X-Git-Url: https://git.distorted.org.uk/~mdw/catacomb/blobdiff_plain/1e8abdf50f7f376628db1f084f1fd64279765dcb..HEAD:/math/mpx-mul4-test.c

diff --git a/math/mpx-mul4-test.c b/math/mpx-mul4-test.c
index 355c7ceb..5325db0f 100644
--- a/math/mpx-mul4-test.c
+++ b/math/mpx-mul4-test.c
@@ -29,6 +29,10 @@
 
 #include "config.h"
 
+#ifdef ENABLE_ASM_DEBUG
+#  include "regdump.h"
+#endif
+
 #include <assert.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -57,6 +61,18 @@ static int cpu_features_p(void) { return (cpu_feature_p(CPUFEAT_X86_SSE2)); }
 static int cpu_features_p(void) { return (cpu_feature_p(CPUFEAT_X86_SSE2)); }
 #endif
 
+#if CPUFAM_ARMEL
+#  define VARIANT _arm_neon
+#  define REPR_32
+static int cpu_features_p(void) { return (cpu_feature_p(CPUFEAT_ARM_NEON)); }
+#endif
+
+#if CPUFAM_ARM64
+#  define VARIANT _arm64_simd
+#  define REPR_32
+static int cpu_features_p(void) { return (1); }
+#endif
+
 #ifndef VARIANT
 #  error "Unsupported CPU family."
 #endif
@@ -80,7 +96,9 @@ typedef struct { mpd w[6]; } carry;
   _(dmul4, NIL,  CARRY, P128, P128, P128, P128, P128, NIL,  CARRY)	\
   _(dmla4, P128, CARRY, P128, P128, P128, P128, P128, NIL,  CARRY)	\
   _(mul4,  NIL,	 CARRY, NIL,  P128, NIL,  P128, P128, NIL,  CARRY)	\
+  _(mul4zc,NIL,	 NIL,	NIL,  P128, NIL,  P128, P128, NIL,  CARRY)	\
   _(mla4,  P128, CARRY, NIL,  P128, NIL,  P128, P128, NIL,  CARRY)	\
+  _(mla4zc,P128, NIL,	NIL,  P128, NIL,  P128, P128, NIL,  CARRY)	\
   _(mmul4, NIL,	 NIL,	P128, P128, P128, P128, P128, X128, CARRY)	\
   _(mmla4, P128, NIL,	P128, P128, P128, P128, P128, X128, CARRY)	\
   _(mont4, P128, NIL,	NIL,  P128, NIL,  P128, P128, X128, CARRY)
@@ -105,7 +123,37 @@ TESTOPS(DECLSTUB)
 
 /*----- Conversion functions ----------------------------------------------*/
 
-#define DEFTYPE(ty, ld, st, nby)					\
+static mp *combine_mpw(mp *d, const mpw *v, size_t n, unsigned off)
+{
+  size_t i;
+  unsigned o;
+  mp m, *t = d;
+  mpw w[1];
+
+  d = MP_ZERO;
+  for (i = 0, o = 0; i < n; i++, o += off) {
+    w[0] = v[i]; mp_build(&m, w, w + 1);
+    t = mp_lsl(t, &m, o); d = mp_add(d, d, t);
+  }
+  mp_drop(t); return (d);
+}
+
+static mp *combine_mpd(mp *d, const mpd *v, size_t n, unsigned off)
+{
+  size_t i;
+  unsigned o;
+  mp m, *t = d;
+  mpw w[2];
+
+  d = MP_ZERO;
+  for (i = 0, o = 0; i < n; i++, o += off) {
+    w[0] = MPW(v[i]); w[1] = MPW(v[i] >> MPW_BITS); mp_build(&m, w, w + 2);
+    t = mp_lsl(t, &m, o); d = mp_add(d, d, t);
+  }
+  mp_drop(t); return (d);
+}
+
+#define DEFTYPE(ty, ld, st, nby, combfn, off)				\
 									\
   static void cvt_##ty(const char *buf, dstr *d)			\
   {									\
@@ -116,7 +164,7 @@ TESTOPS(DECLSTUB)
 									\
     type_hex.cvt(buf, &dd);						\
     if (dd.len != N(x->w)*nby) die(1, "invalid length for " #ty);	\
-    dstr_ensure(d, sizeof(*x));						\
+    dstr_ensure(d, sizeof(*x)); d->len = sizeof(*x);			\
     x = (ty *)d->buf; p = (const octet *)dd.buf;			\
     for (i = 0; i < N(x->w); i++) { x->w[i] = ld(p); p += nby; }	\
     dstr_destroy(&dd);							\
@@ -127,6 +175,7 @@ TESTOPS(DECLSTUB)
     dstr dd = DSTR_INIT;						\
     int i;								\
     const ty *x = (const ty *)d->buf;					\
+    mp *xx = combfn(MP_NEW, x->w, N(x->w), off);			\
     octet *p;								\
 									\
     dstr_ensure(&dd, N(x->w)*nby); p = (octet *)dd.buf;			\
@@ -134,22 +183,26 @@ TESTOPS(DECLSTUB)
     dd.len = N(x->w)*nby;						\
     type_hex.dump(&dd, fp);						\
     dstr_destroy(&dd);							\
+									\
+    fputs(" = 0x", fp); mp_writefile(xx, fp, 16);			\
+    fputs(" = ", fp); mp_writefile(xx, fp, 10);				\
+    MP_DROP(xx);							\
   }									\
 									\
   static int eq_##ty(const ty *x, const ty *y)				\
   {									\
-    int i;								\
-									\
-    for (i = 0; i < N(x->w); i++)					\
-      if (x->w[i] != y->w[i]) return (0);				\
-    return (1);								\
+    mp *xx = combfn(MP_NEW, x->w, N(x->w), off),			\
+      *yy = combfn(MP_NEW, y->w, N(y->w), off);				\
+    int rc = MP_EQ(xx, yy);						\
+    MP_DROP(xx); MP_DROP(yy);						\
+    return (rc);							\
   }									\
 									\
   static const struct test_type type_##ty = { cvt_##ty, dump_##ty };
 
-DEFTYPE(p128, LDW, STW, NWBY)
-DEFTYPE(x128, LDW, STW, NWBY)
-DEFTYPE(carry, LDD, STD, NDBY)
+DEFTYPE(p128, LDW, STW, NWBY, combine_mpw, MPW_BITS)
+DEFTYPE(x128, LDW, STW, NWBY, combine_mpw, MPW_BITS/2)
+DEFTYPE(carry, LDD, STD, NDBY, combine_mpd, MPW_BITS/2)
 
 /*----- Test functions ----------------------------------------------------*/
 
@@ -282,6 +335,9 @@ static test_chunk tests[] = {
 int main(int argc, char *argv[])
 {
   sub_init();
+#ifdef ENABLE_ASM_DEBUG
+  regdump_init();
+#endif
   if (!cpu_features_p())
     { fprintf(stderr, "required cpu feature not available\n"); exit(77); }
   test_run(argc, argv, tests, SRCDIR "/t/mpx-mul4");