THUNDERX2T99: Add threaded DDOT Implementation

kernel/arm64/KERNEL.THUNDERX2T99

@@ -2,7 +2,11 @@ include $(KERNELDIR)/KERNEL.CORTEXA57
 
 DAXPYKERNEL=daxpy_thunderx2t99.S
 
+ifndef SMP
 DDOTKERNEL=ddot_thunderx2t99.S
+else
+DDOTKERNEL=ddot_thunderx2t99.c
+endif
 
 ifeq ($(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N), 8x4)
 DGEMMKERNEL    =  dgemm_kernel_8x4_thunderx2t99.S

kernel/arm64/ddot_thunderx2t99.c

@@ -0,0 +1,258 @@
+/***************************************************************************
+Copyright (c) 2017, The OpenBLAS Project
+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_neon.h>
+
+#define	N	"x0"	/* vector length */
+#define	X	"x1"	/* "X" vector address */
+#define	INC_X	"x2"	/* "X" stride */
+#define	Y	"x3"	/* "Y" vector address */
+#define	INC_Y	"x4"	/* "Y" stride */
+#define J	"x5"	/* loop variable */
+
+#define REG0	"xzr"
+#define DOTF	"d0"
+#define TMPX	"d16"
+#define LD1VX	"{v16.d}[0]"
+#define TMPY	"d24"
+#define LD1VY	"{v24.d}[0]"
+#define SZ	"8"
+
+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);
+
+
+static FLOAT ddot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+{
+	FLOAT  dot = 0.0 ;
+
+	if ( n < 0 )  return(dot);
+
+	__asm__ __volatile__ (
+	"	mov	"N", %[N_]			\n"
+	"	mov	"X", %[X_]			\n"
+	"	mov	"INC_X", %[INCX_]		\n"
+	"	mov	"Y", %[Y_]			\n"
+	"	mov	"INC_Y", %[INCY_]		\n"
+	"	fmov	"DOTF", "REG0"		\n"
+	"	fmov	d1, "REG0"			\n"
+	"	fmov	d2, "REG0"			\n"
+	"	fmov	d3, "REG0"			\n"
+	"	fmov	d4, "REG0"			\n"
+	"	fmov	d5, "REG0"			\n"
+	"	fmov	d6, "REG0"			\n"
+	"	fmov	d7, "REG0"			\n"
+
+	"	cmp	"N", xzr			\n"
+	"	ble	9f	//dot_kernel_L999	\n"
+
+	"	cmp	"INC_X", #1			\n"
+	"	bne	5f	//dot_kernel_S_BEGIN	\n"
+	"	cmp	"INC_Y", #1			\n"
+	"	bne	5f	//dot_kernel_S_BEGIN	\n"
+
+	"1:	//dot_kernel_F_BEGIN			\n"
+	"	asr	"J", "N", #5			\n"
+	"	cmp	"J", xzr			\n"
+	"	beq	3f	//dot_kernel_F1		\n"
+
+	"	.align 5				\n"
+	"2:	//dot_kernel_F32			\n"
+	"	ldp	q16, q17, ["X"]			\n"
+	"	ldp	q24, q25, ["Y"]			\n"
+	"	ldp	q18, q19, ["X", #32]		\n"
+	"	ldp	q26, q27, ["Y", #32]		\n"
+	"	fmla	v0.2d, v16.2d, v24.2d		\n"
+	"	fmla	v1.2d, v17.2d, v25.2d		\n"
+	"	ldp	q20, q21, ["X", #64]		\n"
+	"	ldp	q28, q29, ["Y", #64]		\n"
+	"	fmla	v2.2d, v18.2d, v26.2d		\n"
+	"	fmla	v3.2d, v19.2d, v27.2d		\n"
+	"	ldp	q22, q23, ["X", #96]		\n"
+	"	ldp	q30, q31, ["Y", #96]		\n"
+	"	add	"Y", "Y", #128			\n"
+	"	add	"X", "X", #128			\n"
+	"	fmla	v4.2d, v20.2d, v28.2d		\n"
+	"	fmla	v5.2d, v21.2d, v29.2d		\n"
+	"	PRFM	PLDL1KEEP, ["X", #896]		\n"
+	"	PRFM	PLDL1KEEP, ["Y", #896]		\n"
+	"	PRFM	PLDL1KEEP, ["X", #896+64]	\n"
+	"	PRFM	PLDL1KEEP, ["Y", #896+64]	\n"
+	"	fmla	v6.2d, v22.2d, v30.2d		\n"
+	"	fmla	v7.2d, v23.2d, v31.2d		\n"
+
+	"	ldp	q16, q17, ["X"]			\n"
+	"	ldp	q24, q25, ["Y"]			\n"
+	"	ldp	q18, q19, ["X", #32]		\n"
+	"	ldp	q26, q27, ["Y", #32]		\n"
+	"	fmla	v0.2d, v16.2d, v24.2d		\n"
+	"	fmla	v1.2d, v17.2d, v25.2d		\n"
+	"	ldp	q20, q21, ["X", #64]		\n"
+	"	ldp	q28, q29, ["Y", #64]		\n"
+	"	fmla	v2.2d, v18.2d, v26.2d		\n"
+	"	fmla	v3.2d, v19.2d, v27.2d		\n"
+	"	ldp	q22, q23, ["X", #96]		\n"
+	"	ldp	q30, q31, ["Y", #96]		\n"
+	"	add	"Y", "Y", #128			\n"
+	"	add	"X", "X", #128			\n"
+	"	fmla	v4.2d, v20.2d, v28.2d		\n"
+	"	fmla	v5.2d, v21.2d, v29.2d		\n"
+	"	PRFM	PLDL1KEEP, ["X", #896]		\n"
+	"	PRFM	PLDL1KEEP, ["Y", #896]		\n"
+	"	PRFM	PLDL1KEEP, ["X", #896+64]	\n"
+	"	PRFM	PLDL1KEEP, ["Y", #896+64]	\n"
+	"	fmla	v6.2d, v22.2d, v30.2d		\n"
+	"	fmla	v7.2d, v23.2d, v31.2d		\n"
+
+	"	subs	"J", "J", #1			\n"
+	"	bne	2b	//dot_kernel_F32	\n"
+
+	"	fadd	v0.2d, v0.2d, v1.2d		\n"
+	"	fadd	v2.2d, v2.2d, v3.2d		\n"
+	"	fadd	v4.2d, v4.2d, v5.2d		\n"
+	"	fadd	v6.2d, v6.2d, v7.2d		\n"
+	"	fadd	v0.2d, v0.2d, v2.2d		\n"
+	"	fadd	v4.2d, v4.2d, v6.2d		\n"
+	"	fadd	v0.2d, v0.2d, v4.2d		\n"
+	"	faddp	"DOTF", v0.2d			\n"
+
+	"3:	//dot_kernel_F1				\n"
+	"	ands	"J", "N", #31			\n"
+	"	ble	9f	//dot_kernel_L999	\n"
+
+	"4:	//dot_kernel_F10			\n"
+	"	ldr	"TMPX", ["X"]			\n"
+	"	ldr	"TMPY", ["Y"]			\n"
+	"	add	"X", "X", #"SZ"			\n"
+	"	add	"Y", "Y", #"SZ"			\n"
+	"	fmadd	"DOTF", "TMPX", "TMPY", "DOTF"	\n"
+
+	"	subs	"J", "J", #1			\n"
+	"	bne	4b	//dot_kernel_F10	\n"
+
+	"	b	9f	//dot_kernel_L999	\n"
+
+	"5:	//dot_kernel_S_BEGIN			\n"
+	"	lsl	"INC_X", "INC_X", #3		\n"
+	"	lsl	"INC_Y", "INC_Y", #3		\n"
+	"	asr	"J", "N", #2			\n"
+	"	cmp	"J", xzr			\n"
+	"	ble	7f	//dot_kernel_S1		\n"
+
+	"6:	//dot_kernel_S4:			\n"
+	"	ld1	"LD1VX", ["X"], "INC_X"		\n"
+	"	ld1	"LD1VY", ["Y"], "INC_Y"		\n"
+	"	fmadd	"DOTF", "TMPX", "TMPY", "DOTF"	\n"
+	"	ld1	"LD1VX", ["X"], "INC_X"		\n"
+	"	ld1	"LD1VY", ["Y"], "INC_Y"		\n"
+	"	fmadd	"DOTF", "TMPX", "TMPY", "DOTF"	\n"
+	"	ld1	"LD1VX", ["X"], "INC_X"		\n"
+	"	ld1	"LD1VY", ["Y"], "INC_Y"		\n"
+	"	fmadd	"DOTF", "TMPX", "TMPY", "DOTF"	\n"
+	"	ld1	"LD1VX", ["X"], "INC_X"		\n"
+	"	ld1	"LD1VY", ["Y"], "INC_Y"		\n"
+	"	fmadd	"DOTF", "TMPX", "TMPY", "DOTF"	\n"
+	"	subs	"J", "J", #1			\n"
+	"	bne	6b	//dot_kernel_S4		\n"
+
+	"7:	//dot_kernel_S1:			\n"
+	"	ands	"J", "N", #3			\n"
+	"	ble	9f	//dot_kernel_L999	\n"
+
+	"8:	//dot_kernel_S10			\n"
+	"	ld1	"LD1VX", ["X"], "INC_X"		\n"
+	"	ld1	"LD1VY", ["Y"], "INC_Y"		\n"
+	"	fmadd	"DOTF", "TMPX", "TMPY", "DOTF"	\n"
+	"	subs	"J", "J", #1			\n"
+	"	bne	8b	//dot_kernel_S10	\n"
+
+	"9:	//dot_kernel_L999			\n"
+	"	fmov	%[DOT_], "DOTF"			\n"
+	: [DOT_]  "=r" (dot)		//%0
+	: [N_]    "r"  (n),		//%1
+	  [X_]    "r"  (x),		//%2
+	  [INCX_] "r"  (inc_x),		//%3
+	  [Y_]    "r"  (y),		//%4
+	  [INCY_] "r"  (inc_y)		//%5
+	: "cc",
+	  "memory",
+	  "x0", "x1", "x2", "x3", "x4", "x5",
+	  "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"
+	);
+
+	return(dot);
+}
+
+static int ddot_thread_function(BLASLONG n, BLASLONG dummy0,
+	BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *y,
+	BLASLONG inc_y, FLOAT *result, BLASLONG dummy3)
+{
+	*result = ddot_compute(n, x, inc_x, y, inc_y);
+
+	return 0;
+}
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+{
+	int nthreads;
+	FLOAT dot = 0.0;
+	FLOAT dummy_alpha;
+
+	nthreads = num_cpu_avail(1);
+
+	if (inc_x == 0 || inc_y == 0)
+		nthreads = 1;
+
+	if (n <= 10000)
+		nthreads = 1;
+
+	if (nthreads == 1) {
+		dot = ddot_compute(n, x, inc_x, y, inc_y);
+	} else {
+		int mode, i;
+		char result[MAX_CPU_NUMBER * sizeof(double) * 2];
+		FLOAT *ptr;
+
+		mode = BLAS_DOUBLE  | BLAS_REAL;
+
+		blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
+				   x, inc_x, y, inc_y, result, 0,
+				   ( void *)ddot_thread_function, nthreads);
+
+		ptr = (FLOAT *)result;
+		for (i = 0; i < nthreads; i++) {
+			dot = dot + (*ptr);
+			ptr = (FLOAT *)(((char *)ptr) + sizeof(double) * 2);
+		}
+	}
+
+	return dot;
+}