chiark
/
gitweb
/
~mdw
/
catacomb
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
math/mpx-mul4-{arm-neon,arm64-simd}.S, etc.: Add ARM versions of `mul4'.
[catacomb]
/
math
/
mpmont.c
diff --git
a/math/mpmont.c
b/math/mpmont.c
index 3edb58fbe8b4ac5880ce8785a931e1e5a2b95a11..6109c8569770fe82a518cd31f90c221bbdcad39f 100644
(file)
--- a/
math/mpmont.c
+++ b/
math/mpmont.c
@@
-98,6
+98,14
@@
static void simple_redccore(mpw *dv, mpw *dvl, const mpw *mv,
MAYBE_REDC4(amd64_avx)
#endif
MAYBE_REDC4(amd64_avx)
#endif
+#if CPUFAM_ARMEL
+ MAYBE_REDC4(arm_neon)
+#endif
+
+#if CPUFAM_ARM64
+ MAYBE_REDC4(arm64_simd)
+#endif
+
static redccore__functype *pick_redccore(void)
{
#if CPUFAM_X86
static redccore__functype *pick_redccore(void)
{
#if CPUFAM_X86
@@
-111,6
+119,13
@@
static redccore__functype *pick_redccore(void)
cpu_feature_p(CPUFEAT_X86_AVX));
DISPATCH_PICK_COND(mpmont_reduce, maybe_redc4_amd64_sse2,
cpu_feature_p(CPUFEAT_X86_SSE2));
cpu_feature_p(CPUFEAT_X86_AVX));
DISPATCH_PICK_COND(mpmont_reduce, maybe_redc4_amd64_sse2,
cpu_feature_p(CPUFEAT_X86_SSE2));
+#endif
+#if CPUFAM_ARMEL
+ DISPATCH_PICK_COND(mpmont_reduce, maybe_redc4_arm_neon,
+ cpu_feature_p(CPUFEAT_ARM_NEON));
+#endif
+#if CPUFAM_ARM64
+ DISPATCH_PICK_COND(mpmont_reduce, maybe_redc4_arm64_simd, 1);
#endif
DISPATCH_PICK_FALLBACK(mpmont_reduce, simple_redccore);
}
#endif
DISPATCH_PICK_FALLBACK(mpmont_reduce, simple_redccore);
}
@@
-204,6
+219,14
@@
static void simple_mulcore(mpw *dv, mpw *dvl,
MAYBE_MUL4(amd64_avx)
#endif
MAYBE_MUL4(amd64_avx)
#endif
+#if CPUFAM_ARMEL
+ MAYBE_MUL4(arm_neon)
+#endif
+
+#if CPUFAM_ARM64
+ MAYBE_MUL4(arm64_simd)
+#endif
+
static mulcore__functype *pick_mulcore(void)
{
#if CPUFAM_X86
static mulcore__functype *pick_mulcore(void)
{
#if CPUFAM_X86
@@
-217,6
+240,13
@@
static mulcore__functype *pick_mulcore(void)
cpu_feature_p(CPUFEAT_X86_AVX));
DISPATCH_PICK_COND(mpmont_mul, maybe_mul4_amd64_sse2,
cpu_feature_p(CPUFEAT_X86_SSE2));
cpu_feature_p(CPUFEAT_X86_AVX));
DISPATCH_PICK_COND(mpmont_mul, maybe_mul4_amd64_sse2,
cpu_feature_p(CPUFEAT_X86_SSE2));
+#endif
+#if CPUFAM_ARMEL
+ DISPATCH_PICK_COND(mpmont_mul, maybe_mul4_arm_neon,
+ cpu_feature_p(CPUFEAT_ARM_NEON));
+#endif
+#if CPUFAM_ARM64
+ DISPATCH_PICK_COND(mpmont_mul, maybe_mul4_arm64_simd, 1);
#endif
DISPATCH_PICK_FALLBACK(mpmont_mul, simple_mulcore);
}
#endif
DISPATCH_PICK_FALLBACK(mpmont_mul, simple_mulcore);
}