refs #55. Added DTB_ENTRIES into dynamic arch setting parameters. Now, it can read DTB_ENTRIES on runtime.

This commit is contained in:
traits 2011-09-05 17:37:07 +08:00
parent 12d77deeee
commit 9fc6764fa7
9 changed files with 57 additions and 51 deletions

View File

@ -44,6 +44,7 @@
#ifdef DYNAMIC_ARCH #ifdef DYNAMIC_ARCH
typedef struct { typedef struct {
int dtb_entries;
int offsetA, offsetB, align; int offsetA, offsetB, align;
int sgemm_p, sgemm_q, sgemm_r; int sgemm_p, sgemm_q, sgemm_r;
@ -813,6 +814,7 @@ BLASLONG (*ixamin_k)(BLASLONG, xdouble *, BLASLONG);
extern gotoblas_t *gotoblas; extern gotoblas_t *gotoblas;
#define DTB_ENTRIES gotoblas -> dtb_entries
#define GEMM_OFFSET_A gotoblas -> offsetA #define GEMM_OFFSET_A gotoblas -> offsetA
#define GEMM_OFFSET_B gotoblas -> offsetB #define GEMM_OFFSET_B gotoblas -> offsetB
#define GEMM_ALIGN gotoblas -> align #define GEMM_ALIGN gotoblas -> align
@ -863,6 +865,8 @@ extern gotoblas_t *gotoblas;
#else #else
#define DTB_ENTRIES DTB_DEFAULT_ENTRIES
#define GEMM_OFFSET_A GEMM_DEFAULT_OFFSET_A #define GEMM_OFFSET_A GEMM_DEFAULT_OFFSET_A
#define GEMM_OFFSET_B GEMM_DEFAULT_OFFSET_B #define GEMM_OFFSET_B GEMM_DEFAULT_OFFSET_B
#define GEMM_ALIGN GEMM_DEFAULT_ALIGN #define GEMM_ALIGN GEMM_DEFAULT_ALIGN
@ -997,14 +1001,14 @@ extern gotoblas_t *gotoblas;
#endif #endif
#ifdef XDOUBLE #ifdef XDOUBLE
#define GEMM3M_UNROLL_M QGEMM_DEFAULT_UNROLL_M #define GEMM3M_UNROLL_M QGEMM_UNROLL_M
#define GEMM3M_UNROLL_N QGEMM_DEFAULT_UNROLL_N #define GEMM3M_UNROLL_N QGEMM_UNROLL_N
#elif defined(DOUBLE) #elif defined(DOUBLE)
#define GEMM3M_UNROLL_M DGEMM_DEFAULT_UNROLL_M #define GEMM3M_UNROLL_M DGEMM_UNROLL_M
#define GEMM3M_UNROLL_N DGEMM_DEFAULT_UNROLL_N #define GEMM3M_UNROLL_N DGEMM_UNROLL_N
#else #else
#define GEMM3M_UNROLL_M SGEMM_DEFAULT_UNROLL_M #define GEMM3M_UNROLL_M SGEMM_UNROLL_M
#define GEMM3M_UNROLL_N SGEMM_DEFAULT_UNROLL_N #define GEMM3M_UNROLL_N SGEMM_UNROLL_N
#endif #endif

View File

@ -72,7 +72,7 @@ void get_cpuconfig(void){
printf("#define L1_DATA_LINESIZE 32\n"); printf("#define L1_DATA_LINESIZE 32\n");
printf("#define L2_SIZE 2097152\n"); printf("#define L2_SIZE 2097152\n");
printf("#define L2_LINESIZE 32\n"); printf("#define L2_LINESIZE 32\n");
printf("#define DTB_ENTRIES 32\n"); printf("#define DTB_DEFAULT_ENTRIES 32\n");
printf("#define DTB_SIZE 8192\n"); printf("#define DTB_SIZE 8192\n");
break; break;
@ -81,7 +81,7 @@ void get_cpuconfig(void){
printf("#define L1_DATA_LINESIZE 32\n"); printf("#define L1_DATA_LINESIZE 32\n");
printf("#define L2_SIZE 2097152\n"); printf("#define L2_SIZE 2097152\n");
printf("#define L2_LINESIZE 64\n"); printf("#define L2_LINESIZE 64\n");
printf("#define DTB_ENTRIES 64\n"); printf("#define DTB_DEFAULT_ENTRIES 64\n");
printf("#define DTB_SIZE 8192\n"); printf("#define DTB_SIZE 8192\n");
break; break;
@ -90,7 +90,7 @@ void get_cpuconfig(void){
printf("#define L1_DATA_LINESIZE 64\n"); printf("#define L1_DATA_LINESIZE 64\n");
printf("#define L2_SIZE 4194304\n"); printf("#define L2_SIZE 4194304\n");
printf("#define L2_LINESIZE 64\n"); printf("#define L2_LINESIZE 64\n");
printf("#define DTB_ENTRIES 64\n"); printf("#define DTB_DEFAULT_ENTRIES 64\n");
printf("#define DTB_SIZE 8192\n"); printf("#define DTB_SIZE 8192\n");
break; break;
} }

View File

@ -133,6 +133,6 @@ void get_cpuconfig(void){
printf("#define L2_SIZE 1572864\n"); printf("#define L2_SIZE 1572864\n");
printf("#define L2_LINESIZE 128\n"); printf("#define L2_LINESIZE 128\n");
printf("#define DTB_SIZE 16384\n"); printf("#define DTB_SIZE 16384\n");
printf("#define DTB_ENTRIES 128\n"); printf("#define DTB_DEFAULT_ENTRIES 128\n");
} }

View File

@ -146,7 +146,7 @@ void get_cpuconfig(void){
printf("#define L1_DATA_LINESIZE 32\n"); printf("#define L1_DATA_LINESIZE 32\n");
printf("#define L2_SIZE 512488\n"); printf("#define L2_SIZE 512488\n");
printf("#define L2_LINESIZE 32\n"); printf("#define L2_LINESIZE 32\n");
printf("#define DTB_ENTRIES 64\n"); printf("#define DTB_DEFAULT_ENTRIES 64\n");
printf("#define DTB_SIZE 4096\n"); printf("#define DTB_SIZE 4096\n");
printf("#define L2_ASSOCIATIVE 4\n"); printf("#define L2_ASSOCIATIVE 4\n");
}else{ }else{
@ -155,7 +155,7 @@ void get_cpuconfig(void){
printf("#define L1_DATA_LINESIZE 32\n"); printf("#define L1_DATA_LINESIZE 32\n");
printf("#define L2_SIZE 512488\n"); printf("#define L2_SIZE 512488\n");
printf("#define L2_LINESIZE 32\n"); printf("#define L2_LINESIZE 32\n");
printf("#define DTB_ENTRIES 32\n"); printf("#define DTB_DEFAULT_ENTRIES 32\n");
printf("#define DTB_SIZE 4096\n"); printf("#define DTB_SIZE 4096\n");
printf("#define L2_ASSOCIATIVE 8\n"); printf("#define L2_ASSOCIATIVE 8\n");
} }

View File

@ -165,7 +165,7 @@ void get_cpuconfig(void){
printf("#define L1_DATA_LINESIZE 128\n"); printf("#define L1_DATA_LINESIZE 128\n");
printf("#define L2_SIZE 524288\n"); printf("#define L2_SIZE 524288\n");
printf("#define L2_LINESIZE 128 \n"); printf("#define L2_LINESIZE 128 \n");
printf("#define DTB_ENTRIES 128\n"); printf("#define DTB_DEFAULT_ENTRIES 128\n");
printf("#define DTB_SIZE 4096\n"); printf("#define DTB_SIZE 4096\n");
printf("#define L2_ASSOCIATIVE 8\n"); printf("#define L2_ASSOCIATIVE 8\n");

View File

@ -50,7 +50,7 @@ void get_subdirname(void){
void get_cpuconfig(void){ void get_cpuconfig(void){
printf("#define V9\n"); printf("#define V9\n");
printf("#define DTB_ENTRIES 32\n"); printf("#define DTB_DEFAULT_ENTRIES 32\n");
} }
void get_libname(void){ void get_libname(void){

View File

@ -1393,7 +1393,7 @@ void get_cpuconfig(void){
if (info.size > 0) { if (info.size > 0) {
printf("#define DTB_SIZE %d\n", info.size * 1024); printf("#define DTB_SIZE %d\n", info.size * 1024);
printf("#define DTB_ASSOCIATIVE %d\n", info.associative); printf("#define DTB_ASSOCIATIVE %d\n", info.associative);
printf("#define DTB_ENTRIES %d\n", info.linesize); printf("#define DTB_DEFAULT_ENTRIES %d\n", info.linesize);
} }
features = get_cputype(GET_FEATURE); features = get_cputype(GET_FEATURE);
@ -1422,7 +1422,7 @@ void get_cpuconfig(void){
features = get_coretype(); features = get_coretype();
if (features > 0) printf("#define CORE_%s\n", corename[features]); if (features > 0) printf("#define CORE_%s\n", corename[features]);
} else { } else {
printf("#define DTB_ENTRIES 16\n"); printf("#define DTB_DEFAULT_ENTRIES 16\n");
printf("#define L1_CODE_SIZE 8192\n"); printf("#define L1_CODE_SIZE 8192\n");
printf("#define L1_DATA_SIZE 8192\n"); printf("#define L1_DATA_SIZE 8192\n");
printf("#define L2_SIZE 0\n"); printf("#define L2_SIZE 0\n");

View File

@ -130,7 +130,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPENTIUM2 " \ #define ARCHCONFIG "-DPENTIUM2 " \
"-DL1_DATA_SIZE=16384 -DL1_DATA_LINESIZE=32 " \ "-DL1_DATA_SIZE=16384 -DL1_DATA_LINESIZE=32 " \
"-DL2_SIZE=512488 -DL2_LINESIZE=32 " \ "-DL2_SIZE=512488 -DL2_LINESIZE=32 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 " \ "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
"-DHAVE_CMOV -DHAVE_MMX" "-DHAVE_CMOV -DHAVE_MMX"
#define LIBNAME "p2" #define LIBNAME "p2"
#define CORENAME "P5" #define CORENAME "P5"
@ -144,7 +144,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPENTIUM3 " \ #define ARCHCONFIG "-DPENTIUM3 " \
"-DL1_DATA_SIZE=16384 -DL1_DATA_LINESIZE=32 " \ "-DL1_DATA_SIZE=16384 -DL1_DATA_LINESIZE=32 " \
"-DL2_SIZE=262144 -DL2_LINESIZE=32 " \ "-DL2_SIZE=262144 -DL2_LINESIZE=32 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 " \ "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE " "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE "
#define LIBNAME "coppermine" #define LIBNAME "coppermine"
#define CORENAME "COPPERMINE" #define CORENAME "COPPERMINE"
@ -158,7 +158,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPENTIUM3 " \ #define ARCHCONFIG "-DPENTIUM3 " \
"-DL1_DATA_SIZE=16384 -DL1_DATA_LINESIZE=32 " \ "-DL1_DATA_SIZE=16384 -DL1_DATA_LINESIZE=32 " \
"-DL2_SIZE=524288 -DL2_LINESIZE=32 " \ "-DL2_SIZE=524288 -DL2_LINESIZE=32 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 " \ "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE " "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE "
#define LIBNAME "katmai" #define LIBNAME "katmai"
#define CORENAME "KATMAI" #define CORENAME "KATMAI"
@ -172,7 +172,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPENTIUM4 " \ #define ARCHCONFIG "-DPENTIUM4 " \
"-DL1_DATA_SIZE=8192 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=8192 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=524288 -DL2_LINESIZE=64 " \ "-DL2_SIZE=524288 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " \ "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 " "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 "
#define LIBNAME "northwood" #define LIBNAME "northwood"
#define CORENAME "NORTHWOOD" #define CORENAME "NORTHWOOD"
@ -186,7 +186,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPENTIUM4 " \ #define ARCHCONFIG "-DPENTIUM4 " \
"-DL1_DATA_SIZE=16384 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=16384 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \ "-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " \ "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3" "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3"
#define LIBNAME "prescott" #define LIBNAME "prescott"
#define CORENAME "PRESCOTT" #define CORENAME "PRESCOTT"
@ -200,7 +200,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPENTIUMM " \ #define ARCHCONFIG "-DPENTIUMM " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \ "-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 " \ "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 " "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 "
#define LIBNAME "banias" #define LIBNAME "banias"
#define CORENAME "BANIAS" #define CORENAME "BANIAS"
@ -214,7 +214,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPENTIUMM " \ #define ARCHCONFIG "-DPENTIUMM " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \ "-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 " \ "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 " "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 "
#define LIBNAME "yonah" #define LIBNAME "yonah"
#define CORENAME "YONAH" #define CORENAME "YONAH"
@ -228,7 +228,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DCORE2 " \ #define ARCHCONFIG "-DCORE2 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \ "-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=256 -DDTB_SIZE=4096 " \ "-DDTB_DEFAULT_ENTRIES=256 -DDTB_SIZE=4096 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3" "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3"
#define LIBNAME "core2" #define LIBNAME "core2"
#define CORENAME "CORE2" #define CORENAME "CORE2"
@ -242,7 +242,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPENRYN " \ #define ARCHCONFIG "-DPENRYN " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \ "-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=256 -DDTB_SIZE=4096 " \ "-DDTB_DEFAULT_ENTRIES=256 -DDTB_SIZE=4096 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3 -DHAVE_SSE4_1" "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3 -DHAVE_SSE4_1"
#define LIBNAME "penryn" #define LIBNAME "penryn"
#define CORENAME "PENRYN" #define CORENAME "PENRYN"
@ -257,7 +257,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \ "-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \
"-DL3_SIZE=16777216 -DL3_LINESIZE=64 " \ "-DL3_SIZE=16777216 -DL3_LINESIZE=64 " \
"-DDTB_ENTRIES=256 -DDTB_SIZE=4096 " \ "-DDTB_DEFAULT_ENTRIES=256 -DDTB_SIZE=4096 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3 -DHAVE_SSE4_1" "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3 -DHAVE_SSE4_1"
#define LIBNAME "dunnington" #define LIBNAME "dunnington"
#define CORENAME "DUNNINGTON" #define CORENAME "DUNNINGTON"
@ -271,7 +271,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DNEHALEM " \ #define ARCHCONFIG "-DNEHALEM " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=262144 -DL2_LINESIZE=64 " \ "-DL2_SIZE=262144 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 " \ "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3 -DHAVE_SSE4_1 -DHAVE_SSE4_2" "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3 -DHAVE_SSE4_1 -DHAVE_SSE4_2"
#define LIBNAME "nehalem" #define LIBNAME "nehalem"
#define CORENAME "NEHALEM" #define CORENAME "NEHALEM"
@ -285,7 +285,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DATOM " \ #define ARCHCONFIG "-DATOM " \
"-DL1_DATA_SIZE=24576 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=24576 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=524288 -DL2_LINESIZE=64 " \ "-DL2_SIZE=524288 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=4 " \ "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=4 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3" "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3"
#define LIBNAME "atom" #define LIBNAME "atom"
#define CORENAME "ATOM" #define CORENAME "ATOM"
@ -299,7 +299,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DATHLON " \ #define ARCHCONFIG "-DATHLON " \
"-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \ "-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=32 -DDTB_SIZE=4096 -DHAVE_3DNOW " \ "-DDTB_DEFAULT_ENTRIES=32 -DDTB_SIZE=4096 -DHAVE_3DNOW " \
"-DHAVE_3DNOWEX -DHAVE_MMX -DHAVE_SSE " "-DHAVE_3DNOWEX -DHAVE_MMX -DHAVE_SSE "
#define LIBNAME "athlon" #define LIBNAME "athlon"
#define CORENAME "ATHLON" #define CORENAME "ATHLON"
@ -313,7 +313,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DOPTERON " \ #define ARCHCONFIG "-DOPTERON " \
"-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \ "-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=32 -DDTB_SIZE=4096 -DHAVE_3DNOW " \ "-DDTB_DEFAULT_ENTRIES=32 -DDTB_SIZE=4096 -DHAVE_3DNOW " \
"-DHAVE_3DNOWEX -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 " "-DHAVE_3DNOWEX -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 "
#define LIBNAME "opteron" #define LIBNAME "opteron"
#define CORENAME "OPTERON" #define CORENAME "OPTERON"
@ -327,7 +327,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DOPTERON " \ #define ARCHCONFIG "-DOPTERON " \
"-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \ "-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=32 -DDTB_SIZE=4096 -DHAVE_3DNOW " \ "-DDTB_DEFAULT_ENTRIES=32 -DDTB_SIZE=4096 -DHAVE_3DNOW " \
"-DHAVE_3DNOWEX -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3" "-DHAVE_3DNOWEX -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3"
#define LIBNAME "opteron" #define LIBNAME "opteron"
#define CORENAME "OPTERON" #define CORENAME "OPTERON"
@ -341,7 +341,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DBARCELONA " \ #define ARCHCONFIG "-DBARCELONA " \
"-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=524288 -DL2_LINESIZE=64 -DL3_SIZE=2097152 " \ "-DL2_SIZE=524288 -DL2_LINESIZE=64 -DL3_SIZE=2097152 " \
"-DDTB_ENTRIES=48 -DDTB_SIZE=4096 -DHAVE_3DNOW " \ "-DDTB_DEFAULT_ENTRIES=48 -DDTB_SIZE=4096 -DHAVE_3DNOW " \
"-DHAVE_3DNOWEX -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 " \ "-DHAVE_3DNOWEX -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 " \
"-DHAVE_SSE4A -DHAVE_MISALIGNSSE -DHAVE_128BITFPU -DHAVE_FASTMOVU" "-DHAVE_SSE4A -DHAVE_MISALIGNSSE -DHAVE_128BITFPU -DHAVE_FASTMOVU"
#define LIBNAME "barcelona" #define LIBNAME "barcelona"
@ -356,7 +356,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DGENERIC " \ #define ARCHCONFIG "-DGENERIC " \
"-DL1_DATA_SIZE=16384 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=16384 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=524288 -DL2_LINESIZE=64 " \ "-DL2_SIZE=524288 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " \ "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2" "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2"
#define LIBNAME "generic" #define LIBNAME "generic"
#define CORENAME "GENERIC" #define CORENAME "GENERIC"
@ -370,7 +370,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DVIAC3 " \ #define ARCHCONFIG "-DVIAC3 " \
"-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=32 " \ "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=32 " \
"-DL2_SIZE=65536 -DL2_LINESIZE=32 " \ "-DL2_SIZE=65536 -DL2_LINESIZE=32 " \
"-DDTB_ENTRIES=128 -DDTB_SIZE=4096 " \ "-DDTB_DEFAULT_ENTRIES=128 -DDTB_SIZE=4096 " \
"-DHAVE_MMX -DHAVE_SSE " "-DHAVE_MMX -DHAVE_SSE "
#define LIBNAME "viac3" #define LIBNAME "viac3"
#define CORENAME "VIAC3" #define CORENAME "VIAC3"
@ -384,7 +384,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DNANO " \ #define ARCHCONFIG "-DNANO " \
"-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \ "-DL2_SIZE=1048576 -DL2_LINESIZE=64 " \
"-DDTB_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " \ "-DDTB_DEFAULT_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " \
"-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3" "-DHAVE_CMOV -DHAVE_MMX -DHAVE_SSE -DHAVE_SSE2 -DHAVE_SSE3 -DHAVE_SSSE3"
#define LIBNAME "nano" #define LIBNAME "nano"
#define CORENAME "NANO" #define CORENAME "NANO"
@ -398,7 +398,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPOWER3 " \ #define ARCHCONFIG "-DPOWER3 " \
"-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=128 " \ "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=128 " \
"-DL2_SIZE=2097152 -DL2_LINESIZE=128 " \ "-DL2_SIZE=2097152 -DL2_LINESIZE=128 " \
"-DDTB_ENTRIES=256 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " "-DDTB_DEFAULT_ENTRIES=256 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 "
#define LIBNAME "power3" #define LIBNAME "power3"
#define CORENAME "POWER3" #define CORENAME "POWER3"
#endif #endif
@ -411,7 +411,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPOWER4 " \ #define ARCHCONFIG "-DPOWER4 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \
"-DL2_SIZE=1509949 -DL2_LINESIZE=128 " \ "-DL2_SIZE=1509949 -DL2_LINESIZE=128 " \
"-DDTB_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=6 " "-DDTB_DEFAULT_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=6 "
#define LIBNAME "power4" #define LIBNAME "power4"
#define CORENAME "POWER4" #define CORENAME "POWER4"
#endif #endif
@ -424,7 +424,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPOWER5 " \ #define ARCHCONFIG "-DPOWER5 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \
"-DL2_SIZE=1509949 -DL2_LINESIZE=128 " \ "-DL2_SIZE=1509949 -DL2_LINESIZE=128 " \
"-DDTB_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=6 " "-DDTB_DEFAULT_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=6 "
#define LIBNAME "power5" #define LIBNAME "power5"
#define CORENAME "POWER5" #define CORENAME "POWER5"
#endif #endif
@ -437,7 +437,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPOWER6 " \ #define ARCHCONFIG "-DPOWER6 " \
"-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=128 " \ "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=128 " \
"-DL2_SIZE=4194304 -DL2_LINESIZE=128 " \ "-DL2_SIZE=4194304 -DL2_LINESIZE=128 " \
"-DDTB_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " "-DDTB_DEFAULT_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 "
#define LIBNAME "power6" #define LIBNAME "power6"
#define CORENAME "POWER6" #define CORENAME "POWER6"
#endif #endif
@ -450,7 +450,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPPCG4 " \ #define ARCHCONFIG "-DPPCG4 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=32 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=32 " \
"-DL2_SIZE=262144 -DL2_LINESIZE=32 " \ "-DL2_SIZE=262144 -DL2_LINESIZE=32 " \
"-DDTB_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " "-DDTB_DEFAULT_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 "
#define LIBNAME "ppcg4" #define LIBNAME "ppcg4"
#define CORENAME "PPCG4" #define CORENAME "PPCG4"
#endif #endif
@ -463,7 +463,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPPC970 " \ #define ARCHCONFIG "-DPPC970 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \
"-DL2_SIZE=512488 -DL2_LINESIZE=128 " \ "-DL2_SIZE=512488 -DL2_LINESIZE=128 " \
"-DDTB_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " "-DDTB_DEFAULT_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 "
#define LIBNAME "ppc970" #define LIBNAME "ppc970"
#define CORENAME "PPC970" #define CORENAME "PPC970"
#endif #endif
@ -476,7 +476,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPPC970 " \ #define ARCHCONFIG "-DPPC970 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \
"-DL2_SIZE=1024976 -DL2_LINESIZE=128 " \ "-DL2_SIZE=1024976 -DL2_LINESIZE=128 " \
"-DDTB_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " "-DDTB_DEFAULT_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 "
#define LIBNAME "ppc970mp" #define LIBNAME "ppc970mp"
#define CORENAME "PPC970" #define CORENAME "PPC970"
#endif #endif
@ -489,7 +489,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPPC440 " \ #define ARCHCONFIG "-DPPC440 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=32 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=32 " \
"-DL2_SIZE=16384 -DL2_LINESIZE=128 " \ "-DL2_SIZE=16384 -DL2_LINESIZE=128 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=16 " "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=16 "
#define LIBNAME "ppc440" #define LIBNAME "ppc440"
#define CORENAME "PPC440" #define CORENAME "PPC440"
#endif #endif
@ -502,7 +502,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DPPC440FP2 " \ #define ARCHCONFIG "-DPPC440FP2 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=32 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=32 " \
"-DL2_SIZE=16384 -DL2_LINESIZE=128 " \ "-DL2_SIZE=16384 -DL2_LINESIZE=128 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=16 " "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=16 "
#define LIBNAME "ppc440FP2" #define LIBNAME "ppc440FP2"
#define CORENAME "PPC440FP2" #define CORENAME "PPC440FP2"
#endif #endif
@ -515,7 +515,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DCELL " \ #define ARCHCONFIG "-DCELL " \
"-DL1_DATA_SIZE=262144 -DL1_DATA_LINESIZE=128 " \ "-DL1_DATA_SIZE=262144 -DL1_DATA_LINESIZE=128 " \
"-DL2_SIZE=512488 -DL2_LINESIZE=128 " \ "-DL2_SIZE=512488 -DL2_LINESIZE=128 " \
"-DDTB_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " "-DDTB_DEFAULT_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 "
#define LIBNAME "cell" #define LIBNAME "cell"
#define CORENAME "CELL" #define CORENAME "CELL"
#endif #endif
@ -528,7 +528,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DSICORTEX " \ #define ARCHCONFIG "-DSICORTEX " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=32 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=32 " \
"-DL2_SIZE=512488 -DL2_LINESIZE=32 " \ "-DL2_SIZE=512488 -DL2_LINESIZE=32 " \
"-DDTB_ENTRIES=32 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " "-DDTB_DEFAULT_ENTRIES=32 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 "
#define LIBNAME "mips" #define LIBNAME "mips"
#define CORENAME "sicortex" #define CORENAME "sicortex"
#endif #endif
@ -542,7 +542,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DLOONGSON3A " \ #define ARCHCONFIG "-DLOONGSON3A " \
"-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=32 " \ "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=32 " \
"-DL2_SIZE=512488 -DL2_LINESIZE=32 " \ "-DL2_SIZE=512488 -DL2_LINESIZE=32 " \
"-DDTB_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=4 " "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=4 "
#define LIBNAME "loongson3a" #define LIBNAME "loongson3a"
#define CORENAME "LOONGSON3A" #define CORENAME "LOONGSON3A"
#else #else
@ -555,7 +555,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define SUBDIRNAME "ia64" #define SUBDIRNAME "ia64"
#define ARCHCONFIG "-DITANIUM2 " \ #define ARCHCONFIG "-DITANIUM2 " \
"-DL1_DATA_SIZE=262144 -DL1_DATA_LINESIZE=128 " \ "-DL1_DATA_SIZE=262144 -DL1_DATA_LINESIZE=128 " \
"-DL2_SIZE=1572864 -DL2_LINESIZE=128 -DDTB_SIZE=16384 -DDTB_ENTRIES=128 " "-DL2_SIZE=1572864 -DL2_LINESIZE=128 -DDTB_SIZE=16384 -DDTB_DEFAULT_ENTRIES=128 "
#define LIBNAME "itanium2" #define LIBNAME "itanium2"
#define CORENAME "itanium2" #define CORENAME "itanium2"
#endif #endif
@ -567,7 +567,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define SUBDIRNAME "sparc" #define SUBDIRNAME "sparc"
#define ARCHCONFIG "-DSPARC -DV9 " \ #define ARCHCONFIG "-DSPARC -DV9 " \
"-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=1572864 -DL2_LINESIZE=64 -DDTB_SIZE=8192 -DDTB_ENTRIES=64 " "-DL2_SIZE=1572864 -DL2_LINESIZE=64 -DDTB_SIZE=8192 -DDTB_DEFAULT_ENTRIES=64 "
#define LIBNAME "sparc" #define LIBNAME "sparc"
#define CORENAME "sparc" #define CORENAME "sparc"
#endif #endif
@ -579,7 +579,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define SUBDIRNAME "sparc" #define SUBDIRNAME "sparc"
#define ARCHCONFIG "-DSPARC -DV7 " \ #define ARCHCONFIG "-DSPARC -DV7 " \
"-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \ "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=1572864 -DL2_LINESIZE=64 -DDTB_SIZE=8192 -DDTB_ENTRIES=64 " "-DL2_SIZE=1572864 -DL2_LINESIZE=64 -DDTB_SIZE=8192 -DDTB_DEFAULT_ENTRIES=64 "
#define LIBNAME "sparcv7" #define LIBNAME "sparcv7"
#define CORENAME "sparcv7" #define CORENAME "sparcv7"
#endif #endif
@ -592,7 +592,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define ARCHCONFIG "-DGENERIC " \ #define ARCHCONFIG "-DGENERIC " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \ "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \
"-DL2_SIZE=512488 -DL2_LINESIZE=128 " \ "-DL2_SIZE=512488 -DL2_LINESIZE=128 " \
"-DDTB_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 " "-DDTB_DEFAULT_ENTRIES=128 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=8 "
#define LIBNAME "generic" #define LIBNAME "generic"
#define CORENAME "generic" #define CORENAME "generic"
#endif #endif

View File

@ -49,6 +49,8 @@
static void init_parameter(void); static void init_parameter(void);
gotoblas_t TABLE_NAME = { gotoblas_t TABLE_NAME = {
DTB_DEFAULT_ENTRIES ,
GEMM_DEFAULT_OFFSET_A, GEMM_DEFAULT_OFFSET_B, GEMM_DEFAULT_ALIGN, GEMM_DEFAULT_OFFSET_A, GEMM_DEFAULT_OFFSET_B, GEMM_DEFAULT_ALIGN,
0, 0, 0, 0, 0, 0,