Torbjörn Granlund <t...@gmplib.org> writes:
> Your version is faster than my versions (where I tested them).
>
> I made some minor changes to your code.
Nice!
Made a few additional tweaks, and tried to get it in the right place.
Attaching a patch that adds the file under coreibwl, as suggested, with
an include_mpn from zen. Took out the undefine hack (which I added
because first attempt at using the mulx macros failed), and copied the
windows support stuff from the old implementation. Complete patch
attached, does it look good enough? It now gives appr 30% speedup
compared to mul_1 + addmul_1 on my machine. I'd prefer to next have a
look at addsubmul_1msb0, before trying to optimize this loop further.
diff -r 3dee30523768 ChangeLog
--- a/ChangeLog Fri Oct 08 16:05:05 2021 +0200
+++ b/ChangeLog Fri Oct 08 16:41:04 2021 +0200
@@ -1,5 +1,8 @@
2021-10-08 Niels Möller <ni...@lysator.liu.se>
+ * mpn/x86_64/coreibwl/addaddmul_1msb0.asm: New assembly
+ implementation, using mulx and adox instructions.
+
* tests/mpn/t-addaddmul.c: Unit test for mpn_addaddmul_1msb0.
2021-10-07 Niels Möller <ni...@lysator.liu.se>
diff -r 3dee30523768 mpn/x86_64/coreibwl/addaddmul_1msb0.asm
--- /dev/null Thu Jan 01 00:00:00 1970 +0000
+++ b/mpn/x86_64/coreibwl/addaddmul_1msb0.asm Fri Oct 08 16:41:04 2021 +0200
@@ -0,0 +1,99 @@
+dnl AMD64 mpn_addaddmul_1msb0, R = Au + Bv, u,v < 2^63.
+
+dnl Contributed to the GNU project by Niels Möller and Torbjörn Granlund.
+
+dnl Copyright 2021 Free Software Foundation, 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 INPUT PARAMETERS
+define(`rp', `%rdi')
+define(`ap', `%rsi')
+define(`bp_param', `%rdx')
+define(`n', `%rcx')
+define(`u0', `%r8')
+define(`v0', `%r9')
+
+define(`bp', `%rbx')
+
+define(`c0', `%rax') C carry limb, and return value
+define(`l0', `%r10')
+define(`l1', `%r11')
+define(`hi', `%rbp')
+
+ABI_SUPPORT(DOS64)
+ABI_SUPPORT(STD64)
+
+ASM_START()
+ TEXT
+ ALIGN(16)
+PROLOGUE(mpn_addaddmul_1msb0)
+ FUNC_ENTRY(4)
+IFDOS(` mov 56(%rsp), %r8 ')
+IFDOS(` mov 64(%rsp), %r9 ')
+
+ push %rbx
+ push %rbp
+
+ lea (ap,n,8), ap
+ lea (bp_param,n,8), bp
+ lea (rp,n,8), rp
+ neg n
+
+ xor R32(c0), R32(c0) C Also clears CF and OF
+ test $1, R32(n)
+ jnz L(mid)
+
+ ALIGN(16)
+L(top): mov (ap,n,8), %rdx
+ mulx( u0, l0, hi)
+ mov (bp,n,8), %rdx
+ adox( c0, l0)
+ mulx( v0, l1, c0)
+ adox( hi, c0)
+ adc l0, l1
+ mov l1, (rp,n,8)
+ inc n C Clears OF (since n != 2^63 - 1)
+L(mid): mov (ap,n,8), %rdx
+ mulx( u0, l0, hi)
+ mov (bp,n,8), %rdx
+ adox( c0, l0)
+ mulx( v0, l1, c0)
+ adox( hi, c0)
+ adc l0, l1
+ mov l1, (rp,n,8)
+ inc n
+ jnz L(top)
+
+L(end): adc $0, c0
+ pop %rbp
+ pop %rbx
+ FUNC_EXIT()
+ ret
+EPILOGUE()
diff -r 3dee30523768 mpn/x86_64/zen/addaddmul_1msb0.asm
--- /dev/null Thu Jan 01 00:00:00 1970 +0000
+++ b/mpn/x86_64/zen/addaddmul_1msb0.asm Fri Oct 08 16:41:04 2021 +0200
@@ -0,0 +1,37 @@
+dnl AMD64 mpn_addaddmul_1msb0.
+
+dnl Copyright 2021 Free Software Foundation, 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')
+
+ABI_SUPPORT(DOS64)
+ABI_SUPPORT(STD64)
+
+MULFUNC_PROLOGUE(mpn_addaddmul_1msb0)
+include_mpn(`x86_64/coreibwl/addaddmul_1msb0.asm')
> dnl AMD64 mpn_addsubmul_1msb0, R = Au - Bv, u,v < 2^63.
This comment obviously wrong ;-)
But that function could be implemented by adding two "not %rdx" in the
right places of the loop, plus small adjustment just before and after
the loop.
Since
Au - Bv = Au + (2^{64 n} - 1 - B) v - 2^{64 n} v + v
So complement B on the fly, set initial carry limb to v, and subtract v from
the return value. (Same trick as in arm/v7a/cora15/submul_1).
Should definitely be worth a try, before trying some completely
different loop.
Regards,
/Niels
--
Niels Möller. PGP-encrypted email is preferred. Keyid 368C6677.
Internet email is subject to wholesale government surveillance.
_______________________________________________
gmp-devel mailing list
gmp-devel@gmplib.org
https://gmplib.org/mailman/listinfo/gmp-devel
