Below is largely based on following the RustCUDA quickstart run on Ubuntu laptop with an NVIDA P2000 Max-Q.
This isn't 100% Rust code. The kernel loaded on the GPU is actually C-like CUDA code in
resources/add.cu. The Rust code loads it as a kernel along with data for process and
then fetches the result. Rust does everything but made the C/CUDA kernel itself.
Example add vectors code is in resources/add.cu and it compiled
to resources/add.ptx for loading as a kernel.
cd resources
nvcc --ptx resources/add.cu
Rust code that'll load the kernel, send data, run the kernel then return results is in
src/main.rs. This uses the RustCUDA code that exposes NVIDA's CUDA APIs.
# build the code
cargo build
Compiling proc-macro2 v1.0.27
Compiling unicode-xid v0.2.2
Compiling syn v1.0.73
Compiling bitflags v1.2.1
Compiling cuda-sys v0.2.0
Compiling rustacuda_core v0.1.2
Compiling quote v1.0.9
Compiling rustacuda_derive v0.1.2
Compiling rustacuda v0.1.2
Compiling rust_add v0.1.0 (/home/jfalkner/nvidia-cuda-examples/rust_add)
Finished dev [unoptimized + debuginfo] target(s) in 7.59s
Runing is done same as a normal binary. The CUDA-specific calls allocate memory on the GPU, copy data and run the vector adding kernel.
# run the example on the GPU
cargo run
Finished dev [unoptimized + debuginfo] target(s) in 0.04s
Running `target/debug/rust_add`
Sum is 30
Here are the profiling results using nvprof on the results of cargo build.
sudo nvprof target/debug/rust_add
==1025456== NVPROF is profiling process 1025456, command: target/debug/rust_add
Sum is 30
==1025456== Profiling application: target/debug/rust_add
==1025456== Profiling result:
Type Time(%) Time Calls Avg Min Max Name
GPU activities: 42.52% 4.0000us 3 1.3330us 1.2480us 1.5040us [CUDA memcpy HtoD]
39.46% 3.7120us 1 3.7120us 3.7120us 3.7120us sum
18.03% 1.6960us 1 1.6960us 1.6960us 1.6960us [CUDA memcpy DtoH]
API calls: 62.84% 230.71ms 1 230.71ms 230.71ms 230.71ms cuCtxCreate
21.24% 77.995ms 1 77.995ms 77.995ms 77.995ms cuCtxDestroy
15.80% 58.002ms 1 58.002ms 58.002ms 58.002ms cuModuleLoadData
0.05% 184.43us 3 61.477us 2.6090us 178.42us cuMemAlloc
0.03% 111.57us 3 37.188us 2.1850us 103.86us cuMemFree
0.01% 30.154us 1 30.154us 30.154us 30.154us cuModuleUnload
0.01% 29.947us 1 29.947us 29.947us 29.947us cuDeviceGetPCIBusId
0.01% 24.795us 1 24.795us 24.795us 24.795us cuStreamCreateWithPriority
0.01% 20.762us 3 6.9200us 5.4590us 9.4750us cuMemcpyHtoD
0.00% 15.790us 1 15.790us 15.790us 15.790us cuLaunchKernel
0.00% 14.271us 1 14.271us 14.271us 14.271us cuMemcpyDtoH
0.00% 9.3690us 1 9.3690us 9.3690us 9.3690us cuStreamSynchronize
0.00% 5.0340us 1 5.0340us 5.0340us 5.0340us cuStreamDestroy
0.00% 4.1600us 2 2.0800us 509ns 3.6510us cuDeviceGetCount
0.00% 3.0670us 3 1.0220us 643ns 1.6180us cuDeviceGetAttribute
0.00% 2.3470us 2 1.1730us 490ns 1.8570us cuDeviceGet
0.00% 508ns 1 508ns 508ns 508ns cuModuleGetFunction