From 47d5170f7e1f7491db902b8b07703dd395c304be Mon Sep 17 00:00:00 2001
From: Max Bergfelt <max.bergfelt@arm.com>
Date: Tue, 27 May 2025 10:01:53 +0200
Subject: [PATCH] MLBEDSW-10832: Align ElementwiseMul quantization scaling with
 reference

Adjusted quantization scaling of ElementwiseMul to bit-exactly match
the scaling calculation of QuantizedConvolutionMultipler in the
reference. Scaling of BatchMatMul is consequently aligned.

Change-Id: I00cc55d04ed84acdd8785e9f3610f6198423c65e
Signed-off-by: Max Bergfelt <max.bergfelt@arm.com>
---
 ethosu/regor/architecture/ethosu85/ethos_u85_scaling.cpp | 2 +-
 ethosu/regor/common/scaling.hpp                          | 6 +++---
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/ethosu/regor/architecture/ethosu85/ethos_u85_scaling.cpp b/ethosu/regor/architecture/ethosu85/ethos_u85_scaling.cpp
index b7171177..166774d0 100644
--- a/ethosu/regor/architecture/ethosu85/ethos_u85_scaling.cpp
+++ b/ethosu/regor/architecture/ethosu85/ethos_u85_scaling.cpp
@@ -99,7 +99,7 @@ void RescaleConvolution(HLCOperation *op)
         }
         else
         {
-            outScale = ElementwiseMulScale(ifm1Scale, ifm2Scale, ofmScale);
+            outScale = ElementwiseMulScale<float, double>(ifm1Scale, ifm2Scale, ofmScale);
         }
     }
 
diff --git a/ethosu/regor/common/scaling.hpp b/ethosu/regor/common/scaling.hpp
index 0b4d0360..f288dd54 100644
--- a/ethosu/regor/common/scaling.hpp
+++ b/ethosu/regor/common/scaling.hpp
@@ -50,15 +50,15 @@ public:
 };
 
 /* Calculate elementwise Mul OFM QuantizedScale */
-template<typename T = float>
+template<typename T = float, typename TDIV = T>
 QuantizedScale ElementwiseMulScale(double inputScale, double input2Scale, double outputScale)
 {
     // clamp to single-point precision
     T ifm1Scale = ClampToType<T>(inputScale);
     T ifm2Scale = ClampToType<T>(input2Scale);
-    T outScale = ClampToType<T>(outputScale);
+    TDIV outScale = ClampToType<TDIV>(outputScale);
 
-    T outputRescale = (ifm1Scale * ifm2Scale) / outScale;
+    TDIV outputRescale = (ifm1Scale * ifm2Scale) / outScale;
     return QuantizedScale(outputRescale);
 }
 
-- 
GitLab