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 3edb58f..6109c85 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
 
+#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
@@ -112,6 +120,13 @@ static redccore__functype *pick_redccore(void)
   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);
 }
 
@@ -204,6 +219,14 @@ static void simple_mulcore(mpw *dv, mpw *dvl,
   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
@@ -218,6 +241,13 @@ static mulcore__functype *pick_mulcore(void)
   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);
 }