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
|
// Copyright (C) 2018 Intel Corporation
//
// SPDX-License-Identifier: Apache-2.0
//
#include "mkldnn_permute_node.h"
#include <ie_layers.h>
#include <string>
#include <mkldnn_types.h>
#include <mkldnn_extension_utils.h>
using namespace mkldnn;
using namespace MKLDNNPlugin;
using namespace InferenceEngine;
MKLDNNPermuteNode::MKLDNNPermuteNode(const InferenceEngine::CNNLayerPtr& layer, const mkldnn::engine& eng) : MKLDNNNode(layer, eng) {}
void MKLDNNPermuteNode::getSupportedDescriptors() {
if (getParentEdges().size() != 1)
THROW_IE_EXCEPTION << "Incorrect number of input edges.";
if (!getChildEdges().size())
THROW_IE_EXCEPTION << "Incorrect number of output edges.";
auto& layer = getCnnLayer();
if (!layer) {
THROW_IE_EXCEPTION << "Cannot get CNNLayer.";
}
order.clear();
std::vector<int> layerOrder = layer->GetParamAsInts("order");
for (auto ord : layerOrder)
order.push_back(static_cast<size_t>(ord));
}
void MKLDNNPermuteNode::initSupportedPrimitiveDescriptors() {
if (!supportedPrimitiveDescriptors.empty())
return;
InferenceEngine::Precision precision = getCnnLayer()->insData[0].lock()->getPrecision();
if (precision != InferenceEngine::Precision::FP32)
precision = InferenceEngine::Precision::FP32;
auto inputDataType = MKLDNNExtensionUtils::IEPrecisionToDataType(precision);
precision = getCnnLayer()->outData[0]->getPrecision();
if (precision != InferenceEngine::Precision::FP32)
precision = InferenceEngine::Precision::FP32;
auto outputDataType = MKLDNNExtensionUtils::IEPrecisionToDataType(precision);
InferenceEngine::LayerConfig config;
config.dynBatchSupport = true;
config.inConfs.resize(1);
config.outConfs.resize(1);
config.inConfs[0].inPlace = -1;
config.inConfs[0].constant = false;
config.outConfs[0].inPlace = -1;
config.outConfs[0].constant = false;
if (getParentEdgeAt(0)->getDims().ndims() == 4) {
config.inConfs[0].desc = MKLDNNMemoryDesc(getParentEdgeAt(0)->getDims(), inputDataType, memory::nchw);
config.outConfs[0].desc = MKLDNNMemoryDesc(getChildEdgeAt(0)->getDims(), outputDataType, memory::nchw);
supportedPrimitiveDescriptors.push_back({config, impl_desc_type::unknown});
auto srcDims = getParentEdgeAt(0)->getDims();
if (srcDims[1] % 8 == 0) {
config.inConfs[0].desc = MKLDNNMemoryDesc(getParentEdgeAt(0)->getDims(), inputDataType, memory::nChw8c);
supportedPrimitiveDescriptors.push_back({config, impl_desc_type::unknown});
}
if (srcDims[1] % 16 == 0) {
config.inConfs[0].desc = MKLDNNMemoryDesc(getParentEdgeAt(0)->getDims(), inputDataType, memory::nChw16c);
supportedPrimitiveDescriptors.push_back({config, impl_desc_type::unknown});
}
} else {
config.inConfs[0].desc = MKLDNNMemoryDesc(getParentEdgeAt(0)->getDims(), inputDataType, memory::any);
config.outConfs[0].desc = MKLDNNMemoryDesc(getChildEdgeAt(0)->getDims(), outputDataType,
MKLDNNMemory::GetPlainFormat(getChildEdgeAt(0)->getDims()));
supportedPrimitiveDescriptors.push_back({config, impl_desc_type::unknown});
}
}
void MKLDNNPermuteNode::createPrimitive() {
auto& dstMemPtr = getChildEdgeAt(0)->getMemoryPtr();
auto& srcMemPtr = getParentEdgeAt(0)->getMemoryPtr();
if (!dstMemPtr || !dstMemPtr->GetPrimitivePtr())
THROW_IE_EXCEPTION << "Destination memory didn't allocate.";
if (!srcMemPtr || !srcMemPtr->GetPrimitivePtr())
THROW_IE_EXCEPTION << "Input memory didn't allocate.";
if (getSelectedPrimitiveDescriptor() == nullptr)
THROW_IE_EXCEPTION << "Preferable primitive descriptor does not set.";
}
static void permute_to_0231(int MB, MKLDNNMemoryPtr& srcMemPtr, MKLDNNMemoryPtr& dstMemPtr) {
auto src_data = reinterpret_cast<const float *>(srcMemPtr->GetData());
auto dst_data = reinterpret_cast<float *>(dstMemPtr->GetData());
src_data += srcMemPtr->GetDescriptor().data.layout_desc.blocking.offset_padding;
dst_data += dstMemPtr->GetDescriptor().data.layout_desc.blocking.offset_padding;
// Supports only NCHW to NHWC
int block_size = 1;
if (!MKLDNNMemory::IsPlainFormat(srcMemPtr->GetFormat())) {
block_size = srcMemPtr->GetDescriptor().data.layout_desc.blocking.block_dims[1];
}
const int C = srcMemPtr->GetDims()[1];
const int H = srcMemPtr->GetDims()[2];
const int W = srcMemPtr->GetDims()[3];
// NHWC
const int src_stride = H * W * block_size;
int src_off = 0;
int dst_off = 0;
for (int n = 0; n < MB; n++) {
for (int h = 0; h < H; h++) {
for (int w = 0; w < W; w++) {
src_off = n * C * H * W + (h * W + w) * block_size;
for (int c = 0; c < C; c += block_size) {
for (int b = 0; b < block_size; b++) {
dst_data[dst_off] = src_data[src_off + b];
dst_off++;
}
src_off += src_stride;
}
}
}
}
}
template <size_t scale_H = 0, size_t scale_W = 0>
static void permute_to_014253(int MB, MKLDNNMemoryPtr& srcMemPtr, MKLDNNMemoryPtr& dstMemPtr) {
auto src_data = reinterpret_cast<const float *>(srcMemPtr->GetData());
auto dst_data = reinterpret_cast<float *>(dstMemPtr->GetData());
src_data += srcMemPtr->GetDescriptor().data.layout_desc.blocking.offset_padding;
dst_data += dstMemPtr->GetDescriptor().data.layout_desc.blocking.offset_padding;
const int C = srcMemPtr->GetDims()[1];
const int CH = scale_H > 0 ? static_cast<int>(scale_H) : srcMemPtr->GetDims()[2];
const int CW = scale_W > 0 ? static_cast<int>(scale_W) : srcMemPtr->GetDims()[3];
const int H = srcMemPtr->GetDims()[4];
const int W = srcMemPtr->GetDims()[5];
int src_off = 0;
int dst_off = 0;
for (int n = 0; n < MB; n++) {
for (int c = 0; c < C; c++) {
for (int h = 0; h < H; h++) {
for (int ch = 0; ch < CH; ch++) {
for (int w = 0; w < W; w++) {
for (int cw = 0; cw < CW; cw++) {
src_off = n * C * CH * CW * H * W +
c * CH * CW * H * W +
ch * CW * H * W +
cw * H * W +
h * W +
w;
dst_data[dst_off] = src_data[src_off];
dst_off++;
}
}
}
}
}
}
}
static void permute_to_3012(int MB, MKLDNNMemoryPtr& srcMemPtr, MKLDNNMemoryPtr& dstMemPtr) {
auto src_data = reinterpret_cast<const float *>(srcMemPtr->GetData());
auto dst_data = reinterpret_cast<float *>(dstMemPtr->GetData());
src_data += srcMemPtr->GetDescriptor().data.layout_desc.blocking.offset_padding;
dst_data += dstMemPtr->GetDescriptor().data.layout_desc.blocking.offset_padding;
const int C = srcMemPtr->GetDims()[1];
const int H = srcMemPtr->GetDims()[2];
const int W = srcMemPtr->GetDims()[3];
int src_off = 0;
int dst_off = 0;
for (int w = 0; w < W; w++) {
for (int n = 0; n < MB; n++) {
for (int c = 0; c < C; c++) {
for (int h = 0; h < H; h++) {
src_off = n * C * H * W +
c * H * W +
h * W +
w;
dst_data[dst_off] = src_data[src_off];
dst_off++;
}
}
}
}
}
std::map<InferenceEngine::SizeVector, MKLDNNPermuteNode::PermuteImpl> MKLDNNPermuteNode::OptimizedCases = {
{{0, 2, 3, 1}, MKLDNNPermuteNode::PermuteImpl(permute_to_0231, [](MKLDNNMemoryPtr& srcMemPtr, MKLDNNMemoryPtr& dstMemPtr) {
return true;
})}, // NCHW -> NHWC case
{{0, 1, 4, 2, 5, 3}, MKLDNNPermuteNode::PermuteImpl(permute_to_014253<2, 2>, [](MKLDNNMemoryPtr& srcMemPtr, MKLDNNMemoryPtr& dstMemPtr) {
return MKLDNNMemory::IsPlainFormat(srcMemPtr->GetFormat()) && srcMemPtr->GetDims()[2] == 2 && srcMemPtr->GetDims()[3] == 2;
})}, // Dense upsample convolution case (scale = 2)
{{0, 1, 4, 2, 5, 3}, MKLDNNPermuteNode::PermuteImpl(permute_to_014253<0, 0>, [](MKLDNNMemoryPtr& srcMemPtr, MKLDNNMemoryPtr& dstMemPtr) {
return MKLDNNMemory::IsPlainFormat(srcMemPtr->GetFormat());
})}, // Dense upsample convolution case (generic)
{{3, 0, 1, 2}, MKLDNNPermuteNode::PermuteImpl(permute_to_3012, [](MKLDNNMemoryPtr& srcMemPtr, MKLDNNMemoryPtr& dstMemPtr) {
return MKLDNNMemory::IsPlainFormat(srcMemPtr->GetFormat());
})}, // LPR case
};
void MKLDNNPermuteNode::execute(mkldnn::stream strm) {
auto &dstMemPtr = getChildEdgeAt(0)->getMemoryPtr();
auto &srcMemPtr = getParentEdgeAt(0)->getMemoryPtr();
auto src_data = reinterpret_cast<const float *>(srcMemPtr->GetData());
auto dst_data = reinterpret_cast<float *>(dstMemPtr->GetData());
auto perm = OptimizedCases.find(order);
if (perm != OptimizedCases.end() && perm->second.isValidParams(srcMemPtr, dstMemPtr)) {
perm->second.execute(batchToProcess(), srcMemPtr, dstMemPtr);
} else {
auto srcBlob = getParentEdgeAt(0)->getBlob();
TensorDesc srcDesc = srcBlob->getTensorDesc();
SizeVector& dims = srcDesc.getDims();
InferenceEngine::SizeVector orderedDims;
for (auto ord : order) {
orderedDims.push_back(dims[ord]);
}
TensorDesc dstDesc(InferenceEngine::Precision::FP32, dims, {orderedDims, order});
size_t dataSize = srcBlob->size() / srcDesc.getDims()[0] * batchToProcess();
#pragma omp parallel for
for (size_t i = 0; i < dataSize; i++) {
dst_data[dstDesc.offset(i)] = src_data[srcDesc.offset(i)];
}
}
}
bool MKLDNNPermuteNode::created() const {
return getType() == Permute;
}
|
