Merge pull request #40 from InfiniTensor/dev

Dev

.gitignore

@@ -1,11 +1,16 @@
 .vscode/
+.cache/
 /build/
 
 *.egg-info/
 __pycache__/
 *.so
 *.log
 *.onnx
+*.pb
+*.bin
+*.npy
+*.meta
 
 /scripts/*.py
 !/scripts/format.py

.gitmodules

@@ -12,7 +12,7 @@
 	url = git@github.com:MengRao/fmtlog.git
 [submodule "3rd-party/result"]
 	path = 3rd-party/result
-	url = git@github.com:oktal/result.git
+	url = git@github.com:willowell/result.git
 [submodule "3rd-party/abseil-cpp"]
 	path = 3rd-party/abseil-cpp
 	url = git@github.com:abseil/abseil-cpp.git

CMakeLists.txt

@@ -1,12 +1,15 @@
-cmake_minimum_required(VERSION 3.8 FATAL_ERROR)
+cmake_minimum_required(VERSION 3.12 FATAL_ERROR)
 project(refactor_graph VERSION 0.0.0 LANGUAGES CXX)
+message(STATUS "Project " ${PROJECT_NAME} " version " ${PROJECT_VERSION})
 
 option(ABSL_PROPAGATE_CXX_STD "Abseil need this option" ON)
-option(USE_CUDA "Build with cuda" OFF)
+option(USE_CUDA "Support Nvidia GPU" OFF)
+option(USE_KUNLUN "Support Baidu Kunlunxin" OFF)
 
-set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
 set(CMAKE_CXX_EXTENSIONS OFF)
+set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 
 if(USE_CUDA)
     add_compile_definitions(USE_CUDA)
@@ -24,9 +27,21 @@ if(USE_CUDA)
     message(STATUS "CMAKE_CUDA_ARCHITECTURES set to " ${CMAKE_CUDA_ARCHITECTURES})
 endif()
 
-# add_compile_options(-march=native) # this will cause error in some machine
+if(USE_KUNLUN)
+    add_compile_definitions(USE_KUNLUN)
+    if (DEFINED KUNLUN_HOME)
+        set(KUNLUN_HOME ${KUNLUN_HOME} CACHE STRING "KUNLUN_HOME directory for Kunlun development")
+    elseif(DEFINED ENV{KUNLUN_HOME})
+        set(KUNLUN_HOME $ENV{KUNLUN_HOME} CACHE STRING "KUNLUN_HOME directory for Kunlun development")
+    else()
+        message(FATAL_ERROR "KUNLUN_HOME is not defined from cmake or env")
+    endif()
+    message(STATUS "KUNLUN_HOME: ${KUNLUN_HOME}")
+endif()
+
+add_compile_options(-march=native) # this will cause error in some machine
 add_compile_options(-mtune=native)
-add_compile_options(-fPIC)
+add_compile_options(-Wall)
 
 add_subdirectory(3rd-party/backward-cpp)
 
@@ -39,20 +54,16 @@ add_subdirectory(3rd-party/googletest)
 include_directories(3rd-party/fmtlog)
 add_definitions(-D FMTLOG_HEADER_ONLY)
 
-# disable warning in 3rd-party/abseil-cpp
-# see <https://github.com/abseil/abseil-cpp/pull/1287>
-add_compile_options(-w)
 include_directories(3rd-party/abseil-cpp)
 add_subdirectory(3rd-party/abseil-cpp)
-add_compile_options(-Wall)
 
 include_directories(3rd-party/result)
 
 enable_testing()
 
 add_subdirectory(src/00common)
 add_subdirectory(src/01graph_topo)
-add_subdirectory(src/02mem_manager)
+add_subdirectory(src/02hardware)
 add_subdirectory(src/03runtime)
 add_subdirectory(src/04kernel)
 add_subdirectory(src/05computation)

Makefile

@@ -1,13 +1,15 @@
 .PHONY : build install-python reconfig clean clean-log format test
 
-TYPE ?= debug
+TYPE ?= Debug
 CUDA ?= OFF
+KUNLUN ?= OFF
 
-FORMAT_ORIGIN ?=
+CMAKE_EXTRA =
+# CMAKE_EXTRA += -DCMAKE_CXX_COMPILER=
 
 build:
 	mkdir -p build
-	cmake -DCMAKE_BUILD_TYPE=$(TYPE) -DUSE_CUDA=$(CUDA) -Bbuild
+	cmake -Bbuild -DCMAKE_BUILD_TYPE=$(TYPE) -DUSE_CUDA=$(CUDA) -DUSE_KUNLUN=$(KUNLUN) $(CMAKE_EXTRA)
 	make -j -C build
 
 install-python: build
@@ -27,6 +29,3 @@ clean-log:
 
 test:
 	make test -j -Cbuild
-
-format:
-	@python3 scripts/format.py $(FORMAT_ORIGIN)

README.md

@@ -8,6 +8,7 @@
 
 - [安装](#安装)
 - [使用前端](#使用前端)
+  - [调试功能](#调试功能)
 - [项目结构](#项目结构)
   - [构建系统](#构建系统)
   - [子项目简介](#子项目简介)
@@ -25,17 +26,18 @@
 import sys
 import numpy as np
 from onnx import load
-from refactor_graph.onnx import make_compiler
+from refactor_graph.onnx import make_compiler, find_device
 from onnxruntime import InferenceSession
 
 model = load(sys.argv[1])  # ------------------------------------ 加载模型
 input = np.random.random((10, 3, 224, 224)).astype(np.float32)  # 加载测试样本
 
 compiler = make_compiler(model)  # ------------------------------ 模型导入到编译器
 compiler.substitute("N", 10)  # --------------------------------- 代换输入中的变量
+find_device("nvidia", 0)  # ------------------------------------- 初始化指定加速硬件
 executor = compiler.compile("cuda", "default", [])  # ----------- 编译模型（选择平台、分配器和优化选项）
 executor.set_input(0, input)  # --------------------------------- 设置输入
-executor.prepare()  # ------------------------------------------- 准备推理（分配输出空间）
+executor.dispatch(find_device("nvidia", 1), "default")  # ------- 执行器可以随时调度到另一个硬件
 executor.run()  # ----------------------------------------------- 推理
 
 session = InferenceSession(model.SerializeToString())  # -------- 与 onnxruntime 对比结果以验证推理
@@ -59,6 +61,46 @@ executor = compiler.compile("cuda", "default", [])  # -------- 编译模型
 # 下同
 ```
 
+### 调试功能
+
+项目现已依托前端提供多种调试功能。
+
+1. 列出算子信息
+
+   ```python
+   executor.dbg()
+   ```
+
+2. 保存运行中间结果
+
+   ```python
+   executor.trace("path_to_store_files", "data_file_format")
+   ```
+
+   调用这个方法将启动一次模型推理，并在每个算子推理完成后将算子的所有输入输出张量保存到 `path_to_store_files` 参数指示的目录中。
+   如果目录不存在，将创建此目录。每个张量保存到一个文件，已存在的同名文件将被删除。
+   同时，为每个算子创建一个元信息文本文件，命名为 `node<N>.meta`，其内容具有下述格式：
+
+   ```plaintext
+   <NodeName>\t<N>
+   <input/output>\t<K>\t<EdgeName>\t[FileName]
+   ```
+
+   - `NodeName`: 节点的名字；
+   - `N`: 节点序号；
+   - `input/output`: 张量是节点的输入/输出；
+   - `K`: 输入/输出的序号；
+   - `EdgeName`: 张量的名字；
+   - `FileName`: (optional) 数据文件名。如果张量无效，不会保存数据文件，则文件名为空；
+
+3. 逐算子计时
+
+   ```python
+   executor.bench(<sync>)
+   ```
+
+   对每次推理计时。`sync` 是一个指示是否在每次推理后插入同步的布尔参数，若设置为 `False`，则计时可能是推理异步启动的时间。
+
 ## 项目结构
 
 ### 构建系统
@@ -122,8 +164,8 @@ executor = compiler.compile("cuda", "default", [])  # -------- 编译模型
 - [fmtlog v2.2.1](https://github.com/MengRao/fmtlog/releases/tag/v2.2.1)
 - [googletest v1.14.0](https://github.com/google/googletest/releases/tag/v1.14.0)
 - [backward-cpp v1.6](https://github.com/bombela/backward-cpp/releases/tag/v1.6)
-- [result master](https://github.com/oktal/result)
-- [abseil-cpp 20230802.0](https://github.com/abseil/abseil-cpp/releases/tag/20230802.0)
+- [result master](https://github.com/willowell/result)
+- [abseil-cpp 20230802.1](https://github.com/abseil/abseil-cpp/releases/tag/20230802.1)
 
 ## 技术要点
 

scripts/format.py

@@ -1,4 +1,4 @@
-import sys
+import sys, os
 from pathlib import Path
 from subprocess import run
 
@@ -48,3 +48,15 @@ def format_file(file):
             assert files[0][:2] == "a/"
             assert files[1][:2] == "b/"
             format_file(files[1][2:])
+
+
+def format_everything(dir):
+    if os.path.isdir(dir):
+        for root, _, files in os.walk(dir):
+            for file in files:
+                format_file(os.path.join(root, file))
+    else:
+        format_file(dir)
+
+
+format_everything("src")

src/00common/CMakeLists.txt

@@ -1,5 +1,6 @@
-cmake_minimum_required(VERSION 3.10)
-project(common)
+cmake_minimum_required(VERSION 3.12 FATAL_ERROR)
+project(common VERSION 0.0.0 LANGUAGES CXX)
+message(STATUS "Project " ${PROJECT_NAME} " version " ${PROJECT_VERSION})
 
 file(GLOB_RECURSE COMMON_SRC src/*.cc)
 add_library(common STATIC ${COMMON_SRC})

src/00common/include/common/bf16_t.h

@@ -18,9 +18,9 @@ namespace refactor {
         constexpr static uint16_t MASK_SIGN16 = 0b1'00000'00000'00000;
 
     public:
-        constexpr bf16_t() noexcept : code(converter{0.0}.u16[1]) {}
         constexpr bf16_t(uint16_t code) noexcept : code(code) {}
-        constexpr bf16_t(float value) noexcept : code(converter{value}.u16[1]) {}
+        constexpr bf16_t(float value) noexcept : bf16_t(converter{value}.u16[1]) {}
+        constexpr bf16_t() noexcept : bf16_t(0.f) {}
         constexpr bf16_t(bf16_t const &) noexcept = default;
         constexpr bf16_t(bf16_t &&) noexcept = default;
 

src/00common/include/common/error_handler.h

@@ -25,7 +25,9 @@ namespace refactor {
     [&]() -> T {                                                                                     \
         throw refactor::UnreachableError(ERROR_MSG(fmt::format("Unreachable: " #F, ##__VA_ARGS__))); \
     }()
-#define UNREACHABLE() UNREACHABLEX(void, "no message")
+#define UNREACHABLE()                 \
+    UNREACHABLEX(void, "no message"); \
+    std::abort()
 
 #ifndef DISABLE_ASSERT
 #define ASSERT(CONDITION, F, ...)                                                      \

src/00common/include/common/range.h

@@ -44,7 +44,9 @@ namespace refactor {
     };
 
     template<class t = size_t> range_t<t> range0_(t end) {
-        ASSERT(end >= 0, "end must be greater than 0");
+        if constexpr (std::is_signed_v<t>) {
+            ASSERT(end >= 0, "end must be greater than 0");
+        }
         return {0, end};
     }
     template<class t = size_t> range_t<t> range(t begin, t end) {

src/00common/include/common/rc.hpp

@@ -1,9 +1,7 @@
 #ifndef RC_HPP
 #define RC_HPP
 
-#include <cstddef>
 #include <functional>
-#include <utility>
 
 namespace refactor {
 
@@ -20,7 +18,7 @@ namespace refactor {
         T *_value;
         struct Counter {
             size_t strong, weak;
-        } *_counter;
+        } * _counter;
 
         Rc(T *ptr, Counter *counter) noexcept
             : _value(ptr), _counter(counter) { inc(); }
@@ -38,7 +36,7 @@ namespace refactor {
         }
 
     public:
-        explicit Rc(T *ptr) noexcept
+        explicit Rc(T *ptr)
             : _value(ptr),
               _counter(nullptr) {
             if (!ptr) { return; }

src/00common/include/common/slice.h

@@ -1,6 +1,8 @@
 #ifndef SLICE_H
 #define SLICE_H
 
+#include <cstddef>
+
 namespace refactor {
     template<class t>
     struct slice_t {

src/00common/src/data_type.cc

@@ -1,6 +1,5 @@
 #include "common/data_type.h"
 #include "common/error_handler.h"
-#include <fmt/core.h>
 #include <unordered_set>
 
 namespace refactor {
@@ -66,35 +65,35 @@ namespace refactor {
 
     bool DT::isIeee754() const noexcept {
         static const std::unordered_set<Enum> set{DT::F32, DT::FP16, DT::F64};
-        return set.find(internal) != set.end();
+        return set.contains(internal);
     }
     bool DT::isFloat() const noexcept {
         static const std::unordered_set<Enum> set{DT::F32, DT::FP16, DT::F64, DT::BF16};
-        return set.find(internal) != set.end();
+        return set.contains(internal);
     }
     bool DT::isSignedLarge() const noexcept {
         static const std::unordered_set<Enum> set{
             DT::F32, DT::FP16, DT::BF16, DT::F64, DT::I32, DT::I64};
-        return set.find(internal) != set.end();
+        return set.contains(internal);
     }
     bool DT::isSigned() const noexcept {
         static const std::unordered_set<Enum> set{
             DT::F32, DT::FP16, DT::BF16, DT::F64,
             DT::I8, DT::I16, DT::I32, DT::I64};
-        return set.find(internal) != set.end();
+        return set.contains(internal);
     }
     bool DT::isNumberic() const noexcept {
         static const std::unordered_set<Enum> set{
             DT::F32, DT::U8, DT::I8, DT::U16, DT::I16,
             DT::I32, DT::I64, DT::FP16, DT::F64,
             DT::U32, DT::U64, DT::BF16};
-        return set.find(internal) != set.end();
+        return set.contains(internal);
     }
     bool DT::isCpuNumberic() const noexcept {
         static const std::unordered_set<Enum> set{
             DT::F32, DT::U8, DT::I8, DT::U16, DT::I16,
             DT::I32, DT::I64, DT::F64, DT::U32, DT::U64};
-        return set.find(internal) != set.end();
+        return set.contains(internal);
     }
     bool DT::isBool() const noexcept {
         return internal == DT::Bool;

src/01graph_topo/CMakeLists.txt

@@ -1,5 +1,6 @@
-cmake_minimum_required(VERSION 3.10)
-project(graph_topo)
+cmake_minimum_required(VERSION 3.12 FATAL_ERROR)
+project(graph_topo VERSION 0.0.0 LANGUAGES CXX)
+message(STATUS "Project " ${PROJECT_NAME} " version " ${PROJECT_VERSION})
 
 file(GLOB_RECURSE GRAPH_TOPO_SRC src/*.cc src/*.cpp)
 add_library(graph_topo STATIC ${GRAPH_TOPO_SRC})

src/01graph_topo/include/graph_topo.h

@@ -4,6 +4,6 @@
 #include "graph_topo/builder.hpp"
 #include "graph_topo/inplace_modifier.h"
 #include "graph_topo/polymorph_graph.hpp"
-#include "graph_topo/searcher.h"
+#include "graph_topo/searcher.hh"
 
 #endif// GRAPH_TOPO_GRAPH_TOPO_H