main update: support cyclesim, rtlsim, gvm backend......

.clang-format

@@ -0,0 +1,9 @@
+BasedOnStyle: LLVM          # 基于LLVM的代码风格
+IndentWidth: 4              # 缩进宽度为4个空格
+AccessModifierOffset: -4    # 访问修饰符的偏移量为-4
+ColumnLimit: 100            # 每行代码的最大列数
+AlwaysBreakAfterDefinitionReturnType: None
+PenaltyReturnTypeOnItsOwnLine: 1000000
+AlignAfterOpenBracket: BlockIndent 
+AllowShortIfStatementsOnASingleLine: true
+

.gitignore

@@ -9,4 +9,6 @@ test
 html/
 cmake-build*
 build
-obj_dir*
+obj_dir*
+compile_commands.json
+.cache

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "dependencies/membox"]
+	path = dependencies/membox
+	url = https://github.com/THU-DSP-LAB/membox.git

CMakeLists.txt

@@ -1,51 +1,24 @@
 cmake_minimum_required(VERSION 3.22)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
 set(PROJECT ventus_driver)
 project(${PROJECT})
 
 set(CMAKE_CXX_STANDARD 17)
-# set(CMAKE_CXX_FLAGS -Wl,--whole-archive -Wl,--no-whole-archive)
-
-include_directories(common)
-include_directories(include)
-include_directories(devices)
-include_directories(devices/verilating_device/page_table)
-include_directories(tests)
-include_directories(driver)
-
-
-set(CMAKE_BUILD_TYPE "Debug")
-
-if(CMAKE_BUILD_TYPE AND (CMAKE_BUILD_TYPE STREQUAL "Debug"))
-    set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -Wall -O0 -g -ggdb")
-    message("Debug mode:${CMAKE_CXX_FLAGS_DEBUG}")
-#    add_executable(test_debug ${src_dirs})
-
-
-elseif(CMAKE_BUILD_TYPE AND (CMAKE_BUILD_TYPE STREQUAL "Release"))
-    set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -Wall -O3")
-    message("Release mode:${CMAKE_CXX_FLAGS_RELEASE}")
-#    add_executable(test_release ${src_dirs})
-else()
-    message("else:${CMAKE_BUILD_TYPE}")
-    message("else:${CMAKE_CXX_FLAGS_RELEASE}")
-#    add_executable(test_release ${src_dirs})
-endif()
-
-if(EXISTS "${CMAKE_SOURCE_DIR}/test.cpp")
-    set(CODING_TEST test)
-    add_executable(${CODING_TEST} test.cpp)
-endif()
-
-option(ENABLE_INSTALL "if install driver library to install dir" OFF)
-
-set(DRIVER_LIB_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/)
-message(STATUS "DRIVER_LIB_INSTALL_DIR:" ${DRIVER_LIB_INSTALL_DIR} "\n")
 
-option(ENABLE_VERILATOR "if add verilated rtl device" OFF)
+set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -Wall")
+set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -Wall")
 
-add_subdirectory(devices)
+add_subdirectory(dependencies/membox)
+add_subdirectory(common)
 add_subdirectory(driver)
-if(ENABLE_VERILATOR)
-add_subdirectory(tests)
-endif()
-add_subdirectory(codetests)
+add_subdirectory(codetests)
+
+set(DRIVER_PUBLIC_HEADERS
+    ${CMAKE_CURRENT_SOURCE_DIR}/include/ventus.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/common/vt_utils.h
+)
+install(FILES ${DRIVER_PUBLIC_HEADERS}
+    DESTINATION include COMPONENT "include"
+)

README.md

@@ -1,4 +1,97 @@
 # ventus-driver
-driver code for [ventus-gpgpu](https://github.com/THU-DSP-LAB/ventus-gpgpu)
 
-To build this repository, see [llvm-ventus](https://github.com/THU-DSP-LAB/llvm-project).
+## English
+[中文版 Chinese](#中文)
+
+Driver code for Ventus GPGPU project [ventus-gpgpu](https://github.com/THU-DSP-LAB/ventus-gpgpu). It is intended to be used together with other Ventus toolchain projects. See [ventus-env](https://github.com/THU-DSP-LAB/ventus-env).
+
+### Install
+- Recommended: use [ventus-env](https://github.com/THU-DSP-LAB/ventus-env) to deploy the Ventus environment and build via the script build-ventus.sh.
+
+- Manual CMake build:
+```bash
+cmake -G Ninja -B build/ -S . \
+  -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+  -DCMAKE_INSTALL_PREFIX=../install \
+  -DVENTUS_INSTALL_PREFIX=../install \
+  -DSPIKE_SRC_DIR=../spike \
+  -DDRIVER_ENABLE_AUTOSELECT=ON \
+  -DDRIVER_ENABLE_RTLSIM=ON \
+  -DDRIVER_ENABLE_CYCLESIM=ON
+cmake --build build/
+cmake --install build/
+```
+
+### Usage
+This repository provides the driver implementation for Ventus GPGPU, mainly handling memory management and device connectivity.
+
+- Upstream (host): connects to an OpenCL implementation ([POCL](https://github.com/THU-DSP-LAB/pocl)).
+- Downstream (devices): supports multiple (simulation) backends:
+  - Instruction-level simulator [spike](https://github.com/THU-DSP-LAB/ventus-gpgpu-isa-simulator), path: driver/spike_device
+  - SystemC-based cycle-accurate simulator [cyclesim](https://github.com/THU-DSP-LAB/ventus-gpgpu-cpp-simulator), path: driver/cyclesim_device
+  - Verilator-based [Chisel RTL](https://github.com/THU-DSP-LAB/ventus-gpgpu) simulation framework [sim-verilator](https://github.com/THU-DSP-LAB/ventus-gpgpu/tree/master/sim-verilator), path: driver/rtlsim_device
+  - The helper backend driver/auto_select allows choosing a backend via the environment variable VENTUS_BACKEND (see Usage).
+- Both upstream and downstream connections are via shared libraries.
+
+This repository builds multiple shared libraries installed into the specified prefix. After configuring environment variables as in [ventus-env](https://github.com/THU-DSP-LAB/ventus-env), you can run OpenCL programs and the driver will be invoked automatically.
+
+### Environment variables
+- `VENTUS_BACKEND` selects the device backend. Allowed values: spike/isa, rtl/rtlsim/gpgpu, cycle/cyclesim/systemc/simulator.
+- `VENTUS_WAVEFORM=1` enables waveform dump: fst for rtlsim, vcd for cyclesim.
+- `VENTUS_WAVEFORM_BEGIN` and `VENTUS_WAVEFORM_END` define a time window to limit waveform dump for rtlsim (speeds up simulation). Not supported by cyclesim.
+- `VENTUS_DUMP_RESULT=filename.json` saves all device-to-host copied data and their device addresses into a JSON file for debugging.
+- `VENTUS_TIMING_DDR=0` disables DDR timing modeling in cyclesim (enabled by default). RTL does not support DDR timing yet.
+
+### Example
+```bash
+VENTUS_BACKEND=rtl VENTUS_DUMP_RESULT=app.rtl.json VENTUS_WAVEFORM=1 ./OpenCLapp.out 2>&1 | tee rtl.log
+```
+
+---
+
+## 中文
+[English Version](#english)
+
+这是[ventus-gpgpu](https://github.com/THU-DSP-LAB/ventus-gpgpu)的驱动程序，目前仅支持仿真环境，需与其他Ventus工具链项目配合使用，参见 [ventus-env](https://github.com/THU-DSP-LAB/ventus-env)。
+
+### Install
+推荐使用 [ventus-env](https://github.com/THU-DSP-LAB/ventus-env) 部署 Ventus 环境，使用其中的 build-ventus.sh 脚本来编译安装。
+
+手动 cmake 编译命令：
+```bash
+cmake -G Ninja -B build/ -S . \
+  -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+  -DCMAKE_INSTALL_PREFIX=../install \
+  -DVENTUS_INSTALL_PREFIX=../install \
+  -DSPIKE_SRC_DIR=../spike \
+  -DDRIVER_ENABLE_AUTOSELECT=ON \
+  -DDRIVER_ENABLE_RTLSIM=ON \
+  -DDRIVER_ENABLE_CYCLESIM=ON
+cmake --build build/
+cmake --install build/
+```
+
+### Usage
+本仓库作为乘影 Ventus GPGPU 的驱动实现，主要完成内存管理与设备连接的功能。
+- 向上层连接 OpenCL 实现（[POCL](https://github.com/THU-DSP-LAB/pocl)）。
+- 向底层连接多种（仿真）设备，目前支持：
+  - 指令级仿真器 [spike](https://github.com/THU-DSP-LAB/ventus-gpgpu-isa-simulator)，对应 driver/spike_device
+  - 基于 SystemC 的周期级仿真器 [cyclesim](https://github.com/THU-DSP-LAB/ventus-gpgpu-cpp-simulator)，对应 driver/cyclesim_device
+  - 基于 Verilator 搭建的 [Chisel RTL](https://github.com/THU-DSP-LAB/ventus-gpgpu) 仿真框架 [sim-verilator](https://github.com/THU-DSP-LAB/ventus-gpgpu/tree/master/sim-verilator)，对应 driver/rtlsim_device
+  - 另有 driver/auto_select，用户可通过环境变量 VENTUS_BACKEND 指定具体底层设备（见 Usage）。
+- 向上层与向底层的连接均以动态库链接的形式完成。
+
+本仓库会编译出多个动态库安装到指定目录下。用户参照 [ventus-env](https://github.com/THU-DSP-LAB/ventus-env) 配置好环境变量后运行 OpenCL 程序即可自动调用。
+
+### 环境变量
+- VENTUS_BACKEND 选择底层设备，可选值：spike/isa，rtl/rtlsim/gpgpu，cycle/cyclesim/systemc/simulator。
+- VENTUS_WAVEFORM=1 时，rtlsim 导出 fst 波形，cyclesim 导出 vcd 波形。
+- 设定 VENTUS_WAVEFORM_BEGIN 与 VENTUS_WAVEFORM_END（数字）可使 rtlsim 仅导出该时间段内波形，加速仿真；cyclesim 不支持。
+- VENTUS_DUMP_RESULT=filename.json 将所有从 device 端拷回 host 端的数据及其设备端地址保存到指定 JSON 文件，辅助调试。
+- VENTUS_TIMING_DDR=0 关闭 cyclesim 中的 DDR 时序仿真（默认开启）。RTL 暂不支持 DDR 时序仿真。
+
+### 示例
+```bash
+VENTUS_BACKEND=rtl VENTUS_DUMP_RESULT=app.rtl.json VENTUS_WAVEFORM=1 ./OpenCLapp.out 2>&1 | tee rtl.log
+```
+

common/CMakeLists.txt

@@ -0,0 +1,20 @@
+set(PROJECT driver_common_utils)
+project(${PROJECT})
+
+file(GLOB_RECURSE SRCS ./loadelf.cpp)
+
+add_library(${PROJECT} STATIC ${SRCS})
+
+target_link_libraries(${PROJECT} PRIVATE elf)
+target_link_libraries(${PROJECT} PRIVATE spdlog)
+target_link_libraries(${PROJECT} PRIVATE fmt)
+target_include_directories(${PROJECT} PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
+
+set_target_properties(${PROJECT} PROPERTIES OUTPUT_NAME "${PROJECT}")
+set_target_properties(${PROJECT} PROPERTIES CLEAN_DIRECT_OUTPUT 1)
+set_target_properties(${PROJECT} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+
+# if(ENABLE_INSTALL)
+#     message(STATUS "DRIVER_LIB_INSTALL_DIR:" ${DRIVER_LIB_INSTALL_DIR})
+#     install(TARGETS ${PROJECT} LIBRARY DESTINATION ${DRIVER_LIB_INSTALL_DIR} COMPONENT "lib")
+# endif()

common/loadelf.cpp

@@ -0,0 +1,94 @@
+#include "loadelf.hpp"
+#include <cstdio>
+#include <cstdlib>
+#include <fcntl.h>
+#include <gelf.h>
+#include <libelf.h>
+#include <memory>
+#include <spdlog/spdlog.h>
+#include <unistd.h>
+
+// 分析 ELF 文件，返回所有加载到内存中的段信息
+std::vector<MemBlock> get_data_from_elf(const char *filename, std::shared_ptr<spdlog::logger> logger) {
+    std::vector<MemBlock> blocks;
+
+    // 初始化 libelf 库
+    if (elf_version(EV_CURRENT) == EV_NONE) {
+        logger->error("ELF: cannot initialize libelf");
+        return std::vector<MemBlock>();
+    }
+
+    // 打开 ELF 文件
+    int fd = open(filename, O_RDONLY);
+    if (fd < 0) {
+        const char *errstr = strerrordesc_np(errno);
+        logger->error("ELF: cannot open file '{}': {}", filename, errstr);
+        return std::vector<MemBlock>();
+    }
+
+    // 开始 ELF 处理
+    Elf *e = elf_begin(fd, ELF_C_READ, nullptr);
+    if (!e) {
+        logger->error("ELF: elf_begin failed: {}", elf_errmsg(-1));
+        close(fd);
+        return std::vector<MemBlock>();
+    }
+
+    // 读取 ELF 头部
+    GElf_Ehdr ehdr;
+    if (gelf_getehdr(e, &ehdr) == nullptr) {
+        logger->error("ELF: gelf_getehdr failed: {}", elf_errmsg(-1));
+        elf_end(e);
+        close(fd);
+        return std::vector<MemBlock>();
+    }
+
+    // 获取程序头表中的段数量
+    size_t phnum = 0;
+    if (elf_getphdrnum(e, &phnum) != 0) {
+        logger->error("ELF: elf_getphdrnum failed: {}", elf_errmsg(-1));
+        elf_end(e);
+        close(fd);
+        return std::vector<MemBlock>();
+    }
+
+    // 遍历所有程序头
+    for (size_t i = 0; i < phnum; ++i) {
+        GElf_Phdr phdr;
+        if (gelf_getphdr(e, i, &phdr) != &phdr) {
+            logger->error("ELF: gelf_getphdr header {} failed: {}", i, elf_errmsg(-1));
+            return std::vector<MemBlock>();
+        }
+
+        // 只考虑加载段
+        if (phdr.p_type == PT_LOAD) {
+            MemBlock block;
+            block.vaddr = phdr.p_vaddr;
+            block.memsz = phdr.p_memsz;
+
+            // 如果段在文件中有初始化数据，则读取数据
+            if (phdr.p_filesz > 0) {
+                block.data.resize(phdr.p_filesz);
+                if (lseek(fd, phdr.p_offset, SEEK_SET) == -1) {
+                    const char *errstr = strerrordesc_np(errno);
+                    logger->error("ELF: failed seeking to offset {}: {}", phdr.p_offset, errstr);
+                    return std::vector<MemBlock>();
+                } else {
+                    ssize_t bytesRead = read(fd, block.data.data(), phdr.p_filesz);
+                    if (bytesRead != (ssize_t)phdr.p_filesz) {
+                        const char *errstr = strerrordesc_np(errno);
+                        logger->error(
+                            "ELF: failed reading {} bytes data: {}", phdr.p_filesz, errstr
+                        );
+                        return std::vector<MemBlock>();
+                    }
+                }
+            }
+            blocks.push_back(block);
+        }
+    }
+
+    elf_end(e);
+    close(fd);
+    return blocks;
+}

common/loadelf.hpp

@@ -0,0 +1,16 @@
+#pragma once
+
+#include <cstdint>
+#include <memory>
+#include <spdlog/spdlog.h>
+#include <vector>
+
+// 解析ELF后返回的需要分配的内存块信息
+typedef struct MemBlock {
+    uint64_t vaddr;            // 内存块的起始地址
+    size_t memsz;              // 内存块所需分配大小
+    std::vector<uint8_t> data; // 此内存块的初始化数据
+    // 若(filesz=)data.size < memsz则需要补0到memsz大小
+} MemBlock;
+
+std::vector<MemBlock> get_data_from_elf(const char *filename, std::shared_ptr<spdlog::logger> logger);

common/utils.hpp

@@ -0,0 +1,45 @@
+#include <algorithm>
+#include <cctype>
+#include <cerrno>
+#include <cstdint>
+#include <cstdlib>
+#include <optional>
+#include <string>
+
+inline std::optional<bool> parse_bool(std::string str) {
+    // transform to lowercase safely (unsigned char cast to avoid UB)
+    std::transform(str.begin(), str.end(), str.begin(), [](char c) {
+        return static_cast<char>(std::tolower(c));
+    });
+    if (str == "true" || str == "1" || str == "yes" || str == "on") return true;
+    if (str == "false" || str == "0" || str == "no" || str == "off") return false;
+    return std::nullopt;
+}
+inline std::optional<bool> parse_bool(const char *str) {
+    if (str == nullptr) return std::nullopt;
+    return parse_bool(std::string(str));
+}
+
+inline std::optional<uint64_t> parse_u64(const char *s) {
+    if (!s || *s == '\0') return std::nullopt;
+
+    // 手动处理二进制 0b/0B
+    if (s[0] == '0' && (s[1] == 'b' || s[1] == 'B')) {
+        const char *p = s + 2;
+        if (*p == '\0') return std::nullopt;
+        uint64_t v = 0;
+        while (*p == '0' || *p == '1') {
+            v = (v << 1) | (*p - '0');
+            ++p;
+        }
+        if (*p != '\0') return std::nullopt;
+        return v;
+    }
+
+    errno = 0;
+    char *endp = nullptr;
+    // base = 0 自动识别: 0x/0X -> 16, 前导0 -> 8, 否则 -> 10
+    unsigned long long v = std::strtoull(s, &endp, 0);
+    if (errno != 0 || endp == s || *endp != '\0') return std::nullopt;
+    return static_cast<uint64_t>(v);
+};