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.

// 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
#include <CL/cl.h> /* was: cuda_runtime.h */
#include <device_functions.h>

#define WARP 32

__constant float c_scale[4];

__device  float fast_scale(float v, int idx) {
    return v * c_scale[idx & 3];
}

__device  float slow_path(float v) {
    if (v < 0.0f) return 0.0f;
    return __sinf(v);
}

__launch_bounds__(256, 2)
__kernel void apply(const float* in, float* out, int n) {
    int i = get_group_id(0) * get_local_size(0) + get_local_id(0);
    if (i < n) {
        float v = fast_scale(in[i], i);
        out[i] = slow_path(v);
    }
}

__kernel void warp_sum(const float* in, float* out, int n) {
    int lane = get_sub_group_id() * get_sub_group_size() + get_sub_group_local_id()();
    float v = (get_local_id(0) < n) ? in[get_local_id(0)] : 0.0f;
    barrier(CLK_LOCAL_MEM_FENCE) /* approx */;
    // In-warp sum via shared memory.
    __local float partial[WARP];
    partial[lane] = v;
    barrier(CLK_LOCAL_MEM_FENCE);
    if (lane == 0) {
        float s = 0.0f;
        for (int i = 0; i < WARP; ++i) s += partial[i];
        out[get_group_id(0)] = 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);
    clEnqueueNDRangeKernel(queue, apply_kernel, 1, NULL, (size_t[1]){grid}, (size_t[1]){block}, 0, NULL, NULL) /* args: d_in, d_out, N */;
    clEnqueueNDRangeKernel(queue, warp_sum_kernel, 1, NULL, (size_t[1]){N / WARP}, (size_t[1]){WARP}, 0, NULL, NULL) /* args: d_in, d_warp, N */;

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