# Copyright (c) 2023-2026, Songlin Yang, Yu Zhang, Zhiyuan Li
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
# For a list of all contributors, visit:
#   https://github.com/fla-org/flash-linear-attention/graphs/contributors

import torch
import triton
import triton.language as tl

from fla.utils import autocast_custom_bwd, autocast_custom_fwd, input_guard

# Based: An Educational and Effective Sequence Mixer
# https://hazyresearch.stanford.edu/blog/2023-12-11-zoology2-based


@triton.jit(do_not_specialize=['T'])
def parallel_based_fwd_kernel(
    q,
    k,
    v,
    o,
    z,
    scale,
    T,
    B: tl.constexpr,
    H: tl.constexpr,
    K: tl.constexpr,
    V: tl.constexpr,
    BTL: tl.constexpr,
    BTS: tl.constexpr,
    BK: tl.constexpr,
    BV: tl.constexpr,
):
    # i_c: chunk index. used for sequence parallelism
    i_kv, i_c, i_bh = tl.program_id(0), tl.program_id(1).to(tl.int64), tl.program_id(2)
    NV = tl.cdiv(V, BV)
    i_k = i_kv // (NV)
    i_v = i_kv % (NV)

    o_t = i_c * BTL + tl.arange(0, BTL)
    o_kk = i_k * BK + tl.arange(0, BK)
    o_vv = i_v * BV + tl.arange(0, BV)
    m_t = o_t < T
    m_q = m_t[:, None] & (o_kk[None, :] < K)
    m_o = m_t[:, None] & (o_vv[None, :] < V)
    p_q = q + i_bh * T*K + o_t[:, None] * K + o_kk[None, :]

    # [BQ, BD] block Q, in the shared memory throughout the whole kernel
    b_q = tl.load(p_q, mask=m_q, other=0.0)
    b_q = (b_q * scale).to(b_q.dtype)
    b_o = tl.zeros([BTL, BV], dtype=tl.float32)
    b_z = tl.zeros([BTL], dtype=tl.float32)

    # Q block and K block have no overlap
    # no need for mask, thereby saving flops
    for i_s in range(0, i_c * BTL, BTS):
        o_s = i_s + tl.arange(0, BTS)
        m_k = (o_kk[:, None] < K) & (o_s[None, :] < T)
        m_v = (o_s[:, None] < T) & (o_vv[None, :] < V)
        p_k = k + i_bh * T*K + o_kk[:, None] + o_s[None, :] * K
        p_v = v + i_bh * T*V + o_s[:, None] * V + o_vv[None, :]
        # [BK, BTS]
        b_k = tl.load(p_k, mask=m_k, other=0.0)

        # [BTS, BV]
        b_v = tl.load(p_v, mask=m_v, other=0.0)
        # [BTL, BTS]
        b_s = tl.dot(b_q, (b_k), allow_tf32=False)
        b_s = 1 + b_s + 0.5 * b_s * b_s
        b_z += tl.sum(b_s, axis=1)

        # [BQ, BD]
        b_o = b_o + tl.dot(b_s.to(b_v.dtype), b_v, allow_tf32=False)

    # # rescale interchunk output
    tl.debug_barrier()
    o_q = tl.arange(0, BTL)
    # # sync threads, easy for compiler to optimize
    # tl.debug_barrier()

    o_k = tl.arange(0, BTS)
    # Q block and K block have overlap. masks required
    for i_s in range(i_c * BTL, (i_c + 1) * BTL, BTS):
        o_s = i_s + tl.arange(0, BTS)
        m_k = (o_kk[:, None] < K) & (o_s[None, :] < T)
        m_v = (o_s[:, None] < T) & (o_vv[None, :] < V)
        p_k = k + i_bh * T*K + o_kk[:, None] + o_s[None, :] * K
        p_v = v + i_bh * T*V + o_s[:, None] * V + o_vv[None, :]
        # [BK, BTS]
        b_k = tl.load(p_k, mask=m_k, other=0.0)
        # [BTS, BV]
        b_v = tl.load(p_v, mask=m_v, other=0.0)
        # [BTL, BTS]
        m_s = o_q[:, None] >= o_k[None, :]
        b_s = tl.dot(b_q, b_k, allow_tf32=False)
        b_s = 1 + b_s + 0.5 * b_s * b_s
        b_s = tl.where(m_s, b_s, 0)
        b_z += tl.sum(b_s, axis=1)
        # [BTL, BV]
        b_o += tl.dot(b_s.to(b_q.dtype), b_v, allow_tf32=False)

        o_k += BTS

    p_o = o + (i_bh + B * H * i_k) * T*V + o_t[:, None] * V + o_vv[None, :]
    p_z = z + (i_bh + B * H * i_k) * T + i_c * BTL + tl.arange(0, BTL)
    tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=m_o)
    tl.store(p_z, b_z.to(p_z.dtype.element_ty), mask=((i_c * BTL + tl.arange(0, BTL)) < T))


@triton.jit
def _parallel_based_bwd_dq(
    i_bh,
    i_c,
    i_k,
    i_v,
    q,
    k,
    v,
    do,
    dz,
    dq,
    scale,
    T,
    B: tl.constexpr,
    H: tl.constexpr,
    BTL: tl.constexpr,
    BTS: tl.constexpr,
    BK: tl.constexpr,
    BV: tl.constexpr,
    K: tl.constexpr,
    V: tl.constexpr,
):
    o_t = i_c * BTL + tl.arange(0, BTL)
    o_kk = i_k * BK + tl.arange(0, BK)
    o_vv = i_v * BV + tl.arange(0, BV)
    m_t = o_t < T
    m_q = m_t[:, None] & (o_kk[None, :] < K)
    m_o = m_t[:, None] & (o_vv[None, :] < V)
    p_do = do + i_bh * T*V + o_t[:, None] * V + o_vv[None, :]
    p_q = q + (i_bh) * T*K + o_t[:, None] * K + o_kk[None, :]
    b_q = tl.load(p_q, mask=m_q, other=0.0)
    b_q = (b_q * scale).to(b_q.dtype)

    b_do = tl.load(p_do, mask=m_o, other=0.0).to(b_q.dtype)
    b_dq = tl.zeros([BTL, BK], dtype=tl.float32)
    p_dz = dz + i_bh * T + i_c * BTL + tl.arange(0, BTL)
    b_dz = tl.load(p_dz, mask=(i_c * BTL + tl.arange(0, BTL)) < T)

    for i_s in range(0, i_c * BTL, BTS):
        o_s = i_s + tl.arange(0, BTS)
        m_k = (o_s[:, None] < T) & (o_kk[None, :] < K)
        m_v = (o_vv[:, None] < V) & (o_s[None, :] < T)
        p_k = k + i_bh * T*K + o_s[:, None] * K + o_kk[None, :]
        p_v = v + i_bh * T*V + o_vv[:, None] + o_s[None, :] * V
        # [BTS, BK]
        b_k = tl.load(p_k, mask=m_k, other=0.0)
        # [BV, BTS]
        b_v = tl.load(p_v, mask=m_v, other=0.0)
        # [BTL, BTS]
        b_ds = tl.dot(b_do, b_v, allow_tf32=False)
        if i_v == 0:
            b_ds += b_dz[:, None]
        else:
            b_ds = b_ds
        b_s = tl.dot(b_q, tl.trans(b_k), allow_tf32=False)
        # [BQ, BD]
        b_dq += tl.dot((b_ds * (1 + b_s)).to(b_v.dtype), b_k, allow_tf32=False)

    b_dq *= scale
    o_q = tl.arange(0, BTL)
    o_k = tl.arange(0, BTS)
    # Q block and K block have overlap. masks required
    for i_s in range(i_c * BTL, (i_c + 1) * BTL, BTS):
        o_s = i_s + tl.arange(0, BTS)
        m_k = (o_s[:, None] < T) & (o_kk[None, :] < K)
        m_v = (o_vv[:, None] < V) & (o_s[None, :] < T)
        p_k = k + i_bh * T*K + o_s[:, None] * K + o_kk[None, :]
        p_v = v + i_bh * T*V + o_vv[:, None] + o_s[None, :] * V
        # [BTS, BK]
        b_k = tl.load(p_k, mask=m_k, other=0.0)
        # [BV, BTS]
        b_v = tl.load(p_v, mask=m_v, other=0.0)
        # [BTL, BTS]
        m_s = o_q[:, None] >= o_k[None, :]
        b_ds = tl.dot(b_do, b_v, allow_tf32=False)
        if i_v == 0:
            b_ds += b_dz[:, None]
        else:
            b_ds = b_ds
        b_ds = tl.where(m_s, b_ds, 0) * scale
        b_s = tl.dot(b_q, tl.trans(b_k), allow_tf32=False)
        b_s = tl.where(m_s, b_s, 0)
        # [BTL, BK]
        b_dq += tl.dot((b_ds + b_ds * b_s).to(b_k.dtype), b_k, allow_tf32=False)
        o_k += BTS
    p_dq = dq + (i_bh + B * H * i_v) * T*K + o_t[:, None] * K + o_kk[None, :]
    tl.store(p_dq, b_dq.to(p_dq.dtype.element_ty), mask=m_q)
    return


@triton.jit
def _parallel_based_bwd_dkv(
    i_bh,
    i_c,
    i_k,
    i_v,
    q,
    k,
    v,
    do,
    dz,
    dk,
    dv,
    scale,
    T,
    B: tl.constexpr,
    H: tl.constexpr,
    BTL: tl.constexpr,
    BTS: tl.constexpr,
    BK: tl.constexpr,
    BV: tl.constexpr,
    K: tl.constexpr,
    V: tl.constexpr,
):
    # compute dk dv
    o_t = i_c * BTL + tl.arange(0, BTL)
    o_kk = i_k * BK + tl.arange(0, BK)
    o_vv = i_v * BV + tl.arange(0, BV)
    m_t = o_t < T
    m_k = m_t[:, None] & (o_kk[None, :] < K)
    m_v = m_t[:, None] & (o_vv[None, :] < V)
    p_k = k + i_bh * T*K + o_t[:, None] * K + o_kk[None, :]
    p_v = v + i_bh * T*V + o_t[:, None] * V + o_vv[None, :]
    b_k, b_v = tl.load(p_k, mask=m_k, other=0.0), tl.load(p_v, mask=m_v, other=0.0)
    b_dk, b_dv = tl.zeros([BTL, BK], dtype=tl.float32), tl.zeros([BTL, BV], dtype=tl.float32)

    for i in range((tl.cdiv(T, BTS) * BTS)-BTS, (i_c + 1) * BTL - BTS, -BTS):
        o_s = i + tl.arange(0, BTS)
        m_q = (o_kk[:, None] < K) & (o_s[None, :] < T)
        m_o = (o_vv[:, None] < V) & (o_s[None, :] < T)
        p_q = q + i_bh * T*K + o_kk[:, None] + o_s[None, :] * K
        p_do = do + i_bh * T*V + o_vv[:, None] + o_s[None, :] * V
        p_dz = dz + i_bh * T + i + tl.arange(0, BTS)
        b_q = tl.load(p_q, mask=m_q, other=0.0)  # [BK, BTS]
        b_do = tl.load(p_do, mask=m_o, other=0.0).to(b_q.dtype)  # [BV, BTS]
        b_dz = tl.load(p_dz, mask=(i + tl.arange(0, BTS)) < T)
        b_s = tl.dot(b_k.to(b_q.dtype), b_q, allow_tf32=False) * scale  # [BTL, BTS]
        b_s2 = 1 + b_s + 0.5 * b_s * b_s
        b_dv += tl.dot(b_s2.to(b_q.dtype), tl.trans(b_do), allow_tf32=False)
        b_ds = tl.dot(b_v, b_do, allow_tf32=False) * scale
        if i_v == 0:
            b_ds += b_dz[None, :] * scale
        else:
            b_ds = b_ds
        b_dk += tl.dot((b_ds + b_ds * b_s).to(b_q.dtype), tl.trans(b_q), allow_tf32=False)

    tl.debug_barrier()
    o_q, o_k = tl.arange(0, BTS), tl.arange(0, BTL)
    for i in range(i_c*BTL, (i_c+1)*BTL, BTS):
        o_s = i + tl.arange(0, BTS)
        m_q = (o_kk[:, None] < K) & (o_s[None, :] < T)
        m_o = (o_vv[:, None] < V) & (o_s[None, :] < T)
        p_q = q + i_bh * T*K + o_kk[:, None] + o_s[None, :] * K
        p_do = do + i_bh * T*V + o_vv[:, None] + o_s[None, :] * V
        p_dz = dz + i_bh * T + i + tl.arange(0, BTS)
        b_q = tl.load(p_q, mask=m_q, other=0.0)  # [BD, BQ]
        b_do = tl.load(p_do, mask=m_o, other=0.0).to(b_q.dtype)
        b_dz = tl.load(p_dz, mask=(i + tl.arange(0, BTS)) < T)
        # [BK, BQ]
        m_s = o_k[:, None] <= o_q[None, :]
        b_s = tl.dot(b_k, b_q, allow_tf32=False) * scale
        b_s2 = 1 + b_s + 0.5 * b_s * b_s
        b_s = tl.where(m_s, b_s, 0)
        b_s2 = tl.where(m_s, b_s2, 0)

        b_ds = tl.dot(b_v, b_do, allow_tf32=False)
        if i_v == 0:
            b_ds += b_dz[None, :]
        else:
            b_ds = b_ds
        b_ds = tl.where(m_s, b_ds, 0) * scale
        # [BK, BD]
        b_dv += tl.dot(b_s2.to(b_q.dtype), tl.trans(b_do), allow_tf32=False)
        b_dk += tl.dot((b_ds + b_ds * b_s).to(b_q.dtype), tl.trans(b_q), allow_tf32=False)
        o_q += BTS

    p_dk = dk + (i_bh + B * H * i_v) * T*K + o_t[:, None] * K + o_kk[None, :]
    p_dv = dv + (i_bh + B * H * i_k) * T*V + o_t[:, None] * V + o_vv[None, :]
    tl.store(p_dk, b_dk.to(p_dk.dtype.element_ty), mask=m_k)
    tl.store(p_dv, b_dv.to(p_dv.dtype.element_ty), mask=m_v)
    return


@triton.jit(do_not_specialize=['T'])
def parallel_based_bwd_kernel(
    q,
    k,
    v,
    do,
    dz,
    dq,
    dk,
    dv,
    scale,
    T,
    B: tl.constexpr,
    H: tl.constexpr,
    K: tl.constexpr,
    V: tl.constexpr,
    BTL: tl.constexpr,
    BTS: tl.constexpr,
    BK: tl.constexpr,
    BV: tl.constexpr,
):
    i_kv, i_c, i_bh = tl.program_id(0), tl.program_id(1).to(tl.int64), tl.program_id(2)
    NV = tl.cdiv(V, BV)
    i_k = i_kv // (NV)
    i_v = i_kv % NV
    _parallel_based_bwd_dq(
        i_bh, i_c, i_k, i_v,
        q, k, v, do, dz, dq,
        scale, T, B, H, BTL, BTS, BK, BV, K, V,
    )
    tl.debug_barrier()
    _parallel_based_bwd_dkv(
        i_bh, i_c, i_k, i_v,
        q, k, v, do, dz, dk, dv,
        scale, T, B, H, BTL, BTS, BK, BV, K, V,
    )


class ParallelBasedFunction(torch.autograd.Function):

    @staticmethod
    @input_guard
    @autocast_custom_fwd
    def forward(ctx, q, k, v, scale):
        BTL, BTS = 128, 32
        assert BTL % BTS == 0
        # assert q.shape[-1] % 16 == 0
        BK = min(128, max(triton.next_power_of_2(k.shape[-1]), 16))
        BV = min(128, max(triton.next_power_of_2(v.shape[-1]), 16))
        B, H, T, K, V = *k.shape, v.shape[-1]
        num_stages = 2
        num_warps = 4
        NK = triton.cdiv(K, BK)
        NV = triton.cdiv(V, BV)
        grid = (NK * NV, triton.cdiv(T, BTL), B * H)

        assert NK == 1, "will encounter some synchronization issue if not."

        o = torch.empty(NK, B, H, T, V, device=q.device)
        z = torch.empty(NK, B, H, T, device=q.device)
        parallel_based_fwd_kernel[grid](
            q, k, v, o, z,
            scale,
            B=B,
            H=H,
            T=T,
            K=K,
            V=V,
            BTL=BTL,
            BTS=BTS,
            BK=BK,
            BV=BV,
            num_warps=num_warps,
            num_stages=num_stages,
        )
        ctx.save_for_backward(q, k, v)
        ctx.scale = scale
        return o.sum(0).to(q.dtype), z.sum(0).to(q.dtype)

    @staticmethod
    @input_guard
    @autocast_custom_bwd
    def backward(ctx, do, dz):
        q, k, v = ctx.saved_tensors
        scale = ctx.scale
        BTL, BTS = 64, 32
        assert BTL % BTS == 0
        BK = min(128, max(triton.next_power_of_2(k.shape[-1]), 16))
        BV = min(128, max(triton.next_power_of_2(v.shape[-1]), 16))
        B, H, T, K, V = *k.shape, v.shape[-1]
        num_stages = 2
        num_warps = 4
        NK = triton.cdiv(K, BK)
        NV = triton.cdiv(V, BV)
        grid = (NK * NV, triton.cdiv(T, BTL), B * H)

        assert NK == 1, "will encounter some synchronization issue if not"

        dq = torch.empty(NV, B, H, T, K, dtype=q.dtype, device=q.device)
        dk = torch.empty(NV, B, H, T, K, dtype=q.dtype, device=q.device)
        dv = torch.empty(NK, B, H, T, V, dtype=q.dtype, device=q.device)

        parallel_based_bwd_kernel[grid](
            q, k, v, do, dz, dq, dk, dv,
            scale,
            B=B,
            H=H,
            T=T,
            K=K,
            V=V,
            BTL=BTL,
            BTS=BTS,
            BK=BK,
            BV=BV,
            num_warps=num_warps,
            num_stages=num_stages,
        )

        return dq.sum(0).to(q.dtype), dk.sum(0).to(k.dtype), dv.sum(0).to(v.dtype), None


triton_parallel_based = ParallelBasedFunction.apply


def parallel_based(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    scale: float | None = None,
    use_norm: bool = True,
    head_first: bool = False,
):
    assert q.shape[-1] <= 128, "only support feature dim up to 128"
    if scale is None:
        scale = q.shape[-1] ** -0.5
    if not head_first:
        q, k, v = map(lambda x: x.transpose(1, 2), (q, k, v))
    o, z = triton_parallel_based(q, k, v, scale)
    if use_norm:
        o = o / (z[..., None] + 1e-6)
    if not head_first:
        o = o.transpose(1, 2)
    return o.to(q.dtype)