Changes between Version 1 and Version 2 of Ticket #58179, comment 20


Ignore:
Timestamp:
Mar 7, 2019, 5:54:50 PM (5 years ago)
Author:
kencu (Ken)
Comment:

Legend:

Unmodified
Added
Removed
Modified

  • Ticket #58179, comment 20

    v1 v2  
    1 I think this might be the 970 issue, in `param.h` some #defines seem to be missing -- possibly the Power4 ones would work for the PPC970:
     1I think this might be the 970 issue, in `param.h` some #defines seem to be missing, or may be guarded by a Linux blocker:
    22{{{
    3 #ifdef PPCG4
    4 #define GEMM_DEFAULT_OFFSET_A    0
    5 #define GEMM_DEFAULT_OFFSET_B 1024
    6 #define GEMM_DEFAULT_ALIGN 0x0ffffUL
    7 
    8 #define SGEMM_DEFAULT_UNROLL_M 16
    9 #define SGEMM_DEFAULT_UNROLL_N 4
    10 #define DGEMM_DEFAULT_UNROLL_M 4
    11 #define DGEMM_DEFAULT_UNROLL_N 4
    12 #define CGEMM_DEFAULT_UNROLL_M 8
    13 #define CGEMM_DEFAULT_UNROLL_N 2
    14 #define ZGEMM_DEFAULT_UNROLL_M 2
    15 #define ZGEMM_DEFAULT_UNROLL_N 2
    16 
    17 #define SGEMM_DEFAULT_P 256
    18 #define DGEMM_DEFAULT_P 128
    19 #define CGEMM_DEFAULT_P 128
    20 #define ZGEMM_DEFAULT_P  64
    21 
    22 #define SGEMM_DEFAULT_Q 256
    23 #define DGEMM_DEFAULT_Q 256
    24 #define CGEMM_DEFAULT_Q 256
    25 #define ZGEMM_DEFAULT_Q 256
    26 
    27 #define SYMV_P   4
    28 #endif
    29 
    303#ifdef PPC970
    314
    … …  
    4518#define ZGEMM_DEFAULT_UNROLL_M 2
    4619#define ZGEMM_DEFAULT_UNROLL_N 2
     20
     21#ifdef OS_LINUX
     22#if L2_SIZE == 1024976
     23#define SGEMM_DEFAULT_P 320
     24#define DGEMM_DEFAULT_P 256
     25#define CGEMM_DEFAULT_P 256
     26#define ZGEMM_DEFAULT_P 256
     27#else
     28#define SGEMM_DEFAULT_P 176
     29#define DGEMM_DEFAULT_P 176
     30#define CGEMM_DEFAULT_P 176
     31#define ZGEMM_DEFAULT_P 176
     32#endif
     33#endif
     34
     35#define SGEMM_DEFAULT_Q 512
     36#define DGEMM_DEFAULT_Q 256
     37#define CGEMM_DEFAULT_Q 256
     38#define ZGEMM_DEFAULT_Q 128
     39
     40#define SYMV_P   4
     41
     42#endif
    4743}}}