profiling-api

Profiling API — Instrumenting Kit-Based Code

How to add profiling zones, metrics, and annotations to C++ and Python code in the Carbonite/Kit ecosystem. For capturing traces, see the profiling skill. For analyzing them, see nsys-analyze.

C++ Profiling Macros

Source: carb/profiler/Profile.h

Scope-Based Zone (most common)

#include <carb/profiler/Profile.h>
constexpr const uint64_t kProfilerMask = 1;

void myFunction() {
    CARB_PROFILE_ZONE(kProfilerMask, "My C++ function");
    doHeavyWork();  // zone closes automatically at scope exit (RAII)
}

Parameters: (maskOrChannel, zoneName, ...variadic_args)

• No variadic args → ProfileZoneStatic (pre-registered, faster)
• With variadic args → ProfileZoneDynamic (printf formatting)

Auto Function Name

void myFunction() {
    CARB_PROFILE_FUNCTION(kProfilerMask);
    // zone name = function's pretty-printed name
}

Manual Begin/End

auto zoneId = CARB_PROFILE_BEGIN(kProfilerMask, "Manual zone");
// ... work ...
CARB_PROFILE_END(kProfilerMask, zoneId);

Prefer RAII style (CARB_PROFILE_ZONE) over manual begin/end.

GPU Zones

Kit's RTX renderer uses query-based GPU zone capture:

auto gpuCtx = CARB_PROFILE_CREATE_GPU_CONTEXT("Vulkan GPU", cpuTs, gpuTs, gpuPeriod, "vulkan");
CARB_PROFILE_GPU_QUERY_BEGIN(kProfilerMask, gpuCtx, queryId, "RTX Render Pass");
// ... submit GPU commands ...
CARB_PROFILE_GPU_QUERY_END(kProfilerMask, gpuCtx, queryId);
CARB_PROFILE_GPU_SET_QUERY_VALUE(kProfilerMask, gpuCtx, queryId, gpuTimestamp);

Enable GPU zones in Tracy:

--/profiler/gpu/tracyInject/enabled=true
--/rtx/addTileGpuAnnotations=true

Python Profiling API

Decorator (simplest)

import carb.profiler

@carb.profiler.profile
def my_function():
    do_something()

Manual begin/end

carb.profiler.begin(1, "My Python operation")
# ... work ...
carb.profiler.end(1)

Full IProfiler Interface

profiler = carb.profiler.acquire_profiler_interface()

profiler.begin(mask, name)                      # zone start
profiler.end(mask)                              # zone end
profiler.set_capture_mask(mask) -> int          # returns previous mask
profiler.get_capture_mask() -> int
profiler.value_float(mask, value, name)         # Tracy Plot (float)
profiler.value_int(mask, value, name)           # Tracy Plot (int)
profiler.value_uint(mask, value, name)          # Tracy Plot (uint)
profiler.instant(mask, type, name)              # instant event
profiler.flow(mask, type, id, name)             # cross-thread flow
profiler.frame(mask, name)                      # frame marker
profiler.set_python_profiling_enabled(bool)     # toggle auto-profiling
profiler.is_python_profiling_enabled() -> bool

Types:

carb.profiler.InstantType.THREAD    # thread timeline
carb.profiler.InstantType.PROCESS   # process-wide timeline
carb.profiler.FlowType.BEGIN / END  # flow start/end

Profiler Mask

64-bit bitmask controlling which zones are captured: (zone_mask & capture_mask) != 0

constexpr uint64_t kCaptureMaskNone    = 0;              // nothing
constexpr uint64_t kCaptureMaskAll     = (uint64_t)-1;   // everything (default when no mask arg)
constexpr uint64_t kCaptureMaskDefault = uint64_t(1);    // bit 0
constexpr uint64_t kCaptureMaskProfiler = uint64_t(1) << 63; // profiler internals

If a zone uses mask 0, Carbonite treats it as kCaptureMaskDefault (1).

Workflow: Start with --/app/profilerMask=1 (major spans only, minimal overhead). If more detail needed, remove the arg (defaults to ALL). Always start coarse, then zoom in.

Profiler Channels

Higher-level abstraction over masks, toggled at runtime via settings:

Declaring a Channel (C++)

CARB_PROFILE_DECLARE_CHANNEL("myext.rendering", 1, true, g_myRenderingChannel);
CARB_PROFILE_ZONE(g_myRenderingChannel, "My rendering work");

Runtime Toggle

--/profiler/channels/<name>/enabled=true|false

Commonly disabled during benchmarks (too noisy):

--/profiler/channels/carb.events/enabled=false
--/profiler/channels/carb.tasking/enabled=false

Memory channels:

--/profiler/channels/cpu.memory/enabled=true
--/profiler/channels/cpu.virtualmemory/enabled=true
--/profiler/channels/graphics.memory/enabled=true

Tracy Plot Data (Numeric Metrics)

Record time-series values displayed as graphs in Tracy's Plot view.

C++

float gpuFrameTimeMs = 8.5f;
CARB_PROFILE_VALUE(gpuFrameTimeMs, 1, "GPU Frame Time (ms)");

int32_t triangleCount = 1500000;
CARB_PROFILE_VALUE(triangleCount, 1, "Triangle Count");

uint32_t gpuMemoryMB = 4096;
CARB_PROFILE_VALUE(gpuMemoryMB, 1, "GPU Memory (MB)");

int gpuIndex = 0;
CARB_PROFILE_VALUE(gpuFrameTimeMs, 1, "GPU %d Frame Time", gpuIndex);

Python

profiler.value_float(1, 8.5, "GPU Frame Time (ms)")
profiler.value_int(1, 1500000, "Triangle Count")
profiler.value_uint(1, 4096, "GPU Memory (MB)")

Event Annotations

Instant Events

// C++
CARB_PROFILE_EVENT(1, carb::profiler::InstantType::Thread, "Scene loading started");
CARB_PROFILE_EVENT(1, carb::profiler::InstantType::Process, "Phase transition: WARM -> BENCHMARK");

# Python
profiler.instant(1, carb.profiler.InstantType.THREAD, "Scene loading started")
profiler.instant(1, carb.profiler.InstantType.PROCESS, "Phase transition")

Display as Tracy messages (recommended):

--/plugins/carb.profiler-tracy.plugin/instantEventsAsMessages=true

command_macro.core Annotations

The omni.kit.command_macro.core extension auto-inserts [command_macro][Measurement] Start/End - <tag> events around benchmark measurements.

Automatic Python Function Capture

Capture all Python function calls without per-function instrumentation:

export CARB_PROFILING_PYTHON=1

Or programmatically:

profiler.set_python_profiling_enabled(True)

Performance warning: Significant overhead. Tracy file size ~4x larger (measured: 275MB → 1.2GB). Never use during benchmark measurement — only in the TRACY analysis phase.

Profiling Backend Summary

CPU backend can be toggled on/off at runtime for targeted capture:

profiler.set_capture_mask(1)   # start
# ... section to profile ...
profiler.set_capture_mask(0)   # stop