Merge pull request #4472 from sergei-lewis/dev/slewis/merge-from-riscv

Merge risc-v branch to develop

Author: martin-frbg

Makefile.prebuild

@@ -59,6 +59,22 @@ ifeq ($(TARGET), CK860FV)
 TARGET_FLAGS = -march=ck860v -mcpu=ck860fv -mfdivdu -mhard-float
 endif
 
+ifeq ($(TARGET), x280)
+TARGET_FLAGS = -march=rv64imafdcv_zba_zbb_zfh -mabi=lp64d
+endif
+
+ifeq ($(TARGET), RISCV64_ZVL256B)
+TARGET_FLAGS = -march=rv64imafdcv -mabi=lp64d
+endif
+
+ifeq ($(TARGET), RISCV64_ZVL128B)
+TARGET_FLAGS = -march=rv64imafdcv -mabi=lp64d
+endif
+
+ifeq ($(TARGET), RISCV64_GENERIC)
+TARGET_FLAGS = -march=rv64imafdc -mabi=lp64d
+endif
+
 all: getarch_2nd
 	./getarch_2nd  0 >> $(TARGET_MAKE)
 	./getarch_2nd  1 >> $(TARGET_CONF)

Makefile.riscv64

@@ -2,3 +2,19 @@ ifeq ($(CORE), C910V)
 CCOMMON_OPT += -march=rv64imafdcv0p7_zfh_xtheadc -mabi=lp64d -mtune=c920
 FCOMMON_OPT += -march=rv64imafdcv0p7_zfh_xtheadc -mabi=lp64d -mtune=c920 -static
 endif
+ifeq ($(CORE), x280)
+CCOMMON_OPT += -march=rv64imafdcv_zba_zbb_zfh_zvl512b -mabi=lp64d -ffast-math
+FCOMMON_OPT += -march=rv64imafdcv_zba_zbb_zfh -mabi=lp64d -static
+endif
+ifeq ($(CORE), RISCV64_ZVL256B)
+CCOMMON_OPT += -march=rv64imafdcv_zvl256b -mabi=lp64d
+FCOMMON_OPT += -march=rv64imafdcv -mabi=lp64d -static
+endif
+ifeq ($(CORE), RISCV64_ZVL128B)
+CCOMMON_OPT += -march=rv64imafdcv -mabi=lp64d 
+FCOMMON_OPT += -march=rv64imafdcv -mabi=lp64d -static
+endif
+ifeq ($(CORE), RISCV64_GENERIC)
+CCOMMON_OPT += -march=rv64imafdc -mabi=lp64d
+FCOMMON_OPT += -march=rv64imafdc -mabi=lp64d -static
+endif

README.md

@@ -198,6 +198,11 @@ Please read `GotoBLAS_01Readme.txt` for older CPU models already supported by th
   ```
   (also known to work on C906 as long as you use only single-precision functions - its instruction set support appears to be incomplete in double precision)
 
+- **x280**: Level-3 BLAS and Level-1,2 are optimized by RISC-V Vector extension 1.0.
+  ```sh
+  make HOSTCC=gcc TARGET=x280 NUM_THREADS=8 CC=riscv64-unknown-linux-gnu-clang FC=riscv64-unknown-linux-gnu-gfortran
+  ```
+
 ### Support for multiple targets in a single library
 
 OpenBLAS can be built for multiple targets with runtime detection of the target cpu by specifiying `DYNAMIC_ARCH=1` in Makefile.rule, on the gmake command line or as `-DDYNAMIC_ARCH=TRUE` in cmake.

TargetList.txt

@@ -118,8 +118,11 @@ Z13
 Z14
 
 10.RISC-V 64:
-RISCV64_GENERIC
+RISCV64_GENERIC (e.g. PolarFire Soc/SiFive U54)
+RISCV64_ZVL128B
 C910V
+x280
+RISCV64_ZVL256B
 
 11.LOONGARCH64:
 LOONGSONGENERIC

benchmark/Makefile

@@ -37,6 +37,12 @@ ESSL=/opt/ibm/lib
 #LIBESSL = -lesslsmp $(ESSL)/libxlomp_ser.so.1 $(ESSL)/libxlf90_r.so.1 $(ESSL)/libxlfmath.so.1 $(ESSL)/libxlsmp.so.1 /opt/ibm/xlC/13.1.3/lib/libxl.a
 LIBESSL = -lesslsmp  $(ESSL)/libxlf90_r.so.1 $(ESSL)/libxlfmath.so.1 $(ESSL)/libxlsmp.so.1 /opt/ibm/xlC/13.1.3/lib/libxl.a
 
+# x280 temporary workaround for gfortran
+ifeq ($(TARGET), x280)
+CCOMMON_OPT:=$(filter-out -mllvm --riscv-v-vector-bits-min=512,$(CCOMMON_OPT))
+endif
+
+
 ifneq ($(NO_LAPACK), 1)
 GOTO_LAPACK_TARGETS=slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
 		    scholesky.goto dcholesky.goto ccholesky.goto zcholesky.goto \
@@ -265,9 +271,9 @@ goto :: sgemm.goto dgemm.goto cgemm.goto zgemm.goto \
        ismax.goto idmax.goto \
        isamin.goto idamin.goto icamin.goto izamin.goto \
        ismin.goto idmin.goto \
-       samax.goto damax.goto scamax.goto dzamax.goto \
+       samax.goto damax.goto camax.goto zamax.goto \
        smax.goto dmax.goto \
-       samin.goto damin.goto scamin.goto dzamin.goto \
+       samin.goto damin.goto camin.goto zamin.goto \
        smin.goto dmin.goto \
        saxpby.goto daxpby.goto caxpby.goto zaxpby.goto \
        snrm2.goto dnrm2.goto scnrm2.goto dznrm2.goto $(GOTO_LAPACK_TARGETS) $(GOTO_HALF_TARGETS)
@@ -2832,12 +2838,12 @@ samax.goto : samax.$(SUFFIX) ../$(LIBNAME)
 damax.goto : damax.$(SUFFIX) ../$(LIBNAME)
 	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
-############################################## SCAMAX ##############################################
-scamax.goto : scamax.$(SUFFIX) ../$(LIBNAME)
+############################################## CAMAX ##############################################
+camax.goto : camax.$(SUFFIX) ../$(LIBNAME)
 	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
-############################################## DZAMAX ##############################################
-dzamax.goto : dzamax.$(SUFFIX) ../$(LIBNAME)
+############################################## ZAMAX ##############################################
+zamax.goto : zamax.$(SUFFIX) ../$(LIBNAME)
 	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
 ############################################## SMAX ##############################################
@@ -2856,12 +2862,12 @@ samin.goto : samin.$(SUFFIX) ../$(LIBNAME)
 damin.goto : damin.$(SUFFIX) ../$(LIBNAME)
 	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
-############################################## SCAMIN ##############################################
-scamin.goto : scamin.$(SUFFIX) ../$(LIBNAME)
+############################################## CAMIN ##############################################
+camin.goto : camin.$(SUFFIX) ../$(LIBNAME)
 	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
-############################################## DZAMIN ##############################################
-dzamin.goto : dzamin.$(SUFFIX) ../$(LIBNAME)
+############################################## ZAMIN ##############################################
+zamin.goto : zamin.$(SUFFIX) ../$(LIBNAME)
 	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 
 ############################################## SMIN ##############################################
@@ -3383,10 +3389,10 @@ samax.$(SUFFIX) : amax.c
 damax.$(SUFFIX) : amax.c
 	$(CC) $(CFLAGS) -c -UCOMPLEX  -DDOUBLE -o $(@F) $^
 
-scamax.$(SUFFIX) : amax.c
+camax.$(SUFFIX) : amax.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX  -UDOUBLE -o $(@F) $^
 
-dzamax.$(SUFFIX) : amax.c
+zamax.$(SUFFIX) : amax.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX  -DDOUBLE -o $(@F) $^
 
 
@@ -3403,10 +3409,10 @@ samin.$(SUFFIX) : amin.c
 damin.$(SUFFIX) : amin.c
 	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
 
-scamin.$(SUFFIX) : amin.c
+camin.$(SUFFIX) : amin.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^
 
-dzamin.$(SUFFIX) : amin.c
+zamin.$(SUFFIX) : amin.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
 
 
@@ -3436,4 +3442,4 @@ smallscaling: smallscaling.c ../$(LIBNAME)
 clean ::
 	@rm -f *.goto *.mkl *.acml *.atlas *.veclib *.essl smallscaling
 
-include $(TOPDIR)/Makefile.tail
+include $(TOPDIR)/Makefile.tail

cblas.h

@@ -303,6 +303,15 @@ void cblas_zgemm(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLA
 void cblas_zgemm3m(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransA, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransB, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint N, OPENBLAS_CONST blasint K,
 		 OPENBLAS_CONST void *alpha, OPENBLAS_CONST void *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST void *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST void *beta, void *C, OPENBLAS_CONST blasint ldc);
 
+void cblas_sgemmt(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_UPLO Uplo, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransA, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransB, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint K,
+		 OPENBLAS_CONST float alpha, OPENBLAS_CONST float *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST float *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST float beta, float *C, OPENBLAS_CONST blasint ldc);
+void cblas_dgemmt(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_UPLO Uplo, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransA, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransB, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint K,
+		 OPENBLAS_CONST double alpha, OPENBLAS_CONST double *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST double *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST double beta, double *C, OPENBLAS_CONST blasint ldc);
+void cblas_cgemmt(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_UPLO Uplo, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransA, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransB, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint K,
+		 OPENBLAS_CONST void *alpha, OPENBLAS_CONST void *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST void *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST void *beta, void *C, OPENBLAS_CONST blasint ldc);
+void cblas_zgemmt(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_UPLO Uplo, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransA, OPENBLAS_CONST enum CBLAS_TRANSPOSE TransB, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint K,
+		 OPENBLAS_CONST void *alpha, OPENBLAS_CONST void *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST void *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST void *beta, void *C, OPENBLAS_CONST blasint ldc);
+
 void cblas_ssymm(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_SIDE Side, OPENBLAS_CONST enum CBLAS_UPLO Uplo, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint N,
                  OPENBLAS_CONST float alpha, OPENBLAS_CONST float *A, OPENBLAS_CONST blasint lda, OPENBLAS_CONST float *B, OPENBLAS_CONST blasint ldb, OPENBLAS_CONST float beta, float *C, OPENBLAS_CONST blasint ldc);
 void cblas_dsymm(OPENBLAS_CONST enum CBLAS_ORDER Order, OPENBLAS_CONST enum CBLAS_SIDE Side, OPENBLAS_CONST enum CBLAS_UPLO Uplo, OPENBLAS_CONST blasint M, OPENBLAS_CONST blasint N,

common_interface.h

@@ -498,6 +498,15 @@ void BLASFUNC(zgemm3m)(char *, char *, blasint *, blasint *, blasint *, double *
 void BLASFUNC(xgemm3m)(char *, char *, blasint *, blasint *, blasint *, xdouble *,
 	   xdouble *, blasint *, xdouble *, blasint *, xdouble *, xdouble *, blasint *);
 
+void BLASFUNC(sgemmt)(char*, char *, char *, blasint *, blasint *, float *,
+	   float  *, blasint *, float  *, blasint *, float  *, float  *, blasint *);
+void BLASFUNC(dgemmt)(char*, char *, char *, blasint *, blasint *, double *,
+	   double *, blasint *, double *, blasint *, double *, double *, blasint *);
+void BLASFUNC(cgemmt)(char*, char *, char *, blasint *, blasint *, float *,
+	   float  *, blasint *, float  *, blasint *, float  *, float  *, blasint *);
+void BLASFUNC(zgemmt)(char*, char *, char *, blasint *, blasint *, double *,
+	   double *, blasint *, double *, blasint *, double *, double *, blasint *);
+
 int BLASFUNC(sge2mm)(char *, char *, char *, blasint *, blasint *,
 		     float *, float  *, blasint *, float  *, blasint *,
 		     float *, float  *, blasint *);

common_riscv64.h

@@ -91,8 +91,26 @@ static inline int blas_quickdivide(blasint x, blasint y){
 #define BUFFER_SIZE     ( 32 << 20)
 #define SEEK_ADDRESS
 
-#if defined(C910V)
-#include <riscv_vector.h>
+#if defined(C910V) || (defined(RISCV64_ZVL256B) && (defined(__clang__) || defined(RVV_COMPATIBLE_GCC))) || defined(RISCV64_ZVL128B) || defined(x280)
+# include <riscv_vector.h>
+#endif
+
+#if defined( __riscv_xtheadc ) && defined( __riscv_v ) && ( __riscv_v <= 7000 )
+// t-head toolchain uses obsolete rvv intrinsics, can't build for C910V without this
+#define RISCV_0p10_INTRINSICS
+#define RISCV_RVV(x) x
+#else
+#define RISCV_RVV(x) __riscv_ ## x
+#endif
+
+#if defined(C910V) || defined(RISCV64_ZVL256B)
+# if !defined(DOUBLE)
+#  define EXTRACT_FLOAT(v) RISCV_RVV(vfmv_f_s_f32m1_f32)(v)
+# else
+#  define EXTRACT_FLOAT(v) RISCV_RVV(vfmv_f_s_f64m1_f64)(v)
+# endif
+#else
+# define EXTRACT_FLOAT(v) (v[0])
 #endif
 
 #endif

cpuid_riscv64.c

@@ -70,12 +70,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 /* or implied, of The University of Texas at Austin.                 */
 /*********************************************************************/
 
-#define CPU_GENERIC   0
-#define CPU_C910V     1
+#define CPU_GENERIC         0
+#define CPU_C910V           1
+#define CPU_x280            2
+#define CPU_RISCV64_ZVL256B 3
+#define CPU_RISCV64_ZVL128B 4
 
 static char *cpuname[] = {
   "RISCV64_GENERIC",
-  "C910V"
+  "C910V",
+  "x280",
+  "CPU_RISCV64_ZVL256B",
+  "CPU_RISCV64_ZVL128B"
 };
 
 int detect(void){

ctest/c_cblat1.f

@@ -96,7 +96,7 @@ SUBROUTINE CHECK1(SFAC)
       INTEGER           ICAMAXTEST
       EXTERNAL          SCASUMTEST, SCNRM2TEST, ICAMAXTEST
 *     .. External Subroutines ..
-      EXTERNAL          CSCAL, CSSCALTEST, CTEST, ITEST1, STEST1
+      EXTERNAL          CSCALTEST, CSSCALTEST, CTEST, ITEST1, STEST1
 *     .. Intrinsic Functions ..
       INTRINSIC         MAX
 *     .. Common blocks ..
@@ -214,8 +214,8 @@ SUBROUTINE CHECK1(SFAC)
                CALL STEST1(SCASUMTEST(N,CX,INCX),STRUE4(NP1),
      +                     STRUE4(NP1),SFAC)
             ELSE IF (ICASE.EQ.8) THEN
-*              .. CSCAL ..
-               CALL CSCAL(N,CA,CX,INCX)
+*              .. CSCALTEST ..
+               CALL CSCALTEST(N,CA,CX,INCX)
                CALL CTEST(LEN,CX,CTRUE5(1,NP1,INCX),CTRUE5(1,NP1,INCX),
      +                    SFAC)
             ELSE IF (ICASE.EQ.9) THEN
@@ -236,14 +236,14 @@ SUBROUTINE CHECK1(SFAC)
 *
       INCX = 1
       IF (ICASE.EQ.8) THEN
-*        CSCAL
+*        CSCALTEST
 *        Add a test for alpha equal to zero.
          CA = (0.0E0,0.0E0)
          DO 80 I = 1, 5
             MWPCT(I) = (0.0E0,0.0E0)
             MWPCS(I) = (1.0E0,1.0E0)
    80    CONTINUE
-         CALL CSCAL(5,CA,CX,INCX)
+         CALL CSCALTEST(5,CA,CX,INCX)
          CALL CTEST(5,CX,MWPCT,MWPCS,SFAC)
       ELSE IF (ICASE.EQ.9) THEN
 *        CSSCALTEST