Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

vulkan: add fwht support for Intel with shmem reduction (#23964)

Browse Source 
* vulkan: add fwht support for Intel with shmem reduction

* don't use N as workgroup size

* disable subgroup shuffle on MoltenVK AMD

* disable fwht shader on Intel Windows due to driver bug
This commit is contained in:
Ruben Ortlam
2026-06-05 19:44:40 +02:00
committed by GitHub
parent c4a278d68e
commit e82beaa60d
3 changed files with 76 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+13
View File
@@ -5084,6 +5084,14 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
}
++idx;
}
} else if (device->driver_id != vk::DriverId::eIntelProprietaryWindows) {
// Disabled on Intel Windows due to a driver bug: https://github.com/ggml-org/llama.cpp/pull/23964#issuecomment-4598226147
int idx = 0;
for (uint32_t n : {64, 128, 256, 512}) {
const uint32_t block_size = std::min(device->subgroup_size, n);
ggml_vk_create_pipeline(device, device->pipeline_fwht_f32[idx], "fwht_shmem_f32", fwht_shmem_f32_len, fwht_shmem_f32_data, "main", 2, sizeof(vk_op_fwht_push_constants), {1, 1, 1}, { block_size, n }, 1);
++idx;
}
}
const uint32_t cumsum_elem_per_thread = (device->vendor_id == VK_VENDOR_ID_AMD || device->vendor_id == VK_VENDOR_ID_INTEL) ? 2 : 4;
@@ -5630,6 +5638,11 @@ static vk_device ggml_vk_get_device(size_t idx) {
#endif
device->subgroup_shuffle = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eShuffle);
#ifdef __APPLE__
if (device->vendor_id == VK_VENDOR_ID_AMD) {
device->subgroup_shuffle = false;
}
#endif
device->subgroup_clustered = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
(vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eClustered);