x-jddlZddlZddlmZddlmZddlmZddlmZddl m Z ddl m Z gZ dd Zd Zd Zd ZGd dZdZdZdZddZdS)N)_C_ops)core)check_variable_and_dtype)EagerParamBase) LayerHelper)in_dynamic_or_pir_modec(trtj||St|dgddt d it }||j|j}| dd|gid|gid|i|S) aV This OP takes in the Tensor :attr:`x` and change it to output with aclFormat with int value. This API is only used for Ascend NPU. Args: x(Tensor): An input N-D Tensor with data type bool, float16, float32, float64, int32, int64, int16, int8, uint8. format(int): Storage data format of the output in aclFormat, default value is -1. Returns: Tensor: A Tensor with acl storage format on Ascend NPU. Examples: .. code-block:: pycon >>> # doctest: +REQUIRES(env:NPU) >>> import paddle >>> paddle.device.set_device('npu') >>> x = paddle.ones(shape=[6]) >>> y = paddle.incubate._npu_identity(x, 3) # ACL_FORMAT_NC1HWC0 = 3 >>> print(y.shape) paddle.Size([1, 1, 1, 1, 16]) x) boolint8uint8int16int32int64float16float32float64 npu_identity)dtype stop_gradientoutformat)typeinputsoutputsattrsN)r) rrrrrlocals"create_variable_for_type_inferencerr append_op)r rhelperrs c/var/www/html/banglarbhumi/venv/lib/python3.11/site-packages/paddle/incubate/tensor/manipulation.py _npu_identityr#s6"1f---        "88vxx8877'8   !:SENV$     cBt|trAtj|j}t|j|jfi|}|||}||fSt|tjr.tj }|||}||fSdS)zt Helper to create a new destination tensor and call 'func(dst, src)' which is either offload or reload. N) isinstancercopydeepcopy__dict__shaperpaddleTensorreager) src_tensorfuncstate new_paramtask new_varbases r"_load_reload_implr4Zs *n--! j122":#3Z5EOOOO tIz**$ J . .!j'')) tK,,D  !!r$ctj}tjrt jSt d|DrdSt jS)zf Constructs a new AsyncLoad object. It is used to load/reload data asynchronously on GPU. c3>K|]}tj|VdSNr+is_compiled_with_custom_device.0devs r" z$create_async_load..rs>  7:-c22      r$N)r+deviceget_all_custom_device_typeis_compiled_with_xpur XpuAsyncLoadany AsyncLoad)custom_devicess r"create_async_loadrEjsz ]==??N "$$  """   >L      t~r$c(tjS)zf Constructs a new AsyncLoad object. It is used to load/reload data asynchronously on XPU. )rrAr$r"create_xpu_async_loadrHzs    r$c$eZdZdZdZdZdZdS)_NoopAsyncTasku(A dummy Task for sync‐fallback on XPU.cdS)NTrGselfs r" is_completedz_NoopAsyncTask.is_completedstr$cdSr7rGrLs r"cpu_waitz_NoopAsyncTask.cpu_wait r$cdSr7rGrLs r"xpu_waitz_NoopAsyncTask.xpu_waitrQr$N)__name__ __module__ __qualname____doc__rNrPrSrGr$r"rJrJsG22        r$rJc*tjo(t|do|j}tj}td|Do*t|do|j|v}|s|r]| }tj | tj }|tfSt||jS)a Loads the source tensor into the destination tensor asynchronously. Args: src_tensor (EagerParamBase|paddle.Tensor): The source tensor. async_load (core.AsyncLoad): The AsyncLoad object. Returns: tuple: A tuple containing two elements: - dest_tensor (EagerParamBase|paddle.Tensor): The destination tensor. - task (Task): The task that loads the source tensor into the destination tensor. placec3>K|]}tj|VdSr7r8r:s r"r=z async_offload..s>  ;>F 1# 6 6      r$rY)r+r@hasattrrY is_xpu_placer>r?rBcustom_device_typecpu to_tensornumpyCPUPlacerJr4offload)r. async_load is_xpu_tensorrDis_custom_tensor host_tensorrs r" async_offloadrhs# #%% , J ( ( ,   ) ) + +]==??N   BP      D J ( ( D   / / 1 1^ C %(% nn&& {0022&/:K:KLLLN$$$$ Z); < <rs@ <<<<<<000000000000333333  9999x ! ! !             #=#=#=L<<<44 : : : : : :r$