commit f81652dc970a8071f41c4138508c39684ed9111a
author steniu01 <steven.niu@arm.com> Mon Sep 11 15:29:12 2017 +0100
committer Anthony Barbier <anthony.barbier@arm.com> Fri Nov 02 16:35:24 2018 +0000
tree 6bbda484b46784f1e9ddf132849c15d16ee61230
parent 9fe414430c3c989b1cdc79d41e031495aed2cb7c

COMPMID-516 Increase tolerance rate of Scale, Conv, fully connected and GEMM

This patch also fix the scale kernel issue where it was calcuated the
scale factor inside the gpu but now in the CPU. The GPU and CPU gave
different result for simple float division operation

Change-Id: Ib6709cb6c41dcf4fc0fa4eb79e481430695bf40e
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/87266
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Moritz Pflanzer <moritz.pflanzer@arm.com>

src/core/CL/kernels/CLScaleKernel.cpp

diff --git a/src/core/CL/kernels/CLScaleKernel.cpp b/src/core/CL/kernels/CLScaleKernel.cpp
index 66afc3d..82ebe64 100644
--- a/src/core/CL/kernels/CLScaleKernel.cpp
+++ b/src/core/CL/kernels/CLScaleKernel.cpp
@@ -98,9 +98,12 @@
     ICLKernel::configure(win);
 
     // Set static kernel arguments
+    const float scale_x = static_cast<float>(input->info()->dimension(0)) / output->info()->dimension(0);
+    const float scale_y = static_cast<float>(input->info()->dimension(1)) / output->info()->dimension(1);
+
     unsigned int idx = 2 * num_arguments_per_2D_tensor(); //Skip the input and output parameters
     _kernel.setArg<float>(idx++, input->info()->dimension(0));
     _kernel.setArg<float>(idx++, input->info()->dimension(1));
-    _kernel.setArg<float>(idx++, output->info()->dimension(0));
-    _kernel.setArg<float>(idx++, output->info()->dimension(1));
+    _kernel.setArg<float>(idx++, scale_x);
+    _kernel.setArg<float>(idx++, scale_y);
 }