test_topi_sparse.py 15 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 21 22 23 24 25
"""Test code for sparse operator"""
import numpy as np
import tvm
import topi
import topi.testing
from topi.util import get_const_tuple
import tvm.contrib.sparse as tvmsp
from collections import namedtuple
import time
26
import scipy.sparse as sp
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

def verify_dynamic_csrmv(batch, in_dim, out_dim, use_bias=True):
    nr, nc, n = tvm.var("nr"), tvm.var("nc"), tvm.var("n")
    dtype = 'float32'
    A = tvmsp.placeholder(shape=(nr, nc), nonzeros=n, dtype=dtype, name='A')
    B = tvm.placeholder((in_dim, 1), name='B')
    C = tvm.placeholder((nr,), name='C')
    D = topi.sparse.csrmv(A, B, C if use_bias else None)
    s = tvm.create_schedule(D.op)
    dtype = A.dtype

    # get the test data
    def get_ref_data():
        a_np = np.maximum(np.random.uniform(size=(batch, in_dim)).astype(dtype)-0.5, 0.)
        b_np = np.random.uniform(size=(in_dim, 1)).astype(dtype)-0.5
        c_np = np.random.uniform(size=(batch, )).astype(dtype)
        if use_bias:
            d_np = np.dot(a_np, b_np) + c_np.reshape((batch, 1))
        else:
            d_np = np.dot(a_np, b_np)
        return (a_np, b_np, c_np, d_np)
    a_np, b_np, c_np, d_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)
        a = tvmsp.array(a_np, ctx)
        _nr, _nc, _n = a.shape[0], a.shape[1], a.data.shape[0]
        assert a.shape[0] == a.indptr.shape[0]-1
        b = tvm.nd.array(b_np, ctx)
        c = tvm.nd.array(c_np, ctx)
        d = tvm.nd.array(np.zeros((_nr, 1), dtype=dtype), ctx)
        assert a.data.dtype == A.data.dtype
        assert a.indices.dtype == A.indices.dtype
        assert a.indptr.dtype == A.indptr.dtype
        f = tvm.build(s, [nr, A.data, A.indices, A.indptr, B, C, D], device, name="csrmv")
        f(_nr, a.data, a.indices, a.indptr, b, c, d)
67
        tvm.testing.assert_allclose(d.asnumpy(), d_np, rtol=1e-4, atol=1e-4)
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

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

def verify_dynamic_csrmm(batch, in_dim, out_dim, use_bias=True):
    nr, nc, n = tvm.var("nr"), tvm.var("nc"), tvm.var("n")
    dtype = 'float32'
    A = tvmsp.placeholder(shape=(nr, nc), nonzeros=n, dtype=dtype, name='A')
    B = tvm.placeholder((in_dim, out_dim), name='B')
    C = tvm.placeholder((nr,), name='C')
    D = topi.sparse.csrmm(A, B, C if use_bias else None)
    s = tvm.create_schedule(D.op)
    dtype = A.dtype

    # get the test data
    def get_ref_data():
        a_np = np.maximum(np.random.uniform(size=(batch, in_dim)).astype(dtype)-0.5, 0.)
        b_np = np.random.uniform(size=(in_dim, out_dim)).astype(dtype)-0.5
        c_np = np.random.uniform(size=(batch, )).astype(dtype)
        if use_bias:
            d_np = np.dot(a_np, b_np) + c_np.reshape((batch, 1))
        else:
            d_np = np.dot(a_np, b_np)
        return (a_np, b_np, c_np, d_np)
    a_np, b_np, c_np, d_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)
        a = tvmsp.array(a_np, ctx)
        _nr, _nc, _n = a.shape[0], a.shape[1], a.data.shape[0]
        assert a.shape[0] == a.indptr.shape[0]-1
        b = tvm.nd.array(b_np, ctx)
        c = tvm.nd.array(c_np, ctx)
        d = tvm.nd.array(np.zeros((_nr, out_dim), dtype=dtype), ctx)
        f = tvm.build(s, [nr, A.data, A.indices, A.indptr, B, C, D], device, name="csrmm")

        f(_nr, a.data, a.indices, a.indptr, b, c, d)
109
        tvm.testing.assert_allclose(d.asnumpy(), d_np, rtol=1e-2, atol=1e-2)
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

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

def verify_dense_si(batch, in_dim, out_dim, use_bias=True, dtype='float32'):
    nonzeros = tvm.var('nonzeros')
    A = tvmsp.placeholder(shape=(batch, in_dim), nonzeros=nonzeros, dtype=dtype, name='A')
    B = tvm.placeholder((out_dim, in_dim), dtype=dtype, name='B')
    C = tvm.placeholder((out_dim,), dtype=dtype, name='C')
    D = topi.sparse.dense(A, B, C if use_bias else None)
    s = tvm.create_schedule(D.op)

    # get the test data
    def get_ref_data():
        mag = 10.
        a_np = np.maximum(mag*(np.random.uniform(size=(batch, in_dim)).astype('float32')-0.5), 0.).astype(dtype)
        b_np = (mag*(np.random.uniform(size=(out_dim, in_dim)).astype('float32')-.5)).astype(dtype)
        c_np = (mag*(np.random.uniform(size=(out_dim,)).astype('float32')-.5)).astype(dtype)
        if use_bias:
            d_np = np.dot(a_np, b_np.T) + c_np
        else:
            d_np = np.dot(a_np, b_np.T)
        return (a_np, b_np, c_np, d_np)
    a_np, b_np, c_np, d_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)
        a = tvmsp.array(a_np, ctx)
        b = tvm.nd.array(b_np, ctx)
        c = tvm.nd.array(c_np, ctx)
        d = tvm.nd.array(np.zeros(get_const_tuple(D.shape), dtype=dtype), ctx)
        f = tvm.build(s, [A.data, A.indices, A.indptr, B, C, D], device, name="dense")
        f(a.data, a.indices, a.indptr, b, c, d)
147
        tvm.testing.assert_allclose(d.asnumpy(), d_np, rtol=1e-4, atol=1e-4)
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

    check_device('llvm')

def verify_dense_sw(batch, in_dim, out_dim, use_bias=True, dtype='float32'):
    nonzeros = tvm.var('nonzeros')
    A = tvm.placeholder((batch, in_dim), dtype=dtype, name='A')
    B = tvmsp.placeholder(shape=(out_dim, in_dim), nonzeros=nonzeros, dtype=dtype, name='B')
    C = tvm.placeholder((out_dim,), dtype=dtype, name='C')
    D = topi.sparse.dense(A, B, C if use_bias else None)
    s = tvm.create_schedule(D.op)

    # get the test data
    def get_ref_data():
        mag = 10.
        a_np = (mag*(np.random.uniform(size=(batch, in_dim)).astype('float32')-.5)).astype(dtype)
        b_np = np.maximum(mag*(np.random.uniform(size=(out_dim, in_dim)).astype('float32')-0.5), 0.).astype(dtype)
        c_np = (mag*(np.random.uniform(size=(out_dim,)).astype('float32')-.5)).astype(dtype)
        if use_bias:
            d_np = np.dot(a_np, b_np.T) + c_np
        else:
            d_np = np.dot(a_np, b_np.T)
        return (a_np, b_np, c_np, d_np)
    a_np, b_np, c_np, d_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)
        a = tvm.nd.array(a_np, ctx)
        b = tvmsp.array(b_np, ctx)
        c = tvm.nd.array(c_np, ctx)
        d = tvm.nd.array(np.zeros(get_const_tuple(D.shape), dtype=dtype), ctx)
        f = tvm.build(s, [A, B.data, B.indices, B.indptr, C, D], device, name="dense")
        f(a, b.data, b.indices, b.indptr, c, d)
184
        tvm.testing.assert_allclose(d.asnumpy(), d_np, rtol=1e-4, atol=1e-4)
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

    check_device('llvm')

def test_csrmv():
    verify_dynamic_csrmv(batch=5, in_dim=7, out_dim=1, use_bias=False)
    verify_dynamic_csrmv(batch=5, in_dim=7, out_dim=1, use_bias=True)

def test_csrmm():
    M, K, N = 5, 7, 2
    verify_dynamic_csrmm(batch=M, in_dim=K, out_dim=N, use_bias=False)
    verify_dynamic_csrmm(batch=M, in_dim=K, out_dim=N, use_bias=True)

def test_dense_si():
    M, K, N = 3, 5, 2
    verify_dense_si(batch=M, in_dim=K, out_dim=N, use_bias=False, dtype='float32')
    verify_dense_si(batch=M, in_dim=K, out_dim=N, use_bias=True, dtype='float32')
    verify_dense_si(batch=M, in_dim=K, out_dim=N, use_bias=False, dtype='int32')
    verify_dense_si(batch=M, in_dim=K, out_dim=N, use_bias=True, dtype='int32')
    verify_dense_si(batch=M, in_dim=K, out_dim=N, use_bias=False, dtype='int16')
    verify_dense_si(batch=M, in_dim=K, out_dim=N, use_bias=True, dtype='int16')

def test_dense_sw():
    M, K, N = 3, 5, 2
    verify_dense_sw(batch=M, in_dim=K, out_dim=N, use_bias=False, dtype='float32')
    verify_dense_sw(batch=M, in_dim=K, out_dim=N, use_bias=True, dtype='float32')
    verify_dense_sw(batch=M, in_dim=K, out_dim=N, use_bias=False, dtype='int32')
    verify_dense_sw(batch=M, in_dim=K, out_dim=N, use_bias=True, dtype='int32')
    verify_dense_sw(batch=M, in_dim=K, out_dim=N, use_bias=False, dtype='int16')
    verify_dense_sw(batch=M, in_dim=K, out_dim=N, use_bias=True, dtype='int16')

def test_dense():
    test_dense_si()
    test_dense_sw()

219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237

def test_sparse_dense_csr():
    M, N, K, density = 1, 17, 47, 0.2
    X_np = np.random.randn(M, K).astype("float32")
    W_sp_np = sp.random(N, K, density=density, format='csr', dtype="float32")
    W_np = W_sp_np.todense()
    Y_np = X_np.dot(W_np.T)

    W_data = tvm.placeholder(shape=W_sp_np.data.shape, dtype=str(W_sp_np.data.dtype))
    W_indices = tvm.placeholder(shape=W_sp_np.indices.shape, dtype=str(W_sp_np.indices.dtype))
    W_indptr = tvm.placeholder(shape=W_sp_np.indptr.shape, dtype=str(W_sp_np.indptr.dtype))
    X = tvm.placeholder(shape=X_np.shape, dtype=str(X_np.dtype))
    Y = topi.nn.sparse_dense(X, W_data, W_indices, W_indptr)
    s = tvm.create_schedule(Y.op)
    func = tvm.build(s, [X, W_data, W_indices, W_indptr, Y])
    Y_tvm = tvm.ndarray.array(np.zeros(Y_np.shape, dtype=Y_np.dtype))
    func(tvm.ndarray.array(X_np), tvm.ndarray.array(W_sp_np.data), tvm.ndarray.array(W_sp_np.indices), tvm.ndarray.array(W_sp_np.indptr), Y_tvm)
    tvm.testing.assert_allclose(Y_tvm.asnumpy(), Y_np, atol=1e-4, rtol=1e-4)

238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263
def test_sparse_transpose_csr():
    N, density = 1023, 0.3

    X_sp = sp.random(N, N, density=density, format='csr', dtype='float32')

    X_sp_T = X_sp.transpose()
    X_np_T = X_sp_T.todense()

    X_data = tvm.placeholder(shape=X_sp.data.shape, dtype=str(X_sp.data.dtype))
    X_indices = tvm.placeholder(shape=X_sp.indices.shape, dtype=str(X_sp.indices.dtype))
    X_indptr = tvm.placeholder(shape=X_sp.indptr.shape, dtype=str(X_sp.indptr.dtype))
    
    X_T_data, X_T_indices, X_T_indptr = topi.nn.sparse_transpose(X_data, X_indices, X_indptr)
    s = tvm.create_schedule([X_T_data.op, X_T_indices.op, X_T_indptr.op])
    func = tvm.build(s, [X_data, X_indices, X_indptr, X_T_data, X_T_indices, X_T_indptr])


    X_T_data_tvm = tvm.ndarray.array(np.zeros(X_sp_T.data.shape, dtype=X_sp_T.data.dtype))
    X_T_indices_tvm = tvm.ndarray.array(np.zeros(X_sp_T.indices.shape, dtype=X_sp_T.indices.dtype))
    X_T_indptr_tvm = tvm.ndarray.array(np.zeros(X_sp_T.indptr.shape, dtype=X_sp_T.indptr.dtype))

    func(tvm.ndarray.array(X_sp.data), tvm.ndarray.array(X_sp.indices), tvm.ndarray.array(X_sp.indptr),
        X_T_data_tvm,  X_T_indices_tvm, X_T_indptr_tvm)

    X_T_out = sp.csr_matrix((X_T_data_tvm.asnumpy(), X_T_indices_tvm.asnumpy(), X_T_indptr_tvm.asnumpy()), shape=(N,N)).todense()
    tvm.testing.assert_allclose(X_np_T, X_T_out, atol=1e-4, rtol=1e-4)
264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340

def random_bsr_matrix(M, N, BS_R, BS_C, density, dtype):
    import itertools
    Y = np.zeros((M, N), dtype=dtype)
    assert M % BS_R == 0
    assert N % BS_C == 0
    nnz = int(density * M * N)
    num_blocks = int(nnz / (BS_R * BS_C)) + 1
    candidate_blocks = np.asarray(list(itertools.product(range(0, M, BS_R), range(0, N, BS_C))))
    assert candidate_blocks.shape[0] == M // BS_R * N // BS_C
    chosen_blocks = candidate_blocks[np.random.choice(candidate_blocks.shape[0], size=num_blocks, replace=False)]
    for i in range(len(chosen_blocks)):
        r, c = chosen_blocks[i]
        Y[r:r + BS_R, c:c + BS_C] = np.random.randn(BS_R, BS_C)
    s = sp.bsr_matrix(Y, blocksize=(BS_R, BS_C))
    assert s.data.shape == (num_blocks, BS_R, BS_C)
    assert s.indices.shape == (num_blocks, )
    assert s.indptr.shape == (M // BS_R + 1, )
    return s

def test_sparse_dense_bsr():
    M, N, K, BS_R, BS_C, density = 1, 64, 128, 8, 16, 0.9
    X_np = np.random.randn(M, K).astype("float32")
    W_sp_np = random_bsr_matrix(N, K, BS_R, BS_C, density=density, dtype="float32")
    W_np = W_sp_np.todense()
    Y_np = X_np.dot(W_np.T)

    W_data = tvm.placeholder(shape=W_sp_np.data.shape, dtype=str(W_sp_np.data.dtype))
    W_indices = tvm.placeholder(shape=W_sp_np.indices.shape, dtype=str(W_sp_np.indices.dtype))
    W_indptr = tvm.placeholder(shape=W_sp_np.indptr.shape, dtype=str(W_sp_np.indptr.dtype))
    X = tvm.placeholder(shape=X_np.shape, dtype=str(X_np.dtype))
    Y = topi.nn.sparse_dense(X, W_data, W_indices, W_indptr)
    s = tvm.create_schedule(Y.op)
    func = tvm.build(s, [X, W_data, W_indices, W_indptr, Y])
    Y_tvm = tvm.ndarray.array(np.zeros(Y_np.shape, dtype=Y_np.dtype))
    func(tvm.ndarray.array(X_np),
         tvm.ndarray.array(W_sp_np.data),
         tvm.ndarray.array(W_sp_np.indices),
         tvm.ndarray.array(W_sp_np.indptr),
         Y_tvm)
    tvm.testing.assert_allclose(Y_tvm.asnumpy(), Y_np, atol=1e-4, rtol=1e-4)

def test_sparse_dense_bsr_randomized():
    for _ in range(20):
        BS_R = np.random.randint(1, 16)
        BS_C = np.random.randint(1, 16)
        M = np.random.randint(1, 32)
        N = int(np.random.randint(1, 16) * BS_R)
        K = int(np.random.randint(1, 16) * BS_C)
        density = np.clip(np.random.random(), 0.1, 0.9)
        X_np = np.random.randn(M, K).astype("float32")
        W_sp_np = random_bsr_matrix(N, K, BS_R, BS_C, density=density, dtype="float32")

        W_np = W_sp_np.todense()
        Y_np = np.array(X_np.dot(W_np.T))

        W_data = tvm.placeholder(shape=W_sp_np.data.shape, dtype=str(W_sp_np.data.dtype))
        W_indices = tvm.placeholder(shape=W_sp_np.indices.shape, dtype=str(W_sp_np.indices.dtype))
        W_indptr = tvm.placeholder(shape=W_sp_np.indptr.shape, dtype=str(W_sp_np.indptr.dtype))
        X = tvm.placeholder(shape=X_np.shape, dtype=str(X_np.dtype))
        Y = topi.nn.sparse_dense(X, W_data, W_indices, W_indptr)
        s = tvm.create_schedule(Y.op)
        func = tvm.build(s, [X, W_data, W_indices, W_indptr, Y])
        Y_tvm = tvm.ndarray.array(np.zeros(Y_np.shape, dtype=Y_np.dtype))
        func(tvm.ndarray.array(X_np),
             tvm.ndarray.array(W_sp_np.data),
             tvm.ndarray.array(W_sp_np.indices),
             tvm.ndarray.array(W_sp_np.indptr),
             Y_tvm)
        tvm.testing.assert_allclose(Y_tvm.asnumpy(), Y_np, atol=1e-5, rtol=1e-5)


def test_sparse_dense():
    test_sparse_dense_csr()
    test_sparse_dense_bsr()
    test_sparse_dense_bsr_randomized()

341 342 343 344
if __name__ == "__main__":
    test_csrmv()
    test_csrmm()
    test_dense()
345
    test_sparse_dense()
346
    test_sparse_transpose_csr()