An Example: Adding a Custom Analysis Tool

This section demonstrates how to extend AccelProf by adding a new tool for basic GPU performance analysis. The steps include integrating the tool model, implementing analysis logic, registering the tool, adding a GPU-side patch, and enabling it in the backend runtime.

---

Step 1: Add Tool Model

Add app_analysis to the tool detection logic and internal enumeration types.

Modify the Tool Dispatcher

In bin/accelprof:

...
+elif [ ${TOOL} == "app_analysis" ]
+then
+    export YOSEMITE_TOOL_NAME=app_analysis
...

Add to tool_type.h

typedef enum {
    CODE_CHECK = 0,
    APP_METRICE = 1,
    MEM_TRACE = 2,
    HOT_ANALYSIS = 3,
    UVM_ADVISOR = 4,
+    APP_ANALYSIS = 5,
    TOOL_NUMS = 6
} AnalysisTool_t;

Add to sanalyzer.h

typedef enum {
    GPU_NO_PATCH = 0,
    GPU_PATCH_APP_METRIC = 1,
    GPU_PATCH_MEM_TRACE = 2,
    GPU_PATCH_HOT_ANALYSIS = 3,
    GPU_PATCH_UVM_ADVISOR = 4,
+    GPU_PATCH_APP_ANALYSIS = 5,
} SanitizerPatchName_t;

---

Step 2: Implement Analysis Logic

Define the structure of your analysis tool by inheriting from the Tool base class.

app_analysis.h

// ...header file defining the AppAnalysis class

app_analysis.cpp

// ...implementation of event callbacks and GPU-side analysis logic

Each callback handles a specific type of event (e.g., memory allocations, kernel launches). You can modify or extend the callbacks for custom behavior.

---

Step 3: Enable the Tool in Sanalyzer

In sanalyzer.cpp, register the new tool and its associated patch.

+ #include "tools/app_analysis.h"

+ else if (std::string(tool_name) == "app_analysis") {
+    tool = APP_ANALYSIS;
+    _tools.emplace(APP_ANALYSIS, std::make_shared<AppAnalysis>());
+ }

+ else if (tool == APP_ANALYSIS) {
+    options.patch_name = GPU_PATCH_APP_ANALYSIS;
+    options.patch_file = "gpu_patch_app_analysis.fatbin";
+ }

---

Step 4: Add a GPU Patch

Write a patch file that will be injected during kernel execution.

gpu_patch_app_analysis.cu

// ...implementation of CommonCallback and memory access tracking functions

These patch functions log global, shared, and local memory accesses during kernel execution and update associated tracking data structures.

---

Step 5: Enable Patch Logic in the Backend

In your backend runtime patching and tracking logic, add conditions for GPU_PATCH_APP_ANALYSIS.

+ else if (sanitizer_options.patch_name == GPU_PATCH_APP_ANALYSIS) {
+     // Patch instructions for memory access
+ }

+ else if (sanitizer_options.patch_name == GPU_PATCH_APP_ANALYSIS) {
+     // Memory state allocation
+ }

+ else if (sanitizer_options.patch_name == GPU_PATCH_UVM_ADVISOR) {
+     // Initialization of access state and memory copy
+ }

+ else if (sanitizer_options.patch_name == GPU_PATCH_UVM_ADVISOR) {
+     // Final analysis and data transfer from device to host
+ }

---

Summary

By following these steps, you can extend AccelProf with a new tool for custom GPU analysis:

• Add enum entries and shell hooks

• Implement event-driven analysis logic

• Register and activate your tool

• Write a custom memory patch and wire it into the runtime backend

This modular design enables scalable and flexible extension of profiling capabilities.