decuda 0.1.0

CUDA to HIP, SYCL, OpenCL, and Rust GPU migration tool — automatic source-code translator for porting CUDA C++ kernels to AMD ROCm HIP, Intel oneAPI SYCL, Khronos OpenCL, and Rust GPU (cust / rust-gpu)
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87

// Generated by decuda.
// Host-side runtime calls are translated to the `cust` crate. Kernels
// are emitted as `TODO(decuda)` blocks: rust-gpu translation requires
// the kernel to be authored as a Rust fn. See `examples/` for a
// scaffolded SPIR-V kernel module you can flesh out.
//
// Add to your Cargo.toml:
//   [dependencies]
//   cust = "0.3"

// Complex fixture: __device__ helper functions, __forceinline__ / __noinline__,
// __constant__ memory, warp intrinsics, and a __launch_bounds__ hint.
//
// Exercises:
//   - __device__ helper functions (one __forceinline__, one __noinline__)
//   - __constant__ memory arrays
//   - __launch_bounds__ (flagged in the report, not auto-translated)
//   - __laneid, __syncwarp, __syncthreads
//   - threadIdx / blockIdx / blockDim
//   - cudaMalloc / cudaFree / cudaMemcpy
//   - device_functions.h and cuda_runtime.h headers
// was: #include cuda_runtime.h  ->  cust::cuda_build_setup() /* TODO: import cust crate */
#include <device_functions.h>

#define WARP 32

/* constant */ float c_scale[4];

/* device function */ #[inline(always)] float fast_scale(float v, int idx) {
    return v * c_scale[idx & 3];
}

/* device function */ #[inline(never)] float slow_path(float v) {
    if (v < 0.0f) return 0.0f;
    return __sinf(v);
}

__launch_bounds__(256, 2)
// TODO(decuda): rewrite as rust-gpu kernel fn
 void apply(const float* in, float* out, int n) {
    int i = block_idx * block_dim + thread_idx;
    if (i < n) {
        float v = fast_scale(in[i], i);
        out[i] = slow_path(v);
    }
}

// TODO(decuda): rewrite as rust-gpu kernel fn
 void warp_sum(const float* in, float* out, int n) {
    int lane = lane_id;
    float v = (thread_idx < n) ? in[thread_idx] : 0.0f;
    /* syncwarp: only lane=0 of warp at once */;
    // In-warp sum via shared memory.
    /* shared -> rust-gpu group_memory */ float partial[WARP];
    partial[lane] = v;
    group.sync();
    if (lane == 0) {
        float s = 0.0f;
        for (int i = 0; i < WARP; ++i) s += partial[i];
        out[block_idx] = s;
    }
}

int main(void) {
    const int N = 1 << 18;
    float* d_in = nullptr;
    float* d_out = nullptr;
    float* d_warp = nullptr;

    cudaMalloc((void**)&d_in, N * sizeof(float));
    cudaMalloc((void**)&d_out, N * sizeof(float));
    cudaMalloc((void**)&d_warp, (N / WARP) * sizeof(float));

    float scale_init[4] = {1.0f, 2.0f, 3.0f, 4.0f};
    cudaMemcpyToSymbol(c_scale, scale_init, sizeof(scale_init));

    dim3 grid(N / 256);
    dim3 block(256);
    { /* decuda cust launch */ let _kernel = modules.get_function("apply"); unsafe { let _ = launch!( _kernel<<<grid as grid_size, block as block_size, 0 as usize, default>>>(d_in, d_out, N) ); } };
    { /* decuda cust launch */ let _kernel = modules.get_function("warp_sum"); unsafe { let _ = launch!( _kernel<<<N / WARP as grid_size, WARP as block_size, 0 as usize, default>>>(d_in, d_warp, N) ); } };

    cudaDeviceSynchronize();
    cudaFree(d_in);
    cudaFree(d_out);
    cudaFree(d_warp);
    return 0;
}