def new(shape, dtype='float32'):
'''Allocate on the GPU the necessary memory to hold an array of given
shape and dtype.
Return an XND container whose data points to that memory, augmented with
some GPU attributes.
'''
size = reduce(operator.mul, shape, 1)
if dtype == 'float32':
addr = gpu_mem.new_array_float32(size)
elif dtype == 'float64':
addr = gpu_mem.new_array_float64(size)
a = np.ctypeslib.as_array(addr, shape=shape)
x = xnd.from_buffer(a)
x.gpu_addr = addr
x.gpu_dtype = dtype
x.gpu_shape = shape
x.gpu_size = size
x.gpu_copy_from_buffer = types.MethodType(copy_from_buffer, x)
x.gpu_copy_to_buffer = types.MethodType(copy_to_buffer, x)
x.gpu_delete = types.MethodType(delete, x)
return x
def test_bench_exp():
sizes = [2**i + 1 for i in range(1, 20)]
sizes = np.linspace(1, 3_000_000, 100, dtype=int)
xnd_timings_exp = []
xnd_timings_exp_inout = []
xnd_timings_myexp = []
xnd_timings_myexp_inout = []
numpy_timings_exp = []
numpy_timings_exp_inout = []
xnd_timings_abs = []
numpy_timings_abs = []
xnd_timings_nothing = []
xnd_timings_nothing_inout = []
xnd_timings_copy = []
xnd_timings_copy_inout = []
xnd_timings_mycopy = []
xnd_timings_mycopy_inout = []
for size in sizes:
print(f'size={size}')
a = np.random.uniform(0, 1, size=size)
r = np.empty(dtype=a.dtype, shape=size)
xa = xnd.from_buffer(a)
xr = xnd.from_buffer(r)
if 0:
xnd_timings_copy.append(timeit(vml.copy, (xa, )))
xnd_timings_copy_inout.append(timeit(vml.copy_inout, (xa, xr)))
if 0:
xnd_timings_mycopy.append(timeit(vml.mycopy, (xa, )))
xnd_timings_mycopy_inout.append(timeit(vml.mycopy_inout, (xa, xr)))
if 0:
xnd_timings_nothing.append(timeit(vml.nothing, (xa, )))
xnd_timings_nothing_inout.append(
timeit(vml.nothing_inout, (xa, xr)))
if 0:
xnd_timings_myexp.append(timeit(vml.myexp, (xa, )))
xnd_timings_myexp_inout.append(timeit(vml.myexp_inout, (xa, xr)))
if 1:
xnd_timings_exp.append(timeit(vml.exp, (xa, )))
xnd_timings_exp_inout.append(timeit(vml.exp_inout, (xa, xr)))
if 1:
numpy_timings_exp.append(timeit(np.exp, (a, )))
numpy_timings_exp_inout.append(timeit(np.exp, (a, r)))
#xnd_timings_abs.append(timeit(vml.abs, (xa,)))
#numpy_timings_abs.append(timeit(np.abs, (a,)))
try:
from matplotlib import pyplot as plt
except ImportError:
return
if 0:
plt.plot(sizes, xnd_timings_copy, '+', label='mkl_vml.copy [XND]')
plt.plot(sizes,
xnd_timings_copy_inout,
'+',
label='mkl_vml.copy_inout [XND]')
if 0:
plt.plot(sizes, xnd_timings_mycopy, '+', label='mkl_vml.mycopy [XND]')
plt.plot(sizes,
xnd_timings_mycopy_inout,
'+',
label='mkl_vml.mycopy_inout [XND]')
if 0:
plt.plot(sizes,
xnd_timings_nothing,
'+',
label='mkl_vml.nothing [XND]')
plt.plot(sizes,
xnd_timings_nothing_inout,
'+',
label='mkl_vml.nothing_inout [XND]')
if 0:
plt.plot(sizes, xnd_timings_myexp, '+', label='mkl_vml.myexp [XND]')
plt.plot(sizes,
xnd_timings_myexp_inout,
'x',
label='mkl_vml.myexp_inout [XND]')
if 1:
plt.plot(sizes,
xnd_timings_exp,
'+',
label='mkl_vml.exp(xa)->xr [XND]')
plt.plot(sizes,
xnd_timings_exp_inout,
'x',
label='mkl_vml.exp_inout(xa, xr) [XND]')
if 1:
plt.plot(sizes, numpy_timings_exp, '.', label='numpy.exp(a)->r')
plt.plot(sizes,
numpy_timings_exp_inout,
'.',
label='numpy.exp(a, out=r)')
#plt.plot(sizes, xnd_timings_abs, 'x', label='mkl_vml.abs [XND]')
#plt.plot(sizes, numpy_timings_abs, '.', label='numpy.abs')
plt.legend()
plt.xlabel('Array size')
plt.ylabel('Timing, ms per call')
plt.savefig('mkl_vml_exp.png')
plt.show()
def h(*args, **kwargs):
out = g(*args, **kwargs)
view = xnd.from_buffer(out)
ret = xnd.empty(view.type)
np.copyto(np.array(ret, copy=False), out)
return ret
operations = {'binary': binary_op, 'unary': unary_op}
size = 10_000_000
in0 = np.random.uniform(0, 1, size=size).astype('float64')
in1 = np.random.uniform(0, 1, size=size).astype('float64')
dummy = np.empty_like(in0)
# binary operations
xgin0 = xnd_gpu(from_buffer=in0)
xgin1 = xnd_gpu(from_buffer=in1)
xgout = xnd_gpu(empty_like=in0)
xin0 = xnd.from_buffer(in0)
xin1 = xnd.from_buffer(in1)
index = []
xgtime = []
xtime = []
ntime = []
at_least = 10 # run at least 10 seconds
def run(op, args, kwargs={}, at_least=10):
done = False
n = 0
t0 = time()
while not done:
for _ in range(100):