test_topi_conv2d_winograd.py 6.08 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 18 19 20
"""Example code to do convolution."""

import numpy as np
import tvm
21
from tvm import te
22 23 24 25 26
from tvm import autotvm
from tvm.autotvm.task.space import FallbackConfigEntity
import topi
import topi.testing
from tvm.contrib.pickle_memoize import memoize
27
from topi.nn.util import get_pad_tuple
28 29 30
from topi.util import get_const_tuple


31 32 33 34 35 36 37
_conv2d_nchw_winograd_implement = {
    "arm_cpu": (topi.arm_cpu.conv2d_nchw_winograd, topi.arm_cpu.schedule_conv2d_nchw_winograd),
    "cuda": (topi.cuda.conv2d_nchw_winograd, topi.cuda.schedule_conv2d_nchw_winograd),
    "mali": (topi.mali.conv2d_nchw_winograd, topi.mali.schedule_conv2d_nchw_winograd),
}


hlu1 committed
38 39
def verify_conv2d_nchw(batch, in_channel, in_size, num_filter, kernel, stride, padding, dilation=1, add_bias=False, add_relu=False,
        devices=['cuda', 'llvm -device=arm_cpu', 'opencl -device=mali']):
40 41 42
    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))
43 44 45

    in_height = in_width = in_size

46 47 48
    A = te.placeholder((batch, in_channel, in_height, in_width), name='A')
    W = te.placeholder((num_filter, in_channel, kernel, kernel), name='W')
    bias = te.placeholder((num_filter, 1, 1), name='bias')
49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77

    a_shape = get_const_tuple(A.shape)
    w_shape = get_const_tuple(W.shape)
    bias_shape = get_const_tuple(bias.shape)
    dtype = A.dtype

    @memoize("topi.tests.test_topi_conv2d_nchw.verify_conv2d_nchw")
    def get_ref_data():
        a_np = np.random.uniform(size=a_shape).astype(dtype)
        w_np = np.random.uniform(size=w_shape).astype(dtype)
        b_np = np.random.uniform(size=bias_shape).astype(dtype)
        dw_np = topi.testing.dilate_python(w_np, (1, 1, dilation, dilation))
        c_np = topi.testing.conv2d_nchw_python(a_np, dw_np, stride, padding)
        if add_bias:
            b_np = np.random.uniform(size=bias_shape).astype(dtype)
            c_np += b_np
        if add_relu:
            c_np = np.maximum(c_np, 0)
        return a_np, w_np, b_np, c_np

    a_np, w_np, b_np, c_np = get_ref_data()

    def check_device(device):
        ctx = tvm.context(device, 0)
        if not ctx.exist:
            print("Skip because %s is not enabled" % device)
            return
        print("Running on target: %s" % device)
        with tvm.target.create(device):
78 79
            fcompute, fschedule = topi.testing.dispatch(device, _conv2d_nchw_winograd_implement)
            C = fcompute(A, W, stride, padding, dilation, dtype)
80 81 82 83
            if add_bias:
                C = topi.add(C, bias)
            if add_relu:
                C = topi.nn.relu(C)
84
            s = fschedule([C])
85 86 87 88 89 90

        a = tvm.nd.array(a_np, ctx)
        w = tvm.nd.array(w_np, ctx)
        b = tvm.nd.array(b_np, ctx)
        c = tvm.nd.array(np.zeros(get_const_tuple(C.shape), dtype=C.dtype), ctx)
        if add_bias:
91
            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))
92 93
            func(a, w, b, c)
        else:
94
            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))
95
            func(a, w, c)
96

97
        rtol = 1e-3
98
        tvm.testing.assert_allclose(c.asnumpy(), c_np, rtol=rtol)
99 100


hlu1 committed
101
    for device in devices:
102 103 104 105
        check_device(device)


def test_conv2d_nchw():
106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143
    # inception v3 workloads
    verify_conv2d_nchw(1, 128, 17, 192, 7, 1, 3, devices=['cuda'])
    verify_conv2d_nchw(1, 128, 17, 128, 7, 1, 3, devices=['cuda'])
    verify_conv2d_nchw(1, 160, 17, 160, 7, 1, 3, devices=['cuda'])

    # resnet 18 workloads
    verify_conv2d_nchw(1, 64, 56, 64, 3, 1, 1)
    verify_conv2d_nchw(1, 128, 28, 128, 3, 1, 1)
    verify_conv2d_nchw(1, 256, 14, 256, 3, 1, 1)
    verify_conv2d_nchw(1, 512, 7, 512, 3, 1, 1)
    verify_conv2d_nchw(1, 48,  35, 64, 5, 1, 2, devices=['cuda'])

    # batch size = 2
    verify_conv2d_nchw(2, 64, 56, 64, 3, 1, 1)

    # relu, bias
    verify_conv2d_nchw(2, 64, 56, 64, 3, 1, 1, add_bias=True)
    verify_conv2d_nchw(2, 64, 56, 64, 3, 1, 1, add_relu=True)
    verify_conv2d_nchw(2, 64, 56, 64, 3, 1, 1, add_relu=True, add_bias=True)

    # weird workloads
    verify_conv2d_nchw(1, 1, 1, 1, 3, 1, 1)
    verify_conv2d_nchw(3, 3, 3, 3, 3, 1, 1)
    verify_conv2d_nchw(2, 13, 71, 59, 3, 1, 1)

    # Asymmetric padding
    verify_conv2d_nchw(1,  48, 56,  48, 3, 1, (1, 1, 1, 1))
    verify_conv2d_nchw(1,  64, 28,  64, 3, 1, (1, 1, 1, 1))
    verify_conv2d_nchw(1, 128, 14, 128, 3, 1, (1, 1))
    verify_conv2d_nchw(1, 512,  7, 512, 3, 1, "SAME")
    verify_conv2d_nchw(2, 13,  71,  59, 3, 1, (1, 1, 1, 1))
    verify_conv2d_nchw(2,  48, 56,  48, 3, 1, (1, 1, 1, 1), add_bias=True)
    verify_conv2d_nchw(2,  48, 56,  48, 3, 1, (1, 1), add_relu=True)
    verify_conv2d_nchw(2,  48, 56,  48, 3, 1, "SAME", add_relu=True, add_bias=True)
    verify_conv2d_nchw(1,  64, 17, 192, 7, 1, (3, 1), devices=['cuda'])
    verify_conv2d_nchw(1,  64, 17,  64, 7, 1, (3, 3, 2, 2), devices=['cuda'])
    verify_conv2d_nchw(1, 160, 17, 160, 7, 1, "SAME", devices=['cuda'])
    verify_conv2d_nchw(1,  48, 35,  48, 5, 1, "VALID", devices=['cuda'])
144 145


146 147
if __name__ == "__main__":
    test_conv2d_nchw()