from cython cimport final from cython.operator cimport dereference as deref from cython.parallel cimport parallel, prange from libcpp.vector cimport vector from ...utils._cython_blas cimport _dot from ...utils._openmp_helpers cimport omp_get_thread_num from ...utils._typedefs cimport intp_t, float32_t, float64_t, int32_t import numpy as np from scipy.sparse import issparse from numbers import Integral from sklearn import get_config from sklearn.utils import check_scalar from ...utils._openmp_helpers import _openmp_effective_n_threads ##################### cdef float64_t[::1] _sqeuclidean_row_norms64_dense( const float64_t[:, ::1] X, intp_t num_threads, ): """Compute the squared euclidean norm of the rows of X in parallel. This is faster than using np.einsum("ij, ij->i") even when using a single thread. """ cdef: # Casting for X to remove the const qualifier is needed because APIs # exposed via scipy.linalg.cython_blas aren't reflecting the arguments' # const qualifier. # See: https://github.com/scipy/scipy/issues/14262 float64_t * X_ptr = &X[0, 0] intp_t idx = 0 intp_t n = X.shape[0] intp_t d = X.shape[1] float64_t[::1] squared_row_norms = np.empty(n, dtype=np.float64) for idx in prange(n, schedule='static', nogil=True, num_threads=num_threads): squared_row_norms[idx] = _dot(d, X_ptr + idx * d, 1, X_ptr + idx * d, 1) return squared_row_norms cdef float64_t[::1] _sqeuclidean_row_norms32_dense( const float32_t[:, ::1] X, intp_t num_threads, ): """Compute the squared euclidean norm of the rows of X in parallel. This is faster than using np.einsum("ij, ij->i") even when using a single thread. """ cdef: # Casting for X to remove the const qualifier is needed because APIs # exposed via scipy.linalg.cython_blas aren't reflecting the arguments' # const qualifier. # See: https://github.com/scipy/scipy/issues/14262 float32_t * X_ptr = &X[0, 0] intp_t i = 0, j = 0 intp_t thread_num intp_t n = X.shape[0] intp_t d = X.shape[1] float64_t[::1] squared_row_norms = np.empty(n, dtype=np.float64) # To upcast the i-th row of X from float32 to float64 vector[vector[float64_t]] X_i_upcast = vector[vector[float64_t]]( num_threads, vector[float64_t](d) ) with nogil, parallel(num_threads=num_threads): thread_num = omp_get_thread_num() for i in prange(n, schedule='static'): # Upcasting the i-th row of X from float32 to float64 for j in range(d): X_i_upcast[thread_num][j] = deref(X_ptr + i * d + j) squared_row_norms[i] = _dot( d, X_i_upcast[thread_num].data(), 1, X_i_upcast[thread_num].data(), 1, ) return squared_row_norms cdef float64_t[::1] _sqeuclidean_row_norms64_sparse( const float64_t[:] X_data, const int32_t[:] X_indptr, intp_t num_threads, ): cdef: intp_t n = X_indptr.shape[0] - 1 int32_t X_i_ptr, idx = 0 float64_t[::1] squared_row_norms = np.zeros(n, dtype=np.float64) for idx in prange(n, schedule='static', nogil=True, num_threads=num_threads): for X_i_ptr in range(X_indptr[idx], X_indptr[idx+1]): squared_row_norms[idx] += X_data[X_i_ptr] * X_data[X_i_ptr] return squared_row_norms {{for name_suffix in ["64", "32"]}} from ._datasets_pair cimport DatasetsPair{{name_suffix}} cpdef float64_t[::1] _sqeuclidean_row_norms{{name_suffix}}( X, intp_t num_threads, ): if issparse(X): # TODO: remove this instruction which is a cast in the float32 case # by moving squared row norms computations in MiddleTermComputer. X_data = np.asarray(X.data, dtype=np.float64) X_indptr = np.asarray(X.indptr, dtype=np.int32) return _sqeuclidean_row_norms64_sparse(X_data, X_indptr, num_threads) else: return _sqeuclidean_row_norms{{name_suffix}}_dense(X, num_threads) cdef class BaseDistancesReduction{{name_suffix}}: """ Base float{{name_suffix}} implementation template of the pairwise-distances reduction backends. Implementations inherit from this template and may override the several defined hooks as needed in order to easily extend functionality with minimal redundant code. """ def __init__( self, DatasetsPair{{name_suffix}} datasets_pair, chunk_size=None, strategy=None, ): cdef: intp_t X_n_full_chunks, Y_n_full_chunks if chunk_size is None: chunk_size = get_config().get("pairwise_dist_chunk_size", 256) self.chunk_size = check_scalar(chunk_size, "chunk_size", Integral, min_val=20) self.effective_n_threads = _openmp_effective_n_threads() self.datasets_pair = datasets_pair self.n_samples_X = datasets_pair.n_samples_X() self.X_n_samples_chunk = min(self.n_samples_X, self.chunk_size) X_n_full_chunks = self.n_samples_X // self.X_n_samples_chunk X_n_samples_remainder = self.n_samples_X % self.X_n_samples_chunk self.X_n_chunks = X_n_full_chunks + (X_n_samples_remainder != 0) if X_n_samples_remainder != 0: self.X_n_samples_last_chunk = X_n_samples_remainder else: self.X_n_samples_last_chunk = self.X_n_samples_chunk self.n_samples_Y = datasets_pair.n_samples_Y() self.Y_n_samples_chunk = min(self.n_samples_Y, self.chunk_size) Y_n_full_chunks = self.n_samples_Y // self.Y_n_samples_chunk Y_n_samples_remainder = self.n_samples_Y % self.Y_n_samples_chunk self.Y_n_chunks = Y_n_full_chunks + (Y_n_samples_remainder != 0) if Y_n_samples_remainder != 0: self.Y_n_samples_last_chunk = Y_n_samples_remainder else: self.Y_n_samples_last_chunk = self.Y_n_samples_chunk if strategy is None: strategy = get_config().get("pairwise_dist_parallel_strategy", 'auto') if strategy not in ('parallel_on_X', 'parallel_on_Y', 'auto'): raise RuntimeError(f"strategy must be 'parallel_on_X, 'parallel_on_Y', " f"or 'auto', but currently strategy='{self.strategy}'.") if strategy == 'auto': # This is a simple heuristic whose constant for the # comparison has been chosen based on experiments. # parallel_on_X has less synchronization overhead than # parallel_on_Y and should therefore be used whenever # n_samples_X is large enough to not starve any of the # available hardware threads. if self.n_samples_Y < self.n_samples_X: # No point to even consider parallelizing on Y in this case. This # is in particular important to do this on machines with a large # number of hardware threads. strategy = 'parallel_on_X' elif 4 * self.chunk_size * self.effective_n_threads < self.n_samples_X: # If Y is larger than X, but X is still large enough to allow for # parallelism, we might still want to favor parallelizing on X. strategy = 'parallel_on_X' else: strategy = 'parallel_on_Y' self.execute_in_parallel_on_Y = strategy == "parallel_on_Y" # Not using less, not using more. self.chunks_n_threads = min( self.Y_n_chunks if self.execute_in_parallel_on_Y else self.X_n_chunks, self.effective_n_threads, ) @final cdef void _parallel_on_X(self) noexcept nogil: """Perform computation and reduction in parallel on chunks of X. This strategy dispatches tasks statically on threads. Each task processes exactly only one chunk of X, computing and reducing distances matrices between vectors of this chunk and vectors of all chunks of Y, one chunk of Y at a time. This strategy is embarrassingly parallel with no intermediate data structures synchronization at all. Private datastructures are modified internally by threads. Private template methods can be implemented on subclasses to interact with those datastructures at various stages. """ cdef: intp_t Y_start, Y_end, X_start, X_end, X_chunk_idx, Y_chunk_idx intp_t thread_num with nogil, parallel(num_threads=self.chunks_n_threads): thread_num = omp_get_thread_num() # Allocating thread datastructures self._parallel_on_X_parallel_init(thread_num) for X_chunk_idx in prange(self.X_n_chunks, schedule='static'): X_start = X_chunk_idx * self.X_n_samples_chunk if X_chunk_idx == self.X_n_chunks - 1: X_end = X_start + self.X_n_samples_last_chunk else: X_end = X_start + self.X_n_samples_chunk # Reinitializing thread datastructures for the new X chunk self._parallel_on_X_init_chunk(thread_num, X_start, X_end) for Y_chunk_idx in range(self.Y_n_chunks): Y_start = Y_chunk_idx * self.Y_n_samples_chunk if Y_chunk_idx == self.Y_n_chunks - 1: Y_end = Y_start + self.Y_n_samples_last_chunk else: Y_end = Y_start + self.Y_n_samples_chunk self._parallel_on_X_pre_compute_and_reduce_distances_on_chunks( X_start, X_end, Y_start, Y_end, thread_num, ) self._compute_and_reduce_distances_on_chunks( X_start, X_end, Y_start, Y_end, thread_num, ) # Adjusting thread datastructures on the full pass on Y self._parallel_on_X_prange_iter_finalize(thread_num, X_start, X_end) # end: for X_chunk_idx # Deallocating thread datastructures self._parallel_on_X_parallel_finalize(thread_num) # end: with nogil, parallel return @final cdef void _parallel_on_Y(self) noexcept nogil: """Perform computation and reduction in parallel on chunks of Y. This strategy is a sequence of embarrassingly parallel subtasks: chunks of X are iterated over sequentially, and for each chunk of X, tasks are dispatched statically on threads. Each task processes one and only one chunk of Y, computing and reducing distances matrices between vectors of the chunk of X and vectors of the Y. It comes with lock-free and parallelized intermediate data structures that synchronize at each iteration of the sequential outer loop on X chunks. Private datastructures are modified internally by threads. Private template methods can be implemented on subclasses to interact with those datastructures at various stages. """ cdef: intp_t Y_start, Y_end, X_start, X_end, X_chunk_idx, Y_chunk_idx intp_t thread_num # Allocating datastructures shared by all threads self._parallel_on_Y_init() for X_chunk_idx in range(self.X_n_chunks): X_start = X_chunk_idx * self.X_n_samples_chunk if X_chunk_idx == self.X_n_chunks - 1: X_end = X_start + self.X_n_samples_last_chunk else: X_end = X_start + self.X_n_samples_chunk with nogil, parallel(num_threads=self.chunks_n_threads): thread_num = omp_get_thread_num() # Initializing datastructures used in this thread self._parallel_on_Y_parallel_init(thread_num, X_start, X_end) for Y_chunk_idx in prange(self.Y_n_chunks, schedule='static'): Y_start = Y_chunk_idx * self.Y_n_samples_chunk if Y_chunk_idx == self.Y_n_chunks - 1: Y_end = Y_start + self.Y_n_samples_last_chunk else: Y_end = Y_start + self.Y_n_samples_chunk self._parallel_on_Y_pre_compute_and_reduce_distances_on_chunks( X_start, X_end, Y_start, Y_end, thread_num, ) self._compute_and_reduce_distances_on_chunks( X_start, X_end, Y_start, Y_end, thread_num, ) # end: prange # end: with nogil, parallel # Synchronizing the thread datastructures with the main ones self._parallel_on_Y_synchronize(X_start, X_end) # end: for X_chunk_idx # Deallocating temporary datastructures and adjusting main datastructures self._parallel_on_Y_finalize() return # Placeholder methods which have to be implemented cdef void _compute_and_reduce_distances_on_chunks( self, intp_t X_start, intp_t X_end, intp_t Y_start, intp_t Y_end, intp_t thread_num, ) noexcept nogil: """Compute the pairwise distances on two chunks of X and Y and reduce them. This is THE core computational method of BaseDistancesReduction{{name_suffix}}. This must be implemented in subclasses agnostically from the parallelization strategies. """ return def _finalize_results(self, bint return_distance): """Callback adapting datastructures before returning results. This must be implemented in subclasses. """ return None # Placeholder methods which can be implemented cdef void compute_exact_distances(self) noexcept nogil: """Convert rank-preserving distances to exact distances or recompute them.""" return cdef void _parallel_on_X_parallel_init( self, intp_t thread_num, ) noexcept nogil: """Allocate datastructures used in a thread given its number.""" return cdef void _parallel_on_X_init_chunk( self, intp_t thread_num, intp_t X_start, intp_t X_end, ) noexcept nogil: """Initialize datastructures used in a thread given its number. In this method, EuclideanDistance specialisations of subclass of BaseDistancesReduction _must_ call: self.middle_term_computer._parallel_on_X_init_chunk( thread_num, X_start, X_end, ) to ensure the proper upcast of X[X_start:X_end] to float64 prior to the reduction with float64 accumulator buffers when X.dtype is float32. """ return cdef void _parallel_on_X_pre_compute_and_reduce_distances_on_chunks( self, intp_t X_start, intp_t X_end, intp_t Y_start, intp_t Y_end, intp_t thread_num, ) noexcept nogil: """Initialize datastructures just before the _compute_and_reduce_distances_on_chunks. In this method, EuclideanDistance specialisations of subclass of BaseDistancesReduction _must_ call: self.middle_term_computer._parallel_on_X_pre_compute_and_reduce_distances_on_chunks( X_start, X_end, Y_start, Y_end, thread_num, ) to ensure the proper upcast of Y[Y_start:Y_end] to float64 prior to the reduction with float64 accumulator buffers when Y.dtype is float32. """ return cdef void _parallel_on_X_prange_iter_finalize( self, intp_t thread_num, intp_t X_start, intp_t X_end, ) noexcept nogil: """Interact with datastructures after a reduction on chunks.""" return cdef void _parallel_on_X_parallel_finalize( self, intp_t thread_num ) noexcept nogil: """Interact with datastructures after executing all the reductions.""" return cdef void _parallel_on_Y_init( self, ) noexcept nogil: """Allocate datastructures used in all threads.""" return cdef void _parallel_on_Y_parallel_init( self, intp_t thread_num, intp_t X_start, intp_t X_end, ) noexcept nogil: """Initialize datastructures used in a thread given its number. In this method, EuclideanDistance specialisations of subclass of BaseDistancesReduction _must_ call: self.middle_term_computer._parallel_on_Y_parallel_init( thread_num, X_start, X_end, ) to ensure the proper upcast of X[X_start:X_end] to float64 prior to the reduction with float64 accumulator buffers when X.dtype is float32. """ return cdef void _parallel_on_Y_pre_compute_and_reduce_distances_on_chunks( self, intp_t X_start, intp_t X_end, intp_t Y_start, intp_t Y_end, intp_t thread_num, ) noexcept nogil: """Initialize datastructures just before the _compute_and_reduce_distances_on_chunks. In this method, EuclideanDistance specialisations of subclass of BaseDistancesReduction _must_ call: self.middle_term_computer._parallel_on_Y_pre_compute_and_reduce_distances_on_chunks( X_start, X_end, Y_start, Y_end, thread_num, ) to ensure the proper upcast of Y[Y_start:Y_end] to float64 prior to the reduction with float64 accumulator buffers when Y.dtype is float32. """ return cdef void _parallel_on_Y_synchronize( self, intp_t X_start, intp_t X_end, ) noexcept nogil: """Update thread datastructures before leaving a parallel region.""" return cdef void _parallel_on_Y_finalize( self, ) noexcept nogil: """Update datastructures after executing all the reductions.""" return {{endfor}}