Skip to content

LLM inference in C/C++

C++ ggml

vulkan: disable large matmul tile on Samsung GPUs with 32KB shared memory (#28531)

Jiwoong Song9e258a6e0afa
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan-types.h b/ggml/src/ggml-vulkan/ggml-vulkan-types.h
index ee50f51a38..4158ff7793 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan-types.h
+++ b/ggml/src/ggml-vulkan/ggml-vulkan-types.h
@@ -209,6 +209,8 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
#define VK_VENDOR_ID_QUALCOMM 0x5143
+#define VK_VENDOR_ID_SAMSUNG 0x144d
+
#define VK_DEVICE_DESCRIPTOR_POOL_SIZE 256
#define VK_CHECK(err, msg, dev) \
@@ -1451,4 +1453,3 @@ struct ggml_backend_vk_device_context {
std::string pci_bus_id;
int op_offload_min_batch_size;
};
-
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index a45d5e1f0e..033c117417 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -4823,6 +4823,14 @@ vk_device ggml_vk_get_device(size_t idx) {
device->mul_mat_id_m[i] = true;
device->mul_mat_id_s[i] = !device->coopmat_support;
break;
+ case VK_VENDOR_ID_SAMSUNG:
+ device->mul_mat_l[i] = false;
+ device->mul_mat_m[i] = true;
+ device->mul_mat_s[i] = true;
+ device->mul_mat_id_l[i] = false;
+ device->mul_mat_id_m[i] = true;
+ device->mul_mat_id_s[i] = true;
+ break;
#endif
default:
device->mul_mat_l[i] = true;