SIMD Support Seems Broken

[ertza]

Description

When compiling and running an OpenCV-based face detection and cropping function inside hyperlight-wasm, the output image is malformed, despite the function executing without error. The same WASM when compiled and executed under wasmtime produces the correct output. Both output files attached at the end.

Reproduction Context

• Commit: bd4423a
• Input: 512x512 RGB image lena.jpg
• Output: 100x100 RGB face crop
• Toolchain:
  • C++ OpenCV application (face detection via Haarcascade), compiled to WASM using Emscripten
• Compiler + Runtimes tested:
  • wasmtime (correct image output)
  • hyperlight-wasm (malformed output)

SIMD Instruction Comparison

The following instruction counts were extracted from both compiled WASMs:

Runtime	SSE	SSE2	SSE3	SSSE3	SSE4.1	SSE4.2	AVX	AVX2	AVX-512	Total SIMD
Wasmtime	4,798	24,118	0	114	4,631	148	4,391	17,654	0	55,854
Hyperlight	6,638	93,573	0	0	0	0	0	0	0	100,211

Observations

• The image output buffer length and dimensions are correct (100 x 100 x 3 = 30000 bytes).
• The exact same WASM module produces correct output under wasmtime.
• The issue disappears when OpenCV is compiled with SIMD disabled entirely, indicating the failure is tied to SIMD instruction execution or code generation.

Expected Behavior

Hyperlight should produce correct image output when the binary is built with SIMD enabled.

Actual Behavior

Image output is corrupted or malformed under hyperlight-wasm, also no SIMD instructions beyond SSE2 appear in the compiled binary.

Image Outputs

Hyperlight-wasm

Wasmtime

[ertza]

To add, running the hyperlight-wasm compiled binary using wasmtime also outputs the correct image. The binary only has upto SSE2 instructions as it was compiled using hyperlight-wasm, as stated above.

[simongdavies]

My suspicion is that this is due corruption in the serialisation/deserialisation of either the input or the results since this works as expected targeting x84-64_unknown_none and running in wasmtime @ertza can you share the solution that you used to get these results please?

[ertza]

I have created a repo with the code and instructions for this here https://github.com/ertza/wasmtime-runner