From: lksj92hs <redacted>
Date: Tue, 9 Sep 2025 12:01:15 +0000 (+0300)
Subject: Workaround for subgroup arithmetic failing on MoltenVK with AMD GPUs (issue 15846... 
X-Git-Tag: upstream/0.0.6527~98
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=ed54e32558ffe0f2c3b1d31f3227211fae05bd49;p=pkg%2Fggml%2Fsources%2Fllama.cpp

Workaround for subgroup arithmetic failing on MoltenVK with AMD GPUs (issue 15846) (#15886)
---

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index f0fa9e66..6f130d47 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -3736,6 +3736,12 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->subgroup_arithmetic = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
                                       (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic);
+#ifdef __APPLE__
+        // Workaround for subgroup arithmetic failing on MoltenVK with AMD GPUs (issue 15846)
+        if (device->vendor_id == VK_VENDOR_ID_AMD) {
+            device->subgroup_arithmetic = false;
+        }
+#endif
         device->subgroup_shuffle = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
                                    (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eShuffle);
         device->subgroup_clustered = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&