#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_out_nocheck(
at::Tensor& result,
const at::Tensor& self,
int64_t dim)
{
c10::SmallVector<int64_t, N> dim_list = {dim};
at_npu::native::OpCommand cmd;
cmd.Name("SoftmaxV2")
.Input(self)
.Output(result)
.Attr("axes", dim_list)
.Run();
return result;
}
}
at::Tensor softmax(
const at::Tensor &self,
int64_t dim,
c10::optional<at::ScalarType> dtype)
{
auto result = [&]() {
at::NoNamesGuard guard;
at::Tensor converted = dtype.has_value() ? at_npu::native::custom_ops::_npu_dtype_cast(self, dtype.value()) : self;
return at::_softmax(converted, dim, false);
}();
at::namedinference::propagate_names(result, self);
return result;
}
at::Tensor softmax(
const at::Tensor &self,
at::Dimname dim,
c10::optional<at::ScalarType> dtype)
{
return acl_op::softmax(self, dimname_to_position(self, dim), dtype);
}
at::Tensor _softmax(const at::Tensor &self, int64_t dim, bool half_to_float)
{
at::Tensor result;
if (half_to_float) {
result = npu_preparation::apply_tensor(self, self.options().dtype(at::ScalarType::Float));
} else {
result = npu_preparation::apply_tensor(self);
}
c10::optional<at::ScalarType> dtype = result.scalar_type();
at::ScalarType dst_type;
if (dtype.has_value()) {
dst_type = dtype.value();
} else if (result.defined()) {
dst_type = result.scalar_type();
} else {
dst_type = self.scalar_type();
}
at::Tensor self_cast = dst_type == self.scalar_type() ?
self : at_npu::native::custom_ops::_npu_dtype_cast(self, dst_type);
softmax_out_nocheck(result, self_cast, dim);
return result;
}
at::Tensor& _softmax_out(
const at::Tensor& self,
int64_t dim,
bool half_to_float,
at::Tensor& out)
{
auto dst_type = half_to_float ? at::kFloat : self.scalar_type();
npu_preparation::CheckOut(
{self},
out,
npu_preparation::get_tensor_npu_format(out),
dst_type,
self.sizes());
auto self_dtype = self.scalar_type();
if (half_to_float) {
TORCH_CHECK(self_dtype == at::kHalf, "conversion is supported for Half type only" + OPS_ERROR(ErrCode::TYPE));
} else {
TORCH_CHECK(at::isFloatingType(self_dtype), "_softmax_npu not implemented for '", toString(self_dtype),
"'" + OPS_ERROR(ErrCode::NOT_SUPPORT));
}
at::Tensor self_cast = dst_type == self.scalar_type() ?
self : at_npu::native::custom_ops::_npu_dtype_cast(self, dst_type);
if (!npu_utils::check_match(&out)) {
at::Tensor contiguous_result = npu_utils::format_contiguous(out);
softmax_out_nocheck(contiguous_result, self_cast, dim);
npu_utils::format_fresh_view(out, contiguous_result);
} else {
softmax_out_nocheck(out, self_cast, dim);
}
return out;
}
}