Browse Source

Merge pull request #850 from wernsaar/develop

Bugfixes and enhancements for EXCAVATOR
tags/v0.2.19^2
Werner Saar 10 years ago
parent
commit
6abec09eb4
28 changed files with 65 additions and 139 deletions
  1. +1
    -1
      driver/others/dynamic.c
  2. +2
    -2
      driver/others/parameter.c
  3. +17
    -0
      kernel/setparam-ref.c
  4. +19
    -17
      kernel/x86_64/KERNEL.EXCAVATOR
  5. +1
    -1
      kernel/x86_64/caxpy.c
  6. +1
    -1
      kernel/x86_64/cdot.c
  7. +1
    -1
      kernel/x86_64/cgemv_n_4.c
  8. +1
    -1
      kernel/x86_64/cgemv_t_4.c
  9. +1
    -1
      kernel/x86_64/cscal.c
  10. +1
    -1
      kernel/x86_64/daxpy.c
  11. +1
    -1
      kernel/x86_64/ddot.c
  12. +1
    -1
      kernel/x86_64/dgemv_n_4.c
  13. +1
    -1
      kernel/x86_64/dgemv_t_4.c
  14. +1
    -1
      kernel/x86_64/dscal.c
  15. +1
    -1
      kernel/x86_64/dsymv_L.c
  16. +1
    -1
      kernel/x86_64/dsymv_U.c
  17. +1
    -1
      kernel/x86_64/saxpy.c
  18. +1
    -1
      kernel/x86_64/sdot.c
  19. +2
    -2
      kernel/x86_64/sgemv_n_4.c
  20. +2
    -2
      kernel/x86_64/sgemv_t_4.c
  21. +1
    -1
      kernel/x86_64/ssymv_L.c
  22. +1
    -1
      kernel/x86_64/ssymv_U.c
  23. +1
    -1
      kernel/x86_64/zaxpy.c
  24. +1
    -1
      kernel/x86_64/zdot.c
  25. +1
    -1
      kernel/x86_64/zgemv_n_4.c
  26. +1
    -1
      kernel/x86_64/zgemv_t_4.c
  27. +1
    -1
      kernel/x86_64/zscal.c
  28. +1
    -94
      param.h

+ 1
- 1
driver/others/dynamic.c View File

@@ -439,7 +439,7 @@ static gotoblas_t *force_coretype(char *coretype){
char message[128];
//char mname[20];

for ( i=1 ; i <= 21; i++)
for ( i=1 ; i <= 22; i++)
{
if (!strncasecmp(coretype,corename[i],20))
{


+ 2
- 2
driver/others/parameter.c View File

@@ -167,7 +167,7 @@ int get_L2_size(void){
#if defined(ATHLON) || defined(OPTERON) || defined(BARCELONA) || defined(BOBCAT) || defined(BULLDOZER) || \
defined(CORE_PRESCOTT) || defined(CORE_CORE2) || defined(PENRYN) || defined(DUNNINGTON) || \
defined(CORE_NEHALEM) || defined(CORE_SANDYBRIDGE) || defined(ATOM) || defined(GENERIC) || \
defined(PILEDRIVER) || defined(HASWELL) || defined(STEAMROLLER)
defined(PILEDRIVER) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR)

cpuid(0x80000006, &eax, &ebx, &ecx, &edx);

@@ -251,7 +251,7 @@ int get_L2_size(void){
void blas_set_parameter(void){

int factor;
#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(SANDYBRIDGE) || defined(NEHALEM) || defined(HASWELL) || defined(STEAMROLLER)
#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(SANDYBRIDGE) || defined(NEHALEM) || defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR)
int size = 16;
#else
int size = get_L2_size();


+ 17
- 0
kernel/setparam-ref.c View File

@@ -933,6 +933,23 @@ static void init_parameter(void) {
#endif
#endif

#ifdef EXCAVATOR

#ifdef DEBUG
fprintf(stderr, "Excavator\n");
#endif

TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
#ifdef EXPRECISION
TABLE_NAME.qgemm_p = QGEMM_DEFAULT_P;
TABLE_NAME.xgemm_p = XGEMM_DEFAULT_P;
#endif
#endif


#ifdef PILEDRIVER

#ifdef DEBUG


+ 19
- 17
kernel/x86_64/KERNEL.EXCAVATOR View File

@@ -1,3 +1,7 @@
DSCALKERNEL = dscal.c
CSCALKERNEL = cscal.c
ZSCALKERNEL = zscal.c

SAXPYKERNEL = saxpy.c
DAXPYKERNEL = daxpy.c
CAXPYKERNEL = caxpy.c
@@ -20,7 +24,7 @@ SGEMVTKERNEL = sgemv_t_4.c
DGEMVNKERNEL = dgemv_n_4.c
DGEMVTKERNEL = dgemv_t_4.c

ZGEMVNKERNEL = zgemv_n_dup.S
ZGEMVNKERNEL = zgemv_n_4.c
ZGEMVTKERNEL = zgemv_t_4.c

DCOPYKERNEL = dcopy_bulldozer.S
@@ -68,25 +72,23 @@ ZGEMMOTCOPYOBJ = zgemm_otcopy$(TSUFFIX).$(SUFFIX)
CGEMM3MKERNEL = zgemm3m_kernel_8x4_barcelona.S
ZGEMM3MKERNEL = zgemm3m_kernel_4x4_barcelona.S

STRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
STRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
STRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
STRSMKERNEL_RT = ../generic/trsm_kernel_RT.c

STRSMKERNEL_LN = strsm_kernel_LN_bulldozer.c
STRSMKERNEL_LT = strsm_kernel_LT_bulldozer.c
STRSMKERNEL_RN = strsm_kernel_RN_bulldozer.c
STRSMKERNEL_RT = strsm_kernel_RT_bulldozer.c

DTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
DTRSMKERNEL_LN = dtrsm_kernel_LN_bulldozer.c
DTRSMKERNEL_LT = dtrsm_kernel_LT_8x2_bulldozer.S
DTRSMKERNEL_RN = dtrsm_kernel_RN_8x2_bulldozer.S
DTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c

CTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
CTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
CTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
CTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
DTRSMKERNEL_RT = dtrsm_kernel_RT_bulldozer.c

ZTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
CTRSMKERNEL_LN = ctrsm_kernel_LN_bulldozer.c
CTRSMKERNEL_LT = ctrsm_kernel_LT_bulldozer.c
CTRSMKERNEL_RN = ctrsm_kernel_RN_bulldozer.c
CTRSMKERNEL_RT = ctrsm_kernel_RT_bulldozer.c

ZTRSMKERNEL_LN = ztrsm_kernel_LN_bulldozer.c
ZTRSMKERNEL_LT = ztrsm_kernel_LT_bulldozer.c
ZTRSMKERNEL_RN = ztrsm_kernel_RN_bulldozer.c
ZTRSMKERNEL_RT = ztrsm_kernel_RT_bulldozer.c


+ 1
- 1
kernel/x86_64/caxpy.c View File

@@ -29,7 +29,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"


#if defined(PILEDRIVER) || defined(STEAMROLLER)
#if defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "caxpy_microk_steamroller-2.c"
#elif defined(BULLDOZER)
#include "caxpy_microk_bulldozer-2.c"


+ 1
- 1
kernel/x86_64/cdot.c View File

@@ -32,7 +32,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#if defined(BULLDOZER)
#include "cdot_microk_bulldozer-2.c"
#elif defined(STEAMROLLER) || defined(PILEDRIVER)
#elif defined(STEAMROLLER) || defined(PILEDRIVER) || defined(EXCAVATOR)
#include "cdot_microk_steamroller-2.c"
#elif defined(HASWELL)
#include "cdot_microk_haswell-2.c"


+ 1
- 1
kernel/x86_64/cgemv_n_4.c View File

@@ -31,7 +31,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#if defined(HASWELL)
#include "cgemv_n_microk_haswell-4.c"
#elif defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
#elif defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "cgemv_n_microk_bulldozer-4.c"
#endif



+ 1
- 1
kernel/x86_64/cgemv_t_4.c View File

@@ -30,7 +30,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#if defined(HASWELL)
#include "cgemv_t_microk_haswell-4.c"
#elif defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
#elif defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "cgemv_t_microk_bulldozer-4.c"
#endif



+ 1
- 1
kernel/x86_64/cscal.c View File

@@ -32,7 +32,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "cscal_microk_haswell-2.c"
#elif defined(BULLDOZER) || defined(PILEDRIVER)
#include "cscal_microk_bulldozer-2.c"
#elif defined(STEAMROLLER)
#elif defined(STEAMROLLER) || defined(EXCAVATOR)
#include "cscal_microk_steamroller-2.c"
#elif defined(SANDYBRIDGE)
#include "cscal_microk_bulldozer-2.c"


+ 1
- 1
kernel/x86_64/daxpy.c View File

@@ -33,7 +33,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "daxpy_microk_nehalem-2.c"
#elif defined(BULLDOZER)
#include "daxpy_microk_bulldozer-2.c"
#elif defined(STEAMROLLER)
#elif defined(STEAMROLLER) || defined(EXCAVATOR)
#include "daxpy_microk_steamroller-2.c"
#elif defined(PILEDRIVER)
#include "daxpy_microk_piledriver-2.c"


+ 1
- 1
kernel/x86_64/ddot.c View File

@@ -31,7 +31,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#if defined(BULLDOZER)
#include "ddot_microk_bulldozer-2.c"
#elif defined(STEAMROLLER)
#elif defined(STEAMROLLER) || defined(EXCAVATOR)
#include "ddot_microk_steamroller-2.c"
#elif defined(PILEDRIVER)
#include "ddot_microk_piledriver-2.c"


+ 1
- 1
kernel/x86_64/dgemv_n_4.c View File

@@ -31,7 +31,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#if defined(NEHALEM)
#include "dgemv_n_microk_nehalem-4.c"
#elif defined(HASWELL) || defined(STEAMROLLER)
#elif defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "dgemv_n_microk_haswell-4.c"
#endif



+ 1
- 1
kernel/x86_64/dgemv_t_4.c View File

@@ -28,7 +28,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#include "common.h"

#if defined(HASWELL) || defined(STEAMROLLER)
#if defined(HASWELL) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "dgemv_t_microk_haswell-4.c"
#endif



+ 1
- 1
kernel/x86_64/dscal.c View File

@@ -27,7 +27,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#include "common.h"

#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "dscal_microk_bulldozer-2.c"
#elif defined(SANDYBRIDGE)
#include "dscal_microk_sandy-2.c"


+ 1
- 1
kernel/x86_64/dsymv_L.c View File

@@ -28,7 +28,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#include "common.h"

#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "dsymv_L_microk_bulldozer-2.c"
#elif defined(HASWELL)
#include "dsymv_L_microk_haswell-2.c"


+ 1
- 1
kernel/x86_64/dsymv_U.c View File

@@ -29,7 +29,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"


#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "dsymv_U_microk_bulldozer-2.c"
#elif defined(HASWELL)
#include "dsymv_U_microk_haswell-2.c"


+ 1
- 1
kernel/x86_64/saxpy.c View File

@@ -35,7 +35,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "saxpy_microk_haswell-2.c"
#elif defined(SANDYBRIDGE)
#include "saxpy_microk_sandy-2.c"
#elif defined(PILEDRIVER) || defined(STEAMROLLER)
#elif defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "saxpy_microk_piledriver-2.c"
#endif



+ 1
- 1
kernel/x86_64/sdot.c View File

@@ -30,7 +30,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#if defined(BULLDOZER)
#include "sdot_microk_bulldozer-2.c"
#elif defined(STEAMROLLER) || defined(PILEDRIVER)
#elif defined(STEAMROLLER) || defined(PILEDRIVER) || defined(EXCAVATOR)
#include "sdot_microk_steamroller-2.c"
#elif defined(NEHALEM)
#include "sdot_microk_nehalem-2.c"


+ 2
- 2
kernel/x86_64/sgemv_n_4.c View File

@@ -29,7 +29,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"


#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "sgemv_n_microk_bulldozer-4.c"
#elif defined(NEHALEM)
#include "sgemv_n_microk_nehalem-4.c"
@@ -39,7 +39,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "sgemv_n_microk_haswell-4.c"
#endif

#if defined(STEAMROLLER)
#if defined(STEAMROLLER) || defined(EXCAVATOR)
#define NBMAX 2048
#else
#define NBMAX 4096


+ 2
- 2
kernel/x86_64/sgemv_t_4.c View File

@@ -30,7 +30,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#if defined(NEHALEM)
#include "sgemv_t_microk_nehalem-4.c"
#elif defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
#elif defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "sgemv_t_microk_bulldozer-4.c"
#elif defined(SANDYBRIDGE)
#include "sgemv_t_microk_sandy-4.c"
@@ -38,7 +38,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "sgemv_t_microk_haswell-4.c"
#endif

#if defined(STEAMROLLER)
#if defined(STEAMROLLER) || defined(EXCAVATOR)
#define NBMAX 2048
#else
#define NBMAX 4096


+ 1
- 1
kernel/x86_64/ssymv_L.c View File

@@ -28,7 +28,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#include "common.h"

#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "ssymv_L_microk_bulldozer-2.c"
#elif defined(NEHALEM)
#include "ssymv_L_microk_nehalem-2.c"


+ 1
- 1
kernel/x86_64/ssymv_U.c View File

@@ -29,7 +29,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"


#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "ssymv_U_microk_bulldozer-2.c"
#elif defined(NEHALEM)
#include "ssymv_U_microk_nehalem-2.c"


+ 1
- 1
kernel/x86_64/zaxpy.c View File

@@ -31,7 +31,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#if defined(BULLDOZER)
#include "zaxpy_microk_bulldozer-2.c"
#elif defined(PILEDRIVER) || defined(STEAMROLLER)
#elif defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "zaxpy_microk_steamroller-2.c"
#elif defined(HASWELL)
#include "zaxpy_microk_haswell-2.c"


+ 1
- 1
kernel/x86_64/zdot.c View File

@@ -32,7 +32,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#if defined(BULLDOZER)
#include "zdot_microk_bulldozer-2.c"
#elif defined(STEAMROLLER) || defined(PILEDRIVER)
#elif defined(STEAMROLLER) || defined(PILEDRIVER) || defined(EXCAVATOR)
#include "zdot_microk_steamroller-2.c"
#elif defined(HASWELL)
#include "zdot_microk_haswell-2.c"


+ 1
- 1
kernel/x86_64/zgemv_n_4.c View File

@@ -34,7 +34,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "zgemv_n_microk_haswell-4.c"
#elif defined(SANDYBRIDGE)
#include "zgemv_n_microk_sandy-4.c"
#elif defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
#elif defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "zgemv_n_microk_bulldozer-4.c"
#endif



+ 1
- 1
kernel/x86_64/zgemv_t_4.c View File

@@ -29,7 +29,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"


#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER)
#if defined(BULLDOZER) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
#include "zgemv_t_microk_bulldozer-4.c"
#elif defined(HASWELL)
#include "zgemv_t_microk_haswell-4.c"


+ 1
- 1
kernel/x86_64/zscal.c View File

@@ -32,7 +32,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "zscal_microk_haswell-2.c"
#elif defined(BULLDOZER) || defined(PILEDRIVER)
#include "zscal_microk_bulldozer-2.c"
#elif defined(STEAMROLLER)
#elif defined(STEAMROLLER) || defined(EXCAVATOR)
#include "zscal_microk_steamroller-2.c"
#endif



+ 1
- 94
param.h View File

@@ -410,100 +410,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#endif

#ifdef STEAMROLLER
#define SNUMOPT 8
#define DNUMOPT 4

#define GEMM_DEFAULT_OFFSET_A 64
#define GEMM_DEFAULT_OFFSET_B 832
#define GEMM_DEFAULT_ALIGN 0x0fffUL



#define QGEMM_DEFAULT_UNROLL_N 2
#define CGEMM_DEFAULT_UNROLL_N 2
#define ZGEMM_DEFAULT_UNROLL_N 2
#define XGEMM_DEFAULT_UNROLL_N 1

#ifdef ARCH_X86
#define SGEMM_DEFAULT_UNROLL_N 4
#define DGEMM_DEFAULT_UNROLL_N 4
#define SGEMM_DEFAULT_UNROLL_M 4
#define DGEMM_DEFAULT_UNROLL_M 2
#define QGEMM_DEFAULT_UNROLL_M 2
#define CGEMM_DEFAULT_UNROLL_M 2
#define ZGEMM_DEFAULT_UNROLL_M 1
#define XGEMM_DEFAULT_UNROLL_M 1
#else
#define SGEMM_DEFAULT_UNROLL_N 2
#define DGEMM_DEFAULT_UNROLL_N 2
#define SGEMM_DEFAULT_UNROLL_M 16
#define DGEMM_DEFAULT_UNROLL_M 8
#define QGEMM_DEFAULT_UNROLL_M 2
#define CGEMM_DEFAULT_UNROLL_M 4
#define ZGEMM_DEFAULT_UNROLL_M 2
#define XGEMM_DEFAULT_UNROLL_M 1
#define CGEMM3M_DEFAULT_UNROLL_N 4
#define CGEMM3M_DEFAULT_UNROLL_M 8
#define ZGEMM3M_DEFAULT_UNROLL_N 4
#define ZGEMM3M_DEFAULT_UNROLL_M 4
#define GEMV_UNROLL 8
#endif

#if defined(ARCH_X86_64)
#define SGEMM_DEFAULT_P 768
#define DGEMM_DEFAULT_P 576
#define ZGEMM_DEFAULT_P 288
#define CGEMM_DEFAULT_P 576
#else
#define SGEMM_DEFAULT_P 448
#define DGEMM_DEFAULT_P 480
#define ZGEMM_DEFAULT_P 112
#define CGEMM_DEFAULT_P 224
#endif
#define QGEMM_DEFAULT_P 112
#define XGEMM_DEFAULT_P 56

#if defined(ARCH_X86_64)
#define SGEMM_DEFAULT_Q 192
#define DGEMM_DEFAULT_Q 160
#define ZGEMM_DEFAULT_Q 160
#define CGEMM_DEFAULT_Q 160
#else
#define SGEMM_DEFAULT_Q 224
#define DGEMM_DEFAULT_Q 224
#define ZGEMM_DEFAULT_Q 224
#define CGEMM_DEFAULT_Q 224
#endif
#define QGEMM_DEFAULT_Q 224
#define XGEMM_DEFAULT_Q 224

#define CGEMM3M_DEFAULT_P 448
#define ZGEMM3M_DEFAULT_P 224
#define XGEMM3M_DEFAULT_P 112
#define CGEMM3M_DEFAULT_Q 224
#define ZGEMM3M_DEFAULT_Q 224
#define XGEMM3M_DEFAULT_Q 224
#define CGEMM3M_DEFAULT_R 12288
#define ZGEMM3M_DEFAULT_R 12288
#define XGEMM3M_DEFAULT_R 12288

#define SGEMM_DEFAULT_R 12288
#define QGEMM_DEFAULT_R qgemm_r
#define DGEMM_DEFAULT_R 12288
#define CGEMM_DEFAULT_R cgemm_r
#define ZGEMM_DEFAULT_R zgemm_r
#define XGEMM_DEFAULT_R xgemm_r

#define SYMV_P 16
#define HAVE_EXCLUSIVE_CACHE

#define GEMM_THREAD gemm_thread_mn

#endif


#ifdef EXCAVATOR
#if defined(STEAMROLLER) || defined(EXCAVATOR)
#define SNUMOPT 8
#define DNUMOPT 4



Loading…
Cancel
Save