decuda 0.1.1

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
88
89
90
91
92
93
94
95
96
97
98
99
100

// Generated by decuda. Edit with care.
// HIP is mostly source-compatible with CUDA at the kernel level.
// Compare against the original .cu file for sanity.

// Advanced fixture: warp-level primitives and cooperative patterns.
//
// Exercises:
//   - __shfl_sync (NOT a known builtin -> preserved, flagged)
//   - __ballot_sync (NOT a known builtin -> preserved, flagged)
//   - __any_sync (NOT a known builtin -> preserved, flagged)
//   - __all_sync (NOT a known builtin -> preserved, flagged)
//   - __activemask (NOT a known builtin -> preserved, flagged)
//   - __laneid (known builtin -> rewritten per target)
//   - __syncwarp (known builtin -> rewritten per target)
//   - warpSize (known builtin -> rewritten per target)
//   - __syncthreads (known builtin -> rewritten per target)
//   - atomicAdd, atomicMin, atomicMax (known atomics -> preserved/rewritten)
//   - __global__, __device__, __shared__ qualifiers
//   - threadIdx.x, blockIdx.x, blockDim.x, gridDim.x
//   - cuda_runtime.h header
#include <hip/hip_runtime.h> /* was: cuda_runtime.h */

#define WARP 32

__device__ __forceinline__ int warp_sum(int v) {
    // __shfl_sync is NOT in decuda's builtin table — preserved verbatim.
    for (int offset = WARP / 2; offset > 0; offset /= 2) {
        v += __shfl_sync(0xFFFFFFFFu, v, __laneid()() - offset);
    }
    return v;
}

__device__ __forceinline__ int warp_ballot(int predicate) {
    // __ballot_sync is NOT in decuda's builtin table — preserved verbatim.
    return __ballot_sync(0xFFFFFFFFu, predicate);
}

__device__ __forceinline__ int warp_any(int predicate) {
    // __any_sync is NOT in decuda's builtin table — preserved verbatim.
    return __any_sync(0xFFFFFFFFu, predicate);
}

__device__ __forceinline__ int warp_all(int predicate) {
    // __all_sync is NOT in decuda's builtin table — preserved verbatim.
    return __all_sync(0xFFFFFFFFu, predicate);
}

__global__ void warp_demo(const int* in, int* out, int n) {
    __shared__ int shared[WARP];
    int tid = threadIdx.x;
    int gid = blockIdx.x * blockDim.x + tid;
    int lane = __laneid()();

    // Load value and compute predicate.
    int v = (gid < n) ? in[gid] : 0;
    int pred = (v > 0) ? 1 : 0;

    // Warp-level vote: how many lanes have pred == 1?
    int ballot = warp_ballot(pred);
    int any_pos = warp_any(pred);
    int all_pos = warp_all(pred);
    int active = __activemask();

    // Warp-level sum via shuffle.
    v = warp_sum(v);
    __syncwarp(0xFFFFFFFFu);

    if (lane == 0) {
        shared[tid / WARP] = v;
        atomicAdd(out, v);
        atomicMin(out + 1, ballot);
        atomicMax(out + 2, active);
    }
    __syncthreads();

    // Record vote results from lane 0 of the first warp.
    if (tid == 0) {
        atomicAdd(out + 3, any_pos);
        atomicAdd(out + 4, all_pos);
    }
}

int main(void) {
    const int N = 1 << 16;
    int* d_in = nullptr;
    int* d_out = nullptr;

    hipMalloc((void**)&d_in, N * sizeof(int));
    hipMalloc((void**)&d_out, 5 * sizeof(int));
    hipMemset(d_out, 0, 5 * sizeof(int));

    dim3 grid(N / 256);
    dim3 block(256);
    hipLaunchKernelGGL(warp_demo, dim3(grid), dim3(block), 0, 0, d_in, d_out, N);

    hipDeviceSynchronize();
    hipFree(d_in);
    hipFree(d_out);
    return 0;
}