1 /*
2 * Copyright (C) 2017 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "CompilationBuilder.h"
18 #include "ExecutionPlan.h"
19 #include "GraphDump.h"
20 #include "HalInterfaces.h"
21 #include "Manager.h"
22 #include "ModelBuilder.h"
23 #include "NeuralNetworks.h"
24 #include "NeuralNetworksOEM.h"
25 #include "NeuralNetworksWrapper.h"
26 #include "SampleDriver.h"
27 #include "Utils.h"
28 #include "ValidateHal.h"
29
30 #include <gtest/gtest.h>
31
32 #include <map>
33 #include <queue>
34
35 // Uncomment the following line to generate some debugging output that
36 // may be useful when analyzing failures:
37 //
38 // #define VERBOSE VERBOSE
39
40 // Uncomment the following line to generate DOT graphs.
41 //
42 // #define GRAPH GRAPH
43
44 // These tests do whitebox testing of the graph partitioning
45 // algorithm. It is "whitebox" in the sense that we're not evaluating
46 // whether a particular partitioning is legal, or "good enough"
47 // according to some metric, but whether it exactly matches the
48 // expected behavior of the current partitioning algorithm.
49 //
50 // A key part of the current partitioning algorithm is to determine
51 // which device among the available devices should be the one to
52 // execute a particular operation from the graph. This determination
53 // is made "locally" -- i.e., it does not depend on the graph
54 // topology, only on the properties of the operation in question.
55 // IDevice::getSupportedOperations() indicates which operations in a
56 // graph can be executed on a device, and IDevice::getCapabilities()
57 // indicates how "good" that device is for executing particular kinds
58 // of operations. For each operation, the partitioning algorithm
59 // picks the "best" device that is capable of executing that
60 // operation; if no device can do so, then the algorithm picks the
61 // cpu.
62 //
63 // As part of this testing approach, we want to make it easy to
64 // specify which operations in a test graph can be executed on which
65 // devices. We accomplish this with an abstraction: There are eight
66 // different kinds of operations (each of which has two inputs and one
67 // output), and when we instantiate a device for testing purposes, we
68 // specify what subset of those eight kinds of operations the device
69 // is able to execute.
70 //
71 // The eight kinds of operations are represented in the graph as ADD
72 // or MUL with a particular activation function -- two opcodes times
73 // four activation functions means eight available operation kinds.
74 // This is a low-level representation detail -- when we specify the
75 // behavior of the device or build a graph, we do so in terms of
76 // operation encodings 0..7.
77 //
78 // In order to determine whether or not a partitioning matches the
79 // expected partitioning, we check the number of partitions, check
80 // which device each partition targets, and compare each partition's
81 // subgraph, model inputs, model outputs, submodel inputs, and
82 // submodel outputs against what is expected. In order to perform
83 // that comparison, we build a model to compare against a partition's
84 // submodel and run a graph comparison algorithm on it. The graph
85 // comparison and the inputs and outputs comparisons are syntactic
86 // rather than semantic comparisons -- they don't allow for
87 // reorderings of inputs and outputs. Because of this, we need to
88 // know exactly how the partitioning algorithm orders inputs and
89 // outputs in order to construct the models and operand lists to
90 // compare against. Here are some relevant behaviors of the
91 // partitioning algorithm:
92 //
93 // - It builds a subgraph by walking operations in forward topological
94 // order, and adding each operation's input operands and output
95 // operands in index order (input followed by output) when that
96 // operation is added. (It does not add an input that has already
97 // been added.)
98 // - It finds model inputs, model outputs, and submodel inputs in
99 // the order the corresponding operands were added to the subgraph
100 // (see ExecutionStep methods getModelInputs(), getModelOutputs(),
101 // getTempsAsSubModelInputs(), getOutputsAsSubModelInputs()).
102 // - It finds temps as submodel outputs in numerical order of corresponding
103 // operand number in the original model (see ExecutionStep method
104 // getTempsAsSubModelOutputs()).
105 // - When it calls identifyInputsAndOutputs() on the submodel, it
106 // passes inputs from getModelInputs() in order, followed by temps as
107 // submodel inputs from getTempsAsSubModelInputs() in order,
108 // followed by outputs as submodel inputs from
109 // getOutputsAsSubModelInputs() in order; and it passes outputs from
110 // getModelOutputs() in order followed by submodel outputs from
111 // getTempsAsSubModelOutputs() in order.
112 //
113 // TODO: Maybe the logic for comparing a partition to an expected
114 // model should be changed to tolerate reorderings of inputs and
115 // outputs, so that when we build models and lists to compare
116 // against, we don't need to worry about input and output
117 // orderings. But is there a way to do this that still lets us
118 // verify that we have the correct relationships between
119 // an (original) model's inputs and outputs and each submodel's
120 // inputs and outputs, as well as the correct relationship
121 // between submodel inputs and outputs across partitions?
122
123 namespace {
124
125 using CompilationBuilder = ::android::nn::CompilationBuilder;
126 using Device = ::android::nn::Device;
127 using DeviceManager = ::android::nn::DeviceManager;
128 using ExecutePreference = ::android::nn::wrapper::ExecutePreference;
129 using ExecutionPlan = ::android::nn::ExecutionPlan;
130 using ExecutionStep = ::android::nn::ExecutionStep;
131 using HidlModel = ::android::hardware::neuralnetworks::V1_1::Model;
132 using ModelBuilder = ::android::nn::ModelBuilder;
133 using Result = ::android::nn::wrapper::Result;
134 using SampleDriver = ::android::nn::sample_driver::SampleDriver;
135 using WrapperCompilation = ::android::nn::wrapper::Compilation;
136 using WrapperModel = ::android::nn::wrapper::Model;
137 using WrapperOperandType = ::android::nn::wrapper::OperandType;
138 using WrapperType = ::android::nn::wrapper::Type;
139
140 template <typename T> using sp = ::android::sp<T>;
141
142 // We employ an operation numbering scheme:
143 // - 0..FuseCode-1 = ADD with the appropriate activation function
144 // - FuseCode..2*FuseCode-1 = MUL with the appropriate activation function
145 const uint32_t kNumFuseCodes = 4;
146 const uint32_t kBadOperation = ~0;
147
148 // Look up the operation with the specified index in a graph, and
149 // return the operation encoding -- 0..7; or, if for some reason this
150 // is not one of the encoded operations, then return kBadOperation.
lookupOperation(std::function<const Operation & (uint32_t)> getOperation,std::function<const Operand & (uint32_t)> getOperand,std::function<const uint8_t * (uint32_t)> getValue,uint32_t operationIndex)151 uint32_t lookupOperation(std::function<const Operation&(uint32_t)> getOperation,
152 std::function<const Operand&(uint32_t)> getOperand,
153 std::function<const uint8_t*(uint32_t)> getValue,
154 uint32_t operationIndex) {
155 const Operation& operation = getOperation(operationIndex);
156 switch (operation.type) {
157 case OperationType::ADD:
158 case OperationType::MUL: {
159 // input2 is the fused activation function
160 const Operand& input2 = getOperand(operation.inputs[2]);
161 if ((input2.type == OperandType::INT32) &&
162 (input2.lifetime == OperandLifeTime::CONSTANT_COPY)) {
163 int32_t value;
164 memcpy(&value,
165 getValue(input2.location.offset),
166 input2.location.length);
167 if (operation.type == OperationType::MUL) {
168 value += kNumFuseCodes;
169 }
170 return value;
171 }
172 break;
173 }
174 default:
175 break;
176 }
177 return kBadOperation;
178 }
179
lookupOperation(const HidlModel & model,uint32_t operationIndex)180 uint32_t lookupOperation(const HidlModel& model, uint32_t operationIndex) {
181 return lookupOperation(
182 [&model](uint32_t index) -> const Operation& {
183 return model.operations[index];
184 },
185 [&model](uint32_t index) -> const Operand& {
186 return model.operands[index];
187 },
188 [&model](uint32_t offset) {return &model.operandValues[offset];},
189 operationIndex);
190 }
191
192 #ifdef VERBOSE
193 // This is a debugging utility function
dump(const char * name,const ModelBuilder * model)194 void dump(const char* name, const ModelBuilder* model) {
195 HidlModel hidlModel;
196 model->setHidlModel(&hidlModel);
197 std::cout << name << ": " << toString(hidlModel) << std::endl;
198 std::cout << "inputs: " << toString(hidlModel.inputIndexes) << std::endl;
199 std::cout << "outputs: " << toString(hidlModel.outputIndexes) << std::endl;
200 for (size_t i = 0, e = hidlModel.operations.size(); i < e; i++) {
201 std::cout << "operation[" << i << "]: " << toString(hidlModel.operations[i]) << std::endl;
202 }
203 }
204 #endif
205
206 #ifdef GRAPH
hidlGraphDump(const char * name,const HidlModel & model)207 inline void hidlGraphDump(const char* name, const HidlModel& model) {
208 ::android::nn::graphDump(name, model);
209 }
210 #endif
211
graphDump(const char * name,const WrapperModel & model)212 void graphDump([[maybe_unused]] const char* name, [[maybe_unused]] const WrapperModel& model) {
213 #ifdef GRAPH
214 HidlModel hidlModel;
215 reinterpret_cast<const ModelBuilder*>(model.getHandle())->setHidlModel(&hidlModel);
216 hidlGraphDump(name, hidlModel);
217 #endif
218 }
219
220 // This is an IDevice for testing purposes. It only has a few
221 // interesting properties, all of which are specified as constructor
222 // arguments: device capabilities; which subset of operation kinds
223 // (0..7) does the device support; does the device support the OEM
224 // operation. The subset is represented with a bitmask, in which
225 // operation kind K corresponds to the bit (1 << K).
226 class PartitioningDriver : public SampleDriver {
227 private:
228 // Dummy class -- a prepared model must not be nullptr.
229 class PartitioningPreparedModel : public IPreparedModel {
230 public:
execute(const Request &,const sp<IExecutionCallback> &)231 Return<ErrorStatus> execute(const Request&,
232 const sp<IExecutionCallback>&) override {
233 return ErrorStatus::DEVICE_UNAVAILABLE;
234 }
235 };
236 public:
237 enum OEM { OEMNo, OEMYes };
238
PartitioningDriver(const char * name,Capabilities capabilities,uint32_t operationMask,OEM oem=OEMNo)239 PartitioningDriver(const char *name, Capabilities capabilities,
240 uint32_t operationMask, OEM oem = OEMNo) :
241 SampleDriver(name), mCapabilities(capabilities),
242 mOperationMask(operationMask), mOEM(oem) {}
~PartitioningDriver()243 ~PartitioningDriver() override {}
244
prepareModel_1_1(const Model &,ExecutionPreference,const sp<IPreparedModelCallback> & cb)245 Return<ErrorStatus> prepareModel_1_1(const Model&, ExecutionPreference,
246 const sp<IPreparedModelCallback>& cb) override {
247 cb->notify(ErrorStatus::NONE, new PartitioningPreparedModel);
248 return ErrorStatus::NONE;
249 }
250
getStatus()251 Return<DeviceStatus> getStatus() override {
252 return DeviceStatus::AVAILABLE;
253 }
254
getCapabilities_1_1(getCapabilities_1_1_cb cb)255 Return<void> getCapabilities_1_1(getCapabilities_1_1_cb cb) override {
256 cb(ErrorStatus::NONE, mCapabilities);
257 return Void();
258 }
259
getSupportedOperations_1_1(const Model & model,getSupportedOperations_cb cb)260 Return<void> getSupportedOperations_1_1(const Model& model,
261 getSupportedOperations_cb cb) override {
262 if (!android::nn::validateModel(model)) {
263 cb(ErrorStatus::INVALID_ARGUMENT, std::vector<bool>());
264 return Void();
265 }
266
267 const size_t count = model.operations.size();
268 std::vector<bool> supported(count);
269 for (size_t i = 0; i < count; i++) {
270 if (model.operations[i].type == OperationType::OEM_OPERATION) {
271 supported[i] = (mOEM == OEMYes);
272 continue;
273 }
274 supported[i] = false;
275 uint32_t operation = lookupOperation(model, i);
276 if ((operation != kBadOperation) && (mOperationMask & (1 << operation))) {
277 supported[i] = true;
278 }
279 }
280 cb(ErrorStatus::NONE, supported);
281 return Void();
282 }
283
284 private:
285 Capabilities mCapabilities;
286 uint32_t mOperationMask;
287 OEM mOEM;
288 };
289
290 // This class adds some simple abstractions and utilities on top of
291 // ::android::nn::wrapper::Model. For example, it provides methods
292 // that work in terms of operation kind (0..7); and because we care
293 // about graph topology rather than details of operand types and
294 // values, it greatly simplifies the process of creating operands.
295 class PartitioningModel : public WrapperModel {
296 public:
297 // Create a tensor operand of the specified type, and return the
298 // corresponding operand index.
addFloatOperand()299 uint32_t addFloatOperand() {
300 static const WrapperOperandType type(WrapperType::TENSOR_FLOAT32, { 1 });
301 return addOperand(&type);
302 }
addQuantOperand()303 uint32_t addQuantOperand() {
304 static const WrapperOperandType type(WrapperType::TENSOR_QUANT8_ASYMM, { 1 });
305 return addOperand(&type);
306 }
307
308 // Create an operation with two inputs and one output, specifying
309 // the operation kind (0..7) and the input operand indexes.
310 // Returns the output operand index.
311 enum class Dimensioned { NO, YES };
addOperation2To1(uint32_t operation,const uint32_t input0,const uint32_t input1,Dimensioned dimensionedOutput=Dimensioned::YES)312 uint32_t addOperation2To1(uint32_t operation, const uint32_t input0, const uint32_t input1,
313 Dimensioned dimensionedOutput = Dimensioned::YES) {
314 ANeuralNetworksOperationType type =
315 (operation < kNumFuseCodes ? ANEURALNETWORKS_ADD : ANEURALNETWORKS_MUL);
316 int32_t fuseCode = (operation < kNumFuseCodes ? operation : operation - kNumFuseCodes);
317 uint32_t input2 = addIntOperand(fuseCode);
318 uint32_t output = addOperandOfSameType(input0, dimensionedOutput);
319 addOperation(type, { input0, input1, input2 }, { output });
320 return output;
321 }
322
323 // Create an OEM operation with one input and one output,
324 // specifying the input operand index. Returns the output operand
325 // index.
addOperationOEM1To1(const uint32_t input,Dimensioned dimensionedOutput=Dimensioned::YES)326 uint32_t addOperationOEM1To1(const uint32_t input,
327 Dimensioned dimensionedOutput = Dimensioned::YES) {
328 uint32_t output = addOperandOfSameType(input, dimensionedOutput);
329 addOperation(ANEURALNETWORKS_OEM_OPERATION, { input }, { output });
330 return output;
331 }
332
333 // Run the partitioning algorithm to create an ExecutionPlan.
partitionTheWork(const std::vector<std::shared_ptr<Device>> & devices,ExecutePreference preference,ExecutionPlan * plan)334 int partitionTheWork(const std::vector<std::shared_ptr<Device>>& devices,
335 ExecutePreference preference, ExecutionPlan* plan) {
336 return reinterpret_cast<ModelBuilder*>(getHandle())->partitionTheWork(
337 devices, static_cast<uint32_t>(preference), plan);
338 }
339
340 #ifdef VERBOSE
341 // This is a debugging utility function.
dump(const char * name) const342 void dump(const char* name) const {
343 const ModelBuilder* mb = reinterpret_cast<const ModelBuilder*>(getHandle());
344 ::dump(name, mb);
345 }
346 #endif
347
348 private:
349
350 // Create a scalar integer operand of the specified value, and
351 // return the corresponding operand index.
addIntOperand(int32_t value)352 uint32_t addIntOperand(int32_t value) {
353 static const WrapperOperandType type(WrapperType::INT32, { });
354 uint32_t operand = addOperand(&type);
355 setOperandValue(operand, &value, sizeof(value));
356 return operand;
357 }
358
359 // Create an operand of the same type as the specified operand,
360 // and return the operand index of the new operand.
addOperandOfSameType(uint32_t operand,Dimensioned dimensioned=Dimensioned::YES)361 uint32_t addOperandOfSameType(uint32_t operand, Dimensioned dimensioned = Dimensioned::YES) {
362 const Operand& operandStruct =
363 reinterpret_cast<const ModelBuilder*>(getHandle())->getOperand(operand);
364 WrapperOperandType type(static_cast<WrapperType>(operandStruct.type), { 1 });
365 if (dimensioned == Dimensioned::NO) {
366 for (auto& dimension : type.dimensions) {
367 dimension = 0;
368 }
369 }
370 return addOperand(&type);
371 }
372 };
373
374 // This class adds some utilities on top of ::android::nn::wrapper::Compilation.
375 class PartitioningCompilation : public WrapperCompilation {
376 public:
PartitioningCompilation(const WrapperModel * model)377 PartitioningCompilation(const WrapperModel* model) : WrapperCompilation(model) { }
378
setPartitioning(uint32_t partitioning)379 Result setPartitioning(uint32_t partitioning) {
380 return static_cast<Result>(builder()->setPartitioning(partitioning));
381 }
382
383 using WrapperCompilation::finish;
finish(const std::vector<std::shared_ptr<Device>> & devices)384 Result finish(const std::vector<std::shared_ptr<Device>>& devices) {
385 return static_cast<Result>(builder()->finish(devices));
386 }
387
getExecutionPlan() const388 const ExecutionPlan& getExecutionPlan() const {
389 return builder()->forTest_getExecutionPlan();
390 }
391
392 private:
builder()393 CompilationBuilder* builder() {
394 return reinterpret_cast<CompilationBuilder*>(getHandle());
395 }
396
builder() const397 const CompilationBuilder* builder() const {
398 return reinterpret_cast<const CompilationBuilder*>(getHandle());
399 }
400 };
401
402 #ifdef VERBOSE
403 #define RETURN_TRUE() \
404 { \
405 std::cerr << "returning true from " << __LINE__ << std::endl; \
406 return true; \
407 }
408 #else
409 #define RETURN_TRUE() \
410 { \
411 return true; \
412 }
413 #endif
414 #ifdef VERBOSE
415 #define RETURN_FALSE(MESSAGE) \
416 { \
417 std::cerr << "returning false from " << __LINE__ MESSAGE << std::endl; \
418 return false; \
419 }
420 #else
421 #define RETURN_FALSE(MESSAGE) \
422 { \
423 return false; \
424 }
425 #endif
426
427 class PartitioningTest : public ::testing::Test {
428 protected:
429 using RemapVectorType = ExecutionStep::RemapVectorType;
430 using SubModelOutputSetType = ExecutionStep::SubModelOutputSetType;
431
SetUp()432 virtual void SetUp() {
433 }
434
435 // From a vector of DeviceSpecification, create a vector of
436 // Devices.
437 struct DeviceSpecification {
DeviceSpecification__anon83dd480d0111::PartitioningTest::DeviceSpecification438 DeviceSpecification(const std::string &name, Capabilities capabilities,
439 uint32_t operationMask,
440 PartitioningDriver::OEM oem = PartitioningDriver::OEMNo) :
441 mName(name), mCapabilities(capabilities),
442 mOperationMask(operationMask), mOEM(oem) { }
443 std::string mName;
444 Capabilities mCapabilities;
445 uint32_t mOperationMask;
446 PartitioningDriver::OEM mOEM;
447 };
448 static std::vector<std::shared_ptr<Device>>
makeDevices(std::vector<DeviceSpecification> specifications)449 makeDevices(std::vector<DeviceSpecification> specifications) {
450 std::vector<std::shared_ptr<Device>> devices;
451 for (const auto& specification : specifications) {
452 devices.push_back(std::make_shared<Device>(
453 specification.mName,
454 new PartitioningDriver(specification.mName.c_str(),
455 specification.mCapabilities,
456 specification.mOperationMask,
457 specification.mOEM)));
458 if (!devices.back()->initialize()) {
459 EXPECT_NE("failed to initialize device", nullptr);
460 return {};
461 }
462 }
463 return devices;
464 }
465
466 /*-- Graph comparision ----------------------------------------------------------------*/
467
468 // An operand with certain values for its lifetime does not have a
469 // defining operation in the graph. For the purposes of the graph
470 // comparison algorithm, we encode the "defining operation" index of
471 // such an operand as follows:
472 // - NO_VALUE kPseudoDefiningOperationNoValue
473 // - MODEL_INPUT kPseudoDefiningOperationModelInput0 + (position in list of inputs)
474 // - CONSTANT_COPY kPseudoDefiningOperationConstantCopy0 + (constant value)
475 // Note: For the graphs we build in this test, we
476 // only expect to see 4-byte constants within
477 // a very restricted range, so we only make
478 // room for such constants in our encoding
479 // space.
480 // We do not expect to see CONSTANT_REFERENCE, and so we do not handle
481 // it.
482 //
483 // The encoding is intended to be relatively human readable; it is not
484 // designed to represent some optimal balance of ranges for the items
485 // within its scope (actual operations, inputs, constants).
486
487 enum PseudoDefiningOperationEncodings : uint32_t {
488 kPseudoDefiningOperationModelInput0 = 0x80000000U,
489 kPseudoDefiningOperationConstantCopy0 = 0x90000000U,
490 kPseudoDefiningOperationNoValue = 0xeeeeeeeeU,
491
492 // lowest value for special encoding
493 kPseudoDefiningOperationBase = 0x80000000U,
494
495 // range of encoded input or constant
496 kPseudoDefiningOperationRange = 0x10000000U,
497 };
498
499 // Build a map from operand to defining operation.
500 // TODO: Replace map with vector?
buildDefinitionMap(const ModelBuilder * model,std::map<uint32_t,uint32_t> * defMap)501 void buildDefinitionMap(const ModelBuilder* model,
502 std::map<uint32_t, uint32_t>* defMap) {
503 // actual definitions
504 ASSERT_LT(model->operationCount(), kPseudoDefiningOperationBase);
505 for (uint32_t i = 0, e = model->operationCount(); i < e; i++) {
506 const Operation& operation = model->getOperation(i);
507 for (uint32_t output : operation.outputs) {
508 (*defMap)[output] = i;
509 }
510 }
511 // inputs
512 ASSERT_LT(model->inputCount(), kPseudoDefiningOperationRange);
513 for (uint32_t i = 0, e = model->inputCount(); i < e; i++) {
514 (*defMap)[model->getInputOperandIndex(i)] = kPseudoDefiningOperationModelInput0 + i;
515 }
516 // look for NO_VALUE and CONSTANT_COPY
517 for (uint32_t i = 0, e = model->operandCount(); i < e; i++) {
518 const Operand& operand = model->getOperand(i);
519 switch (operand.lifetime) {
520 case OperandLifeTime::NO_VALUE:
521 (*defMap)[i] = kPseudoDefiningOperationNoValue;
522 break;
523 case OperandLifeTime::CONSTANT_COPY: {
524 ASSERT_EQ(operand.location.length, sizeof(uint32_t));
525 uint32_t value;
526 memcpy(&value, model->getPointerToOperandValue(operand.location.offset), sizeof(uint32_t));
527 ASSERT_LT(value, kPseudoDefiningOperationNoValue);
528 (*defMap)[i] = kPseudoDefiningOperationConstantCopy0 + value;
529 break;
530 }
531 case OperandLifeTime::TEMPORARY_VARIABLE:
532 case OperandLifeTime::MODEL_INPUT:
533 case OperandLifeTime::MODEL_OUTPUT:
534 // already handled
535 break;
536 default:
537 FAIL();
538 break;
539 }
540 }
541 // sanity check
542 ASSERT_EQ(model->operandCount(), defMap->size());
543 }
544
545 #ifdef VERBOSE
dump(const char * name,const std::map<uint32_t,uint32_t> * aMap)546 void dump(const char* name, const std::map<uint32_t, uint32_t>* aMap) {
547 auto writeNum = [](uint32_t num) {
548 if (num >= kPseudoDefiningOperationBase) {
549 std::cout << "0x" << std::hex << num << std::dec;
550 } else {
551 std::cout << num;
552 }
553 };
554
555 std::cout << name << ": { ";
556 bool gotOne = false;
557 for (const auto& entry : *aMap) {
558 if (gotOne) {
559 std::cout << ", ";
560 } else {
561 gotOne = true;
562 }
563 std::cout << "(";
564 writeNum(entry.first);
565 std::cout << ", ";
566 writeNum(entry.second);
567 std::cout << ")";
568 }
569 std::cout << " }" << std::endl;
570 }
571 #endif
572
compare(const Operand & operandA,const Operand & operandB)573 bool compare(const Operand& operandA, const Operand& operandB) {
574 if (operandA.type != operandB.type ||
575 operandA.dimensions != operandB.dimensions ||
576 operandA.numberOfConsumers != operandB.numberOfConsumers ||
577 operandA.scale != operandB.scale ||
578 operandA.zeroPoint != operandB.zeroPoint) {
579 return false;
580 }
581 return true;
582 }
583
584 // Compare two graphs. We ignore operand and operation indexes (i.e.,
585 // two nodes can be the same even if they are numbered differently)
586 // but we also ignore semantics (e.g., even if an operation kind is
587 // such that the operand is commutative, we still pay attention to the
588 // order of its input operands).
589 //
590 // The comparison algorithm works by walking modelA from outputs
591 // towards inputs, along the edge from each operand to its
592 // defining operation, and then along the edges to the operation's
593 // input operands. At each step along the way, we try to match up
594 // operands and operations from modelA with equivalent operands
595 // and operations from modelB.
596 //
597 // We start by assuming that modelA's outputs and modelB's outputs
598 // match positionally (e.g., modelA's first output operand is
599 // equivalent to modelB's first output operand). Once we've
600 // discovered two equivalent operands (such as those outputs), we
601 // place them in a work queue. We repeatedly pull operands off
602 // the queue and compare their defining operations and those
603 // operations' input operands, to discover more pairs of
604 // equivalent operands. If we ever find operations that do not
605 // match (e.g., because operation kind differs), or operands that
606 // do not match (e.g., because operand type differs); or if we
607 // ever find a conflict (we've already decided that operand A's
608 // equivalent operand is B0, but it looks like we need its
609 // equivalent operand to be B1); then the graphs compare unequal.
610 // Otherwise, we'll eventually exhaust the work queue, and
611 // conclude that the graphs compare equal.
compare(const ModelBuilder * modelA,const ModelBuilder * modelB)612 bool compare(const ModelBuilder* modelA, const ModelBuilder* modelB) {
613 #ifdef VERBOSE
614 ::dump("compare(A)", modelA);
615 ::dump("compare(B)", modelB);
616 #endif
617
618 if (modelA->operandCount() != modelB->operandCount() ||
619 modelA->operationCount() != modelB->operationCount() ||
620 modelA->inputCount() != modelB->inputCount() ||
621 modelA->outputCount() != modelB->outputCount()) {
622 RETURN_FALSE();
623 }
624
625 // Maps from operand index to index of defining operation.
626 std::map<uint32_t, uint32_t> defsA, defsB;
627 buildDefinitionMap(modelA, &defsA);
628 buildDefinitionMap(modelB, &defsB);
629 if (HasFatalFailure()) return false;
630
631 // Maps from operand index in modelA to equivalent operand index
632 // in modelB; and from operation index in modelA to equivalent
633 // operation index in modelB.
634 std::map<uint32_t, uint32_t> equivalentOperandsAToB;
635 std::map<uint32_t, uint32_t> equivalentOperationsAToB;
636
637 // Queue of operand indexes from modelA, each of whose defining
638 // operations are to be checked for equivalence with modelB.
639 std::queue<uint32_t> workQueueOperandsA;
640
641 // Seed operand equivalence map and work queue from model outputs.
642 for (uint32_t i = 0, e = modelA->outputCount(); i < e; i++) {
643 uint32_t outputA = modelA->getOutputOperandIndex(i);
644 uint32_t outputB = modelB->getOutputOperandIndex(i);
645 if (!compare(modelA->getOperand(outputA), modelB->getOperand(outputB))) {
646 RETURN_FALSE();
647 }
648 equivalentOperandsAToB[outputA] = outputB;
649 workQueueOperandsA.push(outputA);
650 }
651
652 #ifdef VERBOSE
653 dump("defsA", &defsA);
654 dump("defsB", &defsB);
655 #endif
656
657 // Process the queue.
658 uint32_t pseudoDefinitionCount = 0;
659 while (!workQueueOperandsA.empty()) {
660 #ifdef VERBOSE
661 dump("equivalentOperandsAToB", &equivalentOperandsAToB);
662 dump("equivalentOperationsAToB", &equivalentOperationsAToB);
663 #endif
664 uint32_t operandIndexA = workQueueOperandsA.front();
665 #ifdef VERBOSE
666 std::cout << "operandIndexA: " << operandIndexA << std::endl;
667 #endif
668 workQueueOperandsA.pop();
669 uint32_t operandIndexB = equivalentOperandsAToB.at(operandIndexA);
670
671 uint32_t operationIndexA = defsA.at(operandIndexA);
672 uint32_t operationIndexB = defsB.at(operandIndexB);
673 auto it = equivalentOperationsAToB.find(operationIndexA);
674 if (it != equivalentOperationsAToB.end()) {
675 if (it->second != operationIndexB) {
676 RETURN_FALSE();
677 }
678 continue;
679 }
680
681 // We haven't identified an equivalent operation for
682 // operationIndexA.
683
684 if ((operationIndexA >= kPseudoDefiningOperationBase) !=
685 (operationIndexB >= kPseudoDefiningOperationBase)) {
686 RETURN_FALSE();
687 }
688 // Either both operands have pseudo-definitions, or neither
689 // does.
690 if (operationIndexA >= kPseudoDefiningOperationBase) {
691 // Both operands have pseudo-definitions.
692 if (operationIndexA != operationIndexB) {
693 RETURN_FALSE();
694 }
695 equivalentOperationsAToB[operationIndexA] = operationIndexB;
696 ++pseudoDefinitionCount;
697 continue;
698 }
699
700 // If we get here, neither operation A nor operation B is a
701 // pseudo-definition.
702
703 const Operation& operationA = modelA->getOperation(operationIndexA);
704 const Operation& operationB = modelB->getOperation(operationIndexB);
705 if (operationA.type != operationB.type ||
706 operationA.inputs.size() != operationB.inputs.size() ||
707 operationA.outputs.size() != operationB.outputs.size()) {
708 RETURN_FALSE();
709 }
710 equivalentOperationsAToB[operationIndexA] = operationIndexB;
711 for (uint32_t i = 0, e = operationA.inputs.size(); i < e; i++) {
712 uint32_t inputA = operationA.inputs[i];
713 uint32_t inputB = operationB.inputs[i];
714 auto it = equivalentOperandsAToB.find(inputA);
715 if (it != equivalentOperandsAToB.end()) {
716 if (it->second != inputB) {
717 RETURN_FALSE();
718 }
719 continue;
720 }
721 // We haven't identified an equivalent operand for inputA.
722 if (!compare(modelA->getOperand(inputA), modelB->getOperand(inputB))) {
723 RETURN_FALSE();
724 }
725 equivalentOperandsAToB[inputA] = inputB;
726 workQueueOperandsA.push(inputA);
727 }
728 }
729
730 // Sanity check
731 if (modelA->operandCount() != defsA.size() ||
732 modelA->operandCount() != defsB.size() ||
733 modelA->operandCount() != equivalentOperandsAToB.size() ||
734 modelA->operationCount() + pseudoDefinitionCount != equivalentOperationsAToB.size()) {
735 RETURN_FALSE();
736 }
737
738 RETURN_TRUE();
739 }
740
741 /*-------------------------------------------------------------------------------------*/
742
compare(std::shared_ptr<const ExecutionStep> step,const WrapperModel * model,std::shared_ptr<Device> device)743 bool compare(std::shared_ptr<const ExecutionStep> step,
744 const WrapperModel* model, std::shared_ptr<Device> device) {
745 return (step->getDevice() == device) &&
746 compare(step->getSubModel(),
747 reinterpret_cast<const ModelBuilder*>(model->getHandle()));
748 }
749 };
750
TEST_F(PartitioningTest,SimpleModel)751 TEST_F(PartitioningTest, SimpleModel) {
752 PartitioningModel model;
753 uint32_t opnd0 = model.addFloatOperand();
754 uint32_t opnd1 = model.addFloatOperand();
755 uint32_t opnd2 = model.addOperation2To1(0, opnd0, opnd1);
756 uint32_t opnd3 = model.addFloatOperand();
757 uint32_t opnd4 = model.addOperation2To1(1, opnd2, opnd3);
758 model.identifyInputsAndOutputs({ opnd0, opnd1, opnd3 }, { opnd4 });
759 model.finish();
760 ASSERT_TRUE(model.isValid());
761 graphDump("SimpleModel", model);
762
763 // Simple partition (two devices are each capable of everything, one is the best).
764 const auto devicesA = makeDevices(
765 {
766 {"bad", { .float32Performance = { .execTime = 0.9, .powerUsage = 0.9 },
767 .quantized8Performance = { .execTime = 0.9, .powerUsage = 0.9 } }, ~0U},
768 {"good", { .float32Performance = { .execTime = 0.5, .powerUsage = 0.5 },
769 .quantized8Performance = { .execTime = 0.5, .powerUsage = 0.5 } }, ~0U}
770 });
771 ExecutionPlan planA;
772 ASSERT_EQ(model.partitionTheWork(devicesA, ExecutePreference::PREFER_LOW_POWER, &planA),
773 ANEURALNETWORKS_NO_ERROR);
774 ASSERT_EQ(planA.forTest_getKind(), ExecutionPlan::Kind::SIMPLE);
775 ASSERT_NE(planA.forTest_simpleGetDevice().get(), nullptr);
776 ASSERT_EQ(planA.forTest_simpleGetDevice()->getName(), "good");
777
778 // Simple partition (two devices are each capable of everything, none better than CPU).
779 const auto devicesC = makeDevices(
780 {
781 {"bad", { .float32Performance = { .execTime = 1.1, .powerUsage = 1.1 },
782 .quantized8Performance = { .execTime = 1.1, .powerUsage = 1.1 } }, ~0U},
783 {"bad2", { .float32Performance = { .execTime = 1.0, .powerUsage = 1.0 },
784 .quantized8Performance = { .execTime = 1.0, .powerUsage = 1.0 } }, ~0U}
785 });
786 ExecutionPlan planC;
787 ASSERT_EQ(model.partitionTheWork(devicesC, ExecutePreference::PREFER_LOW_POWER, &planC),
788 ANEURALNETWORKS_NO_ERROR);
789 ASSERT_EQ(planC.forTest_getKind(), ExecutionPlan::Kind::SIMPLE);
790 ASSERT_EQ(planC.forTest_simpleGetDevice(), nullptr);
791
792 // Compound partition (two devices, each is capable of one of the
793 // two operations). We could do more extensive checking here --
794 // for example, verify that each step within the plan has the
795 // correct (model and submodel)x(inputs and outputs).
796 const auto devicesB = makeDevices(
797 {
798 {"0", { .float32Performance = { .execTime = 0.9, .powerUsage = 0.9 },
799 .quantized8Performance = { .execTime = 0.9, .powerUsage = 0.9 } }, 1<<0},
800 {"1", { .float32Performance = { .execTime = 0.5, .powerUsage = 0.5 },
801 .quantized8Performance = { .execTime = 0.5, .powerUsage = 0.5 } }, 1<<1}
802 });
803 ExecutionPlan planB;
804 ASSERT_EQ(model.partitionTheWork(devicesB, ExecutePreference::PREFER_LOW_POWER, &planB),
805 ANEURALNETWORKS_NO_ERROR);
806 ASSERT_EQ(planB.forTest_getKind(), ExecutionPlan::Kind::COMPOUND);
807 const auto& stepsB = planB.forTest_compoundGetSteps();
808 ASSERT_EQ(stepsB.size(), size_t(2));
809 {
810 // Build a model to compare against the submodel from stepsB[0].
811 PartitioningModel modelB0;
812 uint32_t b0Opnd0 = modelB0.addFloatOperand();
813 uint32_t b0Opnd1 = modelB0.addFloatOperand();
814 uint32_t b0Opnd2 = modelB0.addOperation2To1(0, b0Opnd0, b0Opnd1);
815 modelB0.identifyInputsAndOutputs({ b0Opnd0, b0Opnd1 }, { b0Opnd2 });
816 modelB0.finish();
817 ASSERT_TRUE(modelB0.isValid());
818 ASSERT_NO_FATAL_FAILURE(ASSERT_TRUE(compare(stepsB[0], &modelB0, devicesB[0])));
819 ASSERT_EQ(stepsB[0]->getModelInputs(),
820 (RemapVectorType{ { opnd0, b0Opnd0 }, { opnd1, b0Opnd1 } }));
821 ASSERT_EQ(stepsB[0]->getModelOutputs(),
822 (RemapVectorType{}));
823 ASSERT_EQ(stepsB[0]->getTempsAsSubModelInputs(),
824 (RemapVectorType{}));
825 ASSERT_EQ(stepsB[0]->getTempsAsSubModelOutputs(),
826 (SubModelOutputSetType{ { opnd2, b0Opnd2 } }));
827 ASSERT_EQ(stepsB[0]->getOutputsAsSubModelInputs(),
828 (RemapVectorType{}));
829 }
830 {
831 // Build a model to compare against the submodel from stepsB[1].
832 PartitioningModel modelB1;
833 uint32_t b1Opnd2 = modelB1.addFloatOperand();
834 uint32_t b1Opnd3 = modelB1.addFloatOperand();
835 uint32_t b1Opnd4 = modelB1.addOperation2To1(1, b1Opnd2, b1Opnd3);
836 // Note: In the partitioning algorithm, submodel inputs follow
837 // model inputs. In the original model "model", opnd2 is not
838 // an input; so in the submodel "modelB1", the corresponding
839 // input b1Opnd2 is a submodel input, and must follow the
840 // model input b1Opnd3.
841 modelB1.identifyInputsAndOutputs({ b1Opnd3, b1Opnd2 }, { b1Opnd4 });
842 modelB1.finish();
843 ASSERT_TRUE(modelB1.isValid());
844 ASSERT_NO_FATAL_FAILURE(ASSERT_TRUE(compare(stepsB[1], &modelB1, devicesB[1])));
845 ASSERT_EQ(stepsB[1]->getModelInputs(),
846 (RemapVectorType{ { opnd3, b1Opnd3 } }));
847 ASSERT_EQ(stepsB[1]->getModelOutputs(),
848 (RemapVectorType{ { opnd4, b1Opnd4 } }));
849 ASSERT_EQ(stepsB[1]->getTempsAsSubModelInputs(),
850 (RemapVectorType{ { opnd2, b1Opnd2 } }));
851 ASSERT_EQ(stepsB[1]->getTempsAsSubModelOutputs(),
852 (SubModelOutputSetType{}));
853 ASSERT_EQ(stepsB[1]->getOutputsAsSubModelInputs(),
854 (RemapVectorType{}));
855 }
856 }
857
TEST_F(PartitioningTest,Cpu)858 TEST_F(PartitioningTest, Cpu) {
859 // Here's a model where some operations execute only on the Cpu.
860 // To make things interesting, we produce three partitions --
861 // device, cpu, same-device.
862
863 static const uint32_t kCpuOp = 1;
864 static const uint32_t kDevOp = 2;
865
866 const auto devices = makeDevices(
867 {
868 {"1", { .float32Performance = { .execTime = 0.5, .powerUsage = 0.5 },
869 .quantized8Performance = { .execTime = 0.5, .powerUsage = 0.5 } }, 1<<kDevOp}
870 });
871
872 PartitioningModel model;
873
874 uint32_t opnd0 = model.addFloatOperand();
875 uint32_t opnd1 = model.addFloatOperand();
876
877 uint32_t opnd2 = model.addOperation2To1(kDevOp, opnd0, opnd1);
878 uint32_t opnd3 = model.addOperation2To1(kDevOp, opnd0, opnd2);
879
880 uint32_t opnd4 = model.addOperation2To1(kCpuOp, opnd0, opnd3);
881 uint32_t opnd5 = model.addOperation2To1(kCpuOp, opnd2, opnd4);
882
883 uint32_t opnd6 = model.addFloatOperand();
884
885 uint32_t opnd7 = model.addOperation2To1(kDevOp, opnd3, opnd5);
886 uint32_t opnd8 = model.addOperation2To1(kDevOp, opnd6, opnd7);
887
888 model.identifyInputsAndOutputs({ opnd0, opnd1, opnd6 }, { opnd4, opnd8 });
889 model.finish();
890 ASSERT_TRUE(model.isValid());
891
892 ExecutionPlan plan;
893 ASSERT_EQ(model.partitionTheWork(devices, ExecutePreference::PREFER_LOW_POWER, &plan),
894 ANEURALNETWORKS_NO_ERROR);
895 ASSERT_EQ(plan.forTest_getKind(), ExecutionPlan::Kind::COMPOUND);
896 const auto& steps = plan.forTest_compoundGetSteps();
897 ASSERT_EQ(steps.size(), size_t(3));
898 {
899 const auto& step0 = steps[0];
900
901 // Build a model to compare against the submodel from steps[0].
902 PartitioningModel model0;
903 uint32_t m0Opnd0 = model0.addFloatOperand();
904 uint32_t m0Opnd1 = model0.addFloatOperand();
905 uint32_t m0Opnd2 = model0.addOperation2To1(kDevOp, m0Opnd0, m0Opnd1);
906 uint32_t m0Opnd3 = model0.addOperation2To1(kDevOp, m0Opnd0, m0Opnd2);
907 model0.identifyInputsAndOutputs({ m0Opnd0, m0Opnd1 }, { m0Opnd2, m0Opnd3 });
908 model0.finish();
909 ASSERT_TRUE(model0.isValid());
910 ASSERT_NO_FATAL_FAILURE(ASSERT_TRUE(compare(step0, &model0, devices[0])));
911 ASSERT_EQ(step0->getModelInputs(),
912 (RemapVectorType{ { opnd0, m0Opnd0 }, { opnd1, m0Opnd1 } }));
913 ASSERT_EQ(step0->getModelOutputs(),
914 (RemapVectorType{}));
915 ASSERT_EQ(step0->getTempsAsSubModelInputs(),
916 (RemapVectorType{}));
917 ASSERT_EQ(step0->getTempsAsSubModelOutputs(),
918 (SubModelOutputSetType{ { opnd2, m0Opnd2 }, { opnd3, m0Opnd3 } }));
919 ASSERT_EQ(step0->getOutputsAsSubModelInputs(),
920 (RemapVectorType{}));
921 }
922 {
923 const auto& step1 = steps[1];
924
925 // Build a model to compare against the submodel from steps[1].
926 PartitioningModel model1;
927 uint32_t m1Opnd0 = model1.addFloatOperand();
928 uint32_t m1Opnd3 = model1.addFloatOperand();
929 uint32_t m1Opnd4 = model1.addOperation2To1(kCpuOp, m1Opnd0, m1Opnd3);
930 uint32_t m1Opnd2 = model1.addFloatOperand();
931 uint32_t m1Opnd5 = model1.addOperation2To1(kCpuOp, m1Opnd2, m1Opnd4);
932 model1.identifyInputsAndOutputs({ m1Opnd0, m1Opnd3, m1Opnd2 }, { m1Opnd4, m1Opnd5 });
933 model1.finish();
934 ASSERT_TRUE(model1.isValid());
935 ASSERT_NO_FATAL_FAILURE(ASSERT_TRUE(compare(step1, &model1, nullptr)));
936 ASSERT_EQ(step1->getModelInputs(),
937 (RemapVectorType{ { opnd0, m1Opnd0 } }));
938 ASSERT_EQ(step1->getModelOutputs(),
939 (RemapVectorType{ { opnd4, m1Opnd4 } }));
940 ASSERT_EQ(step1->getTempsAsSubModelInputs(),
941 (RemapVectorType{ { opnd3, m1Opnd3 }, { opnd2, m1Opnd2 } }));
942 ASSERT_EQ(step1->getTempsAsSubModelOutputs(),
943 (SubModelOutputSetType{ { opnd5, m1Opnd5 } }));
944 ASSERT_EQ(step1->getOutputsAsSubModelInputs(),
945 (RemapVectorType{}));
946 }
947 {
948 const auto& step2 = steps[2];
949
950 // Build a model to compare against the submodel from steps[2].
951 PartitioningModel model2;
952 uint32_t m2Opnd3 = model2.addFloatOperand();
953 uint32_t m2Opnd5 = model2.addFloatOperand();
954 uint32_t m2Opnd7 = model2.addOperation2To1(kDevOp, m2Opnd3, m2Opnd5);
955 uint32_t m2Opnd6 = model2.addFloatOperand();
956 uint32_t m2Opnd8 = model2.addOperation2To1(kDevOp, m2Opnd6, m2Opnd7);
957 model2.identifyInputsAndOutputs({ m2Opnd6, m2Opnd3, m2Opnd5 }, { m2Opnd8 });
958 model2.finish();
959 ASSERT_TRUE(model2.isValid());
960 ASSERT_NO_FATAL_FAILURE(ASSERT_TRUE(compare(step2, &model2, devices[0])));
961 ASSERT_EQ(step2->getModelInputs(),
962 (RemapVectorType{ { opnd6, m2Opnd6 } }));
963 ASSERT_EQ(step2->getModelOutputs(),
964 (RemapVectorType{ { opnd8, m2Opnd8 } }));
965 ASSERT_EQ(step2->getTempsAsSubModelInputs(),
966 (RemapVectorType{ { opnd3, m2Opnd3 }, { opnd5, m2Opnd5 } }));
967 ASSERT_EQ(step2->getTempsAsSubModelOutputs(),
968 (SubModelOutputSetType{}));
969 ASSERT_EQ(step2->getOutputsAsSubModelInputs(),
970 (RemapVectorType{}));
971 }
972 }
973
TEST_F(PartitioningTest,SetPartitioning)974 TEST_F(PartitioningTest, SetPartitioning) {
975 PartitioningModel model;
976 uint32_t opnd0 = model.addFloatOperand();
977 uint32_t opnd1 = model.addFloatOperand();
978 uint32_t opnd2 = model.addOperation2To1(0, opnd0, opnd1, PartitioningModel::Dimensioned::NO);
979 uint32_t opnd3 = model.addFloatOperand();
980 uint32_t opnd4 = model.addOperation2To1(1, opnd2, opnd3);
981 model.identifyInputsAndOutputs({ opnd0, opnd1, opnd3 }, { opnd4 });
982 model.finish();
983 ASSERT_TRUE(model.isValid());
984
985 // We expect that we cannot successfully partition, because we
986 // have an intermediate operand (opnd2) without dimensions, and
987 // this is not currently handled.
988
989 // One device that can and should execute operation 0.
990 const auto devices = makeDevices({
991 {"hw", { .float32Performance = { .execTime = 0.5, .powerUsage = 0.5 },
992 .quantized8Performance = { .execTime = 0.5, .powerUsage = 0.5 } }, (1<<0)},
993 });
994
995 // Test kPartitioningNo. We should not even attempt partitioning,
996 // so there should be no execution plan.
997 PartitioningCompilation cPNo(&model);
998 ASSERT_EQ(cPNo.setPartitioning(DeviceManager::kPartitioningNo), Result::NO_ERROR);
999 ASSERT_EQ(cPNo.finish(devices), Result::NO_ERROR);
1000 ASSERT_EQ(cPNo.getExecutionPlan().forTest_getKind(), ExecutionPlan::Kind::EMPTY);
1001
1002 // Test kPartitioningWithFallback. We should attempt
1003 // partitioning, reach the end of the partitioning process (so we
1004 // have an execution plan), discover the dimensionless
1005 // intermediate operand, and still return success (because of
1006 // fallback).
1007 PartitioningCompilation cPWithFallback(&model);
1008 ASSERT_EQ(cPWithFallback.setPartitioning(DeviceManager::kPartitioningWithFallback), Result::NO_ERROR);
1009 ASSERT_EQ(cPWithFallback.finish(devices), Result::NO_ERROR);
1010 ASSERT_EQ(cPWithFallback.getExecutionPlan().forTest_getKind(), ExecutionPlan::Kind::ERROR);
1011
1012 // Test kPartitioningWithoutFallback. We should attempt
1013 // partitioning, and fail.
1014 PartitioningCompilation cPWithoutFallback(&model);
1015 ASSERT_EQ(cPWithoutFallback.setPartitioning(DeviceManager::kPartitioningWithoutFallback), Result::NO_ERROR);
1016 ASSERT_EQ(cPWithoutFallback.finish(devices), Result::OP_FAILED);
1017 ASSERT_TRUE(cPWithoutFallback.getExecutionPlan().forTest_hasSubModelOutputsOfUnknownSize());
1018 ASSERT_EQ(cPWithoutFallback.getExecutionPlan().forTest_getKind(), ExecutionPlan::Kind::ERROR);
1019 }
1020
1021 // Regression test for http://b/69166603:
1022 // "partitioned compilation and execution yields wrong results when model output is submodel input"
TEST_F(PartitioningTest,ModelOutputAsSubmodelInput)1023 TEST_F(PartitioningTest, ModelOutputAsSubmodelInput) {
1024 PartitioningModel model;
1025 uint32_t opnd0 = model.addFloatOperand();
1026 uint32_t opnd1 = model.addFloatOperand();
1027 uint32_t opnd2 = model.addOperation2To1(0, opnd0, opnd1);
1028 uint32_t opnd3 = model.addOperation2To1(1, opnd2, opnd2);
1029 model.identifyInputsAndOutputs({ opnd0, opnd1 }, { opnd2, opnd3 });
1030 model.finish();
1031 ASSERT_TRUE(model.isValid());
1032
1033 // Compound partition (two devices, each is capable of one of the
1034 // two operations). We could do more extensive checking here --
1035 // for example, verify that each step within the plan has the
1036 // correct (model and submodel)x(inputs and outputs).
1037 const auto devices = makeDevices(
1038 {
1039 {"0", { .float32Performance = { .execTime = 0.5, .powerUsage = 0.5 },
1040 .quantized8Performance = { .execTime = 0.5, .powerUsage = 0.5 } }, 1<<0},
1041 {"1", { .float32Performance = { .execTime = 0.5, .powerUsage = 0.5 },
1042 .quantized8Performance = { .execTime = 0.5, .powerUsage = 0.5 } }, 1<<1}
1043 });
1044 ExecutionPlan plan;
1045 ASSERT_EQ(model.partitionTheWork(devices, ExecutePreference::PREFER_LOW_POWER, &plan),
1046 ANEURALNETWORKS_NO_ERROR);
1047 ASSERT_EQ(plan.forTest_getKind(), ExecutionPlan::Kind::COMPOUND);
1048 const auto& steps = plan.forTest_compoundGetSteps();
1049 ASSERT_EQ(steps.size(), size_t(2));
1050 {
1051 // Build a model to compare against the submodel from steps[0].
1052 PartitioningModel model0;
1053 uint32_t m0Opnd0 = model0.addFloatOperand();
1054 uint32_t m0Opnd1 = model0.addFloatOperand();
1055 uint32_t m0Opnd2 = model0.addOperation2To1(0, m0Opnd0, m0Opnd1);
1056 model0.identifyInputsAndOutputs({ m0Opnd0, m0Opnd1 }, { m0Opnd2 });
1057 model0.finish();
1058 ASSERT_TRUE(model0.isValid());
1059 ASSERT_NO_FATAL_FAILURE(ASSERT_TRUE(compare(steps[0], &model0, devices[0])));
1060 ASSERT_EQ(steps[0]->getModelInputs(),
1061 (RemapVectorType{ { opnd0, m0Opnd0 }, { opnd1, m0Opnd1 } }));
1062 ASSERT_EQ(steps[0]->getModelOutputs(),
1063 (RemapVectorType{ { opnd2, m0Opnd2 } }));
1064 ASSERT_EQ(steps[0]->getTempsAsSubModelInputs(),
1065 (RemapVectorType{}));
1066 ASSERT_EQ(steps[0]->getTempsAsSubModelOutputs(),
1067 (SubModelOutputSetType{}));
1068 ASSERT_EQ(steps[0]->getOutputsAsSubModelInputs(),
1069 (RemapVectorType{}));
1070 }
1071 {
1072 // Build a model to compare against the submodel from steps[1].
1073 PartitioningModel model1;
1074 uint32_t m1Opnd2 = model1.addFloatOperand();
1075 uint32_t m1Opnd3 = model1.addOperation2To1(1, m1Opnd2, m1Opnd2);
1076 model1.identifyInputsAndOutputs({ m1Opnd2 }, { m1Opnd3 });
1077 model1.finish();
1078 ASSERT_TRUE(model1.isValid());
1079 ASSERT_NO_FATAL_FAILURE(ASSERT_TRUE(compare(steps[1], &model1, devices[1])));
1080 ASSERT_EQ(steps[1]->getModelInputs(),
1081 (RemapVectorType{}));
1082 ASSERT_EQ(steps[1]->getModelOutputs(),
1083 (RemapVectorType{ { opnd3, m1Opnd3 } }));
1084 ASSERT_EQ(steps[1]->getTempsAsSubModelInputs(),
1085 (RemapVectorType{}));
1086 ASSERT_EQ(steps[1]->getTempsAsSubModelOutputs(),
1087 (SubModelOutputSetType{}));
1088 ASSERT_EQ(steps[1]->getOutputsAsSubModelInputs(),
1089 (RemapVectorType{ { opnd2, m1Opnd2 } }));
1090 }
1091 }
1092
TEST_F(PartitioningTest,OemOperations)1093 TEST_F(PartitioningTest, OemOperations) {
1094 // Trivial model consisting solely of OEM operation.
1095 PartitioningModel model;
1096 uint32_t opndIn = model.addFloatOperand();
1097 uint32_t opndOut = model.addOperationOEM1To1(opndIn);
1098 model.identifyInputsAndOutputs({ opndIn }, { opndOut });
1099 model.finish();
1100 ASSERT_TRUE(model.isValid());
1101
1102 // Verify that the best driver than can run an OEM operation is
1103 // used, even if it is not better than the CPU.
1104 const auto devicesBestOEM = makeDevices(
1105 {
1106 {"badOEM", { .float32Performance = { .execTime = 1.5, .powerUsage = 1.5 },
1107 .quantized8Performance = { .execTime = 1.5, .powerUsage = 1.5 } },
1108 ~0U, PartitioningDriver::OEMYes},
1109 {"noOEM", { .float32Performance = { .execTime = 0.5, .powerUsage = 0.5 },
1110 .quantized8Performance = { .execTime = 0.5, .powerUsage = 0.5 } },
1111 ~0U, PartitioningDriver::OEMNo},
1112 {"goodOEM", { .float32Performance = { .execTime = 1.2, .powerUsage = 1.2 },
1113 .quantized8Performance = { .execTime = 1.2, .powerUsage = 1.2 } },
1114 ~0U, PartitioningDriver::OEMYes}
1115 });
1116 PartitioningCompilation compilationBestOEM(&model);
1117 ASSERT_EQ(compilationBestOEM.finish(devicesBestOEM), Result::NO_ERROR);
1118 const auto& planBestOEM = compilationBestOEM.getExecutionPlan();
1119 ASSERT_EQ(planBestOEM.forTest_getKind(), ExecutionPlan::Kind::SIMPLE);
1120 ASSERT_NE(planBestOEM.forTest_simpleGetDevice().get(), nullptr);
1121 ASSERT_EQ(planBestOEM.forTest_simpleGetDevice()->getName(), "goodOEM");
1122
1123 // Verify that we get an error if no driver can run an OEM operation.
1124 const auto devicesNoOEM = makeDevices(
1125 {
1126 {"noOEM", { .float32Performance = { .execTime = 0.5, .powerUsage = 0.5 },
1127 .quantized8Performance = { .execTime = 0.5, .powerUsage = 0.5 } },
1128 ~0U, PartitioningDriver::OEMNo}
1129 });
1130 PartitioningCompilation compilationNoOEM(&model);
1131 ASSERT_EQ(compilationNoOEM.finish(devicesNoOEM), Result::BAD_DATA);
1132
1133 // Verify that we get an error if there are no drivers (only CPU fallback).
1134 PartitioningCompilation compilationNoDrivers(&model);
1135 ASSERT_EQ(compilationNoDrivers.finish({} /* no drivers */), Result::BAD_DATA);
1136 }
1137
TEST_F(PartitioningTest,RelaxedFP)1138 TEST_F(PartitioningTest, RelaxedFP) {
1139 const auto devices = makeDevices(
1140 {
1141 // Best choice for non-relaxed model.
1142 {"f32", { .float32Performance = { .execTime = 0.8, .powerUsage = 0.8 },
1143 .relaxedFloat32toFloat16Performance = { .execTime = 0.9, .powerUsage = 0.9 }},
1144 ~0U},
1145 // Best choice for relaxed model.
1146 {"f16", { .float32Performance = { .execTime = 0.9, .powerUsage = 0.9 },
1147 .relaxedFloat32toFloat16Performance = { .execTime = 0.8, .powerUsage = 0.8 }},
1148 ~0U}
1149 });
1150
1151 auto TrivialTest = [&devices](bool doRelax, const char* expectDevice) {
1152 // Trivial model consisting solely of one operation.
1153 SCOPED_TRACE(expectDevice);
1154 PartitioningModel model;
1155 uint32_t opnd0 = model.addFloatOperand();
1156 uint32_t opnd1 = model.addFloatOperand();
1157 uint32_t opnd2 = model.addOperation2To1(0, opnd0, opnd1);
1158 model.identifyInputsAndOutputs({ opnd0, opnd1 }, { opnd2 });
1159 model.relaxComputationFloat32toFloat16(doRelax);
1160 model.finish();
1161 ASSERT_TRUE(model.isValid());
1162 // Verify that the model will be executed on the appropriate device.
1163 ExecutionPlan plan;
1164 ASSERT_EQ(model.partitionTheWork(devices, ExecutePreference::PREFER_LOW_POWER, &plan),
1165 ANEURALNETWORKS_NO_ERROR);
1166 ASSERT_EQ(plan.forTest_getKind(), ExecutionPlan::Kind::SIMPLE);
1167 ASSERT_EQ(plan.forTest_simpleGetDevice()->getName(), expectDevice);
1168 };
1169
1170 ASSERT_NO_FATAL_FAILURE(TrivialTest(false, "f32"));
1171 ASSERT_NO_FATAL_FAILURE(TrivialTest(true, "f16"));
1172 }
1173
1174 } // namespace
1175