

    [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
        [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
            const bool is_in_bounds = dr + (cm_row + 1) * TM <= p.M && dc + (cm_col + 1) * TN <= p.N;

            if (is_aligned && is_in_bounds) {
                // Full coopMat is within bounds and stride_d is aligned with 16B
                coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> cm_dtype = coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(sums[cm_col * cms_per_row + cm_row]);
                coopMatStore(cm_dtype, data_d, offsets + (dc + cm_col * TN) * p.stride_d + dr + cm_row * TM, p.stride_d, gl_CooperativeMatrixLayoutColumnMajor);
            } else if (is_in_bounds) {
                // Full coopMat is within bounds, but stride_d is not aligned
                coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);

                controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
                [[unroll]] for (uint col = 0; col < TN; col += storestride) {
                    data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
                }
                controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
            } else if (dr + cm_row * TM < p.M && dc + cm_col * TN < p.N) {
                // Partial coopMat is within bounds
                coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);

                controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
                [[unroll]] for (uint col = 0; col < TN; col += storestride) {
                    if (dr + cm_row * TM + store_r < p.M && dc + cm_col * TN + col + store_c < p.N) {
                        data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
                    }
                }
                controlBarrier(gl_ScopeSubgroup, gl_ScopeSubgroup, gl_StorageSemanticsShared, gl_SemanticsAcquireRelease);
            }
        }
    }
#endif // MUL_MAT_ID
#else
    [[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 - ic * BN];
#endif // MUL_MAT_ID
                [[unroll]] for (uint cr = 0; cr < TM / 2; cr++) {
                    const uint sums_idx = (wsic * TN + cc) * WMITER * (TM / 2) + wsir * (TM / 2) + cr;
#ifdef MUL_MAT_ID
                    if (dr_warp + 2 * cr < p.M) {
                        data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + 2 * cr] = D_TYPE(sums[sums_idx].x);
                    }
                    if (dr_warp + 2 * cr + 1 < p.M) {
                        data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + 2 * cr + 1] = D_TYPE(sums[sums_idx].y);
                    }
#else
                    if (dr_warp + 2 * cr < p.M && dc_warp + cc < p.N) {
                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + 2 * cr] = D_TYPE(sums[sums_idx].x);
                    }
                    if (dr_warp + 2 * cr + 1 < p.M && dc_warp + cc < p.N) {
                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + 2 * cr + 1] = D_TYPE(sums[sums_idx].y);
                    }
#endif // MUL_MAT_ID
                }
            }
        }
    }
#endif // COOPMAT
}