#nvbit #gpgpu #ffi #native-bindings #nvidia

sys nvbit-sys

Native bindings to the NVIDIA NVBIT binary instrumentation library

31 releases

0.0.38 Oct 10, 2023
0.0.35 Sep 8, 2023
0.0.32 Jul 29, 2023
0.0.19 Mar 22, 2023
0.0.6 Nov 28, 2022

#323 in Profiling

Download history 955/week @ 2023-12-04 623/week @ 2023-12-11 150/week @ 2023-12-18 335/week @ 2023-12-25 497/week @ 2024-01-01 1509/week @ 2024-01-15 1707/week @ 2024-01-22 4185/week @ 2024-01-29 2993/week @ 2024-02-05 1600/week @ 2024-02-12 530/week @ 2024-02-19 10/week @ 2024-02-26 2049/week @ 2024-03-04 13/week @ 2024-03-11 160/week @ 2024-03-18

2,232 downloads per month
Used in nvbit-rs

Custom license

1.5MB
3K SLoC

CUDA 1.5K SLoC // 0.4% comments Rust 1K SLoC // 0.0% comments

Contains (static library, 1.5MB) nvbit_release/core/libnvbit.a, (ELF exe/lib, 64KB) mem_trace.o, (ELF exe/lib, 24KB) inject_funcs.o

nvbit-rs

Test it out
make -j -B -C test-apps/
# build the mem trace example tracer
cargo build --release -p mem_trace

# trace a sample application
LD_PRELOAD=./target/release/libmem_trace.so ./test-apps/vectoradd/vectoradd 100

Done:

  • implement messagepack and json trace dumping
  • that does not work: clean up the nvbit api such that there the context is managed in nvbit-rs and the hooks are just functions
Accelsim reference
make -B -j -C ./tracer_nvbit
LD_PRELOAD=./tracer_nvbit/tracer_tool/tracer_tool.so ./nvbit-sys/nvbit_release/test-apps/vectoradd/vectoradd

This will generate files: ./tracer_nvbit/tracer_tool/traces/kernelslist and ./tracer_nvbit/tracer_tool/traces/stats.csv.

Our implementation
cargo build --release
make -B -j -C ./examples/accelsim
LD_PRELOAD=./target/release/libaccelsim.so ./test-apps/vectoradd/vectoradd 100
LD_PRELOAD=./target/release/libmem_trace.so ./test-apps/vectoradd/vectoradd 100
docker build --platform linux/amd64 -t builder .
docker run --rm -i -v "$PWD":/src -v "$PWD"/buildcache:/cache builder cargo build
nvcc -D_FORCE_INLINES -dc -c -std=c++11 -I../nvbit_release/core -Xptxas -cloning=no -Xcompiler -w  -O3 -Xcompiler -fPIC tracer_tool.cu -o tracer_tool.o

nvcc -D_FORCE_INLINES -I../nvbit_release/core -maxrregcount=24 -Xptxas -astoolspatch --keep-device-functions -c inject_funcs.cu -o inject_funcs.o

nvcc -D_FORCE_INLINES -O3 tracer_tool.o inject_funcs.o -L../nvbit_release/core -lnvbit -lcuda -shared -o tracer_tool.so

now is a good time to introduce workspaces make the examples individual crates with cargo.toml and build.rs write the custom tracing kernels per example this way we might finally include the symbol

  • a possible way is to provide a wrapper that copies to a cxx::Vec<CUfuncton> i guess (see this example)

  • since we are tracing, and this would need to be performed for each unseen function, this copy overhead is not acceptable

    • TODO: find out how often it is called and maybe still do it and measure
      • UPDATE: get_related_functions is only called once, try it in rust
  • other approach: only receive stuff from the channel, a simple struct...

    • if that works: how can we decide which tracing function to use
      • (since we cannot write new ones in rust)
  • figure out if we can somehow reuse the same nvbit names by using a namespace??

  • or wrap the calls in rust which calls the ffi::rust_* funcs.

  • IMPORTANT OBSERVATION:

    • i almost gave up on cxx, because it was only giving me Unsupported type errors
    • i dont import cxx::UniquePtr or cxx::CxxVector in the ffi module, so i was assuming i need to use cxx:: to reference the types.
    • they dont in the docs, but use them in the top level module ...
    • turns out you need to omit the cxx:: prefix because this is all macro magic ...

Done

  • we must include the CUDA inject funcs? and the nvbit_tool.h into the binary somehow.
    • maybe statically compile them in the build script
    • then link them with nvbit-sys crate
    • then, nvbit_at_init should be called - hopefully

Example

The current goal is to get a working example of a tracer written in rust. Usage should be:

# install lld
sudo apt-get install -y lld
# create a shared dynamic library that implements the nvbit hooks
cargo build -p accelsim
# run the nvbit example CUDA application with the tracer
LD_PRELOAD=./target/debug/libaccelsim.so nvbit-sys/nvbit_release/test-apps/vectoradd/vectoradd

Notes

check the clang versions installed

apt list --installed | grep clang

When running clang nvbit.h, it also complains about missing cassert. -std=c++11 -I$(NVBIT_PATH)

is C++ STL, so we need: clang++ -std=c++11 nvbit.h.

bindgen does not work that well with C++ code, check this.

we need some clang stuff so that bindgen can find #include <cassert>.

We will also need to include nvbit.h, nvbit_tool.h, and tracing injected functions, which require .cu files to be compiled and linked with the binary.

this example shows how .cu can be compiled and linked with the cc crate.

Make sure that the C function hooks of nvbit are not mangled in the shared library:

nm -D ./target/debug/examples/libtracer.so
nm -D ./target/debug/build/nvbit-sys-08fdef510bde07a0/out/libinstrumentation.a

Problem: we need the instrument_inst function to be present in the binary, just like for the example:

nm -D /home/roman/dev/nvbit-sys/tracer_nvbit/tracer_tool/tracer_tool.so | grep instrument

# for a static library:
nm --debug-syms target/debug/build/accelsim-a67c1762e4619dad/out/libinstrumentation.a | grep instrument

Currently, its not :(

Make sure that we link statically:

ldd ./target/debug/examples/libtracer.so

Check what includes cxx generated:

tre target/debug/build/nvbit-sys-*/out/cxxbridge/

Dependencies

~0.7–5MB
~93K SLoC