AMD Gpu traning not working

[rickytrevor]

Hello, I'm trying to run femtoGPT on my rx 6600 under Ubuntu Linux, I've installed the required Rocm OpenCL drivers but, when I run the program using

cargo run --release --features gpu

I get an index out of bounds exception

thread 'main' panicked at 'index out of bounds: the len is 0 but the index is 0', src/graph/gpu/mod.rs:119:22 note: run with RUST_BACKTRACE=1 environment variable to display a backtrace

Edit

This is the log with RUST_BACKTRACE=1

0: rust_begin_unwind at /rustc/5680fa18feaa87f3ff04063800aec256c3d4b4be/library/std/src/panicking.rs:593:5 1: core::panicking::panic_fmt at /rustc/5680fa18feaa87f3ff04063800aec256c3d4b4be/library/core/src/panicking.rs:67:14 2: core::panicking::panic_bounds_check at /rustc/5680fa18feaa87f3ff04063800aec256c3d4b4be/library/core/src/panicking.rs:162:5 3: femto_gpt::graph::gpu::GpuGraph::new 4: femto_gpt::main note: Some details are omitted, run with RUST_BACKTRACE=full for a verbose backtrace.

[keyvank]

@rickytrevor Ah yes you're right. I haven't tried this on AMD. Try changing the brand on line 119 (graph/gpu/mod.rs) to Brand::Amd and recompile

[rickytrevor]

I've tried changing it to Brand::Amd and it just panics to "Not supported", I've briefly looked at the Impl for Brand and the implementation for amd's missing in the get_bus_id function

   pub fn get_bus_id(&self, d: ocl::Device) -> ocl::Result<u32> {
        match self {
            Brand::Nvidia => {
                const CL_DEVICE_PCI_BUS_ID_NV: u32 = 0x4008;
                let result = d.info_raw(CL_DEVICE_PCI_BUS_ID_NV)?;
                Ok(u32::from_le_bytes(result[..].try_into().unwrap()))
            }
            Brand::Amd => panic!("Not supported!"),
        }
    }
}

[rickytrevor]

Update, I've managed to implement the case for Amd

    pub fn get_bus_id(&self, d: ocl::Device) -> ocl::Result<u32> {
        match self {
            Brand::Nvidia => {
                const CL_DEVICE_PCI_BUS_ID_NV: u32 = 0x4008;
                let result = d.info_raw(CL_DEVICE_PCI_BUS_ID_NV)?;
                Ok(u32::from_le_bytes(result[..].try_into().unwrap()))
            }
            Brand::Amd => {
                const CL_DEVICE_PCIE_ID_AMD: u32 = 0x4034;
                let result = d.info_raw(CL_DEVICE_PCIE_ID_AMD)?;
                Ok(u32::from_le_bytes(result[..].try_into().unwrap()))
            }
        }
    }
}

The problem is that right after starting it crashes because of this error, any clues?

Error: GpuError(OclError(Kernel(ArgTypeMismatch { idx: 0, arg_name: "out", ty_name: "float*", ty: ArgType { base_type: Float, cardinality: One, is_ptr: true } })))

[keyvank]

@rickytrevor I don't have a AMD card so can't debug :smile: Happy to get a PR for AMD support though :wink: