[cherubrock-seb]

Problem

The CUDA examples under examples/vector_add and examples/vector_add_tiling currently crash before they can validate their kernels.

On an RTX 3080 with CUDA 12.0, both exact upstream binaries at 67c6589aed8653629209f87d35627ac51dec3ca8 terminate with SIGSEGV (exit 139). PR #16 independently notes the same empty-vector failure and intentionally leaves these examples unchanged.

Reproduction

Build the examples in Release mode and run:

• kernel_float_vecadd
• kernel_float_vecadd_tiling

Both index zero-sized output_expected / output_result host vectors. The copy-back cudaMemcpy is also reversed, so the host result buffer is never populated correctly.

After fixing those two memory issues, the examples expose a second stale check: their expected values still use i + constant, while the kernels actually compute:

• vector_add: half-precision fma(input, input, constant)
• vector_add_tiling: the half-rounded square multiplied by constant

Root cause

The example host-side validation code drifted from the kernel behavior:

1. output vectors are declared without size;
2. the post-kernel copy uses device as destination and host as source;
3. expected values no longer match the implemented square/FMA operations.

Minimal fix

Only the two example files are changed:

• size both host output vectors to n;
• copy output_dev back into output_result.data();
• compute expected values using the same half-precision semantics as each kernel.

No library implementation or public API is changed.

Validation

Validated exact candidate head: 1b1feaa8cbb1b678193f98a70580496e7558691c

Against upstream base: 67c6589aed8653629209f87d35627ac51dec3ca8

Environment:

• NVIDIA GeForce RTX 3080
• CUDA Toolkit 12.0.140
• Release build
• CMAKE_CUDA_ARCHITECTURES=80

Results:

• baseline kernel_float_vecadd: SIGSEGV, exit 139
• baseline kernel_float_vecadd_tiling: SIGSEGV, exit 139
• candidate build: PASS
• candidate kernel_float_vecadd: PASS, prints result correct
• candidate kernel_float_vecadd_tiling: PASS, prints result correct
• git diff --check: PASS

Scope and limitations

This fixes only the two standalone CUDA examples and their validation logic. It does not alter kernel_float arithmetic, tests, or runtime behavior outside these examples.