B \N: @sdZddlmZmZddlZddlZddlZddlZddlm Z ddl m Z ddl ZddlmmZddlmZddlmZddlmZddlmZmZmZmZdd ZeZ d d Z!Gd d d ej"Z#ddZ$Gdddej%Z&ddZ'e Z(y e Z)WnXe*k rTZ+z8dZ,e-e,e.e+ddl/m0Z0e0ddZ1e1Z)WddZ+[+XYnXda2da3ddZ4ddZ5dej6kZ7e8ej69ddZ:e7re:re5dS)a This file implements the code-generator for parallel-vectorize. ParallelUFunc is the platform independent base class for generating the thread dispatcher. This thread dispatcher launches threads that execute the generated function of UFuncCore. UFuncCore is subclassed to specialize for the input/output types. The actual workload is invoked inside the function generated by UFuncCore. UFuncCore also defines a work-stealing mechanism that allows idle threads to steal works from other threads. )print_functionabsolute_importN)RLock) ufuncbuilder)as_dtype)typesutilscgutilsconfigcCstj}|dkrtd|S)z* Gets the available thread count. z(Number of threads specified must be > 0.)r ZNUMBA_NUM_THREADS ValueError)tr6lib/python3.7/site-packages/numba/npyufunc/parallel.pyget_thread_countsrcstjd}tj||tjtjtjtjtjtjg}| d}| d}|j |dt |d} | d} t| | j\} } } }|}|}|}fddt|jd }tjtjgd gd }|j|d d}|tj|}fd d| | | |gD}||g|fdd||fD|||||||||| j | j fS)aWrap the original CPU ufunc/gufunc with a parallel dispatcher. This function will wrap gufuncs and ufuncs something like. Args ---- ctx numba's codegen context innerfunc llvm function of the original CPU gufunc sig type signature of the gufunc inner_ndim inner dimension of the gufunc (this is len(sig.args) in the case of a ufunc) Details ------- The kernel signature looks like this: void kernel(char **args, npy_intp *dimensions, npy_intp* steps, void* data) args - the input arrays + output arrays dimensions - the dimensions of the arrays steps - the step size for the array (this is like sizeof(type)) data - any additional data The parallel backend then stages multiple calls to this kernel concurrently across a number of threads. Practically, for each item of work, the backend duplicates `dimensions` and adjusts the first entry to reflect the size of the item of work, it also forms up an array of pointers into the args for offsets to read/write from/to with respect to its position in the items of work. This allows the same kernel to be used for each item of work, with simply adjusted reads/writes/domain sizes and is safe by virtue of the domain partitioning. NOTE: The execution backend is passed the requested thread count, but it can choose to ignore it (TBB)! Zparallelgufuncwrapperzparallel.gufunc.wrapperz.kernel.)namecs |S)N)Zbitcast)arg)builder byte_ptr_trr as_void_ptrssz(build_gufunc_kernel..as_void_ptrr numba_parallel_forcsg|] }|qSrr).0x)rrr sz'build_gufunc_kernel..csg|] }|qSrr)rr)intp_trrrs)!lcZTypeintZpointerZget_value_typerZintpZfunctionZvoidZcodegenZcreate_libraryZcreate_ir_moduleZ add_functionstrZappend_basic_blockZBuilderargsZget_python_apiZ gil_ensureZ save_threadlenZget_or_insert_functionZ get_constantZuintpZinttoptrZcallZrestore_threadZ gil_releaseZret_voidZ add_ir_moduleZadd_linking_libraryZget_pointer_to_functionr)libraryctx innerfuncsig inner_ndimZbyte_tZfntyZ wrapperlibmodZlfuncZbb_entryr"Z dimensionsZstepsdataZpyapiZ gil_stateZ thread_stateZ array_countZparallel_for_ty parallel_forZfnptrZ innerargsr)rrrrrbuild_gufunc_kernel,sB,              r,c@seZdZddZdS)ParallelUFuncBuilderc Csdt|j}|j}|j}|jj}t|||||}dd|jD}|t |j j j d} ||| fS)NcSsg|]}t|jjqSr)npdtypernum)rarrrrsz.ParallelUFuncBuilder.build..r)_launch_threadstarget_context signaturer$ZfndescZllvm_func_namebuild_ufunc_wrapperr"appendr.r/Z return_typerr0) selfcresr'r%r4r$fnameptr dtypenumsZ keepaliverrrbuildszParallelUFuncBuilder.buildN)__name__ __module__ __qualname__r<rrrrr-sr-cCs4tj||||d|d}t||||t|j\}}|S)NF)Zobjmoder8)rr5r,r#r")r$r%r9r4r8r&r:rrrrr5s   r5cs,eZdZddiffdd ZddZZS)ParallelGUFuncBuilderNFcs.|tddtt|j|||||ddS)NT)Znopython)py_funcr4identitycache targetoptions)updatedictsuperr@__init__)r7rAr4rBrCrD) __class__rrrHszParallelGUFuncBuilder.__init__cCsntt|j||j|j|jd\}}}g}x8|jjD],}t|t j rL|j }n|}| t |jq4W|||fS)zJ Returns (dtype numbers, function ptr, EnvironmentObject) )rC)r2build_gufunc_wrapperrAsinsoutrCr4r" isinstancerZArrayr/r6rr0)r7r8r:env wrapper_namer;r1Ztyrrrr<s zParallelGUFuncBuilder.build)r=r>r?rHr< __classcell__rr)rIrr@s r@cCsz|j}|j}|j}tj|||||d\}} } tdd|D} tdd|D} t| | B} t||||| \}}|| |fS)N)rCcss|]}|D] }|Vq qdS)Nr)rtermsymrrr sz'build_gufunc_wrapper..css|]}|D] }|Vq qdS)Nr)rrQrRrrrrSs)r$r3r4rrJsetr#r,)rAr8rKrLrCr$r%r4r&rNrOZsym_inZsym_outr(r:rrrrrJs rJa@Could not obtain multiprocessing lock due to OS level error: %s A likely cause of this problem is '/dev/shm' is missing orread-only such that necessary semaphores cannot be written. *** The responsibility of ensuring multiprocessing safe access to this initialization sequence/module import is deferred to the user! *** )contextmanagerccs dVdS)NrrrrrnopsrVFcCstdkrtdntSdS)zM Get the name of the threading layer in use for parallel CPU targets Nz#Threading layer is not initialized.)_threading_layerr rrrrthreading_layers rXc s6t&ttrdSddlm}m}m}ddfdd}ttj  }ddd g}d}t d k}t d k}t d d <dd<g} fdd} ||krވ|}|s|dkr| d n|dkr|r| d|} n|dkr~dg} | d |dkrnl|dkr*|r| d| dnF|dkr`|sD| d|rT| d| d nd} t| ||| \}} nH|dkr||\}} |s| d |r| dnd} t| ||s| | td|jtd|jtd|j|d||j}|t| adaWdQRXWdQRXdS)Nr) CFUNCTYPEc_void_pc_intcSsd}|dr6yddlm}Wqtk r2YqXnZ|drhyddlm}Wqtk rdYqXn(|drddlm}nd }t|||S) zZ Loads a specific threading layer backend based on string Ntbbr )tbbpoolomp)omppool workqueue)r`z/Unknown value specified for threading layer: %s) startswithrr] ImportErrorr_r`r )backendlibmsgrrrselect_known_backends     z-_launch_threads..select_known_backendcs0d}x"|D]}|}|dk r Pq Wd}||fS)z_ Selects from presented backends and returns the first working Nrr)Zbackendsrdrc)rfrrselect_from_backends,s z-_launch_threads..select_from_backendsr\r^r`ZDarwinZLinuxz3Intel TBB is required, try: $ conda/pip install tbbZTBBz?Intel OpenMP is required, try: $ conda/pip install intel-openmpZOSX_OMPcstd}d}t|dkrd}t|dkr4||d}t|dkrddfdd|D}|d |}t||dS) Nz&No threading layer could be loaded. %szHINT: %srrr z OR csg|] }|qSrr)rr) err_helpersrrrNsz<_launch_threads..raise_with_hint..z One of: %s)r#joinr )ZrequirederrmsgZhintmsgZhintZoptions)rhrrraise_with_hintFs    z(_launch_threads..raise_with_hint) threadsafeforksafesafernrlrmz+No threading layer available for purpose %sdefaultz7The threading layer requested '%s' is unknown to Numba.rdo_scheduling_signeddo_scheduling_unsignedT)_backend_init_process_lock_backend_init_thread_lock_is_initializedZctypesrYrZr[r!r ZTHREADING_LAYERlowerplatformsystemrFr6r llZ add_symbolr+rprqlaunch_threads NUM_THREADSrW)rYrZr[rgr Z namedbackendsrdZ_IS_OSXZ _IS_LINUXZ requirementsrkZlibnameZ availablereryr)rhrfrr2 sx                        r2ZNUMBA_DYLD_WORKAROUND);__doc__Z __future__rrsysosrvwarningsZ threadingrZ threadRLockZmultiprocessingZ procRLockZnumpyr.Zllvmlite.llvmpy.coreZllvmpyZcorerZllvmlite.bindingZbindingrxZnumba.npyufuncrZnumba.numpy_supportrZnumbarrr r rrzr,Z UFuncBuilderr-r5Z GUFuncBuilderr@rJrsrrOSErrorerewarnr! contextlibrUrVrtrWrXr2environZ_DYLD_WORKAROUND_SETr getZ_DYLD_WORKAROUND_VALrrrr sL      f '