[FORK][FIX] Restore choose_loop_order logic for weights decompression case

Author: dmitrygo

src/cpu/x64/jit_brgemm_inner_product_utils.cpp

@@ -18,6 +18,8 @@
 
 #include "cpu/x64/brgemm/brgemm.hpp"
 
+#include <iostream>
+
 namespace dnnl {
 namespace impl {
 namespace cpu {
@@ -1885,7 +1887,8 @@ void jit_brgemm_ip_fwd_conf_t::choose_loop_order() {
     const bool is_int8 = one_of(src_dt, u8, s8) && wei_dt == s8;
     const bool is_compute_amx = (is_xf16 || is_int8) && is_amx;
 
-    if ((os_block < 32 || do_occ_osc) && (is_compute_amx || is_f32_avx2))
+    // Better to keep ocb loop outermost for weights_decompression case due to overhead on weights unpack into intermediate buffer
+    if ((os_block < 32 || do_occ_osc) && (is_compute_amx || is_f32_avx2) && !weights_decompression)
         loop_order = icc_occ_osc_ocb_osb;
 }