commit 14cbfb2921990d8bf125231e350e2ac8dcd95a8b
author Michele Di Giorgio <michele.digiorgio@arm.com> Wed Oct 23 10:53:10 2019 +0100
committer Michele Di Giorgio <michele.digiorgio@arm.com> Thu Nov 28 10:02:15 2019 +0000
tree 9bec073d72c44c480c8807601889481d9b89ee7e
parent ed7b27dd7cbdae57b880029840ad0235523848e0

COMPMID-2609: Enable quantization with multiplier greater than 1 on OpenCL

Change-Id: I050f1f84e214e61f7cbb0197a672b68a4940edae
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2158
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>

src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp

diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
index a2f4a91..d9705de 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
@@ -297,6 +297,14 @@
         build_opts.add_option_if(is_quantized_per_channel, "-DPER_CHANNEL_QUANTIZATION");
         build_opts.add_option_if(is_dot8_supported, "-DIS_DOT8");
 
+        // Compute non-per-channel multiplier and shift anyway to make OpenCL kernel simpler
+        float multiplier        = iq_info.scale * wq_info.scale / oq_info.scale;
+        int   output_multiplier = 0;
+        int   output_shift      = 0;
+        quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
+        build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
+        build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
+
         if(act_info.enabled())
         {
             const int a_val = quantize_qasymm8(act_info.a(), oq_info);