B \j0@sdZddlmZmZmZddlZddlZddlZddlZddl m Z ddl Z ddl mZddlmZddlmZmZmZdd lmZyeWnek reZYnXd d Zd d ZddZGdddeZGdddeZ Gddde j!Z"d ddZ#dZ$ddZ%ddZ&d!ddZ'dS)"z A HSA dGPU backed ND Array is recognized by checking the __hsa_memory__ attribute on the object. If it exists and evaluate to True, it must define shape, strides, dtype and size attributes similar to a NumPy ndarray. )print_functionabsolute_importdivisionN)c_void_p)driver)devices) dummyarraytypes numpy_support)HsaContextMismatchErrorcCs t|ddS)z#Check if an object is a HSA ndarray__hsa_ndarray__F)getattr)objr;lib/python3.7/site-packages/numba/roc/hsadrv/devicearray.pyis_hsa_ndarraysrcsFtfdd}|dt|dt|dtj|dttfdS)z+Verify the HSA ndarray interface for an objcs6t|st|tt||s2td||fdS)Nz%s must be of type %s)hasattrAttributeError isinstancer)attrtyp)rrr requires_attr!s z3verify_hsa_ndarray_interface..requires_attrshapestridesdtypesizeN)require_hsa_ndarraytuplenprintlong)rrr)rrverify_hsa_ndarray_interfaces     r"cCst|stddS)z8Raises ValueError if is_hsa_ndarray(obj) evaluates Falsezrequire an hsa ndarray objectN)r ValueError)rrrrr-src@sbeZdZdZdZdZdddZeddZedd Z ed d Z dd d Z dddZ ddZ dS)DeviceNDArrayBasezGBase class for an on dGPU NDArray representation cf. numpy.ndarray TNcCst|ttfr|f}t|ttfr(|f}t||_t||jkrHtdtjd|||j |_ t ||_ t ||_ t||_tt|j |_|jdkr |dkrddlm}||j |j |jj |_t|j}n0t|dd}|dkrtd|dkst||_n d}d|_||_dS)a Args ---- shape array shape. strides array strides. dtype data type as numpy.dtype. dgpu_data user provided device memory for the ndarray data buffer zstrides not match ndimrN)_memory_size_from_infoZ _hsa_memsize_z'dgpu_data as no _hsa_memsize_ attribute)rr r!lenndimr#r ArrayZ from_descitemsize_dummyrrrrrZprodrZ numba.roc.apir% alloc_sizerZ get_contextZ mempoolallocrAssertionError dgpu_data)selfrrrr-r%szrrr__init__9s6         zDeviceNDArrayBase.__init__cCs|jjS)N)r-context)r.rrr_contexthszDeviceNDArrayBase._contextcCst|j}t||jdS)zn Magic attribute expected by Numba to get the numba type that represents this object. A)r Z from_dtyperr r(r')r.rrrr _numba_type_ls zDeviceNDArrayBase._numba_type_cCs|jdkrtdS|jjSdS)z:Returns the ctypes pointer to the GPU data buffer Nr)r-rdevice_ctypes_pointer)r.rrrr5us z'DeviceNDArrayBase.device_ctypes_pointercCs|jdkrdS|dk rF|jdk rL|j|}}|j|jkrLt||dn|j}|j}|dkrtjt |t rt |j|||qt |j|||nBt |t rtj |j|j||||dntj|jt||||ddS)zCopy `ary` to `self`. If `ary` is a HSA memory, perform a device-to-device transfer. Otherwise, perform a a host-to-device transfer. If `stream` is a stream object, an async copy to used. rN)expectgot)dst_ctxsrc_ctxdstsrcrstream)rr-r2Zunproxyr r+_driverhsa implicit_syncr DeviceNDArrayZ dGPU_to_dGPUZ host_to_dGPUZasync_dGPU_to_dGPUZasync_host_to_dGPUrget_cpu_context)r.aryr<r1r6r7r/rrrcopy_to_device~s.         z DeviceNDArrayBase.copy_to_devicecCsf|dkrtj|jtjd}n|j|jkr0td|j|jkrhd}|j|krT|j|kshtd|j|jf|j|jkrd|jjff}|j|kr|j|kstd|j|jf|}|jdkst d |j }|jdkr|j}|dkrt j t ||||nt jt|j ||||d |dkr^|jdkrDtj|j|j|d }ntj|j|j|j|d }n|}|S) aCopy ``self`` to ``ary`` or create a new Numpy ndarray if ``ary`` is ``None``. The transfer is synchronous: the function returns after the copy is finished. Always returns the host array. Example:: import numpy as np from numba import hsa arr = np.arange(1000) d_arr = hsa.to_device(arr) my_kernel[100, 100](d_arr) result_array = d_arr.copy_to_host() N)rrzincompatible dtype)r)rz&incompatible shape; device %s; host %srz(incompatible strides; device %s; host %srzNegative memory size)r8r9r:r;rr<)rrbuffer)rrrrD)remptyr+Zbyter TypeErrorrrr)r,r2r=r>r?Z dGPU_to_hostZasync_dGPU_to_hostrrArndarray)r.rBr<ZhostaryZ scalshapesZ scalstridesr1r/rrr copy_to_hostsF              zDeviceNDArrayBase.copy_to_hostcCs|jS)zEReturns a device memory object that is used as the argument. )r-)r.rrr as_hsa_argszDeviceNDArrayBase.as_hsa_arg)N)NN)NN)__name__ __module__ __qualname____doc____hsa_memory__r r0propertyr2r4r5rCrHrIrrrrr$3s /  1 Gr$c@s2eZdZdZddZddZddZd d d Zd S) r@z An on-dGPU array type cCs|jjS)zA Return true if the array is Fortran-contiguous. )r*Z is_f_contig)r.rrris_f_contiguousszDeviceNDArray.is_f_contiguouscCs|jjS)z; Return true if the array is C-contiguous. )r*Z is_c_contig)r.rrris_c_contiguousszDeviceNDArray.is_c_contiguouscOst|dkr&t|dttfr&|d}t|}||jkrP||j|j|j|jdS|j j ||\}}||j j gkr||j|j|j|jdSt ddS)z Reshape the array without changing its contents, similarly to :meth:`numpy.ndarray.reshape`. Example:: d_arr = d_arr.reshape(20, 50, order='F') rr)rrrr-zoperation requires copyingN) r&rrlisttyperrrr-r*reshapeextentNotImplementedError)r.ZnewshapeZkwsclsnewarrextentsrrrrT s   zDeviceNDArray.reshapeCcCsLt|}|jj|d\}}||jjgkr@||j|j|j|jdStddS)zr Flatten the array without changing its contents, similar to :meth:`numpy.ndarray.ravel`. )order)rrrr-zoperation requires copyingN) rSr*ravelrUrrrr-rV)r.r[rWrXrYrrrr\$s  zDeviceNDArray.ravelN)rZ)rJrKrLrMrPrQrTr\rrrrr@s r@c@seZdZdZeddZdS) HostArrayTcCs |jtS)N)ctypesZdata_asr)r.rrrr57szHostArray.device_ctypes_pointerN)rJrKrLrNrOr5rrrrr]4sr]cCs*|jdkr|d}t|j|j|j|dS)z/Create a DeviceNDArray object that is like ary.rr)r-)r'rTr@rrr)rBr-rrrfrom_array_like<s  r_zArray contains non-contiguous buffer and cannot be transferred as a single memory region. Please ensure contiguous buffer with numpy .ascontiguousarray()cCs*t|j}|j||j|}||jkS)N)rZargmaxrrnbytes)rBirrrr_single_bufferKs rbcCsF|jdsB|jdsB|jddkr.t|dSt|r:dSttdS)NZ C_CONTIGUOUSZ F_CONTIGUOUSrT)flagsrsentry_contiguousrbr#errmsg_contiguous_buffer)rBrrrrdQs  rdTcCsBt|r|dfSt|t|}|r6|j|||d|dfSdS)z Create a DeviceArray like obj and optionally copy data from host to device. If obj already represents device memory, it is returned and no copy is made. F)r<r1TN)r=Zis_device_memoryrdr_rC)rr1r<copyZdevobjrrr auto_device^s rg)N)NT)(rMZ __future__rrrwarningsZmathrfweakrefr^rZnumpyrZnumba.roc.hsadrvrr=rZnumbar r r errorr r! NameErrorr rr"robjectr$r@rGr]r_rerbrdrgrrrrs6     J8