From 63c9cab6a5dde2890f4e7ab5d150e950e149eb87 Mon Sep 17 00:00:00 2001
From: Jegor van Opdorp <jegorvanopdorp@gmail.com>
Date: Fri, 16 Oct 2026 08:20:19 +0200
Subject: [PATCH] Skip FP16 tensor core tuning on Mali GPUs

Mali's OpenCL compiler crashes the process on the wmma kernels, so autotuning dies on phones like the Pixel 7a.
---
 cpp/neuralnet/opencltuner.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/cpp/neuralnet/opencltuner.cpp b/cpp/neuralnet/opencltuner.cpp
index a13e025d4e..7f5ba0d135 100644
--- a/cpp/neuralnet/opencltuner.cpp
+++ b/cpp/neuralnet/opencltuner.cpp
@@ -4542,7 +4542,8 @@ void OpenCLTuner::tune(
       static constexpr double FP16_TENSORCORE_REQUIRED_SPEEDUP = 0.9;
       bool foundGoodFP16 = false;
 
-      bool shouldTestFP16TensorCores = testFP16TensorCoresMode == enabled_t::True || (testFP16TensorCoresMode == enabled_t::Auto && !foundGoodFP16);
+      //Mali's compiler aborts the whole process on the wmma kernels instead of failing the build
+      bool shouldTestFP16TensorCores = testFP16TensorCoresMode == enabled_t::True || (testFP16TensorCoresMode == enabled_t::Auto && !foundGoodFP16 && device->info.name.find("Mali") == string::npos);
       if(shouldTestFP16TensorCores) {
         {
           OpenCLTuneParams result16;