diff --git a/kernel/mips/KERNEL.P5600 b/kernel/mips/KERNEL.P5600 index 9a6e06d67..2e1e3390d 100644 --- a/kernel/mips/KERNEL.P5600 +++ b/kernel/mips/KERNEL.P5600 @@ -35,7 +35,7 @@ DSUMKERNEL = ../mips/sum.c CSUMKERNEL = ../mips/zsum.c ZSUMKERNEL = ../mips/zsum.c -ifdef HAVE_MSA +ifndef NO_MSA SASUMKERNEL = ../mips/sasum_msa.c DASUMKERNEL = ../mips/dasum_msa.c CASUMKERNEL = ../mips/casum_msa.c @@ -47,7 +47,7 @@ CASUMKERNEL = ../mips/zasum.c ZASUMKERNEL = ../mips/zasum.c endif -ifdef HAVE_MSA +ifndef NO_MSA SAXPYKERNEL = ../mips/saxpy_msa.c DAXPYKERNEL = ../mips/daxpy_msa.c CAXPYKERNEL = ../mips/caxpy_msa.c @@ -59,7 +59,7 @@ CAXPYKERNEL = ../mips/zaxpy.c ZAXPYKERNEL = ../mips/zaxpy.c endif -ifdef HAVE_MSA +ifndef NO_MSA SCOPYKERNEL = ../mips/scopy_msa.c DCOPYKERNEL = ../mips/dcopy_msa.c CCOPYKERNEL = ../mips/ccopy_msa.c @@ -71,7 +71,7 @@ CCOPYKERNEL = ../mips/zcopy.c ZCOPYKERNEL = ../mips/zcopy.c endif -ifdef HAVE_MSA +ifndef NO_MSA SDOTKERNEL = ../mips/sdot_msa.c DDOTKERNEL = ../mips/ddot_msa.c CDOTKERNEL = ../mips/cdot_msa.c @@ -88,7 +88,7 @@ DNRM2KERNEL = ../mips/nrm2.c CNRM2KERNEL = ../mips/znrm2.c ZNRM2KERNEL = ../mips/znrm2.c -ifdef HAVE_MSA +ifndef NO_MSA SROTKERNEL = ../mips/srot_msa.c DROTKERNEL = ../mips/drot_msa.c CROTKERNEL = ../mips/crot_msa.c @@ -100,7 +100,7 @@ CROTKERNEL = ../mips/zrot.c ZROTKERNEL = ../mips/zrot.c endif -ifdef HAVE_MSA +ifndef NO_MSA SSCALKERNEL = ../mips/sscal_msa.c DSCALKERNEL = ../mips/dscal_msa.c CSCALKERNEL = ../mips/cscal_msa.c @@ -112,7 +112,7 @@ CSCALKERNEL = ../mips/zscal.c ZSCALKERNEL = ../mips/zscal.c endif -ifdef HAVE_MSA +ifndef NO_MSA SSWAPKERNEL = ../mips/sswap_msa.c DSWAPKERNEL = ../mips/dswap_msa.c CSWAPKERNEL = ../mips/cswap_msa.c @@ -124,7 +124,7 @@ CSWAPKERNEL = ../mips/zswap.c ZSWAPKERNEL = ../mips/zswap.c endif -ifdef HAVE_MSA +ifndef NO_MSA SGEMVNKERNEL = ../mips/sgemv_n_msa.c DGEMVNKERNEL = ../mips/dgemv_n_msa.c CGEMVNKERNEL = ../mips/cgemv_n_msa.c @@ -136,7 +136,7 @@ CGEMVNKERNEL = ../mips/zgemv_n.c ZGEMVNKERNEL = ../mips/zgemv_n.c endif -ifdef HAVE_MSA +ifndef NO_MSA SGEMVTKERNEL = ../mips/sgemv_t_msa.c DGEMVTKERNEL = ../mips/dgemv_t_msa.c CGEMVTKERNEL = ../mips/cgemv_t_msa.c @@ -148,7 +148,7 @@ CGEMVTKERNEL = ../mips/zgemv_t.c ZGEMVTKERNEL = ../mips/zgemv_t.c endif -ifdef HAVE_MSA +ifndef NO_MSA SGEMMKERNEL = ../mips/sgemm_kernel_8x8_msa.c SGEMMONCOPY = ../mips/sgemm_ncopy_8_msa.c SGEMMOTCOPY = ../mips/sgemm_tcopy_8_msa.c @@ -162,7 +162,7 @@ SGEMMONCOPYOBJ = sgemm_oncopy.o SGEMMOTCOPYOBJ = sgemm_otcopy.o endif -ifdef HAVE_MSA +ifndef NO_MSA DGEMMKERNEL = ../mips/dgemm_kernel_8x4_msa.c DGEMMINCOPY = ../mips/dgemm_ncopy_8_msa.c DGEMMITCOPY = ../mips/dgemm_tcopy_8_msa.c @@ -180,7 +180,7 @@ DGEMMONCOPYOBJ = dgemm_oncopy.o DGEMMOTCOPYOBJ = dgemm_otcopy.o endif -ifdef HAVE_MSA +ifndef NO_MSA CGEMMKERNEL = ../mips/cgemm_kernel_8x4_msa.c CGEMMINCOPY = ../mips/cgemm_ncopy_8_msa.c CGEMMITCOPY = ../mips/cgemm_tcopy_8_msa.c @@ -198,7 +198,7 @@ CGEMMONCOPYOBJ = cgemm_oncopy.o CGEMMOTCOPYOBJ = cgemm_otcopy.o endif -ifdef HAVE_MSA +ifndef NO_MSA ZGEMMKERNEL = ../mips/zgemm_kernel_4x4_msa.c ZGEMMONCOPY = ../mips/zgemm_ncopy_4_msa.c ZGEMMOTCOPY = ../mips/zgemm_tcopy_4_msa.c @@ -212,7 +212,7 @@ ZGEMMONCOPYOBJ = zgemm_oncopy.o ZGEMMOTCOPYOBJ = zgemm_otcopy.o endif -ifdef HAVE_MSA +ifndef NO_MSA STRSMKERNEL_LN = ../mips/strsm_kernel_LN_8x8_msa.c STRSMKERNEL_LT = ../mips/strsm_kernel_LT_8x8_msa.c STRSMKERNEL_RN = ../mips/strsm_kernel_RN_8x8_msa.c @@ -224,7 +224,7 @@ STRSMKERNEL_RN = ../generic/trsm_kernel_RN.c STRSMKERNEL_RT = ../generic/trsm_kernel_RT.c endif -ifdef HAVE_MSA +ifndef NO_MSA DTRSMKERNEL_LN = ../mips/dtrsm_kernel_LN_8x4_msa.c DTRSMKERNEL_LT = ../mips/dtrsm_kernel_LT_8x4_msa.c DTRSMKERNEL_RN = ../mips/dtrsm_kernel_RN_8x4_msa.c @@ -236,7 +236,7 @@ DTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c DTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c endif -ifdef HAVE_MSA +ifndef NO_MSA CTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c CTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c CTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c @@ -248,7 +248,7 @@ CTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c CTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c endif -ifdef HAVE_MSA +ifndef NO_MSA ZTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c diff --git a/kernel/mips64/KERNEL.LOONGSON3R4 b/kernel/mips64/KERNEL.LOONGSON3R4 index b81e5441d..1149d97f1 100644 --- a/kernel/mips64/KERNEL.LOONGSON3R4 +++ b/kernel/mips64/KERNEL.LOONGSON3R4 @@ -1,4 +1,4 @@ -ifdef HAVE_MSA +ifndef NO_MSA SAXPYKERNEL = ../mips/saxpy_msa.c DAXPYKERNEL = ../mips/daxpy_msa.c CAXPYKERNEL = ../mips/caxpy_msa.c @@ -8,14 +8,14 @@ SAXPYKERNEL = axpy_loongson3a.S DAXPYKERNEL = daxpy_loongson3a_simd.S endif -ifdef HAVE_MSA +ifndef NO_MSA SCOPYKERNEL = ../mips/scopy_msa.c DCOPYKERNEL = ../mips/dcopy_msa.c CCOPYKERNEL = ../mips/ccopy_msa.c ZCOPYKERNEL = ../mips/zcopy_msa.c endif -ifdef HAVE_MSA +ifndef NO_MSA SDOTKERNEL = ../mips/sdot_msa.c DDOTKERNEL = ../mips/ddot_msa.c CDOTKERNEL = ../mips/cdot_msa.c @@ -23,21 +23,21 @@ ZDOTKERNEL = ../mips/zdot_msa.c endif DSDOTKERNEL = ../mips/dot.c -ifdef HAVE_MSA +ifndef NO_MSA SROTKERNEL = ../mips/srot_msa.c DROTKERNEL = ../mips/drot_msa.c CROTKERNEL = ../mips/crot_msa.c ZROTKERNEL = ../mips/zrot_msa.c endif -ifdef HAVE_MSA +ifndef NO_MSA SSCALKERNEL = ../mips/sscal_msa.c DSCALKERNEL = ../mips/dscal_msa.c CSCALKERNEL = ../mips/cscal_msa.c ZSCALKERNEL = ../mips/zscal_msa.c endif -ifdef HAVE_MSA +ifndef NO_MSA SGEMVNKERNEL = ../mips/sgemv_n_msa.c DGEMVNKERNEL = ../mips/dgemv_n_msa.c SGEMVTKERNEL = ../mips/sgemv_t_msa.c @@ -57,21 +57,21 @@ ZGEMVNKERNEL = zgemv_n_loongson3a.c ZGEMVTKERNEL = zgemv_t_loongson3a.c endif -ifdef HAVE_MSA +ifndef NO_MSA SASUMKERNEL = ../mips/sasum_msa.c DASUMKERNEL = ../mips/dasum_msa.c CASUMKERNEL = ../mips/casum_msa.c ZASUMKERNEL = ../mips/zasum_msa.c endif -ifdef HAVE_MSA +ifndef NO_MSA SSWAPKERNEL = ../mips/sswap_msa.c DSWAPKERNEL = ../mips/dswap_msa.c CSWAPKERNEL = ../mips/cswap_msa.c ZSWAPKERNEL = ../mips/zswap_msa.c endif -ifdef HAVE_MSA +ifndef NO_MSA SGEMMKERNEL = ../mips/sgemm_kernel_8x8_msa.c SGEMMONCOPY = ../mips/sgemm_ncopy_8_msa.c SGEMMOTCOPY = ../mips/sgemm_tcopy_8_msa.c @@ -89,7 +89,7 @@ SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX) SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX) endif -ifdef HAVE_MSA +ifndef NO_MSA DGEMMKERNEL = ../mips/dgemm_kernel_8x4_msa.c DGEMMINCOPY = ../mips/dgemm_ncopy_8_msa.c DGEMMITCOPY = ../mips/dgemm_tcopy_8_msa.c @@ -107,7 +107,7 @@ DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX) DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX) endif -ifdef HAVE_MSA +ifndef NO_MSA CGEMMKERNEL = ../mips/cgemm_kernel_8x4_msa.c CGEMMINCOPY = ../mips/cgemm_ncopy_8_msa.c CGEMMITCOPY = ../mips/cgemm_tcopy_8_msa.c @@ -129,7 +129,7 @@ CGEMMONCOPYOBJ = cgemm_oncopy$(TSUFFIX).$(SUFFIX) CGEMMOTCOPYOBJ = cgemm_otcopy$(TSUFFIX).$(SUFFIX) endif -ifdef HAVE_MSA +ifndef NO_MSA ZGEMMKERNEL = ../mips/zgemm_kernel_4x4_msa.c ZGEMMONCOPY = ../mips/zgemm_ncopy_4_msa.c ZGEMMOTCOPY = ../mips/zgemm_tcopy_4_msa.c @@ -143,7 +143,7 @@ ZGEMMONCOPYOBJ = zgemm_oncopy$(TSUFFIX).$(SUFFIX) ZGEMMOTCOPYOBJ = zgemm_otcopy$(TSUFFIX).$(SUFFIX) endif -ifdef HAVE_MSA +ifndef NO_MSA STRSMKERNEL_LN = ../mips/strsm_kernel_LN_8x8_msa.c STRSMKERNEL_LT = ../mips/strsm_kernel_LT_8x8_msa.c STRSMKERNEL_RN = ../mips/strsm_kernel_RN_8x8_msa.c @@ -155,7 +155,7 @@ STRSMKERNEL_RN = ../generic/trsm_kernel_RN.c STRSMKERNEL_RT = ../generic/trsm_kernel_RT.c endif -ifdef HAVE_MSA +ifndef NO_MSA DTRSMKERNEL_LN = ../mips/dtrsm_kernel_LN_8x4_msa.c DTRSMKERNEL_LT = ../mips/dtrsm_kernel_LT_8x4_msa.c DTRSMKERNEL_RN = ../mips/dtrsm_kernel_RN_8x4_msa.c @@ -167,7 +167,7 @@ DTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c DTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c endif -ifdef HAVE_MSA +ifndef NO_MSA CTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c CTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c CTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c @@ -179,7 +179,7 @@ CTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c CTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c endif -ifdef HAVE_MSA +ifndef NO_MSA ZTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c diff --git a/param.h b/param.h index 547463b2f..a4672ce07 100644 --- a/param.h +++ b/param.h @@ -2748,19 +2748,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define GEMM_DEFAULT_OFFSET_B 0 #define GEMM_DEFAULT_ALIGN (BLASLONG)0x03fffUL -#ifdef HAVE_MSA -#define SGEMM_DEFAULT_UNROLL_M 8 -#define SGEMM_DEFAULT_UNROLL_N 8 - -#define DGEMM_DEFAULT_UNROLL_M 8 -#define DGEMM_DEFAULT_UNROLL_N 4 - -#define CGEMM_DEFAULT_UNROLL_M 8 -#define CGEMM_DEFAULT_UNROLL_N 4 - -#define ZGEMM_DEFAULT_UNROLL_M 4 -#define ZGEMM_DEFAULT_UNROLL_N 4 -#else +#if defined(NO_MSA) #define SGEMM_DEFAULT_UNROLL_M 8 #define SGEMM_DEFAULT_UNROLL_N 4 @@ -2772,6 +2760,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define ZGEMM_DEFAULT_UNROLL_M 2 #define ZGEMM_DEFAULT_UNROLL_N 2 +#else +#define SGEMM_DEFAULT_UNROLL_M 8 +#define SGEMM_DEFAULT_UNROLL_N 8 + +#define DGEMM_DEFAULT_UNROLL_M 8 +#define DGEMM_DEFAULT_UNROLL_N 4 + +#define CGEMM_DEFAULT_UNROLL_M 8 +#define CGEMM_DEFAULT_UNROLL_N 4 + +#define ZGEMM_DEFAULT_UNROLL_M 4 +#define ZGEMM_DEFAULT_UNROLL_N 4 #endif #define SGEMM_DEFAULT_P 64 @@ -2961,19 +2961,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define GEMM_DEFAULT_OFFSET_B 0 #define GEMM_DEFAULT_ALIGN (BLASLONG) 0x03fffUL -#if defined(HAVE_MSA) -#define SGEMM_DEFAULT_UNROLL_M 8 -#define SGEMM_DEFAULT_UNROLL_N 8 - -#define DGEMM_DEFAULT_UNROLL_M 8 -#define DGEMM_DEFAULT_UNROLL_N 4 - -#define CGEMM_DEFAULT_UNROLL_M 8 -#define CGEMM_DEFAULT_UNROLL_N 4 - -#define ZGEMM_DEFAULT_UNROLL_M 4 -#define ZGEMM_DEFAULT_UNROLL_N 4 -#else +#if defined(NO_MSA) #define SGEMM_DEFAULT_UNROLL_M 2 #define SGEMM_DEFAULT_UNROLL_N 2 @@ -2985,6 +2973,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. #define ZGEMM_DEFAULT_UNROLL_M 2 #define ZGEMM_DEFAULT_UNROLL_N 2 +#else +#define SGEMM_DEFAULT_UNROLL_M 8 +#define SGEMM_DEFAULT_UNROLL_N 8 + +#define DGEMM_DEFAULT_UNROLL_M 8 +#define DGEMM_DEFAULT_UNROLL_N 4 + +#define CGEMM_DEFAULT_UNROLL_M 8 +#define CGEMM_DEFAULT_UNROLL_N 4 + +#define ZGEMM_DEFAULT_UNROLL_M 4 +#define ZGEMM_DEFAULT_UNROLL_N 4 #endif #define SGEMM_DEFAULT_P 128