Release v0.3.2: GPU profiling infrastructure

Add comprehensive GPU profiling capabilities for CUDA backend:

- CUDA event wrappers (CudaEvent, GpuTimer, GpuTimerPool)
- NVTX integration for Nsight Systems/Compute timeline visualization
- Kernel metrics collection (grid/block dims, occupancy, registers)
- Memory tracking with allocation profiling and leak detection
- Chrome trace format export for GPU timeline visualization
- Feature-gated via 'profiling' feature flag

New files:
- crates/ringkernel-cuda/src/profiling/mod.rs
- crates/ringkernel-cuda/src/profiling/events.rs
- crates/ringkernel-cuda/src/profiling/nvtx.rs
- crates/ringkernel-cuda/src/profiling/metrics.rs
- crates/ringkernel-cuda/src/profiling/memory_tracker.rs
- crates/ringkernel-cuda/src/profiling/chrome_trace.rs

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: mivertowski

CHANGELOG.md

@@ -7,6 +7,61 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.3.2] - 2026-01-20
+
+### Added
+
+#### GPU Profiling Infrastructure
+
+- **CUDA Profiling Module** (`ringkernel-cuda/src/profiling/`) - **NEW MODULE**
+  - Feature-gated via `profiling` feature flag
+  - Comprehensive GPU profiling capabilities for performance analysis
+
+- **CUDA Event Wrappers** (`profiling/events.rs`)
+  - `CudaEvent` - RAII wrapper for CUDA events with timing support
+  - `CudaEventFlags` - Event configuration (blocking sync, disable timing, interprocess)
+  - `GpuTimer` - Start/stop timer using CUDA events with microsecond precision
+  - `GpuTimerPool` - Pool of reusable timers with interior mutability for concurrent access
+
+- **NVTX Integration** (`profiling/nvtx.rs`)
+  - `CudaNvtxProfiler` - Real NVTX profiler using cudarc's nvtx module
+  - Timeline visualization in Nsight Systems and Nsight Compute
+  - `NvtxCategory` - Predefined categories (Kernel, Transfer, Memory, Sync, Queue, User)
+  - `NvtxRange` - RAII wrapper for automatic range end on drop
+  - `NvtxPayload` - Typed payloads for markers (I32, I64, U32, U64, F32, F64)
+  - Implements `GpuProfiler` trait for integration with ringkernel-core
+
+- **Kernel Metrics** (`profiling/metrics.rs`)
+  - `KernelMetrics` - Execution metadata (grid/block dims, GPU time, occupancy, registers)
+  - `TransferMetrics` - Memory transfer stats with bandwidth calculation
+  - `TransferDirection` - HostToDevice, DeviceToHost, DeviceToDevice
+  - `ProfilingSession` - Collects kernel and transfer events with timestamps
+  - `KernelAttributes` - Query kernel attributes via cuFuncGetAttribute
+
+- **Memory Tracking** (`profiling/memory_tracker.rs`)
+  - `CudaMemoryTracker` - Track GPU memory allocations with timing
+  - `TrackedAllocation` - Allocation metadata (ptr, size, kind, label, timestamp)
+  - `CudaMemoryKind` - Device, Pinned, Mapped, Managed memory types
+  - Peak usage tracking and allocation statistics
+  - Integration with `GpuMemoryDashboard` from ringkernel-core
+
+- **Chrome Trace Export** (`profiling/chrome_trace.rs`)
+  - `GpuTraceEvent` - Chrome trace format event structure
+  - `GpuEventArgs` - Rich event metadata (grid/block dims, occupancy, bandwidth)
+  - `GpuChromeTraceBuilder` - Build Chrome trace JSON from profiling sessions
+  - Support for kernel events, transfer events, NVTX ranges, memory allocations
+  - Process/thread naming for multi-GPU and multi-stream visualization
+  - Compatible with chrome://tracing, Perfetto UI, and Nsight Systems
+
+### Changed
+
+- **Dependencies** - Added `nvtx` feature to cudarc dependency
+- **ringkernel-cuda/Cargo.toml** - Added optional `serde` and `serde_json` for Chrome trace export
+
+### Fixed
+
+- Added `ProfilerRange::stub()` public constructor in ringkernel-core for external profiler implementations
+
 ## [0.3.1] - 2026-01-19
 
 ### Added
@@ -771,7 +826,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - CLAUDE.md with build commands and architecture overview
 - Code examples for all major features
 
-[Unreleased]: https://github.com/mivertowski/RustCompute/compare/v0.3.1...HEAD
+[Unreleased]: https://github.com/mivertowski/RustCompute/compare/v0.3.2...HEAD
+[0.3.2]: https://github.com/mivertowski/RustCompute/compare/v0.3.1...v0.3.2
 [0.3.1]: https://github.com/mivertowski/RustCompute/compare/v0.3.0...v0.3.1
 [0.3.0]: https://github.com/mivertowski/RustCompute/compare/v0.2.0...v0.3.0
 [0.2.0]: https://github.com/mivertowski/RustCompute/compare/v0.1.3...v0.2.0

CLAUDE.md

@@ -94,7 +94,7 @@ The project is a Cargo workspace with these crates:
 - **`ringkernel-core`** - Core traits and types (RingMessage, MessageQueue, HlcTimestamp, ControlBlock, RingContext, RingKernelRuntime, K2K messaging, PubSub)
 - **`ringkernel-derive`** - Proc macros: `#[derive(RingMessage)]`, `#[ring_kernel]`, `#[derive(GpuType)]`
 - **`ringkernel-cpu`** - CPU backend implementation (always available, used for testing/fallback)
-- **`ringkernel-cuda`** - NVIDIA CUDA backend with cooperative groups support (feature-gated)
+- **`ringkernel-cuda`** - NVIDIA CUDA backend with cooperative groups and GPU profiling support (feature-gated)
 - **`ringkernel-wgpu`** - WebGPU cross-platform backend (feature-gated)
 - **`ringkernel-metal`** - Apple Metal backend (feature-gated, macOS only, scaffold implementation with runtime/buffer/pipeline)
 - **`ringkernel-codegen`** - GPU kernel code generation
@@ -539,8 +539,9 @@ Main crate (`ringkernel`) features:
 - `metal` - Apple Metal backend
 - `all-backends` - All GPU backends
 
-CUDA-specific features:
+CUDA-specific features (`ringkernel-cuda`):
 - `cooperative` - Enable CUDA cooperative groups for grid-wide synchronization (`grid.sync()`). Requires nvcc at build time for PTX compilation.
+- `profiling` - GPU profiling infrastructure (CUDA events, NVTX, memory tracking, Chrome trace export). Requires nvToolsExt library.
 
 Core crate (`ringkernel-core`) enterprise features:
 - `crypto` - Real cryptography (AES-256-GCM, ChaCha20-Poly1305, Argon2)
@@ -805,7 +806,7 @@ let _ = device.poll(wgpu::PollType::wait_indefinitely());
 
 ## Dependency Versions
 
-Key workspace dependencies (as of v0.3.1):
+Key workspace dependencies (as of v0.3.2):
 
 | Category | Package | Version | Notes |
 |----------|---------|---------|-------|

Cargo.lock

@@ -26,7 +26,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.3.1"
+version = "0.3.2"
 edition = "2021"
 authors = ["Michael Ivertowski <mivertowski@outlook.com>"]
 license = "Apache-2.0"

Cargo.toml

@@ -643,13 +643,55 @@ cargo run -p ringkernel-txmon --bin txmon-benchmark --release --features cuda-co
 
 Performance varies significantly by hardware and workload.
 
+## GPU Profiling
+
+RingKernel v0.3.2 includes comprehensive GPU profiling infrastructure. Enable with the `profiling` feature:
+
+```toml
+[dependencies]
+ringkernel-cuda = { version = "0.3", features = ["profiling"] }
+```
+
+### GPU Timer and Events
+
+```rust
+use ringkernel_cuda::profiling::{GpuTimer, CudaNvtxProfiler, ProfilingSession};
+
+// GPU-side timing with CUDA events
+let mut timer = GpuTimer::new()?;
+timer.start(stream)?;
+// ... kernel execution ...
+timer.stop(stream)?;
+println!("Kernel time: {:.3} ms", timer.elapsed_ms()?);
+
+// NVTX profiling for Nsight Systems/Compute
+let profiler = CudaNvtxProfiler::new();
+{
+    let _range = profiler.push_range("compute_phase", ProfilerColor::CYAN);
+    // ... kernel execution ...
+} // Range automatically ends
+
+// Export to Chrome trace format
+let session = ProfilingSession::new();
+// ... record kernel/transfer events ...
+let builder = GpuChromeTraceBuilder::from_session(&session);
+std::fs::write("gpu_trace.json", builder.build())?;
+```
+
+Features include:
+- **CUDA Events**: GPU-side timing without CPU overhead
+- **NVTX Integration**: Timeline visualization in Nsight Systems/Compute
+- **Kernel Metrics**: Grid/block dims, occupancy, registers per thread
+- **Memory Tracking**: Allocation profiling with leak detection
+- **Chrome Trace Export**: GPU timeline visualization in chrome://tracing
+
 ## Enterprise Security
 
-RingKernel v0.3.1 includes comprehensive enterprise security features. Enable with the `enterprise` feature:
+RingKernel v0.3.2 includes comprehensive enterprise security features. Enable with the `enterprise` feature:
 
 ```toml
 [dependencies]
-ringkernel-core = { version = "0.3.1", features = ["enterprise"] }
+ringkernel-core = { version = "0.3", features = ["enterprise"] }
 ```
 
 ### Authentication & Authorization

README.md

@@ -1162,6 +1162,15 @@ impl ProfilerRange {
         }
     }
 
+    /// Create a stub profiler range for external profiler implementations.
+    ///
+    /// This is used by custom profiler implementations (like CUDA NVTX) that
+    /// manage their own range lifecycle but need to return a ProfilerRange
+    /// for API compatibility.
+    pub fn stub(name: impl Into<String>, backend: GpuProfilerBackend) -> Self {
+        Self::new(name, backend)
+    }
+
     /// Get elapsed duration.
     pub fn elapsed(&self) -> Duration {
         self.start.elapsed()

crates/ringkernel-core/src/observability.rs

@@ -17,7 +17,8 @@ build = "build.rs"
 ringkernel-core = { workspace = true }
 
 # CUDA bindings - auto-detect CUDA version from build system
-cudarc = { version = "0.18.2", optional = true, features = ["cuda-version-from-build-system"] }
+# The nvtx feature enables NVTX profiling integration
+cudarc = { version = "0.18.2", optional = true, features = ["cuda-version-from-build-system", "nvtx"] }
 
 # Async runtime
 tokio = { workspace = true }
@@ -33,6 +34,10 @@ tracing = { workspace = true }
 # Synchronization
 parking_lot = { workspace = true }
 
+# Serialization (for profiling chrome trace export)
+serde = { workspace = true, optional = true }
+serde_json = { workspace = true, optional = true }
+
 [dev-dependencies]
 tokio = { workspace = true, features = ["test-util", "macros", "rt-multi-thread"] }
 
@@ -45,3 +50,5 @@ cuda = ["cudarc"]
 # Cooperative groups support - requires nvcc at build time for PTX compilation
 # Enables grid-wide synchronization via cuLaunchCooperativeKernel
 cooperative = ["cuda"]
+# GPU profiling support - NVTX integration, CUDA events, Chrome trace export
+profiling = ["cuda", "serde", "serde_json"]