test_lang_tensor.py 9.73 KB
Newer Older
tqchen committed
1
import tvm
2
from topi.nn.pooling import pool
tqchen committed
3 4

def test_tensor():
5 6 7
    m = tvm.var('m')
    n = tvm.var('n')
    l = tvm.var('l')
tqchen committed
8 9
    A = tvm.placeholder((m, l), name='A')
    B = tvm.placeholder((n, l), name='B')
10
    T = tvm.compute((m, n, l), lambda i, j, k: A[i, k] * B[j, k])
tqchen committed
11
    print(T)
tqchen committed
12
    print(T.op.body)
tqchen committed
13
    assert(tuple(T.shape) == (m, n, l))
14
    assert(isinstance(A.op, tvm.tensor.PlaceholderOp))
15 16 17 18 19
    assert(A == A)
    assert(T.op.output(0) == T)
    assert(T.op.output(0).__hash__() == T.__hash__())
    d = {T.op.output(0) : 1}
    assert(d[T] == 1)
ziheng committed
20
    assert(T[0][0][0].astype('float16').dtype == 'float16')
21

tqchen committed
22

23 24 25 26 27 28 29 30 31 32 33
def test_rank_zero():
    m = tvm.var('m')
    A = tvm.placeholder((m,), name='A')
    scale = tvm.placeholder((), name='s')
    k = tvm.reduce_axis((0, m), name="k")
    T = tvm.compute((), lambda : tvm.sum(A[k] * scale(), axis=k))
    print(T)
    print(T.op.body)
    assert(tuple(T.shape) == ())


34
def test_conv1d():
35
    n = tvm.var('n')
36 37 38 39 40 41 42 43
    A = tvm.placeholder((n+2), name='A')
    def computeB(ii):
        i = ii + 1
        return A[i-1] + A[i] + A[i+1]
    B = tvm.compute(n, computeB)


def test_tensor_slice():
44
    n = tvm.var('n')
45 46 47 48
    A = tvm.compute((n, n), lambda i, j: 1)
    B = tvm.compute((n,), lambda i: A[0][i] + A[0][i])


49 50 51 52 53 54 55 56 57 58
def test_tensor_reduce_multi_axis():
    m = tvm.var('m')
    n = tvm.var('n')
    A = tvm.placeholder((m, n), name='A')
    k1 = tvm.reduce_axis((0, n), "k")
    k2 = tvm.reduce_axis((0, m), "k")
    C = tvm.compute((1,), lambda _: tvm.sum(A[k1, k2], axis=(k1, k2)))
    C = tvm.compute((1,), lambda _: tvm.sum(A[k1, k2], axis=[k1, k2]))


ziheng committed
59 60 61 62 63 64 65 66 67 68 69 70 71 72
def test_tensor_comm_reducer():
    m = tvm.var('m')
    n = tvm.var('n')
    A = tvm.placeholder((m, n), name='A')
    k = tvm.reduce_axis((0, n), "k")
    mysum = tvm.comm_reducer(lambda x, y: x+y, lambda t: tvm.const(0, dtype=t))
    C = tvm.compute((m,), lambda i: mysum(A[i, k], axis=k))

def test_tensor_comm_reducer_overload():
    m = tvm.var('m')
    n = tvm.var('n')
    mysum = tvm.comm_reducer(lambda x, y: x+y, lambda t: tvm.const(0, dtype=t))
    sum_res = mysum(m, n)

tqchen committed
73
def test_tensor_reduce():
74 75 76
    m = tvm.var('m')
    n = tvm.var('n')
    l = tvm.var('l')
tqchen committed
77 78
    A = tvm.placeholder((m, l), name='A')
    B = tvm.placeholder((n, l), name='B')
79
    T = tvm.compute((m, n, l), lambda i, j, k: A[i, k] * B[j, k])
80
    rv = tvm.reduce_axis((0, A.shape[1]), "k")
81
    C = tvm.compute((m, n), lambda i, j: tvm.sum(T(i, j, rv+1), axis=rv))
82 83 84 85 86
    # json load save
    C_json = tvm.save_json(C)
    C_loaded = tvm.load_json(C_json)
    assert(isinstance(C_loaded, tvm.tensor.Tensor))
    assert(str(C_loaded) == str(C))
tqchen committed
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
def test_tensor_compute1():
    m = 1024
    factor = 16
    dtype = 'float32'

    def intrin_vadd(n):
        x = tvm.placeholder((n,))
        y = tvm.placeholder((n,))
        z = tvm.compute(x.shape, lambda i: x[i] + y[i])

        def intrin_func(ins, outs):
            ib = tvm.ir_builder.create()
            ib.emit(tvm.call_extern(outs[0].dtype, 'vadd', ins[0].access_ptr("r"), ins[1].access_ptr('r'), outs[0].access_ptr('wr')))
            return ib.get()

        with tvm.build_config(offset_factor=n):
            return tvm.decl_tensor_intrin(z.op, intrin_func)

    vadd = intrin_vadd(factor)

    A = tvm.placeholder((m//factor, factor), name="A", dtype=dtype)
    B = tvm.placeholder((m//factor, factor), name="B", dtype=dtype)
    C = tvm.compute((m//factor, factor),
          lambda i: vadd(A[i, 0:factor], B[i, 0:factor]))

    s = tvm.create_schedule(C.op)
    stmt = tvm.lower(s, [A, B, C], simple_mode=True)
    assert isinstance(stmt.body.body, tvm.stmt.Evaluate)

def test_tensor_compute2():
    M = 2048
    N = 1024
    L = 1024
    factor = 16
    factor1 = 32
    factor2 = 32
    dtype = 'float32'

    def intrin_gemm(m, n, l):
        k = tvm.reduce_axis((0, l))
        x = tvm.placeholder((m, l))
        y = tvm.placeholder((n, l))
        # in theory, no relation
        z = tvm.compute((m, n), lambda i, j: tvm.sum(x[i][k] * y[j][k], axis=k))

        def intrin_func(ins, outs):
            x_ptr = ins[0].access_ptr("r")
            y_ptr = ins[1].access_ptr("r")
            z_ptr = outs[0].access_ptr("w")
            body = tvm.call_packed(
                "gemv", x_ptr, y_ptr, z_ptr, m, n, l)
            reset = tvm.call_packed(
                "fill_zero", z_ptr, m, n)
            update = tvm.call_packed(
                "gemv_add", x_ptr, y_ptr, z_ptr, m, n, l)
            return body, reset, update

        with tvm.build_config(offset_factor=n):
            return tvm.decl_tensor_intrin(z.op, intrin_func)

    vgemm = intrin_gemm(factor1, factor2, factor)

    A = tvm.placeholder((M//factor1, L//factor, factor1, factor), name="A", dtype=dtype)
    B = tvm.placeholder((N//factor2, L//factor, factor2, factor), name="B", dtype=dtype)
    k = tvm.reduce_axis((0, L//factor), name='k')
    C = tvm.compute((M//factor1, N//factor2, factor1, factor2),
          lambda i, j: vgemm(A[i, k, 0:factor1, 0:factor], B[j, k, 0:factor2, 0:factor], reduce_axis=k))

    s = tvm.create_schedule(C.op)
    stmt = tvm.lower(s, [A, B, C], simple_mode=True)
    assert isinstance(stmt.body.body.body.first, tvm.stmt.Evaluate)
    assert isinstance(stmt.body.body.body.rest.body, tvm.stmt.Evaluate)
160

161
def test_tensor_scan():
162 163
    m = tvm.var("m")
    n = tvm.var("n")
164 165
    x = tvm.placeholder((m, n))
    s = tvm.placeholder((m, n))
166 167
    res = tvm.scan(tvm.compute((1, n), lambda _, i: x[0, i]),
                   tvm.compute((m, n), lambda t, i: s[t-1, i] + x[t, i]),
168 169 170
                   s)
    assert tuple(res.shape) == (m, n)

171
def test_scan_multi_out():
172 173
    m = tvm.var("m")
    n = tvm.var("n")
174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191
    x1 = tvm.placeholder((m, n))
    s1 = tvm.placeholder((m, n))
    x2 = tvm.placeholder((m, n))
    s2 = tvm.placeholder((m, n))
    s1_init = tvm.compute((1, n), lambda _, i: x1[0, i])
    s2_init = tvm.compute((1, n), lambda _, i: x2[0, i])
    s1_update = tvm.compute((m, n), lambda t, i: s1[t-1, i] + s2[t-1, i] + x1[t, i])
    s2_update = tvm.compute((m, n), lambda t, i: x2[t, i] + s2[t-1,i])

    r0, r1 = tvm.scan([s1_init, s2_init],
                      [s1_update, s2_update],
                      [s1, s2])
    assert(r0.value_index == 0)
    assert(r1.value_index == 1)
    json_str = tvm.save_json(r0.op)
    zz = tvm.load_json(json_str)
    assert isinstance(zz, tvm.tensor.ScanOp)

192
def test_extern():
193
    m = tvm.var('m')
194 195 196 197 198 199 200 201 202 203
    A = tvm.placeholder((m,), name='A')

    def extern_func(ins, outs):
        assert(isinstance(ins[0], tvm.schedule.Buffer))
        return tvm.call_packed("myadd", ins[0].data, outs[0].data, m)
    B = tvm.extern((m,), [A], extern_func)
    assert(tuple(B.shape) == (m,))


def test_extern_multi_out():
204
    m = tvm.var('m')
205 206 207 208 209 210 211 212 213 214 215
    A = tvm.placeholder((m,), name='A')
    B = tvm.compute((m,), lambda i: A[i] * 10)

    def extern_func(ins, outs):
        assert(isinstance(ins[0], tvm.schedule.Buffer))
        return tvm.call_packed(
            "myadd", ins[0].data, outs[0].data, outs[1].data, m)
    res = tvm.extern([A.shape, A.shape], [A, B], extern_func)
    assert(len(res) == 2)
    assert(res[1].value_index == 1)

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252
def test_tuple_inputs():
    m = tvm.var('m')
    n = tvm.var('n')
    A0 = tvm.placeholder((m, n), name='A0')
    A1 = tvm.placeholder((m, n), name='A1')
    T0, T1 = tvm.compute((m, n), lambda i, j: (A0[i, j] * 2, A1[i, j] * 3), name='T')
    s = tvm.create_schedule(T0.op)

    for i in range(len(T0.shape)):
      assert(T0.shape[i] == T1.shape[i])
    assert(T0.op == T1.op)
    assert(T0.value_index == 0)
    assert(T1.value_index == 1)

def test_tuple_with_different_deps():
    m = tvm.var('m')
    n = tvm.var('n')
    A0 = tvm.placeholder((m, n), name='A1')
    A1 = tvm.placeholder((m, n), name='A2')
    B0, B1 = tvm.compute((m, n), lambda i, j: (A0[i, j] * 2, A1[i, j] * 3), name='B')
    C = tvm.compute((m, n), lambda i, j: B0[i, j] + 4, name='C')

    s = tvm.create_schedule(C.op)
    xo, xi = s[C].split(C.op.axis[0], factor=10)
    s[B0.op].compute_at(s[C], xo)
    sch = s.normalize()
    bounds = tvm.schedule.InferBound(sch)
    stmt = tvm.schedule.ScheduleOps(sch, bounds)

    def get_B1_realize(x):
        if isinstance(x, tvm.stmt.Realize) and \
           x.func == B1.op and x.value_index == 1:
            ret.append(x)
    ret = []
    tvm.ir_pass.PostOrderVisit(stmt, get_B1_realize)

    assert stmt.node == C.op and len(ret) == 1
253

254 255 256 257 258 259

def test_tensor_inputs():
    x = tvm.placeholder((1,), name='x')
    y = tvm.compute(x.shape, lambda i: x[i] + x[i])
    assert tuple(y.op.input_tensors) == (x,)

260

261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288
def test_tensor_pool():
    def intrin_pool():
        A = tvm.placeholder((64, 16, 16), name='A')
        kh = tvm.reduce_axis((0, 3), name='kh')
        kw = tvm.reduce_axis((0, 3), name='kw')
        P = tvm.compute((64, 14, 14),
                        lambda c, oh, ow: tvm.max(A[c, oh + kh, ow + kw],
                                                  axis=[kh, kw]),
                        name='p')

        def intrin_func(ins, outs):
            dinp = ins[0]
            dout = outs[0]
            return tvm.call_packed("op", dinp, dout)

        with tvm.build_config(offset_factor=1):
            return tvm.decl_tensor_intrin(P.op, intrin_func)

    A = tvm.placeholder((1, 64, 16, 16), name='A')
    P = pool(data=A, kernel=(3, 3), stride=(1, 1), padding=(0, 0, 0, 0),
             pool_type='max')
    s = tvm.create_schedule(P.op)
    _, oh, _, _ = P.op.axis
    intrin = intrin_pool()
    s[P].tensorize(oh, intrin)
    tvm.lower(s, [A, P])


tqchen committed
289
if __name__ == "__main__":
290
    test_rank_zero()
291
    test_tensor_inputs()
292
    test_tensor_reduce_multi_axis()
293 294
    test_conv1d()
    test_tensor_slice()
tqchen committed
295
    test_tensor()
296 297
    test_tensor_compute1()
    test_tensor_compute2()
tqchen committed
298
    test_tensor_reduce()
299
    test_tensor_scan()
300
    test_scan_multi_out()
301 302
    test_extern()
    test_extern_multi_out()
303 304
    test_tuple_inputs()
    test_tuple_with_different_deps()
305
    test_tensor_pool()