commit a400334f60c28a464ba44170350ec475af73a94a
author Alan Kelly <me@alankelly.dev> Wed Apr 10 21:37:48 2019 +0200
committer Georgios Pinitas <georgios.pinitas@arm.com> Fri May 17 16:24:50 2019 +0000
tree ee43a78fed8d0db44143bc99daeae7adc9d429d6
parent 32bd4dd1b8f39161ad4bbbe118ecfba563d70f90

Optimizes NEBox3x3Kernel by replacing floating point computations with fixed point
Signed-off-by: Alan Kelly <me@alankelly.dev>

Change-Id: Ic48a2db1053433f47eaa5a2d9dd1fa12e33c3953
Reviewed-on: https://review.mlplatform.org/c/1045
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

src/core/NEON/kernels/NEBox3x3Kernel.cpp

diff --git a/src/core/NEON/kernels/NEBox3x3Kernel.cpp b/src/core/NEON/kernels/NEBox3x3Kernel.cpp
index 7a53f93..a401aa7 100644
--- a/src/core/NEON/kernels/NEBox3x3Kernel.cpp
+++ b/src/core/NEON/kernels/NEBox3x3Kernel.cpp
@@ -158,7 +158,9 @@
     unsigned char *const input_mid_ptr = _input->ptr_to_element(Coordinates(-1, 0));
     unsigned char *const input_bot_ptr = _input->ptr_to_element(Coordinates(-1, +1));
 
-    const float32x4_t oneovernine = vdupq_n_f32(1.0f / 9.0f);
+    const int       shift       = 19;
+    int             value       = (1 << shift) / 9 + 1; //58255 / (2^19) ~= 1/9
+    const int32x4_t oneovernine = vdupq_n_s32(value);
 
     execute_window_loop(window, [&](const Coordinates &)
     {
@@ -207,14 +209,15 @@
         //bot right
         out = vaddq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
 
-        float32x4_t outfloathigh = vcvtq_f32_s32(vmovl_s16(vget_high_s16(out)));
-        float32x4_t outfloatlow  = vcvtq_f32_s32(vmovl_s16(vget_low_s16(out)));
+        int32x4_t outfloathigh = vmovl_s16(vget_high_s16(out));
+        int32x4_t outfloatlow  = vmovl_s16(vget_low_s16(out));
 
-        outfloathigh = vmulq_f32(outfloathigh, oneovernine);
-        outfloatlow  = vmulq_f32(outfloatlow, oneovernine);
-
-        out = vcombine_s16(vqmovn_s32(vcvtq_s32_f32(outfloatlow)),
-                           vqmovn_s32(vcvtq_s32_f32(outfloathigh)));
+        outfloathigh = vmulq_s32(outfloathigh, oneovernine);
+        outfloatlow  = vmulq_s32(outfloatlow, oneovernine);
+        outfloathigh = vshrq_n_s32(outfloathigh, shift);
+        outfloatlow  = vshrq_n_s32(outfloatlow, shift);
+        out          = vcombine_s16(vqmovn_s32((outfloatlow)),
+                           vqmovn_s32((outfloathigh)));
 
         vst1_u8(output.ptr(), vqmovun_s16(out));
     },