[Snippets][CPU] Moved N_tail processing to the end in BrgemmCopyBKernel (#28664)

### Details:
- *The performance experiments (see the mentioned ticket please) show
that `N_Tail` processing should be at the end of `BrgemmCopyBKernel`.
The current PR moves tail processing from the beginning to the end in
kernel*

### Tickets:
 - *CVS-161315*

Author: a-sidorova

src/plugins/intel_cpu/src/emitters/snippets/x64/kernel_executors/brgemm_base.cpp

@@ -258,7 +258,7 @@ void BrgemmBaseKernelExecutor::update_config(const ov::snippets::lowered::Expres
     OV_CPU_JIT_EMITTER_ASSERT(brgemm_node, "Got invalid node type in update_config");
     // In case of data repacking LDB is chosen in accordance with repacking buffer size
     if (with_repacking(brgemm_node->get_type())) {
-        LDB = DIM_CAST(brgemm_utils::repacking::compute_LDB(LDB, brgemm_node->get_input_element_type(1)));
+        LDB = DIM_CAST(brgemm_utils::repacking::compute_repacked_n_dim(LDB, brgemm_node->get_input_element_type(1)));
     }
 
     config.update(DIM_CAST(M), DIM_CAST(N), DIM_CAST(K), LDA, LDB, LDC, beta);

src/plugins/intel_cpu/src/emitters/snippets/x64/kernel_executors/brgemm_copy_b.cpp

@@ -227,21 +227,13 @@ void BrgemmCopyBKernel::generate() {
     size_t start_out = 0;
     size_t start_comp = 0;
 
-    auto add_ptr_increments = [&](size_t current_N) {
+    for (size_t nb = 0; nb < div_up(N_blk, wei_N_blk); nb++) {
+        const auto current_N = N_blk - nb * wei_N_blk < wei_N_blk ? wei_N_tail : wei_N_blk;
+        emit_brgemm_copy_b_kernel_call(current_N, K, start_in, start_out, start_comp);
+
         start_in += is_transpose ? K * current_N * wei_data_size : current_N * wei_data_size;
         start_out += current_N * vnni_factor * wei_data_size;
         start_comp += is_with_comp ? current_N * sizeof(int32_t) : 0;
-    };
-
-    // OneDNN requires tail handling before main iterations
-    if (wei_N_tail != 0) {
-        emit_brgemm_copy_b_kernel_call(wei_N_tail, K, start_in, start_out, start_comp);
-        add_ptr_increments(wei_N_tail);
-    }
-
-    for (auto nb = wei_N_tail; nb < N_blk; nb += wei_N_blk) {
-        emit_brgemm_copy_b_kernel_call(wei_N_blk, K, start_in, start_out, start_comp);
-        add_ptr_increments(wei_N_blk);
     }
 
     postamble();
@@ -389,7 +381,7 @@ void BrgemmCopyBKernelExecutor::update_config(const ov::snippets::lowered::Expre
     init(N_dim, N_blk, 0);
 
     const auto& brg_weight_etype = expr->get_node()->get_input_element_type(0);
-    const auto LDB = brgemm_utils::repacking::compute_LDB(N_dim, brg_weight_etype);
+    const auto LDB = brgemm_utils::repacking::compute_repacked_n_dim(N_dim, brg_weight_etype);
     const auto copy_B_wei_stride =
         ov::snippets::utils::get_dim_stride(expr->get_input_port(0), config.is_transposed_B() ? 0 : 1) *
         brg_weight_etype.size();

src/plugins/intel_cpu/src/transformations/snippets/x64/op/brgemm_utils.hpp

@@ -60,18 +60,13 @@ namespace repacking {
 size_t compute_inner_n_block(const ov::element::Type& precision);
 /// \brief  Computes inner K block size used by OneDNN implementation. Depends on tensor precision
 size_t compute_inner_k_block(const ov::element::Type& precision);
-/**
- * @brief Computes leading dimension (LDB) which must be used in brgemm and brgemm_copy_b emitters
- * @param n_block N block size shared between BrgemmCPU and BrgemmCopyB node
- * @param precision tensor precision
- */
+
+/// \brief  Computes N dim in output blocked shape of BrgemmCopyB. Depends on tensor precision
 template <
     typename T,
     typename = typename std::enable_if<(std::is_same<T, size_t>::value || std::is_same<T, int64_t>::value), bool>::type>
-T compute_LDB(T n_block, const ov::element::Type& precision) {
-    return snippets::utils::is_dynamic_value<T>(n_block)
-               ? n_block
-               : std::max(n_block, static_cast<T>(compute_inner_n_block(precision)));
+inline T compute_repacked_n_dim(T n, const ov::element::Type& precision) {
+    return ov::snippets::utils::rnd_up(n, static_cast<T>(compute_inner_n_block(precision)));
 }
 /**
  * @brief Retrieves the expression pointer for the brgemm_copy_b expression corresponding to the given BrgemmCPU

src/plugins/intel_cpu/src/transformations/snippets/x64/pass/lowered/adjust_brgemm_copy_b_loop_ports.cpp

@@ -35,7 +35,8 @@ bool pass::AdjustBrgemmCopyBLoopPorts::update_loop_info(
                     // K blocking loop: account for zero padding
                     if (loop_port.get_dim_idx() == 1) {
                         const auto ptr_incr = loop_desc.ptr_increment;
-                        const auto blocked_shape_ptr_inc = brgemm_utils::repacking::compute_LDB(ptr_incr, precision);
+                        const auto blocked_shape_ptr_inc =
+                            brgemm_utils::repacking::compute_repacked_n_dim(ptr_incr, precision);
                         if (ptr_incr != 0 && ptr_incr != blocked_shape_ptr_inc) {
                             loop_desc.ptr_increment = blocked_shape_ptr_inc;
                             OPENVINO_ASSERT(loop_desc.finalization_offset % ptr_incr == 0,

src/plugins/intel_cpu/src/transformations/snippets/x64/pass/lowered/expressions/brgemm_copy_b_buffer_expressions.cpp

@@ -45,7 +45,7 @@ void RepackedWeightsBufferExpression::init_allocation_size(
 
     const auto& precision = get_node()->get_input_element_type(0);
     // Repacking buffer shape is set in accordance to OneDNN requirements
-    const size_t N_dim = std::max(n_blk, compute_inner_n_block(precision));
+    const size_t N_dim = compute_repacked_n_dim(n_blk, precision);
     if (!in_layout.empty() && in_layout.back() != in_layout.size() - 1) {
         // In case of transpose, K dimension must be rounded-up to number of elems in vector register
         // For the details, please see 'transpose16x8' and 'fixup16x16' implementations and usage in
@@ -88,13 +88,9 @@ void CompensationsBufferExpression::init_allocation_size(
     // Compensations are computed during repacking, so we need to round-up allocation shape according to m_inner_n_block
     // because of OneDNN implementation nuances (as in get_repacking_buffer_size).
     // However, the compensations are computed by N dimension, so K dimension doesn't affect the compensations buffer
+    const auto& precision = parent_expr->get_node()->get_input_element_type(0);
     const size_t n_blk = *ov::snippets::utils::get_projected_subtensor(parent_expr->get_input_port(0)).rbegin();
-    if (snippets::utils::is_dynamic_value(n_blk)) {
-        m_allocation_size = snippets::utils::get_dynamic_value<size_t>();
-    } else {
-        const auto& precision = parent_expr->get_node()->get_input_element_type(0);
-        m_allocation_size = std::max(n_blk, compute_inner_n_block(precision));
-    }
+    m_allocation_size = compute_repacked_n_dim(n_blk, precision);
 }
 
 }  // namespace intel_cpu

src/plugins/intel_cpu/src/transformations/snippets/x64/pass/lowered/external_repacking_adjuster.cpp

@@ -57,7 +57,7 @@ VectorDims BrgemmExternalRepackingAdjuster::get_blk_shape(const VectorDims& plan
     const auto K = *++planar_shape.rbegin();
     const auto N = *planar_shape.rbegin();
     const auto new_K = snippets::utils::div_up(K, vnni_factor);
-    const auto new_N = std::max(N, brgemm_utils::repacking::compute_inner_n_block(prc));
+    const auto new_N = brgemm_utils::repacking::compute_repacked_n_dim(N, prc);
     VectorDims blk_shape(planar_shape.begin(), planar_shape.end() - brgemm_kernel_rank);
     blk_shape.insert(blk_shape.end(), {new_K, new_N, vnni_factor});
     return blk_shape;
@@ -73,7 +73,8 @@ void BrgemmExternalRepackingAdjuster::update_kernel(const RepackExecutorPtr& exe
     auto config = static_cast<BrgemmCopyBKernelConfig*>(generic_config.get());
     const auto idx = config->is_transposed_B() ? 0 : 1;
     const auto copy_wei_stride = ov::snippets::utils::get_dim_in_stride(shape, layout, idx) * prc.size();
-    config->update(N, N, K, K, copy_wei_stride, brgemm_utils::repacking::compute_LDB(N, prc));
+    const auto LDB = brgemm_utils::repacking::compute_repacked_n_dim(N, prc);
+    config->update(N, N, K, K, copy_wei_stride, LDB);
     executor->update_by_config(*config);
 }