File size: 11,787 Bytes
dffe123
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.

#include "../precomp.hpp"
#include <opencv2/dnn/shape_utils.hpp>
#include "./cpu_kernels/fast_norm.hpp"

// CANN backend
#include "../op_cann.hpp"

// OpenVINO backend
#include "../op_inf_engine.hpp"
#include "../ie_ngraph.hpp"

// CUDA backend
#include "../op_cuda.hpp"
#ifdef HAVE_CUDA
#include "../cuda4dnn/primitives/instance_norm.hpp"
using namespace cv::dnn::cuda4dnn;
#endif

// OpenCL backend
#ifdef HAVE_OPENCL
#include "../ocl4dnn/include/math_functions.hpp"
#include "opencl_kernels_dnn.hpp"
#endif

namespace cv { namespace dnn {

// https://github.com/onnx/onnx/blob/main/docs/Operators.md#InstanceNormalization
class InstanceNormLayerImpl CV_FINAL : public InstanceNormLayer {
public:
    InstanceNormLayerImpl(const LayerParams &params) {
        setParamsFrom(params);

        epsilon = params.get<float>("epsilon", 1e-5);
    }

    virtual bool supportBackend(int backendId) CV_OVERRIDE {
#ifdef HAVE_INF_ENGINE
        if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
            return true;
#endif
        return backendId == DNN_BACKEND_OPENCV ||
               backendId == DNN_BACKEND_CUDA;
            //    backendId == DNN_BACKEND_CANN; // not supported due to 1d mat shape issue
    }

    bool getMemoryShapes(const std::vector<MatShape> &inputs,
                         const int requiredOutputs,
                         std::vector<MatShape> &outputs,
                         std::vector<MatShape> &internals) const CV_OVERRIDE {
        const auto &input = inputs[0];
        const auto &scale = inputs[1];
        const auto &bias = inputs[2];
        CV_CheckGE(input.size(), static_cast<size_t>(3), "DNN/InstanceNorm: input dimension >= 3 is required");

        int C = input[1];
        int scale_dim = std::accumulate(scale.begin(), scale.end(), 1, std::multiplies<int>());
        CV_CheckEQ(scale_dim, C, "DNN/InstanceNorm: scale must be a 1d tensor and match the channel of input");
        int bias_dim = std::accumulate(bias.begin(), bias.end(), 1, std::multiplies<int>());
        CV_CheckEQ(bias_dim, C, "DNN/InstanceNorm: bias must be a 1d tensor and match the channel of input");

        outputs.assign(1, inputs[0]);
        return false;
    }

    void forward(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr, OutputArrayOfArrays internals_arr) CV_OVERRIDE {
        CV_TRACE_FUNCTION();
        CV_TRACE_ARG_VALUE(name, "name", name.c_str());

        CV_OCL_RUN(IS_DNN_OPENCL_TARGET(preferableTarget),
                   forward_ocl(inputs_arr, outputs_arr, internals_arr))

        if (inputs_arr.depth() == CV_16F)
        {
            forward_fallback(inputs_arr, outputs_arr, internals_arr);
            return;
        }

        std::vector<Mat> inputs, outputs;
        inputs_arr.getMatVector(inputs);
        outputs_arr.getMatVector(outputs);

        const auto &input = inputs[0];
        const auto &scale = inputs[1];
        const auto &bias = inputs[2];

        fastNormChannel(input, scale, bias, outputs[0], epsilon);
    }

#ifdef HAVE_OPENCL
    bool forward_ocl(InputArrayOfArrays inputs_, OutputArrayOfArrays outputs_, OutputArrayOfArrays internals_) {
        std::vector<UMat> inputs;
        std::vector<UMat> outputs;

        inputs_.getUMatVector(inputs);
        outputs_.getUMatVector(outputs);

        const auto &input = inputs[0], &scale = inputs[1], &bias = inputs[2];
        auto &output = outputs[0];

        const auto input_shape = shape(input);
        size_t N = input_shape[0], C = input_shape[1],
               loops = N * C, norm_size = static_cast<size_t>(total(input_shape, 2));
        float inv_norm_size = 1.f / norm_size;

        // no fp16 support
        if (input.depth() == CV_16F) {
            return false;
        }

        String base_opts = format(" -DT=float -DT4=float4 -Dconvert_T=convert_float4");

        // Calculate mean
        UMat one = UMat::ones(norm_size, 1, CV_32F);
        UMat mean = UMat(loops, 1, CV_32F);
        UMat mean_square = UMat(loops, 1, CV_32F);
        UMat tmp = UMat(loops, norm_size, CV_32F);
        bool ret = ocl4dnn::ocl4dnnGEMV<float>(ocl4dnn::CblasNoTrans, loops, norm_size, inv_norm_size,
                                               input, 0, one, 0, 0.f, mean, 0);
        if (!ret) {
            return false;
        }
        // Calculate mean_square
        int num_vector = (norm_size % 8 == 0) ? 8 : ((norm_size % 4 == 0) ? 4 : 1);
        size_t global[] = {loops, static_cast<size_t>(norm_size / num_vector)};
        String build_opt = format(" -DNUM=%d", num_vector) + base_opts;
        String mean_square_kernel_name = format("calc_mean%d", num_vector);
        ocl::Kernel mean_square_kernel(mean_square_kernel_name.c_str(), ocl::dnn::mvn_oclsrc, build_opt + " -DKERNEL_MEAN");
        if (mean_square_kernel.empty()) {
            return false;
        }
        mean_square_kernel.set(0, ocl::KernelArg::PtrReadOnly(input));
        mean_square_kernel.set(1, (int)loops);
        mean_square_kernel.set(2, (int)norm_size);
        mean_square_kernel.set(3, ocl::KernelArg::PtrReadOnly(mean));
        mean_square_kernel.set(4, ocl::KernelArg::PtrWriteOnly(tmp));
        ret = mean_square_kernel.run(2, global, NULL, false);
        if (!ret) {
            return false;
        }
        ret = ocl4dnn::ocl4dnnGEMV<float>(ocl4dnn::CblasNoTrans, loops, norm_size, inv_norm_size,
                                          tmp, 0, one, 0, 0.f, mean_square, 0);
        if (!ret) {
            return false;
        }
        // Calculate instance norm: output = scale * (x - mean) / sqrt(var + eps) + bias
        String mvn_kernel_name = format("mvn%d", num_vector);
        build_opt += " -DNORM_VARIANCE -DFUSE_BATCH_NORM -DKERNEL_MVN";
        ocl::Kernel mvn_kernel(mvn_kernel_name.c_str(), ocl::dnn::mvn_oclsrc, build_opt);
        if (mvn_kernel.empty()) {
            return false;
        }
        mvn_kernel.set(0, ocl::KernelArg::PtrReadOnly(input));
        mvn_kernel.set(1, (int)loops);
        mvn_kernel.set(2, (int)norm_size);
        mvn_kernel.set(3, (float)epsilon);
        mvn_kernel.set(4, ocl::KernelArg::PtrReadOnly(mean));
        mvn_kernel.set(5, ocl::KernelArg::PtrReadOnly(mean_square));
        mvn_kernel.set(6, ocl::KernelArg::PtrReadOnly(scale));
        mvn_kernel.set(7, ocl::KernelArg::PtrReadOnly(bias));
        mvn_kernel.set(8, (int)C);
        mvn_kernel.set(9, (float)0.f);
        mvn_kernel.set(10, ocl::KernelArg::PtrWriteOnly(output));
        ret = mvn_kernel.run(2, global, NULL, false);
        if (!ret) {
            return false;
        }

        return true;
    }
#endif

#ifdef HAVE_CANN
    virtual Ptr<BackendNode> initCann(const std::vector<Ptr<BackendWrapper> > &inputs,
                                      const std::vector<Ptr<BackendWrapper> > &outputs,
                                      const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE {
        auto input_tensor_wrapper = inputs[0].dynamicCast<CannBackendWrapper>();
        auto input_tensor_desc = input_tensor_wrapper->getTensorDesc();

        auto scale_tensor_wrapper = inputs[1].dynamicCast<CannBackendWrapper>();
        auto scale_tensor_desc = scale_tensor_wrapper->getTensorDesc();

        auto bias_tensor_wrapper = inputs[2].dynamicCast<CannBackendWrapper>();
        auto bias_tensor_desc = bias_tensor_wrapper->getTensorDesc();

        auto last_node = nodes[0].dynamicCast<CannBackendNode>()->getOp();
        auto scale_node = nodes[1].dynamicCast<CannBackendNode>()->getOp();
        auto bias_node = nodes[2].dynamicCast<CannBackendNode>()->getOp();

        auto op = std::make_shared<ge::op::InstanceNorm>(name);

        // set attrs
        op->set_attr_epsilon(epsilon);

        // set inputs
        // set inputs : x
        op->set_input_x_by_name(*last_node, input_tensor_wrapper->name.c_str());
        op->update_input_desc_x(*input_tensor_desc);
        // set inputs : gamma
        op->set_input_gamma_by_name((*scale_node), scale_tensor_wrapper->name.c_str());
        op->update_input_desc_gamma(*scale_tensor_desc);
        // set inputs : beta
        op->set_input_beta_by_name(*bias_node, bias_tensor_wrapper->name.c_str());
        op->update_input_desc_beta(*bias_tensor_desc);

        // set outputs
        auto output_desc_y = std::make_shared<ge::TensorDesc>(ge::Shape(), ge::FORMAT_NCHW, ge::DT_FLOAT);
        op->update_output_desc_y(*output_desc_y);
        auto output_desc_mean = std::make_shared<ge::TensorDesc>(ge::Shape(), ge::FORMAT_NCHW, ge::DT_FLOAT);
        op->update_output_desc_mean(*output_desc_mean);
        auto output_desc_var = std::make_shared<ge::TensorDesc>(ge::Shape(), ge::FORMAT_NCHW, ge::DT_FLOAT);
        op->update_output_desc_variance(*output_desc_var);

        return Ptr<BackendNode>(new CannBackendNode(op));
    }
#endif // HAVE_CANN

#ifdef HAVE_DNN_NGRAPH
    virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs,
                                        const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE {
        // onnx to openvino convertion: https://github.com/openvinotoolkit/openvino/blob/2023.1.0/src/frontends/onnx/frontend/src/op/instance_norm.cpp

        auto ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
        const auto &input_shape = ieInpNode.get_shape();
        std::shared_ptr<ov::Node> mvn, result;

        // mvn
        // https://docs.openvino.ai/2023.1/openvino_docs_ops_normalization_MVN_6.html
        std::vector<int64_t> axes_v(input_shape.size() - 2);
        std::iota(axes_v.begin(), axes_v.end(), 2); // {2, 3, ...} for nd input tensor, n>=3
        auto axes = std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{axes_v.size()}, axes_v.data());
        bool normalize_variance = true;
        mvn = std::make_shared<ov::op::v6::MVN>(ieInpNode, axes, normalize_variance, epsilon, ov::op::MVNEpsMode::INSIDE_SQRT);

        // instance norm = scale * mvn + bias
        auto scale = nodes[1].dynamicCast<InfEngineNgraphNode>()->node;
        std::vector<int64_t> shared_shape_v(input_shape.size(), 1);
        shared_shape_v[1] = -1;
        auto shared_shape = std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{shared_shape_v.size()}, shared_shape_v.data());
        scale  = std::make_shared<ov::op::v1::Reshape>(scale, shared_shape, true);
        result = std::make_shared<ov::op::v1::Multiply>(mvn, scale);
        auto bias = nodes[2].dynamicCast<InfEngineNgraphNode>()->node;
        bias  = std::make_shared<ov::op::v1::Reshape>(bias, shared_shape, true);
        result = std::make_shared<ov::op::v1::Add>(result, bias);

        return Ptr<BackendNode>(new InfEngineNgraphNode(result));
    }
#endif // HAVE_DNN_NGRAPH

#ifdef HAVE_CUDA
    Ptr<BackendNode> initCUDA(void *context_,
                              const std::vector<Ptr<BackendWrapper>>& inputs,
                              const std::vector<Ptr<BackendWrapper>>& outputs) override {
        auto context = reinterpret_cast<csl::CSLContext*>(context_);

        auto input_wrapper = inputs[0].dynamicCast<CUDABackendWrapper>();
        auto input_shape = input_wrapper->getShape();
        size_t loops = static_cast<size_t>(total(input_shape, 0, 2));

        return make_cuda_node<cuda4dnn::InstanceNormOp>(preferableTarget, std::move(context->stream), epsilon, loops);
    }
#endif // HAVE_CUDA

};

Ptr<InstanceNormLayer> InstanceNormLayer::create(const LayerParams &params) {
    return Ptr<InstanceNormLayer>(new InstanceNormLayerImpl(params));
}

}} // cv::dnn