forked from intel/nn-hal
-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathModelManager.cpp
339 lines (301 loc) · 14 KB
/
ModelManager.cpp
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
#include "ModelManager.h"
#undef LOG_TAG
#define LOG_TAG "ModelManager"
namespace android::hardware::neuralnetworks::nnhal {
bool NnapiModelInfo::updateOutputshapes(size_t outputIndex, std::vector<size_t>& outputDims,
bool isLengthSufficient) {
auto& outputShapeDims = mOutputShapes[outputIndex].dimensions;
mOutputShapes[outputIndex].isSufficient = isLengthSufficient;
if (outputDims.size() < outputShapeDims.size()) {
return false;
}
for (size_t i = 0; i < outputShapeDims.size(); i++) {
if (outputShapeDims[i] != outputDims[i]) {
ALOGD("%s Updating dim(%zu) at Output index(%zu)", __func__, i, outputIndex);
outputShapeDims[i] = outputDims[i];
}
}
return true;
}
bool NnapiModelInfo::initializeRunTimeOperandInfo() {
// initialize runtime operand info from model.
const size_t count = mModel.main.operands.size();
ALOGD("Operand size = %zu\n", count);
if (!count) {
ALOGE("NNERR:Operand Count is 0");
return false;
}
mOperands.resize(count);
mOutputShapes.resize(mModel.main.outputIndexes.size());
// Start by setting the runtime info to what's in the model.
for (size_t i = 0; i < count; i++) {
const Operand& from = mModel.main.operands[i];
dumpOperand(i, mModel.main);
RunTimeOperandInfo& to = mOperands[i];
to.dimensions.resize(from.dimensions.size());
for (size_t j = 0; j < from.dimensions.size(); j++) {
to.dimensions[j] = from.dimensions[j];
}
to.scale = from.scale;
ALOGV("OperandType = %d\n", from.type);
switch (from.type) {
case OperandType::TENSOR_FLOAT32:
case OperandType::FLOAT32:
to.type = OperandType::TENSOR_FLOAT32;
break;
case OperandType::INT32:
case OperandType::UINT32:
case OperandType::BOOL:
nnAssert(to.scale == 0);
FALLTHROUGH_INTENDED;
case OperandType::TENSOR_INT32:
to.type = from.type;
break;
case OperandType::TENSOR_FLOAT16:
case OperandType::TENSOR_QUANT16_SYMM:
case OperandType::TENSOR_QUANT16_ASYMM:
case OperandType::FLOAT16:
to.type = from.type;
break;
case OperandType::TENSOR_BOOL8:
to.type = from.type;
break;
case OperandType::TENSOR_QUANT8_ASYMM:
case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
case OperandType::TENSOR_QUANT8_SYMM:
case OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL:
to.type = from.type;
break;
default:
ALOGE("wrong operand type %d", from.type);
return false;
}
to.length = from.location.length;
to.lifetime = from.lifetime;
to.zeroPoint = from.zeroPoint;
switch (from.lifetime) {
case OperandLifeTime::TEMPORARY_VARIABLE:
to.buffer = nullptr;
to.length = sizeOfData(to.type, to.dimensions);
to.numberOfUsesLeft = from.numberOfConsumers;
break;
case OperandLifeTime::CONSTANT_COPY:
to.buffer = const_cast<uint8_t*>(&mModel.operandValues[from.location.offset]);
to.numberOfUsesLeft = 0;
break;
case OperandLifeTime::CONSTANT_REFERENCE: {
auto poolIndex = from.location.poolIndex;
nnAssert(poolIndex < mPoolInfos.size());
auto& r = mPoolInfos[poolIndex];
to.buffer = r.buffer + from.location.offset;
to.numberOfUsesLeft = 0;
break;
}
case OperandLifeTime::SUBGRAPH_INPUT:
case OperandLifeTime::SUBGRAPH_OUTPUT:
case OperandLifeTime::NO_VALUE:
to.buffer = nullptr;
to.numberOfUsesLeft = 0;
break;
default:
return false;
break;
}
}
for (uint32_t i = 0; i < mModel.main.outputIndexes.size(); i++) {
const uint32_t operandIndex = mModel.main.outputIndexes[i];
const RunTimeOperandInfo& from = mOperands[operandIndex];
mOutputShapes[i].dimensions = from.dimensions;
mOutputShapes[i].isSufficient = true;
}
return true;
}
// TODO: Move it to Utils class
template <typename T>
T NnapiModelInfo::GetConstFromBuffer(const uint8_t* buf, uint32_t len) {
// ALOGD("buf: %p, len: %d", buf, len);
if (len != sizeof(T)) {
ALOGE("fix me: typeid(T).name() should be %lu bytes", sizeof(T));
// fix me if buffer is of type float and if float and OperandLifeTime::CONSTANT_REFERENCE
nnAssert(false);
}
return *(T*)(buf);
}
const uint8_t* NnapiModelInfo::GetOperandMemory(int index, uint32_t& lenOut) {
ALOGV("%s", __func__);
const auto op = mModel.main.operands[index];
lenOut = op.location.length;
if (op.lifetime == OperandLifeTime::CONSTANT_COPY) {
ALOGV("operand lifetime OperandLifeTime::CONSTANT_COPY");
if (op.location.poolIndex != 0) {
// ALOGE("CONSTANT_COPY expects poolIndex to be 0");
nnAssert(false);
}
return (const_cast<uint8_t*>(&mModel.operandValues[op.location.offset]));
} else if (op.lifetime == OperandLifeTime::CONSTANT_REFERENCE) {
ALOGV("operand lifetime OperandLifeTime::CONSTANT_REFERENCE");
auto poolIndex = op.location.poolIndex;
auto& r = mPoolInfos[poolIndex];
return (const_cast<uint8_t*>(r.buffer + op.location.offset));
} else if (op.lifetime == OperandLifeTime::TEMPORARY_VARIABLE ||
op.lifetime == OperandLifeTime::SUBGRAPH_INPUT ||
op.lifetime == OperandLifeTime::SUBGRAPH_OUTPUT ||
op.lifetime == OperandLifeTime::NO_VALUE) {
// ALOGD(
// "operand lifetime "
// "OperandLifeTime::MODEL_INPUT||MODEL_OUTPUT||NO_VALUE||TEMPORARY_VARIABLE");
lenOut = sizeOfData(op.type, op.dimensions);
ALOGV("operand lifetime(%d), type(%d), lenOut(%d)", op.lifetime, op.type, lenOut);
return nullptr;
}
ALOGE("operand is expected to be const, but lifetime is %d", op.lifetime);
nnAssert(false); // temp fix since some time const operand set as TEMPORARY_VARIABLE
return nullptr;
}
V1_3::ErrorStatus NnapiModelInfo::setRunTimePoolInfosFromHidlMemories(
const hidl_vec<V1_3::Request::MemoryPool>& pools) {
ALOGV("Number of pools: %zu", pools.size());
mRequestPoolInfos.resize(pools.size());
for (size_t i = 0; i < pools.size(); i++) {
auto& poolInfo = mRequestPoolInfos[i];
switch (pools[i].getDiscriminator()) {
case V1_3::Request::MemoryPool::hidl_discriminator::hidlMemory:
if (!poolInfo.set(pools[i].hidlMemory())) {
ALOGE("Could not map memory pool !!!");
return V1_3::ErrorStatus::GENERAL_FAILURE;
}
break;
case V1_3::Request::MemoryPool::hidl_discriminator::token:
ALOGE(
"%s NNHAL 1.3 driver does not yet support driver buffer allocation. Returning "
"failure",
__func__);
return V1_3::ErrorStatus::INVALID_ARGUMENT;
break;
}
}
return V1_3::ErrorStatus::NONE;
}
ErrorStatus NnapiModelInfo::setRunTimePoolInfosFromHidlMemories(
const hidl_vec<hidl_memory>& pools) {
ALOGV("Number of pools: %zu", pools.size());
mRequestPoolInfos.resize(pools.size());
for (size_t i = 0; i < pools.size(); i++) {
auto& poolInfo = mRequestPoolInfos[i];
if (!poolInfo.set(pools[i])) {
ALOGE("Could not map memory pool !!!");
return ErrorStatus::GENERAL_FAILURE;
}
}
return ErrorStatus::NONE;
}
void* NnapiModelInfo::getBlobFromMemoryPoolIn(const Request& request, uint32_t index,
uint32_t& rBufferLength) {
RunTimeOperandInfo& operand = mOperands[mModel.main.inputIndexes[index]];
const V1_0::RequestArgument& arg = request.inputs[index];
auto poolIndex = arg.location.poolIndex;
nnAssert(poolIndex < mRequestPoolInfos.size());
auto& r = mRequestPoolInfos[poolIndex];
if (arg.dimensions.size() > 0) {
// It's the responsibility of the caller to validate that
// from.dimensions only modifies the dimensions that were
// unspecified in the model. That's the case in SampleDriver.cpp
// with the call to validateRequest().
operand.dimensions = arg.dimensions;
}
operand.buffer = r.buffer + arg.location.offset;
operand.length = arg.location.length;
ALOGV("%s Operand length:%d pointer:%p offset:%d pool index: %d", __func__, operand.length,
(r.buffer + arg.location.offset), arg.location.offset, poolIndex);
rBufferLength = operand.length;
return (r.buffer + arg.location.offset);
}
void* NnapiModelInfo::getBlobFromMemoryPoolIn(const V1_3::Request& request, uint32_t index,
uint32_t& rBufferLength) {
RunTimeOperandInfo& operand = mOperands[mModel.main.inputIndexes[index]];
const V1_0::RequestArgument& arg = request.inputs[index];
auto poolIndex = arg.location.poolIndex;
nnAssert(poolIndex < mRequestPoolInfos.size());
auto& r = mRequestPoolInfos[poolIndex];
if (arg.dimensions.size() > 0) {
// It's the responsibility of the caller to validate that
// from.dimensions only modifies the dimensions that were
// unspecified in the model. That's the case in SampleDriver.cpp
// with the call to validateRequest().
operand.dimensions = arg.dimensions;
}
operand.buffer = r.buffer + arg.location.offset;
operand.length = arg.location.length;
ALOGV("%s Operand length:%d pointer:%p offset:%d pool index: %d", __func__, operand.length,
(r.buffer + arg.location.offset), arg.location.offset, poolIndex);
rBufferLength = operand.length;
return (r.buffer + arg.location.offset);
}
void* NnapiModelInfo::getBlobFromMemoryPoolOut(const Request& request, uint32_t index,
uint32_t& rBufferLength) {
RunTimeOperandInfo& operand = mOperands[mModel.main.outputIndexes[index]];
const V1_0::RequestArgument& arg = request.outputs[index];
auto poolIndex = arg.location.poolIndex;
nnAssert(poolIndex < mRequestPoolInfos.size());
auto& r = mRequestPoolInfos[poolIndex];
ALOGV("%s lifetime:%d location offset:%d length:%d pool index:%d", __func__, operand.lifetime,
arg.location.offset, arg.location.length, poolIndex);
if (arg.dimensions.size() > 0) {
// It's the responsibility of the caller to validate that
// from.dimensions only modifies the dimensions that were
// unspecified in the model. That's the case in SampleDriver.cpp
// with the call to validateRequest().
operand.dimensions = arg.dimensions;
}
operand.buffer = r.buffer + arg.location.offset;
operand.length = arg.location.length;
rBufferLength = operand.length;
ALOGV("%s Operand length:%d pointer:%p", __func__, operand.length,
(r.buffer + arg.location.offset));
return (r.buffer + arg.location.offset);
}
void* NnapiModelInfo::getBlobFromMemoryPoolOut(const V1_3::Request& request, uint32_t index,
uint32_t& rBufferLength) {
RunTimeOperandInfo& operand = mOperands[mModel.main.outputIndexes[index]];
const V1_0::RequestArgument& arg = request.outputs[index];
auto poolIndex = arg.location.poolIndex;
nnAssert(poolIndex < mRequestPoolInfos.size());
auto& r = mRequestPoolInfos[poolIndex];
ALOGV("%s lifetime:%d location offset:%d length:%d pool index:%d", __func__, operand.lifetime,
arg.location.offset, arg.location.length, poolIndex);
if (arg.dimensions.size() > 0) {
// It's the responsibility of the caller to validate that
// from.dimensions only modifies the dimensions that were
// unspecified in the model. That's the case in SampleDriver.cpp
// with the call to validateRequest().
operand.dimensions = arg.dimensions;
}
operand.buffer = r.buffer + arg.location.offset;
operand.length = arg.location.length;
rBufferLength = operand.length;
ALOGV("%s Operand length:%d pointer:%p", __func__, operand.length,
(r.buffer + arg.location.offset));
return (r.buffer + arg.location.offset);
}
bool NnapiModelInfo::isOmittedInput(int operationIndex, uint32_t index) {
uint32_t inputIndex = mModel.main.operations[operationIndex].inputs[index];
const auto op = mModel.main.operands[inputIndex];
if (op.lifetime == OperandLifeTime::NO_VALUE) {
ALOGD("index %d has life time NO_VALUE", index);
return true;
}
return false;
}
template int NnapiModelInfo::GetConstOperand<int>(unsigned int);
template float NnapiModelInfo::GetConstOperand<float>(unsigned int);
template uint8_t NnapiModelInfo::GetConstOperand<uint8_t>(unsigned int);
template int8_t NnapiModelInfo::GetConstOperand<int8_t>(unsigned int);
template uint32_t NnapiModelInfo::GetConstOperand<uint32_t>(unsigned int);
template _Float16 NnapiModelInfo::GetConstOperand<_Float16>(unsigned int);
template int NnapiModelInfo::GetConstFromBuffer<int>(unsigned char const*, unsigned int);
template float NnapiModelInfo::GetConstFromBuffer<float>(unsigned char const*, unsigned int);
template uint8_t NnapiModelInfo::GetConstFromBuffer<uint8_t>(unsigned char const*, unsigned int);
template int8_t NnapiModelInfo::GetConstFromBuffer<int8_t>(unsigned char const*, unsigned int);
template uint32_t NnapiModelInfo::GetConstFromBuffer<uint32_t>(unsigned char const*, unsigned int);
template _Float16 NnapiModelInfo::GetConstFromBuffer<_Float16>(unsigned char const*, unsigned int);
} // namespace android::hardware::neuralnetworks::nnhal