Skip to content

T
tic
Switch branch/tag
Find file
Normal viewHistoryPermalink
test_topi_group_conv2d.py 8.73 KB
Newer Older
[TOPI][CUDA] int8 group conv2d (#2075)
5c410c4c
 
Wuwei Lin committed
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 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 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 
"""Example code to do group convolution."""

import numpy as np
import tvm
from tvm import autotvm
from tvm.autotvm.task.space import FallbackConfigEntity
import topi
import topi.testing
from tvm.contrib.pickle_memoize import memoize
from topi.util import get_const_tuple

from common import get_all_backend, NCHWcInt8Fallback


def verify_group_conv2d_nchw(batch, in_channel, in_size, num_filter, kernel, stride, padding, dilation, groups, add_bias=False, add_relu=False):
    print("Workload: (%d, %d, %d, %d, %d, %d, %d, %d, %d)" %
        (batch, in_channel, in_size, num_filter,
         kernel, stride, padding, dilation, groups))

    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 // groups, kernel, kernel), name='W')
    bias = tvm.placeholder((num_filter, 1, 1), name='bias')

    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_group_conv2d.verify_group_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, groups).astype(dtype)

        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
        if device == "cuda" and not tvm.contrib.nvcc.have_int8(ctx.compute_version):
            print("Skip because int8 intrinsics are not available")
            return

        print("Running on target: %s" % device)
        with tvm.target.create(device):
            C = topi.nn.group_conv2d_nchw(A, W, stride, padding, dilation, groups, out_dtype=dtype)
            if add_bias:
                C = topi.add(C, bias)
            if add_relu:
                C = topi.nn.relu(C)
            s = topi.generic.schedule_group_conv2d_nchw([C])

        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:
            func = tvm.build(s, [A, W, bias, C], device, name="relu_%d_%d_%d_%d_%d_%d_%d_%d_%d" %\
                (batch, in_channel, in_size, num_filter, kernel, stride, padding, dilation, groups))
            func(a, w, b, c)
        else:
            func = tvm.build(s, [A, W, C], device, name="relu_%d_%d_%d_%d_%d_%d_%d_%d_%d" % \
            (batch, in_channel, in_size, num_filter, kernel, stride, padding, dilation, groups))
            func(a, w, c)
        tvm.testing.assert_allclose(c.asnumpy(), c_np, rtol=1e-5)

    for device in ["llvm"]:
        check_device(device)


oc_block_factor = 4


def verify_group_conv2d_NCHWc_int8(batch, in_channel, in_size, num_filter, kernel, stride, padding, dilation, groups, add_bias=False, add_relu=False):
    print("Workload: (%d, %d, %d, %d, %d, %d, %d, %d, %d)" %
        (batch, in_channel, in_size, num_filter,
         kernel, stride, padding, dilation, groups))

    in_height = in_width = in_size

    A = tvm.placeholder((batch, in_channel, in_height, in_width), name='A', dtype='int8')
    W = tvm.placeholder((num_filter, in_channel // groups, kernel, kernel), name='W', dtype='int8')
    bias = tvm.placeholder((num_filter // oc_block_factor, 1, 1, oc_block_factor), name='bias',
                            dtype='int8')

    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_group_conv2d.verify_group_conv2d_NCHWc_int8")
    def get_ref_data():
        a_np = np.random.randint(low=-128, high=127, size=a_shape).astype(dtype)
        w_np = np.random.randint(low=-128, high=128, 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, groups).astype(dtype)

        # convert to NCHWc
        _, _, out_height, out_width = c_np.shape
        c_np = c_np.reshape((batch, num_filter // oc_block_factor, oc_block_factor, \
                out_height, out_width)).transpose(0, 1, 3, 4, 2)

        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
        if device == "cuda" and not tvm.contrib.nvcc.have_int8(ctx.compute_version):
            print("Skip because int8 intrinsics are not available")
            return

        print("Running on target: %s" % device)
        with tvm.target.create(device):
            C = topi.nn.group_conv2d_nchw(A, W, stride, padding, dilation, groups, out_dtype=dtype)
            if add_bias:
                C = topi.add(C, bias)
            if add_relu:
                C = topi.nn.relu(C)
            s = topi.generic.schedule_group_conv2d_nchw([C])

        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:
            func = tvm.build(s, [A, W, bias, C], device, name="relu_%d_%d_%d_%d_%d_%d_%d_%d_%d" %\
                (batch, in_channel, in_size, num_filter, kernel, stride, padding, dilation, groups))
            func(a, w, b, c)
        else:
            func = tvm.build(s, [A, W, C], device, name="relu_%d_%d_%d_%d_%d_%d_%d_%d_%d" % \
            (batch, in_channel, in_size, num_filter, kernel, stride, padding, dilation, groups))
            func(a, w, c)
        tvm.testing.assert_allclose(c.asnumpy(), c_np, rtol=1e-5)

    for device in ["cuda"]:
        check_device(device)


def test_group_conv2d_nchw():
    # ResNeXt-50 workload
    verify_group_conv2d_nchw(1, 128, 56, 128, 3, 1, 1, 1, 32)
    verify_group_conv2d_nchw(1, 256, 56, 256, 3, 2, 1, 1, 32)
    verify_group_conv2d_nchw(1, 256, 28, 256, 3, 1, 1, 1, 32)
    verify_group_conv2d_nchw(1, 512, 28, 512, 3, 2, 1, 1, 32)
    verify_group_conv2d_nchw(1, 512, 14, 512, 3, 1, 1, 1, 32)
    verify_group_conv2d_nchw(1, 1024, 14, 1024, 3, 2, 1, 1, 32)
    verify_group_conv2d_nchw(1, 1024, 7, 1024, 3, 1, 1, 1, 32)

    # bias, relu
    verify_group_conv2d_nchw(1, 128, 56, 128, 3, 1, 1, 1, 32, add_relu=True)
    verify_group_conv2d_nchw(1, 128, 56, 128, 3, 1, 1, 1, 32, add_bias=True)
    verify_group_conv2d_nchw(1, 128, 56, 128, 3, 1, 1, 1, 32, add_relu=True,
                             add_bias=True)

    # dilation
    verify_group_conv2d_NCHWc_int8(1, 128, 56, 128, 3, 1, 1, 2, 32)

    # batch size
    verify_group_conv2d_nchw(2, 128, 56, 128, 3, 1, 1, 1, 32)
    verify_group_conv2d_nchw(9, 128, 56, 128, 3, 1, 1, 1, 32)



def test_group_conv2d_NCHWc_int8():
    with NCHWcInt8Fallback():
        # ResNeXt-50 workload
        verify_group_conv2d_NCHWc_int8(1, 128, 56, 128, 3, 1, 1, 1, 32)
        verify_group_conv2d_NCHWc_int8(1, 256, 56, 256, 3, 2, 1, 1, 32)
        verify_group_conv2d_NCHWc_int8(1, 256, 28, 256, 3, 1, 1, 1, 32)
        verify_group_conv2d_NCHWc_int8(1, 512, 28, 512, 3, 2, 1, 1, 32)
        verify_group_conv2d_NCHWc_int8(1, 512, 14, 512, 3, 1, 1, 1, 32)
        verify_group_conv2d_NCHWc_int8(1, 1024, 14, 1024, 3, 2, 1, 1, 32)
        verify_group_conv2d_NCHWc_int8(1, 1024, 7, 1024, 3, 1, 1, 1, 32)

        # bias, relu
        verify_group_conv2d_NCHWc_int8(1, 128, 56, 128, 3, 1, 1, 1, 32, add_relu=True)
        verify_group_conv2d_NCHWc_int8(1, 128, 56, 128, 3, 1, 1, 1, 32, add_bias=True)
        verify_group_conv2d_NCHWc_int8(1, 128, 56, 128, 3, 1, 1, 1, 32, add_relu=True,
                                       add_bias=True)
        # dilation
        verify_group_conv2d_NCHWc_int8(1, 128, 56, 128, 3, 1, 1, 2, 32)

        # batch size
        verify_group_conv2d_NCHWc_int8(2, 128, 56, 128, 3, 1, 1, 1, 32)
        verify_group_conv2d_NCHWc_int8(9, 128, 56, 128, 3, 1, 1, 1, 32)


if __name__ == "__main__":
    test_group_conv2d_nchw()
    test_group_conv2d_NCHWc_int8()