Crate iro_cuda_ffi_kernels 

Reference CUDA kernels for iro-cuda-ffi.

This crate provides sample kernels that demonstrate proper iro-cuda-ffi usage patterns and serve as integration tests for the iro-cuda-ffi core crate.

Available Kernels

• vector_add_f32: Element-wise vector addition
• fma_chain_f32: Deep compute chain (FMA)
• saxpy_f32: Single-precision A*X + Y
• daxpy_f64: Double-precision A*X + Y
• scale_f32: Vector scaling
• reduce_sum_f32: Parallel sum reduction
• reduce_max_f32: Parallel max reduction

Example

use iro_cuda_ffi::prelude::*;
use iro_cuda_ffi_kernels::vector_add_f32;

let stream = Stream::new()?;

let a = DeviceBuffer::from_slice_sync(&stream, &[1.0f32, 2.0, 3.0, 4.0])?;
let b = DeviceBuffer::from_slice_sync(&stream, &[5.0f32, 6.0, 7.0, 8.0])?;
let mut c = DeviceBuffer::<f32>::zeros(4)?;

vector_add_f32(&stream, &a, &b, &mut c)?;

let result = c.to_vec(&stream)?;
assert_eq!(result, vec![6.0, 8.0, 10.0, 12.0]);

Device-to-device copy

use iro_cuda_ffi::prelude::*;

let stream = Stream::new()?;
let src = DeviceBuffer::from_slice_sync(&stream, &[1.0f32, 2.0, 3.0])?;
let mut dst = DeviceBuffer::<f32>::alloc(src.len())?;

dst.copy_from_device_sync(&stream, &src)?;
let result = dst.to_vec(&stream)?;
assert_eq!(result, vec![1.0, 2.0, 3.0]);

Constants

MAX_GRID_X

Maximum grid_x dimension per CUDA spec (2^31 - 1).

Functions

daxpy_f64

DAXPY operation: y = a * x + y (in-place, double precision)

fma_chain_f32

Deep compute chain: out = fma_chain(a, b, iters)

reduce_max_f32

Parallel max reduction (first pass).

reduce_max_full

Computes the maximum of all elements in the input vector.

reduce_sum_f32

Parallel sum reduction (first pass).

reduce_sum_full

Computes the sum of all elements in the input vector.

reduction_output_size

Returns the number of output elements needed for reduction.

saxpy_f32

SAXPY operation: y = a * x + y (in-place)

scale_f32

Scale vector: out = a * x

vector_add_f32

Element-wise vector addition: out = a + b

verify_abi_linked

Ensures the ABI asserts translation unit is linked.