Skip to main content

GPU Blocks

The gpu { } block offloads tensor operations to the GPU, with automatic CPU fallback.

Basic Syntax​

gpu {
    result is a + b;
    product is a * b;
}

Everything inside the block runs on the GPU when available.

CPU Fallback​

If no GPU is available, the same code runs on the CPU — no code changes needed.

Real Example​

From GpuBlockTensorOps.mx:

Init method() {
    a is Tensor<float>.Random(2, 2);
    b is Tensor<float>.Random(2, 2);
    c is Tensor<float>.Random(2, 2);

    cpuSum is a + b;
    cpuMul is a * b;

    gpu {
        gpuSum is a + b;
        gpuMul is a * b;
        gpuFmaLike is (a * b) + c;
    }

    Print("gpu block prefers GPU and falls back to CPU.");
}

Restrictions​

  • gpu { } blocks are only allowed inside method bodies
  • Operations must be tensor-based (element-wise, matmul)
  • No control flow (if/while/for) inside gpu blocks

Next Steps​