aboutsummaryrefslogtreecommitdiff
path: root/1.3/ArmnnDriverImpl.cpp
blob: 6a69765caee387217944923d3cdb1d7f4559ef1a (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
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
//
// Copyright © 2020 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//

#include "ArmnnDriverImpl.hpp"
#include "../ArmnnPreparedModel_1_3.hpp"
#include "../ModelToINetworkConverter.hpp"
#include "../SystemPropertiesUtils.hpp"

#include <log/log.h>

namespace
{
const char *g_RelaxedFloat32toFloat16PerformanceExecTime    = "ArmNN.relaxedFloat32toFloat16Performance.execTime";
const char *g_RelaxedFloat32toFloat16PerformancePowerUsage  = "ArmNN.relaxedFloat32toFloat16Performance.powerUsage";

const char *g_ifPerformanceExecTime                         = "ArmNN.ifPerformance.execTime";
const char *g_ifPerformancePowerUsage                       = "ArmNN.ifPerformance.powerUsage";

const char *g_whilePerformanceExecTime                      = "ArmNN.whilePerformance.execTime";
const char *g_whilePerformancePowerUsage                    = "ArmNN.whilePerformance.powerUsage";

const char *g_OperandTypeTensorFloat32PerformanceExecTime   = "Armnn.operandTypeTensorFloat32Performance.execTime";
const char *g_OperandTypeTensorFloat32PerformancePowerUsage = "Armnn.operandTypeTensorFloat32Performance.powerUsage";

const char *g_OperandTypeFloat32PerformanceExecTime         = "Armnn.operandTypeFloat32Performance.execTime";
const char *g_OperandTypeFloat32PerformancePowerUsage       = "Armnn.operandTypeFloat32Performance.powerUsage";

const char *g_OperandTypeTensorFloat16PerformanceExecTime   = "Armnn.operandTypeTensorFloat16Performance.execTime";
const char *g_OperandTypeTensorFloat16PerformancePowerUsage = "Armnn.operandTypeTensorFloat16Performance.powerUsage";

const char *g_OperandTypeFloat16PerformanceExecTime         = "Armnn.operandTypeFloat16Performance.execTime";
const char *g_OperandTypeFloat16PerformancePowerUsage       = "Armnn.operandTypeFloat16Performance.powerUsage";

const char *g_OperandTypeTensorQuant8AsymmPerformanceExecTime =
        "Armnn.operandTypeTensorQuant8AsymmPerformance.execTime";
const char *g_OperandTypeTensorQuant8AsymmPerformancePowerUsage =
        "Armnn.operandTypeTensorQuant8AsymmPerformance.powerUsage";

const char *g_OperandTypeTensorQuant8AsymmSignedPerformanceExecTime =
    "Armnn.operandTypeTensorQuant8AsymmSignedPerformance.execTime";
const char *g_OperandTypeTensorQuant8AsymmSignedPerformancePowerUsage =
    "Armnn.operandTypeTensorQuant8AsymmSignedPerformance.powerUsage";

const char *g_OperandTypeTensorQuant16SymmPerformanceExecTime =
        "Armnn.operandTypeTensorQuant16SymmPerformance.execTime";
const char *g_OperandTypeTensorQuant16SymmPerformancePowerUsage =
        "Armnn.operandTypeTensorQuant16SymmPerformance.powerUsage";

const char *g_OperandTypeTensorQuant8SymmPerformanceExecTime =
        "Armnn.operandTypeTensorQuant8SymmPerformance.execTime";
const char *g_OperandTypeTensorQuant8SymmPerformancePowerUsage =
        "Armnn.operandTypeTensorQuant8SymmPerformance.powerUsage";

const char *g_OperandTypeTensorQuant8SymmPerChannelPerformanceExecTime =
    "Armnn.operandTypeTensorQuant8SymmPerChannelPerformance.execTime";
const char *g_OperandTypeTensorQuant8SymmPerChannelPerformancePowerUsage =
    "Armnn.operandTypeTensorQuant8SymmPerChannelPerformance.powerUsage";


const char *g_OperandTypeTensorInt32PerformanceExecTime     = "Armnn.operandTypeTensorInt32Performance.execTime";
const char *g_OperandTypeTensorInt32PerformancePowerUsage   = "Armnn.operandTypeTensorInt32Performance.powerUsage";

const char *g_OperandTypeInt32PerformanceExecTime           = "Armnn.operandTypeInt32Performance.execTime";
const char *g_OperandTypeInt32PerformancePowerUsage         = "Armnn.operandTypeInt32Performance.powerUsage";


void NotifyCallbackAndCheck(const android::sp<V1_3::IPreparedModelCallback>& callback,
                            V1_3::ErrorStatus errorStatus,
                            const android::sp<V1_3::IPreparedModel>& preparedModelPtr)
{
    Return<void> returned = callback->notify_1_3(errorStatus, preparedModelPtr);
    // This check is required, if the callback fails and it isn't checked it will bring down the service
    if (!returned.isOk())
    {
        ALOGE("ArmnnDriverImpl::prepareModel: hidl callback failed to return properly: %s ",
              returned.description().c_str());
    }
}

Return<V1_3::ErrorStatus> FailPrepareModel(V1_3::ErrorStatus error,
                                           const std::string& message,
                                           const android::sp<V1_3::IPreparedModelCallback>& callback)
{
    ALOGW("ArmnnDriverImpl::prepareModel: %s", message.c_str());
    NotifyCallbackAndCheck(callback, error, nullptr);
    return error;
}

} // anonymous namespace

namespace armnn_driver
{
namespace hal_1_3
{

Return<V1_3::ErrorStatus> ArmnnDriverImpl::prepareArmnnModel_1_3(
       const armnn::IRuntimePtr& runtime,
       const armnn::IGpuAccTunedParametersPtr& clTunedParameters,
       const DriverOptions& options,
       const V1_3::Model& model,
       const android::sp<V1_3::IPreparedModelCallback>& cb,
       bool float32ToFloat16,
       V1_3::Priority priority)
{
    ALOGV("ArmnnDriverImpl::prepareArmnnModel_1_3()");

    if (cb.get() == nullptr)
    {
        ALOGW("ArmnnDriverImpl::prepareModel: Invalid callback passed to prepareModel");
        return V1_3::ErrorStatus::INVALID_ARGUMENT;
    }

    if (!runtime)
    {
        return FailPrepareModel(V1_3::ErrorStatus::DEVICE_UNAVAILABLE, "Device unavailable", cb);
    }

    if (!android::nn::validateModel(model))
    {
        return FailPrepareModel(V1_3::ErrorStatus::INVALID_ARGUMENT, "Invalid model passed as input", cb);
    }

    // Deliberately ignore any unsupported operations requested by the options -
    // at this point we're being asked to prepare a model that we've already declared support for
    // and the operation indices may be different to those in getSupportedOperations anyway.
    std::set<unsigned int> unsupportedOperations;
    ModelToINetworkConverter<HalPolicy> modelConverter(options.GetBackends(),
                                                       model,
                                                       unsupportedOperations);

    if (modelConverter.GetConversionResult() != ConversionResult::Success)
    {
        FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, "ModelToINetworkConverter failed", cb);
        return V1_3::ErrorStatus::NONE;
    }

    // Serialize the network graph to a .armnn file if an output directory
    // has been specified in the drivers' arguments.
    auto serializedNetworkFileName =
        SerializeNetwork(*modelConverter.GetINetwork(), options.GetRequestInputsAndOutputsDumpDir());

    // Optimize the network
    armnn::IOptimizedNetworkPtr optNet(nullptr, nullptr);
    armnn::OptimizerOptions OptOptions;
    OptOptions.m_ReduceFp32ToFp16 = float32ToFloat16;

    armnn::BackendOptions gpuAcc("GpuAcc",
    {
        { "FastMathEnabled", options.IsFastMathEnabled() },
        { "SaveCachedNetwork", options.SaveCachedNetwork() },
        { "CachedNetworkFilePath", options.GetCachedNetworkFilePath() }
    });
    armnn::BackendOptions cpuAcc("CpuAcc",
    {
        { "FastMathEnabled", options.IsFastMathEnabled() },
        { "NumberOfThreads", options.GetNumberOfThreads() }
    });
    OptOptions.m_ModelOptions.push_back(gpuAcc);
    OptOptions.m_ModelOptions.push_back(cpuAcc);

    std::vector<std::string> errMessages;
    try
    {
        optNet = armnn::Optimize(*modelConverter.GetINetwork(),
                                 options.GetBackends(),
                                 runtime->GetDeviceSpec(),
                                 OptOptions,
                                 errMessages);
    }
    catch (std::exception& e)
    {
        std::stringstream message;
        message << "Exception (" << e.what() << ") caught from optimize.";
        FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, message.str(), cb);
        return V1_3::ErrorStatus::NONE;
    }

    // Check that the optimized network is valid.
    if (!optNet)
    {
        std::stringstream message;
        message << "Invalid optimized network";
        for (const std::string& msg : errMessages)
        {
            message << "\n" << msg;
        }
        FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, message.str(), cb);
        return V1_3::ErrorStatus::NONE;
    }

    // Export the optimized network graph to a dot file if an output dump directory
    // has been specified in the drivers' arguments.
    std::string dotGraphFileName = ExportNetworkGraphToDotFile(*optNet,
                                                               options.GetRequestInputsAndOutputsDumpDir());

    // Load it into the runtime.
    armnn::NetworkId netId = 0;
    try
    {
        if (runtime->LoadNetwork(netId, move(optNet)) != armnn::Status::Success)
        {
            return FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, "Network could not be loaded", cb);
        }
    }
    catch (std::exception& e)
    {
        std::stringstream message;
        message << "Exception (" << e.what()<< ") caught from LoadNetwork.";
        FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, message.str(), cb);
        return V1_3::ErrorStatus::NONE;
    }

    // Now that we have a networkId for the graph rename the exported files to use it
    // so that we can associate the graph file and the input/output tensor exported files
    RenameExportedFiles(serializedNetworkFileName,
                        dotGraphFileName,
                        options.GetRequestInputsAndOutputsDumpDir(),
                        netId);

    std::unique_ptr<ArmnnPreparedModel_1_3<hal_1_3::HalPolicy>> preparedModel(
            new ArmnnPreparedModel_1_3<hal_1_3::HalPolicy>(
                    netId,
                    runtime.get(),
                    model,
                    options.GetRequestInputsAndOutputsDumpDir(),
                    options.IsGpuProfilingEnabled(),
                    priority));

    // Run a single 'dummy' inference of the model. This means that CL kernels will get compiled (and tuned if
    // this is enabled) before the first 'real' inference which removes the overhead of the first inference.
    if (!preparedModel->ExecuteWithDummyInputs())
    {
        return FailPrepareModel(V1_3::ErrorStatus::GENERAL_FAILURE, "Network could not be executed", cb);
    }

    if (clTunedParameters &&
        options.GetClTunedParametersMode() == armnn::IGpuAccTunedParameters::Mode::UpdateTunedParameters)
    {
        // Now that we've done one inference the CL kernel parameters will have been tuned, so save the updated file.
        try
        {
            clTunedParameters->Save(options.GetClTunedParametersFile().c_str());
        }
        catch (std::exception& error)
        {
            ALOGE("ArmnnDriverImpl::prepareModel: Failed to save CL tuned parameters file '%s': %s",
                  options.GetClTunedParametersFile().c_str(), error.what());
        }
    }

    NotifyCallbackAndCheck(cb, V1_3::ErrorStatus::NONE, preparedModel.release());

    return V1_3::ErrorStatus::NONE;
}

Return<void> ArmnnDriverImpl::getCapabilities_1_3(const armnn::IRuntimePtr& runtime,
                                                  V1_3::IDevice::getCapabilities_1_3_cb cb)
{
    ALOGV("hal_1_3::ArmnnDriverImpl::getCapabilities()");

    V1_3::Capabilities capabilities;

    float defaultValue = .1f;

    if (runtime)
    {
        capabilities.relaxedFloat32toFloat16PerformanceScalar.execTime =
                ParseSystemProperty(g_RelaxedFloat32toFloat16PerformanceExecTime, defaultValue);

        capabilities.relaxedFloat32toFloat16PerformanceScalar.powerUsage =
                ParseSystemProperty(g_RelaxedFloat32toFloat16PerformancePowerUsage, defaultValue);

        capabilities.relaxedFloat32toFloat16PerformanceTensor.execTime =
                ParseSystemProperty(g_RelaxedFloat32toFloat16PerformanceExecTime, defaultValue);

        capabilities.relaxedFloat32toFloat16PerformanceTensor.powerUsage =
                ParseSystemProperty(g_RelaxedFloat32toFloat16PerformancePowerUsage, defaultValue);

        capabilities.ifPerformance.execTime =
                ParseSystemProperty(g_ifPerformanceExecTime, defaultValue);

        capabilities.ifPerformance.powerUsage =
                ParseSystemProperty(g_ifPerformancePowerUsage, defaultValue);

        capabilities.whilePerformance.execTime =
                ParseSystemProperty(g_whilePerformanceExecTime, defaultValue);

        capabilities.whilePerformance.powerUsage =
                ParseSystemProperty(g_whilePerformancePowerUsage, defaultValue);

        // Set the base value for all operand types
        capabilities.operandPerformance = nonExtensionOperandPerformance<HalVersion::V1_3>({FLT_MAX, FLT_MAX});

        // Load supported operand types
        update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_FLOAT32,
                {
                    .execTime = ParseSystemProperty(g_OperandTypeTensorFloat32PerformanceExecTime, defaultValue),
                    .powerUsage = ParseSystemProperty(g_OperandTypeTensorFloat32PerformancePowerUsage, defaultValue)
                });

        update(&capabilities.operandPerformance, V1_3::OperandType::FLOAT32,
                {
                    .execTime = ParseSystemProperty(g_OperandTypeFloat32PerformanceExecTime, defaultValue),
                    .powerUsage = ParseSystemProperty(g_OperandTypeFloat32PerformancePowerUsage, defaultValue)
                });

        update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_FLOAT16,
                {
                    .execTime = ParseSystemProperty(g_OperandTypeTensorFloat16PerformanceExecTime, defaultValue),
                    .powerUsage = ParseSystemProperty(g_OperandTypeTensorFloat16PerformancePowerUsage, defaultValue)
                });

        update(&capabilities.operandPerformance, V1_3::OperandType::FLOAT16,
                {
                    .execTime = ParseSystemProperty(g_OperandTypeFloat16PerformanceExecTime, defaultValue),
                    .powerUsage = ParseSystemProperty(g_OperandTypeFloat16PerformancePowerUsage, defaultValue)
                });

        update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_QUANT8_ASYMM,
                {
                    .execTime = ParseSystemProperty(g_OperandTypeTensorQuant8AsymmPerformanceExecTime, defaultValue),
                    .powerUsage = ParseSystemProperty(g_OperandTypeTensorQuant8AsymmPerformancePowerUsage, defaultValue)
                });

        update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_QUANT8_SYMM,
                {
                    .execTime = ParseSystemProperty(g_OperandTypeTensorQuant8SymmPerformanceExecTime, defaultValue),
                    .powerUsage = ParseSystemProperty(g_OperandTypeTensorQuant8SymmPerformancePowerUsage, defaultValue)
                });
        update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_QUANT8_ASYMM_SIGNED,
               {
                   .execTime = ParseSystemProperty(g_OperandTypeTensorQuant8AsymmSignedPerformanceExecTime,
                   defaultValue),
                   .powerUsage = ParseSystemProperty(g_OperandTypeTensorQuant8AsymmSignedPerformancePowerUsage,
                   defaultValue)
               });

        update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_QUANT16_SYMM,
                {
                    .execTime = ParseSystemProperty(g_OperandTypeTensorQuant16SymmPerformanceExecTime, defaultValue),
                    .powerUsage = ParseSystemProperty(g_OperandTypeTensorQuant16SymmPerformancePowerUsage, defaultValue)
                });

        update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL,
               {
                   .execTime =
                   ParseSystemProperty(g_OperandTypeTensorQuant8SymmPerChannelPerformanceExecTime, defaultValue),
                   .powerUsage =
                   ParseSystemProperty(g_OperandTypeTensorQuant8SymmPerChannelPerformancePowerUsage, defaultValue)
               });

        update(&capabilities.operandPerformance, V1_3::OperandType::TENSOR_INT32,
                {
                    .execTime = ParseSystemProperty(g_OperandTypeTensorInt32PerformanceExecTime, defaultValue),
                    .powerUsage = ParseSystemProperty(g_OperandTypeTensorInt32PerformancePowerUsage, defaultValue)
                });

        update(&capabilities.operandPerformance, V1_3::OperandType::INT32,
                {
                    .execTime = ParseSystemProperty(g_OperandTypeInt32PerformanceExecTime, defaultValue),
                    .powerUsage = ParseSystemProperty(g_OperandTypeInt32PerformancePowerUsage, defaultValue)
                });

        cb(V1_3::ErrorStatus::NONE, capabilities);
    }
    else
    {
        capabilities.relaxedFloat32toFloat16PerformanceScalar.execTime   = 0;
        capabilities.relaxedFloat32toFloat16PerformanceScalar.powerUsage = 0;
        capabilities.relaxedFloat32toFloat16PerformanceTensor.execTime   = 0;
        capabilities.relaxedFloat32toFloat16PerformanceTensor.powerUsage = 0;
        capabilities.ifPerformance.execTime      = 0;
        capabilities.ifPerformance.powerUsage    = 0;
        capabilities.whilePerformance.execTime   = 0;
        capabilities.whilePerformance.powerUsage = 0;

        // Set the base value for all operand types
        capabilities.operandPerformance = nonExtensionOperandPerformance<HalVersion::V1_3>({0.f, 0.0f});

        cb(V1_3::ErrorStatus::DEVICE_UNAVAILABLE, capabilities);
    }

    return Void();
}

} // namespace hal_1_3
} // namespace armnn_driver