LLVM Vulkan
Version 1.1.0 — Drop-In Standard Passes & Clang Integration

GPU-Accelerated GVN &
Drop-in Clang Passes

A high-performance C++ implementation of LLVM middle-end passes (GVN, CSE, CFG, DCE) accelerated via Vulkan compute shaders. Features GPU dominator-tree GVN, commutative canonicalization, standard pass aliasing, and zero-configuration Clang -O2 integration.

📽️ View Slide Deck (12 Slides) 📊 Interactive Benchmark Explorer (109 Rows) 📦 Release v1.1.0 Packages
Designed for offline operation • Works standalone in your browser without external servers • Zero CUDA dependencies
📽️

Offline Presentation

A complete 12-slide presentation detailing architecture, execution pipeline, performance numbers, and the 1.0 contract.

Launch Presentation →
📊

Benchmark Dashboard

109 verifiable measurement rows across CPU and GPU hardware, with interactive filtering, sorting, and CSV export.

Explore Benchmarks →
📦

Relocatable Packages

Ready-to-use CMake package (`LLVMVulkan::compute`), public headers, static library, CLI tool, and embedded SPIR-V shaders.

Download v1.0.0 →

Measured Performance

Clean comparison of identical Release executable on AMD Ryzen 9 5950X vs NVIDIA GeForce RTX 5090 (1,048,576 elements)

Workload CPU Median GPU Median Speedup (CPU / GPU)
Integer fold (32-bit scalar arithmetic) 12.657 ms 4.166 ms 3.04×
Integer fold (64-bit with shaderInt64) 16.178 ms 7.103 ms 2.28×
Sort keys (radix sort vs std::sort) 55.964 ms 1.666 ms 33.59×
Hash records (ragged instruction tuples) 3.331 ms 2.646 ms 1.26×
Exclusive scan (prefix sums) 1.281 ms 1.469 ms 0.87× (CPU wins)
Transparent Measurement Disclosure: Ratios above 1.0 favor the GPU. The baseline comparator is a single-threaded C++ scalar implementation. GPU times include upload, dispatch, synchronization, and host readback. The CPU wins every tested graph case and all small 1,024-element workloads due to worklist efficiency over GPU convergence rounds.

Scoped Architecture

Maintaining host LLVM semantic authority with accelerated offload

1. Scoped LLVM Passes

Integrates with LLVM passes: lv-fold for batched scalar operations, lv-cse for instruction record hashing, and lv-cfg for reachability analysis.

2. Standalone C++17 API

Standalone library offering integer folding, CSR graph reachability, dominators, liveness analysis, unsigned scan, key sorting, and record hashing.

3. Portable Vulkan Engine

Fixed 128-thread workgroups, 18 validated embedded SPIR-V shaders, optional subgroup arithmetic, and enforced Vulkan synchronization validation.

Project Documentation

Comprehensive guides, architectural invariants, and validation evidence

Release Guide & Contract Compatibility contract, build instructions, and platform matrix CPU vs GPU Benchmark All 37 workload rows and measurement boundaries on RTX 5090 LLVM Integration Guide Host authority, pass integration, and plugin linkage requirements Coverage & Gap Inventory Documented scope vs upstream compiler features Testing & Reproduction Instructions for running CTest, lavapipe, and hardware suites Validation Records Differential logs, oracle runs, and fault injection recovery