Add SVE implementation for sdot/ddot

This adds an SVE implementation to sdot/ddot when available, falling back to the previous Advanced SIMD kernel where there's no SVE implementation for the kernel.

All the targets were essentially treating `dot_thunderx2t99.c` as the Advanced SIMD implementation so I've renamed it to better fit with the feature detection.

Makefile.arm64

@@ -70,12 +70,12 @@ endif
 ifeq ($(CORE), NEOVERSEN1)
 ifeq (1, $(filter 1,$(GCCVERSIONGTEQ7) $(ISCLANG)))
 ifeq ($(GCCVERSIONGTEQ9), 1)
-CCOMMON_OPT += -march=armv8.2-a -mtune=neoverse-n1
+CCOMMON_OPT += -march=armv8.2-a+sve -mtune=neoverse-n1
 ifneq ($(F_COMPILER), NAG)
 FCOMMON_OPT += -march=armv8.2-a -mtune=neoverse-n1
 endif
 else
-CCOMMON_OPT += -march=armv8.2-a -mtune=cortex-a72
+CCOMMON_OPT += -march=armv8.2-a+sve -mtune=cortex-a72
 ifneq ($(F_COMPILER), NAG)
 FCOMMON_OPT += -march=armv8.2-a -mtune=cortex-a72
 endif
@@ -94,12 +94,12 @@ ifeq ($(CORE), NEOVERSEV1)
 ifeq (1, $(filter 1,$(GCCVERSIONGTEQ7) $(ISCLANG)))
 ifeq ($(GCCVERSIONGTEQ10), 1)
 ifeq (1, $(filter 1,$(GCCMINORVERSIONGTEQ4) $(GCCVERSIONGTEQ11)))
-CCOMMON_OPT += -march=armv8.4-a -mtune=neoverse-v1
+CCOMMON_OPT += -march=armv8.4-a+sve -mtune=neoverse-v1
 ifneq ($(F_COMPILER), NAG)
 FCOMMON_OPT += -march=armv8.4-a -mtune=neoverse-v1
 endif
 else
-CCOMMON_OPT += -march=armv8.4-a -mtune=native
+CCOMMON_OPT += -march=armv8.4-a+sve -mtune=native
 ifneq ($(F_COMPILER), NAG)
 FCOMMON_OPT += -march=armv8.4-a -mtune=native
 endif
@@ -133,7 +133,7 @@ ifneq ($(F_COMPILER), NAG)
 FCOMMON_OPT += -march=armv8.5-a+sve+sve2+bf16 -mtune=neoverse-n2
 endif
 else
-CCOMMON_OPT += -march=armv8.5-a -mtune=native
+CCOMMON_OPT += -march=armv8.5-a+sve -mtune=native
 ifneq ($(F_COMPILER), NAG)
 FCOMMON_OPT += -march=armv8.5-a -mtune=native
 endif

getarch.c

@@ -1410,7 +1410,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
        "-DL2_SIZE=1048576 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=16 " \
        "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
        "-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DHAVE_SVE -DARMV8 " \
-       "-march=armv8.4-a -mtune=neoverse-v1"
+       "-march=armv8.4-a+sve -mtune=neoverse-v1"
 #define LIBNAME   "neoversev1"
 #define CORENAME  "NEOVERSEV1"
 #endif

kernel/arm64/KERNEL.NEOVERSEN1

@@ -96,8 +96,8 @@ DNRM2KERNEL    = dznrm2_thunderx2t99.c
 CNRM2KERNEL    = scnrm2_thunderx2t99.c
 ZNRM2KERNEL    = dznrm2_thunderx2t99.c
 
-DDOTKERNEL     = dot_thunderx2t99.c
+DDOTKERNEL     = dot.c
-SDOTKERNEL     = dot_thunderx2t99.c
+SDOTKERNEL     = dot.c
 CDOTKERNEL     = zdot_thunderx2t99.c
 ZDOTKERNEL     = zdot_thunderx2t99.c
 DSDOTKERNEL    = dot.S

kernel/arm64/KERNEL.NEOVERSEN2

@@ -96,8 +96,8 @@ DNRM2KERNEL    = dznrm2_thunderx2t99.c
 CNRM2KERNEL    = scnrm2_thunderx2t99.c
 ZNRM2KERNEL    = dznrm2_thunderx2t99.c
 
-DDOTKERNEL     = dot_thunderx2t99.c
+DDOTKERNEL     = dot.c
-SDOTKERNEL     = dot_thunderx2t99.c
+SDOTKERNEL     = dot.c
 CDOTKERNEL     = zdot_thunderx2t99.c
 ZDOTKERNEL     = zdot_thunderx2t99.c
 DSDOTKERNEL    = dot.S

kernel/arm64/KERNEL.NEOVERSEV1

@@ -96,8 +96,8 @@ DNRM2KERNEL    = dznrm2_thunderx2t99.c
 CNRM2KERNEL    = scnrm2_thunderx2t99.c
 ZNRM2KERNEL    = dznrm2_thunderx2t99.c
 
-DDOTKERNEL     = dot_thunderx2t99.c
+DDOTKERNEL     = dot.c
-SDOTKERNEL     = dot_thunderx2t99.c
+SDOTKERNEL     = dot.c
 CDOTKERNEL     = zdot_thunderx2t99.c
 ZDOTKERNEL     = zdot_thunderx2t99.c
 DSDOTKERNEL    = dot.S

kernel/arm64/KERNEL.THUNDERX2T99

@@ -161,8 +161,8 @@ DNRM2KERNEL    = dznrm2_thunderx2t99.c
 ZNRM2KERNEL    = dznrm2_thunderx2t99.c
 
 
-DDOTKERNEL     = dot_thunderx2t99.c
+DDOTKERNEL     = dot.c
-SDOTKERNEL     = dot_thunderx2t99.c
+SDOTKERNEL     = dot.c
 CDOTKERNEL     = zdot_thunderx2t99.c
 ZDOTKERNEL     = zdot_thunderx2t99.c
 DSDOTKERNEL    = dot.S

kernel/arm64/KERNEL.THUNDERX3T110

@@ -161,8 +161,8 @@ DNRM2KERNEL    = dznrm2_thunderx2t99.c
 ZNRM2KERNEL    = dznrm2_thunderx2t99.c
 
 
-DDOTKERNEL     = dot_thunderx2t99.c
+DDOTKERNEL     = dot.c
-SDOTKERNEL     = dot_thunderx2t99.c
+SDOTKERNEL     = dot.c
 CDOTKERNEL     = zdot_thunderx2t99.c
 ZDOTKERNEL     = zdot_thunderx2t99.c
 DSDOTKERNEL    = dot.S

kernel/arm64/dot.c

@@ -0,0 +1,111 @@
+/***************************************************************************
+Copyright (c) 2017, The OpenBLAS Project
+Copyright (c) 2022, Arm Ltd
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+#ifdef HAVE_SVE
+#include "dot_kernel_sve.c"
+#endif
+#include "dot_kernel_asimd.c"
+
+#if defined(SMP)
+extern int blas_level1_thread_with_return_value(int mode, BLASLONG m, BLASLONG n,
+	BLASLONG k, void *alpha, void *a, BLASLONG lda, void *b, BLASLONG ldb,
+	void *c, BLASLONG ldc, int (*function)(), int nthreads);
+#endif
+
+static RETURN_TYPE dot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+{
+	RETURN_TYPE  dot = 0.0 ;
+
+	if ( n <= 0 ) return dot;
+
+#ifdef HAVE_SVE
+	if (inc_x == 1 && inc_y == 1) {
+		return dot_kernel_sve(n, x, y);
+	}
+#endif
+
+	return dot_kernel_asimd(n, x, inc_x, y, inc_y);
+}
+
+#if defined(SMP)
+static int dot_thread_function(BLASLONG n, BLASLONG dummy0,
+	BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *y,
+	BLASLONG inc_y, FLOAT *result, BLASLONG dummy3)
+{
+	*(RETURN_TYPE *)result = dot_compute(n, x, inc_x, y, inc_y);
+
+	return 0;
+}
+#endif
+
+RETURN_TYPE CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+{
+#if defined(SMP)
+	int nthreads;
+	FLOAT dummy_alpha;
+#endif
+	RETURN_TYPE dot = 0.0;
+
+#if defined(SMP)
+	if (inc_x == 0 || inc_y == 0 || n <= 10000)
+		nthreads = 1;
+	else
+		nthreads = num_cpu_avail(1);
+
+	if (nthreads == 1) {
+		dot = dot_compute(n, x, inc_x, y, inc_y);
+	} else {
+		int mode, i;
+		char result[MAX_CPU_NUMBER * sizeof(double) * 2];
+		RETURN_TYPE *ptr;
+
+#if !defined(DOUBLE)
+		mode = BLAS_SINGLE  | BLAS_REAL;
+#else
+		mode = BLAS_DOUBLE  | BLAS_REAL;
+#endif
+
+		blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
+				   x, inc_x, y, inc_y, result, 0,
+				   ( void *)dot_thread_function, nthreads);
+
+		ptr = (RETURN_TYPE *)result;
+		for (i = 0; i < nthreads; i++) {
+			dot = dot + (*ptr);
+			ptr = (RETURN_TYPE *)(((char *)ptr) + sizeof(double) * 2);
+		}
+	}
+#else
+	dot = dot_compute(n, x, inc_x, y, inc_y);
+#endif
+
+	return dot;
+}

kernel/arm64/dot_kernel_asimd.c

@@ -260,18 +260,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	"	faddp	"OUT", v0.2d			\n"
 #endif /* !defined(DOUBLE) */
 
-#if defined(SMP)
+static RETURN_TYPE dot_kernel_asimd(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
-extern int blas_level1_thread_with_return_value(int mode, BLASLONG m, BLASLONG n,
-	BLASLONG k, void *alpha, void *a, BLASLONG lda, void *b, BLASLONG ldb,
-	void *c, BLASLONG ldc, int (*function)(), int nthreads);
-#endif
-
-static RETURN_TYPE dot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 {
 	RETURN_TYPE  dot = 0.0;
-
-	if ( n < 0 ) return dot;
-
 	BLASLONG j = 0;
 
 	__asm__ __volatile__ (
@@ -352,58 +343,3 @@ static RETURN_TYPE dot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, B
 
 	return dot;
 }
-
-#if defined(SMP)
-static int dot_thread_function(BLASLONG n, BLASLONG dummy0,
-	BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *y,
-	BLASLONG inc_y, FLOAT *result, BLASLONG dummy3)
-{
-	*(RETURN_TYPE *)result = dot_compute(n, x, inc_x, y, inc_y);
-
-	return 0;
-}
-#endif
-
-RETURN_TYPE CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
-{
-#if defined(SMP)
-	int nthreads;
-	FLOAT dummy_alpha;
-#endif
-	RETURN_TYPE dot = 0.0;
-
-#if defined(SMP)
-	if (inc_x == 0 || inc_y == 0 || n <= 10000)
-		nthreads = 1;
-	else
-		nthreads = num_cpu_avail(1);
-
-	if (nthreads == 1) {
-		dot = dot_compute(n, x, inc_x, y, inc_y);
-	} else {
-		int mode, i;
-		char result[MAX_CPU_NUMBER * sizeof(double) * 2];
-		RETURN_TYPE *ptr;
-
-#if !defined(DOUBLE)
-		mode = BLAS_SINGLE  | BLAS_REAL;
-#else
-		mode = BLAS_DOUBLE  | BLAS_REAL;
-#endif
-
-		blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
-				   x, inc_x, y, inc_y, result, 0,
-				   ( void *)dot_thread_function, nthreads);
-
-		ptr = (RETURN_TYPE *)result;
-		for (i = 0; i < nthreads; i++) {
-			dot = dot + (*ptr);
-			ptr = (RETURN_TYPE *)(((char *)ptr) + sizeof(double) * 2);
-		}
-	}
-#else
-	dot = dot_compute(n, x, inc_x, y, inc_y);
-#endif
-
-	return dot;
-}

kernel/arm64/dot_kernel_sve.c

@@ -0,0 +1,66 @@
+/***************************************************************************
+Copyright (c) 2022, Arm Ltd
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
+GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF
+THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+#include <arm_sve.h>
+
+#ifdef DOUBLE
+#define SVE_TYPE svfloat64_t
+#define SVE_ZERO svdup_f64(0.0)
+#define SVE_WHILELT svwhilelt_b64
+#define SVE_ALL svptrue_b64()
+#define SVE_WIDTH svcntd()
+#else
+#define SVE_TYPE svfloat32_t
+#define SVE_ZERO svdup_f32(0.0)
+#define SVE_WHILELT svwhilelt_b32
+#define SVE_ALL svptrue_b32()
+#define SVE_WIDTH svcntw()
+#endif
+
+static FLOAT dot_kernel_sve(BLASLONG n, FLOAT *x, FLOAT *y) {
+        SVE_TYPE acc_a = SVE_ZERO;
+        SVE_TYPE acc_b = SVE_ZERO;
+
+        BLASLONG sve_width = SVE_WIDTH;
+
+        for (BLASLONG i = 0; i < n; i += sve_width * 2) {
+                svbool_t pg_a = SVE_WHILELT(i, n);
+                svbool_t pg_b = SVE_WHILELT(i + sve_width, n);
+
+                SVE_TYPE x_vec_a = svld1(pg_a, &x[i]);
+                SVE_TYPE y_vec_a = svld1(pg_a, &y[i]);
+                SVE_TYPE x_vec_b = svld1(pg_b, &x[i + sve_width]);
+                SVE_TYPE y_vec_b = svld1(pg_b, &y[i + sve_width]);
+
+                acc_a = svmla_m(pg_a, acc_a, x_vec_a, y_vec_a);
+                acc_b = svmla_m(pg_b, acc_b, x_vec_b, y_vec_b);
+        }
+
+        return svaddv(SVE_ALL, acc_a) + svaddv(SVE_ALL, acc_b);
+}