σ \ K]c@`ssdZddlmZmZmZddlZddlZddlZddlZddl m Z ddl Z ddl mZddlmZddlmZmZmZdd lmZyeWnek rΫeZnXd „Zd „Zd „Zd efd„ƒYZdefd„ƒYZ de j!fd„ƒYZ"dd„Z$dZ%d„Z&d„Z'de(d„Z)dS(sΫ 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. i(tprint_functiontabsolute_importtdivisionN(tc_void_p(tdriveri(tdevices(t dummyarrayttypest numpy_support(tHsaContextMismatchErrorcC`st|dtƒS(s#Check if an object is a HSA ndarrayt__hsa_ndarray__(tgetattrtFalse(tobj((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pytis_hsa_ndarraysc`sZtˆƒ‡fd†}|dtƒ|dtƒ|dtjƒ|dttfƒdS(s+Verify the HSA ndarray interface for an objc`sStˆ|ƒst|ƒ‚nttˆ|ƒ|ƒsOtd||fƒ‚ndS(Ns%s must be of type %s(thasattrtAttributeErrort isinstanceR (tattrttyp(R (s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyt requires_attr!stshapetstridestdtypetsizeN(trequire_hsa_ndarrayttupletnpRtinttlong(R R((R s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pytverify_hsa_ndarray_interfaces    cC`st|ƒstdƒ‚ndS(s8Raises ValueError if is_hsa_ndarray(obj) evaluates Falsesrequire an hsa ndarray objectN(Rt ValueError(R ((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyR-s tDeviceNDArrayBasecB`szeZdZeZeZdd„Zed„ƒZ ed„ƒZ ed„ƒZ ddd„Z ddd„Z d„ZRS( sGBase class for an on dGPU NDArray representation cf. numpy.ndarray cC`s―t|ttfƒr!|f}nt|ttfƒrB|f}nt|ƒ|_t|ƒ|jkrutdƒ‚ntjjd|||j ƒ|_ t |ƒ|_ t |ƒ|_ tj|ƒ|_ttj|j ƒƒ|_|jdkr“|dkrHddlm}||j |j |jj ƒ|_tjƒj|jƒ}q’t|ddƒ}|dkrutdƒ‚n|dks‡t‚||_nd}d|_||_dS(s Args ---- shape array shape. strides array strides. dtype data type as numpy.dtype. dgpu_data user provided device memory for the ndarray data buffer sstrides not match ndimi(t_memory_size_from_infot _hsa_memsize_s'dgpu_data as no _hsa_memsize_ attributeN(RRRtlentndimRRtArrayt from_desctitemsizet_dummyRRRRRtprodRtNonet numba.roc.apiR!t alloc_sizeRt get_contextt mempoolallocR tAssertionErrort dgpu_data(tselfRRRR0R!tsz((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyt__init__9s6       cC`s |jjS(N(R0tcontext(R1((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyt_contexthscC`s(tj|jƒ}tj||jdƒS(sn Magic attribute expected by Numba to get the numba type that represents this object. tA(Rt from_dtypeRRR%R$(R1R((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyt _numba_type_lscC`s'|jdkrtdƒS|jjSdS(s:Returns the ctypes pointer to the GPU data buffer iN(R0R*Rtdevice_ctypes_pointer(R1((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyR9us c C`sa|jdkrdS|dk rn|jdk rw|j|}}|j|jkrktd|d|ƒ‚qkqwn |j}|j}|dkrΰtjj ƒt |t ƒrΔtj |j|||ƒq]tj |j|||ƒn}t |t ƒr&tjd|jd|jd|d|d |d |ƒn7tjd|jdtjƒd|d|d |d |ƒdS( sήCopy `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. iNtexpecttgottdst_ctxtsrc_ctxtdsttsrcRtstream(RR*R0R5tunproxyR R,t_driverthsat implicit_syncRt DeviceNDArrayt dGPU_to_dGPUt host_to_dGPUtasync_dGPU_to_dGPUtasync_host_to_dGPURtget_cpu_context(R1taryR@R4R:R;R2((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pytcopy_to_device~s.       c C`s=|dkr-tjd|jdtjƒ}nμ|j|jkrNtdƒ‚n|j|jkr¬ddf}|j|ko‡|j|ks¬td|j|jfƒ‚q¬n|j|jkrd|jj ff}|j|koξ|j|kstd|j|jfƒ‚qn|}|jdks4t dƒ‚|j }|jdkrΑ|j}|dkr‡t j jƒt j||||ƒqΑt jd tjƒd |j d |d |d |d|ƒn|dkr3|jdkrtjd|jd|jd|ƒ}q9tjd|jd|jd|jd|ƒ}n|}|S(sήCopy ``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() RRsincompatible dtypeis&incompatible shape; device %s; host %ss(incompatible strides; device %s; host %sisNegative memory sizeR<R=R>R?RR@tbufferRN((i((R*RtemptyR,tbyteRt TypeErrorRRR'R/R5RBRCRDt dGPU_to_hosttasync_dGPU_to_hostRRJRtndarray(R1RKR@thostaryt scalshapest scalstridesR4R2((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyt copy_to_host―sF !         cC`s|jS(sEReturns a device memory object that is used as the argument. (R0(R1((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyt as_hsa_argφsN(t__name__t __module__t__doc__tTruet__hsa_memory__R R*R3tpropertyR5R8R9RLRWRX(((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyR 3s /  1GREcB`s5eZdZd„Zd„Zd„Zdd„ZRS(s An on-dGPU array type cC`s |jjS(sA Return true if the array is Fortran-contiguous. (R(t is_f_contig(R1((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pytis_f_contiguousscC`s |jjS(s; Return true if the array is C-contiguous. (R(t is_c_contig(R1((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pytis_c_contiguoussc O`sιt|ƒdkr8t|dttfƒr8|d}nt|ƒ}||jkr~|d|jd|jd|jd|jƒS|j j ||Ž\}}||j j gkrΩ|d|jd|jd|jd|jƒSt dƒ‚dS( sΆ Reshape the array without changing its contents, similarly to :meth:`numpy.ndarray.reshape`. Example:: d_arr = d_arr.reshape(20, 50, order='F') iiRRRR0soperation requires copyingN( R#RRtlistttypeRRRR0R(treshapetextenttNotImplementedError(R1tnewshapetkwstclstnewarrtextents((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyRe s+  tCc C`swt|ƒ}|jjd|ƒ\}}||jjgkrg|d|jd|jd|jd|jƒStdƒ‚dS(sr Flatten the array without changing its contents, similar to :meth:`numpy.ndarray.ravel`. torderRRRR0soperation requires copyingN( RdR(travelRfRRRR0Rg(R1RnRjRkRl((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyRo$s  (RYRZR[R`RbReRo(((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyREόs    t HostArraycB`seZeZed„ƒZRS(cC`s|jjtƒS(N(tctypestdata_asR(R1((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyR97s(RYRZR\R]R^R9(((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyRp4scC`s@|jdkr!|jdƒ}nt|j|j|jd|ƒS(s/Create a DeviceNDArray object that is like ary.iiR0(R$ReRERRR(RKR0((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pytfrom_array_like<ss™Array contains non-contiguous buffer and cannot be transferred as a single memory region. Please ensure contiguous buffer with numpy .ascontiguousarray()cC`s7tj|jƒ}|j||j|}||jkS(N(RtargmaxRRtnbytes(RKtiR((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyt_single_bufferKscC`s`|jd r\|jd r\|jddkr=t|dƒSt|ƒrMtSttƒ‚ndS(Nt C_CONTIGUOUSt F_CONTIGUOUSi(tflagsRtsentry_contiguousRwR\Rterrmsg_contiguous_buffer(RK((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyR{Qs  cC`s_tj|ƒr|tfSt|ƒt|ƒ}|rQ|j|d|d|ƒn|tfSdS(sͺ 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. R@R4N(RBtis_device_memoryR R{RsRLR\(R R4R@tcopytdevobj((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyt auto_device^s   (*R[t __future__RRRtwarningstmathR~tweakrefRqRtnumpyRtnumba.roc.hsadrvRRBtRtnumbaRRRterrorR Rt NameErrorRRRRtobjectR RERSRpR*RsR|RwR{R\R€(((s;lib/python2.7/site-packages/numba/roc/hsadrv/devicearray.pyts4          Ι8