PTX Lint
The ptxlint is a static analyser for NVIDIA PTX. It reads the .ptx your GPU kernels already compile to and reports local memory traffic, FP64 use, register pressure, shared memory budget and estimated occupancy. It needs no GPU, no CUDA install and no dependencies, so a kernel performance regression can fail CI on a machine that has never seen a graphics card.
针对 NVIDIA PTX 的静态分析工具 —— 直接读取 GPU kernel 编译出的 .ptx，报告 local memory 流量、FP64 使用、寄存器压力、shared memory 预算与估算占用率；不需要显卡、不需要安装 CUDA、零依赖，因此 kernel 的性能回归可以在一台从没见过显卡的机器上卡住 CI。
$ ptxlint kernels.ptx
local_array line 5
arch sm_70 regs/thread 176 (virtual, upper bound) shared 0 B local 256 B
occupancy 12% (8 of 64 warps/SM, 1 blocks/SM @ 256 threads/block; limited by registers)
198 instructions · fp32 10 · int 23 · global 65 · local 72
error 256 bytes of local memory, 72 local accesses — this lives in DRAM, not registers [PTX001:42]
An array indexed by a runtime value cannot stay in registers.
Usage
$ cargo install ptxlint
$ ptxlint kernels.ptx # a file, a directory, or - for stdin
$ ptxlint --deny error kernels.ptx # exit 1 on errors, as a CI gate
$ ptxlint --json kernels.ptx # machine-readable
$ ptxlint --ptxas kernels.ptx # exact register counts, if CUDA is installed
Lints
PTX001 |
local memory in use — an array indexed by a runtime value, backed by DRAM | 用到 local memory，运行时下标的数组实际落在显存里 |
PTX002 |
register spills, needs --ptxas |
寄存器溢出，需要 --ptxas |
PTX003 |
FP64 instructions — in Rust a bare 0.5 is f64 |
FP64 指令 —— Rust 里裸写的 0.5 是 f64 |
PTX004 |
integer div/rem — GPUs have no integer divider |
整数除法取模 —— GPU 没有整数除法器 |
PTX005 |
high register pressure | 寄存器压力过高 |
PTX006 |
low estimated occupancy | 估算占用率过低 |
PTX007 |
shared memory over budget, or capping residency | shared memory 超限或压制驻留块数 |
PTX008 |
narrow, non-vectorised global accesses | 窄的、未向量化的 global 访问 |
PTX009 |
no .maxntid/.reqntid launch bounds |
没有 launch bounds |
PTX010 |
calls that were not inlined | 没有被内联的调用 |
Accuracy
PTX only carries virtual registers, which ptxas coalesces on the way to SASS. Local and shared memory, instruction mix, FP64 use and vector widths are therefore exact, while registers and occupancy are an upper bound, and are labelled as such. The scanner is deliberately not a full PTX grammar: an unknown opcode is recorded and matches no lint, rather than failing CI over an instruction NVIDIA shipped last month.
PTX 中只有虚拟寄存器，真实分配要等 ptxas 降级到 SASS 才确定。所以 local/shared memory、指令构成、FP64 使用与向量宽度是精确的，而寄存器数和占用率只是上界，报告中会明确标注。扫描器故意不是完整的 PTX 文法：不认识的指令原样记录、不匹配任何 lint，而不会因为 NVIDIA 上个月新加了一条指令就让 CI 崩掉。
It is not a profiler — for real tuning use Nsight Compute, for races use compute-sanitizer. This is the smoke alarm, not the fire brigade.
它不是 profiler —— 真要调优请用 Nsight Compute，查竞态请用 compute-sanitizer。这是烟雾报警器，不是消防队。
Library Development
Lint Development
$ cargo test
$ cargo clippy --all-targets -- -D warnings
Fixture Kernel Development
Each case in fixtures/examples/ is a real Rust kernel that compiles to its own .ptx, so a fixture only ever triggers the lint it demonstrates. Regenerating them needs a nightly toolchain.
fixtures/examples/ 下每个用例都是一个真实的 Rust kernel，各自编译出独立的 .ptx，因此一份素材只会触发它要演示的那条 lint。重新生成需要 nightly 工具链。
$ ./fixtures/generate.sh