commit 4a24a58cef2a22434118636d5cbadab57704bcd5
author Marat Dukhan <maratek@google.com> Mon Jan 06 13:30:00 2020 -0800
committer XNNPACK Team <xnnpack-github-robot@google.com> Mon Jan 06 13:30:29 2020 -0800
tree 635f748dca22a484d44ee36da59c0c176d5ac305
parent 77221d39511adaf2115d1482ee0af31a5a3052e3

Use 1-step range reduction in NEONFMA Sigmoid micro-kernels

- Parametrize F32 Sigmoid micro-kernels by the number of range reduction steps
- Use 1-step range reduction in NEONFMA implementations (~5% performance
  improvement on Pixel 2 & 3a, ~6% on Mi A2 Lite)

PiperOrigin-RevId: 288364064

src/init.c

diff --git a/src/init.c b/src/init.c
index 85872be..f0bf70f 100644
--- a/src/init.c
+++ b/src/init.c
@@ -242,7 +242,7 @@
     };
     xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__neon;
     xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__neon_x8;
-    xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__neon_lut64_p2_nr2recps_x8;
+    xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8;
     xnn_params.f32.prelu = (struct prelu_parameters) {
       .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__neon_2x8,
       .row_tile = 2,
@@ -546,7 +546,7 @@
     };
     xnn_params.f32.clamp = (xnn_univector_ukernel_function) xnn_f32_clamp_ukernel__neon;
     xnn_params.f32.hswish = (xnn_univector_ukernel_function) xnn_f32_hswish_ukernel__neonfma_x8;
-    xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__neonfma_lut64_p2_nr2recps_x16;
+    xnn_params.f32.sigmoid = (xnn_univector_ukernel_function) xnn_f32_sigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16;
     xnn_params.f32.prelu = (struct prelu_parameters) {
       .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__neon_2x8,
       .row_tile = 2,