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

// Generated by decuda.
// OpenCL device code is the bulk of this file. The host-side calls
// (`cudaXxx`) have been rewritten to OpenCL equivalents inline; the
// surrounding host program still needs a cl_context + cl_queue, not
// included here. Look for TODO(decuda) markers for items requiring
// manual attention.

// Advanced fixture: unified/managed memory patterns.
//
// Exercises:
//   - cudaMallocManaged (in DB -> hipMallocManaged for HIP; flagged for SYCL/Rust/OpenCL)
//   - __managed__ qualifier (in DB -> per-target rewrite)
//   - cudaHostAlloc (in DB -> renamed for HIP)
//   - cudaMallocHost (in DB -> renamed for HIP)
//   - cudaMemcpy (in DB -> renamed for HIP)
//   - cudaDeviceSynchronize (in DB -> hipDeviceSynchronize for HIP; flagged for SYCL/Rust/OpenCL)
//   - __global__ and __device__ qualifiers
//   - threadIdx.x, blockIdx.x, blockDim.x
//   - cuda_runtime.h header
#include <CL/cl.h> /* was: cuda_runtime.h */

// __managed__ memory: accessible from both host and device without explicit
// cudaMemcpy. decuda rewrites __managed__ per target and flags the
// managed-memory runtime APIs for manual review.
/* use SVM: clSVMAlloc */ int managed_counter = 0;

__device  int atomic_increment(int* addr) {
    return atomicAdd(addr, 1);
}

__kernel void increment_kernel(int* counter, int n) {
    int i = get_group_id(0) * get_local_size(0) + get_local_id(0);
    if (i < n) {
        atomic_increment(counter);
    }
}

__kernel void scale_kernel(float* data, float scale, int n) {
    int i = get_group_id(0) * get_local_size(0) + get_local_id(0);
    if (i < n) {
        data[i] *= scale;
    }
}

int main(void) {
    const int N = 1 << 16;

    // Managed memory: no explicit cudaMemcpy needed between host and device.
    // cudaMallocManaged is in the DB for HIP (-> hipMallocManaged);
    // for SYCL/Rust/OpenCL it is preserved verbatim and flagged.
    float* managed_data = nullptr;
    cudaMallocManaged((void**)&managed_data, N * sizeof(float));

    // Initialize on the host (managed memory is directly accessible).
    for (int i = 0; i < N; i++) {
        managed_data[i] = (float)i;
    }

    // Pinned host memory for comparison.
    float* pinned_data = nullptr;
    cudaMallocHost((void**)&pinned_data, N * sizeof(float));

    dim3 grid(N / 256);
    dim3 block(256);

    // Launch on managed memory — no cudaMemcpy required.
    clEnqueueNDRangeKernel(queue, scale_kernel_kernel, 1, NULL, (size_t[1]){grid}, (size_t[1]){block}, 0, NULL, NULL) /* args: managed_data, 2.0f, N */;
    clEnqueueNDRangeKernel(queue, increment_kernel_kernel, 1, NULL, (size_t[1]){grid}, (size_t[1]){block}, 0, NULL, NULL) /* args: &managed_counter, N */;

    // cudaDeviceSynchronize is in the DB for HIP (-> hipDeviceSynchronize);
    // for SYCL/Rust/OpenCL it is preserved verbatim and flagged.
    cudaDeviceSynchronize();

    // Read back managed memory directly on the host.
    printf("counter = %d, data[0] = %f\n", managed_counter, managed_data[0]);

    // Copy to pinned memory for explicit transfer path.
    cudaMemcpy(pinned_data, managed_data, N * sizeof(float), cudaMemcpyDeviceToHost);

    cudaFree(managed_data);
    cudaFreeHost(pinned_data);
    return 0;
}