test_concat

test_concat.py

import numpy as np
import paddle.fluid as fluid

shape1 = [100,1000,100]
shape2 = [100,1000,100]
axis=0

x1 = fluid.layers.data(
    name='x1',
    shape=shape1,
    dtype='float32',
    append_batch_size=False)
x2 = fluid.layers.data(
    name='x2',
    shape=shape2,
    dtype='float32',
    append_batch_size=False)
y = fluid.layers.concat(input=[x1, x2], axis=axis)

place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())

x1 = np.random.random(shape1).astype('float32')
x2 = np.random.random(shape2).astype('float32')
tensor_x = fluid.core.LoDTensor()
tensor_x.set(x1, place)
tensor_y = fluid.core.LoDTensor()
tensor_y.set(x2, place)

out_py = np.concatenate([x1,x2],axis = axis)
out = exe.run(fluid.default_main_program(),
         feed={"x1": tensor_x,
               "x2": tensor_y}, fetch_list=[y])
out_data = np.array(out)
assert np.allclose(out_data, out_py) is True

PASS_NUM = 300 #,
with fluid.profiler.profiler("GPU", 'total',"/temp/Paddle") as prof:
    for pass_id in range(PASS_NUM):
        exe.run(fluid.default_main_program(),
                feed={"x1": tensor_x,
                      "x2": tensor_y})
    print "over"

Experiment Result:
axis = 0

input_shape	copy(D->D)	D->H->concat->D	kernel
1 x 10^6	0.074824667	/	0.100071
1 x 10^7	0.472866667	/	0.568678
1 x 10^8	4.530733333	/	9.6112

axis = 1

input_shape	copy(D->D)	D->H->concat->D	kernel
100 x 10^4	0.893866667	3.3531	0.076095
100 x 10^5	1.227483333	34.148	0.58596
1000 x 10^4	10.02413333	35.25233333	0.5465
100 x 10^6	5.195033333	362.9734997	9.7122
1000 x 10^5	10.5971	391.4569667	9.506333333

axis = 2

input_shape	copy(D->D)	D->H->concat->D	kernel
10^4 x 100	87.12066667	3.2949	0.078481
10^4 x 1000	89.871	39.51366667	0.55338
10^5 x 1000	938.5879967	389.6895	8.571266667