From d850c3b7f47e58556c160f9d03ea20aa52452020 Mon Sep 17 00:00:00 2001 From: rodri Date: Fri, 24 Nov 2023 15:39:06 +0000 Subject: add more avx instructions and a bench9 benchmark file. --- dppd.s | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) (limited to 'dppd.s') diff --git a/dppd.s b/dppd.s index 907a437..7239d9f 100644 --- a/dppd.s +++ b/dppd.s @@ -12,6 +12,13 @@ TEXT dppd(SB), 1, $0 DPPD(rX1, rX0) /* DPPD $0x31, X1, X0 */ RET +TEXT dppda(SB), 1, $0 + MOVQ SP, AX + VMOVUPD_128mr(8, rAX, rX0) /* VMOVUPD a+0(FP), X0 */ + VMOVUPD_128mr(32, rAX, rX1) /* VMOVUPD b+24(FP), X1 */ + VDPPD(rX1, rX0, rX0) /* VDPPD $0x31, X1, X0, X0 */ + RET + TEXT dppd3(SB), 1, $0 MOVQ SP, AX MOVLPD(8, rAX, rX0) /* MOVLPD a+0(FP), X0 */ @@ -25,6 +32,16 @@ TEXT dppd3(SB), 1, $0 DPPD(rX1, rX0) /* DPPD $0x31, X1, X0 */ RET +TEXT dppd3a(SB), 1, $0 + MOVQ SP, AX + VMOVUPD_128mr(8, rAX, rX0) /* VMOVUPD a+0(FP), X0 */ + VMOVUPD_128mr(40, rAX, rX1) /* VMOVUPD b+32(FP), X1 */ + VDPPD(rX1, rX0, rX0) /* VDPPD $0x31, X1, X0, X0 */ + MOVSD a+16(FP), X1 + MOVSD b+48(FP), X2 + VFMADD231SD(rX1, rX2, rX0) + RET + TEXT Pt2b(SB), 1, $0 MOVQ BP, DI MOVSD x+8(FP), X0 -- cgit v1.2.3