test_lang_tensor_overload_op.py 8.51 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
import numpy as np
import tvm
import topi
import topi.testing
from topi.util import get_const_tuple


def test_operator_type_and_tags():
    k = 1
    n = tvm.var('n')
    A = tvm.placeholder((), name='A')
    B = tvm.placeholder((10, 5), name='B')
    B1 = B[0]
    B2 = B[0,0]

    assert isinstance(k + n, tvm.expr.Expr)
    assert isinstance(n + n, tvm.expr.Expr)
18 19 20 21 22
    assert isinstance(k + A, tvm.tensor.Tensor)
    assert isinstance(A + k, tvm.tensor.Tensor)
    assert isinstance(n + A, tvm.tensor.Tensor)
    assert isinstance(A + n, tvm.tensor.Tensor)
    assert isinstance(A + A, tvm.tensor.Tensor)
23 24 25 26 27 28 29 30 31 32 33 34 35

    assert isinstance(k + B, tvm.tensor.Tensor)
    assert isinstance(B + k, tvm.tensor.Tensor)
    assert isinstance(n + B, tvm.tensor.Tensor)
    assert isinstance(B + n, tvm.tensor.Tensor)
    assert isinstance(A + B, tvm.tensor.Tensor)
    assert isinstance(B + A, tvm.tensor.Tensor)
    assert isinstance(B + B, tvm.tensor.Tensor)

    assert (k + B).op.tag == topi.tag.ELEMWISE
    assert (B + k).op.tag == topi.tag.ELEMWISE
    assert (n + B).op.tag == topi.tag.ELEMWISE
    assert (B + n).op.tag == topi.tag.ELEMWISE
36 37
    assert (A + B).op.tag == topi.tag.BROADCAST
    assert (B + A).op.tag == topi.tag.BROADCAST
38 39 40 41 42 43 44
    assert (B + B).op.tag == topi.tag.BROADCAST

    assert isinstance(k + B2, tvm.expr.Expr)
    assert isinstance(B2 + k, tvm.expr.Expr)
    assert isinstance(n + B2, tvm.expr.Expr)
    assert isinstance(B2 + n, tvm.expr.Expr)
    assert isinstance(B2 + B2, tvm.expr.Expr)
45 46
    assert isinstance(B2 + A, tvm.tensor.Tensor)
    assert isinstance(A + B2, tvm.tensor.Tensor)
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248
    assert isinstance(B2 + B, tvm.tensor.Tensor)
    assert isinstance(B + B2, tvm.tensor.Tensor)


def test_combination():
    k = 3
    n = 5
    m = 10
    x = tvm.var('x')
    A = tvm.placeholder((n, m), name='A')
    B = tvm.placeholder((n, m), name='B')
    C = tvm.placeholder((n, m), name='C')
    D = k + A - B * C / x
    s = tvm.create_schedule(D.op)
    foo = tvm.build(s, [x, A, B, C, D], "llvm")
    ctx = tvm.cpu(0)
    x = 2
    a = tvm.nd.array(np.random.uniform(size=(n, m)).astype(A.dtype), ctx)
    b = tvm.nd.array(np.random.uniform(size=(n, m)).astype(B.dtype), ctx)
    c = tvm.nd.array(np.random.uniform(size=(n, m)).astype(C.dtype), ctx)
    d = tvm.nd.array(np.zeros((n, m), dtype=D.dtype), ctx)
    foo(x, a, b, c, d)
    np.testing.assert_allclose(d.asnumpy(), k + a.asnumpy() - b.asnumpy() * c.asnumpy() / x)


def verify_tensor_scalar_bop(shape, typ="add"):
    """Verify non-constant Tensor and scalar binary operations."""
    sh = [tvm.var('n%d' % i) for i in range(0, len(shape))]
    k = tvm.var('k')
    A = tvm.placeholder(sh, name='A')
    if typ == "add":
        B = A + k
    elif typ == "sub":
        B = A - k
    elif typ == "mul":
        B = A * k
    elif typ == "div":
        B = A / k
    else:
        raise NotImplementedError()

    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):
            s = topi.generic.schedule_elemwise(B)

        k_ = 2
        foo = tvm.build(s, [A, B, k] + sh, device, name="tensor_scalar_" + typ)
        a_npy = np.random.uniform(size=shape).astype(A.dtype)
        if typ == "add":
            b_npy = a_npy + k_
        elif typ == "sub":
            b_npy = a_npy - k_
        elif typ == "mul":
            b_npy = a_npy * k_
        elif typ == "div":
            b_npy = a_npy / k_
        else:
            raise NotImplementedError()

        a_nd = tvm.nd.array(a_npy, ctx)
        b_nd = tvm.nd.array(np.empty(b_npy.shape).astype(B.dtype), ctx)
        foo(a_nd, b_nd, k_, *shape)
        np.testing.assert_allclose(b_nd.asnumpy(), b_npy, rtol=1e-5)

    for device in ['llvm', 'cuda', 'opencl', 'metal', 'rocm', 'vulkan']:
        check_device(device)


def verify_broadcast_bop(lhs_shape, rhs_shape, typ="add"):
    A = tvm.placeholder(shape=lhs_shape, name="A")
    B = tvm.placeholder(shape=rhs_shape, name="B")
    if typ == "add":
        C = A + B
    elif typ == "sub":
        C = A - B
    elif typ == "mul":
        C = A * B
    elif typ == "div":
        C = A / B
    else:
        raise NotImplementedError()

    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):
            s = topi.generic.schedule_broadcast(C)

        foo = tvm.build(s, [A, B, C], device, name="broadcast_binary" + "_" + typ)
        lhs_npy = np.random.uniform(size=lhs_shape).astype(A.dtype)
        rhs_npy = np.random.uniform(size=rhs_shape).astype(A.dtype)
        if typ == "add":
            out_npy = lhs_npy + rhs_npy
        elif typ == "sub":
            out_npy = lhs_npy - rhs_npy
        elif typ == "mul":
            out_npy = lhs_npy * rhs_npy
        elif typ == "div":
            rhs_npy = np.abs(rhs_npy) + 0.001
            out_npy = lhs_npy / rhs_npy
        else:
            raise NotImplementedError()

        lhs_nd = tvm.nd.array(lhs_npy, ctx)
        rhs_nd = tvm.nd.array(rhs_npy, ctx)
        out_nd = tvm.nd.array(np.empty(out_npy.shape).astype(B.dtype), ctx)
        for _ in range(1):
            foo(lhs_nd, rhs_nd, out_nd)
        np.testing.assert_allclose(out_nd.asnumpy(), out_npy, rtol=1E-4, atol=1E-4)

    for device in ['llvm', 'cuda', 'opencl', 'metal', 'rocm', 'vulkan']:
        check_device(device)


def verify_conv2d_scalar_bop(batch, in_size, in_channel, num_filter, kernel, stride, padding, typ="add"):
    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)

        k = 10.0
        with tvm.target.create(device):
            A = tvm.placeholder((batch, in_channel, in_size, in_size), name='A')
            W = tvm.placeholder((num_filter, in_channel, kernel, kernel), name='W')
            B = topi.nn.conv2d(A, W, stride, padding)
            if typ == "add":
                C = B + k
            elif typ == "sub":
                C = B - k
            elif typ == "mul":
                C = B * k
            elif typ == "div":
                C = B / k
            else:
                raise NotImplementedError()
            s = topi.generic.schedule_conv2d_nchw([C])

        foo = tvm.build(s, [A, W, B, C], device, name="conv2d_scalar_" + typ)

        a_npy = np.random.uniform(size=get_const_tuple(A.shape)).astype(A.dtype)
        w_npy = np.random.uniform(size=get_const_tuple(W.shape)).astype(W.dtype)
        b_npy = topi.testing.conv2d_nchw_python(a_npy, w_npy, stride, padding)
        c_npy = np.random.uniform(size=get_const_tuple(B.shape)).astype(B.dtype)
        if typ == "add":
            c_npy = b_npy + k
        elif typ == "sub":
            c_npy = b_npy - k
        elif typ == "mul":
            c_npy = b_npy * k
        elif typ == "div":
            c_npy = b_npy / k
        else:
            raise NotImplementedError()

        a_nd = tvm.nd.array(a_npy, ctx)
        w_nd = tvm.nd.array(w_npy, ctx)
        b_nd = tvm.nd.array(np.empty(b_npy.shape).astype(B.dtype), ctx)
        c_nd = tvm.nd.array(np.empty(c_npy.shape).astype(C.dtype), ctx)
        foo(a_nd, w_nd, b_nd, c_nd)
        np.testing.assert_allclose(c_nd.asnumpy(), c_npy, rtol=1E-4, atol=1E-4)

    for device in ['llvm', 'cuda', 'opencl', 'metal', 'rocm', 'vulkan']:
        check_device(device)


def test_tensor_scalar_bop():
    verify_tensor_scalar_bop((1,), typ="add")
    verify_tensor_scalar_bop((3, 5), typ="sub")
    verify_tensor_scalar_bop((1, 3, 5), typ="mul")
    verify_tensor_scalar_bop((2, 3, 1, 32), typ="div")


def test_broadcast_bop():
    verify_broadcast_bop((2, 3), (), typ="add")
    verify_broadcast_bop((5, 2, 3), (1,), typ="add")
    verify_broadcast_bop((1, 32), (64, 32), typ="sub")
    verify_broadcast_bop((5, 64, 128), (2, 5, 64, 1), typ="mul")
    verify_broadcast_bop((2, 3, 1, 32), (64, 32), typ="div")


def test_conv2d_scalar_bop():
    verify_conv2d_scalar_bop(1, 16, 4, 4, 3, 1, 1, typ="add")
    verify_conv2d_scalar_bop(1, 32, 2, 1, 3, 1, 1, typ="sub")
    verify_conv2d_scalar_bop(1, 32, 1, 1, 3, 1, 1, typ="mul")
    verify_conv2d_scalar_bop(1, 16, 2, 1, 3, 1, 1, typ="div")


if __name__ == "__main__":
    test_operator_type_and_tags()
    test_combination()
    test_tensor_scalar_bop()
    test_broadcast_bop()
249
    test_conv2d_scalar_bop()