Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-19 17:24:57 +02:00
Code Issues Packages Projects Releases Wiki Activity

fix compilation

Browse Source
This commit is contained in:
Ruben Ortlam
2026-08-24 14:50:46 +02:00
parent 8f2c87e00b
commit 43d831641c
1 changed files with 6 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+6 -5
View File
@@ -4194,6 +4194,12 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
uint32_t l_align, m_align, s_align;
// RDNA3.5 preferred wave32 here
const bool cm1_use_wave32 = device->vendor_id == VK_VENDOR_ID_AMD &&
device->subgroup_size_control &&
device->subgroup_min_size <= 32 && device->subgroup_max_size >= 32;
const uint32_t cm1_sg = cm1_use_wave32 ? 32 : device->subgroup_size;
vk_pipeline wait_pipeline;
CompileTask claimed_task {};
bool has_claimed_task = false;
@@ -4276,11 +4282,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
m_warptile_mmq_int = { 128, 64, 64, 32, subgroup_size_8, 32, 2, itm_m, itn_m, itk_m, subgroup_size_8 };
s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, itm_s, itn_s, itk_s, subgroup_size_8 };
// RDNA3.5 preferred wave32 here
const bool cm1_use_wave32 = device->vendor_id == VK_VENDOR_ID_AMD &&
device->subgroup_size_control &&
device->subgroup_min_size <= 32 && device->subgroup_max_size >= 32;
const uint32_t cm1_sg = cm1_use_wave32 ? 32 : device->subgroup_size;
const auto cm1_bs = [cm1_sg](uint32_t bm, uint32_t bn) {
return cm1_sg * (bm / std::min(cm1_sg, bm)) * (bn / 32);
};