Add TODO for lookup table

Author: am17an

ggml/src/ggml-cpu/arch/x86/repack.cpp

@@ -764,6 +764,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                         std::is_same_v<block_tx8, block_iq4_nlx8>) {
                     col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
                 } else if constexpr (std::is_same_v<block_tx8, block_mxfp4x8>) {
+                    //TODO: simd-ify
                     col_scale_f32 = _mm512_set_ps(
                         GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[7]),
                         GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[6]),
@@ -974,6 +975,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                         std::is_same_v<block_tx8, block_iq4_nlx8>) {
                     col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
                 } else if constexpr (std::is_same_v<block_tx8, block_mxfp4x8>) {
+                    //TODO: simd-ify
                     col_scale_f32 = _mm512_set_ps(
                         GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[7]),
                         GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[6]),

ggml/src/ggml-cpu/ggml-cpu.c

@@ -2974,7 +2974,25 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         }
 
         if (node_n + 1 < cgraph->n_nodes) {
-            ggml_barrier(state->threadpool);
+            // Skip the barrier when both the current and next computable node have n_tasks == 1,
+            // since only thread 0 does work and no inter-thread synchronization is needed.
+            bool need_barrier = (ggml_get_n_tasks(node, params.nth) > 1);
+            if (!need_barrier) {
+                for (int j = node_n + 1; j < cgraph->n_nodes; j++) {
+                    struct ggml_tensor * next = cgraph->nodes[j];
+                    if (ggml_op_is_empty(next->op)) {
+                        continue;
+                    }
+                    if ((next->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+                        continue;
+                    }
+                    need_barrier = (ggml_get_n_tasks(next, params.nth) > 1);
+                    break;
+                }
+            }
+            if (need_barrier) {
+                ggml_barrier(state->threadpool);
+            }
         }
     }