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

// 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: inline PTX assembly constructs.
//
// Exercises:
//   - asm("...") with output/input operands
//   - asm volatile("...") with no operands
//   - asm with memory clobber
//   - All PTX constructs are flagged as warnings (NVIDIA-specific, no
//     equivalent in HIP/SYCL/OpenCL/Rust — manual rewrite required)
//   - __global__ kernel, threadIdx.x, blockIdx.x, blockDim.x
//   - cuda_runtime.h header
#include <CL/cl.h> /* was: cuda_runtime.h */

__kernel void ptx_bswap(int* data, int n) {
    int i = get_group_id(0) * get_local_size(0) + get_local_id(0);
    if (i >= n) return;

    int x = data[i];

    // Byte-swap via PTX `prmt` instruction.
    int result;
    asm("prmt.b32 %0, %1, 0, 0x0123;" : "=r"(result) : "r"(x));
    data[i] = result;
}

__kernel void ptx_membar(int* data, int n) {
    int i = get_group_id(0) * get_local_size(0) + get_local_id(0);
    if (i >= n) return;

    // Memory barrier via PTX.
    asm volatile("membar.gl;");
    data[i] += 1;
}

__kernel void ptx_clock(unsigned long long* cycles) {
    // Read the GPU clock counter via PTX.
    unsigned long long c;
    asm volatile("mov.u64 %0, %%clock64;" : "=l"(c));
    if (get_group_id(0) * get_local_size(0) + get_local_id(0) == 0) {
        *cycles = c;
    }
}

__kernel void ptx_lanemask(unsigned int* mask) {
    // Get the active lane mask via PTX.
    unsigned int m;
    asm volatile("activemask.b32 %0;" : "=r"(m));
    if (get_group_id(0) * get_local_size(0) + get_local_id(0) == 0) {
        *mask = m;
    }
}

int main(void) {
    const int N = 1024;
    int* d_data = nullptr;
    unsigned long long* d_cycles = nullptr;
    unsigned int* d_mask = nullptr;

    cudaMalloc((void**)&d_data, N * sizeof(int));
    cudaMalloc((void**)&d_cycles, sizeof(unsigned long long));
    cudaMalloc((void**)&d_mask, sizeof(unsigned int));

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

    clEnqueueNDRangeKernel(queue, ptx_bswap_kernel, 1, NULL, (size_t[1]){grid}, (size_t[1]){block}, 0, NULL, NULL) /* args: d_data, N */;
    clEnqueueNDRangeKernel(queue, ptx_membar_kernel, 1, NULL, (size_t[1]){grid}, (size_t[1]){block}, 0, NULL, NULL) /* args: d_data, N */;
    clEnqueueNDRangeKernel(queue, ptx_clock_kernel, 1, NULL, (size_t[1]){1}, (size_t[1]){1}, 0, NULL, NULL) /* args: d_cycles */;
    clEnqueueNDRangeKernel(queue, ptx_lanemask_kernel, 1, NULL, (size_t[1]){1}, (size_t[1]){1}, 0, NULL, NULL) /* args: d_mask */;

    cudaDeviceSynchronize();
    cudaFree(d_data);
    cudaFree(d_cycles);
    cudaFree(d_mask);
    return 0;
}