test_topi_conv2d_nchw.py 9.06 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
# Licensed to the Apache Software Foundation (ASF) under one
# or more contributor license agreements.  See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership.  The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License.  You may obtain a copy of the License at
#
#   http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
# KIND, either express or implied.  See the License for the
# specific language governing permissions and limitations
# under the License.
17
"""Example code to do convolution."""
18

19 20
import numpy as np
import tvm
21
from tvm import autotvm
22
import topi
23
import topi.testing
24
from tvm.contrib.pickle_memoize import memoize
25
from topi.nn.util import get_pad_tuple
26 27
from topi.util import get_const_tuple

28 29
from common import get_all_backend

30 31 32 33 34 35
def verify_conv2d_nchw(batch, in_channel, in_size, num_filter, kernel, stride, padding, dilation=1, add_bias=False, add_relu=False,\
        use_cudnn=False):

    pad_top, pad_left, pad_bottom, pad_right = get_pad_tuple(padding, (kernel, kernel))
    padding_sum = pad_top + pad_left + pad_bottom + pad_right
    print("Workload: (%d, %d, %d, %d, %d, %d, %d, %d)" % (batch, in_channel, in_size, num_filter, kernel, stride, padding_sum, dilation))
36

37 38 39 40
    in_height = in_width = in_size

    A = tvm.placeholder((batch, in_channel, in_height, in_width), name='A')
    W = tvm.placeholder((num_filter, in_channel, kernel, kernel), name='W')
41
    bias = tvm.placeholder((num_filter, 1, 1), name='bias')
42

43 44
    a_shape = get_const_tuple(A.shape)
    w_shape = get_const_tuple(W.shape)
45
    bias_shape = get_const_tuple(bias.shape)
46 47
    dtype = A.dtype

48
    @memoize("topi.tests.test_topi_conv2d_nchw.verify_conv2d_nchw")
49 50 51
    def get_ref_data():
        a_np = np.random.uniform(size=a_shape).astype(dtype)
        w_np = np.random.uniform(size=w_shape).astype(dtype)
52
        b_np = np.random.uniform(size=bias_shape).astype(dtype)
53
        dw_np = topi.testing.dilate_python(w_np, (1, 1, dilation, dilation))
54 55 56 57 58
        c_np = topi.testing.conv2d_nchw_python(a_np, dw_np, stride, padding)
        if add_bias:
            c_np += b_np
        if add_relu:
            c_np = np.maximum(c_np, 0)
59 60 61
        return a_np, w_np, b_np, c_np

    a_np, w_np, b_np, c_np = get_ref_data()
62 63

    def check_device(device):
64 65
        ctx = tvm.context(device, 0)
        if not ctx.exist:
66 67
            print("Skip because %s is not enabled" % device)
            return
68
        print("Running on target: %s" % device)
69
        with tvm.target.create(device):
70
            C = topi.nn.conv2d(A, W, (stride, stride), padding,
71
                               (dilation, dilation), layout='NCHW', out_dtype=dtype)
72 73 74 75 76 77
            if add_bias:
                C = topi.add(C, bias)
            if add_relu:
                C = topi.nn.relu(C)
            s = topi.generic.schedule_conv2d_nchw([C])

78 79
        a = tvm.nd.array(a_np, ctx)
        w = tvm.nd.array(w_np, ctx)
80
        b = tvm.nd.array(b_np, ctx)
81
        c = tvm.nd.array(np.zeros(get_const_tuple(C.shape), dtype=C.dtype), ctx)
82
        if add_bias:
83
            func = tvm.build(s, [A, W, bias, C], device, name="relu_%d_%d_%d_%d_%d_%d_%d_%d" % (batch, in_channel, in_size, num_filter, kernel, stride, padding_sum, dilation))
84 85
            func(a, w, b, c)
        else:
86
            func = tvm.build(s, [A, W, C], device, name="relu_%d_%d_%d_%d_%d_%d_%d_%d" % (batch, in_channel, in_size, num_filter, kernel, stride, padding_sum, dilation))
87
            func(a, w, c)
88
        tvm.testing.assert_allclose(c.asnumpy(), c_np, rtol=1e-4)
89 90

    for device in get_all_backend():
91 92
        with autotvm.tophub.context(device):  # load tophub pre-tuned parameters
            check_device(device)
93

94 95 96
    if use_cudnn:
        check_device("cuda -model=unknown -libs=cudnn")

97 98

def test_conv2d_nchw():
99
    # ResNet18 workloads
100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117
    verify_conv2d_nchw(1,   3, 224,  64, 7, 2, 3)
    verify_conv2d_nchw(1,  64,  56,  64, 3, 1, 1)
    verify_conv2d_nchw(1,  64,  56,  64, 1, 1, 0)
    verify_conv2d_nchw(1,  64,  56, 128, 3, 2, 1)
    verify_conv2d_nchw(1,  64,  56, 128, 1, 2, 0)
    verify_conv2d_nchw(1, 128,  28, 128, 3, 1, 1)
    verify_conv2d_nchw(1, 128,  28, 256, 3, 2, 1)
    verify_conv2d_nchw(1, 128,  28, 256, 1, 2, 0)
    verify_conv2d_nchw(1, 256,  14, 256, 3, 1, 1)
    verify_conv2d_nchw(1, 256,  14, 512, 3, 2, 1)
    verify_conv2d_nchw(1, 256,  14, 512, 1, 2, 0)
    verify_conv2d_nchw(1, 512,   7, 512, 3, 1, 1)

    # bias, relu
    verify_conv2d_nchw(1, 64, 56, 64, 3, 1, 1, add_relu=True)
    verify_conv2d_nchw(1, 64, 56, 64, 3, 1, 1, add_bias=True)
    verify_conv2d_nchw(1, 64, 56, 64, 3, 1, 1, add_bias=True, add_relu=True)

118
    # dilation = 2
119 120
    verify_conv2d_nchw(1, 64, 56, 64, 3, 1, 1, dilation=2)

121 122 123 124
    # batch size
    verify_conv2d_nchw(4, 64, 56, 64, 3, 1, 1)
    verify_conv2d_nchw(9, 64, 56, 64, 3, 1, 1)

125
    # weird workloads
126 127 128 129 130 131 132 133 134 135
    verify_conv2d_nchw(2, 2, 2, 2, 2, 2, 2)
    verify_conv2d_nchw(3, 3, 3, 3, 3, 3, 3)
    verify_conv2d_nchw(4, 4, 4, 4, 4, 4, 4)
    verify_conv2d_nchw(5, 5, 5, 5, 5, 5, 5)
    verify_conv2d_nchw(6, 6, 6, 6, 6, 6, 6)

    # disable these tests due to some bugs of llvm with nvptx
    # verify_conv2d_nchw(1, 1, 1, 1, 1, 1, 1, dilation=1)
    # verify_conv2d_nchw(1, 1, 1, 1, 1, 1, 1, dilation=2)
    # verify_conv2d_nchw(2, 13, 71, 59, 3, 1, 1)
136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153

    # inception v3 workloads
    verify_conv2d_nchw(1,    3, 299,  32, 3, 2, 0)
    verify_conv2d_nchw(1,   32, 149,  32, 3, 1, 0)
    verify_conv2d_nchw(1,   32, 147,  64, 3, 1, 1)
    verify_conv2d_nchw(1,   64,  73,  80, 1, 1, 0)
    verify_conv2d_nchw(1,   80,  73, 192, 3, 1, 0)
    verify_conv2d_nchw(1,  192,  35,  64, 1, 1, 0)
    verify_conv2d_nchw(1,  192,  35,  48, 1, 1, 0)
    verify_conv2d_nchw(1,   48,  35,  64, 5, 1, 2)
    verify_conv2d_nchw(1,   64,  35,  96, 3, 1, 1)
    verify_conv2d_nchw(1,   96,  35,  96, 3, 1, 1)
    verify_conv2d_nchw(1,  192,  35,  32, 1, 1, 0)
    verify_conv2d_nchw(1,  256,  35,  64, 1, 1, 0)
    verify_conv2d_nchw(1,  256,  35,  48, 1, 1, 0)
    verify_conv2d_nchw(1,  288,  35,  64, 1, 1, 0)
    verify_conv2d_nchw(1,  288,  35,  48, 1, 1, 0)
    verify_conv2d_nchw(1,  288,  35, 384, 3, 2, 0)
154 155 156 157 158 159 160 161
    verify_conv2d_nchw(1,   96,  35,  96, 3, 2, 0)
    verify_conv2d_nchw(1,  768,  17, 192, 1, 1, 0)
    verify_conv2d_nchw(1,  768,  17, 128, 1, 1, 0)
    verify_conv2d_nchw(1,  128,  17, 128, 1, 1, 0)
    verify_conv2d_nchw(1,  128,  17, 192, 7, 1, 3)
    verify_conv2d_nchw(1,  128,  17, 128, 7, 1, 3)
    verify_conv2d_nchw(1,  128,  17, 192, 1, 1, 0)
    verify_conv2d_nchw(1,  768,  17, 160, 1, 1, 0)
162 163
    # disable these tests due to some bugs of llvm with nvptx
    # verify_conv2d_nchw(1,  160,  17, 160, 1, 1, 0)
164 165 166 167 168 169 170
    verify_conv2d_nchw(1,  160,  17, 192, 7, 1, 3)
    verify_conv2d_nchw(1,  160,  17, 160, 7, 1, 3)
    verify_conv2d_nchw(1,  160,  17, 192, 1, 1, 0)
    verify_conv2d_nchw(1,  192,  17, 192, 1, 1, 0)
    verify_conv2d_nchw(1,  192,  17, 192, 7, 1, 3)
    verify_conv2d_nchw(1,  192,  17, 320, 3, 2, 0)
    verify_conv2d_nchw(1,  192,  17, 192, 3, 2, 0)
171 172 173 174 175 176 177 178 179 180 181
    verify_conv2d_nchw(1, 1280,   8, 320, 1, 1, 0)
    verify_conv2d_nchw(1, 1280,   8, 384, 1, 1, 0)
    verify_conv2d_nchw(1,  384,   8, 384, 1, 1, 0)
    verify_conv2d_nchw(1,  384,   8, 384, 3, 1, 1)
    verify_conv2d_nchw(1, 1280,   8, 448, 1, 1, 0)
    verify_conv2d_nchw(1,  448,   8, 384, 3, 1, 1)
    verify_conv2d_nchw(1, 1280,   8, 192, 1, 1, 0)
    verify_conv2d_nchw(1, 2048,   8, 320, 1, 1, 0)
    verify_conv2d_nchw(1, 2048,   8, 384, 1, 1, 0)
    verify_conv2d_nchw(1, 2048,   8, 448, 1, 1, 0)
    verify_conv2d_nchw(1, 2048,   8, 192, 1, 1, 0)
182 183 184 185
    verify_conv2d_nchw(1, 1024,  19,  84, 3, 1, 1)
    verify_conv2d_nchw(1, 2048,  10, 126, 3, 1, 1)
    verify_conv2d_nchw(1,  512,   5, 126, 3, 1, 1)
    verify_conv2d_nchw(1,  256,   3, 126, 3, 1, 1)
186

187
    # Asymmetric padding
188 189 190 191 192 193 194 195 196 197
    verify_conv2d_nchw(1,   3,   35,  64,  7, 2, (0, 0, 1, 1))
    verify_conv2d_nchw(1,  64,    8, 128,  3, 1, (3, 3, 2, 2))
    verify_conv2d_nchw(1,  64,    8,  64,  1, 1, (1, 2, 2, 1))
    verify_conv2d_nchw(1,  64,   17, 192,  1, 1, (1, 2))
    verify_conv2d_nchw(1,  64,    8,  64,  3, 1, (3, 1))
    verify_conv2d_nchw(1, 128,    8, 384,  3, 1, (0, 2))
    verify_conv2d_nchw(1,  64,   35,  64,  3, 1, (1, 2), use_cudnn=True)
    verify_conv2d_nchw(1,  64,    8,  64,  1, 1, "VALID")
    verify_conv2d_nchw(1, 388,    8,  64,  3, 1, "VALID")
    verify_conv2d_nchw(1,  64,   10,  48,  3, 1, "VALID", use_cudnn=True)
198
    verify_conv2d_nchw(1, 512,   19,  64,  1, 1, "SAME")
199
    verify_conv2d_nchw(1,  64,    5,  32,  2, 1, "SAME")
200
    verify_conv2d_nchw(1,  64,    8,  64,  3, 1, "SAME", use_cudnn=True)
201 202 203 204
    verify_conv2d_nchw(1,  64,    8,  64,  3, 1, (1, 2, 2, 1), add_relu=True)
    verify_conv2d_nchw(1,  64,    8,  64,  5, 2, (1, 3), add_bias=True)
    verify_conv2d_nchw(1,  64,    8,  64,  3, 1, "VALID", add_bias=True, add_relu=True)
    verify_conv2d_nchw(1,  64,    8,  64, 24, 1, "SAME", add_bias=True, add_relu=True)
205

206 207 208

if __name__ == "__main__":
    test_conv2d_nchw()