aboutsummaryrefslogtreecommitdiff
path: root/ArmnnPreparedModel.cpp
blob: 2f1abef79cc44c8aabaac7a55ec490e026b43e9d (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//

#define LOG_TAG "ArmnnDriver"

#include "ArmnnPreparedModel.hpp"
#include "Utils.hpp"

#include <boost/format.hpp>
#include <log/log.h>
#include <OperationsUtils.h>

#if defined(ARMNN_ANDROID_P) || defined(ARMNN_ANDROID_Q)
// The headers of the ML framework have changed between Android O and Android P.
// The validation functions have been moved into their own header, ValidateHal.h.
#include <ValidateHal.h>
#endif

#include <cassert>
#include <cinttypes>

using namespace android;

namespace
{
using namespace armnn_driver;

void NotifyCallbackAndCheck(const ::android::sp<V1_0::IExecutionCallback>& callback, ErrorStatus errorStatus,
                            std::string callingFunction)
{
    Return<void> returned = callback->notify(errorStatus);
    // This check is required, if the callback fails and it isn't checked it will bring down the service
    if (!returned.isOk())
    {
        ALOGE("ArmnnDriver::%s: hidl callback failed to return properly: %s",
            callingFunction.c_str(), returned.description().c_str());
    }
}

bool ValidateRequestArgument(const RequestArgument& requestArg, const armnn::TensorInfo& tensorInfo)
{
    if (requestArg.dimensions.size() != 0)
    {
        if (requestArg.dimensions.size() != tensorInfo.GetNumDimensions())
        {
            ALOGE("Mismatched dimensions (request argument: %zu, expected: %u)",
                  requestArg.dimensions.size(), tensorInfo.GetNumDimensions());
            return false;
        }

        for (unsigned int d = 0; d < tensorInfo.GetNumDimensions(); ++d)
        {
            if (requestArg.dimensions[d] != tensorInfo.GetShape()[d])
            {
                ALOGE("Mismatched size for dimension %d (request argument: %u, expected %u)",
                    d, requestArg.dimensions[d], tensorInfo.GetShape()[d]);
                return false;
            }
        }
    }

    return true;
}

armnn::Tensor GetTensorForRequestArgument(const RequestArgument& requestArg,
    const armnn::TensorInfo& tensorInfo,
    const std::vector<::android::nn::RunTimePoolInfo>& requestPools)
{
    if (!ValidateRequestArgument(requestArg, tensorInfo))
    {
        return armnn::Tensor();
    }

    return armnn::Tensor(tensorInfo, GetMemoryFromPool(requestArg.location, requestPools));
}

inline std::string BuildTensorName(const char* tensorNamePrefix, std::size_t index)
{
    return tensorNamePrefix + std::to_string(index);
}

} // anonymous namespace

using namespace android::hardware;

namespace armnn_driver
{
template<typename HalVersion>
RequestThread<ArmnnPreparedModel, HalVersion, ArmnnCallback_1_0> ArmnnPreparedModel<HalVersion>::m_RequestThread;

template<typename HalVersion>
template <typename TensorBindingCollection>
void ArmnnPreparedModel<HalVersion>::DumpTensorsIfRequired(char const* tensorNamePrefix,
                                                           const TensorBindingCollection& tensorBindings)
{
    if (!m_RequestInputsAndOutputsDumpDir.empty())
    {
        const std::string requestName = boost::str(boost::format("%1%_%2%.dump") % m_NetworkId % m_RequestCount);
        for (std::size_t i = 0u; i < tensorBindings.size(); ++i)
        {
            DumpTensor(m_RequestInputsAndOutputsDumpDir,
                requestName,
                BuildTensorName(tensorNamePrefix, i),
                tensorBindings[i].second);
        }
    }
}

template<typename HalVersion>
ArmnnPreparedModel<HalVersion>::ArmnnPreparedModel(armnn::NetworkId networkId,
                                                   armnn::IRuntime* runtime,
                                                   const HalModel& model,
                                                   const std::string& requestInputsAndOutputsDumpDir,
                                                   const bool gpuProfilingEnabled)
    : m_NetworkId(networkId)
    , m_Runtime(runtime)
    , m_Model(model)
    , m_RequestCount(0)
    , m_RequestInputsAndOutputsDumpDir(requestInputsAndOutputsDumpDir)
    , m_GpuProfilingEnabled(gpuProfilingEnabled)
{
    // Enable profiling if required.
    m_Runtime->GetProfiler(m_NetworkId)->EnableProfiling(m_GpuProfilingEnabled);
}

template<typename HalVersion>
ArmnnPreparedModel<HalVersion>::~ArmnnPreparedModel()
{
    // Get a hold of the profiler used by this model.
    std::shared_ptr<armnn::IProfiler> profiler = m_Runtime->GetProfiler(m_NetworkId);

    // Unload the network associated with this model.
    m_Runtime->UnloadNetwork(m_NetworkId);

    // Dump the profiling info to a file if required.
    DumpJsonProfilingIfRequired(m_GpuProfilingEnabled, m_RequestInputsAndOutputsDumpDir, m_NetworkId, profiler.get());
}

template<typename HalVersion>
Return<ErrorStatus> ArmnnPreparedModel<HalVersion>::execute(const Request& request,
                                                            const ::android::sp<V1_0::IExecutionCallback>& callback)
{
    ALOGV("ArmnnPreparedModel::execute(): %s", GetModelSummary(m_Model).c_str());
    m_RequestCount++;

    if (callback.get() == nullptr) {
        ALOGE("ArmnnPreparedModel::execute invalid callback passed");
        return ErrorStatus::INVALID_ARGUMENT;
    }

    if (!android::nn::validateRequest(request, m_Model))
    {
        NotifyCallbackAndCheck(callback, ErrorStatus::INVALID_ARGUMENT, "ArmnnPreparedModel::execute");
        return ErrorStatus::INVALID_ARGUMENT;
    }

    if (!m_RequestInputsAndOutputsDumpDir.empty())
    {
        ALOGD("Dumping inputs and outputs for request %" PRIuPTR, reinterpret_cast<std::uintptr_t>(callback.get()));
    }

    // allocate the tensors on the heap, as they are passed to the request thread
    auto pInputTensors = std::make_shared<armnn::InputTensors>();
    auto pOutputTensors = std::make_shared<armnn::OutputTensors>();

    // map the memory pool into shared pointers
    // use a shared memory pools vector on the heap, as it is passed to the request thread
    auto pMemPools = std::make_shared<std::vector<android::nn::RunTimePoolInfo>>();
    if (!setRunTimePoolInfosFromHidlMemories(pMemPools.get(), request.pools))
    {
        NotifyCallbackAndCheck(callback, ErrorStatus::GENERAL_FAILURE, "ArmnnPreparedModel::execute");
        return ErrorStatus::GENERAL_FAILURE;
    }

    // add the inputs and outputs with their data
    try
    {
        pInputTensors->reserve(request.inputs.size());
        for (unsigned int i = 0; i < request.inputs.size(); i++)
        {
            const auto& inputArg = request.inputs[i];

            const armnn::TensorInfo inputTensorInfo = m_Runtime->GetInputTensorInfo(m_NetworkId, i);
            const armnn::Tensor inputTensor = GetTensorForRequestArgument(inputArg, inputTensorInfo, *pMemPools);
            if (inputTensor.GetMemoryArea() == nullptr)
            {
                ALOGE("Cannot execute request. Error converting request input %u to tensor", i);
                return ErrorStatus::GENERAL_FAILURE;
            }

            pInputTensors->emplace_back(i, inputTensor);
        }

        pOutputTensors->reserve(request.outputs.size());
        for (unsigned int i = 0; i < request.outputs.size(); i++)
        {
            const auto& outputArg = request.outputs[i];

            const armnn::TensorInfo outputTensorInfo = m_Runtime->GetOutputTensorInfo(m_NetworkId, i);
            const armnn::Tensor outputTensor = GetTensorForRequestArgument(outputArg, outputTensorInfo, *pMemPools);
            if (outputTensor.GetMemoryArea() == nullptr)
            {
                ALOGE("Cannot execute request. Error converting request output %u to tensor", i);
                return ErrorStatus::GENERAL_FAILURE;
            }

            pOutputTensors->emplace_back(i, outputTensor);
        }
    }
    catch (std::exception& e)
    {
        ALOGW("Exception caught while preparing for EnqueueWorkload: %s", e.what());
        NotifyCallbackAndCheck(callback, ErrorStatus::GENERAL_FAILURE, "ArmnnPreparedModel::execute");
        return ErrorStatus::GENERAL_FAILURE;
    }

    ALOGV("ArmnnPreparedModel::execute(...) before PostMsg");

    auto cb = [callback](ErrorStatus errorStatus, std::string callingFunction)
    {
        NotifyCallbackAndCheck(callback, errorStatus, callingFunction);
    };

    ArmnnCallback_1_0 armnnCb;
    armnnCb.callback = cb;
    // post the request for asynchronous execution
    m_RequestThread.PostMsg(this, pMemPools, pInputTensors, pOutputTensors, armnnCb);
    ALOGV("ArmnnPreparedModel::execute(...) after PostMsg");
    return ErrorStatus::NONE; // successfully queued
}

template<typename HalVersion>
void ArmnnPreparedModel<HalVersion>::ExecuteGraph(
        std::shared_ptr<std::vector<::android::nn::RunTimePoolInfo>>& pMemPools,
        std::shared_ptr<armnn::InputTensors>& pInputTensors,
        std::shared_ptr<armnn::OutputTensors>& pOutputTensors,
        ArmnnCallback_1_0 cb)
{
    ALOGV("ArmnnPreparedModel::ExecuteGraph(...)");

    DumpTensorsIfRequired("Input", *pInputTensors);

    // run it
    try
    {
        armnn::Status status = m_Runtime->EnqueueWorkload(m_NetworkId, *pInputTensors, *pOutputTensors);
        if (status != armnn::Status::Success)
        {
            ALOGW("EnqueueWorkload failed");
            cb.callback(ErrorStatus::GENERAL_FAILURE, "ArmnnPreparedModel::ExecuteGraph");
            return;
        }
    }
    catch (std::exception& e)
    {
        ALOGW("Exception caught from EnqueueWorkload: %s", e.what());
        cb.callback(ErrorStatus::GENERAL_FAILURE, "ArmnnPreparedModel::ExecuteGraph");
        return;
    }

    DumpTensorsIfRequired("Output", *pOutputTensors);

    // Commit output buffers.
    // Note that we update *all* pools, even if they aren't actually used as outputs -
    // this is simpler and is what the CpuExecutor does.
    for (android::nn::RunTimePoolInfo& pool : *pMemPools)
    {
        pool.update();
    }

    cb.callback(ErrorStatus::NONE, "ExecuteGraph");
}

template<typename HalVersion>
bool ArmnnPreparedModel<HalVersion>::ExecuteWithDummyInputs()
{
    std::vector<std::vector<char>> storage;
    armnn::InputTensors inputTensors;
    for (unsigned int i = 0; i < m_Model.inputIndexes.size(); i++)
    {
        const armnn::TensorInfo inputTensorInfo = m_Runtime->GetInputTensorInfo(m_NetworkId, i);
        storage.emplace_back(inputTensorInfo.GetNumBytes());
        const armnn::ConstTensor inputTensor(inputTensorInfo, storage.back().data());

        inputTensors.emplace_back(i, inputTensor);
    }

    armnn::OutputTensors outputTensors;
    for (unsigned int i = 0; i < m_Model.outputIndexes.size(); i++)
    {
        const armnn::TensorInfo outputTensorInfo = m_Runtime->GetOutputTensorInfo(m_NetworkId, i);
        storage.emplace_back(outputTensorInfo.GetNumBytes());
        const armnn::Tensor outputTensor(outputTensorInfo, storage.back().data());

        outputTensors.emplace_back(i, outputTensor);
    }

    try
    {
        armnn::Status status = m_Runtime->EnqueueWorkload(m_NetworkId, inputTensors, outputTensors);
        if (status != armnn::Status::Success)
        {
            ALOGW("ExecuteWithDummyInputs: EnqueueWorkload failed");
            return false;
        }
    }
    catch (std::exception& e)
    {
        ALOGW("ExecuteWithDummyInputs: Exception caught from EnqueueWorkload: %s", e.what());
        return false;
    }
    return true;
}

///
/// Class template specializations
///

template class ArmnnPreparedModel<hal_1_0::HalPolicy>;

#ifdef ARMNN_ANDROID_NN_V1_1
template class ArmnnPreparedModel<hal_1_1::HalPolicy>;
#endif

#ifdef ARMNN_ANDROID_NN_V1_2
template class ArmnnPreparedModel<hal_1_1::HalPolicy>;
template class ArmnnPreparedModel<hal_1_2::HalPolicy>;
#endif
} // namespace armnn_driver