{{def_kernel("A", "B")}}
    M = {{size("A", -2)}}
    N = {{size("B", -1)}}
    K = {{size("A", -1)}}

    stride_aq = {{stride("A", 0)}}
    stride_am = {{stride("A", 1)}}
    stride_ak = {{stride("A", 2)}}

    stride_bq = {{stride("B", 0)}}
    stride_bk = {{stride("B", 1)}}
    stride_bn = {{stride("B", 2)}}

    # based on triton.ops.matmul
    pid = tl.program_id(0).to(INDEX_DTYPE)
    grid_m = (M + BLOCK_M - 1) // BLOCK_M
    grid_n = (N + BLOCK_N - 1) // BLOCK_N

    # re-order program ID for better L2 performance
    width = GROUP_M * grid_n
    group_id = pid // width
    group_size = min(grid_m - group_id * GROUP_M, GROUP_M)
    pid_m = group_id * GROUP_M + (pid % group_size)
    pid_n = (pid % width) // (group_size)
    tl.assume(pid_m >= 0)
    tl.assume(pid_n >= 0)

    rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
    rn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
    if (stride_am == 1 and stride_ak == M) or (stride_am == K and stride_ak == 1):
        ram = tl.max_contiguous(tl.multiple_of(rm % M, BLOCK_M), BLOCK_M)
    else:
        ram = rm % M
    if (stride_bk == 1 and stride_bn == K) or (stride_bk == N and stride_bn == 1):
        rbn = tl.max_contiguous(tl.multiple_of(rn % N, BLOCK_N), BLOCK_N)
    else:
        rbn = rn % N

    rk = tl.arange(0, BLOCK_K)

    idx_q = tl.program_id(1).to(INDEX_DTYPE)  # batch dimension for BMM
    A = A + (ram[:, None] * stride_am + rk[None, :] * stride_ak + idx_q*stride_aq)
    B = B + (rk[:, None] * stride_bk + rbn[None, :] * stride_bn + idx_q*stride_bq)

    acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=ACC_TYPE)
    for k in range(K, 0, -BLOCK_K):
        if EVEN_K:
            a = tl.load(A)
            b = tl.load(B)
        else:
            a = tl.load(A, mask=rk[None, :] < k, other=0.)
            b = tl.load(B, mask=rk[:, None] < k, other=0.)
        acc += tl.dot(a, b, allow_tf32=ALLOW_TF32)
        A += BLOCK_K * stride_ak
        B += BLOCK_K * stride_bk

    # rematerialize rm and rn to save registers
    rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
    rn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
    idx_q = tl.program_id(1).to(INDEX_DTYPE)  # batch dimension for BMM
    idx_m = rm[:, None]
    idx_n = rn[None, :]
    mask = (idx_m < M) & (idx_n < N)

    # inductor generates a suffix
    {{store_output(("idx_q", "idx_m", "idx_n"), "acc", "mask", val_shape=("BLOCK_M", "BLOCK_N"))}}