add implementation of 08_bilateral_filter by Jason-Young123

08_bilateral_filter/Jason-Young123/.gitignore

@@ -0,0 +1,8 @@
+/obj/
+/useful/
+/useless/
+/analysis/
+/runTester
+/tmp_env/
+/result/
+init.mk

08_bilateral_filter/Jason-Young123/Makefile

@@ -0,0 +1,145 @@
+WDIR = .
+SRC_DIR = $(WDIR)/src
+OBJ_DIR = $(WDIR)/obj
+INCLUDES_DIR = $(WDIR)/include
+ANALYSIS_DIR = $(WDIR)/analysis
+ENV_DIR = $(WDIR)/tmp_env
+SCRIPTS_DIR = $(WDIR)/scripts
+MK = $(WDIR)/init.mk
+RESULT_DIR = ./result
+LOG = $(RESULT_DIR)/runtime.log
+
+-include $(MK)
+
+PLATFORM        ?= nvidia
+# resolve platform
+ifeq ($(PLATFORM),nvidia)
+    PLATFORM_DEFINE 	:= -DPLATFORM_NVIDIA
+    PREFIX 		:= /usr/local/cuda-12.8/bin/
+    CXX 		:= $(PREFIX)nvcc
+    GDB 		:= $(PREFIX)cuda-gdb
+    NCU 		:= $(PREFIX)ncu
+    NCU-GUI 		:= $(PREFIX)ncu-ui
+    NSYS 		:= $(PREFIX)nsys
+    CXXFLAGS		:= -Xcompiler -std=c++17 -O3 -use_fast_math -Xcudafe --diag_suppress=611 -Wno-deprecated-gpu-targets -I$(INCLUDES_DIR)
+    SUFFIX		:= cu
+else ifeq ($(PLATFORM),iluvatar)
+    PLATFORM_DEFINE 	:= -DPLATFORM_ILUVATAR
+    CXX          	:= clang++
+    CXXFLAGS	        := -std=c++17 -O3 -I$(INCLUDES_DIR) -Wno-implicit-const-int-float-conversion -Wno-literal-range
+    LIBS		:= -lcudart -I/usr/local/corex/include -L/usr/local/corex/lib64 -fPIC
+    SUFFIX  		:= cu
+else ifeq ($(PLATFORM),moore)
+    PLATFORM_DEFINE 	:= -DPLATFORM_MOORE
+    CXX          	:= mcc
+    CXXFLAGS          	:= -std=c++17 -O3 -I$(INCLUDES_DIR) -I/usr/local/musa/include 
+    LIBS		:= -L$(ENV_DIR)/lib -L/usr/lib/gcc/x86_64-linux-gnu/11/ -L/usr/local/musa/lib -lmusart
+    RUNTIME_LIBS	+= /usr/local/musa/lib
+    SUFFIX  		:= mu
+else ifeq ($(PLATFORM),metax)
+    PLATFORM_DEFINE 	:= -DPLATFORM_METAX
+    CXX          	:= mxcc
+    CXXFLAGS		:= -std=c++17 -O3 -I$(INCLUDES_DIR)
+    SUFFIX  		:= maca
+else
+    $(error Unsupported PLATFORM '$(PLATFORM)' (expected: nvidia, iluvatar, moore, metax))
+endif
+-include $(SCRIPTS_DIR)/$(PLATFORM).mk
+
+
+ifeq ($(HAS_CV_ENV), yes)
+    CXXFLAGS += -DHAS_CV	
+endif
+
+
+
+SRCS = $(wildcard $(SRC_DIR)/*.$(SUFFIX))
+OBJS = $(patsubst $(SRC_DIR)/%.$(SUFFIX), $(OBJ_DIR)/%.o, $(SRCS))
+EXEC = $(WDIR)/runTester
+
+
+.DEFAULT_GOAL := all
+
+all: $(EXEC)
+
+
+init:
+	@rm -rf $(ENV_DIR) $(MK)
+	@echo "[INIT] Checking OpenCV Environment..."
+	@$(PLATFORM_INIT_ENV); \
+	if pkg-config --exists opencv4; then \
+		echo "[INFO] Built-in OpenCV Found."; \
+		echo "HAS_CV_ENV = yes" > $(MK); \
+		echo "CVFLAGS = $$(pkg-config --cflags opencv4)" >> $(MK); \
+		echo "CVLIBS = $$(pkg-config --libs opencv4)" >> $(MK); \
+		echo "CVRUNTIME_LIBS = $${LD_LIBRARY_PATH}" >> $(MK); \
+	elif command -v conda > /dev/null 2>&1; then \
+		echo "[INFO] Using Conda to Create $(ENV_DIR)..."; \
+		conda create --prefix $(ENV_DIR) -y libopencv; \
+		echo "HAS_CV_ENV = yes" > $(MK); \
+		echo "CVFLAGS = -I$(ENV_DIR)/include/opencv4" >> $(MK); \
+		echo "CVLIBS = -L$(ENV_DIR)/lib -lopencv_core -lopencv_imgproc -lopencv_imgcodecs" >> $(MK); \
+		echo "CVRUNTIME_LIBS = $(ENV_DIR)/lib:$${LD_LIBRARY_PATH}" >> $(MK); \
+	else \
+		echo "[WARN] No OpenCV is Available on this Server."; \
+		echo "HAS_CV_ENV = no" > $(MK); \
+	fi
+	@echo "[INFO] Project Init Done."
+	@echo "[INFO] Enter \`make (run) PLATFORM=[nvidia, moore, metax, iluvatar]\` to Compile & Run."
+
+
+
+
+#link: .o -> exec
+$(EXEC): $(OBJS)
+	@echo "[INFO] $^ -> $@"
+	@$(CXX) $(CXXFLAGS) $(CVFLAGS) $(LIBS) $(CVLIBS) -o $@ $^
+
+
+#compile: .cu -> .o
+$(OBJ_DIR)/%.o: $(SRC_DIR)/%.$(SUFFIX)
+	@mkdir -p $(OBJ_DIR)
+	@echo "[INFO] $< -> $@"
+	@$(CXX) $(CXXFLAGS) $(CVFLAGS) $(PLATFORM_DEFINE) -c -o $@ $<
+
+
+run: $(EXEC)
+	@mkdir -p $(RESULT_DIR)
+	@export LD_LIBRARY_PATH=$(CVRUNTIME_LIBS):$$LD_LIBRARY_PATH; \
+	$(MAKE) -s -f $(SCRIPTS_DIR)/$(PLATFORM).mk exec LD_LIBRARY_PATH="$$LD_LIBRARY_PATH" 2>&1 | tee $(LOG)
+	@echo "[INFO] Runtime log has been written to $(LOG)"
+
+
+
+gdb: $(EXEC)
+	@$(GDB) $<
+
+
+
+
+#Nsight System,系统级分析工具
+nsys: $(EXEC)
+	@mkdir -p $(ANALYSIS_DIR)
+	@$(NSYS) profile -t cuda,nvtx,osrt -o $(ANALYSIS_DIR)/$(EXEC) -f true $<
+	@$(NSYS) stats $(ANALYSIS_DIR)/$(EXEC).nsys-rep --force-export=true
+
+#Nsight Compute,内核级分析工具
+ncu: $(EXEC)
+	@mkdir -p $(ANALYSIS_DIR)
+	@$(NCU) --print-details all --nvtx --call-stack --set full $<
+
+ncu-gui: $(EXEC)
+	@mkdir -p $(ANALYSIS_DIR)
+	@$(NCU) --nvtx --call-stack --set full -f --export $(ANALYSIS_DIR)/$(EXEC).ncu-rep $<
+	@$(NCU-GUI) $(ANALYSIS_DIR)/$(EXEC).ncu-rep
+
+
+clean:
+	rm -rf $(OBJ_DIR) $(ANALYSIS_DIR) $(EXEC) $(RESULT_DIR)
+
+cleanup:
+	rm -rf $(OBJ_DIR) $(ANALYSIS_DIR) $(EXEC) $(RESULT_DIR)
+	rm -rf $(MK) $(ENV_DIR)
+
+
+.PHONY:

08_bilateral_filter/Jason-Young123/README.md

@@ -0,0 +1,38 @@
+# Bilateral-Filter
+
+
+## 介绍
+InifiniTensor 2025冬训练营项目阶段, CUDA方向题目八: 基于CUDA的实时图像双边滤波
+
+
+## 使用说明
+### step1: 环境初始化
+在开始编译前需配置本地openCV环境（优先使用系统自带openCV, 若缺失则尝试基于conda建立虚拟环境tmp_env并安装openCV; 若conda也不可用则项目将在缺少openCV对照的情况下直接运行）
+```bash
+make init PLATFORM=[nvidia, moore, metax, iluvatar]
+```
+
+### step2: 配置参数文件
+修改`tester/config.txt`以配置radius, sigma_spatial和sigma_color（当radius设置为 <= 0时，代码将自适应选取滤波半径）;  
+在`tester/gray/4K`和`tester/rgb/4K`路径下存放待测试的4K(3840 * 2160)图片的bin文件，分别包含10张灰白壁纸和15张RGB壁纸（具体jpeg图片参见resource下的相应路径）;  
+你也可以在`tester/gray/4K`和`tester/rgb/4K`中自行添加待测试.bin文件
+
+### step3: 编译运行
+输出.bin文件和性能日志(runtime.log)将自动保存至`result`路径下
+```bash
+make run PLATFORM=[nvidia, moore, metax, iluvatar]
+```
+
+
+## 其他
+清除编译和运行结果
+```bash
+make clean
+```
+
+清除所有生成文件(包括虚拟环境和辅助脚本)
+```bash
+make cleanup
+```
+
+基于各平台的结果和性能分析参见`report`路径

08_bilateral_filter/Jason-Young123/include/auxiliary.h

@@ -0,0 +1,83 @@
+//文件读写等辅助函数
+#pragma once
+
+#include <fstream>
+#include <iostream>
+#include <filesystem>
+#include <vector>
+#include <myPixel.h>
+
+
+bool getCfg(const std::string& cfgPath, int& radius, float& sigma_spatial, float& sigma_color);
+
+bool getBin(const std::string& binPath, int& width, int& height, uint8_t*& src1, myPixel*& src2);
+
+float binDiff(const std::string& binPath_ref, const std::string& binPath_test);
+
+
+//生成bin文件
+template <typename T>
+bool genBin(const std::string& outBinPath, const T* img, int width, int height){
+    std::ofstream file(outBinPath, std::ios::binary);
+    if(!file){
+        std::cerr << "Error: Could not open file: " << outBinPath << std::endl;
+        return false;
+    }
+
+    int channels = std::is_same_v<T, myPixel> ? 3 : 1;
+
+    file.write(reinterpret_cast<char*>(&width), sizeof(int));
+    file.write(reinterpret_cast<char*>(&height), sizeof(int));
+    file.write(reinterpret_cast<char*>(&channels), sizeof(int));
+    if constexpr(std::is_same_v<T, myPixel>){
+        for(size_t i = 0; i < width * height; ++i){
+            uint8_t rgb[3] = {img[i].R(), img[i].G(), img[i].B() };
+            file.write(reinterpret_cast<char*>(rgb), 3);
+        }
+    }
+    else{
+        file.write(reinterpret_cast<const char*>(img), width * height);
+    }
+
+    file.close();
+    //std::cout << "Successfully genBin: " << outBinPath << std::endl;
+    return true;
+}
+
+
+//大图坐标相对小图坐标而言,取值范围在-radius ~ len + radius - 1
+inline int mapReflect101(int p, int len){
+    if(p < 0){//左越界
+        return -p;
+    }
+    else if(p >= len){//右越界
+        return 2 * (len - 1) - p;
+    }
+    else{//不越界
+        return p;
+    }
+}
+
+//将输入图像以reflect101模式进行边缘延拓
+template <typename T>
+T* Reflect101(const T* src, int width, int height, int radius){
+    if(radius < 0 || radius >= width || radius >= height){
+        std::cerr << "Error: unsupported radius" << std::endl;
+        return nullptr;
+    }
+
+    int pixel_count_r101 = (width + 2 * radius) * (height + 2 * radius);
+    T* src_r101 = new T[pixel_count_r101];
+
+    for(int i = 0; i < height + 2 * radius; ++i){
+        for(int j = 0; j < width + 2 * radius; ++j){
+            int relative_i = i - radius;
+            int relative_j = j - radius;
+            int src_y = mapReflect101(relative_i, height);//纵坐标
+            int src_x = mapReflect101(relative_j, width);//横坐标
+
+            src_r101[i * (width + 2 * radius) + j] = src[src_y * width + src_x];
+        }
+    }
+    return src_r101;
+}