1 // Copyright (C) 2018-2019 Intel Corporation
2 // SPDX-License-Identifier: Apache-2.0
5 #include "ext_list.hpp"
6 #include "ext_base.hpp"
13 #include "ie_parallel.hpp"
15 namespace InferenceEngine {
16 namespace Extensions {
19 inline void clipping(int *idx, const int min, const int max) {
20 (*idx) = ((*idx) > min) ? (*idx) : min;
21 (*idx) = ((*idx) < max) ? (*idx) : (max - 1);
25 class StridedSliceImpl: public ExtLayerBase {
27 explicit StridedSliceImpl(const CNNLayer* layer) {
29 if (layer->insData.size() > 4 || layer->outData.size() != 1)
30 THROW_IE_EXCEPTION << layer->name << " Incorrect number of input/output edges!";
32 src_dims = layer->insData[STRIDEDSLICE_DATA].lock()->getTensorDesc().getDims();
36 if (layer->insData.size() > 1) {
37 begin_dims = layer->insData[STRIDEDSLICE_BEGIN].lock()->getTensorDesc().getDims();
38 if (layer->insData[STRIDEDSLICE_BEGIN].lock()->getTensorDesc().getPrecision() != Precision::I32)
39 THROW_IE_EXCEPTION << layer->name << " Incorrect 'begin' input precision. Only I32 is supported!";
40 if (begin_dims.size() > 1)
41 THROW_IE_EXCEPTION << layer->name << " Begin vector should be 1 dimension";
42 bounds_size = begin_dims[0];
45 if (layer->insData.size() > 2) {
46 end_dims = layer->insData[STRIDEDSLICE_END].lock()->getTensorDesc().getDims();
47 if (layer->insData[STRIDEDSLICE_END].lock()->getTensorDesc().getPrecision() != Precision::I32)
48 THROW_IE_EXCEPTION << layer->name << " Incorrect 'end' input precision. Only I32 is supported!";
49 if (end_dims.size() > 1)
50 THROW_IE_EXCEPTION << layer->name << " End vector should be 1 dimension";
51 if (begin_dims[0] != end_dims[0])
52 THROW_IE_EXCEPTION << layer->name << " Begin vector size should be equal end vectror size";
55 if (layer->insData.size() > 3) {
56 stride_dims = layer->insData[STRIDEDSLICE_STRIDE].lock()->getTensorDesc().getDims();
57 if (layer->insData[STRIDEDSLICE_STRIDE].lock()->getTensorDesc().getPrecision() != Precision::I32)
58 THROW_IE_EXCEPTION << layer->name << " Incorrect 'strides' input precision. Only I32 is supported!";
59 if (stride_dims.size() > 1)
60 THROW_IE_EXCEPTION << layer->name << " End vector should be 1 dimension";
61 if (begin_dims[0] != stride_dims[0])
62 THROW_IE_EXCEPTION << layer->name << " Stride vector size should be equal begin vectror size";
64 dst_dims = layer->outData[0]->getTensorDesc().getDims();
66 std::string::size_type i;
67 std::string begin_mask_str = layer->GetParamAsString("begin_mask", "");
68 for (i = 0; i < begin_mask_str.size(); ++i) {
69 if (begin_mask_str[i] == '1') begin_mask.push_back(1);
70 else if (begin_mask_str[i] == '0') begin_mask.push_back(0);
72 for (; i < src_dims.size(); ++i) begin_mask.push_back(1);
74 std::string end_mask_str = layer->GetParamAsString("end_mask", "");
75 for (i = 0; i < end_mask_str.size(); ++i) {
76 if (end_mask_str[i] == '1') end_mask.push_back(1);
77 else if (end_mask_str[i] == '0') end_mask.push_back(0);
79 for (; i < src_dims.size(); ++i) end_mask.push_back(1);
81 std::string ellipsis_mask_str = layer->GetParamAsString("ellipsis_mask", "");
82 size_t ellipsis_mask_counter = 0;
83 for (i = 0; i < ellipsis_mask_str.size(); ++i) {
84 if (ellipsis_mask_str[i] == '1') {
85 ellipsis_mask_counter++;
86 ellipsis_mask.push_back(1);
87 } else if (ellipsis_mask_str[i] == '0') {
88 ellipsis_mask.push_back(0);
91 if (ellipsis_mask_counter > 1)
92 THROW_IE_EXCEPTION << layer->name << " 'Ellipsis_mask' must be a power of two (only one ellipsis)!";
93 for (; i < src_dims.size(); ++i) ellipsis_mask.push_back(0);
95 std::string new_axis_mask_str = layer->GetParamAsString("new_axis_mask", "");
96 for (i = 0; i < new_axis_mask_str.size(); ++i) {
97 if (new_axis_mask_str[i] == '1') new_axis_mask.push_back(1);
98 else if (new_axis_mask_str[i] == '0') new_axis_mask.push_back(0);
100 for (; i < src_dims.size(); ++i) new_axis_mask.push_back(0);
102 std::string shrink_axis_mask_str = layer->GetParamAsString("shrink_axis_mask", "");
103 for (i = 0; i < shrink_axis_mask_str.size(); ++i) {
104 if (shrink_axis_mask_str[i] == '1') shrink_axis_mask.push_back(1);
105 else if (shrink_axis_mask_str[i] == '0') shrink_axis_mask.push_back(0);
107 for (; i < src_dims.size(); ++i) shrink_axis_mask.push_back(0);
111 for (auto& na : new_axis_mask)
115 for (auto& sa : shrink_axis_mask)
117 max_dims = src_dims.size() + new_axis;
119 // ellipsis_mask must be a power of two (only one ellipsis), so to take a first position
120 ellipsis_pos1 = ellipsis_pos2 = max_dims;
121 for (i = 0; i < ellipsis_mask.size(); i++) {
122 if (ellipsis_mask[i] > 0) {
127 bounds_size -= ellipsis_pos1;
128 if (bounds_size > 0 && (max_dims - bounds_size) > ellipsis_pos1)
129 ellipsis_pos2 = max_dims - bounds_size;
131 begin_dms.assign(max_dims, 0);
132 end_dms.assign(max_dims, -1);
133 stride_dms.assign(max_dims, 1);
135 srcStrides = layer->insData[STRIDEDSLICE_DATA].lock()->getTensorDesc().getBlockingDesc().getStrides();
136 dstStrides = layer->outData[0]->getTensorDesc().getBlockingDesc().getStrides();
137 if (layer->insData.size() == 1) {
138 addConfig(layer, { DataConfigurator(ConfLayout::PLN) }, { DataConfigurator(ConfLayout::PLN) });
139 } else if (layer->insData.size() == 2) {
140 addConfig(layer, { DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN) }, { DataConfigurator(ConfLayout::PLN) });
141 } else if (layer->insData.size() == 3) {
142 addConfig(layer, { DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN) },
143 { DataConfigurator(ConfLayout::PLN) });
145 addConfig(layer, { DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN), DataConfigurator(ConfLayout::PLN),
146 DataConfigurator(ConfLayout::PLN) }, { DataConfigurator(ConfLayout::PLN) });
148 } catch (InferenceEngine::details::InferenceEngineException &ex) {
149 errorMsg = ex.what();
153 StatusCode execute(std::vector<Blob::Ptr>& inputs, std::vector<Blob::Ptr>& outputs, ResponseDesc *resp) noexcept override {
154 const float *src_data = inputs[STRIDEDSLICE_DATA]->cbuffer().as<const float *>() +
155 inputs[STRIDEDSLICE_DATA]->getTensorDesc().getBlockingDesc().getOffsetPadding();
156 int *begin = nullptr, *end = nullptr, *stride = nullptr;
157 if (begin_dims.size())
158 begin = inputs[STRIDEDSLICE_BEGIN]->cbuffer().as<int *>() + inputs[STRIDEDSLICE_BEGIN]->getTensorDesc().getBlockingDesc().getOffsetPadding();
160 end = inputs[STRIDEDSLICE_END]->cbuffer().as<int *>() + inputs[STRIDEDSLICE_END]->getTensorDesc().getBlockingDesc().getOffsetPadding();
161 if (stride_dims.size())
162 stride = inputs[STRIDEDSLICE_STRIDE]->cbuffer().as<int *>() + inputs[STRIDEDSLICE_STRIDE]->getTensorDesc().getBlockingDesc().getOffsetPadding();
163 float* dst_data = outputs[0]->cbuffer().as<float *>() +
164 outputs[0]->getTensorDesc().getBlockingDesc().getOffsetPadding();
166 InferenceEngine::SizeVector src_dims = inputs[STRIDEDSLICE_DATA]->getTensorDesc().getDims();
167 InferenceEngine::SizeVector srcStrides = inputs[STRIDEDSLICE_DATA]->getTensorDesc().getBlockingDesc().getStrides();
168 InferenceEngine::SizeVector dst_dims = outputs[0]->getTensorDesc().getDims();
169 InferenceEngine::SizeVector dstStrides = outputs[0]->getTensorDesc().getBlockingDesc().getStrides();
171 size_t i, j, k, bj, ej, sj;
172 InferenceEngine::SizeVector our_dims;
173 InferenceEngine::SizeVector out_dims;
174 for (i = 0, j = 0, k = 0, bj = 0, ej = 0, sj = 0; static_cast<int>(i) < max_dims; i++) {
175 if (static_cast<int>(i) >= ellipsis_pos1 &&
176 static_cast<int>(i) < ellipsis_pos2) {
177 if (new_axis_mask.size() > i && new_axis_mask[i] == 1)
180 end_dms[i] = end_dms[i] >= 0 ? end_dms[i] : src_dims[j++] + end_dms[i];
182 out_dims.push_back(static_cast<int>(ceil(static_cast<float>(abs(end_dms[i] - begin_dms[i]) + 1) / static_cast<float>(abs(stride_dms[i])))));
183 our_dims.push_back(static_cast<int>(ceil(static_cast<float>(abs(end_dms[i] - begin_dms[i]) + 1) / static_cast<float>(abs(stride_dms[i])))));
186 stride_dms[i] = (stride != nullptr && stride_dims[0] > sj && stride[sj] != 0) ? stride[sj++] : 1;
188 if (begin_mask.size() > j && begin_mask[j] == 0)
189 begin_dms[i] = stride_dms[i] > 0 ? 0 : -1;
191 begin_dms[i] = (begin != nullptr && begin_dims[0] > bj) ? begin[bj] : (stride_dms[i] > 0 ? 0 : -1);
193 begin_dms[i] = begin_dms[i] >= 0 ? begin_dms[i] : src_dims[j] + begin_dms[i];
195 clipping(&begin_dms[i], 0, src_dims[j]);
197 if (end_mask.size() > j && end_mask[j] == 0) {
198 end_dms[i] = stride_dms[i] > 0 ? -1 : 0;
200 int end_dms_tmp = (end != nullptr && end_dims[0] > ej) ? (stride_dms[i] > 0 ? end[ej] - 1 : end[ej] + 1)
202 end_dms[i] = (end != nullptr && end_dims[0] > ej) ? end_dms_tmp : (stride_dms[i] > 0 ? -1 : 0);
205 end_dms[i] = end_dms[i] >= 0 ? end_dms[i] : src_dims[j] + end_dms[i];
207 clipping(&end_dms[i], 0, src_dims[j]);
209 if (new_axis_mask.size() > i && new_axis_mask[i] == 1)
214 if (shrink_axis_mask.size() > k && shrink_axis_mask[k] == 1)
215 end_dms[i] = begin_dms[i];
217 out_dims.push_back(static_cast<int>(ceil(static_cast<float>(abs(end_dms[i] - begin_dms[i]) + 1) /
218 static_cast<float>(abs(stride_dms[i])))));
220 our_dims.push_back(static_cast<int>(ceil(static_cast<float>(abs(end_dms[i] - begin_dms[i]) + 1) /
221 static_cast<float>(abs(stride_dms[i])))));
226 for (i = 0; i < std::min(out_dims.size(), dst_dims.size()); i++) {
227 if (out_dims[i] != dst_dims[i])
228 return PARAMETER_MISMATCH;
231 if (static_cast<int>(src_dims.size()) == max_dims && shrink_axis == 0 &&
232 stride_dms[stride_dms.size()-1] == 1 && stride_dms.size() > 1)
233 strided_slice_vp(src_data, dst_data);
234 else if (static_cast<int>(src_dims.size()) == max_dims && shrink_axis == 0)
235 strided_slice_p(src_data, dst_data);
237 strided_slice(src_data, dst_data, our_dims);
243 const size_t STRIDEDSLICE_DATA = 0;
244 const size_t STRIDEDSLICE_BEGIN = 1;
245 const size_t STRIDEDSLICE_END = 2;
246 const size_t STRIDEDSLICE_STRIDE = 3;
248 void strided_slice(const float *src_data, float* dst_data, std::vector<size_t> &dims);
249 void strided_slice_vp(const float *src_data, float* dst_data);
250 void strided_slice_p(const float *src_data, float* dst_data);
252 SizeVector begin_dims;
254 SizeVector stride_dims;
256 SizeVector begin_mask;
258 SizeVector ellipsis_mask;
259 SizeVector new_axis_mask;
260 SizeVector shrink_axis_mask;
265 std::vector<int> begin_dms;
266 std::vector<int> end_dms;
267 std::vector<int> stride_dms;
268 SizeVector srcStrides;
269 SizeVector dstStrides;
272 int ellipsis_pos1, ellipsis_pos2;
275 void StridedSliceImpl::strided_slice(const float *src_data, float* dst_data, std::vector<size_t> &dims) {
276 size_t work_amount_dst = dstStrides[0] * dst_dims[0];
277 parallel_nt(0, [&](const int ithr, const int nthr) {
279 size_t i, start = 0, end = 0;
280 SizeVector counters(max_dims, 0);
281 splitter(work_amount_dst, nthr, ithr, start, end);
282 for (j = max_dims - 1, i = start; j >= 0; j--) {
283 counters[j] = i % dims[j];
286 for (size_t iwork = start; iwork < end; ++iwork) {
288 for (i = 0, j = 0; static_cast<int>(i) < max_dims; ++i) {
289 if (!(new_axis_mask.size() > i && new_axis_mask[i] == 1))
290 src_idx += (begin_dms[i] + counters[i] * stride_dms[i]) * srcStrides[j++];
293 dst_data[iwork] = src_data[src_idx];
295 for (j = max_dims - 1; j >= 0; j--) {
297 if (counters[j] < dims[j])
306 void StridedSliceImpl::strided_slice_vp(const float *src_data, float* dst_data) {
308 size_t dims_size_1 = dst_dims.size() - 1;
309 size_t dataLength = dst_dims[dims_size_1];
310 size_t work_amount_dst = dstStrides[0] * dst_dims[0] / dst_dims[dims_size_1];
312 parallel_nt(0, [&](const int ithr, const int nthr) {
313 size_t start = 0, end = 0;
314 SizeVector counters(dims_size_1, 0);
315 splitter(work_amount_dst, nthr, ithr, start, end);
316 size_t src_idx = begin_dms[dims_size_1];
317 for (int j = dims_size_1 - 1, i = start; j >= 0; j--) {
318 counters[j] = i % dst_dims[j];
319 src_idx += (begin_dms[j] + counters[j] * stride_dms[j]) * srcStrides[j];
323 for (size_t iwork = start, dst_idx = start * dataLength, i = 1; iwork < end; ++iwork, dst_idx += dataLength) {
324 memcpy(&dst_data[dst_idx], &src_data[src_idx], sizeof(float) * dataLength);
325 for (int j = dims_size_1 - 1; j >= 0; j--) {
327 if (counters[j] < dst_dims[j]) {
328 src_idx += stride_dms[j] * srcStrides[j];
335 for (src_idx = begin_dms[dims_size_1]; i < dims_size_1; ++i)
336 src_idx += (begin_dms[i] + counters[i] * stride_dms[i]) * srcStrides[i];
342 void StridedSliceImpl::strided_slice_p(const float *src_data, float* dst_data) {
343 size_t dims_size = dst_dims.size();
344 size_t work_amount_dst = dstStrides[0] * dst_dims[0];
346 parallel_nt(0, [&](const int ithr, const int nthr) {
347 size_t start = 0, end = 0;
348 SizeVector counters(dims_size, 0);
349 splitter(work_amount_dst, nthr, ithr, start, end);
351 for (int j = dims_size - 1, i = start; j >= 0; j--) {
352 counters[j] = i % dst_dims[j];
353 src_idx += (begin_dms[j] + counters[j] * stride_dms[j]) * srcStrides[j];
357 for (size_t iwork = start, dst_idx = start, i = 1; iwork < end; ++iwork, dst_idx++) {
358 dst_data[dst_idx] = src_data[src_idx];
359 for (int j = dims_size - 1; j >= 0; j--) {
361 if (counters[j] < dst_dims[j]) {
362 src_idx += stride_dms[j] * srcStrides[j];
369 for (src_idx = 0; i < dims_size; ++i)
370 src_idx += (begin_dms[i] + counters[i] * stride_dms[i]) * srcStrides[i];
376 REG_FACTORY_FOR(ImplFactory<StridedSliceImpl>, StridedSlice);
379 } // namespace Extensions
380 } // namespace InferenceEngine