| 1 | // ======================================================================== // |
| 2 | // Copyright 2009-2019 Intel Corporation // |
| 3 | // // |
| 4 | // Licensed under the Apache License, Version 2.0 (the "License"); // |
| 5 | // you may not use this file except in compliance with the License. // |
| 6 | // You may obtain a copy of the License at // |
| 7 | // // |
| 8 | // http://www.apache.org/licenses/LICENSE-2.0 // |
| 9 | // // |
| 10 | // Unless required by applicable law or agreed to in writing, software // |
| 11 | // distributed under the License is distributed on an "AS IS" BASIS, // |
| 12 | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // |
| 13 | // See the License for the specific language governing permissions and // |
| 14 | // limitations under the License. // |
| 15 | // ======================================================================== // |
| 16 | |
| 17 | #pragma once |
| 18 | |
| 19 | #include "node.h" |
| 20 | |
| 21 | namespace oidn { |
| 22 | |
| 23 | // 2x2 nearest-neighbor upsampling node |
| 24 | template<int K> |
| 25 | class UpsampleNode : public Node |
| 26 | { |
| 27 | private: |
| 28 | std::shared_ptr<memory> src; |
| 29 | std::shared_ptr<memory> dst; |
| 30 | |
| 31 | public: |
| 32 | UpsampleNode(const std::shared_ptr<memory>& src, |
| 33 | const std::shared_ptr<memory>& dst) |
| 34 | : src(src), |
| 35 | dst(dst) |
| 36 | { |
| 37 | const mkldnn_memory_desc_t& srcDesc = src->get_desc().data; |
| 38 | const mkldnn_memory_desc_t& dstDesc = dst->get_desc().data; |
| 39 | MAYBE_UNUSED(srcDesc); |
| 40 | MAYBE_UNUSED(dstDesc); |
| 41 | assert(memory_desc_matches_tag(srcDesc, mkldnn_format_tag_t(BlockedFormat<K>::nChwKc))); |
| 42 | assert(memory_desc_matches_tag(dstDesc, mkldnn_format_tag_t(BlockedFormat<K>::nChwKc))); |
| 43 | assert(srcDesc.ndims == 4); |
| 44 | assert(dstDesc.ndims == 4); |
| 45 | assert(srcDesc.data_type == memory::data_type::f32); |
| 46 | assert(dstDesc.data_type == memory::data_type::f32); |
| 47 | assert(srcDesc.dims[0] == 1); |
| 48 | assert(dstDesc.dims[0] == 1); |
| 49 | // 2x2 upsampling |
| 50 | assert(dstDesc.dims[2] == srcDesc.dims[2] * 2); |
| 51 | assert(dstDesc.dims[3] == srcDesc.dims[3] * 2); |
| 52 | } |
| 53 | |
| 54 | void execute(stream& sm) override |
| 55 | { |
| 56 | const mkldnn_memory_desc_t& srcDesc = src->get_desc().data; |
| 57 | |
| 58 | const float* srcPtr = (float*)src->get_data_handle(); |
| 59 | float* dstPtr = (float*)dst->get_data_handle(); |
| 60 | |
| 61 | const int C = srcDesc.dims[1]; |
| 62 | const int H = srcDesc.dims[2]; |
| 63 | const int W = srcDesc.dims[3]; |
| 64 | const int CK = C / K; |
| 65 | |
| 66 | parallel_nd(CK, H, [&](int ck, int h) |
| 67 | { |
| 68 | const size_t offset = ck*H*W*K + h*W*K; |
| 69 | const float* srcPtr_line = srcPtr + offset; |
| 70 | float* dstPtr_line0 = dstPtr + offset * 4; |
| 71 | float* dstPtr_line1 = dstPtr_line0 + W*2*K; // next line |
| 72 | |
| 73 | for (int w = 0; w < W; ++w) |
| 74 | { |
| 75 | #pragma unroll |
| 76 | for (int k = 0; k < K; k += 4) |
| 77 | { |
| 78 | const __m128 m = _mm_load_ps(&srcPtr_line[w*K + k]); |
| 79 | |
| 80 | _mm_stream_ps(&dstPtr_line0[w*2*K + k], m); |
| 81 | _mm_stream_ps(&dstPtr_line0[w*2*K+K + k], m); |
| 82 | _mm_stream_ps(&dstPtr_line1[w*2*K + k], m); |
| 83 | _mm_stream_ps(&dstPtr_line1[w*2*K+K + k], m); |
| 84 | } |
| 85 | } |
| 86 | }); |
| 87 | } |
| 88 | |
| 89 | std::shared_ptr<memory> getDst() const override { return dst; } |
| 90 | }; |
| 91 | |
| 92 | } // namespace oidn |
| 93 | |