submission 352282

# submission.py
#
# NVFP4 (e2m1) block-scaled dual GEMM with SiLU gating on NVIDIA Blackwell (B200 / SM100).
#
# Hybrid approach: single-threaded pipeline barriers, collective UMMA execution

from __future__ import annotations

import torch
from task import input_t, output_t

import cutlass
import cutlass.cute as cute
import cutlass.utils as utils
import cutlass.pipeline as pipeline
from cutlass.cute.nvgpu import cpasync, tcgen05
import cutlass.utils.blackwell_helpers as sm100_utils
import cutlass.utils.blockscaled_layout as blockscaled_utils
from cutlass.cute.runtime import make_ptr


# -------------------------
# Tuning knobs
# -------------------------

mma_tiler_mnk = (128, 128, 256)
mma_inst_shape_k = 64

ab_dtype = cutlass.Float4E2M1FN
sf_dtype = cutlass.Float8E4M3FN
c_dtype = cutlass.Float16

sf_vec_size = 16
threads_per_cta = 128

num_acc_stage = 1
num_ab_stage = 2

num_tmem_alloc_cols = 512
cluster_shape = (1, 1, 1)


def ceil_div(a: int, b: int) -> int:
    return (a + b - 1) // b


@cute.kernel
def dual_gemm_kernel(
    tiled_mma: cute.TiledMma,
    tma_atom_a: cute.CopyAtom,
    mA_mkl: cute.Tensor,
    tma_atom_b1: cute.CopyAtom,
    mB1_nkl: cute.Tensor,
    tma_atom_b2: cute.CopyAtom,
    mB2_nkl: cute.Tensor,
    tma_atom_sfa: cute.CopyAtom,
    mSFA_mkl: cute.Tensor,
    tma_atom_sfb1: cute.CopyAtom,
    mSFB1_nkl: cute.Tensor,
    tma_atom_sfb2: cute.CopyAtom,
    mSFB2_nkl: cute.Tensor,
    mC_mnl: cute.Tensor,
    a_smem_layout_staged: cute.ComposedLayout,
    b_smem_layout_staged: cute.ComposedLayout,
    sfa_smem_layout_staged: cute.Layout,
    sfb_smem_layout_staged: cute.Layout,
    num_tma_load_bytes: cutlass.Constexpr[int],
    silu_op: cutlass.Constexpr = lambda x: x * (1.0 / (1.0 + cute.math.exp(-x, fastmath=True))),
):
    warp_idx = cute.arch.warp_idx()
    warp_idx = cute.arch.make_warp_uniform(warp_idx)
    tidx, _, _ = cute.arch.thread_idx()

    bidx, bidy, bidz = cute.arch.block_idx()
    cta_coord = (bidx, bidy, bidz)

    wg_cnt = cute.size(tiled_mma.thr_id.shape)
    mma_tile_coord_v = cta_coord[0] % wg_cnt
    mma_tile_coord_mnl = (
        cta_coord[0] // wg_cnt,
        cta_coord[1],
        cta_coord[2],
    )

    @cute.struct
    class SharedStorage:
        ab_mbar_ptr: cute.struct.MemRange[cutlass.Int64, num_ab_stage * 2]
        acc_mbar_ptr: cute.struct.MemRange[cutlass.Int64, num_acc_stage * 2]
        tmem_holding_buf: cutlass.Int32

    smem = utils.SmemAllocator()
    storage = smem.allocate(SharedStorage)

    sA = smem.allocate_tensor(
        element_type=ab_dtype,
        layout=a_smem_layout_staged.outer,
        byte_alignment=128,
        swizzle=a_smem_layout_staged.inner,
    )
    sB1 = smem.allocate_tensor(
        element_type=ab_dtype,
        layout=b_smem_layout_staged.outer,
        byte_alignment=128,
        swizzle=b_smem_layout_staged.inner,
    )
    sB2 = smem.allocate_tensor(
        element_type=ab_dtype,
        layout=b_smem_layout_staged.outer,
        byte_alignment=128,
        swizzle=b_smem_layout_staged.inner,
    )
    sSFA = smem.allocate_tensor(
        element_type=sf_dtype,
        layout=sfa_smem_layout_staged,
        byte_alignment=128,
    )
    sSFB1 = smem.allocate_tensor(
        element_type=sf_dtype,
        layout=sfb_smem_layout_staged,
        byte_alignment=128,
    )
    sSFB2 = smem.allocate_tensor(
        element_type=sf_dtype,
        layout=sfb_smem_layout_staged,
        byte_alignment=128,
    )

    # REVERTED: Single-threaded pipeline barriers
    ab_pipeline_producer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread)
    ab_pipeline_consumer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread, 1)

    ab_producer, ab_consumer = pipeline.PipelineTmaUmma.create(
        barrier_storage=storage.ab_mbar_ptr.data_ptr(),
        num_stages=num_ab_stage,
        producer_group=ab_pipeline_producer_group,
        consumer_group=ab_pipeline_consumer_group,
        tx_count=num_tma_load_bytes,
    ).make_participants()

    acc_producer, acc_consumer = pipeline.PipelineUmmaAsync.create(
        barrier_storage=storage.acc_mbar_ptr.data_ptr(),
        num_stages=num_acc_stage,
        producer_group=ab_pipeline_producer_group,
        consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, threads_per_cta),
    ).make_participants()

    gA_mkl = cute.local_tile(mA_mkl, cute.slice_(mma_tiler_mnk, (None, 0, None)), (None, None, None))
    gB1_nkl = cute.local_tile(mB1_nkl, cute.slice_(mma_tiler_mnk, (0, None, None)), (None, None, None))
    gB2_nkl = cute.local_tile(mB2_nkl, cute.slice_(mma_tiler_mnk, (0, None, None)), (None, None, None))

    gSFA_mkl = cute.local_tile(mSFA_mkl, cute.slice_(mma_tiler_mnk, (None, 0, None)), (None, None, None))
    gSFB1_nkl = cute.local_tile(mSFB1_nkl, cute.slice_(mma_tiler_mnk, (0, None, None)), (None, None, None))
    gSFB2_nkl = cute.local_tile(mSFB2_nkl, cute.slice_(mma_tiler_mnk, (0, None, None)), (None, None, None))

    gC_mnl = cute.local_tile(mC_mnl, cute.slice_(mma_tiler_mnk, (None, None, 0)), (None, None, None))

    k_tile_cnt = cute.size(gA_mkl, mode=[3])

    thr_mma = tiled_mma.get_slice(mma_tile_coord_v)

    tCgA = thr_mma.partition_A(gA_mkl)
    tCgB1 = thr_mma.partition_B(gB1_nkl)
    tCgB2 = thr_mma.partition_B(gB2_nkl)

    tCgSFA = thr_mma.partition_A(gSFA_mkl)
    tCgSFB1 = thr_mma.partition_B(gSFB1_nkl)
    tCgSFB2 = thr_mma.partition_B(gSFB2_nkl)

    tCgC = thr_mma.partition_C(gC_mnl)

    tAsA,  tAgA  = cpasync.tma_partition(tma_atom_a,  0, cute.make_layout(1), cute.group_modes(sA, 0, 3),   cute.group_modes(tCgA, 0, 3))
    tBsB1, tBgB1 = cpasync.tma_partition(tma_atom_b1, 0, cute.make_layout(1), cute.group_modes(sB1, 0, 3),  cute.group_modes(tCgB1, 0, 3))
    tBsB2, tBgB2 = cpasync.tma_partition(tma_atom_b2, 0, cute.make_layout(1), cute.group_modes(sB2, 0, 3),  cute.group_modes(tCgB2, 0, 3))

    tAsSFA, tAgSFA = cpasync.tma_partition(tma_atom_sfa,  0, cute.make_layout(1), cute.group_modes(sSFA, 0, 3), cute.group_modes(tCgSFA, 0, 3))
    tAsSFA = cute.filter_zeros(tAsSFA)
    tAgSFA = cute.filter_zeros(tAgSFA)

    tBsSFB1, tBgSFB1 = cpasync.tma_partition(tma_atom_sfb1, 0, cute.make_layout(1), cute.group_modes(sSFB1, 0, 3), cute.group_modes(tCgSFB1, 0, 3))
    tBsSFB1 = cute.filter_zeros(tBsSFB1)
    tBgSFB1 = cute.filter_zeros(tBgSFB1)

    tBsSFB2, tBgSFB2 = cpasync.tma_partition(tma_atom_sfb2, 0, cute.make_layout(1), cute.group_modes(sSFB2, 0, 3), cute.group_modes(tCgSFB2, 0, 3))
    tBsSFB2 = cute.filter_zeros(tBsSFB2)
    tBgSFB2 = cute.filter_zeros(tBgSFB2)

    tCrA  = tiled_mma.make_fragment_A(sA)
    tCrB1 = tiled_mma.make_fragment_B(sB1)
    tCrB2 = tiled_mma.make_fragment_B(sB2)

    acc_shape = tiled_mma.partition_shape_C(mma_tiler_mnk[:2])
    tCtAcc_fake = tiled_mma.make_fragment_C(acc_shape)

    tmem_alloc_barrier = pipeline.NamedBarrier(barrier_id=1, num_threads=threads_per_cta)
    tmem = utils.TmemAllocator(storage.tmem_holding_buf, barrier_for_retrieve=tmem_alloc_barrier)
    tmem.allocate(num_tmem_alloc_cols)
    tmem.wait_for_alloc()

    acc_tmem_ptr = tmem.retrieve_ptr(cutlass.Float32)
    tCtAcc1 = cute.make_tensor(acc_tmem_ptr, tCtAcc_fake.layout)

    acc_tmem_ptr2 = cute.recast_ptr(
        acc_tmem_ptr + tcgen05.find_tmem_tensor_col_offset(tCtAcc1),
        dtype=cutlass.Float32,
    )
    tCtAcc2 = cute.make_tensor(acc_tmem_ptr2, tCtAcc_fake.layout)

    tCtSFA_layout = blockscaled_utils.make_tmem_layout_sfa(
        tiled_mma, mma_tiler_mnk, sf_vec_size, cute.slice_(sfa_smem_layout_staged, (None, None, None, 0))
    )
    sfa_tmem_ptr = cute.recast_ptr(
        acc_tmem_ptr
        + tcgen05.find_tmem_tensor_col_offset(tCtAcc1)
        + tcgen05.find_tmem_tensor_col_offset(tCtAcc2),
        dtype=sf_dtype,
    )
    tCtSFA = cute.make_tensor(sfa_tmem_ptr, tCtSFA_layout)

    tCtSFB_layout = blockscaled_utils.make_tmem_layout_sfb(
        tiled_mma, mma_tiler_mnk, sf_vec_size, cute.slice_(sfb_smem_layout_staged, (None, None, None, 0))
    )
    sfb1_tmem_ptr = cute.recast_ptr(
        acc_tmem_ptr
        + tcgen05.find_tmem_tensor_col_offset(tCtAcc1)
        + tcgen05.find_tmem_tensor_col_offset(tCtAcc2)
        + tcgen05.find_tmem_tensor_col_offset(tCtSFA),
        dtype=sf_dtype,
    )
    tCtSFB1 = cute.make_tensor(sfb1_tmem_ptr, tCtSFB_layout)

    sfb2_tmem_ptr = cute.recast_ptr(
        acc_tmem_ptr
        + tcgen05.find_tmem_tensor_col_offset(tCtAcc1)
        + tcgen05.find_tmem_tensor_col_offset(tCtAcc2)
        + tcgen05.find_tmem_tensor_col_offset(tCtSFA)
        + tcgen05.find_tmem_tensor_col_offset(tCtSFB1),
        dtype=sf_dtype,
    )
    tCtSFB2 = cute.make_tensor(sfb2_tmem_ptr, tCtSFB_layout)

    copy_atom_s2t = cute.make_copy_atom(tcgen05.Cp4x32x128bOp(tcgen05.CtaGroup.ONE), sf_dtype)

    tCsSFA_compact = cute.filter_zeros(sSFA)
    tCtSFA_compact = cute.filter_zeros(tCtSFA)
    tiled_copy_s2t_sfa = tcgen05.make_s2t_copy(copy_atom_s2t, tCtSFA_compact)
    thr_copy_s2t_sfa = tiled_copy_s2t_sfa.get_slice(0)
    tCsSFA_compact_s2t_ = thr_copy_s2t_sfa.partition_S(tCsSFA_compact)
    tCsSFA_compact_s2t = tcgen05.get_s2t_smem_desc_tensor(tiled_copy_s2t_sfa, tCsSFA_compact_s2t_)
    tCtSFA_compact_s2t = thr_copy_s2t_sfa.partition_D(tCtSFA_compact)

    tCsSFB1_compact = cute.filter_zeros(sSFB1)
    tCtSFB1_compact = cute.filter_zeros(tCtSFB1)
    tiled_copy_s2t_sfb = tcgen05.make_s2t_copy(copy_atom_s2t, tCtSFB1_compact)
    thr_copy_s2t_sfb = tiled_copy_s2t_sfb.get_slice(0)
    tCsSFB1_compact_s2t_ = thr_copy_s2t_sfb.partition_S(tCsSFB1_compact)
    tCsSFB1_compact_s2t = tcgen05.get_s2t_smem_desc_tensor(tiled_copy_s2t_sfb, tCsSFB1_compact_s2t_)
    tCtSFB1_compact_s2t = thr_copy_s2t_sfb.partition_D(tCtSFB1_compact)

    tCsSFB2_compact = cute.filter_zeros(sSFB2)
    tCtSFB2_compact = cute.filter_zeros(tCtSFB2)
    tCsSFB2_compact_s2t_ = thr_copy_s2t_sfb.partition_S(tCsSFB2_compact)
    tCsSFB2_compact_s2t = tcgen05.get_s2t_smem_desc_tensor(tiled_copy_s2t_sfb, tCsSFB2_compact_s2t_)
    tCtSFB2_compact_s2t = thr_copy_s2t_sfb.partition_D(tCtSFB2_compact)

    tAgA   = tAgA[(None, mma_tile_coord_mnl[0], None, mma_tile_coord_mnl[2])]
    tBgB1  = tBgB1[(None, mma_tile_coord_mnl[1], None, mma_tile_coord_mnl[2])]
    tBgB2  = tBgB2[(None, mma_tile_coord_mnl[1], None, mma_tile_coord_mnl[2])]
    tAgSFA = tAgSFA[(None, mma_tile_coord_mnl[0], None, mma_tile_coord_mnl[2])]
    tBgSFB1 = tBgSFB1[(None, mma_tile_coord_mnl[1], None, mma_tile_coord_mnl[2])]
    tBgSFB2 = tBgSFB2[(None, mma_tile_coord_mnl[1], None, mma_tile_coord_mnl[2])]

    # -------------------------
    # Mainloop: Single warp does pipeline, all threads do UMMA
    # -------------------------
    if warp_idx == 0:
        acc_empty = acc_producer.acquire_and_advance()
        tiled_mma.set(tcgen05.Field.ACCUMULATE, False)

        prefetch = num_ab_stage - 1

        # Prefetch
        for k_prefetch in range(prefetch):
            if k_prefetch < k_tile_cnt:
                ab_empty = ab_producer.acquire_and_advance()

                cute.copy(tma_atom_a,  tAgA[(None, k_prefetch)],  tAsA[(None, ab_empty.index)],  tma_bar_ptr=ab_empty.barrier)
                cute.copy(tma_atom_b1, tBgB1[(None, k_prefetch)], tBsB1[(None, ab_empty.index)], tma_bar_ptr=ab_empty.barrier)
                cute.copy(tma_atom_b2, tBgB2[(None, k_prefetch)], tBsB2[(None, ab_empty.index)], tma_bar_ptr=ab_empty.barrier)

                cute.copy(tma_atom_sfa,  tAgSFA[(None, k_prefetch)],  tAsSFA[(None, ab_empty.index)],  tma_bar_ptr=ab_empty.barrier)
                cute.copy(tma_atom_sfb1, tBgSFB1[(None, k_prefetch)], tBsSFB1[(None, ab_empty.index)], tma_bar_ptr=ab_empty.barrier)
                cute.copy(tma_atom_sfb2, tBgSFB2[(None, k_prefetch)], tBsSFB2[(None, ab_empty.index)], tma_bar_ptr=ab_empty.barrier)

                ab_empty.commit()

        # Main loop
        for k_tile in range(k_tile_cnt):
            next_k = k_tile + prefetch
            if next_k < k_tile_cnt:
                ab_empty = ab_producer.acquire_and_advance()

                cute.copy(tma_atom_a,  tAgA[(None, next_k)],  tAsA[(None, ab_empty.index)],  tma_bar_ptr=ab_empty.barrier)
                cute.copy(tma_atom_b1, tBgB1[(None, next_k)], tBsB1[(None, ab_empty.index)], tma_bar_ptr=ab_empty.barrier)
                cute.copy(tma_atom_b2, tBgB2[(None, next_k)], tBsB2[(None, ab_empty.index)], tma_bar_ptr=ab_empty.barrier)

                cute.copy(tma_atom_sfa,  tAgSFA[(None, next_k)],  tAsSFA[(None, ab_empty.index)],  tma_bar_ptr=ab_empty.barrier)
                cute.copy(tma_atom_sfb1, tBgSFB1[(None, next_k)], tBsSFB1[(None, ab_empty.index)], tma_bar_ptr=ab_empty.barrier)
                cute.copy(tma_atom_sfb2, tBgSFB2[(None, next_k)], tBsSFB2[(None, ab_empty.index)], tma_bar_ptr=ab_empty.barrier)

                ab_empty.commit()

            ab_full = ab_consumer.wait_and_advance()

            s2t_stage_coord = (None, None, None, None, ab_full.index)
            cute.copy(tiled_copy_s2t_sfa, tCsSFA_compact_s2t[s2t_stage_coord], tCtSFA_compact_s2t)
            cute.copy(tiled_copy_s2t_sfb, tCsSFB1_compact_s2t[s2t_stage_coord], tCtSFB1_compact_s2t)
            cute.copy(tiled_copy_s2t_sfb, tCsSFB2_compact_s2t[s2t_stage_coord], tCtSFB2_compact_s2t)

            num_kblocks = cute.size(tCrA, mode=[2])
            for kblock_idx in cutlass.range(num_kblocks, unroll_full=True):
                kblock_coord = (None, None, kblock_idx, ab_full.index)
                sf_kblock_coord = (None, None, kblock_idx)

                tiled_mma.set(tcgen05.Field.SFA, tCtSFA[sf_kblock_coord].iterator)

                tiled_mma.set(tcgen05.Field.SFB, tCtSFB1[sf_kblock_coord].iterator)
                cute.gemm(tiled_mma, tCtAcc1, tCrA[kblock_coord], tCrB1[kblock_coord], tCtAcc1)

                tiled_mma.set(tcgen05.Field.SFB, tCtSFB2[sf_kblock_coord].iterator)
                cute.gemm(tiled_mma, tCtAcc2, tCrA[kblock_coord], tCrB2[kblock_coord], tCtAcc2)

                tiled_mma.set(tcgen05.Field.ACCUMULATE, True)

            ab_full.release()

        acc_empty.commit()

    # Sync all threads before epilogue
    cute.arch.barrier()

    # -------------------------
    # Epilogue
    # -------------------------
    op = tcgen05.Ld32x32bOp(tcgen05.Repetition.x128, tcgen05.Pack.NONE)
    copy_atom_t2r = cute.make_copy_atom(op, cutlass.Float32)
    tiled_copy_t2r = tcgen05.make_tmem_copy(copy_atom_t2r, tCtAcc1)
    thr_copy_t2r = tiled_copy_t2r.get_slice(tidx)

    tTR_tAcc1 = thr_copy_t2r.partition_S(tCtAcc1)
    tTR_tAcc2 = thr_copy_t2r.partition_S(tCtAcc2)
    tTR_gC = thr_copy_t2r.partition_D(tCgC)

    tTR_rAcc1 = cute.make_rmem_tensor(tTR_gC[None, None, None, None, 0, 0, 0].shape, cutlass.Float32)
    tTR_rAcc2 = cute.make_rmem_tensor(tTR_gC[None, None, None, None, 0, 0, 0].shape, cutlass.Float32)
    tTR_rC    = cute.make_rmem_tensor(tTR_gC[None, None, None, None, 0, 0, 0].shape, c_dtype)

    simt_atom = cute.make_copy_atom(cute.nvgpu.CopyUniversalOp(), c_dtype)
    tTR_gC = tTR_gC[(None, None, None, None, *mma_tile_coord_mnl)]

    acc_full = acc_consumer.wait_and_advance()

    cute.copy(tiled_copy_t2r, tTR_tAcc1, tTR_rAcc1)
    cute.copy(tiled_copy_t2r, tTR_tAcc2, tTR_rAcc2)

    acc1 = silu_op(tTR_rAcc1.load())
    acc2 = tTR_rAcc2.load()
    out = acc1 * acc2

    tTR_rC.store(out.to(c_dtype))
    cute.copy(simt_atom, tTR_rC, tTR_gC)

    acc_full.release()

    cute.arch.barrier()
    tmem.free(acc_tmem_ptr)
    return


@cute.jit
def dual_gemm_launch(
    a_ptr: cute.Pointer,
    b1_ptr: cute.Pointer,
    b2_ptr: cute.Pointer,
    sfa_ptr: cute.Pointer,
    sfb1_ptr: cute.Pointer,
    sfb2_ptr: cute.Pointer,
    c_ptr: cute.Pointer,
    problem_size: tuple,
    silu_op: cutlass.Constexpr = lambda x: x * (1.0 / (1.0 + cute.math.exp(-x, fastmath=True))),
):
    m, n, k, l = problem_size

    a_tensor = cute.make_tensor(
        a_ptr,
        cute.make_layout(
            (m, cute.assume(k, 32), l),
            stride=(cute.assume(k, 32), 1, cute.assume(m * k, 32)),
        ),
    )
    b_tensor1 = cute.make_tensor(
        b1_ptr,
        cute.make_layout(
            (n, cute.assume(k, 32), l),
            stride=(cute.assume(k, 32), 1, cute.assume(n * k, 32)),
        ),
    )
    b_tensor2 = cute.make_tensor(
        b2_ptr,
        cute.make_layout(
            (n, cute.assume(k, 32), l),
            stride=(cute.assume(k, 32), 1, cute.assume(n * k, 32)),
        ),
    )
    c_tensor = cute.make_tensor(
        c_ptr,
        cute.make_layout((cute.assume(m, 32), n, l), stride=(n, 1, m * n)),
    )

    sfa_layout = blockscaled_utils.tile_atom_to_shape_SF(a_tensor.shape, sf_vec_size)
    sfa_tensor = cute.make_tensor(sfa_ptr, sfa_layout)

    sfb_layout = blockscaled_utils.tile_atom_to_shape_SF(b_tensor1.shape, sf_vec_size)
    sfb_tensor1 = cute.make_tensor(sfb1_ptr, sfb_layout)
    sfb_tensor2 = cute.make_tensor(sfb2_ptr, sfb_layout)

    mma_op = tcgen05.MmaMXF4NVF4Op(
        sf_dtype,
        (mma_tiler_mnk[0], mma_tiler_mnk[1], mma_inst_shape_k),
        tcgen05.CtaGroup.ONE,
        tcgen05.OperandSource.SMEM,
    )
    tiled_mma = cute.make_tiled_mma(mma_op)

    cluster_layout_vmnk = cute.tiled_divide(cute.make_layout(cluster_shape), (tiled_mma.thr_id.shape,))

    a_smem_layout_staged = sm100_utils.make_smem_layout_a(tiled_mma, mma_tiler_mnk, ab_dtype, num_ab_stage)
    b_smem_layout_staged = sm100_utils.make_smem_layout_b(tiled_mma, mma_tiler_mnk, ab_dtype, num_ab_stage)

    sfa_smem_layout_staged = blockscaled_utils.make_smem_layout_sfa(tiled_mma, mma_tiler_mnk, sf_vec_size, num_ab_stage)
    sfb_smem_layout_staged = blockscaled_utils.make_smem_layout_sfb(tiled_mma, mma_tiler_mnk, sf_vec_size, num_ab_stage)

    atom_thr_size = cute.size(tiled_mma.thr_id.shape)

    a_smem_layout = cute.slice_(a_smem_layout_staged, (None, None, None, 0))
    tma_atom_a, tma_tensor_a = cute.nvgpu.make_tiled_tma_atom_A(
        cpasync.CopyBulkTensorTileG2SOp(tcgen05.CtaGroup.ONE),
        a_tensor,
        a_smem_layout,
        mma_tiler_mnk,
        tiled_mma,
        cluster_layout_vmnk.shape,
    )

    b_smem_layout = cute.slice_(b_smem_layout_staged, (None, None, None, 0))
    tma_atom_b1, tma_tensor_b1 = cute.nvgpu.make_tiled_tma_atom_B(
        cpasync.CopyBulkTensorTileG2SOp(tcgen05.CtaGroup.ONE),
        b_tensor1,
        b_smem_layout,
        mma_tiler_mnk,
        tiled_mma,
        cluster_layout_vmnk.shape,
    )
    tma_atom_b2, tma_tensor_b2 = cute.nvgpu.make_tiled_tma_atom_B(
        cpasync.CopyBulkTensorTileG2SOp(tcgen05.CtaGroup.ONE),
        b_tensor2,
        b_smem_layout,
        mma_tiler_mnk,
        tiled_mma,
        cluster_layout_vmnk.shape,
    )

    sfa_smem_layout = cute.slice_(sfa_smem_layout_staged, (None, None, None, 0))
    tma_atom_sfa, tma_tensor_sfa = cute.nvgpu.make_tiled_tma_atom_A(
        cpasync.CopyBulkTensorTileG2SOp(tcgen05.CtaGroup.ONE),
        sfa_tensor,
        sfa_smem_layout,
        mma_tiler_mnk,
        tiled_mma,
        cluster_layout_vmnk.shape,
        internal_type=cutlass.Int16,
    )

    sfb_smem_layout = cute.slice_(sfb_smem_layout_staged, (None, None, None, 0))
    tma_atom_sfb1, tma_tensor_sfb1 = cute.nvgpu.make_tiled_tma_atom_B(
        cpasync.CopyBulkTensorTileG2SOp(tcgen05.CtaGroup.ONE),
        sfb_tensor1,
        sfb_smem_layout,
        mma_tiler_mnk,
        tiled_mma,
        cluster_layout_vmnk.shape,
        internal_type=cutlass.Int16,
    )
    tma_atom_sfb2, tma_tensor_sfb2 = cute.nvgpu.make_tiled_tma_atom_B(
        cpasync.CopyBulkTensorTileG2SOp(tcgen05.CtaGroup.ONE),
        sfb_tensor2,
        sfb_smem_layout,
        mma_tiler_mnk,
        tiled_mma,
        cluster_layout_vmnk.shape,
        internal_type=cutlass.Int16,
    )

    a_copy_size = cute.size_in_bytes(ab_dtype, a_smem_layout)
    b_copy_size = cute.size_in_bytes(ab_dtype, b_smem_layout)
    sfa_copy_size = cute.size_in_bytes(sf_dtype, sfa_smem_layout)
    sfb_copy_size = cute.size_in_bytes(sf_dtype, sfb_smem_layout)

    num_tma_load_bytes = (
        a_copy_size + (2 * b_copy_size) + sfa_copy_size + (2 * sfb_copy_size)
    ) * atom_thr_size

    grid = (
        cute.ceil_div(c_tensor.shape[0], mma_tiler_mnk[0]),
        cute.ceil_div(c_tensor.shape[1], mma_tiler_mnk[1]),
        c_tensor.shape[2],
    )

    dual_gemm_kernel(
        tiled_mma,
        tma_atom_a,   tma_tensor_a,
        tma_atom_b1,  tma_tensor_b1,
        tma_atom_b2,  tma_tensor_b2,
        tma_atom_sfa, tma_tensor_sfa,
        tma_atom_sfb1, tma_tensor_sfb1,
        tma_atom_sfb2, tma_tensor_sfb2,
        c_tensor,
        a_smem_layout_staged,
        b_smem_layout_staged,
        sfa_smem_layout_staged,
        sfb_smem_layout_staged,
        num_tma_load_bytes,
        silu_op,
    ).launch(
        grid=grid,
        block=[threads_per_cta, 1, 1],
        cluster=cluster_shape,
    )
    return


_compiled_kernel_cache = None


def compile_kernel():
    global _compiled_kernel_cache
    if _compiled_kernel_cache is not None:
        return _compiled_kernel_cache

    a_ptr = make_ptr(ab_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)
    b1_ptr = make_ptr(ab_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)
    b2_ptr = make_ptr(ab_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)

    sfa_ptr = make_ptr(sf_dtype, 0, cute.AddressSpace.gmem, assumed_align=32)
    sfb1_ptr = make_ptr(sf_dtype, 0, cute.AddressSpace.gmem, assumed_align=32)
    sfb2_ptr = make_ptr(sf_dtype, 0, cute.AddressSpace.gmem, assumed_align=32)

    c_ptr = make_ptr(c_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)

    _compiled_kernel_cache = cute.compile(
        dual_gemm_launch,
        a_ptr, b1_ptr, b2_ptr,
        sfa_ptr, sfb1_ptr, sfb2_ptr,
        c_ptr,
        (0, 0, 0, 0),
    )
    return _compiled_kernel_cache


def custom_kernel(data: input_t) -> output_t:
    a, b1, b2, _, _, _, sfa_permuted, sfb1_permuted, sfb2_permuted, c = data

    compiled = compile_kernel()

    _, k_half, _ = a.shape
    m, n, l = c.shape
    k = k_half * 2

    a_ptr = make_ptr(ab_dtype, a.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)
    b1_ptr = make_ptr(ab_dtype, b1.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)
    b2_ptr = make_ptr(ab_dtype, b2.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)

    sfa_ptr = make_ptr(sf_dtype, sfa_permuted.data_ptr(), cute.AddressSpace.gmem, assumed_align=32)
    sfb1_ptr = make_ptr(sf_dtype, sfb1_permuted.data_ptr(), cute.AddressSpace.gmem, assumed_align=32)
    sfb2_ptr = make_ptr(sf_dtype, sfb2_permuted.data_ptr(), cute.AddressSpace.gmem, assumed_align=32)

    c_ptr = make_ptr(c_dtype, c.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)

    compiled(a_ptr, b1_ptr, b2_ptr, sfa_ptr, sfb1_ptr, sfb2_ptr, c_ptr, (m, n, k, l))
    return c