"""
Fused Attention
===============
This is a Triton implementation of the Flash Attention algorithm
(see: Dao et al., https://arxiv.org/pdf/2205.14135v2.pdf; Rabe and Staats https://arxiv.org/pdf/2112.05682v2.pdf)

Sequence Parallel implementation inspired by HazyResearch
(see https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/flash_attn_triton.py)
"""

import torch

from .. import cdiv, jit
from .. import language as tl


@jit
def _fwd_kernel(
    Q, K, V, sm_scale,
    L,
    Out,
    stride_qz, stride_qh, stride_qm, stride_qk,
    stride_kz, stride_kh, stride_kn, stride_kk,
    stride_vz, stride_vh, stride_vk, stride_vn,
    stride_oz, stride_oh, stride_om, stride_on,
    Z, H, N_CTX,
    BLOCK_M: tl.constexpr, BLOCK_DMODEL: tl.constexpr,
    BLOCK_N: tl.constexpr,
    IS_CAUSAL: tl.constexpr,
):
    start_m = tl.program_id(0)
    off_hz = tl.program_id(1)
    qvk_offset = off_hz * stride_qh
    Q_block_ptr = tl.make_block_ptr(
        base=Q + qvk_offset,
        shape=(N_CTX, BLOCK_DMODEL),
        strides=(stride_qm, stride_qk),
        offsets=(start_m * BLOCK_M, 0),
        block_shape=(BLOCK_M, BLOCK_DMODEL),
        order=(1, 0)
    )
    K_block_ptr = tl.make_block_ptr(
        base=K + qvk_offset,
        shape=(BLOCK_DMODEL, N_CTX),
        strides=(stride_kk, stride_kn),
        offsets=(0, 0),
        block_shape=(BLOCK_DMODEL, BLOCK_N),
        order=(0, 1)
    )
    V_block_ptr = tl.make_block_ptr(
        base=V + qvk_offset,
        shape=(N_CTX, BLOCK_DMODEL),
        strides=(stride_vk, stride_vn),
        offsets=(0, 0),
        block_shape=(BLOCK_N, BLOCK_DMODEL),
        order=(1, 0)
    )
    # initialize offsets
    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
    offs_n = tl.arange(0, BLOCK_N)
    # initialize pointer to m and l
    m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
    l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
    acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
    # credits to: Adam P. Goucher (https://github.com/apgoucher):
    # scale sm_scale by 1/log_2(e) and use
    # 2^x instead of exp in the loop because CSE and LICM
    # don't work as expected with `exp` in the loop
    qk_scale = sm_scale * 1.44269504
    # load q: it will stay in SRAM throughout
    q = tl.load(Q_block_ptr)
    q = (q * qk_scale).to(K.dtype.element_ty)
    lo = 0
    hi = (start_m + 1) * BLOCK_M if IS_CAUSAL else N_CTX
    for start_n in range(lo, hi, BLOCK_N):
        # -- load k, v --
        k = tl.load(K_block_ptr)
        v = tl.load(V_block_ptr)
        # -- compute qk ---
        qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
        if IS_CAUSAL:
            qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk, float("-inf"))
        qk += tl.dot(q, k, allow_tf32=True)
        # -- compute scaling constant ---
        m_i_new = tl.maximum(m_i, tl.max(qk, 1))
        alpha = tl.math.exp2(m_i - m_i_new)
        p = tl.math.exp2(qk - m_i_new[:, None])
        # -- scale and update acc --
        acc_scale = l_i * 0 + alpha  # workaround some compiler bug
        acc *= acc_scale[:, None]
        acc += tl.dot(p.to(V.dtype.element_ty), v, allow_tf32=True)
        # -- update m_i and l_i --
        l_i = l_i * alpha + tl.sum(p, 1)
        m_i = m_i_new
        # update pointers
        K_block_ptr = tl.advance(K_block_ptr, (0, BLOCK_N))
        V_block_ptr = tl.advance(V_block_ptr, (BLOCK_N, 0))
    # write back l and m
    acc = acc / l_i[:, None]
    l_ptrs = L + off_hz * N_CTX + offs_m
    tl.store(l_ptrs, m_i + tl.math.log2(l_i))
    # write back O
    O_block_ptr = tl.make_block_ptr(
        base=Out + qvk_offset,
        shape=(N_CTX, BLOCK_DMODEL),
        strides=(stride_om, stride_on),
        offsets=(start_m * BLOCK_M, 0),
        block_shape=(BLOCK_M, BLOCK_DMODEL),
        order=(1, 0)
    )
    tl.store(O_block_ptr, acc.to(K.dtype.element_ty))


@jit
def _bwd_preprocess(
    Out, DO,
    Delta,
    BLOCK_M: tl.constexpr, D_HEAD: tl.constexpr,
):
    off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
    off_n = tl.arange(0, D_HEAD)
    # load
    o = tl.load(Out + off_m[:, None] * D_HEAD + off_n[None, :]).to(tl.float32)
    do = tl.load(DO + off_m[:, None] * D_HEAD + off_n[None, :]).to(tl.float32)
    # compute
    delta = tl.sum(o * do, axis=1)
    # write-back
    tl.store(Delta + off_m, delta)


@jit
def _bwd_kernel_one_col_block(
    Q, K, V, sm_scale, qk_scale,
    Out, DO,
    DQ, DK, DV,
    L,
    D,
    stride_dqa, stride_qz, stride_qh, stride_qm, stride_qk,
    stride_kz, stride_kh, stride_kn, stride_kk,
    stride_vz, stride_vh, stride_vk, stride_vn,
    Z, H, N_CTX,
    off_hz, start_n, num_block,
    BLOCK_M: tl.constexpr, BLOCK_DMODEL: tl.constexpr,
    BLOCK_N: tl.constexpr,
    SEQUENCE_PARALLEL: tl.constexpr,
    CAUSAL: tl.constexpr,
):
    if SEQUENCE_PARALLEL:
        DQ += stride_dqa.to(tl.int64) * start_n
    if CAUSAL:
        lo = start_n * BLOCK_M
    else:
        lo = 0
    # initialize row/col offsets
    offs_qm = lo + tl.arange(0, BLOCK_M)
    offs_n = start_n * BLOCK_M + tl.arange(0, BLOCK_M)
    offs_m = tl.arange(0, BLOCK_N)
    offs_k = tl.arange(0, BLOCK_DMODEL)
    # initialize pointers to value-like data
    q_ptrs = Q + (offs_qm[:, None] * stride_qm + offs_k[None, :] * stride_qk)
    k_ptrs = K + (offs_n[:, None] * stride_kn + offs_k[None, :] * stride_kk)
    v_ptrs = V + (offs_n[:, None] * stride_vk + offs_k[None, :] * stride_vn)
    do_ptrs = DO + (offs_qm[:, None] * stride_qm + offs_k[None, :] * stride_qk)
    dq_ptrs = DQ + (offs_qm[:, None] * stride_qm + offs_k[None, :] * stride_qk)
    # pointer to row-wise quantities in value-like data
    D_ptrs = D + off_hz * N_CTX
    l_ptrs = L + off_hz * N_CTX
    # initialize dv amd dk
    dv = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
    dk = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
    # k and v stay in SRAM throughout
    k = tl.load(k_ptrs)
    v = tl.load(v_ptrs)
    # loop over rows
    for start_m in range(lo, num_block * BLOCK_M, BLOCK_M):
        offs_m_curr = start_m + offs_m
        # load q, k, v, do on-chip
        q = tl.load(q_ptrs)
        # recompute p = softmax(qk, dim=-1).T
        # NOTE: `do` is pre-divided by `l`; no normalization here
        if CAUSAL:
            qk = tl.where(offs_m_curr[:, None] >= (offs_n[None, :]), float(0.), float("-inf"))
        else:
            qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
        qk += tl.dot(q, tl.trans(k))
        qk *= qk_scale
        l_i = tl.load(l_ptrs + offs_m_curr)
        p = tl.math.exp2(qk - l_i[:, None])
        # compute dv
        do = tl.load(do_ptrs)
        dv += tl.dot(tl.trans(p.to(Q.dtype.element_ty)), do, allow_tf32=True)
        # compute dp = dot(v, do)
        Di = tl.load(D_ptrs + offs_m_curr)
        # dp = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32) - Di[:, None]
        dp = tl.dot(do, tl.trans(v), allow_tf32=True)
        # compute ds = p * (dp - delta[:, None])
        ds = (p * (dp - Di[:, None]) * sm_scale).to(Q.dtype.element_ty)
        # compute dk = dot(ds.T, q)
        dk += tl.dot(tl.trans(ds), q, allow_tf32=True)
        # compute dq
        if not SEQUENCE_PARALLEL:
            dq = tl.load(dq_ptrs)
            dq += tl.dot(ds, k, allow_tf32=True)
            tl.store(dq_ptrs, dq)
        elif SEQUENCE_PARALLEL:
            # dq = tl.dot(ds, k, allow_tf32=True)
            dq = tl.trans(tl.dot(tl.trans(k), tl.trans(ds), allow_tf32=True))
            tl.store(dq_ptrs, dq)

        # increment pointers
        dq_ptrs += BLOCK_M * stride_qm
        q_ptrs += BLOCK_M * stride_qm
        do_ptrs += BLOCK_M * stride_qm
    # write-back
    dv_ptrs = DV + (offs_n[:, None] * stride_vk + offs_k[None, :] * stride_vn)
    dk_ptrs = DK + (offs_n[:, None] * stride_kn + offs_k[None, :] * stride_kk)
    tl.store(dv_ptrs, dv)
    tl.store(dk_ptrs, dk)


@jit
def _bwd_kernel(
    # fmt: off
    Q, K, V, sm_scale,
    Out, DO,
    DQ, DK, DV,
    L,
    D,
    stride_dqa, stride_qz, stride_qh, stride_qm, stride_qk,
    stride_kz, stride_kh, stride_kn, stride_kk,
    stride_vz, stride_vh, stride_vk, stride_vn,
    Z, H, N_CTX,
    BLOCK_M: tl.constexpr, BLOCK_DMODEL: tl.constexpr,
    BLOCK_N: tl.constexpr,
    SEQUENCE_PARALLEL: tl.constexpr,
    CAUSAL: tl.constexpr,
    # fmt: on
):
    qk_scale = sm_scale * 1.44269504
    off_hz = tl.program_id(0)
    off_z = off_hz // H
    off_h = off_hz % H
    # offset pointers for batch/head
    Q += off_z * stride_qz + off_h * stride_qh
    K += off_z * stride_kz + off_h * stride_kh
    V += off_z * stride_vz + off_h * stride_vh
    DO += off_z * stride_qz + off_h * stride_qh
    DQ += off_z * stride_qz + off_h * stride_qh
    DK += off_z * stride_kz + off_h * stride_kh
    DV += off_z * stride_vz + off_h * stride_vh

    num_block_n = tl.cdiv(N_CTX, BLOCK_N)
    if not SEQUENCE_PARALLEL:
        for start_n in range(0, num_block_n):
            _bwd_kernel_one_col_block(
                Q, K, V, sm_scale, qk_scale, Out, DO,
                DQ, DK, DV,
                L,
                D,
                stride_dqa, stride_qz, stride_qh, stride_qm, stride_qk,
                stride_kz, stride_kh, stride_kn, stride_kk,
                stride_vz, stride_vh, stride_vk, stride_vn,
                Z, H, N_CTX,
                off_hz, start_n, num_block_n,
                BLOCK_M=BLOCK_M, BLOCK_DMODEL=BLOCK_DMODEL,
                BLOCK_N=BLOCK_N,
                SEQUENCE_PARALLEL=SEQUENCE_PARALLEL,
                CAUSAL=CAUSAL,
            )
    else:
        start_n = tl.program_id(1)
        _bwd_kernel_one_col_block(
            Q, K, V, sm_scale, qk_scale, Out, DO,
            DQ, DK, DV,
            L,
            D,
            stride_dqa, stride_qz, stride_qh, stride_qm, stride_qk,
            stride_kz, stride_kh, stride_kn, stride_kk,
            stride_vz, stride_vh, stride_vk, stride_vn,
            Z, H, N_CTX,
            off_hz, start_n, num_block_n,
            BLOCK_M=BLOCK_M, BLOCK_DMODEL=BLOCK_DMODEL,
            BLOCK_N=BLOCK_N,
            SEQUENCE_PARALLEL=SEQUENCE_PARALLEL,
            CAUSAL=CAUSAL,
        )


class _attention(torch.autograd.Function):

    @staticmethod
    def forward(ctx, q, k, v, causal, sm_scale, sequence_parallel=False):
        # only support for Ampere now
        capability = torch.cuda.get_device_capability()
        if capability[0] < 8:
            raise RuntimeError("Flash attention currently only supported for compute capability >= 80")
        BLOCK_M = 128
        BLOCK_N = 64
        # shape constraints
        Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
        assert Lq == Lk and Lk == Lv
        assert Lk in {16, 32, 64, 128}
        o = torch.empty_like(q)
        grid = (cdiv(q.shape[2], BLOCK_M), q.shape[0] * q.shape[1], 1)
        L = torch.empty((q.shape[0] * q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)
        num_warps = 4 if Lk <= 64 else 8
        _fwd_kernel[grid](
            q, k, v, sm_scale,
            L,
            o,
            q.stride(0), q.stride(1), q.stride(2), q.stride(3),
            k.stride(0), k.stride(1), k.stride(2), k.stride(3),
            v.stride(0), v.stride(1), v.stride(2), v.stride(3),
            o.stride(0), o.stride(1), o.stride(2), o.stride(3),
            q.shape[0], q.shape[1], q.shape[2],
            BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N, BLOCK_DMODEL=Lk,
            IS_CAUSAL=causal,
            num_warps=num_warps,
            num_stages=4)

        ctx.save_for_backward(q, k, v, o, L)
        ctx.grid = grid
        ctx.sm_scale = sm_scale
        ctx.BLOCK_DMODEL = Lk
        ctx.causal = causal
        ctx.sequence_parallel = sequence_parallel
        return o

    @staticmethod
    def backward(ctx, do):
        BLOCK = 128
        q, k, v, o, L = ctx.saved_tensors
        sequence_parallel = ctx.sequence_parallel
        seq_len_kv = k.shape[2]
        do = do.contiguous()
        if sequence_parallel:
            replicas = cdiv(seq_len_kv, BLOCK)
            new_dq_shape = (replicas,) + q.shape
            dq = torch.zeros(new_dq_shape, device=q.device, dtype=q.dtype)
        else:
            dq = torch.zeros_like(q, dtype=torch.float32)
        dk = torch.empty_like(k)
        dv = torch.empty_like(v)
        delta = torch.empty_like(L)
        _bwd_preprocess[(ctx.grid[0] * ctx.grid[1], )](
            o, do,
            delta,
            BLOCK_M=BLOCK, D_HEAD=ctx.BLOCK_DMODEL,
        )
        _bwd_kernel[(ctx.grid[1], cdiv(seq_len_kv, BLOCK) if sequence_parallel else 1)](
            q, k, v, ctx.sm_scale,
            o, do,
            dq, dk, dv,
            L,
            delta,
            o.numel(), q.stride(0), q.stride(1), q.stride(2), q.stride(3),
            k.stride(0), k.stride(1), k.stride(2), k.stride(3),
            v.stride(0), v.stride(1), v.stride(2), v.stride(3),
            q.shape[0], q.shape[1], q.shape[2],
            BLOCK_M=BLOCK, BLOCK_N=BLOCK,
            BLOCK_DMODEL=ctx.BLOCK_DMODEL,
            SEQUENCE_PARALLEL=sequence_parallel,
            CAUSAL=ctx.causal,
            num_warps=8,
            num_stages=1,
        )

        if len(dq.shape) == 5:
            dq = dq.sum(dim=0)
        return dq, dk, dv, None, None, None


attention = _attention.apply