Skip to content

Commit

Permalink
Merge branch 'develop' into dev/slewis/merge-from-riscv
Browse files Browse the repository at this point in the history
  • Loading branch information
@sergei-lewis
sergei-lewis committed Feb 1, 2024
2 parents a3b0ef6 + ec74dcd commit 3ffd686
Show file tree
Hide file tree
Showing 21 changed files with 4,154 additions and 4,079 deletions.
6,888 changes: 3,444 additions & 3,444 deletions benchmark/Makefile
View file

Large diffs are not rendered by default.

13 changes: 13 additions & 0 deletions cblas.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -101,6 +101,16 @@ CBLAS_INDEX cblas_idamin(OPENBLAS_CONST blasint n, OPENBLAS_CONST double *x, OPE
CBLAS_INDEX cblas_icamin(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx);
CBLAS_INDEX cblas_izamin(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx);

float cblas_samax(OPENBLAS_CONST blasint n, OPENBLAS_CONST float *x, OPENBLAS_CONST blasint incx);
double cblas_damax(OPENBLAS_CONST blasint n, OPENBLAS_CONST double *x, OPENBLAS_CONST blasint incx);
float cblas_scamax(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx);
double cblas_dzamax(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx);

float cblas_samin(OPENBLAS_CONST blasint n, OPENBLAS_CONST float *x, OPENBLAS_CONST blasint incx);
double cblas_damin(OPENBLAS_CONST blasint n, OPENBLAS_CONST double *x, OPENBLAS_CONST blasint incx);
float cblas_scamin(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx);
double cblas_dzamin(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx);

CBLAS_INDEX cblas_ismax(OPENBLAS_CONST blasint n, OPENBLAS_CONST float *x, OPENBLAS_CONST blasint incx);
CBLAS_INDEX cblas_idmax(OPENBLAS_CONST blasint n, OPENBLAS_CONST double *x, OPENBLAS_CONST blasint incx);
CBLAS_INDEX cblas_icmax(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx);
Expand All @@ -116,6 +126,9 @@ void cblas_daxpy(OPENBLAS_CONST blasint n, OPENBLAS_CONST double alpha, OPENBLAS
void cblas_caxpy(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *alpha, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx, void *y, OPENBLAS_CONST blasint incy);
void cblas_zaxpy(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *alpha, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx, void *y, OPENBLAS_CONST blasint incy);

void cblas_caxpyc(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *alpha, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx, void *y, OPENBLAS_CONST blasint incy);
void cblas_zaxpyc(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *alpha, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx, void *y, OPENBLAS_CONST blasint incy);

void cblas_scopy(OPENBLAS_CONST blasint n, OPENBLAS_CONST float *x, OPENBLAS_CONST blasint incx, float *y, OPENBLAS_CONST blasint incy);
void cblas_dcopy(OPENBLAS_CONST blasint n, OPENBLAS_CONST double *x, OPENBLAS_CONST blasint incx, double *y, OPENBLAS_CONST blasint incy);
void cblas_ccopy(OPENBLAS_CONST blasint n, OPENBLAS_CONST void *x, OPENBLAS_CONST blasint incx, void *y, OPENBLAS_CONST blasint incy);
Expand Down
2 changes: 2 additions & 0 deletions interface/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -130,6 +130,8 @@ endif ()
foreach (float_type ${FLOAT_TYPES})

if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
GenerateNamedObjects("zaxpy.c" "" "axpyc" ${CBLAS_FLAG} "" "" false ${float_type})

GenerateNamedObjects("zger.c" "" "geru" ${CBLAS_FLAG} "" "" false ${float_type})
GenerateNamedObjects("zger.c" "CONJ" "gerc" ${CBLAS_FLAG} "" "" false ${float_type})
GenerateNamedObjects("zdot.c" "CONJ" "dotc" ${CBLAS_FLAG} "" "" false ${float_type})
Expand Down
49 changes: 44 additions & 5 deletions interface/Makefile
View file
Original file line number Diff line number Diff line change
Expand Up @@ -270,7 +270,8 @@ CSBLAS1OBJS = \
cblas_scopy.$(SUFFIX) cblas_sdot.$(SUFFIX) cblas_sdsdot.$(SUFFIX) cblas_dsdot.$(SUFFIX) \
cblas_srot.$(SUFFIX) cblas_srotg.$(SUFFIX) cblas_srotm.$(SUFFIX) cblas_srotmg.$(SUFFIX) \
cblas_sscal.$(SUFFIX) cblas_sswap.$(SUFFIX) cblas_snrm2.$(SUFFIX) cblas_saxpby.$(SUFFIX) \
cblas_ismin.$(SUFFIX) cblas_ismax.$(SUFFIX) cblas_ssum.$(SUFFIX)
cblas_ismin.$(SUFFIX) cblas_ismax.$(SUFFIX) cblas_ssum.$(SUFFIX) cblas_samax.$(SUFFIX) \
cblas_samin.$(SUFFIX)

CSBLAS2OBJS = \
cblas_sgemv.$(SUFFIX) cblas_sger.$(SUFFIX) cblas_ssymv.$(SUFFIX) cblas_strmv.$(SUFFIX) \
Expand All @@ -295,7 +296,8 @@ CDBLAS1OBJS = \
cblas_dcopy.$(SUFFIX) cblas_ddot.$(SUFFIX) \
cblas_drot.$(SUFFIX) cblas_drotg.$(SUFFIX) cblas_drotm.$(SUFFIX) cblas_drotmg.$(SUFFIX) \
cblas_dscal.$(SUFFIX) cblas_dswap.$(SUFFIX) cblas_dnrm2.$(SUFFIX) cblas_daxpby.$(SUFFIX) \
cblas_idmin.$(SUFFIX) cblas_idmax.$(SUFFIX) cblas_dsum.$(SUFFIX)
cblas_idmin.$(SUFFIX) cblas_idmax.$(SUFFIX) cblas_dsum.$(SUFFIX) cblas_damax.$(SUFFIX) \
cblas_damin.$(SUFFIX)

CDBLAS2OBJS = \
cblas_dgemv.$(SUFFIX) cblas_dger.$(SUFFIX) cblas_dsymv.$(SUFFIX) cblas_dtrmv.$(SUFFIX) \
Expand All @@ -315,7 +317,7 @@ CCBLAS1OBJS = \
cblas_cdotc_sub.$(SUFFIX) cblas_cdotu_sub.$(SUFFIX) \
cblas_cscal.$(SUFFIX) cblas_csscal.$(SUFFIX) \
cblas_cswap.$(SUFFIX) cblas_scnrm2.$(SUFFIX) \
cblas_caxpby.$(SUFFIX) \
cblas_caxpby.$(SUFFIX) cblas_scamax.$(SUFFIX) cblas_caxpyc.$(SUFFIX) cblas_scamin.$(SUFFIX) \
cblas_icmin.$(SUFFIX) cblas_icmax.$(SUFFIX) cblas_scsum.$(SUFFIX) cblas_csrot.$(SUFFIX) cblas_crotg.$(SUFFIX)

CCBLAS2OBJS = \
Expand All @@ -340,12 +342,12 @@ CXERBLAOBJ = \

CZBLAS1OBJS = \
cblas_izamax.$(SUFFIX) cblas_izamin.$(SUFFIX) cblas_dzasum.$(SUFFIX) cblas_zaxpy.$(SUFFIX) \
cblas_zcopy.$(SUFFIX) \
cblas_zcopy.$(SUFFIX) cblas_dzamax.$(SUFFIX) cblas_dzamin.$(SUFFIX) \
cblas_zdotc.$(SUFFIX) cblas_zdotu.$(SUFFIX) \
cblas_zdotc_sub.$(SUFFIX) cblas_zdotu_sub.$(SUFFIX) \
cblas_zscal.$(SUFFIX) cblas_zdscal.$(SUFFIX) \
cblas_zswap.$(SUFFIX) cblas_dznrm2.$(SUFFIX) \
cblas_zaxpby.$(SUFFIX) \
cblas_zaxpby.$(SUFFIX) cblas_zaxpyc.$(SUFFIX) \
cblas_izmin.$(SUFFIX) cblas_izmax.$(SUFFIX) cblas_dzsum.$(SUFFIX) cblas_zdrot.$(SUFFIX) cblas_zrotg.$(SUFFIX)


Expand Down Expand Up @@ -1533,6 +1535,30 @@ cblas_icmin.$(SUFFIX) cblas_icmin.$(PSUFFIX) : imax.c
cblas_izmin.$(SUFFIX) cblas_izmin.$(PSUFFIX) : imax.c
$(CC) $(CFLAGS) -DCBLAS -c -UUSE_ABS -DUSE_MIN $< -o $(@F)

cblas_samax.$(SUFFIX) cblas_samax.$(PSUFFIX) : max.c
$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -UUSE_MIN $< -o $(@F)

cblas_damax.$(SUFFIX) cblas_damax.$(PSUFFIX) : max.c
$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -UUSE_MIN $< -o $(@F)

cblas_scamax.$(SUFFIX) cblas_scamax.$(PSUFFIX) : max.c
$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -UUSE_MIN $< -o $(@F)

cblas_dzamax.$(SUFFIX) cblas_dzamax.$(PSUFFIX) : max.c
$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -UUSE_MIN $< -o $(@F)

cblas_samin.$(SUFFIX) cblas_samin.$(PSUFFIX) : max.c
$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -DUSE_MIN $< -o $(@F)

cblas_damin.$(SUFFIX) cblas_damin.$(PSUFFIX) : max.c
$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -DUSE_MIN $< -o $(@F)

cblas_scamin.$(SUFFIX) cblas_scamin.$(PSUFFIX) : max.c
$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -DUSE_MIN $< -o $(@F)

cblas_dzamin.$(SUFFIX) cblas_dzamin.$(PSUFFIX) : max.c
$(CC) $(CFLAGS) -DCBLAS -c -DUSE_ABS -DUSE_MIN $< -o $(@F)

cblas_sasum.$(SUFFIX) cblas_sasum.$(PSUFFIX) : asum.c
$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)

Expand Down Expand Up @@ -1627,6 +1653,19 @@ cblas_daxpy.$(SUFFIX) cblas_daxpy.$(PSUFFIX) : axpy.c
cblas_caxpy.$(SUFFIX) cblas_caxpy.$(PSUFFIX) : zaxpy.c
$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)

cblas_caxpyc.$(SUFFIX) cblas_caxpyc.$(PSUFFIX) : zaxpy.c
$(CC) $(CFLAGS) -DCBLAS -c -DCONJ $< -o $(@F)

cblas_zaxpyc.$(SUFFIX) cblas_zaxpyc.$(PSUFFIX) : zaxpy.c
$(CC) $(CFLAGS) -DCBLAS -c -DCONJ $< -o $(@F)

cblas_xaxpyc.$(SUFFIX) cblas_xaxpyc.$(PSUFFIX) : zaxpy.c
$(CC) $(CFLAGS) -DCBLAS -c -DCONJ $< -o $(@F)

sscal.$(SUFFIX) sscal.$(PSUFFIX) : scal.c
$(CC) $(CFLAGS) -c $< -o $(@F)

dscal.$(SUFFIX) dscal.$(PSUFFIX) : scal.c
cblas_zaxpy.$(SUFFIX) cblas_zaxpy.$(PSUFFIX) : zaxpy.c
$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)

Expand Down
7 changes: 6 additions & 1 deletion interface/max.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -145,8 +145,13 @@ FLOATRET NAME(blasint *N, FLOAT *x, blasint *INCX){

#else

#ifdef COMPLEX
FLOAT CNAME(blasint n, void *vx, blasint incx){
FLOAT *x = (FLOAT*) vx;
#else
FLOAT CNAME(blasint n, FLOAT *x, blasint incx){

#endif

FLOAT ret;

PRINT_DEBUG_CNAME;
Expand Down
2 changes: 2 additions & 0 deletions kernel/loongarch64/KERNEL.LOONGSON2K1000
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,10 +14,12 @@ ZSCALKERNEL = cscal_lsx.S
SAMAXKERNEL = amax_lsx.S
DAMAXKERNEL = amax_lsx.S
CAMAXKERNEL = camax_lsx.S
ZAMAXKERNEL = camax_lsx.S

SAMINKERNEL = amin_lsx.S
DAMINKERNEL = amin_lsx.S
CAMINKERNEL = camin_lsx.S
ZAMINKERNEL = camin_lsx.S

SMAXKERNEL = max_lsx.S
DMAXKERNEL = max_lsx.S
Expand Down
2 changes: 2 additions & 0 deletions kernel/loongarch64/KERNEL.LOONGSON3R5
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,10 +14,12 @@ ZSCALKERNEL = cscal_lasx.S
SAMAXKERNEL = amax_lasx.S
DAMAXKERNEL = amax_lasx.S
CAMAXKERNEL = camax_lasx.S
ZAMAXKERNEL = camax_lasx.S

SAMINKERNEL = amin_lasx.S
DAMINKERNEL = amin_lasx.S
CAMINKERNEL = camin_lasx.S
ZAMINKERNEL = camin_lasx.S

SMAXKERNEL = max_lsx.S
DMAXKERNEL = max_lsx.S
Expand Down
1 change: 0 additions & 1 deletion kernel/loongarch64/amin_lasx.S
View file
Original file line number Diff line number Diff line change
Expand Up @@ -66,7 +66,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#else
xvldrepl.w VM0, X, 0
#endif
XVFSUB VM0, VM0, VM0
bne INCX, TEMP, .L20

srai.d I, N, 4
Expand Down
1 change: 0 additions & 1 deletion kernel/loongarch64/amin_lsx.S
View file
Original file line number Diff line number Diff line change
Expand Up @@ -66,7 +66,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#else
vldrepl.w VM0, X, 0
#endif
VFSUB VM0, VM0, VM0
bne INCX, TEMP, .L20

srai.d I, N, 3
Expand Down
150 changes: 84 additions & 66 deletions kernel/loongarch64/camax_lasx.S
View file
Original file line number Diff line number Diff line change
Expand Up @@ -63,42 +63,60 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
bge $r0, N, .L999
bge $r0, INCX, .L999
li.d TEMP, 1
li.w I, -1
slli.d TEMP, TEMP, ZBASE_SHIFT
slli.d INCX, INCX, ZBASE_SHIFT
xvreplgr2vr.w neg1, I
xvffint.s.w neg1, neg1
srai.d I, N, 3
bne INCX, TEMP, .L20
bge $r0, I, .L23
.align 3

.L10:
xvld VX0, X, 0 * SIZE
xvld VX1, X, 8 * SIZE
addi.d I, I, -1
xvld VX0, X, 0
xvld VX1, X, 32
#ifdef DOUBLE
xvpickev.d x1, VX1, VX0
xvpickod.d x2, VX1, VX0
#else
xvpickev.w x1, VX1, VX0
xvpickod.w x2, VX1, VX0
xvfmul.s x3, neg1, x1
xvfmul.s x4, neg1, x2
xvfcmp.clt.s VT0, x1, res0
xvfcmp.clt.s VT1, x2, res0
xvbitsel.v x1, x1, x3, VT0
xvbitsel.v x2, x2, x4, VT1
#endif
XVFSUB x3, res0, x1
XVFSUB x4, res0, x2
XVFMAX x1, x1, x3
XVFMAX x2, x2, x4
XVFADD VM1, x1, x2
XVFMAX VM0, VM0, VM1
#ifdef DOUBLE
xvld VX0, X, 64
xvld VX1, X, 96
xvpickev.d x1, VX1, VX0
xvpickod.d x2, VX1, VX0
XVFSUB x3, res0, x1
XVFSUB x4, res0, x2
XVFMAX x1, x1, x3
XVFMAX x2, x2, x4
XVFADD VM1, x1, x2
XVFMAX VM0, VM0, VM1
#endif
addi.d I, I, -1
addi.d X, X, 16 * SIZE
xvfadd.s VM1, x1, x2
xvfmax.s VM0, VM0, VM1
blt $r0, I, .L10
.align 3

.L11:
#ifdef DOUBLE
xvpickve.d x1, VM0, 0
xvpickve.d x2, VM0, 1
XVFMAX VM0, x1, x2
#else
xvpickve.w x1, VM0, 0
xvpickve.w x2, VM0, 1
xvpickve.w x3, VM0, 2
xvpickve.w x4, VM0, 3
xvfmax.s VM1, x1, x2
xvfmax.s VM0, x3, x4
xvfmax.s VM0, VM0, VM1
XVFMAX VM0, x1, x2
XVFMAX VM1, x3, x4
XVFMAX VM0, VM0, VM1
#endif
b .L23
.align 3

Expand All @@ -107,66 +125,66 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
.align 3

.L21:
fld.s t1, X, 0 * SIZE
fld.s t2, X, 1 * SIZE
LD t1, X, 0 * SIZE
LD t2, X, 1 * SIZE
add.d X, X, INCX
fld.s t3, X, 0 * SIZE
fld.s t4, X, 1 * SIZE
LD t3, X, 0 * SIZE
LD t4, X, 1 * SIZE
add.d X, X, INCX
fabs.s t1, t1
fabs.s t2, t2
fabs.s t3, t3
fabs.s t4, t4
fadd.s t1, t1, t2
fadd.s t3, t3, t4
fmax.s s1, t1, t3
fld.s t1, X, 0 * SIZE
fld.s t2, X, 1 * SIZE
FABS t1, t1
FABS t2, t2
FABS t3, t3
FABS t4, t4
ADD t1, t1, t2
ADD t3, t3, t4
FMAX s1, t1, t3
LD t1, X, 0 * SIZE
LD t2, X, 1 * SIZE
add.d X, X, INCX
fld.s t3, X, 0 * SIZE
fld.s t4, X, 1 * SIZE
LD t3, X, 0 * SIZE
LD t4, X, 1 * SIZE
add.d X, X, INCX
fabs.s t1, t1
fabs.s t2, t2
fabs.s t3, t3
fabs.s t4, t4
fadd.s t1, t1, t2
fadd.s t3, t3, t4
fmax.s s1, t1, t3
fld.s t1, X, 0 * SIZE
fld.s t2, X, 1 * SIZE
FABS t1, t1
FABS t2, t2
FABS t3, t3
FABS t4, t4
ADD t1, t1, t2
ADD t3, t3, t4
FMAX s1, t1, t3
LD t1, X, 0 * SIZE
LD t2, X, 1 * SIZE
add.d X, X, INCX
fld.s t3, X, 0 * SIZE
fld.s t4, X, 1 * SIZE
LD t3, X, 0 * SIZE
LD t4, X, 1 * SIZE
add.d X, X, INCX
fabs.s t1, t1
fabs.s t2, t2
fabs.s t3, t3
fabs.s t4, t4
FABS t1, t1
FABS t2, t2
FABS t3, t3
FABS t4, t4
addi.d I, I, -1
fadd.s t1, t1, t2
fadd.s t3, t3, t4
fmax.s s3, t1, t3
fld.s t1, X, 0 * SIZE
fld.s t2, X, 1 * SIZE
ADD t1, t1, t2
ADD t3, t3, t4
FMAX s3, t1, t3
LD t1, X, 0 * SIZE
LD t2, X, 1 * SIZE
add.d X, X, INCX
fld.s t3, X, 0 * SIZE
fld.s t4, X, 1 * SIZE
LD t3, X, 0 * SIZE
LD t4, X, 1 * SIZE
add.d X, X, INCX
fabs.s t1, t1
fabs.s t2, t2
fabs.s t3, t3
fabs.s t4, t4
fadd.s t1, t1, t2
fadd.s t3, t3, t4
fmax.s s4, t1, t3
FABS t1, t1
FABS t2, t2
FABS t3, t3
FABS t4, t4
ADD t1, t1, t2
ADD t3, t3, t4
FMAX s4, t1, t3
blt $r0, I, .L21
.align 3

.L22:
fmax.s s1, s1, s2
fmax.s s3, s3, s4
fmax.s s1, s1, s3
FMAX s1, s1, s2
FMAX s3, s3, s4
FMAX s1, s1, s3
.align 3

.L23: //N<8
Expand All @@ -182,12 +200,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
FABS a1, a1
ADD a0, a0, a1
add.d X, X, INCX
fmax.s s1, a0, s1
FMAX s1, a0, s1
blt $r0, I, .L24
.align 3

.L999:
fmov.s $f0, $f22
MOV $f0, $f22
jirl $r0, $r1, 0x0
.align 3

Expand Down
Loading

0 comments on commit 3ffd686

Please sign in to comment.