mnn-sys 0.1.2

Rust bindings for MNN, a lightweight deep neural network inference engine.
Documentation 
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

//
//  OpenCLTarget.cpp
//  MNN
//
//  Created by MNN on 2022/11/14.
//  Copyright © 2018, Alibaba Group Holding Limited
//
#include <string>
#include <vector>
#include <fstream>
#include <sstream>
#include <unordered_map>

#include "core/TensorUtils.hpp"
#include "MNN_generated.h"

using namespace MNN;

class OpenCLTarget {
public:
    OpenCLTarget(std::vector<Node*>& nodes, int idx) : nodes(nodes) {
        sort(nodes.begin(), nodes.end(), [](Node* x, Node* y) { return x->topoIndex < y->topoIndex; });
        // 1. gen kernel body
        std::stringstream kernelBody;
        kernelBody << "{\n";
        down();
        /*
        kernelBody << getIndent() << "const int x = get_global_id(0), y = get_global_id(1);\n";
        kernelBody << getIndent() << "if (x >= h || y >= w) { return; }\n";
        kernelBody << getIndent() << "const int2 pos = (int2)(x, y);\n";
        */
        kernelBody << getIndent() << "GET_CHECK\n";
        // now just deal elemwise
        kernelBody << addElemwiseOp(nodes);
        //addElemwiseOp(nodes);
        //kernelBody << "\twrite_imagef(output_0, pos, read_imagef(input_0, SAMPLER, pos));\n";
        // kernelBody << "write_imagef(output_0, pos, (float4)(1.0, 1.0, 1.0, 1.0));\n";
        up();
        kernelBody << "}\n";
        // 2. gen kernel prototype
        std::stringstream kernelProto;
        // a) func name
        kernelProto << "__kernel void kernel_" << idx << "(";
        // b) args
        for (auto& input : inputs) {
            kernelProto << "__read_only image2d_t " << varMap[input] << ", ";
        }
        for (auto& output : outputs) {
            kernelProto << "__write_only image2d_t " << varMap[output] << ", ";
        }
        // c) dims info
        kernelProto << "__private const int global_size_dim0, __private const int global_size_dim1, __private const int global_size_dim2)";
        // 3. append to kernel
        kernelCode.append(kernelProto.str());
        kernelCode.append(kernelBody.str());
    }
    std::string codegen() { return kernelCode; }
    std::vector<Tensor*> getInputs() { return inputs; }
    std::vector<Tensor*> getOutputs() { return outputs; }
private:
    std::string addElemwiseOp(std::vector<Node*>& nodes) {
        std::string pos = "pos";
        std::unordered_map<Tensor*, std::string> cacheMap;
        for (auto& node : nodes) {
            std::stringstream ss;
            auto cmd = node->cmd;
            std::vector<std::string> inputs(cmd->inputs.size());
            for (int i = 0; i < cmd->inputs.size(); i++) {
                if (cacheMap.find(cmd->inputs[i]) == cacheMap.end()) {
                    if (cmd->inputs[i]->shape().empty() && TensorUtils::getDescribe(cmd->inputs[i])->usage == Tensor::InsideDescribe::CONSTANT) {
                        float val = cmd->inputs[i]->host<float>()[0];
                        std::stringstream ssval;
                        ssval << "((float4)(" << val << "))";
                        inputs[i] = ssval.str();
                    } else {
                        inputs[i] = readPixel(getNameByTensor(cmd->inputs[i], true), pos);
                    }
                } else {
                    inputs[i] = cacheMap[cmd->inputs[i]];
                    cacheMap.erase(cmd->inputs[i]);
                }
            }
            switch (cmd->op->type()) {
                case MNN::OpType_BinaryOp:
                {
                    auto lhs = inputs[0], rhs = inputs[1];
                    auto type = static_cast<MNN::BinaryOpOperation>(cmd->op->main_as_BinaryOp()->opType());
                    switch (type) {
                        case BinaryOpOperation_ADD:
                            ss << "(" << lhs << "+" << rhs << ")";
                            break;
                        case BinaryOpOperation_SUB:
                            ss << "(" << lhs << "-" << rhs << ")";
                            break;
                        case BinaryOpOperation_MUL:
                            ss << "(" << lhs << "*" << rhs << ")";
                            break;
                        case BinaryOpOperation_POW:
                            ss << "pow(" << lhs << "," << rhs << ")";
                            break;
                        case BinaryOpOperation_DIV:
                            ss << "(" << lhs << "/" << rhs << ")";
                            break;
                        case BinaryOpOperation_MAXIMUM:
                            ss << "fmax(" << lhs << "," << rhs << ")";
                            break;
                        case BinaryOpOperation_MINIMUM:
                            ss << "fmin(" << lhs << "," << rhs << ")";
                            break;
                        case BinaryOpOperation_REALDIV:
                            ss << "(" << lhs << "/" << rhs << ")";
                            break;
                        default:
                            break;
                    }
                    break;
                }
                case MNN::OpType_Eltwise:
                {
                    auto type = cmd->op->main_as_Eltwise()->type();
                    switch (type) {
                        case EltwiseType_SUM:
                        case EltwiseType_SUB:
                        case EltwiseType_PROD:
                        {
                            std::unordered_map<EltwiseType, std::string> elemToOp = {
                                {EltwiseType_PROD, "*"}, {EltwiseType_SUM, "+"}, {EltwiseType_SUB, "-"}
                            };
                            ss << "(" << inputs[0] << elemToOp[type] << inputs[1];
                            for (int i = 2; i < inputs.size(); i++) {
                                ss << elemToOp[type] << inputs[i];
                            }
                            ss << ")";
                            break;
                        }
                        case EltwiseType_MAXIMUM:
                        {
                            std::function<std::string(int)> fmax = [&inputs, &fmax](int d) {
                                if (d == inputs.size() - 1) {
                                    return inputs[d];
                                }
                                return "fmax(" + inputs[d] + ", " + fmax(d+1) + ")";
                            };
                            ss << fmax(0);
                            break;
                        }
                        default:
                            break;
                    }
                    break;
                }
                case MNN::OpType_UnaryOp:
                {
                    auto unary = cmd->op->main_as_UnaryOp();
                    auto type = unary->opType();
                    auto operand = inputs[0];
                    switch (type) {
                        case UnaryOpOperation_SQUARE:
                            ss << operand << " * " << operand;
                            break;
                        case UnaryOpOperation_ERF:
                            ss << "erf(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_ERFC:
                            ss << "erfc(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_SQRT:
                            ss << "sqrt(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_RSQRT:
                            ss << "rsqrt(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_ABS:
                            ss << "fabs(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_SIN:
                            ss << "sin(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_COS:
                            ss << "cos(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_SIGN:
                            ss << "sign(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_EXP:
                            ss << "exp(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_NEG:
                            ss << "-(" << operand << ")";
                            break;
                        case UnaryOpOperation_TAN:
                            ss << "tan(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_CEIL:
                            ss << "ceil(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_LOG1P:
                            ss << "log1p(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_FLOOR:
                            ss << "floor(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_ROUND:
                            ss << "round(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_SIGMOID:
                            ss << "native_recip((float4)1+native_exp(convert_float4(-" << operand << ")))";
                            break;
                        case UnaryOpOperation_TANH:
                            ss << "tanh(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_RECIPROCAL:
                            ss << "native_recip(convert_float4(" << operand << "))";
                            break;
                        case UnaryOpOperation_LOG:
                            ss << "native_log(convert_float4(" << operand << "+(float4)((float)0.0000001)))";
                            break;
                        default:
                            break;
                    }
                    break;
                }
                case MNN::OpType_ReLU6:
                {
                    auto operand = inputs[0];
                    auto relu6 = cmd->op->main_as_Relu6();
                    float minv = relu6->minValue();
                    float maxv = relu6->maxValue();
                    ss << "fmin(fmax(" << operand << "," << getNumVal(minv) << "), " << getNumVal(maxv) << ")";
                    break;
                }
                case MNN::OpType_ReLU:
                {
                    auto operand = inputs[0];
                    auto relu = cmd->op->main_as_Relu();
                    float slope = relu->slope();
                    ss << "fmax(" << operand << "," << getNumVal(0) << ")";
                    break;
                }
                default:
                    break;
            }
            cacheMap[cmd->outputs[0]] = ss.str();
        }
        std::stringstream ss;
        for (auto& iter : cacheMap) {
            auto output = getNameByTensor(iter.first, false);
            ss << writePixel(output, pos, iter.second);
        }
        return ss.str();
    }
    template <typename T>
    std::string getNumVal(T t) {
        return "(float4)((float)" + std::to_string(t) + ")";
    }
    std::string readPixel(std::string img, std::string pos) {
        return "read_imagef(" + img + ", SAMPLER, " + pos + ")";
    }
    std::string writePixel(std::string img, std::string pos, std::string data) {
        return getIndent() + "write_imagef(" + img + ", " + pos + ", " + data + ");\n";
    }
    void down() { indent++; }
    void up() { indent--; }
    std::string getIndent() {
        return std::string(indent*4, ' ');
    }
    std::string getNameByTensor(Tensor* t, bool read) {
        if (varMap.find(t) != varMap.end()) {
            return varMap[t];
        }
        if (read) {
            int idx = inputs.size();
            inputs.push_back(t);
            varMap[t] = "input_" + std::to_string(idx);
            return varMap[t];
        } else {
            int idx = outputs.size();
            outputs.push_back(t);
            varMap[t] = "output_" + std::to_string(idx);;
            return varMap[t];
        }
    }
private:
    std::vector<Node*> nodes;
    std::vector<Tensor*> inputs;
    std::vector<Tensor*> outputs;
    std::unordered_map<const Tensor*, std::string> varMap;
    std::string kernelCode;
    int indent = 0;
};

std::string OpenCLPluginModule::codegen() {
    std::stringstream sourceCode;
    sourceCode << "#define GET_CHECK\\\n\
        const int c = get_global_id(0), w = get_global_id(1), hb = get_global_id(2);\\\n\
        if (c >= global_size_dim0 || w >= global_size_dim1 || hb >= global_size_dim2) { return; }\\\n\
        const int2 pos = (int2)(mad24(c, global_size_dim1, w), hb);\n";
    sourceCode << "__constant sampler_t SAMPLER = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;\n";
    for (int i = 0; i < getFunctionNum(); i++) {
        sourceCode << functions[i]->codegen();
    }
    return sourceCode.str();
}

InOutTensors OpenCLPluginModule::addFunction(std::vector<Node*> nodes) {
    std::unique_ptr<OpenCLPluginFunction> func(new OpenCLPluginFunction(nodes, getFunctionNum()));
    auto res = std::make_pair<std::vector<MNN::Tensor*>, std::vector<MNN::Tensor*>>(func->getInputs(), func->getOutputs());
    functions.emplace_back(std::move(func));
    return res;
}