aboutsummaryrefslogtreecommitdiff
path: root/Utils.cpp
blob: 00d61c7bc6bf743c3b7d1dd3c01fb4b4fea2ccfb (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
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//

#define LOG_TAG "ArmnnDriver"

#include "Utils.hpp"
#include "Half.hpp"

#include <armnnUtils/Permute.hpp>

#include <armnn/Utils.hpp>

#include <cassert>
#include <cerrno>
#include <cinttypes>
#include <sstream>
#include <cstdio>
#include <time.h>



using namespace android;
using namespace android::hardware;
using namespace android::hidl::memory::V1_0;

namespace armnn_driver
{
const armnn::PermutationVector g_DontPermute{};

namespace
{

void SwizzleAndroidNn4dTensorToArmNn(const armnn::TensorShape& inTensorShape, const void* input,
                                     void* output, size_t dataTypeSize, const armnn::PermutationVector& mappings)
{
    assert(inTensorShape.GetNumDimensions() == 4U);

    armnnUtils::Permute(armnnUtils::Permuted(inTensorShape, mappings), mappings, input, output, dataTypeSize);
}

} // anonymous namespace

void SwizzleAndroidNn4dTensorToArmNn(const armnn::TensorInfo& tensor, const void* input, void* output,
                                     const armnn::PermutationVector& mappings)
{
    assert(tensor.GetNumDimensions() == 4U);

    armnn::DataType dataType = tensor.GetDataType();
    switch (dataType)
    {
    case armnn::DataType::Float16:
    case armnn::DataType::Float32:
    case armnn::DataType::QAsymmU8:
    case armnn::DataType::QSymmS8:
    case armnn::DataType::QAsymmS8:
        SwizzleAndroidNn4dTensorToArmNn(tensor.GetShape(), input, output, armnn::GetDataTypeSize(dataType), mappings);
        break;
    default:
        ALOGW("Unknown armnn::DataType for swizzling");
        assert(0);
    }
}

void* GetMemoryFromPool(DataLocation location, const std::vector<android::nn::RunTimePoolInfo>& memPools)
{
    // find the location within the pool
    assert(location.poolIndex < memPools.size());

    const android::nn::RunTimePoolInfo& memPool = memPools[location.poolIndex];

    uint8_t* memPoolBuffer = memPool.getBuffer();

    uint8_t* memory = memPoolBuffer + location.offset;

    return memory;
}

armnn::TensorInfo GetTensorInfoForOperand(const V1_0::Operand& operand)
{
    armnn::DataType type;

    switch (operand.type)
    {
        case V1_0::OperandType::TENSOR_FLOAT32:
            type = armnn::DataType::Float32;
            break;
        case V1_0::OperandType::TENSOR_QUANT8_ASYMM:
            type = armnn::DataType::QAsymmU8;
            break;
        case V1_0::OperandType::TENSOR_INT32:
            type = armnn::DataType::Signed32;
            break;
        default:
            throw UnsupportedOperand<V1_0::OperandType>(operand.type);
    }

    armnn::TensorInfo ret(operand.dimensions.size(), operand.dimensions.data(), type);

    ret.SetQuantizationScale(operand.scale);
    ret.SetQuantizationOffset(operand.zeroPoint);

    return ret;
}

#if defined(ARMNN_ANDROID_NN_V1_2) || defined(ARMNN_ANDROID_NN_V1_3)// Using ::android::hardware::neuralnetworks::V1_2

armnn::TensorInfo GetTensorInfoForOperand(const V1_2::Operand& operand)
{
    using namespace armnn;
    bool perChannel = false;

    DataType type;
    switch (operand.type)
    {
        case V1_2::OperandType::TENSOR_BOOL8:
            type = armnn::DataType::Boolean;
            break;
        case V1_2::OperandType::TENSOR_FLOAT32:
            type = armnn::DataType::Float32;
            break;
        case V1_2::OperandType::TENSOR_FLOAT16:
            type = armnn::DataType::Float16;
            break;
        case V1_2::OperandType::TENSOR_QUANT8_ASYMM:
            type = armnn::DataType::QAsymmU8;
            break;
        case V1_2::OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL:
            perChannel=true;
            ARMNN_FALLTHROUGH;
        case V1_2::OperandType::TENSOR_QUANT8_SYMM:
            type = armnn::DataType::QSymmS8;
            break;
        case V1_2::OperandType::TENSOR_QUANT16_SYMM:
            type = armnn::DataType::QSymmS16;
            break;
        case V1_2::OperandType::TENSOR_INT32:
            type = armnn::DataType::Signed32;
            break;
        default:
            throw UnsupportedOperand<V1_2::OperandType>(operand.type);
    }

    TensorInfo ret(operand.dimensions.size(), operand.dimensions.data(), type);
    if (perChannel)
    {
        // ExtraParams is expected to be of type channelQuant
        BOOST_ASSERT(operand.extraParams.getDiscriminator() ==
                     V1_2::Operand::ExtraParams::hidl_discriminator::channelQuant);

        auto perAxisQuantParams = operand.extraParams.channelQuant();

        ret.SetQuantizationScales(perAxisQuantParams.scales);
        ret.SetQuantizationDim(MakeOptional<unsigned int>(perAxisQuantParams.channelDim));
    }
    else
    {
        ret.SetQuantizationScale(operand.scale);
        ret.SetQuantizationOffset(operand.zeroPoint);
    }

    return ret;
}

#endif

#ifdef ARMNN_ANDROID_NN_V1_3 // Using ::android::hardware::neuralnetworks::V1_3

armnn::TensorInfo GetTensorInfoForOperand(const V1_3::Operand& operand)
{
    using namespace armnn;
    bool perChannel = false;

    DataType type;
    switch (operand.type)
    {
        case V1_3::OperandType::TENSOR_BOOL8:
            type = armnn::DataType::Boolean;
            break;
        case V1_3::OperandType::TENSOR_FLOAT32:
            type = armnn::DataType::Float32;
            break;
        case V1_3::OperandType::TENSOR_FLOAT16:
            type = armnn::DataType::Float16;
            break;
        case V1_3::OperandType::TENSOR_QUANT8_ASYMM:
            type = armnn::DataType::QAsymmU8;
            break;
        case V1_3::OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL:
            perChannel=true;
            ARMNN_FALLTHROUGH;
        case V1_3::OperandType::TENSOR_QUANT8_SYMM:
            type = armnn::DataType::QSymmS8;
            break;
        case V1_3::OperandType::TENSOR_QUANT16_SYMM:
            type = armnn::DataType::QSymmS16;
            break;
        case V1_3::OperandType::TENSOR_INT32:
            type = armnn::DataType::Signed32;
            break;
        case V1_3::OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
            type = armnn::DataType::QAsymmS8;
            break;
        default:
            throw UnsupportedOperand<V1_3::OperandType>(operand.type);
    }

    TensorInfo ret(operand.dimensions.size(), operand.dimensions.data(), type);
    if (perChannel)
    {
        // ExtraParams is expected to be of type channelQuant
        BOOST_ASSERT(operand.extraParams.getDiscriminator() ==
                     V1_2::Operand::ExtraParams::hidl_discriminator::channelQuant);

        auto perAxisQuantParams = operand.extraParams.channelQuant();

        ret.SetQuantizationScales(perAxisQuantParams.scales);
        ret.SetQuantizationDim(MakeOptional<unsigned int>(perAxisQuantParams.channelDim));
    }
    else
    {
        ret.SetQuantizationScale(operand.scale);
        ret.SetQuantizationOffset(operand.zeroPoint);
    }

    return ret;
}

#endif

std::string GetOperandSummary(const V1_0::Operand& operand)
{
    return android::hardware::details::arrayToString(operand.dimensions, operand.dimensions.size()) + " " +
        toString(operand.type);
}

#if defined(ARMNN_ANDROID_NN_V1_2) || defined(ARMNN_ANDROID_NN_V1_3) // Using ::android::hardware::neuralnetworks::V1_2

std::string GetOperandSummary(const V1_2::Operand& operand)
{
    return android::hardware::details::arrayToString(operand.dimensions, operand.dimensions.size()) + " " +
           toString(operand.type);
}

#endif

#ifdef ARMNN_ANDROID_NN_V1_3 // Using ::android::hardware::neuralnetworks::V1_3

std::string GetOperandSummary(const V1_3::Operand& operand)
{
    return android::hardware::details::arrayToString(operand.dimensions, operand.dimensions.size()) + " " +
           toString(operand.type);
}

#endif

using DumpElementFunction = void (*)(const armnn::ConstTensor& tensor,
    unsigned int elementIndex,
    std::ofstream& fileStream);

namespace
{
template <typename ElementType, typename PrintableType = ElementType>
void DumpTensorElement(const armnn::ConstTensor& tensor, unsigned int elementIndex, std::ofstream& fileStream)
{
    const ElementType* elements = reinterpret_cast<const ElementType*>(tensor.GetMemoryArea());
    fileStream << static_cast<PrintableType>(elements[elementIndex]) << ",";
}

constexpr const char* MemoryLayoutString(const armnn::ConstTensor& tensor)
{
    const char* str = "";

    switch (tensor.GetNumDimensions())
    {
        case 4:  { str = "(BHWC) "; break; }
        case 3:  { str = "(HWC) "; break; }
        case 2:  { str = "(HW) "; break; }
        default: { str = ""; break; }
    }

    return str;
}
} // namespace

void DumpTensor(const std::string& dumpDir,
    const std::string& requestName,
    const std::string& tensorName,
    const armnn::ConstTensor& tensor)
{
    // The dump directory must exist in advance.
    const std::string fileName = boost::str(boost::format("%1%/%2%_%3%.dump") % dumpDir % requestName % tensorName);

    std::ofstream fileStream;
    fileStream.open(fileName, std::ofstream::out | std::ofstream::trunc);

    if (!fileStream.good())
    {
        ALOGW("Could not open file %s for writing", fileName.c_str());
        return;
    }

    DumpElementFunction dumpElementFunction = nullptr;

    switch (tensor.GetDataType())
    {
        case armnn::DataType::Float32:
        {
            dumpElementFunction = &DumpTensorElement<float>;
            break;
        }
        case armnn::DataType::QAsymmU8:
        {
            dumpElementFunction = &DumpTensorElement<uint8_t, uint32_t>;
            break;
        }
        case armnn::DataType::Signed32:
        {
            dumpElementFunction = &DumpTensorElement<int32_t>;
            break;
        }
        case armnn::DataType::Float16:
        {
            dumpElementFunction = &DumpTensorElement<armnn::Half>;
            break;
        }
        default:
        {
            dumpElementFunction = nullptr;
        }
    }

    if (dumpElementFunction != nullptr)
    {
        const unsigned int numDimensions = tensor.GetNumDimensions();

        const unsigned int batch = (numDimensions == 4) ? tensor.GetShape()[numDimensions - 4] : 1;

        const unsigned int height = (numDimensions >= 3)
                                    ? tensor.GetShape()[numDimensions - 3]
                                    : (numDimensions >= 2) ? tensor.GetShape()[numDimensions - 2] : 1;

        const unsigned int width = (numDimensions >= 3)
                                   ? tensor.GetShape()[numDimensions - 2]
                                   : (numDimensions >= 1) ? tensor.GetShape()[numDimensions - 1] : 0;

        const unsigned int channels = (numDimensions >= 3) ? tensor.GetShape()[numDimensions - 1] : 1;

        fileStream << "# Number of elements " << tensor.GetNumElements() << std::endl;
        fileStream << "# Dimensions " << MemoryLayoutString(tensor);
        fileStream << "[" << tensor.GetShape()[0];
        for (unsigned int d = 1; d < numDimensions; d++)
        {
            fileStream << "," << tensor.GetShape()[d];
        }
        fileStream << "]" << std::endl;

        for (unsigned int e = 0, b = 0; b < batch; ++b)
        {
            if (numDimensions >= 4)
            {
                fileStream << "# Batch " << b << std::endl;
            }
            for (unsigned int c = 0; c < channels; c++)
            {
                if (numDimensions >= 3)
                {
                    fileStream << "# Channel " << c << std::endl;
                }
                for (unsigned int h = 0; h < height; h++)
                {
                    for (unsigned int w = 0; w < width; w++, e += channels)
                    {
                        (*dumpElementFunction)(tensor, e, fileStream);
                    }
                    fileStream << std::endl;
                }
                e -= channels - 1;
                if (c < channels)
                {
                    e -= ((height * width) - 1) * channels;
                }
            }
            fileStream << std::endl;
        }
        fileStream << std::endl;
    }
    else
    {
        fileStream << "Cannot dump tensor elements: Unsupported data type "
            << static_cast<unsigned int>(tensor.GetDataType()) << std::endl;
    }

    if (!fileStream.good())
    {
        ALOGW("An error occurred when writing to file %s", fileName.c_str());
    }
}

void DumpJsonProfilingIfRequired(bool gpuProfilingEnabled,
                                 const std::string& dumpDir,
                                 armnn::NetworkId networkId,
                                 const armnn::IProfiler* profiler)
{
    // Check if profiling is required.
    if (!gpuProfilingEnabled)
    {
        return;
    }

    // The dump directory must exist in advance.
    if (dumpDir.empty())
    {
        return;
    }

    BOOST_ASSERT(profiler);

    // Set the name of the output profiling file.
    const std::string fileName = boost::str(boost::format("%1%/%2%_%3%.json")
                                            % dumpDir
                                            % std::to_string(networkId)
                                            % "profiling");

    // Open the ouput file for writing.
    std::ofstream fileStream;
    fileStream.open(fileName, std::ofstream::out | std::ofstream::trunc);

    if (!fileStream.good())
    {
        ALOGW("Could not open file %s for writing", fileName.c_str());
        return;
    }

    // Write the profiling info to a JSON file.
    profiler->Print(fileStream);
}

std::string ExportNetworkGraphToDotFile(const armnn::IOptimizedNetwork& optimizedNetwork,
                                        const std::string& dumpDir)
{
    std::string fileName;
    // The dump directory must exist in advance.
    if (dumpDir.empty())
    {
        return fileName;
    }

    std::string timestamp = GetFileTimestamp();
    if (timestamp.empty())
    {
        return fileName;
    }

    // Set the name of the output .dot file.
    fileName = boost::str(boost::format("%1%/%2%_networkgraph.dot")
                          % dumpDir
                          % timestamp);

    ALOGV("Exporting the optimized network graph to file: %s", fileName.c_str());

    // Write the network graph to a dot file.
    std::ofstream fileStream;
    fileStream.open(fileName, std::ofstream::out | std::ofstream::trunc);

    if (!fileStream.good())
    {
        ALOGW("Could not open file %s for writing", fileName.c_str());
        return fileName;
    }

    if (optimizedNetwork.SerializeToDot(fileStream) != armnn::Status::Success)
    {
        ALOGW("An error occurred when writing to file %s", fileName.c_str());
    }
    return fileName;
}

bool IsDynamicTensor(const armnn::TensorInfo& outputInfo)
{
    // Dynamic tensors have at least one 0-sized dimension
    return outputInfo.GetNumElements() == 0u;
}

std::string GetFileTimestamp()
{
    // used to get a timestamp to name diagnostic files (the ArmNN serialized graph
    // and getSupportedOperations.txt files)
    timespec ts;
    int iRet = clock_gettime(CLOCK_MONOTONIC_RAW, &ts);
    std::stringstream ss;
    if (iRet == 0)
    {
        ss << std::to_string(ts.tv_sec) << "_" << std::to_string(ts.tv_nsec);
    }
    else
    {
        ALOGW("clock_gettime failed with errno %s : %s", std::to_string(errno).c_str(), std::strerror(errno));
    }
    return ss.str();
}

void RenameGraphDotFile(const std::string& oldName, const std::string& dumpDir, const armnn::NetworkId networkId)
{
    if (dumpDir.empty())
    {
        return;
    }
    if (oldName.empty())
    {
        return;
    }
    const std::string newFileName = boost::str(boost::format("%1%/%2%_networkgraph.dot")
                                               % dumpDir
                                               % std::to_string(networkId));
    int iRet = rename(oldName.c_str(), newFileName.c_str());
    if (iRet != 0)
    {
        std::stringstream ss;
        ss << "rename of [" << oldName << "] to [" << newFileName << "] failed with errno " << std::to_string(errno)
           << " : " << std::strerror(errno);
        ALOGW(ss.str().c_str());
    }
}

void CommitPools(std::vector<::android::nn::RunTimePoolInfo>& memPools)
{
    if (memPools.empty())
    {
        return;
    }
    // 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 (auto& pool : memPools)
    {
        // Type android::nn::RunTimePoolInfo has changed between Android P & Q and Android R, where
        // update() has been removed and flush() added.
#if defined(ARMNN_ANDROID_R) // Use the new Android implementation.
        pool.flush();
#else
        pool.update();
#endif
    }
}



} // namespace armnn_driver