X-Git-Url: https://git.distorted.org.uk/~mdw/catacomb/blobdiff_plain/444083aef7e70ce9afe893a36d72e1a1a976f1ed..a1a9ee0a7240087e202a7855e470573de0e59c09:/math/mpmont.c

diff --git a/math/mpmont.c b/math/mpmont.c
index 968766d3..2ed1e113 100644
--- a/math/mpmont.c
+++ b/math/mpmont.c
@@ -42,6 +42,8 @@
 
 /* #define MPMONT_DISABLE */
 
+#define MPMONT_KTHRESH (16*MPK_THRESH)
+
 /*----- Low-level implementation ------------------------------------------*/
 
 #ifndef MPMONT_DISABLE
@@ -88,14 +90,28 @@ static void simple_redccore(mpw *dv, mpw *dvl, const mpw *mv,
 
 #if CPUFAM_X86
   MAYBE_REDC4(x86_sse2)
+  MAYBE_REDC4(x86_avx)
+#endif
+
+#if CPUFAM_AMD64
+  MAYBE_REDC4(amd64_sse2)
+  MAYBE_REDC4(amd64_avx)
 #endif
 
 static redccore__functype *pick_redccore(void)
 {
 #if CPUFAM_X86
+  DISPATCH_PICK_COND(mpmont_reduce, maybe_redc4_x86_avx,
+		     cpu_feature_p(CPUFEAT_X86_AVX));
   DISPATCH_PICK_COND(mpmont_reduce, maybe_redc4_x86_sse2,
 		     cpu_feature_p(CPUFEAT_X86_SSE2));
 #endif
+#if CPUFAM_AMD64
+  DISPATCH_PICK_COND(mpmont_reduce, maybe_redc4_amd64_avx,
+		     cpu_feature_p(CPUFEAT_X86_AVX));
+  DISPATCH_PICK_COND(mpmont_reduce, maybe_redc4_amd64_sse2,
+		     cpu_feature_p(CPUFEAT_X86_SSE2));
+#endif
   DISPATCH_PICK_FALLBACK(mpmont_reduce, simple_redccore);
 }
 
@@ -180,20 +196,35 @@ static void simple_mulcore(mpw *dv, mpw *dvl,
 
 #if CPUFAM_X86
   MAYBE_MUL4(x86_sse2)
+  MAYBE_MUL4(x86_avx)
+#endif
+
+#if CPUFAM_AMD64
+  MAYBE_MUL4(amd64_sse2)
+  MAYBE_MUL4(amd64_avx)
 #endif
 
 static mulcore__functype *pick_mulcore(void)
 {
 #if CPUFAM_X86
+  DISPATCH_PICK_COND(mpmont_mul, maybe_mul4_x86_avx,
+		     cpu_feature_p(CPUFEAT_X86_AVX));
   DISPATCH_PICK_COND(mpmont_mul, maybe_mul4_x86_sse2,
 		     cpu_feature_p(CPUFEAT_X86_SSE2));
 #endif
+#if CPUFAM_AMD64
+  DISPATCH_PICK_COND(mpmont_mul, maybe_mul4_amd64_avx,
+		     cpu_feature_p(CPUFEAT_X86_AVX));
+  DISPATCH_PICK_COND(mpmont_mul, maybe_mul4_amd64_sse2,
+		     cpu_feature_p(CPUFEAT_X86_SSE2));
+#endif
   DISPATCH_PICK_FALLBACK(mpmont_mul, simple_mulcore);
 }
 
 /* --- @finish@ --- *
  *
- * Arguments:	@mpmont *mm@ = pointer to a Montgomery reduction context
+ * Arguments:	@const mpmont *mm@ = pointer to a Montgomery reduction
+ *			context
  *		*mp *d@ = pointer to mostly-reduced operand
  *
  * Returns:	---
@@ -205,7 +236,7 @@ static mulcore__functype *pick_mulcore(void)
  *		need to do an additional subtraction if %$d$% is negative.
  */
 
-static void finish(mpmont *mm, mp *d)
+static void finish(const mpmont *mm, mp *d)
 {
   mpw *dv = d->v, *dvl = d->vl;
   size_t n = mm->n;
@@ -310,7 +341,7 @@ void mpmont_destroy(mpmont *mm)
 
 /* --- @mpmont_reduce@ --- *
  *
- * Arguments:	@mpmont *mm@ = pointer to Montgomery reduction context
+ * Arguments:	@const mpmont *mm@ = pointer to Montgomery reduction context
  *		@mp *d@ = destination
  *		@mp *a@ = source, assumed positive
  *
@@ -319,7 +350,7 @@ void mpmont_destroy(mpmont *mm)
 
 #ifdef MPMONT_DISABLE
 
-mp *mpmont_reduce(mpmont *mm, mp *d, mp *a)
+mp *mpmont_reduce(const mpmont *mm, mp *d, mp *a)
 {
   mp_div(0, &d, a, mm->m);
   return (d);
@@ -327,13 +358,13 @@ mp *mpmont_reduce(mpmont *mm, mp *d, mp *a)
 
 #else
 
-mp *mpmont_reduce(mpmont *mm, mp *d, mp *a)
+mp *mpmont_reduce(const mpmont *mm, mp *d, mp *a)
 {
   size_t n = mm->n;
 
   /* --- Check for serious Karatsuba reduction --- */
 
-  if (n > MPK_THRESH * 3) {
+  if (n > MPMONT_KTHRESH) {
     mp al;
     mpw *vl;
     mp *u;
@@ -369,7 +400,7 @@ mp *mpmont_reduce(mpmont *mm, mp *d, mp *a)
 
 /* --- @mpmont_mul@ --- *
  *
- * Arguments:	@mpmont *mm@ = pointer to Montgomery reduction context
+ * Arguments:	@const mpmont *mm@ = pointer to Montgomery reduction context
  *		@mp *d@ = destination
  *		@mp *a, *b@ = sources, assumed positive
  *
@@ -378,7 +409,7 @@ mp *mpmont_reduce(mpmont *mm, mp *d, mp *a)
 
 #ifdef MPMONT_DISABLE
 
-mp *mpmont_mul(mpmont *mm, mp *d, mp *a, mp *b)
+mp *mpmont_mul(const mpmont *mm, mp *d, mp *a, mp *b)
 {
   d = mp_mul(d, a, b);
   mp_div(0, &d, d, mm->m);
@@ -387,11 +418,11 @@ mp *mpmont_mul(mpmont *mm, mp *d, mp *a, mp *b)
 
 #else
 
-mp *mpmont_mul(mpmont *mm, mp *d, mp *a, mp *b)
+mp *mpmont_mul(const mpmont *mm, mp *d, mp *a, mp *b)
 {
   size_t n = mm->n;
 
-  if (n > MPK_THRESH * 3) {
+  if (n > MPMONT_KTHRESH) {
     d = mp_mul(d, a, b);
     d = mpmont_reduce(mm, d, d);
   } else {
@@ -413,6 +444,10 @@ mp *mpmont_mul(mpmont *mm, mp *d, mp *a, mp *b)
 
 #ifdef TEST_RIG
 
+#ifdef ENABLE_ASM_DEBUG
+#  include "regdump.h"
+#endif
+
 static int tcreate(dstr *v)
 {
   mp *m = *(mp **)v[0].buf;
@@ -508,7 +543,6 @@ static int tmul(dstr *v)
     mp_drop(mr);
   }
 
-
   MP_DROP(m);
   MP_DROP(a);
   MP_DROP(b);
@@ -527,6 +561,9 @@ static test_chunk tests[] = {
 int main(int argc, char *argv[])
 {
   sub_init();
+#ifdef ENABLE_ASM_DEBUG
+  regdump_init();
+#endif
   test_run(argc, argv, tests, SRCDIR "/t/mpmont");
   return (0);
 }