Skip to content

vulkan: clamp l_/m_ warptile WM to <= BM (fix wrong matmul on subgroupSize > 64) #25735

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
lkazarin wants to merge 1 commit into from
Closed

vulkan: clamp l_/m_ warptile WM to <= BM (fix wrong matmul on subgroupSize > 64) #25735

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
48 changes: 30 additions & 18 deletions ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3877,11 +3877,23 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {

// some shaders have a minimum subgroup size
const uint32_t subgroup_size_8 = std::max(device->subgroup_size, 8u);
// The l_/m_ warptiles encode an implicit invariant WM <= BM:
// l_warptile: BM=128, WM=warp*2 ; m_warptile: BM=64, WM=warp
// Adreno/turnip reports subgroupSize=128 (double threadsize), making both
// degenerate (WM>BM): BM/WM folds to 0 (div-by-zero in mul_mm.comp), buf_a
// indexing runs past its end into buf_b, and half the output columns are
// never written -> silently wrong results. Clamp the tiling WARP for those
// two tiles to 64 (keeps NUM_WARPS*WM*WN == BM*BN and WM<=BM).
// NOTE: s_warptile must keep the unclamped value -- its BLOCK_SIZE is
// subgroup_size_32, and it needs NUM_WARPS==1 to keep its area at BM*BN.
const uint32_t mm_warp_8 = std::min(subgroup_size_8, 64u);
const uint32_t subgroup_size_16 = std::max(device->subgroup_size, 16u);
const uint32_t subgroup_size_32 = std::max(device->subgroup_size, 32u);

const uint32_t mul_mat_subgroup_size = (device->vendor_id == VK_VENDOR_ID_INTEL && device->subgroup_size_control) ? device->subgroup_min_size : device->subgroup_size;
const uint32_t mul_mat_subgroup_size_8 = std::max(mul_mat_subgroup_size, 8u);
// same WM<=BM clamp as mm_warp_8, for the mmqid l_/m_ warptiles
const uint32_t mul_mat_mm_warp_8 = std::min(mul_mat_subgroup_size_8, 64u);
const uint32_t mul_mat_subgroup_size_16 = std::max(mul_mat_subgroup_size, 16u);
const uint32_t mul_mat_subgroup_size_32 = std::max(mul_mat_subgroup_size, 32u);

Expand Down Expand Up @@ -3964,34 +3976,34 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {

const uint32_t s_warptile_wm = device->subgroup_size == 8 ? 8 : 32;

l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 };
m_warptile = { 128, 64, 64, 16, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
l_warptile = { 128, 128, 128, 16, mm_warp_8 * 2, 64, 2, tm_l, tn_l, tk_l, mm_warp_8 };
m_warptile = { 128, 64, 64, 16, mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mm_warp_8 };
s_warptile = { subgroup_size_32, 32, 32, 16, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 };

l_warptile_mmq = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 };
m_warptile_mmq = { 128, 64, 64, 32, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
l_warptile_mmq = { 128, 128, 128, 32, mm_warp_8 * 2, 64, 2, tm_l, tn_l, tk_l, mm_warp_8 };
m_warptile_mmq = { 128, 64, 64, 32, mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mm_warp_8 };
s_warptile_mmq = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 };

// Integer MMQ has a smaller shared memory profile, but heavier register use
l_warptile_mmq_int = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 };
m_warptile_mmq_int = { 128, 64, 64, 32, subgroup_size_8, 32, 2, 2, 2, 1, subgroup_size_8 };
l_warptile_mmq_int = { 128, 128, 128, 32, mm_warp_8 * 2, 64, 2, 4, 4, 1, mm_warp_8 };
m_warptile_mmq_int = { 128, 64, 64, 32, mm_warp_8, 32, 2, 2, 2, 1, mm_warp_8 };
s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, 2, 1, 1, subgroup_size_8 };

// K-quants use even more registers, mitigate by setting WMITER to 1
l_warptile_mmq_int_k = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 1, 4, 4, 1, subgroup_size_8 };
m_warptile_mmq_int_k = { 128, 64, 64, 32, subgroup_size_8, 32, 1, 2, 2, 1, subgroup_size_8 };
l_warptile_mmq_int_k = { 128, 128, 128, 32, mm_warp_8 * 2, 64, 1, 4, 4, 1, mm_warp_8 };
m_warptile_mmq_int_k = { 128, 64, 64, 32, mm_warp_8, 32, 1, 2, 2, 1, mm_warp_8 };
s_warptile_mmq_int_k = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 1, 2, 1, 1, subgroup_size_8 };

l_warptile_id = { 128, 128, 128, 16, mul_mat_subgroup_size_16 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_subgroup_size_16 };
m_warptile_id = { 128, 64, 64, 16, mul_mat_subgroup_size_16, 32, 2, tm_m, tn_m, tk_m, mul_mat_subgroup_size_16 };
s_warptile_id = { mul_mat_subgroup_size_16, 32, 32, 16, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_16 };

l_warptile_mmqid = { 128, 128, 128, 32, mul_mat_subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_subgroup_size_8 };
m_warptile_mmqid = { 128, 64, 64, 32, mul_mat_subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, mul_mat_subgroup_size_8 };
l_warptile_mmqid = { 128, 128, 128, 32, mul_mat_mm_warp_8 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_mm_warp_8 };
m_warptile_mmqid = { 128, 64, 64, 32, mul_mat_mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mul_mat_mm_warp_8 };
s_warptile_mmqid = { mul_mat_subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, mul_mat_subgroup_size_8 };

l_warptile_mmqid_int = { 128, 128, 128, 32, mul_mat_subgroup_size_8 * 2, 64, 2, 4, 4, 1, mul_mat_subgroup_size_8 };
m_warptile_mmqid_int = { 128, 64, 64, 32, mul_mat_subgroup_size_8, 32, 2, 2, 2, 1, mul_mat_subgroup_size_8 };
l_warptile_mmqid_int = { 128, 128, 128, 32, mul_mat_mm_warp_8 * 2, 64, 2, 4, 4, 1, mul_mat_mm_warp_8 };
m_warptile_mmqid_int = { 128, 64, 64, 32, mul_mat_mm_warp_8, 32, 2, 2, 2, 1, mul_mat_mm_warp_8 };
s_warptile_mmqid_int = { mul_mat_subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, 2, 1, 1, mul_mat_subgroup_size_8 };

l_warptile_mmqid_int_k = { 128, 128, 128, 32, mul_mat_subgroup_size_16 * 2, 64, 1, 4, 4, 1, mul_mat_subgroup_size_16 };
Expand All @@ -4004,13 +4016,13 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
m_warptile_mmqid = m_warptile_mmqid_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
} else if (device->vendor_id == VK_VENDOR_ID_AMD && device->coopmat_support && device->driver_id != vk::DriverId::eAmdProprietary) {
// This is intentionally using tx_m values, slight performance increase
l_warptile = { 256, 128, 128, 16, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
l_warptile_mmq = l_warptile_mmq_int = { 256, 128, 128, 32, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
l_warptile = { 256, 128, 128, 16, mm_warp_8, 64, 2, tm_m, tn_m, tk_m, mm_warp_8 };
l_warptile_mmq = l_warptile_mmq_int = { 256, 128, 128, 32, mm_warp_8, 64, 2, tm_m, tn_m, tk_m, mm_warp_8 };
l_warptile_mmq_int_k = { 256, 128, 128, 32, subgroup_size_16, 64, 1, 4, 2, 1, subgroup_size_16 };
} else if (device->vendor_id == VK_VENDOR_ID_INTEL && device->coopmat_support) {
// 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 };
l_warptile = { 512, 128, 128, 16, mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mm_warp_8 };
l_warptile_mmq = { 512, 128, 128, 32, mm_warp_8, 32, 2, tm_m, tn_m, tk_m, mm_warp_8 };
}

l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
Expand Down Expand Up @@ -4854,8 +4866,8 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
const uint32_t s_warptile_wm = device->subgroup_size == 8 ? 8 : 32;

// use scalar tile sizes
l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 };
m_warptile = { 128, 64, 64, 16, subgroup_size_8, 32, 2, 4, 2, 1, subgroup_size_8 };
l_warptile = { 128, 128, 128, 16, mm_warp_8 * 2, 64, 2, 4, 4, 1, mm_warp_8 };
m_warptile = { 128, 64, 64, 16, mm_warp_8, 32, 2, 4, 2, 1, mm_warp_8 };
s_warptile = { subgroup_size_32, 32, 32, 16, s_warptile_wm, 32, 2, 2, 2, 1, subgroup_size_8 };

l_wg_denoms = {128, 128, 1 };
Expand Down