Conv generates wrong code for dilation > 1 (GPU silent-wrong, CPU segfault)

[harz05]

SOFIE generates incorrect Conv code whenever dilation > 1, on both the alpaka GPU path and the CPU path. This is the same bug fixed upstream in ROOT by root-project/root#22474 (issue root-project/root#22473); the fix has not been ported here, and the GPU path needs an additional fix of its own.

cause

In Initialize, fAttrKernelShape is overwritten with the dilation-expanded kernel size k + (dilation - 1)*(k - 1) (a 3x3 kernel with dilation 2 becomes 5x5). That expanded value is then used again together with fAttrDilations, so dilation is counted twice. With dilation 1 the expansion k + 0 = k is a no-op, which is why no existing test catches it: every Conv model in the suite uses dilation 1.

GPU path (Generate_GPU_Kernel_ALPAKA): silently wrong output

The weight-vectorisation kernel bakes the gap into the _f layout using the expanded kernel, while the im2col kernel re-applies the same gap. For a 3x3 dilation-2 model the generated kernels contain, in the same file:

    // weight-vec: gap folded into _f (kernel treated as effective width 5, hstride = 2*5 = 10)
    std::size_t const f_idx = ... kh * 10u + kw * 2u;
    f[f_idx] = W[elem_idx];

    // im2col: rows decoded with the RAW width 3 over a 25-slot block, dilation applied AGAIN
    std::size_t const kh = k_rem / 3u;
    int64_t const ih_in = static_cast<int64_t>(oh * 1u + kh * 2u) - 0;
    int64_t const iw_in = static_cast<int64_t>(ow * 1u + kw * 2u) - 0;

The weight buffer is spaced by 10 (effective-5 grid) but the column buffer is decoded with width 3 and multiplied by dilation a second time, so the two no longer correspond. The matrix dimensions stay consistent, so there is no crash, just a wrong result.

CPU path (Generate): segfault

Emits Im2col(..., fAttrKernelShape[0], fAttrKernelShape[1], ..., fAttrDilations[0], fAttrDilations[1], ...), i.e. the expanded kernel (5,5) together with dilation (2,2). Im2col already applies dilation internally, so it computes a negative output dimension and writes past the _xcol buffer, segfaulting. This is exactly root-project/root#22473; the one-line fix from root-project/root#22474 (fAttrDilations = {1,1,1} after the weight reorder) is missing here.

Reproducer

A 3x3 dilation-2 Conv, input 1x1x7x7, no padding, unit stride (the same model as the ROOT reproducer in root-project/root#22473). Generated with GenerateGPU_ALPAKA the output does not match the ONNX/PyTorch reference; the CPU Generate segfaults.

Expected

Generated Conv matches the ONNX reference for any valid dilation, the same way it already does for padding and strides.