CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-30 08:42:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: matmul gcn tuning (#13016)

Browse Source 
* tune matmul for gcn

* this one is more power efficient

* Update ggml/src/ggml-vulkan/ggml-vulkan.cpp

Co-authored-by: 0cc4m <picard12@live.de>

* disable this tune for the proprietary driver

---------

Co-authored-by: 0cc4m <picard12@live.de>
This commit is contained in:
Eve
2025-04-24 07:18:33 +00:00
committed by GitHub
parent 5630406959
commit b3b6d862cf
1 changed files with 7 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@@ -246,6 +246,7 @@ struct vk_device_struct {
bool pipeline_robustness; bool pipeline_robustness;
vk::Device device; vk::Device device;
uint32_t vendor_id; uint32_t vendor_id;
vk::DriverId driver_id;
vk_device_architecture architecture; vk_device_architecture architecture;
vk_queue compute_queue; vk_queue compute_queue;
vk_queue transfer_queue; vk_queue transfer_queue;
@@ -1740,6 +1741,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
m_warptile_mmq_int = { 128, 64, 64, 32, subgroup_size_8, 32, 2, 2, 2, 1, subgroup_size_8 }; m_warptile_mmq_int = { 128, 64, 64, 32, subgroup_size_8, 32, 2, 2, 2, 1, subgroup_size_8 };
s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, 32, 32, 2, 2, 1, 1, subgroup_size_8 }; s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, 32, 32, 2, 2, 1, 1, subgroup_size_8 };
// chip specific tuning
if ((device->architecture == AMD_GCN) && (device->driver_id != vk::DriverId::eAmdProprietary)) {
m_warptile_mmq = m_warptile_mmq_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
}
l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 }; l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
m_mmq_wg_denoms = m_wg_denoms = { 64, 64, 1 }; m_mmq_wg_denoms = m_wg_denoms = { 64, 64, 1 };
s_mmq_wg_denoms = s_wg_denoms = { 32, 32, 1 }; s_mmq_wg_denoms = s_wg_denoms = { 32, 32, 1 };
@@ -2658,6 +2664,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
device->physical_device.getProperties2(&props2); device->physical_device.getProperties2(&props2);
device->properties = props2.properties; device->properties = props2.properties;
device->vendor_id = device->properties.vendorID; device->vendor_id = device->properties.vendorID;
device->driver_id = driver_props.driverID;
const char* GGML_VK_FORCE_MAX_ALLOCATION_SIZE = getenv("GGML_VK_FORCE_MAX_ALLOCATION_SIZE"); const char* GGML_VK_FORCE_MAX_ALLOCATION_SIZE = getenv("GGML_VK_FORCE_MAX_ALLOCATION_SIZE");