FM / mkl

/usr/local/lib64/python3.6/site-packages/caffe2/python/mkl/

/usr/local/lib64/python3.6/site-packages/caffe2/python/mkl

Name	Size	Mode	Actions
__pycache__/	-	0755	rm
mkl_concat_op_test.py	1173	0644	edit dl rm
mkl_conv_op_test.py	1684	0644	edit dl rm
mkl_copy_op_test.py	2149	0644	edit dl rm
mkl_elementwise_add_op_test.py	1201	0644	edit dl rm
mkl_elementwise_sum_op_test.py	1334	0644	edit dl rm
mkl_fc_op_test.py	927	0644	edit dl rm
mkl_fc_speed_test.py	3849	0644	edit dl rm
mkl_fill_op_test.py	1019	0644	edit dl rm
mkl_LRN_op_test.py	1160	0644	edit dl rm
mkl_LRN_speed_test.py	3182	0644	edit dl rm
mkl_pool_op_test.py	1332	0644	edit dl rm
mkl_pool_speed_test.py	4210	0644	edit dl rm
mkl_relu_op_test.py	988	0644	edit dl rm
mkl_sbn_op_test.py	3110	0644	edit dl rm
mkl_sbn_speed_test.py	4600	0644	edit dl rm
mkl_sigmoid_op_test.py	833	0644	edit dl rm
mkl_speed_test.py	3079	0644	edit dl rm
mkl_squeeze_op_test.py	968	0644	edit dl rm
rewrite_graph.py	8433	0644	edit dl rm
rewrite_graph_test.py	7889	0644	edit dl rm
__init__.py	0	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/caffe2/python/mkl/mkl_fc_speed_test.py (3849B)

import unittest

import numpy as np
from caffe2.proto import caffe2_pb2
from caffe2.python import core, workspace, test_util

@unittest.skipIf(not workspace.C.has_mkldnn, "Skipping as we do not have mkldnn.")
class TestMKLBasic(test_util.TestCase):
    def testFCSpeed(self):
        # We randomly select a shape to test the speed. Intentionally we
        # test a batch size of 1 since this may be the most frequent use
        # case for MKL during deployment time.
        X = np.random.rand(1, 256, 6, 6).astype(np.float32) - 0.5
        #X = np.random.rand(32, 256*6*6).astype(np.float32) - 0.5
        W = np.random.rand(4096, 9216).astype(np.float32) - 0.5
        b = np.random.rand(4096).astype(np.float32) - 0.5
        mkl_do = core.DeviceOption(caffe2_pb2.MKLDNN)
        # Makes sure that feed works.
        workspace.FeedBlob("X", X)
        workspace.FeedBlob("W", W)
        workspace.FeedBlob("b", b)
        workspace.FeedBlob("X_mkl", X, device_option=mkl_do)
        workspace.FeedBlob("W_mkl", W, device_option=mkl_do)
        workspace.FeedBlob("b_mkl", b, device_option=mkl_do)
        net = core.Net("test")
        # Makes sure that we can run relu.
        net.FC(["X", "W", "b"], "Y")
        net.FC(["X_mkl", "W_mkl", "b_mkl"], "Y_mkl", device_option=mkl_do)

workspace.CreateNet(net)
        workspace.RunNet(net)
        # makes sure that the results are good.
        np.testing.assert_allclose(
            workspace.FetchBlob("Y"),
            workspace.FetchBlob("Y_mkl"),
            atol=1e-2,
            rtol=1e-2)
        runtime = workspace.BenchmarkNet(net.Proto().name, 1, 100, True)

print("FC CPU runtime {}, MKL runtime {}.".format(runtime[1], runtime[2]))

def testConvReluMaxPoolFcSpeed(self):
        # We randomly select a shape to test the speed. Intentionally we
        # test a batch size of 1 since this may be the most frequent use
        # case for MKL during deployment time.
        X = np.random.rand(1, 256, 13, 13).astype(np.float32) - 0.5
        W = np.random.rand(256, 256, 3, 3).astype(np.float32) - 0.5
        b = np.random.rand(256).astype(np.float32) - 0.5

w_fc = np.random.rand(4096, 9216).astype(np.float32) - 0.5
        b_fc = np.random.rand(4096).astype(np.float32) - 0.5
        mkl_do = core.DeviceOption(caffe2_pb2.MKLDNN)
        # Makes sure that feed works.
        workspace.FeedBlob("X", X)
        workspace.FeedBlob("W", W)
        workspace.FeedBlob("b", b)
        workspace.FeedBlob("w_fc", w_fc)
        workspace.FeedBlob("b_fc", b_fc)
        workspace.FeedBlob("X_mkl", X, device_option=mkl_do)
        workspace.FeedBlob("W_mkl", W, device_option=mkl_do)
        workspace.FeedBlob("b_mkl", b, device_option=mkl_do)
        workspace.FeedBlob("w_fc_mkl", w_fc, device_option=mkl_do)
        workspace.FeedBlob("b_fc_mkl", b_fc, device_option=mkl_do)

net = core.Net("test")

net.Conv(["X", "W", "b"], "C", pad=1, stride=1, kernel=3)
        net.Relu("C", "R")
        net.MaxPool("R", "P", stride=2, kernel=3)
        net.FC(["P","w_fc", "b_fc"], "Y")

net.Conv(["X_mkl", "W_mkl", "b_mkl"], "C_mkl",
                 pad=1, stride=1, kernel=3, device_option=mkl_do)
        net.Relu("C_mkl", "R_mkl", device_option=mkl_do)
        net.MaxPool("R_mkl", "P_mkl",
                 stride=2, kernel=3, device_option=mkl_do)
        net.FC(["P_mkl","w_fc_mkl", "b_fc_mkl"], "Y_mkl", device_option=mkl_do)

if __name__ == '__main__':
    unittest.main()

import unittest import numpy as np from caffe2.proto import caffe2_pb2 from caffe2.python import core, workspace, test_util @unittest.skipIf(not workspace.C.has_mkldnn, "Skipping as we do not have mkldnn.") class TestMKLBasic(test_util.TestCase): def testFCSpeed(self): # We randomly select a shape to test the speed. Intentionally we # test a batch size of 1 since this may be the most frequent use # case for MKL during deployment time. X = np.random.rand(1, 256, 6, 6).astype(np.float32) - 0.5 #X = np.random.rand(32, 256*6*6).astype(np.float32) - 0.5 W = np.random.rand(4096, 9216).astype(np.float32) - 0.5 b = np.random.rand(4096).astype(np.float32) - 0.5 mkl_do = core.DeviceOption(caffe2_pb2.MKLDNN) # Makes sure that feed works. workspace.FeedBlob("X", X) workspace.FeedBlob("W", W) workspace.FeedBlob("b", b) workspace.FeedBlob("X_mkl", X, device_option=mkl_do) workspace.FeedBlob("W_mkl", W, device_option=mkl_do) workspace.FeedBlob("b_mkl", b, device_option=mkl_do) net = core.Net("test") # Makes sure that we can run relu. net.FC(["X", "W", "b"], "Y") net.FC(["X_mkl", "W_mkl", "b_mkl"], "Y_mkl", device_option=mkl_do) workspace.CreateNet(net) workspace.RunNet(net) # makes sure that the results are good. np.testing.assert_allclose( workspace.FetchBlob("Y"), workspace.FetchBlob("Y_mkl"), atol=1e-2, rtol=1e-2) runtime = workspace.BenchmarkNet(net.Proto().name, 1, 100, True) print("FC CPU runtime {}, MKL runtime {}.".format(runtime[1], runtime[2])) def testConvReluMaxPoolFcSpeed(self): # We randomly select a shape to test the speed. Intentionally we # test a batch size of 1 since this may be the most frequent use # case for MKL during deployment time. X = np.random.rand(1, 256, 13, 13).astype(np.float32) - 0.5 W = np.random.rand(256, 256, 3, 3).astype(np.float32) - 0.5 b = np.random.rand(256).astype(np.float32) - 0.5 w_fc = np.random.rand(4096, 9216).astype(np.float32) - 0.5 b_fc = np.random.rand(4096).astype(np.float32) - 0.5 mkl_do = core.DeviceOption(caffe2_pb2.MKLDNN) # Makes sure that feed works. workspace.FeedBlob("X", X) workspace.FeedBlob("W", W) workspace.FeedBlob("b", b) workspace.FeedBlob("w_fc", w_fc) workspace.FeedBlob("b_fc", b_fc) workspace.FeedBlob("X_mkl", X, device_option=mkl_do) workspace.FeedBlob("W_mkl", W, device_option=mkl_do) workspace.FeedBlob("b_mkl", b, device_option=mkl_do) workspace.FeedBlob("w_fc_mkl", w_fc, device_option=mkl_do) workspace.FeedBlob("b_fc_mkl", b_fc, device_option=mkl_do) net = core.Net("test") net.Conv(["X", "W", "b"], "C", pad=1, stride=1, kernel=3) net.Relu("C", "R") net.MaxPool("R", "P", stride=2, kernel=3) net.FC(["P","w_fc", "b_fc"], "Y") net.Conv(["X_mkl", "W_mkl", "b_mkl"], "C_mkl", pad=1, stride=1, kernel=3, device_option=mkl_do) net.Relu("C_mkl", "R_mkl", device_option=mkl_do) net.MaxPool("R_mkl", "P_mkl", stride=2, kernel=3, device_option=mkl_do) net.FC(["P_mkl","w_fc_mkl", "b_fc_mkl"], "Y_mkl", device_option=mkl_do) workspace.CreateNet(net) workspace.RunNet(net) # makes sure that the results are good. np.testing.assert_allclose( workspace.FetchBlob("Y"), workspace.FetchBlob("Y_mkl"), atol=1e-2, rtol=1e-2) runtime = workspace.BenchmarkNet(net.Proto().name, 1, 100, True) if __name__ == '__main__': unittest.main()