07d0e60fe4
- Fork cudarc locally (vendor/cudarc): add CudaContext::load_cubin() that calls cuModuleLoadData directly — zero nvrtc dependency - Remove "nvrtc" feature from ml-core, ml-dqn, ml-ppo Cargo.toml - Replace all 89 Ptx::from_binary + load_module calls with load_cubin - ml-core cuda_autograd: wire 9 stub constructors to precompiled cubins (activation, elementwise, linear, loss, reduction, dropout, layer_norm, optimizer) - ml-core build.rs: compile 8 BF16-native CUDA kernels via nvcc - cubin_loader.rs: thin wrapper around CudaContext::load_cubin() - Fix size_of::<f32> in gpu_tensor.rs, stream_ops.rs, layer_norm.rs - Fix test data: Vec<f32> → Vec<half::bf16> for memcpy_htod - Stub ml-ppo/ml-dqn runtime compile_ptx calls (dead code) - backtest_metrics_kernel.cu: full native BF16 rewrite (no float) - backtest_env_kernel.cu: shared memory → __nv_bfloat16 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
63 lines
2.0 KiB
Rust
63 lines
2.0 KiB
Rust
use cudarc::driver::*;
|
|
use cudarc::nvrtc::compile_ptx;
|
|
|
|
use std::thread;
|
|
|
|
const KERNEL_SRC: &str = "
|
|
extern \"C\" __global__ void hello_world(int i) {
|
|
printf(\"Hello from the cuda kernel in thread %d\\n\", i);
|
|
}
|
|
";
|
|
|
|
fn main() -> Result<(), DriverError> {
|
|
{
|
|
// Option 1: sharing ctx & module between threads
|
|
thread::scope(|s| {
|
|
let ptx = compile_ptx(KERNEL_SRC).unwrap();
|
|
let ctx = CudaContext::new(0)?;
|
|
let module = ctx.load_module(ptx)?;
|
|
for i in 0..10i32 {
|
|
let thread_ctx = ctx.clone();
|
|
let thread_module = module.clone();
|
|
s.spawn(move || {
|
|
let stream = thread_ctx.default_stream();
|
|
let f = thread_module.load_function("hello_world")?;
|
|
unsafe {
|
|
stream
|
|
.launch_builder(&f)
|
|
.arg(&i)
|
|
.launch(LaunchConfig::for_num_elems(1))
|
|
}
|
|
});
|
|
}
|
|
Ok(())
|
|
})?;
|
|
}
|
|
|
|
{
|
|
// Option 2: initializing different context in each
|
|
// Note that this will still schedule to the same stream since we are using the
|
|
// default stream here on the same device.
|
|
thread::scope(move |s| {
|
|
for i in 0..10i32 {
|
|
s.spawn(move || {
|
|
let ptx = compile_ptx(KERNEL_SRC).unwrap();
|
|
let ctx = CudaContext::new(0)?;
|
|
let module = ctx.load_module(ptx)?;
|
|
let stream = ctx.default_stream();
|
|
let f = module.load_function("hello_world")?;
|
|
unsafe {
|
|
stream
|
|
.launch_builder(&f)
|
|
.arg(&i)
|
|
.launch(LaunchConfig::for_num_elems(1))
|
|
}
|
|
});
|
|
}
|
|
Ok(())
|
|
})?;
|
|
}
|
|
|
|
Ok(())
|
|
}
|