CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-29 08:41:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: Use coopmat2 for conv2d (#14982)

Browse Source
This commit is contained in:
Jeff Bolz
2025-08-03 07:23:57 -05:00
committed by GitHub
parent 5c0eb5ef54
commit 6c7a441161
3 changed files with 69 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@@ -3096,6 +3096,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
uint32_t conv2d_SHMEM_PAD = 4;
bool conv2d_UNROLL = true;
if (device->coopmat2) {
conv2d_SHMEM_PAD = 8; // 8 float16_t
}
if (device->vendor_id == VK_VENDOR_ID_INTEL) {
conv2d_SHMEM_PAD = 0;
conv2d_UNROLL = false;
@@ -3154,7 +3158,14 @@ static void ggml_vk_load_shaders(vk_device& device) {
std::array<uint32_t, 3> wg_denoms = { conv2d_BS_K, conv2d_BS_NPQ, 1 };
std::vector<uint32_t> spec_constants = { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD };
if (conv2d_UNROLL) {
if (device->coopmat2) {
ggml_vk_create_pipeline(
device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_cm2_len, conv2d_f32_cm2_data, "main", 3,
sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
ggml_vk_create_pipeline(
device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_cm2_len, conv2d_f16_f32_cm2_data, "main", 3,
sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
} else if (conv2d_UNROLL) {
ggml_vk_create_pipeline(
device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_unroll_len, conv2d_f32_unroll_data, "main", 3,
sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);