ó àÆ÷Xc@`s{ddlmZmZmZddlmZddlZd„Zd„Zdd„Z dd„Z dd „Z d „Z dS( i(tabsolute_importtprint_functiontdivision(txrangeNc C`s£d}x–tt||ƒƒD]\}\}}|d|}|dtƒ7}xKt|ƒD]=\}} | dkr†|dtƒ7}qZ|dtƒ7}qZWqW|S(s; Produce code to declare all necessary variables. tslv%is* %(dtype)s* %(var)s_iter; txs  npy_intp %(var)s_n%(value)i; ssize_t %(var)s_stride%(value)i; int %(var)s_jump%(value)i_%(j)i; sB int %(var)s_jump%(value)s_%(j)i; (t enumeratetziptlocals( t loop_orderstdtypestsubtdecltit loop_ordertdtypetvartjtvalue((s;/tmp/pip-build-X4mzal/theano/theano/tensor/elemwise_cgen.pyt make_declares( cC`s×d}xtt||ƒƒD]ÿ\}\}}d|}g|D]}|dkr?|^q?} | r‡t| ƒd} |dtƒ7}nd} x‹ttt|ƒƒƒD]q\} } | dkröddtƒ| f}|d tƒ7}d tƒ} q¦d | }|d tƒ7}d} q¦WqWd}xŸt|ŒD]‘}gt|ƒD]$\} }|dkrE| |f^qE}t|ƒd kr‡q2n|d\}}x)|dD]\} }|dtƒ7}q¢Wq2W||||S(NRs %%(lv%i)sRisÁ if (PyArray_NDIM(%(var)s) < %(min_nd)s) { PyErr_SetString(PyExc_ValueError, "Not enough dimensions on input."); %%(fail)s } t0s (%s) - (%s)s%(var)s_stride%(index)ssy %(var)s_n%(index)s = PyArray_DIMS(%(var)s)[%(index)s]; %(var)s_stride%(index)s = PyArray_STRIDES(%(var)s)[%(index)s] / sizeof(%(dtype)s); %(var)s_jump%(index)s_%(j)s = %(jump)s; //printf("%(var)s_jump%(index)s_%(j)s is:"); //std::cout << %(var)s_jump%(index)s_%(j)s << std::endl; s*%(var)s_n%(index)s*%(var)s_stride%(index)ss-(%s)sÏ %(var)s_jump%(index)s_%(j)s = %(jump)s; //printf("%(var)s_jump%(index)s_%(j)s is:"); //std::cout << %(var)s_jump%(index)s_%(j)s << std::endl; iisã if (%%(lv%(j0)s)s_n%(x0)s != %%(lv%(j)s)s_n%(x)s) { PyErr_Format(PyExc_ValueError, "Input dimension mis-match. (input[%%%%i].shape[%%%%i] = %%%%i, input[%%%%i].shape[%%%%i] = %%%%i)", %(j0)s, %(x0)s, %%(lv%(j0)s)s_n%(x0)s, %(j)s, %(x)s, %%(lv%(j)s)s_n%(x)s ); %%(fail)s } (RRtmaxRtreversedtlisttlen(R R R tinitR RRRRtnonxtmin_ndtadjustRtindextjumptchecktmatchest to_comparetj0tx0((s;/tmp/pip-build-X4mzal/theano/theano/tensor/elemwise_cgen.pyt make_checks(s:( %%  7 Rc C`sÚ|jƒ}|jdƒr0|jddƒ}nt|dƒ}d}x}tt|ŒƒD]i\}}xZt|ƒD];\} } | dkrr|d| } |dtƒ7}PqrqrW|dtƒ7}qYWd ttƒ|S( sGenerate C code to allocate outputs. Parameters ---------- fortran : str A string included in the generated code. If it evaluate to non-zero, an ndarray in fortran order will be created, otherwise it will be c order. tTHEANO_COMPLEXt NPY_COMPLEXiRRslv%is&dims[%(i)s] = %(var)s_n%(candidate)s; sdims[%(i)s] = 1; s” { npy_intp dims[%(nd)s]; //npy_intp* dims = (npy_intp*)malloc(%(nd)s * sizeof(npy_intp)); %(init_dims)s if (!%(olv)s) { %(olv)s = (PyArrayObject*)PyArray_EMPTY(%(nd)s, dims, %(type)s, %(fortran)s); } else { PyArray_Dims new_dims; new_dims.len = %(nd)s; new_dims.ptr = dims; PyObject* success = PyArray_Resize(%(olv)s, &new_dims, 0, NPY_CORDER); if (!success) { // If we can't resize the ndarray we have we can allocate a new one. PyErr_Clear(); Py_XDECREF(%(olv)s); %(olv)s = (PyArrayObject*)PyArray_EMPTY(%(nd)s, dims, %(type)s, 0); } } if (!%(olv)s) { %(fail)s } } (tuppert startswithtreplaceRRRRtdict( R RR tfortranttypetndt init_dimsR t candidatesRt candidateR((s;/tmp/pip-build-X4mzal/theano/theano/tensor/elemwise_cgen.pyt make_allocvs   c`sc‡‡‡fd†}i}x¬tt|ˆƒƒD]•\}\}} x€t|ƒD]G\} } | dkrP|j| dƒ|| cdtƒˆ7\}\} }}||j|dƒ| | |||ƒ} q W| |d7} d| S( s/ Make a nested loop over several arrays and associate specific code to each level of nesting. Parameters ---------- loop_orders : list of N tuples of length M Each value of each tuple can be either the index of a dimension to loop over or the letter 'x' which means there is no looping to be done over that variable at that point (in other words we broadcast over that dimension). If an entry is an integer, it will become an alias of the entry of that rank. loop_tasks : list of M+1 pieces of code The ith loop_task is a pair of strings, the first string is code to be executed before the ith loop starts, the second one contains code to be executed just before going to the next element of the ith dimension. The last element if loop_tasks is a single string, containing code to be executed at the very end. sub : dictionary Maps 'lv#' to a suitable variable name. The 'lvi' variable corresponds to the ith element of loop_orders. c `s¼d|}d}d}x_t|ƒD]Q\}}ˆd|} ˆ|} |dtƒ7}|dkr#dtƒ}q#q#Wˆrštjj} dtƒ} nd} | d tƒ7} d tƒS( NsITER_%iRt1slv%isU%(dtype)s &%(var)s_i = * ( %(var)s_iter + %(iterv)s * %(var)s_jump%(index)s_%(i)s ); Rs%(var)s_n%(index)ssK#pragma omp parallel for if( %(suitable_n)s >=%(openmp_elemwise_minsize)s) s>for (int %(iterv)s = 0; %(iterv)s<%(suitable_n)s; %(iterv)s++)si %(preloop)s %(forloop)s { %(update)s %(code)s } (RRttheanotconfigtopenmp_elemwise_minsize( tprelooptcodetindicesR titervtupdatet suitable_nRRRRR5tforloop(R topenmpR (s;/tmp/pip-build-X4mzal/theano/theano/tensor/elemwise_cgen.pyt loop_overÎs    RRs>%%(lv%(i)s)s_iter = (%(dtype)s*)(PyArray_DATA(%%(lv%(i)s)s)); iiiÿÿÿÿs{%s}( RRt setdefaultRRRRRtget(R R t loop_tasksR R=R>tpreloopsR RRRRtstpre_taskttaskR8((R R=R s;/tmp/pip-build-X4mzal/theano/theano/tensor/elemwise_cgen.pyt make_loopµs( P-c `sét|ƒ}t|dƒ}ˆd|}dtƒ} x`t||ƒD]N\} } | dkrt| dtƒ7} n| dtƒ7} | dtƒ7} qHW| dtƒ7} g} x{tt|ŒƒD]g\} } xKt| ƒD]7\}}|dkr݈d|}d tƒ}PqÝqÝWd }| j|ƒqÄWd td |d dj| ƒƒ}|dtƒ7}x%t|ƒD]} |dtƒ7}qoW‡fd†‰dtd|d |ddj‡fd†t|ƒDƒƒƒ}|dtƒ7}xat|ƒD]S} ˆd| }|dtƒ7}x+tt|ƒƒD]}|dtƒ7}q-WqõWd}x9t|ƒD]+\} }ˆd| }|dtƒ7}q_Wd}xt|ƒD]s\}}ˆd|}|dtƒ7}d}x5tt |ƒƒD]!} d| }|dtƒ7}qåW|d7}q¡W|}x¦tt |ƒƒD]’} d| }d | }d}d}| |d!krp|}n| dkr¥|r¥t j j }|d"tƒ7}q¥n|d#tƒ7}d$tƒ}q1Wd%jd&| ||||d'gƒS((sÂA bit like make_loop, but when only the inner-most loop executes code. All the loops will be reordered so that the loops over the output tensor are executed with memory access as contiguous as possible. For instance, if the output tensor is c_contiguous, the inner-most loop will be on its rows; if it's f_contiguous, it will be on its columns. The output tensor's index among the loop variables is indicated by olv_index. islv%is¦ std::vector< std::pair > %(ovar)s_loops(%(nnested)i); std::vector< std::pair >::iterator %(ovar)s_loops_it = %(ovar)s_loops.begin(); Rs_ %(ovar)s_loops_it->first = abs(PyArray_STRIDES(%(ovar)s)[%(index)i]); s7 %(ovar)s_loops_it->first = 0; sQ %(ovar)s_loops_it->second = %(i)i; ++%(ovar)s_loops_it; s• // rbegin and rend are reversed iterators, so this sorts in decreasing order std::sort(%(ovar)s_loops.rbegin(), %(ovar)s_loops.rend()); s%(var)s_n%(candidate)sR2s6 int init_totals[%(nnested)s] = {%(totals)s}; tnnestedttotalss, s5 %(ovar)s_loops_it = %(ovar)s_loops.begin(); sh int TOTAL_%(i)i = init_totals[%(ovar)s_loops_it->second]; ++%(ovar)s_loops_it; c`sYˆd|}g}x>|D]6}|dkrD|jdtƒƒq|jdƒqW|S(s¨ Returns a list containing a C expression representing the stride for each dimension of the ith variable, in the specified loop_order. slv%iRs%(var)s_stride%(index)sR(tappendR(RR RtrR(R (s;/tmp/pip-build-X4mzal/theano/theano/tensor/elemwise_cgen.pytget_loop_stridesUs  sL int init_strides[%(nvars)i][%(nnested)i] = { %(strides)s };tnvarststridess, c3`s?|]5\}}t|ƒdkrdjˆ||ƒƒVqdS(is, N(Rtjoin(t.0R tlo(RK(s;/tmp/pip-build-X4mzal/theano/theano/tensor/elemwise_cgen.pys ls sR std::vector< std::pair >::reverse_iterator %(ovar)s_loops_rit; s6 %(ovar)s_loops_rit = %(ovar)s_loops.rbegin();sˆ int %(var)s_stride_l%(j)i = init_strides[%(i)i][%(ovar)s_loops_rit->second]; ++%(ovar)s_loops_rit; Rs4%(var)s_iter = (%(dtype)s*)(PyArray_DATA(%(var)s)); s'%(dtype)s &%(var)s_i = * ( %(var)s_itersITER_%is +%(var)s_stride_l%(i)i*%(iterv)ss); sTOTAL_%iisF#pragma omp parallel for if( %(total)s >=%(openmp_elemwise_minsize)s) s8for(int %(iterv)s = 0; %(iterv)s<%(total)s; %(iterv)s++)sƒ %(forloop)s { // begin loop %(i)i %(update)s %(loop)s } // end loop %(i)i s t{s} ( RRRRRIR*RNRRtrangeR3R4R5(tinit_loop_orderst olv_indexR t inner_taskR R=RLRGtovart order_loopsR RRHR/RR0Rttotaltdeclare_totalstdeclare_stridest declare_iterRtpointer_updatettot_jumpR9tloopR:R<R5((RKR s;/tmp/pip-build-X4mzal/theano/theano/tensor/elemwise_cgen.pytmake_reordered_loopúsœ             c`s„‡fd†}i}x¬tt||ƒƒD]•\}\}}x€t|ƒD]G\} } | dkrJ|j| dƒ|| cdtƒˆ7\}\} } }||j|dƒ| | | ||ƒ} q,W| |d7} d| S( s. Make a nested loop over several arrays and associate specific code to each level of nesting. Parameters ---------- loop_orders : list of N tuples of length M Each value of each tuple can be either the index of a dimension to loop over or the letter 'x' which means there is no looping to be done over that variable at that point (in other words we broadcast over that dimension). If an entry is an integer, it will become an alias of the entry of that rank. loop_tasks : list of M+1 pieces of code The ith loop_task is a pair of strings, the first string is code to be executed before the ith loop starts, the second one contains code to be executed just before going to the next element of the ith dimension. The last element if loop_tasks is a single string, containing code to be executed at the very end. sub: dictionary Maps 'lv#' to a suitable variable name. The 'lvi' variable corresponds to the ith element of loop_orders. c `syd|}d}d}xUt|ƒD]G\}}ˆd|} |dtƒ7}|dkr#dtƒ}q#q#WdtƒS( NsITER_%iRR2slv%is-%(var)s_iter += %(var)s_jump%(index)s_%(i)s; Rs%(var)s_n%(index)ssš %(preloop)s for (int %(iterv)s = %(suitable_n)s; %(iterv)s; %(iterv)s--) { %(code)s %(update)s } (RR( R6R7R8R R9R:R;RRR(R (s;/tmp/pip-build-X4mzal/theano/theano/tensor/elemwise_cgen.pyR>ís  RRs>%%(lv%(i)s)s_iter = (%(dtype)s*)(PyArray_DATA(%%(lv%(i)s)s)); iiiÿÿÿÿs{%s}( RRR?RRR@RRR(R R RAR R>RBR RRRRRCRDRER8((R s;/tmp/pip-build-X4mzal/theano/theano/tensor/elemwise_cgen.pytmake_loop_careduceÓs"( P-(t __future__RRRt six.movesRR3RR$R1tNoneRFR_R`(((s;/tmp/pip-build-X4mzal/theano/theano/tensor/elemwise_cgen.pyts  " N ? F Ø