# Copyright (c) 2021 Graphcore Ltd. All rights reserved.

import numpy as np
import popxl
import torch
from transformers.models.bert import BertConfig as HFConfig
from transformers.models.bert.modeling_bert import BertIntermediate, BertOutput

import popxl_addons as addons
from config import BertConfig
from modelling.feed_forward import FeedForward
from popxl.utils import to_numpy


def test_feed_forward_cmp_huggingface(test_config: BertConfig):
    torch.manual_seed(42)

    batch_size = test_config.execution.micro_batch_size
    seq_len = test_config.model.sequence_length
    hidden_size = test_config.model.hidden_size
    intermediate_size = hidden_size * 4

    # HuggingFace
    config = HFConfig(hidden_size=hidden_size, seq_len=seq_len, intermediate_size=intermediate_size)
    hf_intermediate_layer = BertIntermediate(config).eval()
    hf_output_layer = BertOutput(config).eval()

    # HF forward
    input_t = torch.rand((batch_size, seq_len, hidden_size), requires_grad=True)
    intermediate_output = hf_intermediate_layer(input_t)
    output_ = hf_output_layer(intermediate_output, input_t)

    # HF backwards
    grad_wrt = torch.rand(output_.shape)
    output_.backward(gradient=grad_wrt)
    input_grad_HF = input_t.grad.detach().numpy()
    output_HF = output_.detach().numpy()

    # popxl
    ir = popxl.Ir()
    ir.replication_factor = test_config.execution.data_parallel

    main = ir.main_graph

    with main:
        inputs_data, inputs_host_steam, inputs_tensors = zip(
            *[
                addons.host_load(input_t.reshape(-1, config.hidden_size), popxl.float32, name="input"),
            ]
        )
        input_t = inputs_tensors[0]
        args, ff_graph = FeedForward(test_config).create_graph(input_t)
        grad_ff_graph = addons.autodiff(ff_graph)

        ff = args.init("feedforward")
        layer = ff_graph.bind(ff)
        call_info = layer.call_with_info(input_t)
        act, *_ = call_info.outputs
        act_stream = addons.host_store(act)

        # Backwards
        gradient = popxl.constant(grad_wrt.reshape(act.shape).numpy().copy(), act.dtype, "gradient")
        grad_input, *_ = grad_ff_graph.call(gradient, args=grad_ff_graph.grad_graph_info.inputs_dict(call_info))

        grad_stream = addons.host_store(grad_input)

    # Map weights from huggingface
    weights = FeedForward.hf_mapping(test_config, ff, hf_intermediate_layer, hf_output_layer)

    inputs = dict(zip(inputs_host_steam, inputs_data))

    ir.num_host_transfers = test_config.execution.device_iterations

    with popxl.Session(ir, "ipu_hw") as session:
        session.write_variables_data(weights)
        outs = session.run(inputs)

    np.testing.assert_almost_equal(output_HF, outs[act_stream].reshape(output_HF.shape), 3)
    np.testing.assert_almost_equal(input_grad_HF, outs[grad_stream].reshape(input_grad_HF.shape), 3)