happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

vulkan: optimize Mali and Adreno tuning & fix FP16/matrix stability 

- Add VK_VENDOR_ID_ARM (0x13B5) and VK_VENDOR_ID_QUALCOMM (0x5143).
    - Disable 'l_warptile' (Large) paths on ARM and Qualcomm to avoid register pressure/driver crashes.
    - Explicitly disable cooperative matrix (fp16-matrix) on mobile GPUs to prevent crashes.
    - Disable integer dot product (int8-matrix) on Qualcomm due to driver issues.
    - Keep general FP16 enabled for large model support.
    - Limit suballocation block size to 256MB to improve memory stability on Android.
This commit is contained in:
Gong-Mi 2026-01-20 21:13:23 +08:00
parent b5b8fa1c8b
commit e0f3a6b292
1 changed files with 18 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -90,8 +90,10 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
#define VK_VENDOR_ID_AMD 0x1002
#define VK_VENDOR_ID_APPLE 0x106b
#define VK_VENDOR_ID_ARM 0x13B5
#define VK_VENDOR_ID_INTEL 0x8086
#define VK_VENDOR_ID_NVIDIA 0x10de
#define VK_VENDOR_ID_QUALCOMM 0x5143
#define VK_DEVICE_DESCRIPTOR_POOL_SIZE 256
@ -3059,6 +3061,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
device->mul_mat_l[i] = false;
}
if (device->vendor_id == VK_VENDOR_ID_ARM || device->vendor_id == VK_VENDOR_ID_QUALCOMM) {
device->mul_mat_l[i] = false;
device->mul_mat_id_l[i] = false;
}
// Disable mul_mat_id if not enough shared memory is available
if (!ggml_vk_matmul_shmem_support(device, s_warptile_mmqid, true, t)) {
device->mul_mat_id_s[i] = false;
@ -4676,6 +4683,17 @@ static vk_device ggml_vk_get_device(size_t idx) {
// Limit batching of allocations to 1GB by default to avoid fragmentation issues
device->suballocation_block_size = 1024*1024*1024;
}
if (device->vendor_id == VK_VENDOR_ID_ARM) {
// ARM Mali optimization: disable fp16-matrix to prevent crashes, limit memory blocks
device->coopmat_support = false;
device->suballocation_block_size = 256 * 1024 * 1024;
} else if (device->vendor_id == VK_VENDOR_ID_QUALCOMM) {
// Qualcomm Adreno optimization: disable fp16-matrix and int8-dotprod to prevent crashes, limit memory blocks
device->coopmat_support = false;
device->integer_dot_product = false;
device->suballocation_block_size = 256 * 1024 * 1024;
}
device->suballocation_block_size = std::min(device->suballocation_block_size, device->max_memory_allocation_size);
device->subgroup_size = subgroup_props.subgroupSize;