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

// 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: tiled matrix transpose using a shared-memory tile.
//
// Exercises:
//   - 2D grid and block (dim3 with two components)
//   - __shared__ tile with __syncthreads barriers
//   - blockIdx.{x,y} and threadIdx.{x,y} and blockDim.{x,y}
//   - __global__ kernel with pointer + dimension args
//   - cudaMalloc / cudaFree / cudaMemcpy
//   - cuda_runtime.h header
#include <CL/cl.h> /* was: cuda_runtime.h */

#define TILE 16

__kernel void transpose(const float* in, float* out, int width, int height) {
    __local float tile[TILE][TILE];

    int x = get_group_id(0) * get_local_size(0) + get_local_id(0);
    int y = get_group_id(0) * get_local_size(0) + get_local_id(0);

    // Load a tile from global memory into shared memory (coalesced read).
    if (x < width && y < height) {
        tile[get_local_id(0)][get_local_id(0)] = in[y * width + x];
    }
    barrier(CLK_LOCAL_MEM_FENCE);

    // Write the transposed tile back to global memory.
    int ox = get_group_id(0) * get_local_size(0) + get_local_id(0);
    int oy = get_group_id(0) * get_local_size(0) + get_local_id(0);
    if (ox < height && oy < width) {
        out[oy * height + ox] = tile[get_local_id(0)][get_local_id(0)];
    }
}

int main(void) {
    const int W = 1024;
    const int H = 1024;
    float* d_in = nullptr;
    float* d_out = nullptr;

    cudaMalloc((void**)&d_in, W * H * sizeof(float));
    cudaMalloc((void**)&d_out, W * H * sizeof(float));
    cudaMemcpy(d_in, d_in, W * H * sizeof(float), cudaMemcpyDeviceToDevice);

    dim3 grid(W / TILE, H / TILE);
    dim3 block(TILE, TILE);
    clEnqueueNDRangeKernel(queue, transpose_kernel, 1, NULL, (size_t[1]){grid}, (size_t[1]){block}, 0, NULL, NULL) /* args: d_in, d_out, W, H */;

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