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>

tests/validation/reference/DepthwiseConvolutionLayer.cpp

diff --git a/tests/validation/reference/DepthwiseConvolutionLayer.cpp b/tests/validation/reference/DepthwiseConvolutionLayer.cpp
index 608093d..0c7e92b 100644
--- a/tests/validation/reference/DepthwiseConvolutionLayer.cpp
+++ b/tests/validation/reference/DepthwiseConvolutionLayer.cpp
@@ -197,7 +197,7 @@
                 int         output_shift      = 0;
                 const float weights_scale     = (is_quantized_per_channel) ? weights_scale_vec[out_z] : weights_scale_vec[0];
                 const float multiplier        = input_scale * weights_scale / output_scale;
-                arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+                arm_compute::quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
 
                 for(int y = minimum_y; y <= minimum_y + maximum_y; y += conv_info.stride().second)
                 {
@@ -220,9 +220,8 @@
                             }
                         }
                         val += bias_val;
-                        val = asymm_rounding_divide_by_pow2(asymm_int_mult(val, output_multiplier), output_shift);
-                        val += output_offset;
-                        val = utility::clamp<int32_t>(val, 0, 255);
+                        // Quantize down
+                        val = quantize_down_scale_by_fixedpoint(val, output_multiplier, output_shift, output_offset, 0, 255);
 
                         // Store the result
                         dst[out_pos++] = val;