#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& embedding_renorm_gather2d_nocheck(
at::Tensor& result,
const at::Tensor& self,
const at::Tensor& indices)
{
at_npu::native::OpCommand cmd;
cmd.Name("GatherV2D")
.Input(self)
.Input(indices)
.Output(result)
.Attr("axis", static_cast<int64_t>(0))
.Run();
return result;
}
at::Tensor& embedding_renorm_execute_nocheck(
at::Tensor& result,
const at::Tensor& self,
double max_norm,
double norm_type)
{
at_npu::native::OpCommand cmd;
cmd.Name("Renorm")
.Input(self)
.Output(result)
.Attr("p", static_cast<float>(norm_type))
.Attr("dim", static_cast<int64_t>(0))
.Attr("maxnorm", static_cast<float>(max_norm))
.Run();
return result;
}
at::Tensor& embedding_renorm_scatter_update_nocheck(
at::Tensor& result,
const at::Tensor& self,
const at::Tensor& indices,
const at::Tensor& update)
{
at_npu::native::OpCommand cmd;
cmd.Name("ScatterUpdate")
.Input(self)
.Input(indices)
.Input(update)
.Output(result)
.Attr("use_locking", false)
.Run();
return result;
}
at::Tensor& embedding_renorm_out_npu_nocheck(
at::Tensor& result,
const at::Tensor& self,
const at::Tensor& indices,
double max_norm,
double norm_type)
{
at::SmallVector<int64_t, SIZE> mid_size = {indices.size(0), self.size(1)};
at::Tensor mid_input = npu_preparation::apply_tensor(self, mid_size);
at::Tensor mid_output = npu_preparation::apply_tensor(self, mid_size);
embedding_renorm_gather2d_nocheck(mid_input, self, indices);
embedding_renorm_execute_nocheck(mid_output, mid_input, max_norm, norm_type);
auto num_indices = indices.numel();
at::Tensor input_indices;
if (num_indices - 1 == 0) {
input_indices = at_npu::native::custom_ops::_npu_dtype_cast(at::zeros({1}, self.options()), at::kLong);
} else {
input_indices = at_npu::native::custom_ops::_npu_dtype_cast(at::range(0, num_indices - 1, self.options()), at::kLong);
}
auto num_mid_output = mid_output.numel();
mid_output.resize_(num_mid_output);
at::Tensor scalar_out = npu_preparation::apply_tensor(self, {num_indices, 1});
embedding_renorm_gather2d_nocheck(scalar_out, mid_output, input_indices);
at::Tensor out_res = mid_input * scalar_out;
embedding_renorm_scatter_update_nocheck(result, self, indices, out_res);
return result;
}
}
at::Tensor& embedding_renorm_(
at::Tensor& self,
const at::Tensor& indices,
double max_norm,
double norm_type)
{
auto self_arg = at::TensorArg(self, "self", 1);
auto indices_arg = at::TensorArg(indices, "indices", 2);
at::checkDim("embedding_renorm_", self_arg, 2);
at::checkScalarType("embedding_renorm_", indices_arg, at::kLong);
auto num_indices = indices.numel();
TORCH_CHECK(num_indices >= 1, "indices.numel() must be greater than or equal to 1, but got ", num_indices,
OPS_ERROR(ErrCode::PARAM));
at::native::resize_(indices, num_indices);
npu_preparation::CheckMemory({self, indices}, {self});
if (!npu_utils::check_match(&self)) {
at::Tensor contiguous_self = npu_utils::format_contiguous(self);
embedding_renorm_out_npu_nocheck(contiguous_self, contiguous_self, indices, max_norm, norm_type);
npu_utils::format_fresh_view(self, contiguous_self);
} else {
embedding_renorm_out_npu_nocheck(self, self, indices, max_norm, norm_type);
}
return self;
}
}