[FORK][FIX] Limit microkernel unrolling factor

dmitry-gorokhov · dmitry-gorokhov · commit 36a1036bb2ad · 2025-02-24T10:45:05.000+01:00
[FORK][FEATURE] DQ IP: performance enhansments
diff --git a/src/cpu/x64/brgemm/jit_brgemm_kernel.cpp b/src/cpu/x64/brgemm/jit_brgemm_kernel.cpp
@@ -3130,7 +3130,8 @@ void jit_brgemm_kernel_t<Wmm>::ldb_loop(int bd_block2, bool is_bdb_tail,
             }
             rdb_group = rdb_group / brg.rd_block;
             auto rbd_blocks = brg.rdb / rdb_group;
-            auto max_rdb_unroll = 8;
+            // unrolling is limited to prevent exceeding L1i cache
+            auto max_rdb_unroll = brg.with_src_dyn_quant ? 8 : 2;
 
             if (brg.with_wei_decomp && rdb_group <= max_rdb_unroll) {
                 if (rbd_blocks > 0) {

Original file line number	Diff line number	Diff line change
`@@ -3130,7 +3130,8 @@ void jit_brgemm_kernel_t<Wmm>::ldb_loop(int bd_block2, bool is_bdb_tail,`
`3130`	`3130`	`}`
`3131`	`3131`	`rdb_group = rdb_group / brg.rd_block;`
`3132`	`3132`	`auto rbd_blocks = brg.rdb / rdb_group;`
`3133`		`- auto max_rdb_unroll = 8;`
	`3133`	`+ // unrolling is limited to prevent exceeding L1i cache`
	`3134`	`+ auto max_rdb_unroll = brg.with_src_dyn_quant ? 8 : 2;`
`3134`	`3135`
`3135`	`3136`	`if (brg.with_wei_decomp && rdb_group <= max_rdb_unroll) {`
`3136`	`3137`	`if (rbd_blocks > 0) {`