#include "torch_npu/csrc/aten/CustomFunctions.h"
#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/utils/OpAdapter.h"
namespace acl_op {
using npu_preparation = at_npu::native::OpPreparation;
using npu_utils = at_npu::native::NpuUtils;
using tuple_tensor =
std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>;
using tensor_list = std::tuple<at::Tensor &, at::Tensor &, at::Tensor &, at::Tensor &, at::Tensor &, at::Tensor &>;
using tensor_list7 = std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>;
namespace {
tuple_tensor lstm_cell_npu_impl(const at::Tensor &input, const at::Tensor &w_ih, const at::Tensor &w_hh,
const at::Tensor &h, const at::Tensor &c, const at::Tensor &bias)
{
at::Tensor input_reshape = input.reshape({1, input.size(0), input.size(1)});
at::Tensor h_reshape = h.reshape({1, h.size(0), h.size(1)});
at::Tensor c_reshape = c.reshape({1, c.size(0), c.size(1)});
int64_t num_step = input_reshape.size(0);
int64_t batch_size = input_reshape.size(1);
int64_t hidden_size = w_hh.size(1) / 4;
at::SmallVector<int64_t, SIZE> output_size = {num_step, batch_size, hidden_size};
at::Tensor y_output = npu_preparation::apply_tensor(input_reshape, output_size);
at::Tensor h_output = npu_preparation::apply_tensor(input_reshape, output_size);
at::Tensor c_output = npu_preparation::apply_tensor(input_reshape, output_size);
at::Tensor i_output = npu_preparation::apply_tensor_with_format(input_reshape, output_size, ACL_FORMAT_FRACTAL_NZ);
at::Tensor j_output = npu_preparation::apply_tensor_with_format(input_reshape, output_size, ACL_FORMAT_FRACTAL_NZ);
at::Tensor f_output = npu_preparation::apply_tensor_with_format(input_reshape, output_size, ACL_FORMAT_FRACTAL_NZ);
at::Tensor o_output = npu_preparation::apply_tensor_with_format(input_reshape, output_size, ACL_FORMAT_FRACTAL_NZ);
at::Tensor tanhc = npu_preparation::apply_tensor_with_format(input_reshape, output_size, ACL_FORMAT_FRACTAL_NZ);
at_npu::native::OpCommand cmd;
cmd.Name("DynamicRNNV2").Input(input_reshape).Input(w_ih).Input(w_hh);
if (bias.defined()) {
cmd.Input(bias);
} else {
cmd.Input();
}
cmd.Input()
.Input(h_reshape)
.Input(c_reshape)
.Output(y_output)
.Output(h_output)
.Output(c_output)
.Output(i_output)
.Output(j_output)
.Output(f_output)
.Output(o_output)
.Output(tanhc)
.Attr("cell_type", (string) "LSTM")
.Attr("direction", (string) "UNIDIRECTIONAL")
.Attr("cell_depth", (int64_t)1)
.Attr("use_peephole", (bool)false)
.Attr("keep_prob", (float)1.0)
.Attr("cell_clip", (float)-1.0)
.Attr("num_proj", (int64_t)0)
.Attr("time_major", (bool)true)
.Attr("activation", (string) "tanh")
.Attr("forget_bias", (float)0.0)
.Attr("gate_order", (string) "ifco")
.Run();
at::Tensor h_out = h_output[0];
at::Tensor c_out = c_output[0];
return std::make_tuple(y_output, h_out, c_out, i_output, j_output, f_output, o_output, tanhc);
}
tensor_list lstm_cell_backward_npu_impl_nocheck(
at::Tensor &grad_input, at::Tensor &grad_wih, at::Tensor &grad_whh, at::Tensor &grad_bias, at::Tensor &grad_ht,
at::Tensor &grad_ct, const at::Tensor &grad_y, const at::Tensor &grad_h, const at::Tensor &grad_c,
const at::Tensor &input, const at::Tensor &w_ih, const at::Tensor &w_hh, const at::Tensor &h, const at::Tensor &c,
const at::Tensor &y_output, const at::Tensor &h_output, const at::Tensor &c_output, const at::Tensor &i,
const at::Tensor &j, const at::Tensor &f, const at::Tensor &o, const at::Tensor &tanhc)
{
at::Tensor seq_length = at::zeros({}, input.options());
at::Tensor mask = at::zeros({}, input.options().dtype(at::kByte));
at::Tensor wci = at::zeros({}, input.options());
at::Tensor wcf = at::zeros({}, input.options());
at::Tensor wco = at::zeros({}, input.options());
at_npu::native::OpCommand cmd;
cmd.Name("DynamicRNNV2Grad")
.Input(input)
.Input(w_ih)
.Input(w_hh)
.Input(y_output)
.Input(h)
.Input(c)
.Input(h_output)
.Input(c_output)
.Input(grad_y)
.Input(grad_h)
.Input(grad_c)
.Input(i)
.Input(j)
.Input(f)
.Input(o)
.Input(tanhc)
.Input(seq_length)
.Input(wci)
.Input(wcf)
.Input(wco)
.Input(mask)
.Output(grad_wih)
.Output(grad_whh)
.Output(grad_bias)
.Output(grad_input)
.Output(grad_ht)
.Output(grad_ct)
.Attr("cell_type", (string) "LSTM")
.Attr("direction", (string) "UNIDIRECTIONAL")
.Attr("cell_depth", (int64_t)1)
.Attr("use_peephole", (bool)false)
.Attr("keep_prob", (float)1.0)
.Attr("cell_clip", (float)-1.0)
.Attr("num_proj", (int64_t)0)
.Attr("time_major", (bool)true)
.Attr("activation", (string) "tanh")
.Attr("recurrent_activation", (string) "sigmoid")
.Attr("gate_order", (string) "ifjo")
.Attr("stateful", (bool)false)
.Attr("merge_mode", (string) "concat")
.Run();
return std::tie(grad_input, grad_wih, grad_whh, grad_bias, grad_ht, grad_ct);
}
}
tensor_list7 npu_lstm_cell_backward(const c10::optional<at::Tensor> &grad_y_opt,
const c10::optional<at::Tensor> &grad_h_opt,
const c10::optional<at::Tensor> &grad_c_opt, const at::Tensor &input,
const at::Tensor &w_ih, const at::Tensor &w_hh, const at::Tensor &h,
const at::Tensor &c, const at::Tensor &y_output, const at::Tensor &h_output,
const at::Tensor &c_output, const at::Tensor &i, const at::Tensor &j,
const at::Tensor &f, const at::Tensor &o, const at::Tensor &tanhc)
{
const at::Tensor &grad_y_opt_val = c10::value_or_else(grad_y_opt, [] { return at::Tensor(); });
const at::Tensor &grad_h_opt_val = c10::value_or_else(grad_h_opt, [] { return at::Tensor(); });
const at::Tensor &grad_c_opt_val = c10::value_or_else(grad_c_opt, [] { return at::Tensor(); });
auto grad_y = grad_y_opt_val.defined() ? grad_y_opt_val : at::zeros(y_output.sizes(), h.options());
auto grad_h = grad_h_opt_val.defined() ? grad_h_opt_val : at::zeros(h.sizes(), h_output.options());
auto grad_c = grad_c_opt_val.defined() ? grad_c_opt_val : at::zeros(c.sizes(), c_output.options());
int64_t hidden_size = y_output.size(2);
at::SmallVector<int64_t, SIZE> output_size = {4 * hidden_size};
at::Tensor grad_input = npu_preparation::apply_tensor(input);
at::Tensor grad_wih = npu_preparation::apply_tensor(w_ih);
at::Tensor grad_whh = npu_preparation::apply_tensor(w_hh);
at::Tensor grad_bias = npu_preparation::apply_tensor(i, output_size);
at::Tensor grad_ht = npu_preparation::apply_tensor(h);
at::Tensor grad_ct = npu_preparation::apply_tensor(c);
lstm_cell_backward_npu_impl_nocheck(grad_input, grad_wih, grad_whh, grad_bias, grad_ht, grad_ct, grad_y, grad_h,
grad_c, input, w_ih, w_hh, h, c, y_output, h_output, c_output, i, j, f, o,
tanhc);
return std::tie(grad_input, grad_wih, grad_whh, grad_bias, grad_bias, grad_ht, grad_ct);
}
tuple_tensor npu_lstm_cell(const at::Tensor &input, const at::Tensor &w_ih, const at::Tensor &w_hh, const at::Tensor &h,
const at::Tensor &c, const c10::optional<at::Tensor> &b_ih_opt,
const c10::optional<at::Tensor> &b_hh_opt)
{
const at::Tensor &b_ih = c10::value_or_else(b_ih_opt, [] { return at::Tensor(); });
const at::Tensor &b_hh = c10::value_or_else(b_hh_opt, [] { return at::Tensor(); });
at::Tensor bias;
if (b_ih.defined()) {
bias = at::add(b_ih, b_hh).to(input.dtype());
}
return lstm_cell_npu_impl(input, w_ih, w_hh, h, c, bias);
}
std::tuple<at::Tensor, at::Tensor> lstm_cell(const at::Tensor &input, at::TensorList hx, const at::Tensor &w_ih,
const at::Tensor &w_hh, const c10::optional<at::Tensor> &b_ih_opt,
const c10::optional<at::Tensor> &b_hh_opt)
{
at::Tensor weight_ih = w_ih.t().to(input.dtype());
at::Tensor weight_hh = w_hh.t().to(input.dtype());
at::Tensor h = hx[0];
at::Tensor c = hx[1];
auto result = at_npu::native::custom_ops::npu_lstm_cell(input, weight_ih, weight_hh, h, c, b_ih_opt, b_hh_opt);
return std::tuple<at::Tensor, at::Tensor>(std::get<1>(result), std::get<2>(result));
}
}