LLM inference in C/C++
C++ ggml
vulkan: warptiles currently assume warp sizes <= 64, clamp to work around larger warps (#27726)
Ruben Ortlam5e6a37cb115d
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cppindex 31923a2ba0..8108e94c16 100644--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp@@ -4171,10 +4171,16 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {const uint32_t subgroup_size_16 = std::max(device->subgroup_size, 16u);const uint32_t subgroup_size_32 = std::max(device->subgroup_size, 32u);+ // clamp WARP for l_/m_ warptiles so WM <= BM (breaks on subgroupSize > 64)+ const uint32_t mm_warp_8 = std::min(subgroup_size_8, 64u);+ const uint32_t mm_warp_16 = std::min(subgroup_size_16, 64u);+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);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);+ const uint32_t mul_mat_mm_warp_8 = std::min(mul_mat_subgroup_size_8, 64u);+ const uint32_t mul_mat_mm_warp_16 = std::min(mul_mat_subgroup_size_16, 64u);const bool subgroup_min_size_16 = (!device->subgroup_size_control && device->subgroup_size >= 16) ||(device->subgroup_size_control && device->subgroup_max_size >= 16);@@ -4255,39 +4261,39 @@ 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 };- s_warptile = { subgroup_size_32, 32, 32, 16, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, 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 };- s_warptile_mmq = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, tm_s, tn_s, tk_s, 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 };- s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 2, 2, 1, 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 };- s_warptile_mmq_int_k = { subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 1, 2, 1, 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_id = { 128, 128, 128, 16, mul_mat_mm_warp_16 * 2, 64, 2, tm_l, tn_l, tk_l, mul_mat_mm_warp_16 };+ m_warptile_id = { 128, 64, 64, 16, mul_mat_mm_warp_16, 32, 2, tm_m, tn_m, tk_m, mul_mat_mm_warp_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 };- 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 = { 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 };- 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 = { 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 };- m_warptile_mmqid_int_k = { 128, 64, 64, 32, mul_mat_subgroup_size_16, 32, 1, 2, 2, 1, mul_mat_subgroup_size_16 };- s_warptile_mmqid_int_k = { mul_mat_subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 1, 2, 1, 1, mul_mat_subgroup_size_16 };+ l_warptile_mmqid_int_k = { 128, 128, 128, 32, mul_mat_mm_warp_16 * 2, 64, 1, 4, 4, 1, mul_mat_mm_warp_16 };+ m_warptile_mmqid_int_k = { 128, 64, 64, 32, mul_mat_mm_warp_16, 32, 1, 2, 2, 1, mul_mat_mm_warp_16 };+ s_warptile_mmqid_int_k = { mul_mat_subgroup_size_32, 32, 32, 32, s_warptile_wm, 32, 1, 2, 1, 1, mul_mat_subgroup_size_16 };// chip specific tuningif ((device->architecture == AMD_GCN) && (device->driver_id != vk::DriverId::eAmdProprietary)) {@@ -4295,13 +4301,13 @@ 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_mmq_int_k = { 256, 128, 128, 32, subgroup_size_16, 64, 1, 4, 2, 1, subgroup_size_16 };+ 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, mm_warp_16, 64, 1, 4, 2, 1, mm_warp_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 };@@ -5174,8 +5180,8 @@ 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 };