# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang


import torch
import triton
import triton.language as tl

from fla.ops.utils import prepare_chunk_indices
from fla.ops.utils.op import exp
from fla.utils import autotune_cache_kwargs, check_shared_mem

BK_LIST = [32, 64] if check_shared_mem() else [16, 32]
BV_LIST = [64, 128] if check_shared_mem('ampere') else [16, 32]


@triton.heuristics({
    'IS_VARLEN': lambda args: args['cu_seqlens'] is not None,
})
@triton.autotune(
    configs=[
        triton.Config({'BK': BK, 'BV': BV}, num_warps=num_warps, num_stages=num_stages)
        for BK in BK_LIST
        for BV in BV_LIST
        for num_warps in [2, 4, 8]
        for num_stages in [2, 3, 4]
    ],
    key=['BT'],
    **autotune_cache_kwargs,
)
@triton.jit(do_not_specialize=['T'])
def chunk_kda_bwd_kernel_inter(
    q,
    k,
    v,
    g,
    h,
    do,
    dh,
    dq,
    dk,
    dv,
    dw,
    dg,
    cu_seqlens,
    chunk_indices,
    scale,
    T,
    H: tl.constexpr,
    K: tl.constexpr,
    V: tl.constexpr,
    BT: tl.constexpr,
    BK: tl.constexpr,
    BV: tl.constexpr,
    IS_VARLEN: tl.constexpr,
):
    i_k, i_t, i_bh = tl.program_id(0), tl.program_id(1), tl.program_id(2)
    i_b, i_h = i_bh // H, i_bh % H
    if IS_VARLEN:
        i_tg = i_t
        i_n, i_t = tl.load(chunk_indices + i_t * 2).to(tl.int32), tl.load(chunk_indices + i_t * 2 + 1).to(tl.int32)
        bos, eos = tl.load(cu_seqlens + i_n).to(tl.int32), tl.load(cu_seqlens + i_n + 1).to(tl.int32)
        T = eos - bos
        NT = tl.cdiv(T, BT)
    else:
        NT = tl.cdiv(T, BT)
        i_tg = i_b * NT + i_t
        bos, eos = i_b * T, i_b * T + T
    o_k = i_k * BK + tl.arange(0, BK)
    m_k = o_k < K

    q += (bos * H + i_h) * K
    k += (bos * H + i_h) * K
    v += (bos * H + i_h) * V
    g += (bos * H + i_h) * K
    h += (i_tg * H + i_h) * K*V
    do += (bos * H + i_h) * V
    dh += (i_tg * H + i_h) * K*V
    dq += (bos * H + i_h) * K
    dk += (bos * H + i_h) * K
    dw += (bos * H + i_h) * K
    dv += (bos * H + i_h) * V
    dg += (bos * H + i_h) * K

    p_g = tl.make_block_ptr(g, (T, K), (H*K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0))
    b_g = tl.load(p_g, boundary_check=(0, 1))
    p_gn = g + (min(T, i_t * BT + BT) - 1) * H*K + o_k
    b_gn = tl.load(p_gn, mask=m_k, other=0)
    b_dq = tl.zeros([BT, BK], dtype=tl.float32)
    b_dk = tl.zeros([BT, BK], dtype=tl.float32)
    b_dw = tl.zeros([BT, BK], dtype=tl.float32)
    b_dgk = tl.zeros([BK], dtype=tl.float32)

    for i_v in range(tl.cdiv(V, BV)):
        p_v = tl.make_block_ptr(v, (T, V), (H*V, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0))
        p_do = tl.make_block_ptr(do, (T, V), (H*V, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0))
        p_h = tl.make_block_ptr(h, (V, K), (1, V), (i_v * BV, i_k * BK), (BV, BK), (0, 1))
        p_dh = tl.make_block_ptr(dh, (V, K), (1, V), (i_v * BV, i_k * BK), (BV, BK), (0, 1))
        # [BT, BV]
        b_v = tl.load(p_v, boundary_check=(0, 1))
        b_do = tl.load(p_do, boundary_check=(0, 1))
        # [BV, BK]
        b_h = tl.load(p_h, boundary_check=(0, 1))
        b_dh = tl.load(p_dh, boundary_check=(0, 1))

        # [BK]
        b_dgk += tl.sum(b_h * b_dh, axis=0)
        # [BT, BK]
        b_dq += tl.dot(b_do, b_h.to(b_do.dtype))
        b_dk += tl.dot(b_v, b_dh.to(b_v.dtype))

        p_dv = tl.make_block_ptr(dv, (T, V), (H*V, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0))
        b_dv = tl.load(p_dv, boundary_check=(0, 1))
        b_dw += tl.dot(b_dv.to(b_v.dtype), b_h.to(b_v.dtype))

    p_dw = tl.make_block_ptr(dw, (T, K), (H*K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0))
    tl.store(p_dw, -b_dw.to(p_dw.dtype.element_ty), boundary_check=(0, 1))

    b_dgk *= exp(b_gn)
    b_dq *= scale
    b_dq = b_dq * exp(b_g)
    b_dk = b_dk * exp(b_gn[None, :] - b_g)

    p_q = tl.make_block_ptr(q, (T, K), (H*K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0))
    p_k = tl.make_block_ptr(k, (T, K), (H*K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0))
    p_dq = tl.make_block_ptr(dq, (T, K), (H*K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0))
    p_dk = tl.make_block_ptr(dk, (T, K), (H*K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0))
    p_dg = tl.make_block_ptr(dg, (T, K), (H*K, 1), (i_t * BT, i_k * BK), (BT, BK), (1, 0))
    b_q = tl.load(p_q, boundary_check=(0, 1))
    b_k = tl.load(p_k, boundary_check=(0, 1))
    b_dgk += tl.sum(b_dk * b_k, axis=0)
    b_dg = b_q * b_dq - b_k * b_dk
    b_dg = b_dg - tl.cumsum(b_dg, axis=0) + tl.sum(b_dg, axis=0)[None, :] + b_dgk[None, :]

    tl.store(p_dq, b_dq.to(p_dq.dtype.element_ty), boundary_check=(0, 1))
    tl.store(p_dk, b_dk.to(p_dk.dtype.element_ty), boundary_check=(0, 1))
    tl.store(p_dg, b_dg.to(p_dg.dtype.element_ty), boundary_check=(0, 1))


def chunk_kda_bwd_dqkwg(
    q: torch.Tensor,
    k: torch.Tensor,
    w: torch.Tensor,
    v: torch.Tensor,
    h: torch.Tensor,
    g: torch.Tensor,
    do: torch.Tensor,
    dh: torch.Tensor,
    dv: torch.Tensor,
    scale: float | None = None,
    cu_seqlens: torch.LongTensor | None = None,
    chunk_size: int = 64,
):
    B, T, H, K, V = *k.shape, v.shape[-1]
    BT = min(chunk_size, max(16, triton.next_power_of_2(T)))

    chunk_indices = prepare_chunk_indices(cu_seqlens, chunk_size) if cu_seqlens is not None else None
    NT = triton.cdiv(T, BT) if cu_seqlens is None else len(chunk_indices)

    dq = torch.empty_like(q, dtype=torch.float)
    dk = torch.empty_like(k, dtype=torch.float)
    dw = torch.empty_like(w)
    dg = torch.empty_like(g)
    def grid(meta): return (triton.cdiv(K, meta['BK']), NT, B * H)
    chunk_kda_bwd_kernel_inter[grid](
        q=q,
        k=k,
        v=v,
        g=g,
        h=h,
        do=do,
        dh=dh,
        dq=dq,
        dk=dk,
        dv=dv,
        dw=dw,
        dg=dg,
        cu_seqlens=cu_seqlens,
        chunk_indices=chunk_indices,
        scale=scale,
        T=T,
        H=H,
        K=K,
        V=V,
        BT=BT,
    )
    return dq, dk, dw, dg