[Optimization] Zen 4 / EPYC Macro-Blocking (P, Q, R) tuning utilizing COOPERLAKE kernels #5837

Activity

1. martin-frbg commented on Jun 15, 2026

   

   martin-frbg

   on Jun 15, 2026

   Collaborator

   More actions

   Sounds right, but would probably require cloning the COOPERLAKE target to create a dedicated ZEN4. I guess you could try if
   adding a conditional inside the COOPERLAKE section of param.h would be sufficient, something like

   if L3_SIZE == 33554432
   #define SGEMM_DEFAULT_P = <your p for zen4>
   ...
   else
   #define SGEMM_DEFAULT_P = 640
   ...
   endif
   ```
   and similarly set `TABLE_NAME.sgemm_p` etc. depending on the value of `l3` in the COOPERLAKE section of `kernel/setparam-ref.c` for the DYNAMIC_ARCH case. (I do wonder how many different L3 cache configurations we'd have to cater for across the "ZEN4" processor family ?)
   
2. Vasudeva-bit commented on Jun 15, 2026

   

   Vasudeva-bit

   on Jun 15, 2026

   ContributorAuthor

   More actions

   I shall try the conditional workaround to see if it's sufficient. Based on my search and Zen4 Wikipedia table, the majority seem to use the standard 32 MB per CCD, with two other variants: 96MB (per CCD with 3D V-Cache) and 16/8 MB (APUs).

   Also, I think I could scale the new values of P, Q, R accordingly for DGEMM, BFGEMM, etc.
3. Vasudeva-bit commented on Jun 19, 2026

   

   Vasudeva-bit

   on Jun 19, 2026

   Last edited by Vasudeva-bit

   ContributorAuthor

   More actions

   @martin-frbg I tried a few things. I see that the L3_SIZE macro in param.h and the get_l3_size function in setparam-ref.c track L3 cache size as a total over all the chiplets. The following conditioning works for cases with more or fewer than 4 chiplets (9334 has 4). I believe this covers a majority of the Zen4 family.

   // param.h
   #if L3 > 0 and L3_SIZE % 33554432 == 0
   #define SGEMM_DEFAULT_P 320
   #define SGEMM_DEFAULT_Q 512
   #else
   #define SGEMM_DEFAULT_P 640
   #define SGEMM_DEFAULT_Q 320
   #endif
   
   #define SGEMM_DEFAULT_R sgemm_r
   

   &

   // setparam-ref.c
   int l3_kb = get_l3_size();
   unsigned int eax, ebx, ecx, edx;
   cpuid(0, &eax, &ebx, &ecx, &edx);
   
   // 0x68747541 = "Auth" (AuthenticAMD)
   if ((ebx == 0x68747541) && (l3_kb > 0) && (l3_kb % 32768 == 0)) {
       TABLE_NAME.sgemm_p = 320;
       TABLE_NAME.sgemm_q = 512;
       // R is dynamically calculated
   }
   

   I shall make a PR based on this. Kindly let me know your comments, if any.
4. martin-frbg

   linked a pull request that will close this issueOptimize Zen 4 GEMM macro block sizes (P, Q, R) #5868on Jul 4, 2026
5. martin-frbg

   closed this as completedin #5868on Jul 4, 2026