# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

import torch.nn as nn
import math
import torch
if torch.__version__ >= '1.8':
    import torch_npu


class PositionalEncoding(nn.Module):
    """Positional encoding.

    Args:
        d_model: Embedding dimension.
        dropout_rate: Dropout rate.
        max_len: Maximum input length.
        reverse: Whether to reverse the input position.
    """

    def __init__(self, d_model, dropout_rate, max_len=5000, reverse=False):
        """Construct an PositionalEncoding object."""
        super(PositionalEncoding, self).__init__()
        self.d_model = d_model
        self.reverse = reverse
        self.xscale = math.sqrt(self.d_model)
        self.dropout = nn.Dropout(p=dropout_rate)
        self.pe = None
        self.extend_pe(torch.tensor(0.0).expand(1, max_len))

    def extend_pe(self, x):
        """Reset the positional encodings."""
        if self.pe is not None:
            if self.pe.size(1) >= x.size(1):
                if self.pe.dtype != x.dtype or self.pe.device != x.device:
                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
                return
        pe = torch.zeros(x.size(1), self.d_model)
        if self.reverse:
            position = torch.arange(
                x.size(1) - 1, -1, -1.0, dtype=torch.float32
            ).unsqueeze(1)
        else:
            position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, self.d_model, 2, dtype=torch.float32)
            * -(math.log(10000.0) / self.d_model)
        )
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0)
        self.pe = pe.to(device=x.device, dtype=x.dtype)

    def forward(self, x: torch.Tensor):
        """Add positional encoding.
        Args:
            x (torch.Tensor): Input tensor B X T X C
        Returns:
            torch.Tensor: Encoded tensor B X T X C
        """
        self.extend_pe(x)
        x = x * self.xscale + self.pe[:, : x.size(1)]
        return self.dropout(x)

class NpuSlice(torch.autograd.Function):
    @staticmethod
    def forward(ctx, input, x1, x2):
        B, c = input.shape
        ctx.params = (x1, x2)
        result = torch_npu.npu_indexing(input, [0, x1],
                                    [B, x2], [1, 1])
        return result
    @staticmethod
    def backward(ctx, grad):
        _, c = grad.shape
        x1, x2 = ctx.params
        pads = (0, 0, x1, c - x2)
        self_grad = torch_npu.npu_pad(grad, pads)
        return self_grad, None, None

npu_slice = NpuSlice.apply

class RelPositionalEncoding(nn.Module):
    """Relative positional encoding module (new implementation).

    Args:
        d_model: Embedding dimension.
        dropout_rate: Dropout rate.
        max_len: Maximum input length.
    """

    def __init__(self, max_len, d_model):
        """Construct an PositionalEncoding object."""
        super(RelPositionalEncoding, self).__init__()
        self.d_model = d_model
        self.pe = None
        self.extend_pe(torch.tensor(0.0).expand(1, max_len))

    def extend_pe(self, x):
        """Reset the positional encodings."""
        if self.pe is not None:
            # self.pe contains both positive and negative parts
            # the length of self.pe is 2 * input_len - 1
            if self.pe.size(1) >= x.size(1) * 2 - 1:
                if self.pe.dtype != x.dtype or self.pe.device != x.device:
                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
                return
        # Suppose `i` means to the position of query vecotr and `j` means the
        # position of key vector. We use position relative positions when keys
        # are to the left (i>j) and negative relative positions otherwise (i<j).
        pe_positive = torch.zeros(x.size(1), self.d_model)
        pe_negative = torch.zeros(x.size(1), self.d_model)
        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, self.d_model, 2, dtype=torch.float32)
            * -(math.log(10000.0) / self.d_model)
        )
        pe_positive[:, 0::2] = torch.sin(position * div_term)
        pe_positive[:, 1::2] = torch.cos(position * div_term)
        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)

        # Reserve the order of positive indices and concat both positive and
        # negative indices. This is used to support the shifting trick
        # as in https://arxiv.org/abs/1901.02860
        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
        pe_negative = pe_negative[1:].unsqueeze(0)
        pe = torch.cat([pe_positive, pe_negative], dim=1)
        self.pe = pe.to(device=x.device, dtype=x.dtype)

    def forward(self, x: torch.Tensor):
        """Add positional encoding.
        Args:
            x : Input tensor T X B X C.
        Returns:
            torch.Tensor: Encoded tensor T X B X C.

        """
        x = x.transpose(0, 1)  # Change TBC to BTC
        self.extend_pe(x)
        # pos_emb = self.pe[
        #     :,
        #     self.pe.size(1) // 2 - x.size(1) + 1 : self.pe.size(1) // 2 + x.size(1),
        # ]
        pos_emb = npu_slice(self.pe, self.pe.size(1) // 2 - x.size(1) + 1,
                            self.pe.size(1) // 2 + x.size(1))
        pos_emb = pos_emb.transpose(0, 1)  # change to TBC
        return pos_emb