AI Math Gets Faster on Hexagon Chips, Boosting Your Phone's Smarts
Your phone's AI just got a speed boost—here's why that matters.
The softplus function in this backend is now vectorized: softplus_f32 and tile_softplus_f32 previously ran a scalar logf(1.0f + expf(x)) loop, and both are replaced with a shared hvx_softplus_f32_aa() helper built from the backend's existing hvx_vec_exp_f32 and hvx_vec_log_f32 primitives — so no new numerical approximation is introduced. Because hvx_vec_exp_f32 clamps its input to (-88, 88), a naive log(1 + exp(x)) composition would saturate to ~88.72 for large x instead of returning x; the Q6_V_vmux_QVV select keeps the scalar reference's x > 20 guard (ggml_compute_softplus_f32 in ggml-impl.h), which is what makes the vector path match the CPU result.
- Qualcomm's Hexagon chips now compute softplus (an AI math function) faster by processing multiple data points at once.
- This means AI apps on your phone will run smoother and use less battery.
- The improvement is limited to devices with Hexagon chips, but it's a step toward more powerful on-device AI.
Why It Matters
Faster AI on your phone means quicker voice assistants, better photo filters, and longer battery life.