[Gmp-commit] /var/hg/gmp: 2 new changesets
mercurial at gmplib.org
mercurial at gmplib.org
Sun Nov 18 00:27:54 UTC 2018
details: /var/hg/gmp/rev/176728661568
changeset: 17708:176728661568
user: Torbjorn Granlund <tg at gmplib.org>
date: Sun Nov 18 01:27:02 2018 +0100
description:
Optimise.
details: /var/hg/gmp/rev/2c7b911c6329
changeset: 17709:2c7b911c6329
user: Torbjorn Granlund <tg at gmplib.org>
date: Sun Nov 18 01:27:27 2018 +0100
description:
New file.
diffstat:
mpn/powerpc64/mode64/p9/addmul_1.asm | 74 ++--
mpn/powerpc64/mode64/p9/mul_basecase.asm | 416 +++++++++++++++++++++++++++++++
2 files changed, 450 insertions(+), 40 deletions(-)
diffs (truncated from 557 to 300 lines):
diff -r 52bae415aab0 -r 2c7b911c6329 mpn/powerpc64/mode64/p9/addmul_1.asm
--- a/mpn/powerpc64/mode64/p9/addmul_1.asm Fri Nov 16 07:47:53 2018 +0100
+++ b/mpn/powerpc64/mode64/p9/addmul_1.asm Sun Nov 18 01:27:27 2018 +0100
@@ -1,4 +1,4 @@
-dnl POWER9 mpn_addmul_1.
+dnl Power9 mpn_addmul_1.
dnl Copyright 2017, 2018 Free Software Foundation, Inc.
@@ -40,8 +40,9 @@
C POWER9 2.5
C TODO
-C * Schedule for POWER9 pipeline.
-C * Unroll to at least 4x if that proves beneficial.
+C * Schedule for Power9 pipeline.
+C * Unroll 4x if that proves beneficial.
+C * This is marginally faster (but much smaller) than ../aorsmul_1.asm.
C INPUT PARAMETERS
define(`rp', `r3')
@@ -51,41 +52,37 @@
ASM_START()
PROLOGUE(mpn_addmul_1)
- std r31, -8(r1)
-
cmpdi cr6, n, 2
-
addi r0, n, -1 C FIXME: postpone
srdi r0, r0, 1 C FIXME: postpone
mtctr r0 C FIXME: postpone
-
rldicl. r0, n, 0,63 C r0 = n & 3, set cr0
bne cr0, L(b1)
L(b0): ld r10, 0(rp)
ld r12, 0(up)
ld r11, 8(rp)
- ld r31, 8(up)
- maddld( r0, r12, v0, r10)
+ ld r0, 8(up)
+ maddld( r9, r12, v0, r10)
maddhdu(r7, r12, v0, r10)
ble cr6, L(2)
ld r10, 16(rp)
ld r12, 16(up)
- maddld( r8, r31, v0, r11)
- maddhdu(r5, r31, v0, r11)
+ maddld( r8, r0, v0, r11)
+ maddhdu(r5, r0, v0, r11)
addic up, up, 16
addi rp, rp, -8
b L(mid)
L(b1): ld r11, 0(rp)
- ld r31, 0(up)
+ ld r0, 0(up)
ble cr6, L(1)
ld r10, 8(rp)
ld r12, 8(up)
- maddld( r0, r31, v0, r11)
- maddhdu(r5, r31, v0, r11)
+ maddld( r8, r0, v0, r11)
+ maddhdu(r5, r0, v0, r11)
ld r11, 16(rp)
- ld r31, 16(up)
+ ld r0, 16(up)
maddld( r9, r12, v0, r10)
maddhdu(r7, r12, v0, r10)
addic up, up, 24
@@ -94,43 +91,40 @@
ALIGN(16)
L(top): ld r10, 24(rp)
ld r12, 0(up)
- std r0, 0(rp)
- adde r0, r5, r9
- maddld( r8, r31, v0, r11) C W:0,2,4
- maddhdu(r5, r31, v0, r11) C W:1,3,5
+ std r8, 0(rp)
+ adde r9, r5, r9
+ maddld( r8, r0, v0, r11) C W:0,2,4
+ maddhdu(r5, r0, v0, r11) C W:1,3,5
L(mid): ld r11, 32(rp)
- ld r31, 8(up)
- std r0, 8(rp)
- adde r0, r7, r8
+ ld r0, 8(up)
+ std r9, 8(rp)
+ adde r8, r7, r8
maddld( r9, r12, v0, r10) C W:1,3,5
maddhdu(r7, r12, v0, r10) C W:2,4,6
addi rp, rp, 16
addi up, up, 16
bdnz L(top)
-L(end): std r0, 0(rp)
- adde r0, r5, r9
- maddld( r8, r31, v0, r11)
- maddhdu(r5, r31, v0, r11)
- std r0, 8(rp)
- adde r0, r7, r8
- std r0, 16(rp)
+L(end): std r8, 0(rp)
+ maddld( r8, r0, v0, r11)
+ adde r9, r5, r9
+ maddhdu(r5, r0, v0, r11)
+ std r9, 8(rp)
+ adde r8, r7, r8
+ std r8, 16(rp)
addze r3, r5
- ld r31, -8(r1)
blr
-L(2): maddld( r8, r31, v0, r11)
- maddhdu(r5, r31, v0, r11)
- std r0, 0(rp)
- addc r0, r7, r8
- std r0, 8(rp)
+L(2): maddld( r8, r0, v0, r11)
+ maddhdu(r5, r0, v0, r11)
+ std r9, 0(rp)
+ addc r8, r7, r8
+ std r8, 8(rp)
addze r3, r5
- ld r31, -8(r1)
blr
-L(1): maddld( r0, r31, v0, r11)
- std r0, 0(rp)
- maddhdu(r3, r31, v0, r11)
- ld r31, -8(r1)
+L(1): maddld( r8, r0, v0, r11)
+ std r8, 0(rp)
+ maddhdu(r3, r0, v0, r11)
blr
EPILOGUE()
diff -r 52bae415aab0 -r 2c7b911c6329 mpn/powerpc64/mode64/p9/mul_basecase.asm
--- /dev/null Thu Jan 01 00:00:00 1970 +0000
+++ b/mpn/powerpc64/mode64/p9/mul_basecase.asm Sun Nov 18 01:27:27 2018 +0100
@@ -0,0 +1,416 @@
+dnl Power9 mpn_mul_basecase.
+
+dnl Copyright 1999-2001, 2003-2006, 2008, 2017-2018 Free Software Foundation,
+dnl Inc.
+
+dnl This file is part of the GNU MP Library.
+dnl
+dnl The GNU MP Library is free software; you can redistribute it and/or modify
+dnl it under the terms of either:
+dnl
+dnl * the GNU Lesser General Public License as published by the Free
+dnl Software Foundation; either version 3 of the License, or (at your
+dnl option) any later version.
+dnl
+dnl or
+dnl
+dnl * the GNU General Public License as published by the Free Software
+dnl Foundation; either version 2 of the License, or (at your option) any
+dnl later version.
+dnl
+dnl or both in parallel, as here.
+dnl
+dnl The GNU MP Library is distributed in the hope that it will be useful, but
+dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
+dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
+dnl for more details.
+dnl
+dnl You should have received copies of the GNU General Public License and the
+dnl GNU Lesser General Public License along with the GNU MP Library. If not,
+dnl see https://www.gnu.org/licenses/.
+
+include(`../config.m4')
+
+C cycles/limb
+C POWER3/PPC630 -
+C POWER4/PPC970 -
+C POWER5 -
+C POWER6 -
+C POWER7 -
+C POWER8 -
+C POWER9 1.62
+
+C TODO
+C * Check if (inner) loop alignment affects performance.
+C * Could we schedule loads less in addmul_2/mul_2? That would save some regs
+C and make the tail code more manageable.
+C * Postpone some register saves to main loop.
+C * Perhaps write more small operands (3x1, 3x2, 3x3) code.
+C * Consider restoring rp,up after loop using arithmetic, eliminating r20,r21.
+C On the other hand, the current rp,up restore register are useful for OSP.
+C * Do OSP. This should save a lot with the current deep addmul_2 pipeline.
+
+C INPUT PARAMETERS
+define(`rp', `r3')
+define(`up', `r4')
+define(`un', `r5')
+define(`vp', `r6')
+define(`vn', `r7')
+
+define(`v0', `r0')
+define(`v1', `r7')
+define(`rp2', `r24')
+define(`up2', `r25')
+
+ASM_START()
+PROLOGUE(mpn_mul_basecase)
+ cmpdi cr0, un, 2
+ bgt cr0, L(un_gt2)
+ cmpdi cr6, vn, 1
+ ld r7, 0(vp)
+ ld r5, 0(up)
+ mulld r8, r5, r7 C weight 0
+ mulhdu r9, r5, r7 C weight 1
+ std r8, 0(rp)
+ beq cr0, L(2x)
+ std r9, 8(rp)
+ blr
+ ALIGN(16)
+L(2x): ld r0, 8(up)
+ mulld r8, r0, r7 C weight 1
+ mulhdu r10, r0, r7 C weight 2
+ addc r9, r9, r8
+ addze r10, r10
+ bne cr6, L(2x2)
+ std r9, 8(rp)
+ std r10, 16(rp)
+ blr
+ ALIGN(16)
+L(2x2): ld r6, 8(vp)
+ mulld r8, r5, r6 C weight 1
+ mulhdu r11, r5, r6 C weight 2
+ addc r9, r9, r8
+ std r9, 8(rp)
+ adde r11, r11, r10
+ mulld r12, r0, r6 C weight 2
+ mulhdu r0, r0, r6 C weight 3
+ addze r0, r0
+ addc r11, r11, r12
+ addze r0, r0
+ std r11, 16(rp)
+ std r0, 24(rp)
+ blr
+
+L(un_gt2):
+ std r22, -80(r1)
+ std r23, -72(r1)
+ std r24, -64(r1)
+ std r25, -56(r1)
+ std r26, -48(r1)
+ std r27, -40(r1)
+ std r28, -32(r1)
+ std r29, -24(r1)
+ std r30, -16(r1)
+ std r31, -8(r1)
+ mr rp2, r3 C rp
+ mr up2, r4 C up
+ srdi r22, r5, 2 C un
+ subfic r23, r7, 0 C -vn, clear CA
+ subfo r0, r0, r0 C clear OV (and r0)
+
+ cmpdi cr6, un, 3
+ rldicl r0, un, 0, 63 C r0 = un & 1
+ cmpdi cr7, r0, 0
+ rldicl r0, un, 63, 63 C FIXME: unused for vn = 1
+ cmpdi cr5, r0, 0 C FIXME: unused for vn = 1
+
+ ld v0, 0(vp)
+ rldicl. r9, vn, 0, 63
+ beq cr0, L(vn_evn)
+
+L(vn_odd):
+ addi r10, un, -2
+ ld r5, 0(up)
+ srdi r10, r10, 1
+ mtctr r10
+ bne cr7, L(m1_b1)
+
+L(m1_b0):
+ ld r10, 8(up)
+ mulld r9, r5, v0
+ mulhdu r11, r5, v0
+ ld r12, 16(up)
+ mulld r8, r10, v0
+ mulhdu r5, r10, v0
+ addi rp, rp, -8
+ b L(m1_mid)
+
+L(m1_b1):
+ ld r12, 8(up)
+ mulld r8, r5, v0
+ mulhdu r5, r5, v0
+ ld r10, 16(up)
+ mulld r9, r12, v0
+ mulhdu r11, r12, v0
+ addi up, up, 8
+ beq cr6, L(m1_end) C jump taken means un = 3, vn = {1,3}
+
+ ALIGN(16)
+L(m1_top):
More information about the gmp-commit
mailing list