#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
namespace op_api {
static constexpr int OK = 0;
static inline std::tuple<at::Tensor, at::Tensor> expand_outplace_npu(const at::Tensor& to_expand1,
const at::Tensor& to_expand2)
{
if (to_expand1.sizes().equals(to_expand2.sizes())) {
return std::make_tuple(to_expand1, to_expand2);
}
auto expanded_size = at::infer_size(to_expand1.sizes(), to_expand2.sizes());
return std::make_tuple(to_expand1.expand(expanded_size, true), to_expand2.expand(expanded_size, true));
}
static inline at::SmallVector<int64_t, SIZE> masked_select_npu_output_size(const at::Tensor& self,
const at::Tensor& mask)
{
at::Tensor maskCast;
at::Tensor selfCast;
std::tie(maskCast, selfCast) = expand_outplace_npu(mask, self);
auto outputSize = {maskCast.numel()};
return outputSize;
}
static at::Tensor exec_aclnn_masked_select(const at::Tensor& self, const at::Tensor& mask, at::Tensor& out)
{
static auto opApiFuncAddr = GetOpApiFuncAddr("aclGetViewShape");
TORCH_CHECK(opApiFuncAddr != nullptr, "GetOpApiFuncAddr failed.", OPS_ERROR(ErrCode::PTR));
using aclGetViewShapeFunc = int (*)(const aclTensor *tensor, int64_t **view_dims, uint64_t *view_dims_num);
auto aclGetViewShape = reinterpret_cast<aclGetViewShapeFunc>(opApiFuncAddr);
OP_EXEC_LOG(aclnnMaskedSelect, "EXEC_NPU_CMD_SYNC", self, mask, out);
auto npuAclParams = EXEC_NPU_CMD_SYNC(aclnnMaskedSelect, self, mask, out);
int64_t *view_dims = nullptr;
uint64_t view_dim_num = 0;
auto ret = aclGetViewShape(npuAclParams.Get<2>(), &view_dims, &view_dim_num);
TORCH_CHECK(ret == OK, "aclGetViewShape failed", OPS_ERROR(ErrCode::ACL));
at::SmallVector<int64_t, SIZE> outputShapeSize = {};
for (uint64_t i = 0; i < view_dim_num; i++) {
outputShapeSize.push_back(view_dims[i]);
}
out.resize_(outputShapeSize);
delete[] view_dims;
view_dims = nullptr;
return out;
}
at::Tensor masked_select(const at::Tensor& self, const at::Tensor& mask)
{
at::namedinference::compute_broadcast_outnames(self, mask);
DO_COMPATIBILITY(aclnnMaskedSelect, acl_op::masked_select(self, mask));
auto outputSize = masked_select_npu_output_size(self, mask);
at::Tensor out = at_npu::native::OpPreparation::apply_tensor_without_format(self, outputSize);
return exec_aclnn_masked_select(self, mask, out);
}
at::Tensor& masked_select_out(const at::Tensor& self, const at::Tensor& mask,
at::Tensor& result)
{
at::namedinference::compute_broadcast_outnames(self, mask);
DO_COMPATIBILITY(aclnnMaskedSelect, acl_op::masked_select_out(self, mask, result));
auto outputSize = masked_select_npu_output_size(self, mask);
at_npu::native::OpPreparation::check_tensor({self, mask}, result, self.scalar_type(), outputSize);
auto out = result.is_contiguous() ? result : result.contiguous();
auto output = exec_aclnn_masked_select(self, mask, out);
if (!result.is_contiguous()) {
result.copy_(output);
}
return result;
}}