#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
namespace {
};
std::tuple<at::Tensor, at::Tensor> npu_gelu_quant(
const at::Tensor &self,
const c10::optional<at::Tensor> &input_scale,
const c10::optional<at::Tensor> &input_offset,
c10::string_view approximate,
c10::string_view quant_mode,
c10::optional<int64_t> dst_type,
c10::string_view round_mode)
{
const at::Tensor &input_scale_value = c10::value_or_else(input_scale, [] { return at::Tensor(); });
const at::Tensor &input_offset_value = c10::value_or_else(input_offset, [] { return at::Tensor(); });
at::Tensor y;
at::Tensor out_scale;
aclDataType y_acltype;
if (!dst_type.has_value()) {
ASCEND_LOGI("[npu_gelu_quant]: Parameter(dst_type) is None, setting aclTensor y dtype to default: %s",
at_npu::native::AclDataTypeToString(aclDataType::ACL_INT8).c_str());
y_acltype = aclDataType::ACL_INT8;
} else {
ASCEND_LOGI("[npu_gelu_quant]: Getting aclTensor y dtype by Parameter(dst_type): %ld", dst_type.value());
y_acltype = c10_npu::GetAclDataType(dst_type.value());
ASCEND_LOGI("[npu_gelu_quant]: Setting aclTensor y dtype to: %s", at_npu::native::AclDataTypeToString(y_acltype).c_str());
}
at::ScalarType y_scalar_type = npu_preparation::convert_to_scalar_type(y_acltype);
auto y_shape = op_infer::array_to_small_vector(self.sizes());
y = npu_preparation::apply_tensor_without_format(y_shape, c10::dtype(y_scalar_type));
TensorWrapper y_wrapper = {y, y_acltype};
std::string approximate_str = std::string(approximate.data());
std::string quant_mode_str = std::string(quant_mode.data());
std::string round_mode_str = std::string(round_mode.data());
const char* approximate_char = approximate_str.c_str();
const char* quant_mode_char = quant_mode_str.c_str();
const char* round_mode_char = round_mode_str.c_str();
bool quant_mode_valid = (quant_mode_str == "dynamic" || quant_mode_str == "static");
TORCH_CHECK(quant_mode_valid,
"Parameter quant_mode must be 'dynamic' or 'static', got: " + quant_mode_str + OPS_ERROR(ErrCode::PARAM));
if (quant_mode_str == "dynamic") {
at::SmallVector<int64_t, op_infer::SIZE> scale_shape = {};
int scale_dim = self.dim() - 1;
int index = 0;
for (; index < scale_dim; ++index) {
scale_shape.push_back(self.size(index));
}
out_scale = npu_preparation::apply_tensor_without_format(scale_shape, c10::dtype(c10::ScalarType::Float));
TensorWrapper out_scale_wrapper = {out_scale, aclDataType::ACL_FLOAT};
EXEC_NPU_CMD(aclnnGeluQuant, self, input_scale_value, input_offset_value,
approximate_char, quant_mode_char, round_mode_char, y_acltype, y_wrapper, out_scale_wrapper);
return std::make_tuple(y, out_scale);
} else {
EXEC_NPU_CMD(aclnnGeluQuant, self, input_scale_value, input_offset_value,
approximate_char, quant_mode_char, round_mode_char, y_acltype, y_wrapper, out_scale);
return std::make_tuple(y, out_scale);
}
}
}