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

ggml-vulkan: optimize Mali-G720/Adreno tuning & fix stability 

- Implement 4x4 warptile tuning for Mali-G720/Immortalis MC12.
- Optimize tuning parameters for ARM Mali and Qualcomm Adreno.
- Fix matrix multiplication out-of-bounds (OOB) access by moving restrictions to initialization.
- Ensure stability by removing risky subgroup size clamping on Qualcomm devices.
This commit is contained in:
Gong-Mi 2026-02-02 13:03:47 +08:00
parent b5b8fa1c8b
commit 9b86eb6207
1 changed files with 48 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

48
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
@ -3035,6 +3037,25 @@ static void ggml_vk_load_shaders(vk_device& device) {
// Xe2/Xe3 with coopmat enabled - warptile performance tuning
l_warptile = { 512, 128, 128, 16, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
l_warptile_mmq = { 512, 128, 128, 32, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
} else if (device->vendor_id == VK_VENDOR_ID_QUALCOMM) {
uint32_t s = device->subgroup_size;
m_warptile = m_warptile_mmq = m_warptile_mmq_int = m_warptile_mmq_k = m_warptile_mmq_int_k =
m_warptile_id = m_warptile_mmqid = m_warptile_mmqid_int = m_warptile_mmqid_int_k =
{ 64, 64, 64, s, s, 16, 8, 8, 1, 1, s };
s_warptile = s_warptile_mmq = s_warptile_mmq_int = s_warptile_mmq_k = s_warptile_mmq_int_k =
s_warptile_id = s_warptile_mmqid = s_warptile_mmqid_int = s_warptile_mmqid_int_k =
{ 32, 32, 64, s, s, 16, 8, 8, 1, 1, s };
} else if (device->vendor_id == VK_VENDOR_ID_ARM && device->subgroup_size >= 16) {
uint32_t s = device->subgroup_size;
s_warptile = s_warptile_mmq = s_warptile_mmq_int = s_warptile_mmq_k = s_warptile_mmq_int_k =
s_warptile_id = s_warptile_mmqid = s_warptile_mmqid_int = s_warptile_mmqid_int_k =
{ 32, 32, 64, s, s, 16, 4, 4, 1, 1, s };
m_warptile = m_warptile_mmq = m_warptile_mmq_int = m_warptile_mmq_k = m_warptile_mmq_int_k =
m_warptile_id = m_warptile_mmqid = m_warptile_mmqid_int = m_warptile_mmqid_int_k =
{ 64, 64, 64, s, s, 16, 4, 4, 1, 1, s };
l_warptile = l_warptile_mmq = l_warptile_mmq_int = l_warptile_mmq_k = l_warptile_mmq_int_k =
l_warptile_id = l_warptile_mmqid = l_warptile_mmqid_int = l_warptile_mmqid_int_k =
{ 64, 64, 64, s, s, 16, 4, 4, 1, 1, s };
}
l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
@ -3044,6 +3065,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
m_align = 64;
s_align = 32;
if (device->vendor_id == VK_VENDOR_ID_QUALCOMM) {
m_align = 128;
}
for (uint32_t i = 0; i < GGML_TYPE_COUNT; ++i) {
ggml_type t = (ggml_type)i;
// Disable medium and large matrix multiplication if not enough shared memory is available
@ -4676,6 +4701,17 @@ 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;
@ -5154,6 +5190,15 @@ static vk_device ggml_vk_get_device(size_t idx) {
device->mul_mat_id_m[i] = true;
device->mul_mat_id_s[i] = true;
break;
case VK_VENDOR_ID_ARM:
case VK_VENDOR_ID_QUALCOMM:
device->mul_mat_l[i] = false;
device->mul_mat_id_l[i] = false;
device->mul_mat_m[i] = true;
device->mul_mat_s[i] = true;
device->mul_mat_id_m[i] = true;
device->mul_mat_id_s[i] = true;
break;
case VK_VENDOR_ID_APPLE:
device->mul_mat_l[i] = false;
device->mul_mat_m[i] = true;
@ -13728,6 +13773,9 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
// (and scaled down based on model size, so smaller models submit earlier).
// Also submit at least every 100 nodes, in case there are workloads without as much matmul.
int nodes_per_submit = 100;
if (ctx->device->vendor_id == VK_VENDOR_ID_QUALCOMM) {
nodes_per_submit = 1000;
}
int submitted_nodes = 0;
int submit_count = 0;
uint64_t mul_mat_bytes = 0;