2019-04-17 10:49:11 +08:00
|
|
|
//
|
|
|
|
// SizeComputer.hpp
|
|
|
|
// MNN
|
|
|
|
//
|
|
|
|
// Created by MNN on 2019/01/23.
|
|
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
|
|
//
|
|
|
|
|
|
|
|
#ifndef SizeComputer_hpp
|
|
|
|
#define SizeComputer_hpp
|
|
|
|
|
2020-11-05 16:41:56 +08:00
|
|
|
#include <MNN/Tensor.hpp>
|
2019-04-17 10:49:11 +08:00
|
|
|
#include <map>
|
|
|
|
#include <string>
|
|
|
|
#include <vector>
|
|
|
|
#include "MNN_generated.h"
|
2020-11-05 16:41:56 +08:00
|
|
|
#include "core/Execution.hpp"
|
2019-12-27 22:16:57 +08:00
|
|
|
#include "core/TensorUtils.hpp"
|
2019-04-17 10:49:11 +08:00
|
|
|
#define FLOPS_M 1000000.0f
|
|
|
|
|
|
|
|
namespace MNN {
|
|
|
|
|
|
|
|
/** computer for op. calculate input and output tensors' shape. when analyzing model, calculate flops too. */
|
2019-08-22 20:13:46 +08:00
|
|
|
class MNN_PUBLIC SizeComputer {
|
2019-04-17 10:49:11 +08:00
|
|
|
friend class SizeComputerSuite;
|
2020-11-05 16:41:56 +08:00
|
|
|
|
2019-04-17 10:49:11 +08:00
|
|
|
public:
|
2019-08-22 20:13:46 +08:00
|
|
|
void setInputIndex(std::vector<int>&& index) {
|
|
|
|
mNeedContentInputIndex = std::move(index);
|
|
|
|
}
|
2019-04-17 10:49:11 +08:00
|
|
|
/**
|
|
|
|
* @brief deinitializer.
|
|
|
|
*/
|
|
|
|
virtual ~SizeComputer() = default;
|
|
|
|
|
|
|
|
public:
|
|
|
|
/**
|
|
|
|
* @brief calculate input and output tensors' shape for given op.
|
|
|
|
* @param op given op.
|
|
|
|
* @param inputs given input tensors.
|
|
|
|
* @param outputs given output tensors.
|
|
|
|
* @return true if success, false otherwise.
|
|
|
|
*/
|
|
|
|
virtual bool onComputeSize(const MNN::Op* op, const std::vector<Tensor*>& inputs,
|
|
|
|
const std::vector<Tensor*>& outputs) const = 0;
|
|
|
|
|
|
|
|
/**
|
|
|
|
* @brief calculate the flops of this op with the info of inputs size.
|
|
|
|
* @param op given op.
|
|
|
|
* @param inputs given input tensors.
|
|
|
|
* @param outputs given output tensors.
|
|
|
|
* @return the flops in M.
|
|
|
|
*/
|
|
|
|
virtual float onComputeFlops(const MNN::Op* op, const std::vector<Tensor*>& inputs,
|
|
|
|
const std::vector<Tensor*>& outputs) const;
|
|
|
|
|
|
|
|
/**
|
|
|
|
* @brief calculate input and output tensors' shape for any registed op.
|
|
|
|
* @param op given registed op.
|
|
|
|
* @param inputs given input tensors.
|
|
|
|
* @param outputs given output tensors.
|
|
|
|
* @return true if success, false otherwise.
|
|
|
|
*/
|
|
|
|
static bool computeOutputSize(const MNN::Op* op, const std::vector<Tensor*>& inputs,
|
|
|
|
const std::vector<Tensor*>& outputs);
|
- dynamic computation graph (beta)
- add supports (/express)
- add tests
- add benchmarks with it (/benchmark/exprModels)
- Python
- MNN engine and tools were submitted to pip
- available on Windows/macOS/Linux
- Engine/Converter
- add supports for each op benchmarking
- refactor optimizer by separating steps
- CPU
- add supports for Conv3D, Pool3D, ELU, ReverseSequence
- fix ArgMax, Permute, Scale, BinaryOp, Slice, SliceTf
- OpenCL
- add half transform in CPU
- add broadcast supports for binary
- optimize Conv2D, Reshape, Eltwise, Gemm, etc.
- OpenGL
- add sub, real div supports for binary
- add supports for unary
- optimize Conv2D, Reshape
- Vulkan
- add max supports for eltwise
- Metal
- fix metallib missing problem
- Train/Quantization
- use express to refactor training codes
2019-09-26 21:02:07 +08:00
|
|
|
static float computeFlops(const MNN::Op* op, const std::vector<Tensor*>& inputs,
|
2020-11-05 16:41:56 +08:00
|
|
|
const std::vector<Tensor*>& outputs);
|
- dynamic computation graph (beta)
- add supports (/express)
- add tests
- add benchmarks with it (/benchmark/exprModels)
- Python
- MNN engine and tools were submitted to pip
- available on Windows/macOS/Linux
- Engine/Converter
- add supports for each op benchmarking
- refactor optimizer by separating steps
- CPU
- add supports for Conv3D, Pool3D, ELU, ReverseSequence
- fix ArgMax, Permute, Scale, BinaryOp, Slice, SliceTf
- OpenCL
- add half transform in CPU
- add broadcast supports for binary
- optimize Conv2D, Reshape, Eltwise, Gemm, etc.
- OpenGL
- add sub, real div supports for binary
- add supports for unary
- optimize Conv2D, Reshape
- Vulkan
- add max supports for eltwise
- Metal
- fix metallib missing problem
- Train/Quantization
- use express to refactor training codes
2019-09-26 21:02:07 +08:00
|
|
|
|
2025-09-22 23:05:26 +08:00
|
|
|
static bool computeBroadCastDims(const std::vector<Tensor*>& inputs,
|
2021-04-08 15:34:23 +08:00
|
|
|
const std::vector<Tensor*>& outputs);
|
2020-11-05 16:41:56 +08:00
|
|
|
|
2021-04-08 15:34:23 +08:00
|
|
|
static std::vector<int> needInputContent(const MNN::Op* op, int inputSize);
|
2019-08-22 20:13:46 +08:00
|
|
|
private:
|
|
|
|
std::vector<int> mNeedContentInputIndex;
|
2019-04-17 10:49:11 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
/** size computer suite */
|
- build:
- unify schema building in core and converter;
- add more build script for android;
- add linux build script for python;
- ops impl:
- add floor mod support in binary;
- use eltwise impl in add/max/sub/mul binary for optimization;
- remove fake double support in cast;
- fix 5d support for concat;
- add adjX and adjY support for batch matmul;
- optimize conv2d back prop filter;
- add pad mode support for conv3d;
- fix bug in conv2d & conv depthwise with very small feature map;
- optimize binary without broacast;
- add data types support for gather;
- add gather ND support;
- use uint8 data type in gather v2;
- add transpose support for matmul;
- add matrix band part;
- add dim != 4 support for padding, reshape & tensor convert;
- add pad type support for pool3d;
- make ops based on TensorFlow Lite quantization optional;
- add all & any support for reduction;
- use type in parameter as output type in reduction;
- add int support for unary;
- add variable weight support for conv2d;
- fix conv2d depthwise weights initialization;
- fix type support for transpose;
- fix grad outputs count for reduce grad and reshape grad;
- fix priorbox & detection output;
- fix metal softmax error;
- python:
- add runSessionWithCallBackInfo interface;
- add max nodes limit (1400) for visualization tool;
- fix save error in python3;
- align default dim;
- convert:
- add extra design for optimization;
- add more post converting optimizers;
- add caffe v1 weights blob support;
- add cast, unary, conv transpose support for onnx model;
- optimize batchnorm, conv with variable weights, prelu, reshape, slice, upsample for onnx model;
- add cos/sin/atan/tan support for unary for tensorflow model;
- add any/all support for reduction for tensorflow model;
- add elu, conv3d, pool3d support for tensorflow model;
- optimize argmax, batchnorm, concat, batch to space, conv with variable weights, prelu, slice for tensorflow model;
- others:
- fix size computer lock;
- fix thread pool deadlock;
- add express & parameters in express;
- rewrite blitter chooser without static map;
- add tests for expr;
2019-10-29 13:37:26 +08:00
|
|
|
class MNN_PUBLIC SizeComputerSuite {
|
2019-04-17 10:49:11 +08:00
|
|
|
public:
|
|
|
|
/**
|
|
|
|
* @brief deinitializer.
|
|
|
|
*/
|
|
|
|
~SizeComputerSuite();
|
|
|
|
/**
|
|
|
|
* @brief get shared instance.
|
|
|
|
* @return shared instance.
|
|
|
|
*/
|
|
|
|
static SizeComputerSuite* get();
|
- build:
- unify schema building in core and converter;
- add more build script for android;
- add linux build script for python;
- ops impl:
- add floor mod support in binary;
- use eltwise impl in add/max/sub/mul binary for optimization;
- remove fake double support in cast;
- fix 5d support for concat;
- add adjX and adjY support for batch matmul;
- optimize conv2d back prop filter;
- add pad mode support for conv3d;
- fix bug in conv2d & conv depthwise with very small feature map;
- optimize binary without broacast;
- add data types support for gather;
- add gather ND support;
- use uint8 data type in gather v2;
- add transpose support for matmul;
- add matrix band part;
- add dim != 4 support for padding, reshape & tensor convert;
- add pad type support for pool3d;
- make ops based on TensorFlow Lite quantization optional;
- add all & any support for reduction;
- use type in parameter as output type in reduction;
- add int support for unary;
- add variable weight support for conv2d;
- fix conv2d depthwise weights initialization;
- fix type support for transpose;
- fix grad outputs count for reduce grad and reshape grad;
- fix priorbox & detection output;
- fix metal softmax error;
- python:
- add runSessionWithCallBackInfo interface;
- add max nodes limit (1400) for visualization tool;
- fix save error in python3;
- align default dim;
- convert:
- add extra design for optimization;
- add more post converting optimizers;
- add caffe v1 weights blob support;
- add cast, unary, conv transpose support for onnx model;
- optimize batchnorm, conv with variable weights, prelu, reshape, slice, upsample for onnx model;
- add cos/sin/atan/tan support for unary for tensorflow model;
- add any/all support for reduction for tensorflow model;
- add elu, conv3d, pool3d support for tensorflow model;
- optimize argmax, batchnorm, concat, batch to space, conv with variable weights, prelu, slice for tensorflow model;
- others:
- fix size computer lock;
- fix thread pool deadlock;
- add express & parameters in express;
- rewrite blitter chooser without static map;
- add tests for expr;
2019-10-29 13:37:26 +08:00
|
|
|
static void init();
|
2019-04-17 10:49:11 +08:00
|
|
|
|
|
|
|
public:
|
|
|
|
/**
|
|
|
|
* @brief register size computer for designated type
|
|
|
|
* @param computer size computer
|
|
|
|
* @param type designated type
|
|
|
|
*/
|
|
|
|
void insert(SizeComputer* computer, OpType type);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* @brief query size computer for designated type
|
|
|
|
* @param type designated type
|
|
|
|
* @return size computer if found, nullptr otherwise.
|
|
|
|
*/
|
|
|
|
SizeComputer* search(OpType type);
|
|
|
|
|
|
|
|
private:
|
|
|
|
/** shared instance */
|
|
|
|
static SizeComputerSuite* gInstance;
|
|
|
|
/** registered size computer */
|
2021-11-30 10:10:53 +08:00
|
|
|
std::vector<SizeComputer*> mRegistry;
|
2019-04-17 10:49:11 +08:00
|
|
|
};
|
|
|
|
|
|
|
|
/** register for size computer */
|
|
|
|
template <class T>
|
|
|
|
class SizeComputerRegister {
|
|
|
|
public:
|
|
|
|
/**
|
|
|
|
* @brief initializer. register size computer to suite.
|
|
|
|
* @param type designated type
|
|
|
|
*/
|
|
|
|
SizeComputerRegister(OpType type) {
|
|
|
|
T* test = new T;
|
|
|
|
SizeComputerSuite* ts = SizeComputerSuite::get();
|
|
|
|
ts->insert(test, type);
|
|
|
|
}
|
2019-08-22 20:13:46 +08:00
|
|
|
SizeComputerRegister(OpType type, std::vector<int>&& index) {
|
2020-11-05 16:41:56 +08:00
|
|
|
T* test = new T;
|
2019-08-22 20:13:46 +08:00
|
|
|
test->setInputIndex(std::move(index));
|
|
|
|
SizeComputerSuite* ts = SizeComputerSuite::get();
|
|
|
|
ts->insert(test, type);
|
|
|
|
}
|
2019-04-17 10:49:11 +08:00
|
|
|
};
|
|
|
|
} // namespace MNN
|
2019-05-09 19:39:33 +08:00
|
|
|
#define REGISTER_SHAPE(name, op) \
|
|
|
|
void ___##name##__##op##__() { \
|
2021-02-07 10:45:07 +08:00
|
|
|
name* _temp = new name; \
|
2019-05-09 19:39:33 +08:00
|
|
|
SizeComputerSuite* ts = SizeComputerSuite::get(); \
|
2021-02-07 10:45:07 +08:00
|
|
|
ts->insert(_temp, op); \
|
2020-11-05 16:41:56 +08:00
|
|
|
}
|
|
|
|
// Output shape is depent on index-input content data
|
|
|
|
#define REGISTER_SHAPE_INPUTS(name, op, index) \
|
|
|
|
void ___##name##__##op##__() { \
|
|
|
|
SizeComputerSuite* ts = SizeComputerSuite::get(); \
|
2021-02-07 10:45:07 +08:00
|
|
|
name* computer = new name; \
|
|
|
|
computer->setInputIndex(index); \
|
|
|
|
ts->insert(computer, op); \
|
2019-05-09 19:39:33 +08:00
|
|
|
}
|
2019-04-17 10:49:11 +08:00
|
|
|
|
2022-07-19 13:52:07 +08:00
|
|
|
#ifdef MNN_SUPPORT_DEPRECATED_OP
|
|
|
|
#define REGISTER_SHAPE_OLD(name, op) \
|
|
|
|
void ___##name##__##op##__() { \
|
|
|
|
name* _temp = new name; \
|
|
|
|
SizeComputerSuite* ts = SizeComputerSuite::get(); \
|
|
|
|
ts->insert(_temp, op); \
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
#define REGISTER_SHAPE_OLD(name, op) void ___##name##__##op##__() {}
|
|
|
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
2023-12-04 11:12:20 +08:00
|
|
|
#ifdef MNN_SUPPORT_RENDER
|
|
|
|
#define REGISTER_SHAPE_INPUTS_RENDER(name, op, index) \
|
|
|
|
void ___##name##__##op##__() { \
|
|
|
|
SizeComputerSuite* ts = SizeComputerSuite::get(); \
|
|
|
|
name* computer = new name; \
|
|
|
|
computer->setInputIndex(index); \
|
|
|
|
ts->insert(computer, op); \
|
|
|
|
}
|
|
|
|
|
|
|
|
#else
|
|
|
|
#define REGISTER_SHAPE_INPUTS_RENDER(name, op, index) void ___##name##__##op##__() {}
|
|
|
|
|
|
|
|
#endif
|
|
|
|
|
2024-02-29 16:21:40 +08:00
|
|
|
#ifdef MNN_SUPPORT_TRANSFORMER_FUSE
|
|
|
|
#define REGISTER_SHAPE_INPUTS_TRANSFORMER_FUSE(name, op) \
|
|
|
|
void ___##name##__##op##__() { \
|
|
|
|
name* _temp = new name; \
|
|
|
|
SizeComputerSuite* ts = SizeComputerSuite::get(); \
|
|
|
|
ts->insert(_temp, op); \
|
|
|
|
}
|
|
|
|
|
|
|
|
#else
|
|
|
|
#define REGISTER_SHAPE_INPUTS_TRANSFORMER_FUSE(name, op) void ___##name##__##op##__() {}
|
|
|
|
|
|
|
|
#endif
|
|
|
|
|
2019-04-17 10:49:11 +08:00
|
|
|
#endif
|