Equivalent FP32 head experiments
Keep RTMW3D-L trained weights, image resolution and all 133 3D joints while comparing explicit im2col plus MatMul against the original convolution kernels. The split8 candidate increases the batch of independent partial products and sums their FP32 results. Speed is not assumed before GPU measurement. Upstream model and implementation: https://github.com/open-mmlab/mmpose The gpu-decoder control only replaces the finite-score minimum with Less plus Where, which are registered in ORT 1.27 JSEP, to test the graph-capture placement restriction separately. The other two candidates also rewrite the measured final head convolutions. Original FP32 weights are reordered without quantization; all other model parameters are preserved. All three pass small public-fixture CPU conversion checks, including independent hips, knees, ankles and foot gates. This is not target-video accuracy or native WebGPU validation. Use the browser benchmark to measure synchronized inference, output drift and per-kernel GPU time on the target device.