Add run_pipeline option of rechunk.

ggml/src/CMakeLists.txt

@@ -897,6 +897,10 @@ if (GGML_TMAC)
             set(GGML_EXTRA_LIBS_PRIVATE ${GGML_EXTRA_LIBS_PRIVATE} t_mac_no_tvm)
             set(GGML_SOURCES_TMAC ${GGML_SOURCES_TMAC} ${TMAC_KERNELS_SOURCE})
         endif()
+
+        if (GGML_TMAC_RECHUNK)
+            add_compile_definitions(TMAC_RECHUNK)
+        endif()
     else()
         message(WARNING "TMAC not found")
     endif()

ggml/src/ggml-cpu.c

@@ -7577,13 +7577,14 @@ UseGgmlGemm1:;
 
         int64_t dr0 = chunk_size0;
         int64_t dr1 = chunk_size1;
-
+#if defined(TMAC_RECHUNK)
         // Rechunk
         if ((nchunk1 == 1) && (nchunk0 > nth * 4)) {
             // dr0 should be divisible by chunk_size0
             dr0 = (ne0 / (nth * 4) / chunk_size0) * chunk_size0;
             nchunk0 = (nr0 + dr0 - 1) / dr0;
         }
+#endif
 
         int current_chunk = ith;