def __init__(self):
self._lib = None
initialize()
err = lib.clblasSetup()
if err:
raise CLRuntimeError("clblasSetup() failed with error %s" %
CL.get_error_description(err), err)
self._lib = lib # to hold the reference
def dgemm(self, queues, order, transA, transB,
rowsCountA, columnCountB, commonSideLength,
alpha, A, B, beta, C,
offsetA=0, strideA=0,
offsetB=0, strideB=0,
offsetC=0, strideC=0,
wait_for=None, need_event=False):
"""Double precision (double) GEneral Matrix Multiplication.
C = alpha * dot(A, B) + beta * C
C = alpha * dot(A^T, B) + beta * C
C = alpha * dot(A, B^T) + beta * C
C = alpha * dot(A^T, B^T) + beta * C
Parameters:
queues: list of the Queue objects on which this operation
will be enqueued.
order: row/column order (clblasRowMajor, clblasColumnMajor).
transA: how matrix A is to be transposed
(clblasNoTrans, clblasTrans, clblasConjTrans).
transB: how matrix B is to be transposed
(clblasNoTrans, clblasTrans, clblasConjTrans).
rowsCountA: number of rows in matrix A.
columnCountB: number of columns in matrix B.
commonSideLength: length of the common side of the matrices.
alpha: the factor of matrix A.
A: Buffer object storing matrix A.
B: Buffer object storing matrix B.
beta: the factor of matrix C.
C: Buffer object storing matrix C.
offsetA: offset of the first element of the matrix A
in the buffer object, counted in elements.
strideA: leading dimension of matrix A:
((clblasNoTrans, clblasRowMajor) or
(clblasTrans, clblasColumnMajor)): >= commonSideLength,
else: >= rowsCountA.
offsetB: offset of the first element of the matrix B
in the buffer object, counted in elements.
strideB: leading dimension of matrix B:
((clblasNoTrans, clblasRowMajor) or
(clblasTrans, clblasColumnMajor)): >= columnCountB,
else: >= commonSideLength.
offsetC: offset of the first element of the matrix C
in the buffer object, counted in elements.
strideC: leading dimension of matrix C:
clblasRowMajor: >= columnCountB,
else: >= rowsCountA.
wait_for: list of the Event objects to wait.
need_event: return Event object or not.
Returns:
Event object or None if need_event == False.
"""
event = ffi.new("cl_event[]", 1) if need_event else clffi.ffi.NULL
wait_list, n_events = CL.get_wait_list(wait_for)
_queues = ffi.new("cl_command_queue[]", len(queues))
for i, q in enumerate(queues):
_queues[i] = q.handle
if not strideA:
strideA = (
commonSideLength
if ((transA == clblasNoTrans and order == clblasRowMajor) or
(transA != clblasNoTrans and order == clblasColumnMajor))
else rowsCountA)
if not strideB:
strideB = (
columnCountB
if ((transB == clblasNoTrans and order == clblasRowMajor) or
(transB != clblasNoTrans and order == clblasColumnMajor))
else commonSideLength)
if not strideC:
strideC = columnCountB if order == clblasRowMajor else rowsCountA
err = self._lib.clblasDgemm(
order, transA, transB, rowsCountA, columnCountB, commonSideLength,
alpha, A.handle, offsetA, strideA, B.handle, offsetB, strideB,
beta, C.handle, offsetC, strideC, len(queues), _queues,
n_events, wait_list, event)
if err:
raise CLRuntimeError("clblasDgemm() failed with error %s" %
CL.get_error_description(err), err)
return Event(event[0]) if event != clffi.ffi.NULL else None