def duplicate(ctx, data):
"""
# XXX is this the same as transpose???
cannot use pyopencl.array.transpose at the momoment, why???
"""
KERNEL = """
for (int j = 0; j < IN_BLOCK_SIZE; ++j) {
b[OUT_BLOCK_SIZE*__id+j] = a[j];
}
"""
mapper = Blockwise(ctx, map_expr=KERNEL, arguments=[
('a', 'global const', data.dtype, '*a'),
('b', 'global', data.dtype, '*b')
],
in_blocksize=reduce(mul, data.shape),
out_blocksize=reduce(mul, data.shape)
)
mapper.build()
def _kernel(queue, length, b=None):
if b is None:
shape = [length] + list(data.shape[1:])
b = cl.array.empty(queue, tuple(shape), data.dtype)
mapper(queue, length, data.data, b.data)
return b
return _kernel
def rearange_to_block(ctx, strided_data):
"""
# XXX is this the same as transpose???
cannot use pyopencl.array.transpose at the momoment, why???
"""
KERNEL = """
b[OUT_BLOCK_SIZE*__item_id.y+__id] = a[__in_offset+__item_id.y];
"""
mapper = Blockwise(ctx, map_expr=KERNEL, arguments=[
('a', 'global const', strided_data.dtype, '*a'),
('b', 'global', strided_data.dtype, '*b')
],
in_blocksize=strided_data.shape[1],
out_blocksize=strided_data.shape[0],
block_shape=strided_data.shape,
threads=(1, strided_data.shape[1],)
)
mapper.build()
def _kernel(queue, length, b):
return mapper(queue, length, strided_data.data, b)
return _kernel
