ringkernel-cuda-codegen

Rust-to-CUDA transpiler for RingKernel GPU kernels.

Overview

This crate enables writing GPU kernels in a restricted Rust DSL and transpiling them to CUDA C code. It supports three kernel types:

1. Global Kernels - Standard CUDA __global__ functions
2. Stencil Kernels - Tile-based kernels with GridPos abstraction (2D and 3D)
3. Ring Kernels - Persistent actor kernels with message loops

Installation

[dependencies]
ringkernel-cuda-codegen = "0.1"
syn = { version = "2.0", features = ["full"] }

Global Kernels

For general-purpose CUDA kernels:

use ringkernel_cuda_codegen::transpile_global_kernel;
use syn::parse_quote;

let func: syn::ItemFn = parse_quote! {
    fn saxpy(x: &[f32], y: &mut [f32], a: f32, n: i32) {
        let idx = block_idx_x() * block_dim_x() + thread_idx_x();
        if idx >= n { return; }
        y[idx as usize] = a * x[idx as usize] + y[idx as usize];
    }
};

let cuda_code = transpile_global_kernel(&func)?;

Stencil Kernels

For grid-based computations with neighbor access (2D and 3D):

use ringkernel_cuda_codegen::{transpile_stencil_kernel, StencilConfig, Grid};

// 2D stencil
let func: syn::ItemFn = parse_quote! {
    fn fdtd(p: &[f32], p_prev: &mut [f32], c2: f32, pos: GridPos) {
        let lap = pos.north(p) + pos.south(p) + pos.east(p) + pos.west(p)
                  - 4.0 * p[pos.idx()];
        p_prev[pos.idx()] = 2.0 * p[pos.idx()] - p_prev[pos.idx()] + c2 * lap;
    }
};

let config = StencilConfig::new("fdtd")
    .with_grid(Grid::Grid2D)
    .with_tile_size(16, 16)
    .with_halo(1);

let cuda_code = transpile_stencil_kernel(&func, &config)?;

// 3D stencil with up/down neighbors
let func_3d: syn::ItemFn = parse_quote! {
    fn laplacian_3d(p: &[f32], out: &mut [f32], pos: GridPos) {
        let lap = pos.north(p) + pos.south(p) + pos.east(p) + pos.west(p)
                  + pos.up(p) + pos.down(p) - 6.0 * p[pos.idx()];
        out[pos.idx()] = lap;
    }
};

let config_3d = StencilConfig::new("laplacian")
    .with_grid(Grid::Grid3D)
    .with_tile_size(8, 8)
    .with_halo(1);

Ring Kernels

For persistent actor-model kernels:

use ringkernel_cuda_codegen::{transpile_ring_kernel, RingKernelConfig};

let handler: syn::ItemFn = parse_quote! {
    fn process(ctx: &RingContext, msg: &Request) -> Response {
        let tid = ctx.global_thread_id();
        ctx.sync_threads();
        Response { value: msg.value * 2.0, id: tid as u64 }
    }
};

let config = RingKernelConfig::new("processor")
    .with_block_size(128)
    .with_queue_capacity(1024)
    .with_hlc(true)   // Hybrid Logical Clocks
    .with_k2k(true);  // Kernel-to-kernel messaging

let cuda_code = transpile_ring_kernel(&handler, &config)?;

DSL Reference

Thread/Block Indices

• thread_idx_x(), thread_idx_y(), thread_idx_z() → threadIdx.x/y/z
• block_idx_x(), block_idx_y(), block_idx_z() → blockIdx.x/y/z
• block_dim_x(), block_dim_y(), block_dim_z() → blockDim.x/y/z
• grid_dim_x(), grid_dim_y(), grid_dim_z() → gridDim.x/y/z
• warp_size() → warpSize

Stencil Intrinsics (2D)

• pos.idx() - Linear index
• pos.north(buf), pos.south(buf) - Y-axis neighbors
• pos.east(buf), pos.west(buf) - X-axis neighbors
• pos.at(buf, dx, dy) - Relative offset access

Stencil Intrinsics (3D)

• pos.up(buf), pos.down(buf) - Z-axis neighbors
• pos.at(buf, dx, dy, dz) - 3D relative offset access

Synchronization

• sync_threads() → __syncthreads() - Block-level barrier
• sync_threads_count(pred) → __syncthreads_count() - Count threads with predicate
• sync_threads_and(pred) → __syncthreads_and() - AND of predicate
• sync_threads_or(pred) → __syncthreads_or() - OR of predicate
• thread_fence() → __threadfence() - Device memory fence
• thread_fence_block() → __threadfence_block() - Block memory fence
• thread_fence_system() → __threadfence_system() - System memory fence

Atomic Operations (Integer)

• atomic_add(ptr, val) → atomicAdd
• atomic_sub(ptr, val) → atomicSub
• atomic_min(ptr, val) → atomicMin
• atomic_max(ptr, val) → atomicMax
• atomic_exchange(ptr, val) → atomicExch
• atomic_cas(ptr, compare, val) → atomicCAS
• atomic_and(ptr, val) → atomicAnd
• atomic_or(ptr, val) → atomicOr
• atomic_xor(ptr, val) → atomicXor
• atomic_inc(ptr, val) → atomicInc (increment with wrap)
• atomic_dec(ptr, val) → atomicDec (decrement with wrap)

Basic Math Functions

• sqrt(), rsqrt() - Square root, reciprocal sqrt
• abs(), fabs() - Absolute value
• floor(), ceil(), round(), trunc() - Rounding
• fma(), mul_add() - Fused multiply-add
• fmin(), fmax() - Minimum, maximum
• fmod(), remainder() - Modulo operations
• copysign() - Copy sign
• cbrt() - Cube root
• hypot() - Hypotenuse

Trigonometric Functions

• sin(), cos(), tan() - Basic trig
• asin(), acos(), atan(), atan2() - Inverse trig
• sincos() - Combined sine and cosine
• sinpi(), cospi() - Sin/cos of π*x

Hyperbolic Functions

• sinh(), cosh(), tanh() - Hyperbolic
• asinh(), acosh(), atanh() - Inverse hyperbolic

Exponential and Logarithmic Functions

• exp(), exp2(), exp10(), expm1() - Exponentials
• log(), ln(), log2(), log10(), log1p() - Logarithms
• pow(), powf(), powi() - Power
• ldexp(), scalbn() - Load/scale exponent
• ilogb() - Extract exponent
• erf(), erfc(), erfinv(), erfcinv() - Error functions
• lgamma(), tgamma() - Gamma functions

Classification Functions

• is_nan(), isnan() → isnan
• is_infinite(), isinf() → isinf
• is_finite(), isfinite() → isfinite
• is_normal(), isnormal() → isnormal
• signbit() - Check sign bit
• nextafter() - Next representable value
• fdim() - Positive difference

Warp Operations

• warp_active_mask() → __activemask() - Active lane mask
• warp_shfl(mask, val, lane) → __shfl_sync - Shuffle
• warp_shfl_up(mask, val, delta) → __shfl_up_sync
• warp_shfl_down(mask, val, delta) → __shfl_down_sync
• warp_shfl_xor(mask, val, lane_mask) → __shfl_xor_sync
• warp_ballot(mask, pred) → __ballot_sync
• warp_all(mask, pred) → __all_sync
• warp_any(mask, pred) → __any_sync

Warp Match Operations (Volta+)

• warp_match_any(mask, val) → __match_any_sync
• warp_match_all(mask, val) → __match_all_sync

Warp Reduce Operations (SM 8.0+)

• warp_reduce_add(mask, val) → __reduce_add_sync
• warp_reduce_min(mask, val) → __reduce_min_sync
• warp_reduce_max(mask, val) → __reduce_max_sync
• warp_reduce_and(mask, val) → __reduce_and_sync
• warp_reduce_or(mask, val) → __reduce_or_sync
• warp_reduce_xor(mask, val) → __reduce_xor_sync

Bit Manipulation

• popc(), popcount(), count_ones() → __popc - Population count
• clz(), leading_zeros() → __clz - Count leading zeros
• ctz(), trailing_zeros() → __ffs - 1 - Count trailing zeros
• ffs() → __ffs - Find first set
• brev(), reverse_bits() → __brev - Bit reverse
• byte_perm() → __byte_perm - Byte permutation
• funnel_shift_left() → __funnelshift_l
• funnel_shift_right() → __funnelshift_r

Memory Operations

• ldg(ptr), load_global(ptr) → __ldg - Read-only cache load
• prefetch_l1(ptr) → __prefetch_l1 - L1 prefetch
• prefetch_l2(ptr) → __prefetch_l2 - L2 prefetch

Special Functions

• rcp(), recip() → __frcp_rn - Fast reciprocal
• fast_div() → __fdividef - Fast division
• saturate(), clamp_01() → __saturatef - Saturate to [0,1]
• j0(), j1(), jn() - Bessel functions of first kind
• y0(), y1(), yn() - Bessel functions of second kind
• normcdf(), normcdfinv() - Normal CDF
• cyl_bessel_i0(), cyl_bessel_i1() - Cylindrical Bessel functions

Clock and Timing

• clock() → clock() - 32-bit clock counter
• clock64() → clock64() - 64-bit clock counter
• nanosleep(ns) → __nanosleep - Sleep for nanoseconds

RingContext Methods

• ctx.thread_id() → threadIdx.x
• ctx.block_id() → blockIdx.x
• ctx.global_thread_id() → (blockIdx.x * blockDim.x + threadIdx.x)
• ctx.sync_threads() → __syncthreads()
• ctx.lane_id() → (threadIdx.x % 32)
• ctx.warp_id() → (threadIdx.x / 32)

Ring Kernel Intrinsics

• is_active(), should_terminate(), mark_terminated()
• messages_processed(), input_queue_size(), output_queue_size()
• input_queue_empty(), output_queue_empty(), enqueue_response(&resp)
• hlc_tick(), hlc_update(ts), hlc_now() - HLC operations
• k2k_send(target, &msg), k2k_try_recv() - K2K messaging
• k2k_has_message(), k2k_peek(), k2k_pending_count()

Type Mapping

Intrinsic Count

The transpiler supports 120+ GPU intrinsics across 13 categories:

Testing

cargo test -p ringkernel-cuda-codegen

The crate includes 171 tests covering all kernel types, intrinsics, and language features.

License

Apache-2.0