FaceDetection is not working on wasm backend

Hi everyone, first of all, thanks for this excellent library! I ran into an issue while trying to run FaceDetection or FaceLandmarksDetection models on the wasm backend.

Describe the current behavior While this works well on WebGL backend, when I use the wasm backend, I get either empty array as a result or a memory access out of bounds (might be related to #6128)

RuntimeError: memory access out of bounds
    at tfjs-backend-wasm-simd.wasm:0x3ff23
    at tfjs-backend-wasm-simd.wasm:0x2cdc7
    at tfjs-backend-wasm-simd.wasm:0x2dacd
    at tfjs-backend-wasm-simd.wasm:0x1d5d9
    at tfjs-backend-wasm-simd.wasm:0x68da
    at Module._FusedConv2D (tfjs-backend-wasm.js:9:21945)
    at Object.fusedConv2d [as kernelFunc] (FusedConv2D.ts:140:3)
    at kernelFunc (engine.ts:644:22)
    at engine.ts:710:23
    at Engine.scopedRun (engine.ts:478:19)

You can find code sandbox examples with WebGL and wasm below. I also tried this stock example provided by TensorFlow.js by adding tfjs-wasm in the list of model backends, but I get the same issue with nothing detected at all (empty array)

Is there something I am missing or doing wrong here?

Thanks for your help!

Describe the expected behavior I guess the expected behavior is that it works on both wasm and WebGL backend.

Standalone code to reproduce the issue Example with wasm backend: https://codesandbox.io/s/confident-nightingale-b8z48b Example with WebGL backend: https://codesandbox.io/s/elegant-tereshkova-mdqj6k

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow.js): Yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): MacOS Monterey
• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device: Pixel 4a
• TensorFlow.js installed from (npm or script link): npm
• TensorFlow.js version (use command below): 3.18.0
• Browser version: Chrome 102.0.5005.61
• Tensorflow.js Converter Version: 3.18.0

[mattsoulanille]

Quick update to this issue: I built the WASM backend with debug flags set, and here’s its more descriptive error message. Thanks for posting a reproduction!

RuntimeError: memory access out of bounds
    at xnn_f32_igemm_relu_ukernel_4x8__wasmsimd_splat (tfjs-backend-wasm-simd.wasm:0x1608e4)
    at xnn_compute_igemm (tfjs-backend-wasm-simd.wasm:0xd29af)
    at pthreadpool_parallelize_2d_tile_2d (tfjs-backend-wasm-simd.wasm:0x1dc567)
    at xnn_run_operator (tfjs-backend-wasm-simd.wasm:0xd58b0)
    at tfjs::wasm::conv2d(unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, bool, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, bool, FusableActivation, unsigned long, float, unsigned long) (tfjs-backend-wasm-simd.wasm:0x600e9)
    at FusedConv2D (tfjs-backend-wasm-simd.wasm:0x2a2e0)
    at tfjs-backend-wasm.js:2176:1
    at ccall (tfjs-backend-wasm.js:1368:1)
    at tfjs-backend-wasm.js:1383:1
    at Object.fusedConv2d [as kernelFunc] (FusedConv2D.ts:140:1)

[google-ml-butler[bot]]

Are you satisfied with the resolution of your issue? Yes No