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

// 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.

// Richer fixture exercising atomics, shared memory, syncthreads, warp
// intrinsics, 2D launches, and a constant-memory global.
#include <hip/hip_runtime.h> /* was: cuda_runtime.h */

#define N 1024

__constant__ float kCoefficients[8];

__global__ void increment_atomic(int* counter, int delta) {
    int prev = atomicAdd(counter, delta);
    if (threadIdx.x == 0) {
        atomicCAS(counter, prev, prev + 1);
    }
}

__global__ void warp_scan(const int* in, int* out) {
    __shared__ int buf[32];
    int lane = __laneid()();
    int wid = threadIdx.x / 32;
    buf[lane] = in[threadIdx.x];
    __syncwarp(0xFFFFFFFFu);
    if (lane == 0) {
        int s = 0;
        for (int i = 0; i < 32; ++i) s += buf[i];
        out[wid] = s;
    }
    __syncthreads();
}

__global__ void matmul(const float* a, const float* b, float* c,
                       int m, int n, int k) {
    int row = blockIdx.y * blockDim.y + threadIdx.y;
    int col = blockIdx.x * blockDim.x + threadIdx.x;
    if (row < m && col < n) {
        float s = kCoefficients[0] * a[row * k] * b[col];
        c[row * n + col] = s;
    }
}

void launch_examples(int* counter, float* a, float* b, float* c) {
    dim3 grid(N / 32, N / 16);
    dim3 block(32, 16);
    hipLaunchKernelGGL(increment_atomic, dim3(grid), dim3(block), 0, 0, counter, 1);
    hipLaunchKernelGGL(warp_scan, dim3(1), dim3(32), 0, 0, counter, a);
    hipLaunchKernelGGL(matmul, dim3(grid), dim3(block), 0, 0, a, b, c, N, N, N);
}