cpu_fusion.cpp 176 KB
Newer Older
1
//*****************************************************************************
2
// Copyright 2017-2019 Intel Corporation
3 4 5 6 7 8 9 10 11 12 13 14 15
//
// Licensed 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.
//*****************************************************************************
16 17 18 19 20 21 22 23

#include <algorithm>
#include <cstdio>
#include <iostream>
#include <list>
#include <memory>

#include "gtest/gtest.h"
24
#include "misc.hpp"
25
#include "ngraph/autodiff/adjoints.hpp"
Louis Feng's avatar
Louis Feng committed
26
#include "ngraph/file_util.hpp"
27 28 29
#include "ngraph/graph_util.hpp"
#include "ngraph/log.hpp"
#include "ngraph/ngraph.hpp"
30
#include "ngraph/op/batch_norm.hpp"
31
#include "ngraph/op/concat.hpp"
32
#include "ngraph/op/dequantize.hpp"
33
#include "ngraph/op/experimental/quantized_concat.hpp"
34 35
#include "ngraph/op/experimental/quantized_conv.hpp"
#include "ngraph/op/experimental/quantized_conv_bias.hpp"
36
#include "ngraph/op/get_output_element.hpp"
37
#include "ngraph/op/max_pool.hpp"
38
#include "ngraph/op/negative.hpp"
39
#include "ngraph/op/parameter.hpp"
40
#include "ngraph/op/quantize.hpp"
41
#include "ngraph/op/relu.hpp"
Pruthvi's avatar
Pruthvi committed
42
#include "ngraph/op/reverse_sequence.hpp"
43
#include "ngraph/op/sigmoid.hpp"
44
#include "ngraph/op/sum.hpp"
45
#include "ngraph/op/tanh.hpp"
46
#include "ngraph/pass/algebraic_simplification.hpp"
47
#include "ngraph/pass/core_fusion.hpp"
48 49
#include "ngraph/pass/graph_rewrite.hpp"
#include "ngraph/pass/manager.hpp"
Louis Feng's avatar
Louis Feng committed
50 51
#include "ngraph/pass/reshape_elimination.hpp"
#include "ngraph/pass/visualize_tree.hpp"
52 53
#include "ngraph/pattern/matcher.hpp"
#include "ngraph/pattern/op/label.hpp"
54
#include "ngraph/pattern/op/skip.hpp"
Nick Korovaiko's avatar
Nick Korovaiko committed
55
#include "ngraph/runtime/cpu/cpu_layout_descriptor.hpp"
56
#include "ngraph/runtime/cpu/cpu_tensor_view.hpp"
57
#include "ngraph/runtime/cpu/op/batch_dot.hpp"
58
#include "ngraph/runtime/cpu/op/batch_norm_relu.hpp"
59
#include "ngraph/runtime/cpu/op/bounded_relu.hpp"
gaurides's avatar
gaurides committed
60
#include "ngraph/runtime/cpu/op/conv_add.hpp"
61
#include "ngraph/runtime/cpu/op/conv_bias.hpp"
62
#include "ngraph/runtime/cpu/op/conv_relu.hpp"
Nick Korovaiko's avatar
Nick Korovaiko committed
63
#include "ngraph/runtime/cpu/op/convert_layout.hpp"
64
#include "ngraph/runtime/cpu/op/group_conv.hpp"
65
#include "ngraph/runtime/cpu/op/group_conv_bias.hpp"
66
#include "ngraph/runtime/cpu/op/leaky_relu.hpp"
67
#include "ngraph/runtime/cpu/op/loop_kernel.hpp"
68
#include "ngraph/runtime/cpu/op/lstm.hpp"
69
#include "ngraph/runtime/cpu/op/matmul_bias.hpp"
70
#include "ngraph/runtime/cpu/op/rnn.hpp"
Pruthvi's avatar
Pruthvi committed
71
#include "ngraph/runtime/cpu/op/rnn_utils.hpp"
72
#include "ngraph/runtime/cpu/op/sigmoid_mul.hpp"
73
#include "ngraph/runtime/cpu/op/update_slice.hpp"
74
#include "ngraph/runtime/cpu/pass/cpu_fusion.hpp"
75
#include "ngraph/runtime/cpu/pass/cpu_loop_kernel_fusion.hpp"
76
#include "ngraph/runtime/cpu/pass/cpu_mat_fusion.hpp"
Nick Korovaiko's avatar
Nick Korovaiko committed
77
#include "ngraph/runtime/cpu/pass/cpu_post_layout_optimizations.hpp"
78
#include "ngraph/runtime/cpu/pass/cpu_rnn_fusion.hpp"
79
#include "ngraph/runtime/cpu/pass/cpu_workspace_insertion.hpp"
80 81
#include "ngraph/serializer.hpp"
#include "ngraph/util.hpp"
82
#include "nlohmann/json.hpp"
83
#include "util/all_close.hpp"
84
#include "util/all_close_f.hpp"
Pruthvi's avatar
Pruthvi committed
85 86
#include "util/autodiff/backprop_function.hpp"
#include "util/autodiff/numeric_compare.hpp"
87
#include "util/matcher.hpp"
88
#include "util/random.hpp"
89
#include "util/random.hpp"
90
#include "util/test_tools.hpp"
91

92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114
using namespace ngraph;
using namespace std;

TEST(cpu_fusion, gemm_pattern)
{
    Shape shape_w{2, 4};
    Shape shape_x{4, 1};
    Shape shape_b{1};
    auto A = make_shared<op::Parameter>(element::f32, shape_w);
    auto B = make_shared<op::Parameter>(element::f32, shape_x);
    auto C = make_shared<op::Parameter>(element::f32, shape_b);

    auto dot = make_shared<op::Dot>(A, B);
    auto broadcast = make_shared<op::Broadcast>(C, dot->get_shape(), AxisSet{0});
    auto add = dot + broadcast;

    auto W = std::make_shared<pattern::op::Label>(A);
    auto x = std::make_shared<pattern::op::Label>(B);

    auto reshape_pred = [](std::shared_ptr<Node> n) {
        return static_cast<bool>(std::dynamic_pointer_cast<op::Reshape>(n));
    };

115 116
    auto skip_w = std::make_shared<pattern::op::Skip>(W, reshape_pred);
    auto skip_x = std::make_shared<pattern::op::Skip>(x, reshape_pred);
117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137

    auto pdot = make_shared<op::Dot>(skip_w, skip_x);
    auto b = std::make_shared<pattern::op::Label>(C);
    auto pbroadcast = make_shared<op::Broadcast>(b, dot->get_shape(), AxisSet{0});
    auto padd = pdot + pbroadcast;

    TestMatcher n(nullptr);
    ASSERT_TRUE(n.match(padd, add));
    ASSERT_EQ(n.get_pattern_map()[W], A);
    ASSERT_EQ(n.get_pattern_map()[x], B);
    ASSERT_EQ(n.get_pattern_map()[b], C);

    auto reshape_w = make_shared<op::Reshape>(A, AxisVector{1, 0}, W->get_shape());
    auto reshape_x = make_shared<op::Reshape>(B, AxisVector{1, 0}, x->get_shape());
    auto re_dot = make_shared<op::Dot>(reshape_w, reshape_x);
    auto re_add = re_dot + broadcast;
    ASSERT_TRUE(n.match(padd, re_add));
    ASSERT_EQ(n.get_pattern_map()[W], A);
    ASSERT_EQ(n.get_pattern_map()[x], B);
    ASSERT_EQ(n.get_pattern_map()[b], C);

138 139 140 141 142 143 144 145 146 147 148 149
    auto cg = make_shared<op::MatmulBias>(
        W, x, C, W->get_shape(), x->get_shape(), false, false, AxisSet{0});
}

TEST(cpu_fusion, gemm_cpu_broadcast_row)
{
    Shape shapeA{3, 2};
    Shape shapeB{2, 3};
    Shape shapeC{2, 2};
    auto A = make_shared<op::Parameter>(element::f32, shapeA);
    auto B = make_shared<op::Parameter>(element::f32, shapeB);

150
    auto bias = op::Constant::create<float>(element::f32, Shape{2}, std::vector<float>{2.0f, 3.0f});
151 152

    auto cg = make_shared<op::MatmulBias>(
153
        A, B, bias, A->get_shape(), B->get_shape(), true, true, AxisSet{0});
154

155
    auto f = make_shared<Function>(cg, ParameterVector{A, B});
156

157
    auto backend = runtime::Backend::create("CPU");
158

159 160 161
    shared_ptr<runtime::Tensor> a = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> b = backend->create_tensor(element::f32, shapeB);
    shared_ptr<runtime::Tensor> result = backend->create_tensor(element::f32, shapeC);
162 163 164 165 166 167

    vector<float> dataA{1.0f, 4.0f, 1.0f, 4.0f, 1.0f, 4.0f};
    vector<float> dataB{3.0f, 3.0f, 3.0f, 9.0f, 9.0f, 9.0f};
    copy_data(a, dataA);
    copy_data(b, dataB);

168
    auto handle = backend->compile(f);
169
    handle->call_with_validate({result}, {a, b});
170
    vector<float> expected{11, 30, 38, 111};
171
    EXPECT_TRUE(test::all_close_f(read_vector<float>(result), expected, MIN_FLOAT_TOLERANCE_BITS));
172 173
}

174 175 176 177 178 179 180 181
TEST(cpu_fusion, gemm_cpu_broadcast_column)
{
    Shape shapeA{3, 2};
    Shape shapeB{2, 3};
    Shape shapeC{2, 2};
    auto A = make_shared<op::Parameter>(element::f32, shapeA);
    auto B = make_shared<op::Parameter>(element::f32, shapeB);

182
    auto bias = op::Constant::create<float>(element::f32, Shape{2}, std::vector<float>{2.0f, 3.0f});
183 184

    auto cg = make_shared<op::MatmulBias>(
185
        A, B, bias, A->get_shape(), B->get_shape(), true, true, AxisSet{1});
186

187
    auto f = make_shared<Function>(cg, ParameterVector{A, B});
188

189
    auto backend = runtime::Backend::create("CPU");
190

191 192 193
    shared_ptr<runtime::Tensor> a = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> b = backend->create_tensor(element::f32, shapeB);
    shared_ptr<runtime::Tensor> result = backend->create_tensor(element::f32, shapeC);
194 195 196 197 198 199

    vector<float> dataA{1.0f, 4.0f, 1.0f, 4.0f, 1.0f, 4.0f};
    vector<float> dataB{3.0f, 3.0f, 3.0f, 9.0f, 9.0f, 9.0f};
    copy_data(a, dataA);
    copy_data(b, dataB);

200
    auto handle = backend->compile(f);
201
    handle->call_with_validate({result}, {a, b});
202
    vector<float> expected{11, 29, 39, 111};
203
    EXPECT_TRUE(test::all_close_f(read_vector<float>(result), expected, MIN_FLOAT_TOLERANCE_BITS));
204 205 206
}

TEST(cpu_fusion, gemm_cpu_broadcast_matrix)
207 208 209 210 211 212 213 214 215 216 217 218 219
{
    Shape shapeA{3, 2};
    Shape shapeB{2, 3};
    Shape shapeC{2, 2};
    auto A = make_shared<op::Parameter>(element::f32, shapeA);
    auto B = make_shared<op::Parameter>(element::f32, shapeB);

    auto reshape_w = make_shared<op::Reshape>(A, AxisVector{1, 0}, Shape{2, 3});
    auto reshape_x = make_shared<op::Reshape>(B, AxisVector{1, 0}, Shape{3, 2});

    auto one = op::Constant::create<float>(element::f32, Shape{}, std::vector<float>{1.0f});

    auto broadcast = make_shared<op::Broadcast>(one, shapeC, AxisSet{0, 1});
220 221
    auto cg = make_shared<op::MatmulBias>(
        A, B, one, A->get_shape(), B->get_shape(), true, true, AxisSet{0, 1});
222

223
    auto f = make_shared<Function>(cg, ParameterVector{A, B});
224

225
    auto backend = runtime::Backend::create("CPU");
226

227 228 229
    shared_ptr<runtime::Tensor> a = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> b = backend->create_tensor(element::f32, shapeB);
    shared_ptr<runtime::Tensor> result = backend->create_tensor(element::f32, shapeC);
230 231 232 233 234 235

    vector<float> dataA{1.0f, 4.0f, 1.0f, 4.0f, 1.0f, 4.0f};
    vector<float> dataB{3.0f, 3.0f, 3.0f, 9.0f, 9.0f, 9.0f};
    copy_data(a, dataA);
    copy_data(b, dataB);

236
    auto handle = backend->compile(f);
237
    handle->call_with_validate({result}, {a, b});
238
    vector<float> expected{10, 28, 37, 109};
239
    EXPECT_TRUE(test::all_close_f(read_vector<float>(result), expected, MIN_FLOAT_TOLERANCE_BITS));
240 241
}

242 243 244 245 246 247 248 249 250 251 252 253 254 255
TEST(cpu_fusion, gemm_cpu_no_bias)
{
    auto shapeA = Shape{3, 2};
    auto shapeB = Shape{2, 3};
    auto shapeC = Shape{2, 2};
    auto A = make_shared<op::Parameter>(element::f32, shapeA);
    auto B = make_shared<op::Parameter>(element::f32, shapeB);

    auto reshape_w = make_shared<op::Reshape>(A, AxisVector{1, 0}, Shape{2, 3});
    auto reshape_x = make_shared<op::Reshape>(B, AxisVector{1, 0}, Shape{3, 2});

    auto cg =
        make_shared<op::MatmulBias>(A, B, nullptr, A->get_shape(), B->get_shape(), true, true);

256
    auto f = make_shared<Function>(cg, ParameterVector{A, B});
257

258
    auto backend = runtime::Backend::create("CPU");
259

260 261 262
    shared_ptr<runtime::Tensor> a = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> b = backend->create_tensor(element::f32, shapeB);
    shared_ptr<runtime::Tensor> result = backend->create_tensor(element::f32, shapeC);
263 264 265 266 267 268

    vector<float> dataA{1.0f, 4.0f, 1.0f, 4.0f, 1.0f, 4.0f};
    vector<float> dataB{3.0f, 3.0f, 3.0f, 9.0f, 9.0f, 9.0f};
    copy_data(a, dataA);
    copy_data(b, dataB);

269
    auto handle = backend->compile(f);
270
    handle->call_with_validate({result}, {a, b});
271
    vector<float> expected{9, 27, 36, 108};
272
    EXPECT_TRUE(test::all_close_f(read_vector<float>(result), expected, MIN_FLOAT_TOLERANCE_BITS));
273 274
}

275 276 277 278 279 280 281 282 283 284 285 286 287 288 289
TEST(cpu_fusion, cpu_fusion_pass_basic)
{
    Shape shape{};
    Shape shape_w{2, 4};
    Shape shape_x{4, 1};
    Shape shape_b{1};
    auto A = make_shared<op::Parameter>(element::f32, shape_w);
    auto B = make_shared<op::Parameter>(element::f32, shape_x);
    auto C = make_shared<op::Parameter>(element::f32, shape_b);

    auto dot = make_shared<op::Dot>(A, B);
    auto broadcast = make_shared<op::Broadcast>(C, dot->get_shape(), AxisSet{0});
    auto add = dot + broadcast;
    auto graph = make_shared<op::Abs>(add);
    pass::Manager pass_manager;
290
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
291
    auto func = make_shared<Function>(graph, ParameterVector{A, B, C});
292
    pass_manager.run_passes(func);
293
    ASSERT_NE(std::dynamic_pointer_cast<op::MatmulBias>(graph->get_argument(0)), nullptr);
294 295
}

296 297 298 299 300 301 302 303 304 305 306 307 308 309 310
TEST(cpu_fusion, commutative_matmul_bias)
{
    Shape shape{};
    Shape shape_w{2, 4};
    Shape shape_x{4, 1};
    Shape shape_b{1};
    auto A = make_shared<op::Parameter>(element::f32, shape_w);
    auto B = make_shared<op::Parameter>(element::f32, shape_x);
    auto C = make_shared<op::Parameter>(element::f32, shape_b);

    auto dot = make_shared<op::Dot>(A, B);
    auto broadcast = make_shared<op::Broadcast>(C, dot->get_shape(), AxisSet{0});
    auto add = broadcast + dot;
    auto graph = make_shared<op::Abs>(add);
    pass::Manager pass_manager;
311
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
312
    auto func = make_shared<Function>(graph, ParameterVector{A, B, C});
313
    pass_manager.run_passes(func);
314
    ASSERT_NE(std::dynamic_pointer_cast<op::MatmulBias>(graph->get_argument(0)), nullptr);
315 316
}

317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332
TEST(cpu_fusion, cpu_fusion_pass_matmul_bias)
{
    Shape shape_w{2, 4};
    Shape shape_x{4, 1};
    Shape shape_b{1};
    auto W = make_shared<op::Parameter>(element::f32, shape_w);
    auto x = make_shared<op::Parameter>(element::f32, shape_x);
    auto b = make_shared<op::Parameter>(element::f32, shape_b);

    auto mmb = std::make_shared<op::MatmulBias>(
        W, x, nullptr, W->get_shape(), x->get_shape(), false, false);
    auto broadcast = std::make_shared<op::Broadcast>(b, mmb->get_shape(), AxisSet{0});
    auto add = mmb + broadcast;

    auto graph = make_shared<op::Abs>(add);
    pass::Manager pass_manager;
333
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
334
    auto func = make_shared<Function>(graph, ParameterVector{W, x, b});
335
    pass_manager.run_passes(func);
336
    auto gmm = graph->get_argument(0);
337
    ASSERT_TRUE(std::dynamic_pointer_cast<op::MatmulBias>(gmm));
338
    ASSERT_EQ(gmm->get_argument(2), b);
339 340 341 342 343 344 345 346 347 348 349 350 351 352 353
}

TEST(cpu_fusion, cpu_fusion_pass_matmul_no_bias)
{
    Shape shape_w{4, 2};
    Shape shape_x{1, 4};
    auto W = make_shared<op::Parameter>(element::f32, shape_w);
    auto x = make_shared<op::Parameter>(element::f32, shape_x);

    auto reshape_w = std::make_shared<op::Reshape>(W, AxisVector{1, 0}, Shape{2, 4});
    auto reshape_x = std::make_shared<op::Reshape>(x, AxisVector{1, 0}, Shape{4, 1});
    auto re_dot = make_shared<op::Dot>(reshape_w, reshape_x);
    auto graph = make_shared<op::Abs>(re_dot);

    pass::Manager pass_manager;
354
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
355
    auto func = make_shared<Function>(graph, ParameterVector{W, x});
356 357 358 359 360
    pass_manager.run_passes(func);
    size_t mmb = count_ops_of_type<op::MatmulBias>(func);
    ASSERT_EQ(mmb, 1);
}

361 362 363 364 365 366 367
TEST(cpu_fusion, gemm_mlp)
{
    const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/mnist_mlp_forward.json");
    const string json_string = file_util::read_file_to_string(json_path);
    stringstream ss(json_string);
    shared_ptr<Function> func = ngraph::deserialize(ss);
    pass::Manager pass_manager;
368
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
369
    pass_manager.run_passes(func);
370 371
    auto mmbs = count_ops_of_type<op::MatmulBias>(func);
    ASSERT_EQ(mmbs, 3);
372 373 374 375 376 377 378
}

TEST(cpu_fusion, fuse_fprop_bn)
{
    pass::Manager pass_manager;
    pass_manager.register_pass<pass::VisualizeTree>("bn_fprop_before_fusion.png");
    pass_manager.register_pass<ngraph::pass::ReshapeElimination>();
379
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
380 381 382 383 384 385
    pass_manager.register_pass<pass::VisualizeTree>("bn_fprop_after_fusion.png");
    const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/bn_fprop_b2c3h2w2.json");
    const string json_string = file_util::read_file_to_string(json_path);
    stringstream ss(json_string);
    shared_ptr<Function> func = ngraph::deserialize(ss);
    pass_manager.run_passes(func);
386
    size_t ccg = count_ops_of_type<op::BatchNormTraining>(func);
387 388
    ASSERT_EQ(ccg, 1);
}
nikolay.korovaiko's avatar
nikolay.korovaiko committed
389

390 391 392 393 394 395 396 397
TEST(cpu_fusion, zero_padded_reshaped_conv)
{
    auto X = make_shared<op::Parameter>(element::f32, Shape{1, 2, 2, 1});
    auto F = make_shared<op::Parameter>(element::f32, Shape{1, 1, 1, 1});

    auto pad_value = op::Constant::create<float>(element::f32, Shape{}, std::vector<float>{0.0f});

    auto pad =
398
        make_shared<op::Pad>(X, pad_value, CoordinateDiff{0, 1, 0, 0}, CoordinateDiff{0, 0, 1, 0});
399 400 401 402 403 404 405 406 407 408 409

    auto reshape = make_shared<op::Reshape>(pad, AxisVector{0, 3, 1, 2}, Shape{1, 1, 3, 3});

    auto conv = make_shared<op::Convolution>(reshape,
                                             F,
                                             Strides{1, 1},
                                             Strides{1, 1},
                                             CoordinateDiff{0, 0},
                                             CoordinateDiff{0, 0},
                                             Strides{1, 1});

410
    auto func = make_shared<Function>(conv, ParameterVector{X, F});
411 412 413

    ASSERT_EQ(count_ops_of_type<op::Pad>(func), 1);

414 415
    auto backend = runtime::Backend::create("CPU");
    backend->compile(func);
416 417 418 419 420 421 422 423 424 425 426 427

    ASSERT_EQ(count_ops_of_type<op::Pad>(func), 0);
}

TEST(cpu_fusion, zero_padded_conv)
{
    auto X = make_shared<op::Parameter>(element::f32, Shape{1, 1, 2, 2});
    auto F = make_shared<op::Parameter>(element::f32, Shape{1, 1, 1, 1});

    auto pad_value = op::Constant::create<float>(element::f32, Shape{}, std::vector<float>{0.0f});

    auto pad =
428
        make_shared<op::Pad>(X, pad_value, CoordinateDiff{0, 0, 0, 1}, CoordinateDiff{0, 0, 1, 0});
429 430 431 432 433 434 435 436 437

    auto conv = make_shared<op::Convolution>(pad,
                                             F,
                                             Strides{1, 1},
                                             Strides{1, 1},
                                             CoordinateDiff{0, 0},
                                             CoordinateDiff{0, 0},
                                             Strides{1, 1});

438
    auto func = make_shared<Function>(conv, ParameterVector{X, F});
439 440 441

    ASSERT_EQ(count_ops_of_type<op::Pad>(func), 1);

442 443
    auto backend = runtime::Backend::create("CPU");
    backend->compile(func);
444 445 446 447 448 449 450 451 452 453 454 455

    ASSERT_EQ(count_ops_of_type<op::Pad>(func), 0);
}

TEST(cpu_fusion, non_zero_padded_conv)
{
    auto X = make_shared<op::Parameter>(element::f32, Shape{1, 1, 2, 2});
    auto F = make_shared<op::Parameter>(element::f32, Shape{1, 1, 1, 1});

    auto pad_value = op::Constant::create<float>(element::f32, Shape{}, std::vector<float>{1.0f});

    auto pad =
456
        make_shared<op::Pad>(X, pad_value, CoordinateDiff{0, 0, 0, 1}, CoordinateDiff{0, 0, 1, 0});
457 458 459 460 461 462 463 464 465

    auto conv = make_shared<op::Convolution>(pad,
                                             F,
                                             Strides{1, 1},
                                             Strides{1, 1},
                                             CoordinateDiff{0, 0},
                                             CoordinateDiff{0, 0},
                                             Strides{1, 1});

466
    auto func = make_shared<Function>(conv, ParameterVector{X, F});
467 468 469

    ASSERT_EQ(count_ops_of_type<op::Pad>(func), 1);

470 471
    auto backend = runtime::Backend::create("CPU");
    backend->compile(func);
472 473 474

    ASSERT_EQ(count_ops_of_type<op::Pad>(func), 1);
}
475 476 477 478 479 480 481 482 483

TEST(cpu_fusion, zero_padded_conv_backprop_filters)
{
    auto X = make_shared<op::Parameter>(element::f32, Shape{1, 1, 2, 2});
    auto F = make_shared<op::Parameter>(element::f32, Shape{1, 1, 2, 2});

    auto pad_value = op::Constant::create<float>(element::f32, Shape{}, std::vector<float>{0.0f});

    auto pad =
484
        make_shared<op::Pad>(X, pad_value, CoordinateDiff{0, 0, 0, 1}, CoordinateDiff{0, 0, 1, 0});
485 486 487 488 489 490 491 492 493 494

    auto conv = make_shared<op::ConvolutionBackpropFilters>(pad,
                                                            Shape{1, 1, 2, 2},
                                                            F,
                                                            Strides{1, 1},
                                                            Strides{1, 1},
                                                            CoordinateDiff{0, 0},
                                                            CoordinateDiff{0, 0},
                                                            Strides{1, 1});

495
    auto func = make_shared<Function>(conv, ParameterVector{X, F});
496 497 498

    ASSERT_EQ(count_ops_of_type<op::Pad>(func), 1);

499 500
    auto backend = runtime::Backend::create("CPU");
    backend->compile(func);
501 502 503 504

    ASSERT_EQ(count_ops_of_type<op::Pad>(func), 0);
}

nikolay.korovaiko's avatar
nikolay.korovaiko committed
505 506 507 508
TEST(cpu_fusion, fuse_conv_bias)
{
    pass::Manager pass_manager;
    pass_manager.register_pass<ngraph::pass::ReshapeElimination>();
509
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::DIFFERENTIABLE_FUSIONS);
nikolay.korovaiko's avatar
nikolay.korovaiko committed
510 511 512 513 514 515 516 517
    const string json_path = file_util::path_join(SERIALIZED_ZOO, "conv_bias.json");
    const string json_string = file_util::read_file_to_string(json_path);
    stringstream ss(json_string);
    shared_ptr<Function> func = ngraph::deserialize(ss);
    pass_manager.run_passes(func);
    size_t cb = count_ops_of_type<op::ConvolutionBias>(func);
    ASSERT_GT(cb, 0);
}
518

Louis Feng's avatar
Louis Feng committed
519
struct ConvolutionBiasTestData
Louis Feng's avatar
Louis Feng committed
520
{
Louis Feng's avatar
Louis Feng committed
521 522 523 524 525 526
    size_t n{0};
    size_t c{0};
    size_t filter{0};
    size_t kernel_size{0};
    size_t w{0};
    size_t h{0};
527 528 529 530 531 532 533 534
    shared_ptr<runtime::Tensor> data_val;
    shared_ptr<runtime::Tensor> weights_val;
    shared_ptr<runtime::Tensor> bias_val;
    shared_ptr<runtime::Tensor> result_val;
    shared_ptr<runtime::Tensor> delta_val;
    shared_ptr<runtime::Tensor> d_data_val;
    shared_ptr<runtime::Tensor> d_weights_val;
    shared_ptr<runtime::Tensor> d_bias_val;
Louis Feng's avatar
Louis Feng committed
535 536 537 538 539 540 541 542 543 544 545 546 547 548
    vector<float> expected_result_val;
    vector<float> expected_d_data_val;
    vector<float> expected_d_weights_val;
    vector<float> expected_d_bias_val;

    Shape data_shape;
    Shape weights_shape;
    Shape bias_shape;
    Shape result_shape;
    shared_ptr<op::Parameter> data;
    shared_ptr<op::Parameter> weights;
    shared_ptr<op::Parameter> bias;
    shared_ptr<op::Parameter> delta;

549
    void n1c1h3w3(runtime::Backend* backend)
Louis Feng's avatar
Louis Feng committed
550
    {
Louis Feng's avatar
Louis Feng committed
551 552 553 554 555 556 557 558 559 560 561 562 563
        n = 1;
        c = 1;
        filter = 1;
        kernel_size = 3;
        w = 3;
        h = w;

        data_shape = Shape{n, c, h, w};
        data = make_shared<op::Parameter>(element::f32, data_shape);
        weights_shape = Shape{filter, c, kernel_size, kernel_size};
        weights = make_shared<op::Parameter>(element::f32, weights_shape);
        bias_shape = Shape{filter};
        bias = make_shared<op::Parameter>(element::f32, bias_shape);
Louis Feng's avatar
Louis Feng committed
564
        result_shape = Shape{n, filter, 1, 1};
Louis Feng's avatar
Louis Feng committed
565

566
        data_val = backend->create_tensor(element::f32, data_shape);
Louis Feng's avatar
Louis Feng committed
567 568 569 570 571 572 573 574 575 576
        copy_data(data_val,
                  vector<float>{-0.67765152f,
                                0.10073948f,
                                0.57595438f,
                                -0.3469252f,
                                -0.22134334f,
                                -1.80471897f,
                                -0.80642909f,
                                1.22033095f,
                                2.23235631f});
577
        weights_val = backend->create_tensor(element::f32, weights_shape);
Louis Feng's avatar
Louis Feng committed
578 579 580 581 582 583 584 585 586 587
        copy_data(weights_val,
                  vector<float>{0.20070229f,
                                -0.54968649f,
                                -0.19819015f,
                                -0.38577855f,
                                1.37109005f,
                                -0.23789984f,
                                0.14867957f,
                                -0.49851316f,
                                -0.84815776f});
588
        bias_val = backend->create_tensor(element::f32, bias_shape);
Louis Feng's avatar
Louis Feng committed
589 590
        copy_data(bias_val, vector<float>{0.07811152f});

591
        result_val = backend->create_tensor(element::f32, result_shape);
Louis Feng's avatar
Louis Feng committed
592 593 594
        copy_data(result_val, vector<float>{0});

        delta = make_shared<op::Parameter>(element::f32, result_shape);
595
        delta_val = backend->create_tensor(element::f32, result_shape);
Louis Feng's avatar
Louis Feng committed
596 597
        copy_data(delta_val, vector<float>{-2.58936238f});

598
        d_data_val = backend->create_tensor(element::f32, data_shape);
Louis Feng's avatar
Louis Feng committed
599
        copy_data(d_data_val, vector<float>{0, 0, 0, 0, 0, 0, 0, 0, 0});
Louis Feng's avatar
Louis Feng committed
600

601
        d_weights_val = backend->create_tensor(element::f32, weights_shape);
Louis Feng's avatar
Louis Feng committed
602
        copy_data(d_weights_val, vector<float>{0, 0, 0, 0, 0, 0, 0, 0, 0});
Louis Feng's avatar
Louis Feng committed
603

604
        d_bias_val = backend->create_tensor(element::f32, bias_shape);
Louis Feng's avatar
Louis Feng committed
605 606
        copy_data(d_bias_val, vector<float>{0});

Louis Feng's avatar
Louis Feng committed
607
        expected_result_val = vector<float>{-2.58936238f};
Louis Feng's avatar
Louis Feng committed
608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625
        expected_d_data_val = vector<float>{-0.51969099f,
                                            1.42333758f,
                                            0.5131861f,
                                            0.99892044f,
                                            -3.5502491f,
                                            0.61600888f,
                                            -0.3849853f,
                                            1.29083121f,
                                            2.19618773f};
        expected_d_weights_val = vector<float>{1.7546854f,
                                               -0.26085103f,
                                               -1.49135458f,
                                               0.89831507f,
                                               0.57313812f,
                                               4.67307138f,
                                               2.08813715f,
                                               -3.15987897f,
                                               -5.7803793f};
Louis Feng's avatar
Louis Feng committed
626
        expected_d_bias_val = vector<float>{-2.58936238f};
Louis Feng's avatar
Louis Feng committed
627
    }
Louis Feng's avatar
Louis Feng committed
628
};
Louis Feng's avatar
Louis Feng committed
629

Louis Feng's avatar
Louis Feng committed
630
TEST(cpu_fusion, conv_bias_fprop_n1c1h3w3)
Louis Feng's avatar
Louis Feng committed
631
{
632
    auto backend = runtime::Backend::create("CPU");
Louis Feng's avatar
Louis Feng committed
633 634

    ConvolutionBiasTestData conv_test;
635
    conv_test.n1c1h3w3(backend.get());
Louis Feng's avatar
Louis Feng committed
636 637 638 639

    auto convolution = make_shared<op::Convolution>(conv_test.data, conv_test.weights);
    auto convolution_bias = make_shared<op::ConvolutionBias>(convolution, conv_test.bias);

Louis Feng's avatar
Louis Feng committed
640
    auto f = make_shared<Function>(
641
        convolution_bias, ParameterVector{conv_test.data, conv_test.weights, conv_test.bias});
Louis Feng's avatar
Louis Feng committed
642

643
    auto handle = backend->compile(f);
644 645
    handle->call_with_validate({conv_test.result_val},
                               {conv_test.data_val, conv_test.weights_val, conv_test.bias_val});
Louis Feng's avatar
Louis Feng committed
646
    auto result_vec = read_vector<float>(conv_test.result_val);
Louis Feng's avatar
Louis Feng committed
647

Louis Feng's avatar
Louis Feng committed
648 649
    EXPECT_TRUE(
        test::all_close(conv_test.expected_result_val, read_vector<float>(conv_test.result_val)));
Louis Feng's avatar
Louis Feng committed
650 651
}

Louis Feng's avatar
Louis Feng committed
652
TEST(cpu_fusion, conv_bias_bprop_n1c1h3w3)
Louis Feng's avatar
Louis Feng committed
653
{
654
    auto backend = runtime::Backend::create("CPU");
Louis Feng's avatar
Louis Feng committed
655

Louis Feng's avatar
Louis Feng committed
656
    ConvolutionBiasTestData conv_test;
657
    conv_test.n1c1h3w3(backend.get());
Louis Feng's avatar
Louis Feng committed
658 659 660 661

    auto convolution = make_shared<op::Convolution>(conv_test.data, conv_test.weights);
    auto convolution_bias = make_shared<op::ConvolutionBias>(convolution, conv_test.bias);

Louis Feng's avatar
Louis Feng committed
662
    auto f = make_shared<Function>(
663
        convolution_bias, ParameterVector{conv_test.data, conv_test.weights, conv_test.bias});
Louis Feng's avatar
Louis Feng committed
664

665 666 667 668 669
    ngraph::autodiff::Adjoints adjoints(NodeVector{convolution_bias}, NodeVector{conv_test.delta});

    auto d_data = adjoints.backprop_node(conv_test.data);
    auto d_weights = adjoints.backprop_node(conv_test.weights);
    auto d_bias = adjoints.backprop_node(conv_test.bias);
Louis Feng's avatar
Louis Feng committed
670

Louis Feng's avatar
Louis Feng committed
671 672
    auto df = make_shared<Function>(
        NodeVector{d_data, d_weights, d_bias},
673
        ParameterVector{conv_test.data, conv_test.weights, conv_test.bias, conv_test.delta});
674 675 676
    auto handle = backend->compile(df);
    handle->call_with_validate(

677 678
        {conv_test.d_data_val, conv_test.d_weights_val, conv_test.d_bias_val},
        {conv_test.data_val, conv_test.weights_val, conv_test.bias_val, conv_test.delta_val});
Louis Feng's avatar
Louis Feng committed
679

Louis Feng's avatar
Louis Feng committed
680 681 682 683 684 685
    EXPECT_TRUE(
        test::all_close(conv_test.expected_d_data_val, read_vector<float>(conv_test.d_data_val)));
    EXPECT_TRUE(test::all_close(conv_test.expected_d_weights_val,
                                read_vector<float>(conv_test.d_weights_val)));
    EXPECT_TRUE(
        test::all_close(conv_test.expected_d_bias_val, read_vector<float>(conv_test.d_bias_val)));
Louis Feng's avatar
Louis Feng committed
686
}
Pruthvi's avatar
Pruthvi committed
687

688 689 690 691 692 693
TEST(cpu_fusion, conv_bias_bprop)
{
    Shape shape{2, 2, 1, 1};
    auto data_batch = std::make_shared<op::Parameter>(element::f32, shape);
    auto filters = std::make_shared<op::Parameter>(element::f32, shape);
    auto delta = std::make_shared<op::Parameter>(element::f32, shape);
694 695
    auto bias = make_shared<op::Parameter>(element::f32, Shape{shape[0]});
    auto pbroadcast = std::make_shared<op::Broadcast>(bias, shape, AxisSet{1, 2, 3});
696 697 698 699 700 701
    auto conv = std::make_shared<op::Convolution>(data_batch, filters);
    auto conv_bias = std::make_shared<op::Add>(conv, pbroadcast);

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    pass_manager.register_pass<pass::VisualizeTree>("conv_bias_bprop_fusion");
702
    auto f = make_shared<Function>(conv_bias, ParameterVector{data_batch, filters, bias});
703 704 705 706 707 708 709 710

    ngraph::autodiff::Adjoints adjoints(NodeVector{conv_bias}, NodeVector{delta});

    auto d_data = adjoints.backprop_node(data_batch);
    auto d_weights = adjoints.backprop_node(filters);
    auto d_bias = adjoints.backprop_node(bias);

    auto df = make_shared<Function>(NodeVector{d_data, d_weights, d_bias},
711
                                    ParameterVector{data_batch, filters, bias, delta});
712 713 714 715 716 717

    pass_manager.run_passes(df);
    size_t ccg = count_ops_of_type<op::ConvolutionBiasBackpropFiltersBias>(df);
    ASSERT_EQ(ccg, 1);
}

718 719 720 721 722 723 724 725 726 727 728 729
TEST(cpu_fusion, batchnorm_fprop_relu_b1c2h2w2)
{
    auto input_shape = Shape{1, 2, 2, 2};
    auto input = make_shared<op::Parameter>(element::f32, input_shape);
    auto mean_shape = Shape{2};
    auto var_shape = Shape{2};
    auto gamma_shape = Shape{2};
    auto gamma = make_shared<op::Parameter>(element::f32, gamma_shape);
    auto beta_shape = Shape{2};
    auto beta = make_shared<op::Parameter>(element::f32, beta_shape);
    double eps = 0.001;
    auto shape_r = Shape{1, 2, 2, 2};
730
    auto bn = make_shared<op::BatchNormTraining>(input, gamma, beta, eps);
731 732 733 734 735 736 737 738 739 740 741 742 743

    auto output_rt = std::make_shared<op::GetOutputElement>(bn, 0);
    // Note, op::Splice is used to break Relu(BatchNorm) fusion
    // otherwise we will be comparing two BatchNormRelus
    // Unfortunately, we can't use INTERPRETER for
    // verifying the results as it doesn't implement
    // BatchNorm op.
    auto slice =
        std::make_shared<op::Slice>(output_rt, Coordinate{0, 0, 0, 0}, Coordinate{1, 2, 2, 2});
    auto output_relu = std::make_shared<op::Relu>(slice);
    auto mean_rt = std::make_shared<op::GetOutputElement>(bn, 1);
    auto variance_rt = std::make_shared<op::GetOutputElement>(bn, 2);

744
    auto bn_relu = make_shared<op::BatchNormTrainingRelu>(eps, gamma, beta, input);
745 746 747 748 749 750
    auto output_rt_bnr = std::make_shared<op::GetOutputElement>(bn_relu, 0);
    auto mean_rt_bnr = std::make_shared<op::GetOutputElement>(bn_relu, 1);
    auto variance_rt_bnr = std::make_shared<op::GetOutputElement>(bn_relu, 2);

    auto f = make_shared<Function>(
        NodeVector{output_relu, mean_rt, variance_rt, output_rt_bnr, mean_rt_bnr, variance_rt_bnr},
751
        ParameterVector{input, gamma, beta});
752
    auto backend = runtime::Backend::create("CPU");
753 754

    // Create some tensors for input/output
755
    auto input_t = backend->create_tensor(element::f32, Shape{1, 2, 2, 2});
756 757 758 759 760 761 762 763 764 765

    copy_data(input_t,
              vector<float>{0.54881352f,
                            0.71518934f,
                            0.60276335f,
                            0.54488319f,
                            0.42365479f,
                            0.64589411f,
                            0.4375872f,
                            0.89177299f});
766
    auto gamma_t = backend->create_tensor(element::f32, gamma_shape);
767
    copy_data(gamma_t, vector<float>{1.0f, 1.0f});
768
    auto beta_t = backend->create_tensor(element::f32, beta_shape);
769
    copy_data(beta_t, vector<float>{0.0f, 0.0f});
770 771 772 773 774 775 776 777
    auto bn_output = backend->create_tensor(element::f32, shape_r);
    auto result_mean = backend->create_tensor(element::f32, mean_shape);
    auto result_variance = backend->create_tensor(element::f32, var_shape);

    auto bn_output_bnr = backend->create_tensor(element::f32, shape_r);
    auto result_mean_bnr = backend->create_tensor(element::f32, mean_shape);
    auto result_variance_bnr = backend->create_tensor(element::f32, var_shape);

778
    auto handle = backend->compile(f);
779 780 781 782 783 784 785
    handle->call_with_validate({bn_output,
                                result_mean,
                                result_variance,
                                bn_output_bnr,
                                result_mean_bnr,
                                result_variance_bnr},
                               {input_t, gamma_t, beta_t});
786 787 788 789 790 791 792 793

    EXPECT_TRUE(test::all_close(read_vector<float>(bn_output), read_vector<float>(bn_output_bnr)));
    EXPECT_TRUE(
        test::all_close(read_vector<float>(result_mean), read_vector<float>(result_mean_bnr)));
    EXPECT_TRUE(test::all_close(read_vector<float>(result_variance),
                                read_vector<float>(result_variance_bnr)));
}

794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843
static void test_batchnorm_fprop_relu(Shape input_shape)
{
    auto make_bn_relu_function = [&]() {
        auto c_axis = input_shape[1];
        auto input = make_shared<op::Parameter>(element::f32, input_shape);
        auto mean_shape = Shape{c_axis};
        auto var_shape = Shape{c_axis};
        auto gamma_shape = Shape{c_axis};
        auto gamma = make_shared<op::Parameter>(element::f32, gamma_shape);
        auto beta_shape = Shape{c_axis};
        auto beta = make_shared<op::Parameter>(element::f32, beta_shape);
        double eps = 0.001;
        auto shape_r = input_shape;
        auto bn = make_shared<op::BatchNormTraining>(eps, gamma, beta, input);
        auto output_rt = std::make_shared<op::GetOutputElement>(bn, 0);

        auto output_relu = std::make_shared<op::Relu>(output_rt);
        auto mean_rt = std::make_shared<op::GetOutputElement>(bn, 1);
        auto variance_rt = std::make_shared<op::GetOutputElement>(bn, 2);

        auto f = make_shared<Function>(NodeVector{output_relu, mean_rt, variance_rt},
                                       ParameterVector{input, gamma, beta});
        return f;
    };
    auto cpu_f = make_bn_relu_function();
    auto int_f = make_bn_relu_function();
    test::Uniform<float> rng(-10.0f, 10.0f);
    vector<vector<float>> args;

    for (shared_ptr<op::Parameter> param : int_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
    }
}

TEST(cpu_fusion, batchnorm_fprop_relu)
{
    test_batchnorm_fprop_relu(Shape{1, 2, 2, 2});
    test_batchnorm_fprop_relu(Shape{1, 2, 2, 2, 2});
    test_batchnorm_fprop_relu(Shape{2, 2, 2, 4, 4});
}

844 845 846 847 848 849 850 851
TEST(cpu_fusion, fuse_conv_relu)
{
    auto A = std::make_shared<op::Parameter>(element::f32, Shape{2, 1, 2, 2});
    auto weights = std::make_shared<op::Parameter>(element::f32, Shape{1, 1, 2, 2});
    auto convolution = std::make_shared<op::Convolution>(A, weights, Strides{1, 1}, Strides{1, 1});
    auto relu = std::make_shared<op::Relu>(convolution);
    auto abs_node =
        std::make_shared<op::Abs>(std::make_shared<op::Abs>(std::make_shared<op::Abs>(relu)));
852
    auto func = make_shared<Function>(abs_node, ParameterVector{A, weights});
853 854

    pass::Manager pass_manager;
855
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
856 857 858 859 860
    pass_manager.run_passes(func);
    size_t cb = count_ops_of_type<op::ConvolutionRelu>(func);
    ASSERT_GT(cb, 0);
}

861
TEST(cpu_fusion, conv_relu_n2c1h2w2_2)
862 863 864 865
{
    Shape shape_a{2, 1, 6, 6};
    Shape shape_weights{1, 1, 2, 2};

866 867 868 869 870
    auto make_int_function = [shape_a, shape_weights]() {
        auto A = std::make_shared<op::Parameter>(element::f32, shape_a);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto conv = std::make_shared<op::Convolution>(A, weights, Strides{2, 2}, Strides{1, 1});
        auto relu = std::make_shared<op::Relu>(conv);
871
        auto f = make_shared<Function>(NodeVector{relu}, ParameterVector{A, weights});
872 873
        return f;
    };
874

875
    auto int_f = make_int_function();
876

877 878 879 880 881
    auto make_cpu_function = [shape_a, shape_weights]() {
        auto A = std::make_shared<op::Parameter>(element::f32, shape_a);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto conv = std::make_shared<op::Convolution>(A, weights, Strides{2, 2}, Strides{1, 1});
        auto conv_relu = std::make_shared<op::ConvolutionRelu>(conv);
882
        auto f = make_shared<Function>(NodeVector{conv_relu}, ParameterVector{A, weights});
883 884
        return f;
    };
885

886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901
    auto cpu_f = make_cpu_function();

    vector<vector<float>> args{
        {1.25f,  2.25f, 5.25f, 6.25f,  -1.25f, -1.25f, 3.25f, -4.25f, 7.25f,  8.25f,  -1.25f,
         -1.25f, 1.25f, 2.25f, -3.25f, 2.25f,  4.25f,  4.25f, 1.25f,  2.25f,  -4.25f, 2.25f,
         4.25f,  4.25f, 0.f,   0.f,    -1.f,   0.f,    2.f,   2.f,    0.f,    0.f,    0.f,
         0.f,    2.f,   2.f,   1.25f,  2.25f,  5.25f,  6.25f, 1.25f,  1.25f,  3.25f,  4.25f,
         -7.25f, 8.25f, 1.25f, -1.25f, -1.25f, 2.25f,  3.25f, 2.25f,  -4.25f, -4.25f, -1.25f,
         -2.25f, 4.25f, 2.25f, 4.25f,  4.25f,  0.f,    0.f,   1.f,    0.f,    -2.f,   2.f,
         0.f,    0.f,   0.f,   0.f,    -2.f,   -2.f},
        {2., 2., 2., 2.}};

    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0)));
}
902

903 904 905 906 907 908 909 910 911 912 913 914 915 916
TEST(cpu_fusion, conv_bias_relu_n2c1h2w2_2)
{
    Shape shape_a{2, 1, 6, 6};
    Shape shape_weights{1, 1, 2, 2};
    Shape shape_bias{1};

    auto make_int_function = [shape_a, shape_weights, shape_bias]() {
        auto A = std::make_shared<op::Parameter>(element::f32, shape_a);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto conv = std::make_shared<op::Convolution>(A, weights, Strides{2, 2}, Strides{1, 1});
        auto bias = std::make_shared<op::Parameter>(element::f32, shape_bias);
        auto conv_bias =
            conv + std::make_shared<op::Broadcast>(bias, conv->get_shape(), AxisSet{0, 2, 3});
        auto relu = std::make_shared<op::Relu>(conv_bias);
917
        auto f = make_shared<Function>(NodeVector{relu}, ParameterVector{A, weights, bias});
918 919 920 921 922 923 924 925 926 927
        return f;
    };

    auto int_f = make_int_function();

    auto make_cpu_function = [shape_a, shape_weights, shape_bias]() {
        auto A = std::make_shared<op::Parameter>(element::f32, shape_a);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto bias = std::make_shared<op::Parameter>(element::f32, shape_bias);
        auto conv = std::make_shared<op::Convolution>(A, weights, Strides{2, 2}, Strides{1, 1});
928
        auto conv_bias_relu = std::make_shared<op::ConvolutionBias>(conv, bias, true);
929 930
        auto f =
            make_shared<Function>(NodeVector{conv_bias_relu}, ParameterVector{A, weights, bias});
931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951
        return f;
    };

    auto cpu_f = make_cpu_function();

    vector<vector<float>> args{
        {1.25f,  2.25f, 5.25f, 6.25f,  -1.25f, -1.25f, 3.25f, -4.25f, 7.25f,  8.25f,  -1.25f,
         -1.25f, 1.25f, 2.25f, -3.25f, 2.25f,  4.25f,  4.25f, 1.25f,  2.25f,  -4.25f, 2.25f,
         4.25f,  4.25f, 0.f,   0.f,    -1.f,   0.f,    2.f,   2.f,    0.f,    0.f,    0.f,
         0.f,    2.f,   2.f,   1.25f,  2.25f,  5.25f,  6.25f, 1.25f,  1.25f,  3.25f,  4.25f,
         -7.25f, 8.25f, 1.25f, -1.25f, -1.25f, 2.25f,  3.25f, 2.25f,  -4.25f, -4.25f, -1.25f,
         -2.25f, 4.25f, 2.25f, 4.25f,  4.25f,  0.f,    0.f,   1.f,    0.f,    -2.f,   2.f,
         0.f,    0.f,   0.f,   0.f,    -2.f,   -2.f},
        {2., 2., 2., 2.},
        {0.1f}};

    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0)));
}

952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975
TEST(cpu_fusion, conv_horizontal_fusion)
{
    Shape shape_a{2, 1, 6, 6};
    Shape shape_weights{1, 1, 2, 2};
    Shape shape_bias{1};

    auto make_function = [shape_a, shape_weights, shape_bias]() {
        auto A = std::make_shared<op::Parameter>(element::f32, shape_a);
        auto weights1 = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto conv1 = std::make_shared<op::Convolution>(A, weights1, Strides{2, 2}, Strides{1, 1});
        auto bias1 = std::make_shared<op::Parameter>(element::f32, shape_bias);
        auto conv_bias1 =
            conv1 + std::make_shared<op::Broadcast>(bias1, conv1->get_shape(), AxisSet{0, 2, 3});
        auto relu1 = std::make_shared<op::Relu>(conv_bias1);

        auto weights2 = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto conv2 = std::make_shared<op::Convolution>(A, weights2, Strides{2, 2}, Strides{1, 1});
        auto bias2 = std::make_shared<op::Parameter>(element::f32, shape_bias);
        auto conv_bias2 =
            conv2 + std::make_shared<op::Broadcast>(bias2, conv2->get_shape(), AxisSet{0, 2, 3});
        auto relu2 = std::make_shared<op::Relu>(conv_bias2);

        auto concat = std::make_shared<op::Concat>(NodeVector{relu1, relu2}, 1);
        auto f = make_shared<Function>(NodeVector{concat},
976
                                       ParameterVector{A, weights1, bias1, weights2, bias2});
977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002
        return f;
    };
    auto int_f = make_function();
    auto cpu_f = make_function();

    vector<vector<float>> args{
        {1.25f,  2.25f, 5.25f, 6.25f,  -1.25f, -1.25f, 3.25f, -4.25f, 7.25f,  8.25f,  -1.25f,
         -1.25f, 1.25f, 2.25f, -3.25f, 2.25f,  4.25f,  4.25f, 1.25f,  2.25f,  -4.25f, 2.25f,
         4.25f,  4.25f, 0.f,   0.f,    -1.f,   0.f,    2.f,   2.f,    0.f,    0.f,    0.f,
         0.f,    2.f,   2.f,   1.25f,  2.25f,  5.25f,  6.25f, 1.25f,  1.25f,  3.25f,  4.25f,
         -7.25f, 8.25f, 1.25f, -1.25f, -1.25f, 2.25f,  3.25f, 2.25f,  -4.25f, -4.25f, -1.25f,
         -2.25f, 4.25f, 2.25f, 4.25f,  4.25f,  0.f,    0.f,   1.f,    0.f,    -2.f,   2.f,
         0.f,    0.f,   0.f,   0.f,    -2.f,   -2.f},
        {2., 2., 2., 2.},
        {0.1f},
        {3., 3., 3., 3.},
        {0.2f}};

    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0)));

    size_t cpu_cb = count_ops_of_type<op::ConvolutionBias>(cpu_f);
    ASSERT_EQ(cpu_cb, 1);
}

1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018
// ConvolutionBiasAdd relies on an in-place fused MKLDNN kernel.
// Need to ensure that it is fused only when in-place buffer allocation is feasible
shared_ptr<Function> gen_conv_bias_add(bool param_input, bool result_output)
{
    auto A = make_shared<op::Parameter>(element::f32, Shape{2, 1, 2, 2});
    auto weights = make_shared<op::Parameter>(element::f32, Shape{1, 1, 1, 1});
    auto bias = make_shared<op::Parameter>(element::f32, Shape{1});
    auto conv = make_shared<op::Convolution>(A, weights, Strides{1, 1}, Strides{1, 1});
    auto bias_broadcast = make_shared<op::Broadcast>(bias, conv->get_shape(), AxisSet{0, 2, 3});
    auto convbias = conv + bias_broadcast;
    auto B = make_shared<op::Parameter>(element::f32, Shape{2, 1, 2, 2});
    auto abs_B = make_shared<op::Abs>(B);
    auto add =
        param_input ? make_shared<op::Add>(convbias, B) : make_shared<op::Add>(convbias, abs_B);
    auto abs = make_shared<op::Abs>(add);

1019 1020
    return result_output ? make_shared<Function>(add, ParameterVector{A, weights, bias, B})
                         : make_shared<Function>(abs, ParameterVector{A, weights, bias, B});
1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037
}

TEST(cpu_fusion, fuse_conv_bias_add)
{
    auto func_fuse = gen_conv_bias_add(false, false);
    auto func_nofuse1 = gen_conv_bias_add(true, false);
    auto func_nofuse2 = gen_conv_bias_add(false, true);

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    pass_manager.run_passes(func_fuse);
    ASSERT_EQ(count_ops_of_type<op::ConvolutionBiasAdd>(func_fuse), 1);

    pass_manager.run_passes(func_nofuse1);
    ASSERT_EQ(count_ops_of_type<op::ConvolutionBiasAdd>(func_nofuse1), 0);

    pass_manager.run_passes(func_nofuse2);
gaurides's avatar
gaurides committed
1038
    ASSERT_EQ(count_ops_of_type<op::ConvolutionBiasAdd>(func_nofuse2), 1);
1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055
}

TEST(cpu_fusion, conv_bias_add)
{
    auto int_f = gen_conv_bias_add(false, false);
    auto cpu_f = gen_conv_bias_add(false, false);

    vector<vector<float>> args{{1.25f, 2.25f, 5.25f, 6.25f, -1.25f, -1.25f, 3.25f, -4.25f},
                               {-1.25f},
                               {2.25f},
                               {1.25f, 2.25f, -3.25f, 2.25f, 4.25f, 4.25f, 1.25f, 2.25f}};

    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0)));
}

gaurides's avatar
gaurides committed
1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067
// ConvolutionAdd relies on an in-place fused MKLDNN kernel.
// Need to ensure that it is fused only when in-place buffer allocation is feasible
shared_ptr<Function> gen_conv_add(bool param_input, bool result_output)
{
    auto A = make_shared<op::Parameter>(element::f32, Shape{2, 1, 2, 2});
    auto weights = make_shared<op::Parameter>(element::f32, Shape{1, 1, 1, 1});
    auto conv = make_shared<op::Convolution>(A, weights, Strides{1, 1}, Strides{1, 1});
    auto B = make_shared<op::Parameter>(element::f32, Shape{2, 1, 2, 2});
    auto abs_B = make_shared<op::Abs>(B);
    auto add = param_input ? make_shared<op::Add>(conv, B) : make_shared<op::Add>(conv, abs_B);
    auto abs = make_shared<op::Abs>(add);

1068 1069
    return result_output ? make_shared<Function>(add, ParameterVector{A, weights, B})
                         : make_shared<Function>(abs, ParameterVector{A, weights, B});
gaurides's avatar
gaurides committed
1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110
}

TEST(cpu_fusion, fuse_conv_add)
{
    auto func_fuse = gen_conv_add(false, false);
    auto func_nofuse1 = gen_conv_add(true, false);
    auto func_nofuse2 = gen_conv_add(false, true);

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    pass_manager.run_passes(func_fuse);
    ASSERT_EQ(count_ops_of_type<op::ConvolutionAdd>(func_fuse), 1);

    pass_manager.run_passes(func_nofuse1);
    ASSERT_EQ(count_ops_of_type<op::ConvolutionAdd>(func_nofuse1), 0);

    pass_manager.run_passes(func_nofuse2);
    ASSERT_EQ(count_ops_of_type<op::ConvolutionAdd>(func_nofuse2), 1);
}

TEST(cpu_fusion, conv_add)
{
    auto int_f = gen_conv_add(false, false);
    auto cpu_f = gen_conv_add(false, false);

    vector<vector<float>> args{{1.25f, 2.25f, 5.25f, 6.25f, -1.25f, -1.25f, 3.25f, -4.25f},
                               {-1.25f},
                               {1.25f, 2.25f, -3.25f, 2.25f, 4.25f, 4.25f, 1.25f, 2.25f}};

    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0)));

    int_f = gen_conv_add(false, true);
    cpu_f = gen_conv_add(false, true);

    int_results = execute(int_f, args, "INTERPRETER");
    cpu_results = execute(cpu_f, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0)));
}

1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142
shared_ptr<Function> gen_groupconv_batchnorm(const bool add_goe,
                                             const bool with_relu,
                                             const Shape shape_in,
                                             const Shape shape_weights,
                                             const Shape shape_out,
                                             const size_t groups)
{
    auto input = make_shared<op::Parameter>(element::f32, shape_in);
    auto weights = make_shared<op::Parameter>(element::f32, shape_weights);

    unsigned long OC = shape_out.at(1);
    Shape shape_bn{OC};
    auto group_conv = make_shared<op::GroupConvolution>(input,
                                                        weights,
                                                        Strides{1, 1},
                                                        Strides{1, 1},
                                                        CoordinateDiff{0, 0},
                                                        CoordinateDiff{0, 0},
                                                        Strides{1, 1},
                                                        groups,
                                                        shape_out);

    double eps = 0.001;
    auto gamma = std::make_shared<op::Parameter>(element::f32, shape_bn);
    auto beta = std::make_shared<op::Parameter>(element::f32, shape_bn);
    auto mean = std::make_shared<op::Parameter>(element::f32, shape_bn);
    auto var = std::make_shared<op::Parameter>(element::f32, shape_bn);

    auto goe_bn = std::make_shared<op::GetOutputElement>(group_conv, 0);

    // Adding a goe will stop fusion since the patterns wont expect to see this op
    auto bn =
1143 1144
        add_goe ? std::make_shared<op::BatchNormInference>(goe_bn, gamma, beta, mean, var, eps)
                : std::make_shared<op::BatchNormInference>(group_conv, gamma, beta, mean, var, eps);
1145 1146 1147 1148
    if (with_relu)
    {
        auto prelu = std::make_shared<op::Relu>(bn);
        auto f = make_shared<Function>(NodeVector{prelu},
1149
                                       ParameterVector{input, weights, gamma, beta, mean, var});
1150 1151 1152 1153 1154
        return f;
    }
    else
    {
        auto f = make_shared<Function>(NodeVector{bn},
1155
                                       ParameterVector{input, weights, gamma, beta, mean, var});
1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249
        return f;
    }
}

void fuse_groupconv_batchnorm_helper(Shape shape_in,
                                     Shape shape_weights,
                                     Shape shape_r,
                                     size_t groups)
{
    auto func_fuse =
        gen_groupconv_batchnorm(false, false, shape_in, shape_weights, shape_r, groups);
    auto func_fuse2 =
        gen_groupconv_batchnorm(false, true, shape_in, shape_weights, shape_r, groups);

    {
        pass::Manager pass_manager;
        pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
        pass_manager.run_passes(func_fuse);
        ASSERT_EQ(count_ops_of_type<op::GroupConvolutionBias>(func_fuse), 1);
    }

    {
        // test groupconv + batchnorm + relu fusion
        pass::Manager pass_manager;
        pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
        pass_manager.run_passes(func_fuse2);
        ASSERT_EQ(count_ops_of_type<op::GroupConvolutionBias>(func_fuse2), 1);
        ASSERT_EQ(count_ops_of_type<op::Relu>(func_fuse2), 0);
    }
}

void groupconv_batchnorm_test_val_helper(
    const bool with_relu, Shape shape_in, Shape shape_weights, Shape shape_r, size_t groups)
{
    shared_ptr<Function> fuse_func =
        gen_groupconv_batchnorm(false, with_relu, shape_in, shape_weights, shape_r, groups);
    shared_ptr<Function> nofuse_func =
        gen_groupconv_batchnorm(true, with_relu, shape_in, shape_weights, shape_r, groups);

    test::Uniform<float> rng(1.0f, 100.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : fuse_func->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    auto fuse_results = execute(fuse_func, args, "CPU");
    auto nofuse_results = execute(nofuse_func, args, "CPU");

    EXPECT_TRUE(test::all_close(fuse_results.at(0), nofuse_results.at(0)));
}

TEST(cpu_fusion, fuse_groupconv_batchnorm1)
{
    Shape shape_in{1, 20, 5, 5};
    Shape shape_weights{8, 10, 3, 3};
    Shape shape_r{1, 8, 3, 3};
    fuse_groupconv_batchnorm_helper(shape_in, shape_weights, shape_r, 2);
    groupconv_batchnorm_test_val_helper(false, shape_in, shape_weights, shape_r, 2);
    groupconv_batchnorm_test_val_helper(true, shape_in, shape_weights, shape_r, 2);
}

TEST(cpu_fusion, fuse_groupconv_batchnorm2)
{
    Shape shape_in{1, 20, 5, 5};
    Shape shape_weights{5, 4, 3, 3};
    Shape shape_r{1, 5, 3, 3};
    fuse_groupconv_batchnorm_helper(shape_in, shape_weights, shape_r, 5);
    groupconv_batchnorm_test_val_helper(false, shape_in, shape_weights, shape_r, 5);
    groupconv_batchnorm_test_val_helper(true, shape_in, shape_weights, shape_r, 5);
}

TEST(cpu_fusion, fuse_groupconv_batchnorm3)
{
    Shape shape_in{1, 20, 5, 5};
    Shape shape_weights{20, 1, 3, 3};
    Shape shape_r{1, 20, 3, 3};
    fuse_groupconv_batchnorm_helper(shape_in, shape_weights, shape_r, 20);
    groupconv_batchnorm_test_val_helper(false, shape_in, shape_weights, shape_r, 20);
    groupconv_batchnorm_test_val_helper(true, shape_in, shape_weights, shape_r, 20);
}

TEST(cpu_fusion, fuse_groupconv_batchnorm4)
{
    Shape shape_in{1, 20, 4, 4};
    Shape shape_weights{5, 20, 1, 1};
    Shape shape_r{1, 5, 4, 4};
    fuse_groupconv_batchnorm_helper(shape_in, shape_weights, shape_r, 1);
    groupconv_batchnorm_test_val_helper(false, shape_in, shape_weights, shape_r, 1);
    groupconv_batchnorm_test_val_helper(true, shape_in, shape_weights, shape_r, 1);
}

1250 1251 1252 1253 1254 1255 1256 1257
std::vector<shared_ptr<runtime::Tensor>> rnn_matrix_fusion_eval(const size_t time_steps,
                                                                const Shape& data_shape,
                                                                const Shape& weights_shape,
                                                                const Shape& bias_shape,
                                                                const vector<float>& data_val,
                                                                const vector<float>& weights_val,
                                                                const vector<float>& bias_val,
                                                                const bool enable_pass)
1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277
{
    auto data = make_shared<op::Parameter>(element::f32, data_shape);
    auto weights = make_shared<op::Parameter>(element::f32, weights_shape);
    auto bias = make_shared<op::Parameter>(element::f32, bias_shape);

    // results from each time step
    NodeVector results;
    for (size_t t = 0; t < time_steps; ++t)
    {
        auto data_slice = make_shared<op::Slice>(
            data, Coordinate{0, t, 0}, Coordinate{data_shape[0], t + 1, data_shape[2]});
        auto data_reshape = make_shared<op::Reshape>(
            data_slice, AxisVector{0, 1, 2}, Shape{data_shape[0], data_shape[2]});
        auto weights_reshape = make_shared<op::Reshape>(
            weights, AxisVector{1, 0}, Shape{weights_shape[1], weights_shape[0]});
        auto dot = make_shared<op::Dot>(data_reshape, weights_reshape);
        auto bias_broadcast = make_shared<op::Broadcast>(bias, dot->get_shape(), AxisSet{0});
        auto add = make_shared<op::Add>(dot, bias_broadcast);
        results.push_back(add);
    }
1278
    auto func = make_shared<Function>(results, ParameterVector{data, weights, bias});
1279 1280 1281 1282
    if (enable_pass)
    {
        pass::Manager pass_manager;
        pass_manager.register_pass<runtime::cpu::pass::CPURnnMatFusion>();
1283
        pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
1284 1285 1286 1287 1288 1289
        pass_manager.run_passes(func);
        // check all of our dot/add are converted to a single MatmulBias op.
        size_t count = count_ops_of_type<op::MatmulBias>(func);
        EXPECT_EQ(count, 1);
    }

1290
    auto backend = runtime::Backend::create("CPU");
1291

1292
    shared_ptr<runtime::Tensor> data_tensor =
1293
        backend->create_tensor(element::f32, data->get_shape());
1294
    shared_ptr<runtime::Tensor> weights_tensor =
1295
        backend->create_tensor(element::f32, weights->get_shape());
1296
    shared_ptr<runtime::Tensor> bias_tensor =
1297
        backend->create_tensor(element::f32, bias->get_shape());
1298

1299
    std::vector<shared_ptr<runtime::Tensor>> result_tensors;
1300 1301
    for (auto r : results)
    {
1302
        result_tensors.push_back(backend->create_tensor(element::f32, r->get_shape()));
1303 1304 1305 1306 1307
    }

    copy_data(data_tensor, data_val);
    copy_data(weights_tensor, weights_val);
    copy_data(bias_tensor, bias_val);
1308 1309
    auto handle = backend->compile(func);
    handle->call_with_validate(result_tensors, {data_tensor, weights_tensor, bias_tensor});
1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327
    return result_tensors;
}

TEST(cpu_fusion, rnn_matrix_fusion_eval_pass)
{
    const size_t time_steps = 4;
    Shape data_shape{3, time_steps, 5};
    Shape weights_shape{6, data_shape[2]};
    Shape bias_shape{6};

    test::Uniform<float> rng{0, 1, 0};
    vector<float> data_val(shape_size(data_shape));
    vector<float> weights_val(shape_size(weights_shape));
    vector<float> bias_val(shape_size(bias_shape));
    rng.initialize(data_val);
    rng.initialize(weights_val);
    rng.initialize(bias_val);

1328
    std::vector<shared_ptr<runtime::Tensor>> result_expected = rnn_matrix_fusion_eval(
1329
        time_steps, data_shape, weights_shape, bias_shape, data_val, weights_val, bias_val, false);
1330
    std::vector<shared_ptr<runtime::Tensor>> result_fused = rnn_matrix_fusion_eval(
1331 1332 1333 1334 1335 1336
        time_steps, data_shape, weights_shape, bias_shape, data_val, weights_val, bias_val, true);
    for (size_t i = 0; i < result_expected.size(); ++i)
    {
        EXPECT_TRUE(test::all_close<float>(result_expected[i], result_fused[i]));
    }
}
1337 1338 1339 1340 1341

TEST(cpu_fusion, rnn_fusion_from_json_model)
{
    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPURnnMatFusion>();
1342
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
1343 1344 1345 1346 1347 1348 1349
    const string json_path =
        file_util::path_join(SERIALIZED_ZOO, "mxnet/rnn-10-step-fusion-test.json");
    const string json_string = file_util::read_file_to_string(json_path);
    stringstream ss(json_string);
    shared_ptr<Function> func = ngraph::deserialize(ss);
    pass_manager.run_passes(func);
    const size_t NUM_STEPS = 10;
1350
    auto mmb_predicate = [=](std::shared_ptr<Node> node) {
1351
        auto users = node->get_users();
1352
        return (users.size() == NUM_STEPS) &&
1353 1354 1355 1356 1357 1358 1359 1360
               std::all_of(begin(users), end(users), [](std::shared_ptr<Node> n) {
                   return std::dynamic_pointer_cast<op::Slice>(n) != nullptr;
               });
    };

    auto mmbs = get_ops_of_type<op::MatmulBias>(func);
    ASSERT_TRUE(std::any_of(begin(mmbs), end(mmbs), mmb_predicate));
}
Nick Korovaiko's avatar
Nick Korovaiko committed
1361 1362 1363 1364 1365 1366 1367

TEST(cpu_fusion, weight_fusion)
{
    auto param = std::make_shared<op::Parameter>(element::f32, Shape{64});
    auto reshape_conv =
        std::make_shared<ngraph::op::Reshape>(param, AxisVector{0}, Shape{16, 4, 1, 1});
    auto data_conv = std::make_shared<op::Parameter>(element::f32, Shape{16, 4, 7, 7});
Scott Cyphers's avatar
Scott Cyphers committed
1368
    auto tvt = reshape_conv->get_outputs().at(0).get_tensor_ptr().get();
1369
    auto lt_desc = std::make_shared<runtime::cpu::LayoutDescriptor>(*tvt);
Nick Korovaiko's avatar
Nick Korovaiko committed
1370 1371 1372 1373 1374 1375 1376
    auto cvt_lt_conv = std::make_shared<runtime::cpu::op::ConvertLayout>(reshape_conv, lt_desc);
    auto conv = std::make_shared<ngraph::op::Convolution>(
        data_conv, cvt_lt_conv, Strides{1, 1}, Strides{1, 1});

    auto reshape_conv_bprop =
        std::make_shared<op::Reshape>(param, AxisVector{0}, Shape{16, 4, 1, 1});
    auto dummy_arg_conv_bprop = std::make_shared<op::Parameter>(element::f32, Shape{1, 16, 7, 7});
Scott Cyphers's avatar
Scott Cyphers committed
1377
    auto tvt_bprop = reshape_conv_bprop->get_outputs().at(0).get_tensor_ptr().get();
1378
    auto lt_desc_bprop = std::make_shared<runtime::cpu::LayoutDescriptor>(*tvt_bprop);
Nick Korovaiko's avatar
Nick Korovaiko committed
1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393
    auto cvt_lt_conv_bprop =
        std::make_shared<runtime::cpu::op::ConvertLayout>(reshape_conv_bprop, lt_desc_bprop);
    auto conv_bprop = std::make_shared<op::ConvolutionBackpropData>(Shape{1, 4, 7, 7},
                                                                    cvt_lt_conv_bprop,
                                                                    dummy_arg_conv_bprop,
                                                                    Strides{1, 1},
                                                                    Strides{1, 1},
                                                                    CoordinateDiff{0, 0},
                                                                    CoordinateDiff{0, 0},
                                                                    Strides{1, 1});

    auto conv_relu = std::make_shared<op::Relu>(conv);
    auto conv_bprop_abs = std::make_shared<op::Abs>(conv_bprop);

    auto f = make_shared<Function>(NodeVector{conv_relu, conv_bprop_abs},
1394
                                   ParameterVector{param, data_conv, dummy_arg_conv_bprop});
Nick Korovaiko's avatar
Nick Korovaiko committed
1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPUPostLayoutOptimizations>();
    pass_manager.run_passes(f);

    auto new_conv_bprop_data = conv_bprop_abs->get_argument(0);
    auto new_convert_layout = new_conv_bprop_data->get_argument(0);

    ASSERT_EQ(std::dynamic_pointer_cast<runtime::cpu::op::ConvertLayout>(
                  new_convert_layout->get_argument(0)),
              cvt_lt_conv);
}
1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419

TEST(cpu_fusion, max_pool_with_indices)
{
    Shape shape_a{10, 3, 28, 28};
    auto input = std::make_shared<op::Parameter>(element::f32, shape_a);
    Shape window_shape{2, 2};
    auto max_pool = std::make_shared<op::MaxPool>(input, window_shape);
    auto C = std::make_shared<op::Parameter>(element::f32, max_pool->get_shape());

    ngraph::autodiff::Adjoints adjoints(NodeVector{max_pool}, NodeVector{C});

    auto dinput = adjoints.backprop_node(input);

1420
    auto df = std::make_shared<Function>(NodeVector{dinput}, ParameterVector{input, C});
1421

1422
    auto f = std::make_shared<Function>(NodeVector{max_pool}, ParameterVector{input});
1423 1424 1425 1426 1427 1428 1429 1430

    {
        pass::Manager pass_manager;
        pass_manager.register_pass<pass::VisualizeTree>("max_pool_fprop_before.pdf");
        pass_manager.run_passes(f);
    }

    {
1431
        NodeVector nv_cwi;
1432 1433
        pass::Manager pass_manager;
        pass_manager.register_pass<pass::VisualizeTree>("max_pool_bprop_before.pdf");
1434
        pass_manager.register_pass<runtime::cpu::pass::CPUWorkspaceInsertion>(nv_cwi);
1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463
        pass_manager.register_pass<pass::VisualizeTree>("max_pool_bprop_after.pdf");
        pass_manager.run_passes(df);
    }

    {
        pass::Manager pass_manager;
        pass_manager.register_pass<pass::VisualizeTree>("max_pool_fprop_after.pdf");
        pass_manager.run_passes(f);
    }

    auto maxpool_goe_output =
        std::dynamic_pointer_cast<op::GetOutputElement>(f->get_results().at(0)->get_argument(0));
    ASSERT_TRUE(maxpool_goe_output);
    ASSERT_EQ(maxpool_goe_output->get_n(), 0);
    auto maxpool_with_indices = df->get_results().at(0)->get_argument(0);
    auto maxpool_goe_indices =
        std::dynamic_pointer_cast<op::GetOutputElement>(maxpool_with_indices->get_argument(2));
    ASSERT_TRUE(maxpool_goe_indices);
    ASSERT_EQ(maxpool_goe_indices->get_n(), 1);
}

TEST(cpu_fusion, backwards_maxpool_with_indices_n4_c1_hw4_2x2_max)
{
    Shape shape_a{1, 4, 4, 4};
    Shape maxpool_shape{1, 4, 3, 3};
    auto A = std::make_shared<op::Parameter>(element::f32, shape_a);
    Shape window_shape{2, 2};
    auto window_movement_strides = Strides{1, 1};
    auto maxpool = std::make_shared<op::MaxPool>(A, window_shape, window_movement_strides);
1464
    auto f = std::make_shared<Function>(maxpool, ParameterVector{A});
1465 1466

    auto backend = runtime::Backend::create("CPU");
1467
    shared_ptr<runtime::Tensor> ep = backend->create_tensor(element::f32, maxpool_shape);
1468 1469
    vector<float> dataEp(shape_size(maxpool_shape), 4);

1470 1471
    shared_ptr<runtime::Tensor> input = backend->create_tensor(element::f32, shape_a);
    shared_ptr<runtime::Tensor> output = backend->create_tensor(element::f32, shape_a);
1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493

    vector<float> dataInput{11.f, 31.f, 40.f, 47.f, 13.f, 61.f, 48.f, 59.f, 17.f, 39.f, 64.f,
                            62.f, 45.f, 55.f, 36.f, 19.f, 65.f, 33.f, 49.f, 30.f, 56.f, 41.f,
                            53.f, 58.f, 22.f, 35.f, 52.f, 50.f, 63.f, 54.f, 12.f, 26.f, 44.f,
                            21.f, 69.f, 24.f, 46.f, 25.f, 51.f, 29.f, 72.f, 15.f, 73.f, 10.f,
                            16.f, 37.f, 70.f, 32.f, 28.f, 66.f, 57.f, 27.f, 60.f, 42.f, 43.f,
                            71.f, 18.f, 38.f, 67.f, 68.f, 14.f, 20.f, 34.f, 23.f};

    vector<float> expected{0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 12.0f, 0.0f, 4.0f, 0.0f, 0.0f,  16.0f,
                           0.0f, 0.0f, 4.0f, 0.0f, 0.0f, 4.0f,  0.0f, 0.0f, 0.0f, 4.0f,  0.0f,
                           8.0f, 8.0f, 0.0f, 0.0f, 4.0f, 0.0f,  4.0f, 4.0f, 0.0f, 0.0f,  0.0f,
                           0.0f, 8.0f, 0.0f, 4.0f, 0.0f, 0.0f,  0.0f, 8.0f, 0.0f, 16.0f, 0.0f,
                           0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 8.0f,  0.0f, 0.0f, 4.0f, 0.0f,  0.0f,
                           8.0f, 0.0f, 4.0f, 8.0f, 4.0f, 0.0f,  0.0f, 0.0f, 0.0f};

    copy_data(ep, dataEp);
    copy_data(input, dataInput);

    auto C = std::make_shared<op::Parameter>(element::f32, maxpool_shape);
    auto df = autodiff::backprop_function(f);

    {
1494
        NodeVector nv_cwi;
1495 1496
        pass::Manager pass_manager;
        pass_manager.register_pass<pass::VisualizeTree>("max_pool_bprop_before2.pdf");
1497
        pass_manager.register_pass<runtime::cpu::pass::CPUWorkspaceInsertion>(nv_cwi);
1498 1499 1500 1501
        pass_manager.register_pass<pass::VisualizeTree>("max_pool_bprop_after2.pdf");
        pass_manager.run_passes(df);
    }

1502
    auto handle = backend->compile(df);
1503
    handle->call_with_validate({output}, {input, ep});
1504
    EXPECT_TRUE(test::all_close_f(read_vector<float>(output), expected, MIN_FLOAT_TOLERANCE_BITS));
1505
}
1506

1507 1508 1509
#if defined(NGRAPH_HALIDE)

TEST(cpu_fusion, loop_kernel_one_input_one_output_halide)
1510 1511
{
    Shape shapeA{2, 2};
1512 1513 1514
    auto A = make_shared<op::Parameter>(element::f32, shapeA);
    auto relu_a = make_shared<op::Relu>(A);
    auto relu_relu_a = make_shared<op::Relu>(relu_a);
1515
    auto lk = make_shared<runtime::cpu::op::LoopKernel>(
1516
        NodeVector{relu_a, relu_relu_a}, NodeVector{relu_relu_a}, NodeVector{A});
1517
    auto f = make_shared<Function>(NodeVector{lk}, ParameterVector{A});
1518 1519

    auto backend = runtime::Backend::create("CPU");
1520 1521
    shared_ptr<runtime::Tensor> a = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> result = backend->create_tensor(element::f32, shapeA);
1522

1523
    vector<float> dataA{-1, 4, -1, 4};
1524
    copy_data(a, dataA);
1525
    vector<float> expected{0, 4, 0, 4};
1526

1527
    auto handle = backend->compile(f);
1528
    handle->call_with_validate({result}, {a});
1529

1530
    EXPECT_TRUE(test::all_close(read_vector<float>(result), expected));
1531 1532
}

1533
TEST(cpu_fusion, loop_kernel_two_input_two_output_halide)
1534 1535
{
    Shape shapeA{2, 2};
1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560
    auto A = make_shared<op::Parameter>(element::f32, shapeA);
    auto B = make_shared<op::Parameter>(element::f32, shapeA);
    auto relu_a = make_shared<op::Relu>(A);
    auto add_ab = make_shared<op::Add>(relu_a, B);

    auto lk = make_shared<runtime::cpu::op::LoopKernel>(
        NodeVector{relu_a, add_ab}, NodeVector{relu_a, add_ab}, NodeVector{A, B});

    auto goe1 = make_shared<op::GetOutputElement>(lk, 0);
    auto goe2 = make_shared<op::GetOutputElement>(lk, 1);
    auto f = make_shared<Function>(NodeVector{goe1, goe2}, ParameterVector{A, B});

    auto backend = runtime::Backend::create("CPU");
    shared_ptr<runtime::Tensor> a = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> b = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> result_relu = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> result_add = backend->create_tensor(element::f32, shapeA);

    vector<float> dataA{-1, 4, -1, 4};
    vector<float> dataB{0, 4, 0, 4};
    copy_data(a, dataA);
    copy_data(b, dataB);
    vector<float> expected_relu{0, 4, 0, 4};
    vector<float> expected_add{4, 4, 4, 4};

1561 1562
    auto handle = backend->compile(f);
    handle->call_with_validate({result_relu, result_add}, {a, b});
1563 1564 1565 1566 1567 1568 1569 1570 1571

    EXPECT_TRUE(test::all_close(read_vector<float>(result_relu), expected_relu));
}

TEST(cpu_fusion, loop_kernel_embedded_graph_halide)
{
    Shape shapeA{2, 2};
    auto A = make_shared<op::Parameter>(element::f32, shapeA);
    auto B = make_shared<op::Parameter>(element::f32, shapeA);
1572 1573 1574 1575 1576
    auto neg_a = make_shared<op::Negative>(A);
    auto neg_b = make_shared<op::Negative>(B);
    auto add = neg_a + neg_b;
    auto lk = make_shared<runtime::cpu::op::LoopKernel>(
        NodeVector{add}, NodeVector{add}, NodeVector{neg_a, neg_b});
1577
    auto f = make_shared<Function>(NodeVector{lk}, ParameterVector{A, B});
1578 1579

    auto backend = runtime::Backend::create("CPU");
1580 1581 1582
    shared_ptr<runtime::Tensor> a = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> b = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> result = backend->create_tensor(element::f32, shapeA);
1583

1584
    vector<float> dataA{1, 4, 1, 4};
1585
    copy_data(a, dataA);
1586
    vector<float> dataB{1, 2, 3, 4};
1587
    copy_data(b, dataB);
1588
    vector<float> expected{-2, -6, -4, -8};
1589
    auto handle = backend->compile(f);
1590
    handle->call_with_validate({result}, {a, b});
1591
    EXPECT_TRUE(test::all_close_f(read_vector<float>(result), expected, MIN_FLOAT_TOLERANCE_BITS));
1592 1593
}

1594
TEST(cpu_fusion, loop_kernel_two_inputs_one_output_halide)
1595 1596
{
    Shape shapeA{2, 2};
1597 1598
    auto A = make_shared<op::Parameter>(element::f32, shapeA);
    auto B = make_shared<op::Parameter>(element::f32, shapeA);
1599 1600 1601
    auto add = A + B;
    auto lk = make_shared<runtime::cpu::op::LoopKernel>(
        NodeVector{add}, NodeVector{add}, NodeVector{A, B});
1602
    auto f = make_shared<Function>(NodeVector{lk}, ParameterVector{A, B});
1603 1604

    auto backend = runtime::Backend::create("CPU");
1605 1606 1607
    shared_ptr<runtime::Tensor> a = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> b = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> result = backend->create_tensor(element::f32, shapeA);
1608

1609
    vector<float> dataA{1, 4, 1, 4};
1610
    copy_data(a, dataA);
1611
    vector<float> dataB{1, 2, 3, 4};
1612
    copy_data(b, dataB);
1613
    vector<float> expected{2, 6, 4, 8};
1614

1615
    auto handle = backend->compile(f);
1616
    handle->call_with_validate({result}, {a, b});
1617

1618
    EXPECT_TRUE(test::all_close_f(read_vector<float>(result), expected, MIN_FLOAT_TOLERANCE_BITS));
1619 1620
}

1621
TEST(cpu_fusion, loop_kernel_multiple_outputs_halide)
1622 1623
{
    Shape shapeA{2, 2};
1624 1625 1626 1627
    auto A = make_shared<op::Parameter>(element::f32, shapeA);
    auto B = make_shared<op::Parameter>(element::f32, shapeA);
    auto C = make_shared<op::Parameter>(element::f32, shapeA);
    auto D = make_shared<op::Parameter>(element::f32, shapeA);
1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646

    auto neg_a = make_shared<op::Negative>(A);
    auto neg_b = make_shared<op::Negative>(B);
    auto add_ab = neg_a + neg_b;
    auto add_cd = C + B;
    auto add_cd_abs = make_shared<op::Abs>(add_cd);
    auto add_ab_abs = make_shared<op::Abs>(add_ab);
    auto add_aab = add_ab_abs + A;
    auto add_cdd = add_cd_abs + D;

    auto lk = make_shared<runtime::cpu::op::LoopKernel>(
        NodeVector{neg_a, neg_b, add_ab, add_cd, add_cd_abs, add_ab_abs, add_aab, add_cdd},
        NodeVector{add_aab, add_cdd, neg_b},
        NodeVector{A, B, C, D});
    auto add_aab_goe = std::make_shared<op::GetOutputElement>(lk, 0);
    auto add_cdd_goe = std::make_shared<op::GetOutputElement>(lk, 1);
    auto neg_b_goe = std::make_shared<op::GetOutputElement>(lk, 2);

    auto f = make_shared<Function>(NodeVector{add_aab_goe, add_cdd_goe, neg_b_goe},
1647
                                   ParameterVector{A, B, C, D});
1648 1649 1650

    auto backend = runtime::Backend::create("CPU");

1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662
    shared_ptr<runtime::Tensor> a = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> b = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> c = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> d = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> r1 = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> r2 = backend->create_tensor(element::f32, shapeA);
    shared_ptr<runtime::Tensor> r3 = backend->create_tensor(element::f32, shapeA);

    vector<float> dataA{1, 4, 1, 4};
    vector<float> dataB{3, 3, 3, 9};
    vector<float> dataC{1, 2, 3, 4};
    vector<float> dataD{-2, 2, -1, 1};
1663 1664 1665 1666 1667
    copy_data(a, dataA);
    copy_data(b, dataB);
    copy_data(c, dataC);
    copy_data(d, dataD);

1668
    auto handle = backend->compile(f);
1669
    handle->call_with_validate({r1, r2, r3}, {a, b, c, d});
1670

1671 1672 1673
    vector<float> expected1{5, 11, 5, 17};
    vector<float> expected2{2, 7, 5, 14};
    vector<float> expected3{-3, -3, -3, -9};
1674 1675 1676
    EXPECT_TRUE(test::all_close_f(read_vector<float>(r1), expected1, MIN_FLOAT_TOLERANCE_BITS));
    EXPECT_TRUE(test::all_close_f(read_vector<float>(r2), expected2, MIN_FLOAT_TOLERANCE_BITS));
    EXPECT_TRUE(test::all_close_f(read_vector<float>(r3), expected3, MIN_FLOAT_TOLERANCE_BITS));
1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704
}

TEST(cpu_fusion, loop_kernel_copy_with_new_args)
{
    Shape shapeA{2, 2};
    auto A = make_shared<op::Parameter>(element::i32, shapeA);
    auto B = make_shared<op::Parameter>(element::i32, shapeA);
    auto C = make_shared<op::Parameter>(element::i32, shapeA);
    auto D = make_shared<op::Parameter>(element::i32, shapeA);

    auto neg_a = make_shared<op::Negative>(A);
    auto neg_b = make_shared<op::Negative>(B);
    auto add_ab = neg_a + neg_b;
    auto add_cd = C + B;
    auto add_cd_abs = make_shared<op::Abs>(add_cd);
    auto add_ab_abs = make_shared<op::Abs>(add_ab);
    auto add_aab = add_ab_abs + A;
    auto add_cdd = add_cd_abs + D;

    auto lk = make_shared<runtime::cpu::op::LoopKernel>(
        NodeVector{neg_a, neg_b, add_ab, add_cd, add_cd_abs, add_ab_abs, add_aab, add_cdd},
        NodeVector{add_aab, add_cdd, neg_b},
        NodeVector{A, B, C, D});
    auto add_aab_goe = std::make_shared<op::GetOutputElement>(lk, 0);
    auto add_cdd_goe = std::make_shared<op::GetOutputElement>(lk, 1);
    auto neg_b_goe = std::make_shared<op::GetOutputElement>(lk, 2);

    auto f = make_shared<Function>(NodeVector{add_aab_goe, add_cdd_goe, neg_b_goe},
1705
                                   ParameterVector{A, B, C, D});
1706 1707 1708 1709 1710

    auto copy_f = clone_function(*f);

    auto backend = runtime::Backend::create("CPU");

1711 1712 1713 1714 1715 1716 1717 1718 1719 1720
    shared_ptr<runtime::Tensor> a = backend->create_tensor(element::i32, shapeA);
    shared_ptr<runtime::Tensor> b = backend->create_tensor(element::i32, shapeA);
    shared_ptr<runtime::Tensor> c = backend->create_tensor(element::i32, shapeA);
    shared_ptr<runtime::Tensor> d = backend->create_tensor(element::i32, shapeA);
    shared_ptr<runtime::Tensor> r1 = backend->create_tensor(element::i32, shapeA);
    shared_ptr<runtime::Tensor> r2 = backend->create_tensor(element::i32, shapeA);
    shared_ptr<runtime::Tensor> r3 = backend->create_tensor(element::i32, shapeA);
    shared_ptr<runtime::Tensor> copy_r1 = backend->create_tensor(element::i32, shapeA);
    shared_ptr<runtime::Tensor> copy_r2 = backend->create_tensor(element::i32, shapeA);
    shared_ptr<runtime::Tensor> copy_r3 = backend->create_tensor(element::i32, shapeA);
1721 1722 1723 1724 1725 1726 1727 1728 1729 1730

    vector<int> dataA{1, 4, 1, 4};
    vector<int> dataB{3, 3, 3, 9};
    vector<int> dataC{1, 2, 3, 4};
    vector<int> dataD{-2, 2, -1, 1};
    copy_data(a, dataA);
    copy_data(b, dataB);
    copy_data(c, dataC);
    copy_data(d, dataD);

1731
    auto handle = backend->compile(f);
1732 1733 1734
    handle->call_with_validate({r1, r2, r3}, {a, b, c, d});
    auto h1 = backend->compile(copy_f);
    h1->call_with_validate({copy_r1, copy_r2, copy_r3}, {a, b, c, d});
1735 1736 1737 1738 1739

    EXPECT_EQ(read_vector<int>(r1), read_vector<int>(copy_r1));
    EXPECT_EQ(read_vector<int>(r2), read_vector<int>(copy_r2));
    EXPECT_EQ(read_vector<int>(r3), read_vector<int>(copy_r3));
}
1740 1741 1742

#endif

1743 1744 1745 1746 1747 1748 1749 1750
static std::shared_ptr<ngraph::Function> make_forward_function()
{
    Shape shape_a{10, 3, 28, 28};
    auto input = std::make_shared<op::Parameter>(element::f32, shape_a);
    Shape window_shape{2, 2};
    auto max_pool = std::make_shared<op::MaxPool>(input, window_shape);
    auto neg = std::make_shared<op::Negative>(max_pool);
    auto absn = std::make_shared<op::Abs>(max_pool);
1751
    return std::make_shared<Function>(NodeVector{max_pool, neg, absn}, ParameterVector{input});
1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847
}

static std::pair<std::shared_ptr<ngraph::Function>, std::vector<std::shared_ptr<ngraph::Node>>>
    make_backward_function(std::shared_ptr<ngraph::Function> f)
{
    // get parameters
    std::vector<std::shared_ptr<ngraph::op::Parameter>> back_parameters = f->get_parameters();

    ngraph::NodeVector adjoints;
    ngraph::NodeVector outputs;
    for (auto Y : f->get_results())
    {
        // Get the output
        // Create the Adjoint
        auto C = std::make_shared<ngraph::op::Parameter>(Y->get_element_type(), Y->get_shape());
        outputs.push_back(Y);
        adjoints.push_back(C);
    }

    ngraph::autodiff::Adjoints adjoint{outputs, adjoints};

    // Perform autodiff
    std::vector<std::shared_ptr<Node>> dYdXs(back_parameters.size());
    transform(back_parameters.begin(),
              back_parameters.end(),
              dYdXs.begin(),
              [&adjoint](const std::shared_ptr<Node>& X) { return adjoint.backprop_node(X); });

    // create the backward function
    std::vector<std::shared_ptr<ngraph::op::Parameter>> param_adjoints;
    for (auto n : adjoints)
        param_adjoints.push_back(std::dynamic_pointer_cast<ngraph::op::Parameter>(n));
    back_parameters.insert(back_parameters.begin(), param_adjoints.begin(), param_adjoints.end());

    return {std::make_shared<ngraph::Function>(dYdXs, back_parameters), adjoints};
}

void optimize_graph(std::shared_ptr<ngraph::Function>& f, std::shared_ptr<ngraph::Function> bf)
{
    // start by removing excess reshapes
    NodeVector nv_cwi;
    ngraph::pass::Manager pass_manager;
    pass_manager.register_pass<ngraph::pass::ReshapeElimination>();
    pass_manager.register_pass<ngraph::pass::ReshapeElimination>();
    pass_manager.register_pass<runtime::cpu::pass::CPUWorkspaceInsertion>(nv_cwi);
    pass_manager.register_pass<pass::VisualizeTree>("before.fprop_cache.pdf");

    pass_manager.run_passes(f);
    pass_manager.run_passes(bf);
    if (nv_cwi.size() > 0)
    {
        NodeVector new_outputs;
        for (auto r : f->get_results())
        {
            new_outputs.push_back(r->get_argument(0));
        }

        new_outputs.insert(new_outputs.end(), nv_cwi.begin(), nv_cwi.end());
        f = std::make_shared<ngraph::Function>(new_outputs, f->get_parameters());
    }

    ngraph::NodeVector dYdXs;
    for (size_t i = 0; i < bf->get_output_size(); ++i)
    {
        dYdXs.push_back(bf->get_output_op(i)->get_argument(0));
    }

    ngraph::NodeVector combined_outputs;
    for (auto r : f->get_results())
    {
        combined_outputs.push_back(r->get_argument(0));
    }

    combined_outputs.insert(combined_outputs.end(), dYdXs.begin(), dYdXs.end());

    std::vector<std::shared_ptr<ngraph::op::Parameter>> combined_parameters = f->get_parameters();
    std::vector<std::shared_ptr<ngraph::op::Parameter>> back_parameters = bf->get_parameters();

    combined_parameters.insert(
        combined_parameters.end(), back_parameters.begin(), back_parameters.end());
    auto combinedf = std::make_shared<ngraph::Function>(combined_outputs, combined_parameters);
    // rerun Reshape elimination to help simplify the graph again, run CPUFusion
    // this replaces nodes in both f and bf due to shared-ptr - ness
    ngraph::pass::Manager pass_manager_comb;
    pass_manager_comb.register_pass<ngraph::pass::ReshapeElimination>();
    pass_manager_comb.register_pass<ngraph::runtime::cpu::pass::CPUFusion>();
    pass_manager_comb.run_passes(combinedf);
}

TEST(cpu_fusion, maxpool_with_indices_in_mxnet)
{
    auto f = make_forward_function();
    auto bfa = make_backward_function(f);
    auto maybe_bf = bfa.first;
    auto adjoints = bfa.second;
    optimize_graph(f, maybe_bf);
1848
    auto fprop_cache = ngraph::cache_fprop(f, maybe_bf);
1849 1850 1851 1852 1853 1854

    auto mpwi_bprop = fprop_cache.bprop->get_results().at(0)->get_argument(0);
    ASSERT_TRUE(std::dynamic_pointer_cast<op::Parameter>(mpwi_bprop->get_argument(0)));
    ASSERT_TRUE(std::dynamic_pointer_cast<op::Parameter>(mpwi_bprop->get_argument(2)));
}

1855
TEST(cpu_fusion, conv_batch_norm_folding)
1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869
{
    Shape shape_input{1, 8, 3, 3};
    Shape shape_weights{2, 8, 1, 1};
    Shape shape_norm{2};

    auto make_function = [shape_input, shape_weights, shape_norm]() {
        auto input = std::make_shared<op::Parameter>(element::f32, shape_input);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);
        double eps = 0.001;
        auto gamma = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto beta = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto mean = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto var = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto conv = std::make_shared<op::Convolution>(input, weights, Strides{1, 1}, Strides{1, 1});
1870
        auto bn = std::make_shared<op::BatchNormInference>(conv, gamma, beta, mean, var, eps);
1871
        auto f = make_shared<Function>(NodeVector{bn},
1872
                                       ParameterVector{input, weights, gamma, beta, mean, var});
1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912
        return f;
    };

    auto int_f = make_function();
    auto cpu_f = make_function();

    vector<vector<float>> args{
        {1.25f,  2.25f, 5.25f, 6.25f,  -1.25f, -1.25f, 3.25f, -4.25f, 7.25f,  8.25f,  -1.25f,
         -1.25f, 1.25f, 2.25f, -3.25f, 2.25f,  4.25f,  4.25f, 1.25f,  2.25f,  -4.25f, 2.25f,
         4.25f,  4.25f, 0.f,   0.f,    -1.f,   0.f,    2.f,   2.f,    0.f,    0.f,    0.f,
         0.f,    2.f,   2.f,   1.25f,  2.25f,  5.25f,  6.25f, 1.25f,  1.25f,  3.25f,  4.25f,
         -7.25f, 8.25f, 1.25f, -1.25f, -1.25f, 2.25f,  3.25f, 2.25f,  -4.25f, -4.25f, -1.25f,
         -2.25f, 4.25f, 2.25f, 4.25f,  4.25f,  0.f,    0.f,   1.f,    0.f,    -2.f,   2.f,
         0.f,    0.f,   0.f,   0.f,    -2.f,   -2.f},
        {1.25f,
         2.25f,
         5.25f,
         6.25f,
         -1.25f,
         -1.25f,
         3.25f,
         -4.25f,
         7.25f,
         8.25f,
         -1.25f,
         0.f,
         0.f,
         0.f,
         0.f,
         -2.f},
        {-0.9384f, 0.01875f},
        {11.0f, 1.3f},
        {0.12f, 0.31f},
        {0.01f, 0.11f},
    };

    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0)));
}
1913

1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931
TEST(cpu_fusion, convbias_batch_norm_folding)
{
    Shape shape_input{2, 8, 5, 5};
    Shape shape_weights{2, 8, 2, 2};
    Shape shape_norm{2};

    auto make_function = [shape_input, shape_weights, shape_norm]() {
        auto input = std::make_shared<op::Parameter>(element::f32, shape_input);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto bias = std::make_shared<op::Parameter>(element::f32, Shape{2});
        double eps = 1.01;
        auto gamma = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto beta = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto mean = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto var = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto conv = std::make_shared<op::Convolution>(input, weights, Strides{1, 1}, Strides{1, 1});
        auto convbias =
            conv + std::make_shared<op::Broadcast>(bias, conv->get_shape(), AxisSet{0, 2, 3});
1932
        auto bn = std::make_shared<op::BatchNormInference>(convbias, gamma, beta, mean, var, eps);
1933
        auto f = make_shared<Function>(
1934
            NodeVector{bn}, ParameterVector{input, weights, bias, gamma, beta, mean, var});
1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955
        return f;
    };

    auto int_f = make_function();
    auto cpu_f = make_function();

    test::Uniform<float> rng(1.0f, 100.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : cpu_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0)));
}

TEST(cpu_fusion, conv_affine_folding)
1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971
{
    Shape shape_input{1, 8, 3, 3};
    Shape shape_weights{2, 8, 1, 1};
    Shape shape_norm{2};

    auto make_function = [shape_input, shape_weights, shape_norm]() {
        auto input = std::make_shared<op::Parameter>(element::f32, shape_input);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);

        auto a = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto b = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto conv = std::make_shared<op::Convolution>(input, weights, Strides{1, 1}, Strides{1, 1});
        auto out = std::make_shared<op::Add>(
            std::make_shared<op::Multiply>(
                conv, std::make_shared<op::Broadcast>(a, conv->get_shape(), AxisSet{0, 2, 3})),
            std::make_shared<op::Broadcast>(b, conv->get_shape(), AxisSet{0, 2, 3}));
1972
        auto f = make_shared<Function>(NodeVector{out}, ParameterVector{input, weights, a, b});
1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
        return f;
    };

    auto int_f = make_function();
    auto cpu_f = make_function();

    vector<vector<float>> args{
        {1.25f,  2.25f, 5.25f, 6.25f,  -1.25f, -1.25f, 3.25f, -4.25f, 7.25f,  8.25f,  -1.25f,
         -1.25f, 1.25f, 2.25f, -3.25f, 2.25f,  4.25f,  4.25f, 1.25f,  2.25f,  -4.25f, 2.25f,
         4.25f,  4.25f, 0.f,   0.f,    -1.f,   0.f,    2.f,   2.f,    0.f,    0.f,    0.f,
         0.f,    2.f,   2.f,   1.25f,  2.25f,  5.25f,  6.25f, 1.25f,  1.25f,  3.25f,  4.25f,
         -7.25f, 8.25f, 1.25f, -1.25f, -1.25f, 2.25f,  3.25f, 2.25f,  -4.25f, -4.25f, -1.25f,
         -2.25f, 4.25f, 2.25f, 4.25f,  4.25f,  0.f,    0.f,   1.f,    0.f,    -2.f,   2.f,
         0.f,    0.f,   0.f,   0.f,    -2.f,   -2.f},
        {1.25f,
         2.25f,
         5.25f,
         6.25f,
         -1.25f,
         -1.25f,
         3.25f,
         -4.25f,
         7.25f,
         8.25f,
         -1.25f,
         0.f,
         0.f,
         0.f,
         0.f,
         -2.f},
        {-0.9384f, 0.01875f},
        {11.0f, 1.3f},
    };

    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0)));
}

2012
TEST(cpu_fusion, convbias_affine_folding1)
2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032
{
    Shape shape_input{1, 6, 3, 3};
    Shape shape_weights{3, 6, 1, 1};
    Shape shape_norm{3};

    auto make_function = [shape_input, shape_weights, shape_norm]() {
        auto input = std::make_shared<op::Parameter>(element::f32, shape_input);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto bias = std::make_shared<op::Parameter>(element::f32, Shape{3});

        auto a = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto b = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto conv = std::make_shared<op::Convolution>(input, weights, Strides{1, 1}, Strides{1, 1});
        auto convbias =
            conv + std::make_shared<op::Broadcast>(bias, conv->get_shape(), AxisSet{0, 2, 3});
        auto out = std::make_shared<op::Add>(
            std::make_shared<op::Multiply>(
                convbias, std::make_shared<op::Broadcast>(a, conv->get_shape(), AxisSet{0, 2, 3})),
            std::make_shared<op::Broadcast>(b, conv->get_shape(), AxisSet{0, 2, 3}));
        auto f =
2033
            make_shared<Function>(NodeVector{out}, ParameterVector{input, weights, bias, a, b});
2034 2035 2036
        return f;
    };

2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090
    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    auto func = make_function();
    pass_manager.run_passes(func);
    ASSERT_EQ(count_ops_of_type<op::ConvolutionBiasAdd>(func), 1);

    auto int_f = make_function();
    auto cpu_f = make_function();

    test::Uniform<float> rng(20.0f, 300.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : cpu_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0)));
}

TEST(cpu_fusion, convbias_affine_folding2)
{
    Shape shape_input{1, 6, 3, 3};
    Shape shape_weights{3, 6, 1, 1};
    Shape shape_norm{1};

    auto make_function = [shape_input, shape_weights, shape_norm]() {
        auto input = std::make_shared<op::Parameter>(element::f32, shape_input);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto bias = std::make_shared<op::Parameter>(element::f32, Shape{3});

        auto a = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto b = std::make_shared<op::Parameter>(element::f32, shape_norm);
        auto conv = std::make_shared<op::Convolution>(input, weights, Strides{1, 1}, Strides{1, 1});
        auto convbias =
            conv + std::make_shared<op::Broadcast>(bias, conv->get_shape(), AxisSet{0, 2, 3});
        auto out = std::make_shared<op::Add>(
            std::make_shared<op::Multiply>(
                convbias, std::make_shared<op::Broadcast>(a, conv->get_shape(), AxisSet{1, 2, 3})),
            std::make_shared<op::Broadcast>(b, conv->get_shape(), AxisSet{1, 2, 3}));
        auto f =
            make_shared<Function>(NodeVector{out}, ParameterVector{input, weights, bias, a, b});
        return f;
    };

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    auto func = make_function();
    pass_manager.run_passes(func);
    ASSERT_EQ(count_ops_of_type<op::ConvolutionBiasAdd>(func), 1);

2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107
    auto int_f = make_function();
    auto cpu_f = make_function();

    test::Uniform<float> rng(20.0f, 300.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : cpu_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0)));
}

2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140
TEST(cpu_fusion, group_convolution_fusion)
{
    Shape shape_a{1, 32, 2, 2};
    auto A = make_shared<op::Parameter>(element::f32, shape_a);
    Shape shape_b{2, 16, 1, 1};
    auto B = make_shared<op::Parameter>(element::f32, shape_b);
    Shape shape_r{1, 2, 2, 2};

    auto a_slice0 = std::make_shared<op::Slice>(A, Coordinate{0, 0, 0, 0}, Coordinate{1, 16, 2, 2});
    auto a_slice1 =
        std::make_shared<op::Slice>(A, Coordinate{0, 16, 0, 0}, Coordinate{1, 32, 2, 2});

    auto b_slice0 = std::make_shared<op::Slice>(B, Coordinate{0, 0, 0, 0}, Coordinate{1, 16, 1, 1});
    auto b_slice1 = std::make_shared<op::Slice>(B, Coordinate{1, 0, 0, 0}, Coordinate{2, 16, 1, 1});

    auto conv_lower = make_shared<op::Convolution>(a_slice0,
                                                   b_slice0,
                                                   Strides{1, 1},
                                                   Strides{1, 1},
                                                   CoordinateDiff{0, 0},
                                                   CoordinateDiff{0, 0},
                                                   Strides{1, 1});

    auto conv_upper = make_shared<op::Convolution>(a_slice1,
                                                   b_slice1,
                                                   Strides{1, 1},
                                                   Strides{1, 1},
                                                   CoordinateDiff{0, 0},
                                                   CoordinateDiff{0, 0},
                                                   Strides{1, 1});

    auto concat = make_shared<op::Concat>(NodeVector{conv_lower, conv_upper}, 1);

2141
    auto f = make_shared<Function>(NodeVector{concat}, ParameterVector{A, B});
2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195
    pass::Manager pass_manager;
    pass_manager.register_pass<pass::VisualizeTree>("before_group.pdf");
    pass_manager.register_pass<runtime::cpu::pass::CPUBatchFusion>();
    pass_manager.register_pass<pass::VisualizeTree>("after_group.pdf");
    pass_manager.run_passes(f);
    auto gc =
        std::dynamic_pointer_cast<op::GroupConvolution>(f->get_results().at(0)->get_argument(0));
    ASSERT_TRUE(gc);
}

TEST(cpu_fusion, group_convolution)
{
    auto backend = runtime::Backend::create("CPU");
    test::Uniform<float> rng(2.0f, 10.0f);

    const size_t GROUPS = 2;
    Shape shape_a{1, 32, 2, 2};
    auto A = make_shared<op::Parameter>(element::f32, shape_a);
    Shape shape_b{2, 16, 1, 1};
    auto B = make_shared<op::Parameter>(element::f32, shape_b);
    Shape shape_r{1, 2, 2, 2};
    auto group_conv = make_shared<op::GroupConvolution>(A,
                                                        B,
                                                        Strides{1, 1},
                                                        Strides{1, 1},
                                                        CoordinateDiff{0, 0},
                                                        CoordinateDiff{0, 0},
                                                        Strides{1, 1},
                                                        GROUPS,
                                                        shape_r);

    Shape shape_c{1, 16, 2, 2};
    auto C = make_shared<op::Parameter>(element::f32, shape_c);
    Shape shape_d{1, 16, 1, 1};
    auto D = make_shared<op::Parameter>(element::f32, shape_d);
    auto conv_lower = make_shared<op::Convolution>(C,
                                                   D,
                                                   Strides{1, 1},
                                                   Strides{1, 1},
                                                   CoordinateDiff{0, 0},
                                                   CoordinateDiff{0, 0},
                                                   Strides{1, 1});

    auto E = make_shared<op::Parameter>(element::f32, shape_c);
    auto F = make_shared<op::Parameter>(element::f32, shape_d);
    auto conv_upper = make_shared<op::Convolution>(E,
                                                   F,
                                                   Strides{1, 1},
                                                   Strides{1, 1},
                                                   CoordinateDiff{0, 0},
                                                   CoordinateDiff{0, 0},
                                                   Strides{1, 1});

    auto f = make_shared<Function>(NodeVector{group_conv, conv_lower, conv_upper},
2196
                                   ParameterVector{A, B, C, D, E, F});
2197 2198 2199 2200 2201 2202 2203 2204 2205 2206

    auto a_ = rng.initialize(backend->create_tensor(element::f32, shape_a));
    auto b_ = rng.initialize(backend->create_tensor(element::f32, shape_b));

    vector<float> rv(shape_size(shape_r), 0);
    auto group_result = std::dynamic_pointer_cast<ngraph::runtime::cpu::CPUTensorView>(
        backend->create_tensor(element::f32, shape_r, rv.data()));

    auto av = read_vector<float>(a_);
    auto bv = read_vector<float>(b_);
2207 2208
    auto c_ = backend->create_tensor(element::f32, shape_c, av.data()); // lower data
    auto d_ = backend->create_tensor(element::f32, shape_d, bv.data()); // upper data
2209 2210

    auto e_ =
2211
        backend->create_tensor(element::f32, shape_c, av.data() + av.size() / 2); // lower weights
2212
    auto f_ =
2213
        backend->create_tensor(element::f32, shape_d, bv.data() + bv.size() / 2); // upper weights
2214 2215

    Shape shape_ur{1, 1, 2, 2};
2216
    // allocate a contigious storage for both lower and upper halves.
2217 2218 2219 2220 2221
    vector<float> erv(shape_size(shape_r), 0);
    auto lower_result = std::dynamic_pointer_cast<ngraph::runtime::cpu::CPUTensorView>(
        backend->create_tensor(element::f32, shape_ur, erv.data()));
    auto upper_result = std::dynamic_pointer_cast<ngraph::runtime::cpu::CPUTensorView>(
        backend->create_tensor(element::f32, shape_ur, erv.data() + erv.size() / 2));
2222 2223 2224
    auto handle = backend->compile(f);
    handle->call_with_validate({group_result, lower_result, upper_result},
                               {a_, b_, c_, d_, e_, f_});
2225
    EXPECT_TRUE(test::all_close_f(rv, erv));
2226 2227
}

2228 2229 2230 2231
TEST(cpu_fusion, rnn_fprop_1_lstm_cell)
{
    auto src_layer = make_shared<op::Parameter>(element::f32, Shape{10, 100});
    auto src_iter = make_shared<op::Parameter>(element::f32, Shape{20, 100});
2232 2233
    auto weights_layer = make_shared<op::Parameter>(element::f32, Shape{100, 400});
    auto weights_iter = make_shared<op::Parameter>(element::f32, Shape{100, 400});
2234 2235 2236 2237 2238 2239 2240
    auto biases = make_shared<op::Parameter>(element::f32, Shape{400});
    const int number_of_timesteps = 1;
    const int number_of_gates_per_cell = 4;
    const int src_seq_length = 1;
    const int num_rnn_cell_states = 2;
    const int rnn_direction = 1;
    const int num_of_rnn_fused_layer = 1;
Pruthvi's avatar
Pruthvi committed
2241 2242 2243
    ngraph::runtime::cpu::rnn_utils::rnntype rnn_type =
        ngraph::runtime::cpu::rnn_utils::rnntype::vanilla_lstm;

2244 2245 2246 2247 2248 2249 2250 2251 2252 2253
    auto rnn_node = make_shared<op::Rnn>(src_layer,
                                         src_iter,
                                         weights_layer,
                                         weights_iter,
                                         biases,
                                         number_of_timesteps,
                                         number_of_gates_per_cell,
                                         src_seq_length,
                                         num_rnn_cell_states,
                                         rnn_direction,
Pruthvi's avatar
Pruthvi committed
2254 2255 2256
                                         num_of_rnn_fused_layer,
                                         rnn_type);

2257 2258 2259 2260 2261
    auto rnn_ht_output = make_shared<op::GetOutputElement>(rnn_node, 0);
    auto rnn_ct_output = make_shared<op::GetOutputElement>(rnn_node, 1);

    auto func = make_shared<Function>(
        NodeVector{rnn_ht_output, rnn_ct_output},
2262
        ParameterVector{src_layer, src_iter, weights_layer, weights_iter, biases});
2263 2264
    auto backend = runtime::Backend::create("CPU");

2265
    shared_ptr<runtime::Tensor> src_layer_t =
2266
        backend->create_tensor(element::f32, src_layer->get_shape());
2267
    shared_ptr<runtime::Tensor> src_iter_t =
2268
        backend->create_tensor(element::f32, src_iter->get_shape());
2269
    shared_ptr<runtime::Tensor> weights_layer_t =
2270
        backend->create_tensor(element::f32, weights_layer->get_shape());
2271
    shared_ptr<runtime::Tensor> weights_iter_t =
2272
        backend->create_tensor(element::f32, weights_iter->get_shape());
2273
    shared_ptr<runtime::Tensor> biases_t =
2274
        backend->create_tensor(element::f32, biases->get_shape());
2275
    shared_ptr<runtime::Tensor> result_ht = backend->create_tensor(element::f32, {10, 100});
2276
    shared_ptr<runtime::Tensor> result_ct = backend->create_tensor(element::f32, Shape{20, 100});
2277 2278 2279 2280 2281 2282 2283

    copy_data(src_layer_t, vector<float>(1000, 1));
    copy_data(src_iter_t, vector<float>(2000, 1));
    copy_data(weights_layer_t, vector<float>(400 * 100, 1));
    copy_data(weights_iter_t, vector<float>(400 * 100, 1));
    copy_data(biases_t, vector<float>(400, 1));

2284 2285
    auto handle = backend->compile(func);
    handle->call_with_validate(
2286 2287
        {result_ht, result_ct},
        {src_layer_t, src_iter_t, weights_layer_t, weights_iter_t, biases_t});
2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362
    vector<float> expected_ht(10 * 100, 0.964028f);
    vector<float> expected_ct;
    for (size_t i = 0; i < 20 * 100; i++)
    {
        if (i < 1000)
        {
            expected_ct.push_back(0.964028f);
        }
        else
        {
            expected_ct.push_back(2.0f);
        }
    }

    EXPECT_TRUE(test::all_close(expected_ht, read_vector<float>(result_ht)));
    EXPECT_TRUE(test::all_close(expected_ct, read_vector<float>(result_ct)));
}

TEST(cpu_fusion, fuse_lstm_cells)
{
    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::LSTMFusion>();
    const string json_path =
        file_util::path_join(SERIALIZED_ZOO, "mxnet/2rnn_layer_3lstm_cell.json");
    const string json_string = file_util::read_file_to_string(json_path);
    stringstream ss(json_string);
    shared_ptr<Function> func = ngraph::deserialize(ss);
    pass_manager.run_passes(func);
    auto lstm_ops = get_ops_of_type<op::Lstm>(func);
    EXPECT_EQ(lstm_ops.size(), 6);
}

TEST(cpu_fusion, fuse_2_layer_rnn)
{
    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::LSTMFusion>();
    pass_manager.register_pass<runtime::cpu::pass::RNNFusion>();
    const string json_path =
        file_util::path_join(SERIALIZED_ZOO, "mxnet/2rnn_layer_3lstm_cell.json");
    const string json_string = file_util::read_file_to_string(json_path);
    stringstream ss(json_string);
    shared_ptr<Function> func = ngraph::deserialize(ss);
    pass_manager.run_passes(func);
    size_t count = count_ops_of_type<op::Rnn>(func);
    auto rnn_ops = get_ops_of_type<op::Rnn>(func);
    EXPECT_EQ(rnn_ops.size(), count);
    for (auto& node : rnn_ops)
    {
        EXPECT_EQ(node->get_num_timesteps(), node->get_src_sequence_length());
        EXPECT_EQ(node->get_num_cell_states(), node->get_argument(1)->get_arguments().size());
    }
}

TEST(cpu_fusion, fuse_1_layer_rnn)
{
    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::LSTMFusion>();
    pass_manager.register_pass<runtime::cpu::pass::RNNFusion>();
    const string json_path =
        file_util::path_join(SERIALIZED_ZOO, "mxnet/1rnn_layer_3lstm_cell.json");
    const string json_string = file_util::read_file_to_string(json_path);
    stringstream ss(json_string);
    shared_ptr<Function> func = ngraph::deserialize(ss);
    pass_manager.run_passes(func);
    size_t count = count_ops_of_type<op::Rnn>(func);
    auto rnn_ops = get_ops_of_type<op::Rnn>(func);
    EXPECT_EQ(rnn_ops.size(), 1);
    EXPECT_EQ(rnn_ops.size(), count);
    for (auto& node : rnn_ops)
    {
        EXPECT_EQ(node->get_num_timesteps(), node->get_src_sequence_length());
        EXPECT_EQ(node->get_num_cell_states(), node->get_argument(1)->get_arguments().size());
    }
}

2363
TEST(cpu_fusion, rnn_fusion_1lstm_cell)
2364 2365
{
    const std::string file_name("mxnet/1_lstm_cell_forward.json");
2366 2367
    auto cpu_f = make_function_from_file(file_name);
    auto int_f = make_function_from_file(file_name);
2368
    test::Uniform<float> rng(-1.0f, 1.0f);
2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384
    vector<vector<float>> args;

    for (shared_ptr<op::Parameter> param : int_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
    }
}

2385
TEST(cpu_fusion, rnn_fusion_1rnn_layer_3lstm_cell)
2386 2387
{
    const std::string file_name("mxnet/1rnn_layer_3lstm_cell.json");
2388 2389
    auto cpu_f = make_function_from_file(file_name);
    auto int_f = make_function_from_file(file_name);
2390
    test::Uniform<float> rng(-1.0f, 1.0f);
2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406
    vector<vector<float>> args;

    for (shared_ptr<op::Parameter> param : int_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
    }
}

2407
TEST(cpu_fusion, rnn_fusion_2rnn_layer_3lstm_cell)
2408 2409
{
    const std::string file_name("mxnet/2rnn_layer_3lstm_cell.json");
2410 2411
    auto cpu_f = make_function_from_file(file_name);
    auto int_f = make_function_from_file(file_name);
2412
    test::Uniform<float> rng(-1.0f, 1.0f);
2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427
    vector<vector<float>> args;

    for (shared_ptr<op::Parameter> param : int_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
    }
}
2428

2429 2430
#if 0

2431
TEST(cpu_fusion, loop_kernel_fusion_multiple_groups_pruned)
2432
{
2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449
    auto make_function = []() -> std::shared_ptr<Function> {
        Shape shape{};
        auto a = make_shared<op::Parameter>(element::f32, shape);
        auto b = make_shared<op::Parameter>(element::f32, shape);
        auto c = make_shared<op::Parameter>(element::f32, shape);
        auto add_ab = a + b;
        auto add_abs = std::make_shared<op::Abs>(add_ab);
        auto abs_neg = std::make_shared<op::Negative>(add_abs);
        auto sub_c_neg = c - abs_neg;

        auto d = make_shared<op::Parameter>(element::f32, shape);
        auto d_abs = std::make_shared<op::Abs>(d);
        auto add_d = d_abs + add_ab;
        auto neg_d = std::make_shared<op::Negative>(add_d);

        auto mul_cd = neg_d * sub_c_neg;
        auto f =
2450
            std::make_shared<Function>(ngraph::NodeVector{mul_cd}, ParameterVector{a, b, c, d});
2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466

        return f;
    };

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPULoopKernelFusion>(3);
    auto cpu_f = make_function();
    auto int_f = make_function();
    pass_manager.run_passes(cpu_f);
    test::Uniform<float> rng(-100.0f, 100.0f);
    vector<vector<float>> args;

    size_t lkn = count_ops_of_type<runtime::cpu::op::LoopKernel>(cpu_f);
    ASSERT_GT(lkn, 0);

    for (shared_ptr<op::Parameter> param : cpu_f->get_parameters())
2467
    {
2468 2469 2470
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
2471
    }
2472 2473 2474 2475 2476 2477 2478
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
    }
}
2479

2480
TEST(cpu_fusion, loop_kernel_fusion_bounded_relu)
2481
{
2482 2483 2484 2485 2486 2487 2488 2489 2490 2491
    auto make_function = []() -> std::shared_ptr<Function> {
        Shape shape{};
        auto a = make_shared<op::Parameter>(element::f32, shape);
        auto relu = make_shared<op::Relu>(a);
        auto upper_bound =
            op::Constant::create<float>(element::f32, shape, std::vector<float>{6.0f});
        auto minn = make_shared<op::Minimum>(relu, upper_bound);
        auto absn = make_shared<op::Abs>(minn);
        auto negn = std::make_shared<op::Negative>(absn);

2492
        auto f = std::make_shared<Function>(ngraph::NodeVector{negn}, ParameterVector{a});
2493

2494 2495
        return f;
    };
2496

2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508
    pass::Manager pass_manager;
    pass_manager.register_pass<pass::VisualizeTree>("before_relu_fusion.pdf");
    pass_manager.register_pass<runtime::cpu::pass::CPULoopKernelFusion>(3);
    pass_manager.register_pass<pass::VisualizeTree>("after_relu_fusion.pdf");
    auto cpu_f = make_function();
    auto int_f = make_function();
    pass_manager.run_passes(cpu_f);
    test::Uniform<float> rng(-100.0f, 100.0f);
    vector<vector<float>> args;

    size_t lkn = count_ops_of_type<runtime::cpu::op::LoopKernel>(cpu_f);
    ASSERT_GT(lkn, 0);
2509

2510
    for (shared_ptr<op::Parameter> param : cpu_f->get_parameters())
2511
    {
2512 2513 2514
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
2515
    }
2516 2517 2518
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
2519
    {
2520
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
2521
    }
2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542
}

TEST(cpu_fusion, loop_kernel_fusion_multiple_groups)
{
    auto make_function = []() -> std::shared_ptr<Function> {
        Shape shape{};
        auto a = make_shared<op::Parameter>(element::f32, shape);
        auto b = make_shared<op::Parameter>(element::f32, shape);
        auto c = make_shared<op::Parameter>(element::f32, shape);
        auto add_ab = a + b;
        auto add_abs = std::make_shared<op::Abs>(add_ab);
        auto abs_neg = std::make_shared<op::Negative>(add_abs);
        auto sub_c_neg = c - abs_neg;

        auto d = make_shared<op::Parameter>(element::f32, shape);
        auto d_abs = std::make_shared<op::Abs>(d);
        auto add_d = d_abs + add_ab;
        auto neg_d = std::make_shared<op::Negative>(add_d);

        auto mul_cd = neg_d * sub_c_neg;
        auto f =
2543
            std::make_shared<Function>(ngraph::NodeVector{mul_cd}, ParameterVector{a, b, c, d});
2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557

        return f;
    };

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPULoopKernelFusion>(2);
    auto cpu_f = make_function();
    auto int_f = make_function();
    pass_manager.run_passes(cpu_f);
    test::Uniform<float> rng(-100.0f, 100.0f);
    vector<vector<float>> args;

    size_t lkn = count_ops_of_type<runtime::cpu::op::LoopKernel>(cpu_f);
    ASSERT_GT(lkn, 0);
2558

2559
    for (shared_ptr<op::Parameter> param : cpu_f->get_parameters())
2560
    {
2561 2562 2563
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
2564
    }
2565 2566 2567
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
2568
    {
2569
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
2570 2571 2572
    }
}

2573
TEST(cpu_fusion, loop_kernel_fusion_one_group)
2574
{
2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591
    auto make_function = []() -> std::shared_ptr<Function> {
        Shape shape{};
        auto a = make_shared<op::Parameter>(element::f32, shape);
        auto b = make_shared<op::Parameter>(element::f32, shape);
        auto c = make_shared<op::Parameter>(element::f32, shape);
        auto add_ab = a + b;
        auto add_abs = std::make_shared<op::Abs>(add_ab);
        auto abs_neg = std::make_shared<op::Negative>(add_abs);
        auto sub_c_neg = c - abs_neg;
        auto d = make_shared<op::Parameter>(element::f32, shape);
        auto add_d = sub_c_neg + d;
        auto abs_add_d = std::make_shared<op::Abs>(add_d);
        auto e = make_shared<op::Parameter>(element::f32, shape);
        auto add_e = e + abs_add_d;
        auto neg_e = std::make_shared<op::Negative>(add_e);

        auto f = std::make_shared<Function>(ngraph::NodeVector{neg_e},
2592
                                            ParameterVector{a, b, c, d, e});
2593 2594 2595 2596 2597

        return f;

    };

2598
    pass::Manager pass_manager;
2599 2600 2601 2602 2603 2604
    pass_manager.register_pass<runtime::cpu::pass::CPULoopKernelFusion>(2);
    auto cpu_f = make_function();
    auto int_f = make_function();
    pass_manager.run_passes(cpu_f);
    test::Uniform<float> rng(-100.0f, 100.0f);
    vector<vector<float>> args;
2605

2606 2607
    size_t lkn = count_ops_of_type<runtime::cpu::op::LoopKernel>(cpu_f);
    ASSERT_GT(lkn, 0);
2608

2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620
    for (shared_ptr<op::Parameter> param : cpu_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
    }
2621 2622
}

2623 2624
#endif

2625 2626 2627
TEST(cpu_fusion, sigmoid_multiply_fusion)
{
    pass::Manager pass_manager;
2628
    pass_manager.register_pass<pass::CoreFusion>();
2629 2630 2631 2632 2633 2634 2635 2636 2637 2638
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/3_lstm_cell_forward.json");
    const string json_string = file_util::read_file_to_string(json_path);
    stringstream ss(json_string);
    shared_ptr<Function> func = ngraph::deserialize(ss);
    pass_manager.run_passes(func);
    size_t ccg = count_ops_of_type<op::SigmoidMultiply>(func);
    ASSERT_EQ(ccg, 18);
}

2639
void sigmoid_multiply_fusion_forward_compute(runtime::Backend* backend,
2640
                                             const ParameterVector& input_params,
2641 2642 2643 2644 2645 2646 2647
                                             const vector<vector<float>>& input_data,
                                             const vector<Shape>& input_shapes,
                                             const Shape& result_shape,
                                             shared_ptr<Node> input_0_node,
                                             shared_ptr<Node> input_1_node,
                                             const vector<float>& expected)
{
2648
    shared_ptr<runtime::Tensor> result_tensor = backend->create_tensor(element::f32, result_shape);
2649

2650
    vector<shared_ptr<runtime::Tensor>> input_tensors;
2651 2652 2653 2654 2655 2656 2657 2658
    for (int i = 0; i < input_params.size(); ++i)
    {
        input_tensors.push_back(backend->create_tensor(element::f32, input_shapes[i]));
        copy_data(input_tensors[i], input_data[i]);
    }

    auto mul_node = input_0_node * input_1_node;
    auto func = make_shared<Function>(mul_node, input_params);
2659
    auto handle = backend->compile(func);
2660
    handle->call_with_validate({result_tensor}, input_tensors);
2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680
    EXPECT_TRUE(test::all_close(read_vector<float>(result_tensor), expected));
}

TEST(cpu_fusion, sigmoid_multiply_fusion_forward)
{
    auto backend = runtime::Backend::create("CPU");

    Shape data_shape{1, 1, 2, 2};
    Shape const_shape{1};

    vector<float> input_0_data{1.f, 2.f, 3.f, 4.f};
    vector<float> input_1_data{1.2f, 2.3f, 3.5f, 4.7f};
    vector<float> const_data{1.2f};
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_2_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto sigmoid_0 = make_shared<op::Sigmoid>(input_0_param);
        auto sigmoid_1 = make_shared<op::Add>(input_1_param, input_2_param);
        vector<float> expected{1.60833f, 3.78743f, 6.19173f, 8.54352f};
2681
        ParameterVector input_params{input_0_param, input_1_param, input_2_param};
2682 2683
        vector<vector<float>> input_data{input_0_data, input_0_data, input_1_data};
        vector<Shape> input_shapes{data_shape, data_shape, data_shape};
2684
        sigmoid_multiply_fusion_forward_compute(backend.get(),
2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698
                                                input_params,
                                                input_data,
                                                input_shapes,
                                                data_shape,
                                                sigmoid_0,
                                                sigmoid_1,
                                                expected);
    }
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, const_shape);
        auto sigmoid_0 = make_shared<op::Broadcast>(input_1_param, data_shape, AxisSet{1, 2, 3});
        auto sigmoid_1 = make_shared<op::Sigmoid>(input_0_param);
        vector<float> expected{0.87727f, 1.05696f, 1.14309f, 1.17842f};
2699
        ParameterVector input_params{input_0_param, input_1_param};
2700 2701
        vector<vector<float>> input_data{input_0_data, const_data};
        vector<Shape> input_shapes{data_shape, const_shape};
2702
        sigmoid_multiply_fusion_forward_compute(backend.get(),
2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716
                                                input_params,
                                                input_data,
                                                input_shapes,
                                                data_shape,
                                                sigmoid_0,
                                                sigmoid_1,
                                                expected);
    }
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, const_shape);
        auto sigmoid_0 = make_shared<op::Sigmoid>(input_0_param);
        auto sigmoid_1 = make_shared<op::Broadcast>(input_1_param, data_shape, AxisSet{1, 2, 3});
        vector<float> expected{0.87727f, 1.05696f, 1.14309f, 1.17842f};
2717
        ParameterVector input_params{input_0_param, input_1_param};
2718 2719
        vector<vector<float>> input_data{input_0_data, const_data};
        vector<Shape> input_shapes{data_shape, const_shape};
2720
        sigmoid_multiply_fusion_forward_compute(backend.get(),
2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734
                                                input_params,
                                                input_data,
                                                input_shapes,
                                                data_shape,
                                                sigmoid_0,
                                                sigmoid_1,
                                                expected);
    }
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto sigmoid_0 = make_shared<op::Sigmoid>(input_0_param);
        auto sigmoid_1 = make_shared<op::Sigmoid>(input_1_param);
        vector<float> expected{0.561837f, 0.800536f, 0.924652f, 0.973163f};
2735
        ParameterVector input_params{input_0_param, input_1_param};
2736 2737
        vector<vector<float>> input_data{input_0_data, input_1_data};
        vector<Shape> input_shapes{data_shape, data_shape};
2738
        sigmoid_multiply_fusion_forward_compute(backend.get(),
2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752
                                                input_params,
                                                input_data,
                                                input_shapes,
                                                data_shape,
                                                sigmoid_0,
                                                sigmoid_1,
                                                expected);
    }
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto sigmoid_0 = make_shared<op::Sigmoid>(input_0_param);
        auto sigmoid_1 = make_shared<op::Tanh>(input_1_param);
        vector<float> expected{0.60945f, 0.863266f, 0.950838f, 0.981851f};
2753
        ParameterVector input_params{input_0_param, input_1_param};
2754 2755
        vector<vector<float>> input_data{input_0_data, input_1_data};
        vector<Shape> input_shapes{data_shape, data_shape};
2756
        sigmoid_multiply_fusion_forward_compute(backend.get(),
2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770
                                                input_params,
                                                input_data,
                                                input_shapes,
                                                data_shape,
                                                sigmoid_0,
                                                sigmoid_1,
                                                expected);
    }
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto sigmoid_0 = make_shared<op::Tanh>(input_0_param);
        auto sigmoid_1 = make_shared<op::Sigmoid>(input_1_param);
        vector<float> expected{0.585304f, 0.876182f, 0.965887f, 0.990322f};
2771
        ParameterVector input_params{input_0_param, input_1_param};
2772 2773
        vector<vector<float>> input_data{input_0_data, input_1_data};
        vector<Shape> input_shapes{data_shape, data_shape};
2774
        sigmoid_multiply_fusion_forward_compute(backend.get(),
2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788
                                                input_params,
                                                input_data,
                                                input_shapes,
                                                data_shape,
                                                sigmoid_0,
                                                sigmoid_1,
                                                expected);
    }
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto sigmoid_0 = make_shared<op::Tanh>(input_0_param);
        auto sigmoid_1 = make_shared<op::Tanh>(input_1_param);
        vector<float> expected{0.634907f, 0.94484f, 0.993242f, 0.999164f};
2789
        ParameterVector input_params{input_0_param, input_1_param};
2790 2791
        vector<vector<float>> input_data{input_0_data, input_1_data};
        vector<Shape> input_shapes{data_shape, data_shape};
2792
        sigmoid_multiply_fusion_forward_compute(backend.get(),
2793 2794 2795 2796 2797 2798 2799 2800 2801 2802
                                                input_params,
                                                input_data,
                                                input_shapes,
                                                data_shape,
                                                sigmoid_0,
                                                sigmoid_1,
                                                expected);
    }
}

2803
void sigmoid_multiply_fusion_backward_compute(runtime::Backend* backend,
2804
                                              const ParameterVector& input_params,
2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817
                                              const vector<vector<float>>& input_data,
                                              const vector<Shape>& input_shapes,
                                              const vector<float> delta_data,
                                              const Shape& delta_shape,
                                              const Shape& d_input_0_shape,
                                              const Shape& d_input_1_shape,
                                              shared_ptr<Node> input_0_node,
                                              shared_ptr<Node> input_1_node,
                                              shared_ptr<Node> input_0_adjoint,
                                              shared_ptr<Node> input_1_adjoint,
                                              const vector<float>& expected_0,
                                              const vector<float>& expected_1)
{
2818
    vector<shared_ptr<runtime::Tensor>> input_tensors;
2819 2820 2821 2822 2823 2824 2825
    for (int i = 0; i < input_params.size(); ++i)
    {
        input_tensors.push_back(backend->create_tensor(element::f32, input_shapes[i]));
        copy_data(input_tensors[i], input_data[i]);
    }

    auto delta_param = make_shared<op::Parameter>(element::f32, delta_shape);
2826
    shared_ptr<runtime::Tensor> delta_tensor = backend->create_tensor(element::f32, delta_shape);
2827 2828
    copy_data(delta_tensor, delta_data);

2829
    ParameterVector back_params(input_params);
2830 2831 2832
    back_params.push_back(delta_param);
    input_tensors.push_back(delta_tensor);

2833
    shared_ptr<runtime::Tensor> d_input_0_tensor =
2834
        backend->create_tensor(element::f32, d_input_0_shape);
2835
    shared_ptr<runtime::Tensor> d_input_1_tensor =
2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853
        backend->create_tensor(element::f32, d_input_1_shape);

    using FunctionType = op::SigmoidMultiply::FunctionType;
    auto input_0_type = op::SigmoidMultiply::identify_node_type(input_0_node);
    auto input_1_type = op::SigmoidMultiply::identify_node_type(input_1_node);
    // for Identity functions, we use the node itself, otherwise use its input
    // where we will apply the function of input node
    auto input_0_alt =
        (input_0_type == FunctionType::Identity) ? input_0_node : input_0_node->get_argument(0);
    auto input_1_alt =
        (input_1_type == FunctionType::Identity) ? input_1_node : input_1_node->get_argument(0);
    auto sigmoid_mul =
        make_shared<op::SigmoidMultiply>(input_0_alt, input_1_alt, input_0_type, input_1_type);

    ngraph::autodiff::Adjoints adjoints(NodeVector{sigmoid_mul}, NodeVector{delta_param});
    auto d_input_0 = adjoints.backprop_node(input_0_adjoint);
    auto d_input_1 = adjoints.backprop_node(input_1_adjoint);
    auto df = make_shared<Function>(NodeVector{d_input_0, d_input_1}, back_params);
2854 2855
    auto handle = backend->compile(df);
    handle->call_with_validate({d_input_0_tensor, d_input_1_tensor}, input_tensors);
2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879
    EXPECT_TRUE(test::all_close(read_vector<float>(d_input_0_tensor), expected_0));
    EXPECT_TRUE(test::all_close(read_vector<float>(d_input_1_tensor), expected_1));
}

TEST(cpu_fusion, sigmoid_multiply_fusion_backward)
{
    auto backend = runtime::Backend::create("CPU");

    Shape data_shape{1, 1, 2, 2};
    Shape const_shape{1};

    vector<float> input_0_data{1.f, 2.f, 3.f, 4.f};
    vector<float> input_1_data{1.2f, 2.2f, 3.2f, 4.2f};
    vector<float> const_data{1.2f};
    vector<float> delta_data(shape_size(data_shape), 20.0f);

    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_2_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto sigmoid_0 = make_shared<op::Sigmoid>(input_0_param);
        auto sigmoid_1 = make_shared<op::Add>(input_1_param, input_2_param);
        vector<float> expected_0{8.65093f, 8.81946f, 5.60191f, 2.89668f};
        vector<float> expected_1{14.6212f, 17.6159f, 19.0515f, 19.6403f};
2880
        ParameterVector input_params{input_0_param, input_1_param, input_2_param};
2881 2882
        vector<vector<float>> input_data{input_0_data, input_0_data, input_1_data};
        vector<Shape> input_shapes{data_shape, data_shape, data_shape};
2883
        sigmoid_multiply_fusion_backward_compute(backend.get(),
2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904
                                                 input_params,
                                                 input_data,
                                                 input_shapes,
                                                 delta_data,
                                                 data_shape,
                                                 data_shape,
                                                 data_shape,
                                                 sigmoid_0,
                                                 sigmoid_1,
                                                 input_0_param,
                                                 sigmoid_1,
                                                 expected_0,
                                                 expected_1);
    }
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, const_shape);
        auto sigmoid_0 = make_shared<op::Broadcast>(input_1_param, data_shape, AxisSet{1, 2, 3});
        auto sigmoid_1 = make_shared<op::Tanh>(input_0_param);
        vector<float> expected_0{15.2319f, 19.2806f, 19.9011f, 19.9866f};
        vector<float> expected_1{10.0794f, 1.69562f, 0.236785f, 0.0321828f};
2905
        ParameterVector input_params{input_0_param, input_1_param};
2906 2907
        vector<vector<float>> input_data{input_0_data, const_data};
        vector<Shape> input_shapes{data_shape, const_shape};
2908
        sigmoid_multiply_fusion_backward_compute(backend.get(),
2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929
                                                 input_params,
                                                 input_data,
                                                 input_shapes,
                                                 delta_data,
                                                 data_shape,
                                                 data_shape,
                                                 data_shape,
                                                 sigmoid_0,
                                                 sigmoid_1,
                                                 sigmoid_0,
                                                 input_0_param,
                                                 expected_0,
                                                 expected_1);
    }
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, const_shape);
        auto sigmoid_0 = make_shared<op::Tanh>(input_0_param);
        auto sigmoid_1 = make_shared<op::Broadcast>(input_1_param, data_shape, AxisSet{1, 2, 3});
        vector<float> expected_0{10.0794f, 1.69562f, 0.236785f, 0.0321828f};
        vector<float> expected_1{15.2319f, 19.2806f, 19.9011f, 19.9866f};
2930
        ParameterVector input_params{input_0_param, input_1_param};
2931 2932
        vector<vector<float>> input_data{input_0_data, const_data};
        vector<Shape> input_shapes{data_shape, const_shape};
2933
        sigmoid_multiply_fusion_backward_compute(backend.get(),
2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954
                                                 input_params,
                                                 input_data,
                                                 input_shapes,
                                                 delta_data,
                                                 data_shape,
                                                 data_shape,
                                                 data_shape,
                                                 sigmoid_0,
                                                 sigmoid_1,
                                                 input_0_param,
                                                 sigmoid_1,
                                                 expected_0,
                                                 expected_1);
    }
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto sigmoid_0 = make_shared<op::Sigmoid>(input_0_param);
        auto sigmoid_1 = make_shared<op::Sigmoid>(input_1_param);
        vector<float> expected_0{3.02202f, 1.89041f, 0.868146f, 0.348035f};
        vector<float> expected_1{2.60102f, 1.58192f, 0.716941f, 0.285879f};
2955
        ParameterVector input_params{input_0_param, input_1_param};
2956 2957
        vector<vector<float>> input_data{input_0_data, input_1_data};
        vector<Shape> input_shapes{data_shape, data_shape};
2958
        sigmoid_multiply_fusion_backward_compute(backend.get(),
2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979
                                                 input_params,
                                                 input_data,
                                                 input_shapes,
                                                 delta_data,
                                                 data_shape,
                                                 data_shape,
                                                 data_shape,
                                                 sigmoid_0,
                                                 sigmoid_1,
                                                 input_0_param,
                                                 input_1_param,
                                                 expected_0,
                                                 expected_1);
    }
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto sigmoid_0 = make_shared<op::Sigmoid>(input_0_param);
        auto sigmoid_1 = make_shared<op::Tanh>(input_1_param);
        vector<float> expected_0{3.27813f, 2.04894f, 0.900536f, 0.353095f};
        vector<float> expected_1{4.45975f, 0.84425f, 0.126201f, 0.0176579f};
2980
        ParameterVector input_params{input_0_param, input_1_param};
2981 2982
        vector<vector<float>> input_data{input_0_data, input_1_data};
        vector<Shape> input_shapes{data_shape, data_shape};
2983
        sigmoid_multiply_fusion_backward_compute(backend.get(),
2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004
                                                 input_params,
                                                 input_data,
                                                 input_shapes,
                                                 delta_data,
                                                 data_shape,
                                                 data_shape,
                                                 data_shape,
                                                 sigmoid_0,
                                                 sigmoid_1,
                                                 input_0_param,
                                                 input_1_param,
                                                 expected_0,
                                                 expected_1);
    }
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto sigmoid_0 = make_shared<op::Tanh>(input_0_param);
        auto sigmoid_1 = make_shared<op::Sigmoid>(input_1_param);
        vector<float> expected_0{6.45521f, 1.27207f, 0.189593f, 0.0264228f};
        vector<float> expected_1{2.70967f, 1.7314f, 0.748913f, 0.29092f};
3005
        ParameterVector input_params{input_0_param, input_1_param};
3006 3007
        vector<vector<float>> input_data{input_0_data, input_1_data};
        vector<Shape> input_shapes{data_shape, data_shape};
3008
        sigmoid_multiply_fusion_backward_compute(backend.get(),
3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029
                                                 input_params,
                                                 input_data,
                                                 input_shapes,
                                                 delta_data,
                                                 data_shape,
                                                 data_shape,
                                                 data_shape,
                                                 sigmoid_0,
                                                 sigmoid_1,
                                                 input_0_param,
                                                 input_1_param,
                                                 expected_0,
                                                 expected_1);
    }
    {
        auto input_0_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto input_1_param = make_shared<op::Parameter>(element::f32, data_shape);
        auto sigmoid_0 = make_shared<op::Tanh>(input_0_param);
        auto sigmoid_1 = make_shared<op::Tanh>(input_1_param);
        vector<float> expected_0{7.00227f, 1.37874f, 0.196666f, 0.026807f};
        vector<float> expected_1{4.64603f, 0.924027f, 0.131829f, 0.0179692f};
3030
        ParameterVector input_params{input_0_param, input_1_param};
3031 3032
        vector<vector<float>> input_data{input_0_data, input_1_data};
        vector<Shape> input_shapes{data_shape, data_shape};
3033
        sigmoid_multiply_fusion_backward_compute(backend.get(),
3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048
                                                 input_params,
                                                 input_data,
                                                 input_shapes,
                                                 delta_data,
                                                 data_shape,
                                                 data_shape,
                                                 data_shape,
                                                 sigmoid_0,
                                                 sigmoid_1,
                                                 input_0_param,
                                                 input_1_param,
                                                 expected_0,
                                                 expected_1);
    }
}
3049 3050 3051 3052

TEST(cpu_fusion, fuse_batch_dot)
{
    pass::Manager pass_manager;
3053
    pass_manager.register_pass<runtime::cpu::pass::CPUBatchFusion>();
3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065
    const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/batch_dot_3.json");
    const string json_string = file_util::read_file_to_string(json_path);
    stringstream ss(json_string);
    shared_ptr<Function> func = ngraph::deserialize(ss);
    pass_manager.run_passes(func);
    size_t ccg = count_ops_of_type<op::BatchDot>(func);
    ASSERT_EQ(ccg, 1);
}

TEST(cpu_fusion, fuse_batch_dot_forward)
{
    pass::Manager pass_manager;
3066
    pass_manager.register_pass<runtime::cpu::pass::CPUBatchFusion>();
3067 3068

    const std::string file_name("mxnet/batch_dot_3.json");
3069 3070
    auto cpu_f = make_function_from_file(file_name);
    auto int_f = make_function_from_file(file_name);
3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087
    pass_manager.run_passes(cpu_f);
    test::Uniform<float> rng(0.0f, 1.0f);
    vector<vector<float>> args;

    for (shared_ptr<op::Parameter> param : int_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < int_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
    }
}
3088

3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119
TEST(cpu_fusion, fuse_batch_dot_backward)
{
    const std::string file_name("mxnet/batch_dot_3.json");
    auto cpu_f = make_function_from_file(file_name);
    auto int_f = make_function_from_file(file_name);

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPUBatchFusion>();
    pass_manager.run_passes(cpu_f);

    auto int_df = autodiff::backprop_function(int_f);
    auto cpu_df = autodiff::backprop_function(cpu_f);

    test::Uniform<float> rng(-1.0f, 1.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : cpu_df->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    auto int_results = execute(int_df, args, "INTERPRETER");
    auto cpu_results = execute(cpu_df, args, "CPU");

    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
    }
}

3120
TEST(cpu_fusion, fuse_rnn_across_layer_2layer_3timestep)
3121
{
3122
    const std::string file_name("mxnet/2layer_3timestep_ic100oc100.json");
3123 3124
    auto cpu_f = make_function_from_file(file_name);
    auto int_f = make_function_from_file(file_name);
3125
    test::Uniform<float> rng(-1.0f, 1.0f);
3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136
    vector<vector<float>> args;

    for (shared_ptr<op::Parameter> param : int_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");

3137
    EXPECT_EQ(1, count_ops_of_type<op::Rnn>(cpu_f));
3138 3139
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
3140
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
3141 3142
    }
}
3143 3144 3145 3146 3147 3148 3149 3150 3151

static void check_bounded_relu(Shape param_shape, float constant_val)
{
    auto make_function = [](Shape input_shape, float alpha_val) {
        auto relu_input = std::make_shared<op::Parameter>(element::f32, input_shape);
        auto relu = std::make_shared<op::Relu>(relu_input);
        auto alpha = op::Constant::create<float>(
            element::f32, input_shape, std::vector<float>(1.0f, alpha_val));
        auto min = std::make_shared<op::Minimum>(relu, alpha);
3152
        auto f = make_shared<Function>(NodeVector{min}, ParameterVector{relu_input});
3153 3154 3155 3156 3157
        return f;
    };

    auto cpu_f = make_function(param_shape, constant_val);
    auto int_f = make_function(param_shape, constant_val);
3158
    test::Uniform<float> rng(-10.0f, 10.0f);
3159 3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 3174 3175 3176 3177 3178 3179
    vector<vector<float>> args;

    for (shared_ptr<op::Parameter> param : int_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");

    EXPECT_EQ(1, count_ops_of_type<op::BoundedRelu>(cpu_f));
    EXPECT_TRUE(test::all_close(cpu_results.at(0), int_results.at(0), 1.0e-4f, 1.0e-4f));
}

TEST(cpu_fusion, fuse_bounded_relu_inter_vs_cpu)
{
    check_bounded_relu(Shape{4, 3, 2, 2}, 6.0f);
    check_bounded_relu(Shape{4, 3}, 4.0f);
    check_bounded_relu(Shape{4, 3, 2}, 2.0f);
}
3180

3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191 3192 3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205 3206 3207 3208 3209 3210 3211 3212 3213 3214 3215 3216 3217 3218 3219
TEST(cpu_fusion, fuse_leaky_relu)
{
    auto make_function = [](Shape input_shape, vector<float> alpha_val) {
        auto input = std::make_shared<op::Parameter>(element::f32, input_shape);
        auto alpha = op::Constant::create<float>(element::f32, input_shape, alpha_val);
        auto out =
            std::make_shared<op::Maximum>(input, std::make_shared<op::Multiply>(input, alpha));
        auto f = make_shared<Function>(NodeVector{out}, ParameterVector{input});
        return f;
    };

    auto no_fuse1 = make_function(Shape{1, 2, 3}, std::vector<float>(6, -1.0f));
    auto no_fuse2 = make_function(Shape{1, 3}, std::vector<float>{1.4f, 1.2f, 1.4f});

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    pass_manager.run_passes(no_fuse1);
    pass_manager.run_passes(no_fuse2);
    EXPECT_EQ(0, count_ops_of_type<op::LeakyRelu>(no_fuse1));
    EXPECT_EQ(0, count_ops_of_type<op::LeakyRelu>(no_fuse2));

    // non-mkldnn kernel
    auto cpu_f1 = make_function(Shape{1, 2, 3}, std::vector<float>(6, 0.1f));
    // mkldnn kernel
    auto cpu_f2 = make_function(Shape{2, 3}, std::vector<float>(6, 0.1f));

    vector<vector<float>> args;
    args.push_back(std::vector<float>{-1, -2, 0, 1, 2, 3});
    std::vector<float> expected_result{-0.1f, -0.2f, 0.0f, 1.0f, 2.0f, 3.0f};

    auto cpu1_results = execute(cpu_f1, args, "CPU");
    EXPECT_EQ(1, count_ops_of_type<op::LeakyRelu>(cpu_f1));
    EXPECT_TRUE(test::all_close(cpu1_results.at(0), expected_result));

    auto cpu2_results = execute(cpu_f2, args, "CPU");
    EXPECT_EQ(1, count_ops_of_type<op::LeakyRelu>(cpu_f2));
    EXPECT_TRUE(test::all_close(cpu2_results.at(0), expected_result));
}

3220 3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232 3233 3234 3235 3236 3237 3238 3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252 3253 3254 3255 3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267 3268 3269 3270 3271 3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290 3291 3292 3293 3294 3295 3296 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 3308 3309 3310 3311 3312 3313 3314 3315 3316 3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394 3395 3396 3397 3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414
TEST(cpu_fusion, fuse_update_slice)
{
    auto make_function = [](bool fuse = true) {
        auto input = std::make_shared<op::Parameter>(element::f32, Shape{4, 32, 16});
        Shape lower_bounds{1, 0, 0};
        Shape upper_bounds{2, 32, 16};
        auto slice = std::make_shared<op::Slice>(
            input, fuse ? lower_bounds : Shape{3, 0, 0}, fuse ? upper_bounds : Shape{4, 32, 16});
        auto update = std::make_shared<op::Parameter>(element::f32, Shape{1, 32, 16});
        auto add = std::make_shared<op::Add>(slice, update);
        auto out = std::make_shared<op::ReplaceSlice>(input, add, lower_bounds, upper_bounds);
        auto f = make_shared<Function>(NodeVector{out}, ParameterVector{input, update});
        return f;
    };

    auto fuse = make_function(true);
    auto no_fuse = make_function(false);

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    pass_manager.run_passes(fuse);
    pass_manager.run_passes(no_fuse);
    EXPECT_EQ(1, count_ops_of_type<op::UpdateSlice>(fuse));
    EXPECT_EQ(0, count_ops_of_type<op::UpdateSlice>(no_fuse));

    auto int_f = make_function();
    auto cpu_f = make_function();

    test::Uniform<float> rng(0.0f, 1.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : int_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i)));
    }
}

TEST(cpu_fusion, fuse_update_slice_inplace)
{
    auto make_function = [](bool fuse = true) {
        auto input = std::make_shared<op::Parameter>(element::f32, Shape{4, 32, 16});
        auto abs = std::make_shared<op::Abs>(input);
        Shape lower_bounds{1, 0, 0};
        Shape upper_bounds{2, 32, 16};
        auto slice = std::make_shared<op::Slice>(abs, lower_bounds, upper_bounds);
        auto update = std::make_shared<op::Parameter>(element::f32, Shape{1, 32, 16});
        auto add = std::make_shared<op::Add>(slice, update);
        auto rs = std::make_shared<op::ReplaceSlice>(abs, add, lower_bounds, upper_bounds);
        auto out = std::make_shared<op::Abs>(rs);
        if (fuse)
        {
            return make_shared<Function>(NodeVector{out}, ParameterVector{input, update});
        }
        else
        {
            return make_shared<Function>(NodeVector{out, add}, ParameterVector{input, update});
        }
    };

    auto fuse = make_function(true);
    auto no_fuse = make_function(false);

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    pass_manager.run_passes(fuse);
    pass_manager.run_passes(no_fuse);
    EXPECT_EQ(1, count_ops_of_type<op::UpdateSlice>(fuse));
    EXPECT_EQ(0, count_ops_of_type<op::UpdateSlice>(no_fuse));

    auto int_f = make_function();
    auto cpu_f = make_function();

    test::Uniform<float> rng(0.0f, 1.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : int_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i)));
    }
}

TEST(cpu_fusion, fuse_update_slice_strided)
{
    auto make_function = [](bool fuse = true) {
        auto input = std::make_shared<op::Parameter>(element::f32, Shape{4, 32, 16});
        Shape lower_bounds{1, 0, 0};
        Shape upper_bounds{2, 32, 16};
        Strides strides{1, 2, 2};
        auto slice = std::make_shared<op::Slice>(input,
                                                 fuse ? lower_bounds : Shape{3, 0, 0},
                                                 fuse ? upper_bounds : Shape{4, 32, 16},
                                                 strides);
        auto update = std::make_shared<op::Parameter>(element::f32, Shape{1, 16, 8});
        auto add = std::make_shared<op::Add>(slice, update);
        auto out =
            std::make_shared<op::ReplaceSlice>(input, add, lower_bounds, upper_bounds, strides);
        auto f = make_shared<Function>(NodeVector{out}, ParameterVector{input, update});
        return f;
    };

    auto fuse = make_function(true);
    auto no_fuse = make_function(false);

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    pass_manager.run_passes(fuse);
    pass_manager.run_passes(no_fuse);
    EXPECT_EQ(1, count_ops_of_type<op::UpdateSlice>(fuse));
    EXPECT_EQ(0, count_ops_of_type<op::UpdateSlice>(no_fuse));

    auto int_f = make_function();
    auto cpu_f = make_function();

    test::Uniform<float> rng(0.0f, 1.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : int_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i)));
    }
}

TEST(cpu_fusion, fuse_update_slice_strided_inplace)
{
    auto make_function = [](bool fuse = true) {
        auto input = std::make_shared<op::Parameter>(element::f32, Shape{4, 32, 16});
        auto abs = std::make_shared<op::Abs>(input);
        Shape lower_bounds{1, 0, 0};
        Shape upper_bounds{2, 32, 16};
        Strides strides{1, 4, 2};
        auto slice = std::make_shared<op::Slice>(abs, lower_bounds, upper_bounds, strides);
        auto update = std::make_shared<op::Parameter>(element::f32, Shape{1, 8, 8});
        auto add = std::make_shared<op::Add>(slice, update);
        auto rs = std::make_shared<op::ReplaceSlice>(abs, add, lower_bounds, upper_bounds, strides);
        auto out = std::make_shared<op::Abs>(rs);
        if (fuse)
        {
            return make_shared<Function>(NodeVector{out}, ParameterVector{input, update});
        }
        else
        {
            return make_shared<Function>(NodeVector{out, add}, ParameterVector{input, update});
        }
    };

    auto fuse = make_function(true);
    auto no_fuse = make_function(false);

    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    pass_manager.run_passes(fuse);
    pass_manager.run_passes(no_fuse);
    EXPECT_EQ(1, count_ops_of_type<op::UpdateSlice>(fuse));
    EXPECT_EQ(0, count_ops_of_type<op::UpdateSlice>(no_fuse));

    auto int_f = make_function();
    auto cpu_f = make_function();

    test::Uniform<float> rng(0.0f, 1.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : int_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i)));
    }
}

3415 3416 3417 3418 3419 3420 3421 3422 3423
TEST(cpu_fusion, dot_batch_forward)
{
    const Shape shape_a{2, 3, 2};
    const Shape shape_b{2, 3};

    auto generate_func = [&shape_a, &shape_b]() -> shared_ptr<Function> {
        auto a = make_shared<op::Parameter>(element::f32, shape_a);
        auto b = make_shared<op::Parameter>(element::f32, shape_b);
        auto dot = make_shared<op::Dot>(a, b);
3424
        return make_shared<Function>(dot, ParameterVector{a, b});
3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441 3442 3443 3444
    };
    shared_ptr<Function> cpu_func = generate_func();
    shared_ptr<Function> int_func = generate_func();

    test::Uniform<float> rng(0.0f, 1.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : int_func->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    auto int_results = execute(int_func, args, "INTERPRETER");
    auto cpu_results = execute(cpu_func, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
    }
}
Pruthvi's avatar
Pruthvi committed
3445
static std::shared_ptr<Function>
3446
    create_rnn_input_linear_transformation_function(size_t num_timesteps, bool data_is_4d = false)
Pruthvi's avatar
Pruthvi committed
3447 3448 3449
{
    auto W = std::make_shared<op::Parameter>(element::f32, Shape{400, 50});
    auto bias = std::make_shared<op::Parameter>(element::f32, Shape{400});
3450
    ParameterVector params{W, bias};
Pruthvi's avatar
Pruthvi committed
3451
    auto create_graph = [&]() -> std::shared_ptr<Node> {
3452 3453 3454 3455

        auto data_param = (data_is_4d)
                              ? std::make_shared<op::Parameter>(element::f32, Shape{2, 5, 1, 50})
                              : std::make_shared<op::Parameter>(element::f32, Shape{10, 1, 50});
Pruthvi's avatar
Pruthvi committed
3456
        params.push_back(data_param);
3457
        auto reshape_axis_order = data_is_4d ? AxisVector{0, 1, 2, 3} : AxisVector{0, 1, 2};
Pruthvi's avatar
Pruthvi committed
3458
        auto data_param_reshape =
3459
            std::make_shared<op::Reshape>(data_param, reshape_axis_order, Shape{10, 50});
Pruthvi's avatar
Pruthvi committed
3460 3461 3462 3463 3464 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488
        auto W_reshape = std::make_shared<op::Reshape>(W, AxisVector{1, 0}, Shape{50, 400});
        auto dot = std::make_shared<op::Dot>(data_param_reshape, W_reshape);
        auto bias_broadcast = make_shared<op::Broadcast>(bias, dot->get_shape(), AxisSet{0});
        auto add_bias = std::make_shared<op::Add>(dot, bias_broadcast);
        return add_bias;

    };

    NodeVector graph_nodes;
    for (size_t i = 0; i < num_timesteps; i++)
    {
        graph_nodes.push_back(create_graph());
    }
    auto concat = std::make_shared<op::Concat>(graph_nodes, 0);
    return make_shared<Function>(NodeVector{concat}, params);
}

TEST(cpu_fusion, fuse_rnn_input_across_time_steps)
{
    auto func = create_rnn_input_linear_transformation_function(10);
    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPURnnMatFusion>();
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    pass_manager.run_passes(func);
    size_t ref_matmulbias_count = 1;
    auto matmulbias_count = count_ops_of_type<op::MatmulBias>(func);
    EXPECT_EQ(ref_matmulbias_count, matmulbias_count);
}

3489 3490 3491 3492 3493 3494 3495 3496 3497 3498 3499 3500
TEST(cpu_fusion, fuse_rnn_input_across_time_steps_4d_data)
{
    auto func = create_rnn_input_linear_transformation_function(10, true);
    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::CPURnnMatFusion>();
    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>();
    pass_manager.run_passes(func);
    size_t ref_matmulbias_count = 10; // no CPURnnMatFusion transformations
    auto matmulbias_count = count_ops_of_type<op::MatmulBias>(func);
    EXPECT_EQ(ref_matmulbias_count, matmulbias_count);
}

Pruthvi's avatar
Pruthvi committed
3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515 3516 3517 3518 3519 3520 3521
TEST(cpu_fusion, rnn_input_fusion_inter_vs_cpu)
{
    shared_ptr<Function> cpu_func = create_rnn_input_linear_transformation_function(10);
    shared_ptr<Function> int_func = create_rnn_input_linear_transformation_function(10);

    test::Uniform<float> rng(-10.0f, 10.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : int_func->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    auto int_results = execute(int_func, args, "INTERPRETER");
    auto cpu_results = execute(cpu_func, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
    }
}
3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537 3538 3539 3540 3541 3542 3543 3544

TEST(cpu_fusion, validate_fuse_gru_inputs)
{
    const std::string file_name("mxnet/gru_debug.json");
    auto cpu_func = make_function_from_file(file_name);
    auto int_func = make_function_from_file(file_name);

    test::Uniform<float> rng(-10.0f, 10.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : int_func->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    auto int_results = execute(int_func, args, "INTERPRETER");
    auto cpu_results = execute(cpu_func, args, "CPU");
    for (size_t i = 0; i < cpu_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
    }
}
3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601 3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623 3624 3625 3626 3627 3628 3629 3630 3631 3632 3633 3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648 3649 3650 3651 3652 3653 3654 3655 3656 3657 3658 3659 3660 3661 3662

TEST(cpu_quant_fusion, qconv_relu)
{
    auto make_function = []() {
        Shape shape_input{1, 2, 2, 2};
        Shape shape_weights{1, 2, 1, 1};
        auto input = std::make_shared<op::Parameter>(element::f32, shape_input);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto input_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto weights_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto output_scale = op::Constant::create(element::f32, Shape{}, {4.0f});
        auto int8_zero = op::Constant::create(element::i8, Shape{}, {0});
        auto uint8_zero = op::Constant::create(element::u8, Shape{}, {0});

        op::Quantize::RoundMode round_mode = op::Quantize::RoundMode::ROUND_NEAREST_TOWARD_EVEN;
        auto q_input = std::make_shared<op::Quantize>(
            input, input_scale, uint8_zero, element::u8, AxisSet{}, round_mode);
        auto q_weights = std::make_shared<op::Quantize>(
            weights, weights_scale, int8_zero, element::i8, AxisSet{}, round_mode);
        auto requant_scale = (input_scale * weights_scale) / output_scale;
        auto conv = std::make_shared<op::QuantizedConvolution>(q_input,
                                                               q_weights,
                                                               Strides{1, 1},
                                                               Strides{1, 1},
                                                               CoordinateDiff{0, 0},
                                                               CoordinateDiff{0, 0},
                                                               Strides{1, 1},
                                                               requant_scale);
        auto dq = std::make_shared<op::Dequantize>(
            conv, output_scale, int8_zero, element::f32, AxisSet{});
        auto relu = std::make_shared<op::Relu>(dq);
        auto q = std::make_shared<op::Quantize>(
            relu, output_scale, uint8_zero, element::u8, AxisSet{}, round_mode);
        auto q_f =
            std::make_shared<op::Dequantize>(q, output_scale, uint8_zero, element::f32, AxisSet{});
        return make_shared<Function>(NodeVector{q_f}, ParameterVector{input, weights});
    };

    auto cpu_f1 = make_function();
    auto cpu_f2 = make_function();

    test::Uniform<float> rng(2.0f, 2.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : cpu_f1->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:0", 1);
    auto cpu1_results = execute(cpu_f1, args, "CPU");
    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:1", 1);
    auto cpu2_results = execute(cpu_f2, args, "CPU");
    // Expected output - [2, 2, ...]
    EXPECT_TRUE(test::all_close(cpu1_results.at(0), cpu2_results.at(0)));
}

TEST(cpu_quant_fusion, qconvb_relu)
{
    auto make_function = []() {
        Shape shape_input{1, 2, 2, 2};
        Shape shape_weights{1, 2, 1, 1};
        auto input = std::make_shared<op::Parameter>(element::f32, shape_input);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto bias = std::make_shared<op::Parameter>(element::f32, Shape{shape_weights[0]});
        auto input_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto weights_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto output_scale = op::Constant::create(element::f32, Shape{}, {4.0f});
        auto int8_zero = op::Constant::create(element::i8, Shape{}, {0});
        auto int32_zero = op::Constant::create(element::i32, Shape{}, {0});
        auto uint8_zero = op::Constant::create(element::u8, Shape{}, {0});

        op::Quantize::RoundMode round_mode = op::Quantize::RoundMode::ROUND_NEAREST_TOWARD_EVEN;
        auto q_input = std::make_shared<op::Quantize>(
            input, input_scale, uint8_zero, element::u8, AxisSet{}, round_mode);
        auto q_weights = std::make_shared<op::Quantize>(
            weights, weights_scale, int8_zero, element::i8, AxisSet{}, round_mode);
        auto q_bias = std::make_shared<op::Quantize>(
            bias, input_scale * weights_scale, int32_zero, element::i32, AxisSet{}, round_mode);
        auto requant_scale = (input_scale * weights_scale) / output_scale;
        auto conv = std::make_shared<op::QuantizedConvolutionBias>(q_input,
                                                                   q_weights,
                                                                   bias,
                                                                   Strides{1, 1},
                                                                   Strides{1, 1},
                                                                   CoordinateDiff{0, 0},
                                                                   CoordinateDiff{0, 0},
                                                                   Strides{1, 1},
                                                                   requant_scale);
        auto dq = std::make_shared<op::Dequantize>(
            conv, output_scale, int8_zero, element::f32, AxisSet{});
        auto relu = std::make_shared<op::Relu>(dq);
        auto q = std::make_shared<op::Quantize>(
            relu, output_scale, uint8_zero, element::u8, AxisSet{}, round_mode);
        auto q_f =
            std::make_shared<op::Dequantize>(q, output_scale, uint8_zero, element::f32, AxisSet{});
        return make_shared<Function>(NodeVector{q_f}, ParameterVector{input, weights, bias});
    };

    auto cpu_f1 = make_function();
    auto cpu_f2 = make_function();

    test::Uniform<float> rng(2.0f, 2.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : cpu_f1->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:0", 1);
    auto cpu1_results = execute(cpu_f1, args, "CPU");
    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:1", 1);
    auto cpu2_results = execute(cpu_f2, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu1_results.at(0), cpu2_results.at(0)));
}

3663 3664 3665 3666 3667 3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698 3699 3700 3701 3702 3703 3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 3726 3727 3728 3729 3730 3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757 3758 3759 3760 3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790
TEST(cpu_quant_fusion, qavg_pool)
{
    auto make_function = []() {
        Shape shape_input{1, 2, 4, 4};
        auto input = std::make_shared<op::Parameter>(element::f32, shape_input);
        auto input_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto weights_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto int8_zero = op::Constant::create(element::i8, Shape{}, {0});
        auto uint8_zero = op::Constant::create(element::u8, Shape{}, {0});

        op::Quantize::RoundMode round_mode = op::Quantize::RoundMode::ROUND_NEAREST_TOWARD_EVEN;
        auto q_input = std::make_shared<op::Quantize>(
            input, input_scale, uint8_zero, element::u8, AxisSet{}, round_mode);
        auto dq = std::make_shared<op::Dequantize>(
            q_input, input_scale, uint8_zero, element::f32, AxisSet{});
        auto avg_pool = std::make_shared<op::AvgPool>(dq, Shape{2, 2});
        return make_shared<Function>(NodeVector{avg_pool}, ParameterVector{input});
    };

    auto cpu_f1 = make_function();
    auto cpu_f2 = make_function();

    test::Uniform<float> rng(4.0f, 4.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : cpu_f1->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:0", 1);
    auto cpu1_results = execute(cpu_f1, args, "CPU");
    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:1", 1);
    auto cpu2_results = execute(cpu_f2, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu1_results.at(0), cpu2_results.at(0)));
}

TEST(cpu_quant_fusion, qmax_pool)
{
    auto make_function = []() {
        Shape shape_input{1, 2, 4, 4};
        auto input = std::make_shared<op::Parameter>(element::f32, shape_input);
        auto input_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto weights_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto int8_zero = op::Constant::create(element::i8, Shape{}, {0});
        auto uint8_zero = op::Constant::create(element::u8, Shape{}, {0});

        op::Quantize::RoundMode round_mode = op::Quantize::RoundMode::ROUND_NEAREST_TOWARD_EVEN;
        auto q_input = std::make_shared<op::Quantize>(
            input, input_scale, uint8_zero, element::u8, AxisSet{}, round_mode);
        auto dq = std::make_shared<op::Dequantize>(
            q_input, input_scale, uint8_zero, element::f32, AxisSet{});
        auto maxpool = std::make_shared<op::MaxPool>(dq, Shape{2, 2});
        return make_shared<Function>(NodeVector{maxpool}, ParameterVector{input});
    };

    auto cpu_f1 = make_function();
    auto cpu_f2 = make_function();

    test::Uniform<float> rng(1.0f, 10.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : cpu_f1->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:0", 1);
    auto cpu1_results = execute(cpu_f1, args, "CPU");
    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:1", 1);
    auto cpu2_results = execute(cpu_f2, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu1_results.at(0), cpu2_results.at(0)));
}

TEST(cpu_quant_fusion, qconcat)
{
    auto make_function = []() {
        auto get_input_slice = [](std::shared_ptr<op::Parameter>& input) {
            auto input_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
            auto int8_zero = op::Constant::create(element::i8, Shape{}, {0});
            auto uint8_zero = op::Constant::create(element::u8, Shape{}, {0});

            op::Quantize::RoundMode round_mode = op::Quantize::RoundMode::ROUND_NEAREST_TOWARD_EVEN;
            auto q_input = std::make_shared<op::Quantize>(
                input, input_scale, uint8_zero, element::u8, AxisSet{}, round_mode);
            auto dq = std::make_shared<op::Dequantize>(
                q_input, input_scale, uint8_zero, element::f32, AxisSet{});
            return dq;
        };

        NodeVector concat_inputs, concats;
        ParameterVector inputs;
        Shape shape_input{1, 2, 4, 4};
        inputs.push_back(std::make_shared<op::Parameter>(element::f32, shape_input));
        concat_inputs.push_back(get_input_slice(inputs.back()));
        // Concat2  -- Concat7
        for (size_t i = 0; i < 6; i++)
        {
            inputs.push_back(std::make_shared<op::Parameter>(element::f32, shape_input));
            concat_inputs.push_back(get_input_slice(inputs.back()));
            concats.push_back(std::make_shared<op::Concat>(concat_inputs, 0));
        }
        return make_shared<Function>(concats, inputs);
    };

    auto cpu_f1 = make_function();
    auto cpu_f2 = make_function();

    test::Uniform<float> rng(2.0f, 2.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : cpu_f1->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:0", 1);
    auto cpu1_results = execute(cpu_f1, args, "CPU");
    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:1", 1);
    auto cpu2_results = execute(cpu_f2, args, "CPU");
    // Expect Concat2 -- Concat6 to be fused and not Concat7
    ASSERT_EQ(count_ops_of_type<op::QuantizedConcat>(cpu_f2), 5);
    EXPECT_TRUE(test::all_close(cpu1_results.at(0), cpu2_results.at(0)));
}

3791 3792 3793 3794 3795 3796 3797 3798 3799 3800 3801 3802 3803 3804 3805 3806 3807 3808 3809 3810 3811 3812 3813 3814 3815 3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942 3943 3944 3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957 3958 3959 3960 3961 3962 3963 3964 3965 3966 3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992 3993 3994 3995 3996 3997 3998 3999 4000
TEST(cpu_quant_fusion, dq_q)
{
    auto make_function = [](bool match_scales = true, bool match_et = true) {
        Shape shape_input{1, 2, 2};
        auto input = std::make_shared<op::Parameter>(element::i8, shape_input);
        auto dq_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto int8_zero = op::Constant::create(element::i8, Shape{}, {0});
        auto dq =
            std::make_shared<op::Dequantize>(input, dq_scale, int8_zero, element::f32, AxisSet{});
        float q_scalev = 2.0f;
        if (!match_scales)
        {
            q_scalev = 1.0f;
        }
        auto q_scale = op::Constant::create(element::f32, Shape{}, {q_scalev});
        op::Quantize::RoundMode round_mode = op::Quantize::RoundMode::ROUND_NEAREST_TOWARD_EVEN;
        if (match_et)
        {
            auto q = std::make_shared<op::Quantize>(
                dq, q_scale, int8_zero, element::i8, AxisSet{}, round_mode);
            return make_shared<Function>(NodeVector{q}, ParameterVector{input});
        }
        else
        {
            auto uint8_zero = op::Constant::create(element::u8, Shape{}, {0});
            auto q = std::make_shared<op::Quantize>(
                dq, q_scale, uint8_zero, element::u8, AxisSet{}, round_mode);
            return make_shared<Function>(NodeVector{q}, ParameterVector{input});
        }
    };

    auto cpu_f1 = make_function();
    auto cpu_f2 = make_function();

    vector<vector<int8_t>> args;
    args.push_back({-1, 2, 3, 4});

    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:0", 1);
    auto cpu1_results = execute(cpu_f1, args, "CPU");
    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:1", 1);
    auto cpu2_results = execute(cpu_f2, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu1_results.at(0), cpu2_results.at(0)));

    auto backend = runtime::Backend::create("CPU");
    auto fuse = make_function(true, true);
    auto no_fuse1 = make_function(false, true);
    auto no_fuse2 = make_function(true, false);
    backend->compile(fuse);
    backend->compile(no_fuse1);
    backend->compile(no_fuse2);
    ASSERT_EQ(count_ops_of_type<op::Quantize>(fuse), 0);
    ASSERT_EQ(count_ops_of_type<op::Quantize>(no_fuse1), 1);
    ASSERT_EQ(count_ops_of_type<op::Quantize>(no_fuse2), 1);
}

TEST(cpu_quant_fusion, qconvbsa)
{
    auto make_function = []() {
        Shape shape_input{1, 2, 2, 2};
        Shape shape_weights{1, 2, 1, 1};
        Shape shape_summand{1, 1, 2, 2};
        auto input = std::make_shared<op::Parameter>(element::f32, shape_input);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto bias = std::make_shared<op::Parameter>(element::f32, Shape{shape_weights[0]});
        auto summand = std::make_shared<op::Parameter>(element::f32, shape_summand);

        auto input_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto weights_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto output_scale = op::Constant::create(element::f32, Shape{}, {4.0f});
        auto summand_scale = op::Constant::create(element::f32, Shape{}, {2.0f});

        auto int8_zero = op::Constant::create(element::i8, Shape{}, {0});
        auto int32_zero = op::Constant::create(element::i32, Shape{}, {0});
        auto uint8_zero = op::Constant::create(element::u8, Shape{}, {0});

        op::Quantize::RoundMode round_mode = op::Quantize::RoundMode::ROUND_NEAREST_TOWARD_EVEN;
        auto q_input = std::make_shared<op::Quantize>(
            input, input_scale, uint8_zero, element::u8, AxisSet{}, round_mode);
        auto q_weights = std::make_shared<op::Quantize>(
            weights, weights_scale, int8_zero, element::i8, AxisSet{}, round_mode);
        auto q_bias = std::make_shared<op::Quantize>(
            bias, input_scale * weights_scale, int32_zero, element::i32, AxisSet{}, round_mode);
        auto q_summand = std::make_shared<op::Quantize>(
            summand, summand_scale, int8_zero, element::i8, AxisSet{}, round_mode);

        // Left Graph
        auto requant_scale = (input_scale * weights_scale) / output_scale;
        auto conv = std::make_shared<op::QuantizedConvolutionBias>(q_input,
                                                                   q_weights,
                                                                   bias,
                                                                   Strides{1, 1},
                                                                   Strides{1, 1},
                                                                   CoordinateDiff{0, 0},
                                                                   CoordinateDiff{0, 0},
                                                                   Strides{1, 1},
                                                                   requant_scale);
        auto dq_l = std::make_shared<op::Dequantize>(
            conv, output_scale, int8_zero, element::f32, AxisSet{});
        auto r_l = std::make_shared<op::Reshape>(dq_l, AxisVector{0, 1, 2, 3}, Shape{1, 2, 2});
        auto b_l = std::make_shared<op::Broadcast>(r_l, Shape{1, 1, 2, 2}, AxisSet{0});

        // Right Graph
        auto dq_r = std::make_shared<op::Dequantize>(
            q_summand, summand_scale, int8_zero, element::f32, AxisSet{});
        auto r_r = std::make_shared<op::Reshape>(dq_r, AxisVector{0, 1, 2, 3}, Shape{1, 2, 2});
        auto b_r = std::make_shared<op::Broadcast>(r_r, Shape{1, 1, 2, 2}, AxisSet{0});
        auto add = b_l + b_r;
        auto relu = std::make_shared<op::Relu>(add);
        return make_shared<Function>(NodeVector{relu},
                                     ParameterVector{input, weights, bias, summand});
    };

    auto cpu_f1 = make_function();
    auto cpu_f2 = make_function();

    test::Uniform<float> rng(4.0f, 4.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : cpu_f1->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    // Disable CPUQuantFusion
    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:0", 1);
    auto cpu1_results = execute(cpu_f1, args, "CPU");
    // Enable CPUQuantFusion
    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:1", 1);
    auto cpu2_results = execute(cpu_f2, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu1_results.at(0), cpu2_results.at(0)));
}

TEST(cpu_quant_fusion, qconvba)
{
    auto make_function = []() {
        Shape shape_input{1, 2, 2, 2};
        Shape shape_weights{1, 2, 1, 1};
        Shape shape_summand{1, 1, 2, 2};
        auto input = std::make_shared<op::Parameter>(element::f32, shape_input);
        auto weights = std::make_shared<op::Parameter>(element::f32, shape_weights);
        auto bias = std::make_shared<op::Parameter>(element::f32, Shape{shape_weights[0]});
        auto summand = std::make_shared<op::Parameter>(element::f32, shape_summand);

        auto input_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto weights_scale = op::Constant::create(element::f32, Shape{}, {2.0f});
        auto output_scale = op::Constant::create(element::f32, Shape{}, {4.0f});
        auto summand_scale = op::Constant::create(element::f32, Shape{}, {4.0f});

        auto int8_zero = op::Constant::create(element::i8, Shape{}, {0});
        auto int32_zero = op::Constant::create(element::i32, Shape{}, {0});
        auto uint8_zero = op::Constant::create(element::u8, Shape{}, {0});

        op::Quantize::RoundMode round_mode = op::Quantize::RoundMode::ROUND_NEAREST_TOWARD_EVEN;
        auto q_input = std::make_shared<op::Quantize>(
            input, input_scale, uint8_zero, element::u8, AxisSet{}, round_mode);
        auto q_weights = std::make_shared<op::Quantize>(
            weights, weights_scale, int8_zero, element::i8, AxisSet{}, round_mode);
        auto q_bias = std::make_shared<op::Quantize>(
            bias, input_scale * weights_scale, int32_zero, element::i32, AxisSet{}, round_mode);
        auto q_summand = std::make_shared<op::Quantize>(
            summand, summand_scale, uint8_zero, element::u8, AxisSet{}, round_mode);

        // Left Graph
        auto requant_scale = (input_scale * weights_scale) / output_scale;
        auto conv = std::make_shared<op::QuantizedConvolutionBias>(q_input,
                                                                   q_weights,
                                                                   bias,
                                                                   Strides{1, 1},
                                                                   Strides{1, 1},
                                                                   CoordinateDiff{0, 0},
                                                                   CoordinateDiff{0, 0},
                                                                   Strides{1, 1},
                                                                   requant_scale);
        auto dq_l = std::make_shared<op::Dequantize>(
            conv, output_scale, int8_zero, element::f32, AxisSet{});
        auto r_l = std::make_shared<op::Reshape>(dq_l, AxisVector{0, 1, 2, 3}, Shape{1, 2, 2});
        auto b_l = std::make_shared<op::Broadcast>(r_l, Shape{1, 1, 2, 2}, AxisSet{0});

        // Right Graph
        auto dq_r = std::make_shared<op::Dequantize>(
            q_summand, summand_scale, uint8_zero, element::f32, AxisSet{});
        auto r_r = std::make_shared<op::Reshape>(dq_r, AxisVector{0, 1, 2, 3}, Shape{1, 2, 2});
        auto b_r = std::make_shared<op::Broadcast>(r_r, Shape{1, 1, 2, 2}, AxisSet{0});
        auto add = b_l + b_r;
        auto relu = std::make_shared<op::Relu>(add);
        return make_shared<Function>(NodeVector{relu},
                                     ParameterVector{input, weights, bias, summand});
    };

    auto cpu_f1 = make_function();
    auto cpu_f2 = make_function();

    test::Uniform<float> rng(2.0f, 2.0f);
    vector<vector<float>> args;
    for (shared_ptr<op::Parameter> param : cpu_f1->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }

    // Disable CPUQuantFusion
    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:0", 1);
    auto cpu1_results = execute(cpu_f1, args, "CPU");
    // Enable CPUQuantFusion
    set_environment("NGRAPH_PASS_ENABLES", "CPUQuantFusion:1", 1);
    auto cpu2_results = execute(cpu_f2, args, "CPU");
    EXPECT_TRUE(test::all_close(cpu1_results.at(0), cpu2_results.at(0)));
}
Pruthvi's avatar
Pruthvi committed
4001 4002 4003 4004 4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043

TEST(cpu_fusion, fuse_bi_directional_rnn)
{
    pass::Manager pass_manager;
    pass_manager.register_pass<runtime::cpu::pass::LSTMFusion>();
    pass_manager.register_pass<runtime::cpu::pass::RNNFusion>();
    pass_manager.register_pass<ngraph::pass::AlgebraicSimplification>();
    pass_manager.register_pass<runtime::cpu::pass::MultiLayerRNNFusion>();
    pass_manager.register_pass<runtime::cpu::pass::BiDirectionalRnn>();
    const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/lstm_bi_directional.json");
    const string json_string = file_util::read_file_to_string(json_path);
    stringstream ss(json_string);
    shared_ptr<Function> func = ngraph::deserialize(ss);
    pass_manager.run_passes(func);
    // Bidirectional graph pass will folds the reverse seq
    auto rev_seq_ops = get_ops_of_type<op::Reverse>(func);
    auto rnn_ops = get_ops_of_type<op::Rnn>(func);
    EXPECT_EQ(rev_seq_ops.size(), 0);
    // fuse two bi-directional rnn layers in to one MKLDNN Op
    EXPECT_EQ(rnn_ops.size(), 1);
}

TEST(cpu_fusion, bi_rnn_interpreter_vs_cpu)
{
    const std::string file_name("mxnet/lstm_bi_directional.json");
    auto cpu_f = make_function_from_file(file_name);
    auto int_f = make_function_from_file(file_name);
    test::Uniform<float> rng(0.0f, 1.0f);
    vector<vector<float>> args;

    for (shared_ptr<op::Parameter> param : int_f->get_parameters())
    {
        vector<float> tensor_val(shape_size(param->get_shape()));
        rng.initialize(tensor_val);
        args.push_back(tensor_val);
    }
    auto int_results = execute(int_f, args, "INTERPRETER");
    auto cpu_results = execute(cpu_f, args, "CPU");
    for (size_t i = 0; i < int_results.size(); i++)
    {
        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
    }
}