"""
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)
"""

import torch

import triton
import triton.language as tl


@triton.jit
def _fwd_kernel(
    Q, K, V, sm_scale,
    L, M,
    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,
):
    start_m = tl.program_id(0)
    off_hz = tl.program_id(1)
    # initialize offsets
    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
    offs_n = tl.arange(0, BLOCK_N)
    offs_d = tl.arange(0, BLOCK_DMODEL)
    off_q = off_hz * stride_qh + offs_m[:, None] * stride_qm + offs_d[None, :] * stride_qk
    off_k = off_hz * stride_qh + offs_n[None, :] * stride_kn + offs_d[:, None] * stride_kk
    off_v = off_hz * stride_qh + offs_n[:, None] * stride_qm + offs_d[None, :] * stride_qk
    # Initialize pointers to Q, K, V
    q_ptrs = Q + off_q
    k_ptrs = K + off_k
    v_ptrs = V + off_v
    # initialize pointer to m and l
    m_prev = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
    l_prev = tl.zeros([BLOCK_M], dtype=tl.float32)
    acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
    # load q: it will stay in SRAM throughout
    q = tl.load(q_ptrs)
    # loop over k, v and update accumulator
    for start_n in range(0, (start_m + 1) * BLOCK_M, BLOCK_N):
        # -- compute qk ----
        k = tl.load(k_ptrs)
        qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
        qk += tl.dot(q, k)
        qk *= sm_scale
        qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk, float("-inf"))
        # compute new m
        m_curr = tl.maximum(tl.max(qk, 1), m_prev)
        # correct old l
        l_prev *= tl.exp(m_prev - m_curr)
        # attention weights
        p = tl.exp(qk - m_curr[:, None])
        l_curr = tl.sum(p, 1) + l_prev
        # rescale operands of matmuls
        l_rcp = 1. / l_curr
        p *= l_rcp
        acc *= (l_prev * l_rcp)[:, None]
        # update acc
        p = p.to(tl.float16)
        v = tl.load(v_ptrs)
        acc += tl.dot(p, v)
        # update m_i and l_i
        l_prev = l_curr
        m_prev = m_curr
        # update pointers
        k_ptrs += BLOCK_N * stride_kn
        v_ptrs += BLOCK_N * stride_vk
    # rematerialize offsets to save registers
    start_m = tl.program_id(0)
    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
    # write back l and m
    l_ptrs = L + off_hz * N_CTX + offs_m
    m_ptrs = M + off_hz * N_CTX + offs_m
    tl.store(l_ptrs, l_prev)
    tl.store(m_ptrs, m_prev)
    # initialize pointers to output
    offs_n = tl.arange(0, BLOCK_DMODEL)
    off_o = off_hz * stride_oh + offs_m[:, None] * stride_om + offs_n[None, :] * stride_on
    out_ptrs = Out + off_o
    tl.store(out_ptrs, acc)


@triton.jit
def _bwd_preprocess(
    Out, DO, L,
    NewDO, 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)
    denom = tl.load(L + off_m).to(tl.float32)
    # compute
    do = do / denom[:, None]
    delta = tl.sum(o * do, axis=1)
    # write-back
    tl.store(NewDO + off_m[:, None] * D_HEAD + off_n[None, :], do)
    tl.store(Delta + off_m, delta)


@triton.jit
def _bwd_kernel(
    Q, K, V, sm_scale, Out, DO,
    DQ, DK, DV,
    L, M,
    D,
    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,
    num_block,
    BLOCK_M: tl.constexpr, BLOCK_DMODEL: tl.constexpr,
    BLOCK_N: tl.constexpr,
):
    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_qz + off_h * stride_qh
    V += off_z * stride_qz + off_h * stride_qh
    DO += off_z * stride_qz + off_h * stride_qh
    DQ += off_z * stride_qz + off_h * stride_qh
    DK += off_z * stride_qz + off_h * stride_qh
    DV += off_z * stride_qz + off_h * stride_qh
    for start_n in range(0, num_block):
        lo = start_n * BLOCK_M
        # 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_qm + offs_k[None, :] * stride_qk)
        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
        m_ptrs = M + 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
            qk = tl.dot(q, tl.trans(k))
            qk = tl.where(offs_m_curr[:, None] >= (offs_n[None, :]), qk, float("-inf"))
            m = tl.load(m_ptrs + offs_m_curr)
            p = tl.exp(qk * sm_scale - m[:, None])
            # compute dv
            do = tl.load(do_ptrs)
            dv += tl.dot(tl.trans(p.to(tl.float16)), do)
            # 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))
            # compute ds = p * (dp - delta[:, None])
            ds = p * dp * sm_scale
            # compute dk = dot(ds.T, q)
            dk += tl.dot(tl.trans(ds.to(tl.float16)), q)
            # compute dq
            dq = tl.load(dq_ptrs)
            dq += tl.dot(ds.to(tl.float16), k)
            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_qm + offs_k[None, :] * stride_qk)
        dk_ptrs = DK + (offs_n[:, None] * stride_kn + offs_k[None, :] * stride_kk)
        tl.store(dv_ptrs, dv)
        tl.store(dk_ptrs, dk)


class _attention(torch.autograd.Function):

    @staticmethod
    def forward(ctx, q, k, v, sm_scale):
        # 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 = 128
        # 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}
        assert Lk in {64}  # TODO: fix other cases
        o = torch.empty_like(q)
        grid = (triton.cdiv(q.shape[2], BLOCK), 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)
        m = 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, m,
            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, BLOCK_N=BLOCK,
            BLOCK_DMODEL=Lk, num_warps=num_warps,
            num_stages=2,
        )

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

    @staticmethod
    def backward(ctx, do):
        BLOCK = 128
        q, k, v, o, l, m = ctx.saved_tensors
        do = do.contiguous()
        dq = torch.zeros_like(q, dtype=torch.float32)
        dk = torch.empty_like(k)
        dv = torch.empty_like(v)
        do_scaled = torch.empty_like(do)
        delta = torch.empty_like(l)
        _bwd_preprocess[(ctx.grid[0] * ctx.grid[1], )](
            o, do, l,
            do_scaled, delta,
            BLOCK_M=BLOCK, D_HEAD=ctx.BLOCK_DMODEL,
        )
        _bwd_kernel[(ctx.grid[1],)](
            q, k, v, ctx.sm_scale,
            o, do_scaled,
            dq, dk, dv,
            l, m,
            delta,
            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],
            ctx.grid[0],
            BLOCK_M=BLOCK, BLOCK_N=BLOCK,
            BLOCK_DMODEL=ctx.BLOCK_DMODEL, num_warps=8,
            num_stages=1,
        )
        return dq, dk, dv, None


attention = _attention.apply