Imported from PlumBlossomMaid/Insight7 (.qwen/skills/fix-full-kernel-view-offset/SKILL.md). Install upstream with npx skills add PlumBlossomMaid/Insight7 --skill fix-full-kernel-view-offset. Copyright stays with the author.

Fix full kernel for view offset

The Problem

The full kernel (backends/cpu/kernels/creation/full.cpp and .cu) writes to out->data directly, ignoring out->offset. When out is a view (e.g., arr["1,1"]), the data pointer points to the parent array's start, and offset indicates where the view begins. Without adding offset, the kernel writes to position 0 instead of the view's actual position.

// WRONG: ignores offset
float *dst = static_cast<float *>(out->data);
dst[i] = val;

// CORRECT: applies offset
float *dst = static_cast<float *>(out->data) + out->offset;
dst[i] = val;

Symptom

a = ins.from_numpy(np.arange(9).reshape(3,3))
a["1,1"] = 99  # Should write to position [1,1] (offset 4)
# Bug: writes to position [0,0] (offset 0)
print(a.numpy())  # [[99, 1, 2], [3, 4, 5], [6, 7, 8]] ← WRONG

Scalar view a["1,1"] has is_contiguous() == true (0-dim), so fill_() dispatches directly to the full kernel. The kernel writes to out->data[0] which is the parent's position 0.

The Fix

CPU kernel (backends/cpu/kernels/creation/full.cpp)

Add offset to the data pointer in every dtype case:

int64_t offset = out->offset;
// ...
case INSIGHT_DTYPE_F64: {
    double *dst = static_cast<double *>(out->data) + offset;
    for (int64_t i = 0; i < n; ++i)
        dst[i] = fill_val;
    break;
}

CUDA kernel (backends/cuda/kernels/creation/full.cu)

Compute offset-adjusted pointer before the switch:

size_t elem_size = insight_dtype_size(dtype);
char *data_with_offset = static_cast<char *>(out->data) + out->offset * elem_size;
// Then cast data_with_offset in each case:
full_kernel<<<blocks, threads>>>(reinterpret_cast<float *>(data_with_offset), val, n);

How to detect

Test scalar element assignment on a 2D array:

a = ins.from_numpy(np.arange(9, dtype=np.float64).reshape(3,3))
a["1,1"] = 99
assert a.numpy()[1,1] == 99  # Fails without fix (writes to [0,0])

Root cause

The full kernel was designed for creating new arrays (where offset is always 0), not for filling views. When fill_() reuses the full kernel for contiguous views, the offset must be respected.

Related Skills

• add-array-setitem — fill_() dispatches to full kernel for contiguous views
• fix-contiguous-copy-noncontiguous-output — similar pattern (kernel ignores output metadata)
• fix-strided-view-data-access — general category of view offset bugs