[Gmp-commit] /var/hg/gmp: 4 new changesets
mercurial at gmplib.org
mercurial at gmplib.org
Sat Sep 14 11:12:17 UTC 2019
details: /var/hg/gmp/rev/fdb32090830c
changeset: 17894:fdb32090830c
user: Torbjorn Granlund <tg at gmplib.org>
date: Sat Sep 14 13:09:14 2019 +0200
description:
(GMP_ASM_X86_MULX): Set X86_ASM_MULX to config.h.
details: /var/hg/gmp/rev/59d8cef5e18e
changeset: 17895:59d8cef5e18e
user: Torbjorn Granlund <tg at gmplib.org>
date: Sat Sep 14 13:10:15 2019 +0200
description:
Update use of GMP_ASM_X86_MULX.
details: /var/hg/gmp/rev/ce6516e7d7c8
changeset: 17896:ce6516e7d7c8
user: Torbjorn Granlund <tg at gmplib.org>
date: Sat Sep 14 13:10:36 2019 +0200
description:
(x86 umul_ppmm): Test also X86_ASM_MULX for when to use mulx variant.
details: /var/hg/gmp/rev/0bc765c5b024
changeset: 17897:0bc765c5b024
user: Torbjorn Granlund <tg at gmplib.org>
date: Sat Sep 14 13:12:15 2019 +0200
description:
Trivial merge.
diffstat:
ChangeLog | 6 ++
acinclude.m4 | 2 +
configure.ac | 14 ++--
longlong.h | 5 +-
tune/tuneup.c | 172 +++++++++++++++++++++++----------------------------------
5 files changed, 90 insertions(+), 109 deletions(-)
diffs (truncated from 333 to 300 lines):
diff -r edfbbc5d4e48 -r 0bc765c5b024 ChangeLog
--- a/ChangeLog Fri Sep 13 14:32:41 2019 +0200
+++ b/ChangeLog Sat Sep 14 13:12:15 2019 +0200
@@ -1,3 +1,9 @@
+2019-09-13 Niels Möller <nisse at lysator.liu.se>
+
+ * tune/tuneup.c (one_method): New helper function, to measure
+ several functions for a fix size.
+ (tune_hgcd2, tune_div_qr_1, tune_mod_1, tune_jacobi_base): Use it.
+
2019-09-05 Torbjörn Granlund <tg at gmplib.org>
* mpn/arm64/gcd_22.asm: Rewrite to make better use of Arm conditional
diff -r edfbbc5d4e48 -r 0bc765c5b024 acinclude.m4
--- a/acinclude.m4 Fri Sep 13 14:32:41 2019 +0200
+++ b/acinclude.m4 Sat Sep 14 13:12:15 2019 +0200
@@ -2659,6 +2659,8 @@
])
case $gmp_cv_asm_x86_mulx in
yes)
+ AC_DEFINE(X86_ASM_MULX, 1,
+ [Define to 1 if the assembler understands the mulx instruction])
ifelse([$1],,:,[$1])
;;
*)
diff -r edfbbc5d4e48 -r 0bc765c5b024 configure.ac
--- a/configure.ac Fri Sep 13 14:32:41 2019 +0200
+++ b/configure.ac Sat Sep 14 13:12:15 2019 +0200
@@ -1873,11 +1873,13 @@
gcc_cflags_arch="-march=bdver4 -march=bdver3 -march=bdver2 -march=bdver1 -march=amdfam10 -march=k8 -march=k8~-mno-sse2"
path="x86/bd4 x86/bd3 x86/bd2 x86/bd1 x86/k7/mmx x86/k7 x86/mmx x86"
path_64="x86_64/bd4 x86_64/bd3 x86_64/bd2 x86_64/bd1 x86_64/k10 x86_64/k8 x86_64"
+ x86_have_mulx=1
;;
zen | zennoavx)
gcc_cflags_cpu="-mtune=znver1 -mtune=amdfam10 -mtune=k8"
gcc_cflags_arch="-march=znver1 -march=amdfam10 -march=k8"
path="x86/k7/mmx x86/k7 x86/mmx x86"
+ x86_have_mulx=1
path_64="x86_64/zen x86_64"
;;
core2)
@@ -1904,6 +1906,7 @@
gcc_cflags_arch="-march=haswell -march=corei7 -march=core2 -march=core2~-mno-sse2 -march=k8 -march=k8~-mno-sse2"
path="x86/coreisbr x86/p6/sse2 x86/p6/p3mmx x86/p6/mmx x86/p6 x86/mmx x86"
path_64="x86_64/coreihwl x86_64/coreisbr x86_64/coreinhm x86_64/core2 x86_64"
+ x86_have_mulx=1
;;
coreibwl | coreibwlnoavx | broadwell | broadwellnoavx)
gcc_cflags_cpu="-mtune=broadwell -mtune=corei7 -mtune=core2 -mtune=k8"
@@ -1911,6 +1914,7 @@
path="x86/coreisbr x86/p6/sse2 x86/p6/p3mmx x86/p6/mmx x86/p6 x86/mmx x86"
path_64="x86_64/coreibwl x86_64/coreihwl x86_64/coreisbr x86_64/coreinhm x86_64/core2 x86_64"
# extra_functions_64="missing" # enable for bmi2/adx simulation
+ x86_have_mulx=1
;;
skylake | skylakenoavx | kabylake | kabylakenoavx)
gcc_cflags_cpu="-mtune=skylake -mtune=broadwell -mtune=corei7 -mtune=core2 -mtune=k8"
@@ -1919,6 +1923,7 @@
path="x86/coreisbr x86/p6/sse2 x86/p6/p3mmx x86/p6/mmx x86/p6 x86/mmx x86"
path_64="x86_64/skylake x86_64/coreibwl x86_64/coreihwl x86_64/coreisbr x86_64/coreinhm x86_64/core2 x86_64"
# extra_functions_64="missing" # enable for bmi2/adx simulation
+ x86_have_mulx=1
;;
atom) # in-order pipeline atom
gcc_cflags_cpu="-mtune=atom -mtune=pentium3"
@@ -2581,12 +2586,6 @@
*sse2*) GMP_ASM_X86_SSE2( , [GMP_STRIP_PATH(sse2)]) ;;
esac
fi
- case "$path $fat_path" in
- *mulx*) GMP_ASM_X86_MULX( , [GMP_STRIP_PATH(mulx)]) ;;
- esac
- case "$path $fat_path" in
- *adx*) GMP_ASM_X86_ADX( , [GMP_STRIP_PATH(adx)]) ;;
- esac
;;
esac
@@ -3756,6 +3755,9 @@
;;
X86_PATTERN | X86_64_PATTERN)
GMP_ASM_ALIGN_FILL_0x90
+ if test $x86_have_mulx = 1; then
+ GMP_ASM_X86_MULX
+ fi
case $ABI in
32)
GMP_INCLUDE_MPN(x86/x86-defs.m4)
diff -r edfbbc5d4e48 -r 0bc765c5b024 longlong.h
--- a/longlong.h Fri Sep 13 14:32:41 2019 +0200
+++ b/longlong.h Sat Sep 14 13:12:15 2019 +0200
@@ -1058,8 +1058,9 @@
: "=r" (sh), "=&r" (sl) \
: "0" ((UDItype)(ah)), "rme" ((UDItype)(bh)), \
"1" ((UDItype)(al)), "rme" ((UDItype)(bl)))
-#if HAVE_HOST_CPU_haswell || HAVE_HOST_CPU_broadwell || HAVE_HOST_CPU_skylake \
- || HAVE_HOST_CPU_bd4 || HAVE_HOST_CPU_zen
+#if X86_ASM_MULX
+ && (HAVE_HOST_CPU_haswell || HAVE_HOST_CPU_broadwell
+ || HAVE_HOST_CPU_skylake || HAVE_HOST_CPU_bd4 || HAVE_HOST_CPU_zen)
#define umul_ppmm(w1, w0, u, v) \
__asm__ ("mulx\t%3, %0, %1" \
: "=r" (w0), "=r" (w1) \
diff -r edfbbc5d4e48 -r 0bc765c5b024 tune/tuneup.c
--- a/tune/tuneup.c Fri Sep 13 14:32:41 2019 +0200
+++ b/tune/tuneup.c Sat Sep 14 13:12:15 2019 +0200
@@ -519,8 +519,8 @@
}
void
-print_define_with_margin (const char *name, mp_size_t value,
- mp_size_t runner_up, double speedup)
+print_define_with_speedup (const char *name, mp_size_t value,
+ mp_size_t runner_up, double speedup)
{
char buf[100];
snprintf (buf, sizeof(buf), "%.2f%% faster than %ld",
@@ -712,6 +712,48 @@
print_define_end (param->name, *threshold);
}
+void
+one_method (int n, speed_function_t *functions,
+ const char *name, const char *define,
+ const struct param_t *param)
+{
+ double *t;
+ int i;
+ int method;
+ int method_runner_up;
+
+ TMP_DECL;
+ TMP_MARK;
+ t = TMP_ALLOC (n * sizeof (*t));
+
+ for (i = 0; i < n; i++)
+ {
+ t[i] = tuneup_measure (functions[i], param, &s);
+ if (option_trace >= 1)
+ printf ("size=%ld, %s, method %d %.9f\n",
+ (long) s.size, name, i + 1, t[i]);
+ if (t[i] == -1.0)
+ {
+ printf ("Oops, can't measure all %s methods\n", name);
+ abort ();
+ }
+ }
+ method = 0;
+ for (i = 1; i < n; i++)
+ if (t[i] < t[method])
+ method = i;
+
+ method_runner_up = (method == 0);
+ for (i = 0; i < n; i++)
+ if (i != method && t[i] < t[method_runner_up])
+ method_runner_up = i;
+
+ print_define_with_speedup (define, method + 1, method_runner_up + 1,
+ t[method_runner_up] / t[method]);
+
+ TMP_FREE;
+}
+
/* Special probing for the fft thresholds. The size restrictions on the
FFTs mean the graph of time vs size has a step effect. See this for
@@ -1911,47 +1953,15 @@
tune_hgcd2 (void)
{
static struct param_t param;
- double t[3+1];
- int method;
- int runner_up_method;
- double runner_up_ratio;
+ speed_function_t f[3] =
+ {
+ speed_mpn_hgcd2_1,
+ speed_mpn_hgcd2_2,
+ speed_mpn_hgcd2_3,
+ };
s.size = 1;
- t[1] = tuneup_measure (speed_mpn_hgcd2_1, ¶m, &s);
- if (option_trace >= 1)
- printf ("size=%ld, mpn_hgcd2_1 %.9f\n", (long) s.size, t[1]);
-
- t[2] = tuneup_measure (speed_mpn_hgcd2_2, ¶m, &s);
- if (option_trace >= 1)
- printf ("size=%ld, mpn_hgcd2_2 %.9f\n", (long) s.size, t[2]);
-
- t[3] = tuneup_measure (speed_mpn_hgcd2_3, ¶m, &s);
- if (option_trace >= 1)
- printf ("size=%ld, mpn_hgcd2_3 %.9f\n", (long) s.size, t[3]);
-
- if (t[1] == -1.0 || t[2] == -1.0 || t[3] == -1.0)
- {
- printf ("Oops, can't measure all mpn_hgcd2 methods\n");
- abort ();
- }
-
- if (t[1] < t[2] && t[1] < t[3])
- {
- method = 1;
- runner_up_method = (t[2] < t[3]) ? 2 : 3;
- }
- else if (t[2] < t[3])
- {
- method = 2;
- runner_up_method = (t[1] < t[3]) ? 1 : 3;
- }
- else
- {
- method = 3;
- runner_up_method = (t[1] < t[2]) ? 1 : 2;
- }
- print_define_with_margin ("HGCD2_METHOD", method, runner_up_method,
- t[runner_up_method] / t[method]);
+ one_method (3, f, "mpn_hgcd2", "HGCD2_METHOD", ¶m);
}
void
@@ -2227,22 +2237,16 @@
if (!HAVE_NATIVE_mpn_div_qr_1n_pi1)
{
static struct param_t param;
- double t1, t2;
+ speed_function_t f[2] =
+ {
+ speed_mpn_div_qr_1n_pi1_1,
+ speed_mpn_div_qr_1n_pi1_2,
+ };
s.size = 10;
s.r = randlimb_norm ();
- t1 = tuneup_measure (speed_mpn_div_qr_1n_pi1_1, ¶m, &s);
- t2 = tuneup_measure (speed_mpn_div_qr_1n_pi1_2, ¶m, &s);
-
- if (t1 == -1.0 || t2 == -1.0)
- {
- printf ("Oops, can't measure all mpn_div_qr_1n_pi1 methods at %ld\n",
- (long) s.size);
- abort ();
- }
- div_qr_1n_pi1_method = (t1 < t2) ? 1 : 2;
- print_define ("DIV_QR_1N_PI1_METHOD", div_qr_1n_pi1_method);
+ one_method (2, f, "mpn_div_qr_1n_pi1", "DIV_QR_1N_PI1_METHOD", ¶m);
}
{
@@ -2289,22 +2293,15 @@
if (!HAVE_NATIVE_mpn_mod_1_1p)
{
static struct param_t param;
- double t1, t2;
+ speed_function_t f[2] =
+ {
+ speed_mpn_mod_1_1_1,
+ speed_mpn_mod_1_1_2,
+ };
s.size = 10;
s.r = randlimb_half ();
-
- t1 = tuneup_measure (speed_mpn_mod_1_1_1, ¶m, &s);
- t2 = tuneup_measure (speed_mpn_mod_1_1_2, ¶m, &s);
-
- if (t1 == -1.0 || t2 == -1.0)
- {
- printf ("Oops, can't measure all mpn_mod_1_1 methods at %ld\n",
- (long) s.size);
- abort ();
- }
- mod_1_1p_method = (t1 < t2) ? 1 : 2;
- print_define ("MOD_1_1P_METHOD", mod_1_1p_method);
+ one_method (2, f, "mpn_mod_1_1", "MOD_1_1P_METHOD", ¶m);
}
if (UDIV_PREINV_ALWAYS)
@@ -2680,44 +2677,17 @@
tune_jacobi_base (void)
{
static struct param_t param;
- double t1, t2, t3, t4;
- int method;
+ speed_function_t f[4] =
+ {
+ speed_mpn_jacobi_base_1,
+ speed_mpn_jacobi_base_2,
+ speed_mpn_jacobi_base_3,
+ speed_mpn_jacobi_base_4,
+ };
s.size = GMP_LIMB_BITS * 3 / 4;
- t1 = tuneup_measure (speed_mpn_jacobi_base_1, ¶m, &s);
- if (option_trace >= 1)
- printf ("size=%ld, mpn_jacobi_base_1 %.9f\n", (long) s.size, t1);
-
More information about the gmp-commit
mailing list