#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;
namespace {
at::Tensor softmax_backward_out_nocheck(
at::Tensor& grad_input,
const at::Tensor& grad_output,
const at::Tensor& output,
int64_t dim,
at::ScalarType input_dtype) {
c10::SmallVector<int64_t, N> dim_list = {dim};
at_npu::native::OpCommand cmd;
cmd.Name("SoftmaxGrad")
.Input(output)
.Input(grad_output)
.Output(grad_input)
.Attr("axes", dim_list)
.Run();
return grad_input;
}
}
at::Tensor& _softmax_backward_data_out(
const at::Tensor& grad_output,
const at::Tensor& output,
int64_t dim,
at::ScalarType input_dtype,
at::Tensor& result) {
npu_preparation::CheckOut(
{grad_output, output},
result,
grad_output);
if (!npu_utils::check_match(&result)) {
at::Tensor contiguous_result = npu_utils::format_contiguous(result);
softmax_backward_out_nocheck(contiguous_result, grad_output, output, dim, input_dtype);
npu_utils::format_fresh_view(result, contiguous_result);
} else {
softmax_backward_out_nocheck(result, grad_output, output, dim, input_dtype);
}
return result;
}
at::Tensor _softmax_backward_data(
const at::Tensor& grad_output,
const at::Tensor& output,
int64_t dim,
at::ScalarType input_dtype) {
auto output_size = op_infer::input_same_output_size(grad_output);
at::Tensor temp_output = output;
if (npu_preparation::get_tensor_npu_format(temp_output) == ACL_FORMAT_NC1HWC0) {
at_npu::native::custom_ops::npu_format_cast_(temp_output, npu_preparation::get_tensor_npu_format(grad_output));
}
at::Tensor grad_input = npu_preparation::apply_tensor(temp_output, output_size);
softmax_backward_out_nocheck(grad_input, grad_output, temp_output, dim, input_dtype);
return grad_input;
}
}