I have a matrix multiplication kernel that when timed in cuda c is 10x faster than when calling the same functions over rust ffi.
IK I can use cuBLAS but I was using this as an exercise to learn more advanced cuda optimizations.
When looking at nvidia's nsight systems, I see the kernel taking extremely long when using the rust version. Basically no overhead in both tests. This is confusing, as how can same kernel take longer in Rust?
It makes me think it has to be an issue with how I compiled the library for Rust. Especially since the timings in both tests are identical when using cuBLAS.
Here is the build.rs for the matrix library
use cc;
use std::{env, path::Path};
fn main() {
println!("cargo:rerun-if-changed=cuda_kernels/cuda_kernels.cu");
// let cublas_path =
// Path::new("C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.2/lib/x64/cublas.lib");
cc::Build::new()
.cuda(true)
.cudart("static")
// .object(cublas_path)
.file("cuda_kernels/cuda_kernels.cu")
.compile("cuda_kernels");
if let Ok(cuda_path) = env::var("CUDA_HOME") {
println!("cargo:rustc-link-search=native={}/lib64", cuda_path);
} else {
println!("cargo:rustc-link-search=native=/usr/local/cuda/lib64");
}
println!("cargo:rustc-link-lib=dylib=cudart");
println!("cargo:rustc-link-lib=dylib=cublas");
}
And here is the code for my tests.
CUDA Headers for my library
#include <cublas_v2.h>
#include <cuda.h>
// Make sure bindings are not mangled for rust
extern "C" {
// Misc
void cuda_synchronize();
// Matrix Setup API
size_t register_matrix(float* data, size_t rows, size_t cols);
void unregister_matrix(size_t mat_id);
void get_matrix_data(size_t mat_id, int rows, int cols, float* data_buffer);
// Matrix operation API
size_t cuda_matrix_multiply(size_t mat1_id, size_t mat1_rows, size_t mat1_cols, size_t mat2_id, size_t mat2_rows, size_t mat2_cols);
}
CUDA C Test
#include <chrono>
#include <vector>
using namespace std::chrono;
#include "../cuda_kernels.cuh"
int main() {
// This is just for timing, assumes everything is correct.
int dim = 4096;
std::vector<float> data;
for (int i = 0; i < dim * dim; i++) {
data.push_back(23.47);
}
// Register data as a 4096 x 4096 matrix
int mat1 = register_matrix(&data[0], dim, dim);
int mat2 = register_matrix(&data[0], dim, dim);
auto start_host = high_resolution_clock::now();
cudaEvent_t start;
cudaEvent_t end;
cudaEventCreate(&start);
cudaEventCreate(&end);
cudaEventRecord(start);
int num_iter = 100;
for (int i = 0; i < num_iter; i++) {
// Perform multiplication
int result_id = cuda_matrix_multiply(mat1, dim, dim, mat2, dim, dim);
cuda_synchronize();
unregister_matrix(result_id);
}
float gpu_time = 0;
cudaEventRecord(end);
cudaEventSynchronize(start);
cudaEventSynchronize(end);
cudaEventElapsedTime(&gpu_time, start, end);
auto end_host = high_resolution_clock::now();
auto cpu_time = duration_cast<milliseconds>(end_host - start_host);
std::cout << "Average gpu function time was: " << gpu_time / num_iter << " ms" << std::endl;
std::cout << "Including overhead was: " << (float)cpu_time.count() / num_iter << " ms" << std::endl;
// Okay something is wrong with the overhead on rust benchmark. Something taking 184.3 ms here is taking 1.3 seconds there.
}
Now on the Rust Side here is the bindings for the cuda functions
Bindings.rs
use std::ffi::c_float;
extern "C" {
pub fn cuda_synchronize();
pub fn register_matrix(data: *const c_float, rows: usize, cols: usize) -> usize;
pub fn unregister_matrix(mat_id: usize) -> usize;
pub fn cuda_matrix_multiply(
mat1_id: usize,
mat1_rows: usize,
mat1_cols: usize,
mat2_buffer: usize,
mat2_rows: usize,
mat2_cols: usize,
) -> usize;
}
And the Rust benchmark that is the same as CUDA C version
rust_bench_test.rs
// See why there is a 13x discrepancy between speed of rust ffi and c++ benchmarks
use std::time::Instant;
use matrix_lib::bindings::{
cuda_matrix_multiply, cuda_synchronize, register_matrix, unregister_matrix,
};
#[test]
fn mat_mult_benchmark() {
// Random numbers for generation
let mat_dim = 4096;
let id_1;
let id_2;
unsafe {
id_1 = register_matrix(vec![0.0; mat_dim * mat_dim].as_ptr(), mat_dim, mat_dim);
id_2 = register_matrix(vec![0.0; mat_dim * mat_dim].as_ptr(), mat_dim, mat_dim);
}
let num_iterations = 100;
let start = Instant::now();
let mut result_id = 0;
for _ in 0..num_iterations {
unsafe {
result_id = cuda_matrix_multiply(id_1, mat_dim, mat_dim, id_2, mat_dim, mat_dim);
cuda_synchronize();
unregister_matrix(result_id);
}
}
unsafe { cuda_synchronize() }
let elapsed = start.elapsed();
println!(
"\n=================================\nTime per iteration: {} ms\n=================================",
elapsed.as_millis() as f64 / num_iterations as f64
);
print!("{}", result_id);
assert!(false);
}
You wrote your benchmark as a #[test]. cargo test uses the test profile, which is by default the same as the dev profile, which has no optimization and enables debug assertions. This default behavior is because tests often need to be debugged (so optimizations are disabled), should check more edge cases (so debug assertions are enabled), and should give clear stack traces (so optimizations, particularly inlining, are disabled).
You can run cargo test --release, but it is better to work with the right default by making your benchmark be an actual benchmark target. cargo bench instead uses the bench profile, which is by default the same as the release profile, which has opt_level = 3 and no debug assertions.
The currently recommended tool to run benchmarks in Rust is criterion; if you use it as per its documentation then you will have a benchmark target and optimized builds of it by default. You can also skip criterion and continue using your own measurement code; the main thing is that you will need to move the file to benches/rust_bench_test.rs, declare the benchmark target like this
[[bench]]
name = "rust_bench_test"
harness = false
and write a fn main() instead of a #[test] function.