cpu: x64: brgemm: prefetchw removed from c buffer

Author: yair-obodovsky

src/cpu/x64/brgemm/jit_brgemm_amx_uker.cpp

@@ -1124,9 +1124,14 @@ void jit_brgemm_amx_uker_base_t::prefetch_CD_range(brgemm_iteration_t &bi,
             auto ptr_D = EVEX_compress_addr(reg_D, d_offset);
             uni_prefetch(ptr_D, pft, true);
         } else if (are_post_ops_applicable_) {
-            const auto c_offset = C_offset(bi, bdb, bd, ldb_pos);
-            auto ptr_C = EVEX_compress_addr(reg_C, c_offset);
-            uni_prefetch(ptr_C, pft, true);
+            //            TODO: split hints C and D hints
+            //              Using prefetchw for the C matrix is generally harmful
+            //              because the C matrix is frequently reused and remains in the cache.
+            //              However, it is very necessary for the D matrix
+
+            //            const auto c_offset = C_offset(bi, bdb, bd, ldb_pos);
+            //            auto ptr_C = EVEX_compress_addr(reg_C, c_offset);
+            //            uni_prefetch(ptr_C, pft, true);
         } else {
             const auto d_offset = D_offset(bi, bdb, bd, ldb_pos);
             auto ptr_D = EVEX_compress_addr(reg_D, d_offset);

src/cpu/x64/matmul/brgemm_matmul_utils.cpp

@@ -676,7 +676,6 @@ struct matmul_avx512_blocking_params_t {
     }
 };
 
-
 float compute_blocking_heuristic_avx512(brgemm_matmul_conf_t &bgmmc,
         const brgemm_matmul_conf_utils_t &bm_conf_utils,
         const matmul_avx512_blocking_params_t::matmul_params_t &matmul,
@@ -1908,10 +1907,6 @@ void init_scratchpad(memory_tracking::registrar_t &scratchpad,
                 default_data_align);
 }
 
-
-
-
-
 } // namespace matmul
 } // namespace x64
 } // namespace cpu