Merge pull request #110 from aqjune-aws/neontest

Update test,benchmark to assume that arm has the Neon extension

benchmarks/benchmark.c

@@ -767,7 +767,23 @@ void call_sm2_montjadd(void) repeat(sm2_montjadd(b1,b2,b3))
 void call_sm2_montjdouble(void) repeat(sm2_montjdouble(b1,b2))
 void call_sm2_montjmixadd(void) repeat(sm2_montjmixadd(b1,b2,b3))
 
-#ifdef __ARM_NEON
+#ifdef __x86_64__
+
+void call_bignum_copy_row_from_table_8n_neon__32_16(void) {}
+void call_bignum_copy_row_from_table_8n_neon__32_32(void) {}
+void call_bignum_copy_row_from_table_16_neon__32(void) {}
+void call_bignum_copy_row_from_table_32_neon__32(void) {}
+
+void call_bignum_emontredc_8n_neon__32(void) {}
+void call_bignum_kmul_16_32_neon(void) {}
+void call_bignum_ksqr_16_32_neon(void) {}
+void call_bignum_kmul_32_64_neon(void) {}
+void call_bignum_ksqr_32_64_neon(void) {}
+void call_bignum_mul_8_16_neon(void) {}
+void call_bignum_sqr_8_16_neon(void) {}
+
+#else
+
 void call_bignum_copy_row_from_table_8n_neon__32_16(void) \
     repeat(bignum_copy_row_from_table_8n_neon(b0,b1,32,16,0))
 void call_bignum_copy_row_from_table_8n_neon__32_32(void) \
@@ -785,27 +801,13 @@ void call_bignum_ksqr_32_64_neon(void) repeat(bignum_ksqr_32_64_neon(b0,b1,b2))
 void call_bignum_mul_8_16_neon(void) repeat(bignum_mul_8_16_neon(b0,b1,b2))
 void call_bignum_sqr_8_16_neon(void) repeat(bignum_sqr_8_16_neon(b0,b1))
 
-#else
-void call_bignum_copy_row_from_table_8n_neon__32_16(void) {}
-void call_bignum_copy_row_from_table_8n_neon__32_32(void) {}
-void call_bignum_copy_row_from_table_16_neon__32(void) {}
-void call_bignum_copy_row_from_table_32_neon__32(void) {}
-
-void call_bignum_emontredc_8n_neon__32(void) {}
-void call_bignum_kmul_16_32_neon(void) {}
-void call_bignum_ksqr_16_32_neon(void) {}
-void call_bignum_kmul_32_64_neon(void) {}
-void call_bignum_ksqr_32_64_neon(void) {}
-void call_bignum_mul_8_16_neon(void) {}
-void call_bignum_sqr_8_16_neon(void) {}
-
 #endif
 
 int main(int argc, char *argv[])
 {
   int bmi = get_arch_name() == ARCH_AARCH64 || supports_bmi2_and_adx();
   int all = 1;
-  int neon = supports_neon();
+  int arm = get_arch_name() == ARCH_AARCH64;
   char *argending;
   long negreps;
   function_to_test = "";
@@ -910,10 +912,14 @@ int main(int argc, char *argv[])
   timingtest(all,"bignum_copy (32 -> 32)" ,call_bignum_copy__32_32);
   timingtest(all,"bignum_copy_row_from_table (h=32,w=16)",call_bignum_copy_row_from_table__32_16);
   timingtest(all,"bignum_copy_row_from_table (h=32,w=32)",call_bignum_copy_row_from_table__32_32);
-  timingtest(neon,"bignum_copy_row_from_table_8n_neon (h=32,w=16)",call_bignum_copy_row_from_table_8n_neon__32_16);
-  timingtest(neon,"bignum_copy_row_from_table_8n_neon (h=32,w=32)",call_bignum_copy_row_from_table_8n_neon__32_32);
-  timingtest(neon,"bignum_copy_row_from_table_16_neon (h=32)",call_bignum_copy_row_from_table_16_neon__32);
-  timingtest(neon,"bignum_copy_row_from_table_32_neon (h=32)",call_bignum_copy_row_from_table_32_neon__32);
+  timingtest(arm, "bignum_copy_row_from_table_8n_neon (h=32,w=16)",
+             call_bignum_copy_row_from_table_8n_neon__32_16);
+  timingtest(arm, "bignum_copy_row_from_table_8n_neon (h=32,w=32)",
+             call_bignum_copy_row_from_table_8n_neon__32_32);
+  timingtest(arm, "bignum_copy_row_from_table_16_neon (h=32)",
+             call_bignum_copy_row_from_table_16_neon__32);
+  timingtest(arm, "bignum_copy_row_from_table_32_neon (h=32)",
+             call_bignum_copy_row_from_table_32_neon__32);
   timingtest(all,"bignum_ctd (32)" ,call_bignum_ctd__32);
   timingtest(all,"bignum_ctz (32)" ,call_bignum_ctz__32);
   timingtest(bmi,"bignum_deamont_p256",call_bignum_deamont_p256);
@@ -944,7 +950,8 @@ int main(int argc, char *argv[])
   timingtest(all,"bignum_emontredc (12 -> 6)",call_bignum_emontredc__6);
   timingtest(all,"bignum_emontredc (64 -> 32)",call_bignum_emontredc__32);
   timingtest(bmi,"bignum_emontredc_8n (64 -> 32)",call_bignum_emontredc_8n__32);
-  timingtest(neon,"bignum_emontredc_8n_neon (64 -> 32)",call_bignum_emontredc_8n_neon__32);
+  timingtest(arm, "bignum_emontredc_8n_neon (64 -> 32)",
+             call_bignum_emontredc_8n_neon__32);
   timingtest(all,"bignum_eq (32x32)" ,call_bignum_eq__32_32);
   timingtest(all,"bignum_even (32)" ,call_bignum_even__32);
   timingtest(all,"bignum_frombebytes_4",call_bignum_frombebytes_4);
@@ -964,13 +971,13 @@ int main(int argc, char *argv[])
   timingtest(all,"bignum_invsqrt_p25519_alt",call_bignum_invsqrt_p25519_alt);
   timingtest(all,"bignum_iszero (32)" ,call_bignum_iszero__32);
   timingtest(bmi,"bignum_kmul_16_32",call_bignum_kmul_16_32);
-  timingtest(neon,"bignum_kmul_16_32_neon",call_bignum_kmul_16_32_neon);
+  timingtest(arm, "bignum_kmul_16_32_neon", call_bignum_kmul_16_32_neon);
   timingtest(bmi,"bignum_kmul_32_64",call_bignum_kmul_32_64);
-  timingtest(neon,"bignum_kmul_32_64_neon",call_bignum_kmul_32_64_neon);
+  timingtest(arm, "bignum_kmul_32_64_neon", call_bignum_kmul_32_64_neon);
   timingtest(bmi,"bignum_ksqr_16_32",call_bignum_ksqr_16_32);
-  timingtest(neon,"bignum_ksqr_16_32_neon",call_bignum_ksqr_16_32_neon);
+  timingtest(arm, "bignum_ksqr_16_32_neon", call_bignum_ksqr_16_32_neon);
   timingtest(bmi,"bignum_ksqr_32_64",call_bignum_ksqr_32_64);
-  timingtest(neon,"bignum_ksqr_32_64_neon",call_bignum_ksqr_32_64_neon);
+  timingtest(arm, "bignum_ksqr_32_64_neon", call_bignum_ksqr_32_64_neon);
   timingtest(all,"bignum_le (32x32)" ,call_bignum_le__32_32);
   timingtest(all,"bignum_littleendian_4",call_bignum_littleendian_4);
   timingtest(all,"bignum_littleendian_6",call_bignum_littleendian_6);
@@ -1055,7 +1062,7 @@ int main(int argc, char *argv[])
   timingtest(all,"bignum_mul_6_12_alt",call_bignum_mul_6_12_alt);
   timingtest(bmi,"bignum_mul_8_16",call_bignum_mul_8_16);
   timingtest(all,"bignum_mul_8_16_alt",call_bignum_mul_8_16_alt);
-  timingtest(neon,"bignum_mul_8_16_neon",call_bignum_mul_8_16_neon);
+  timingtest(arm, "bignum_mul_8_16_neon", call_bignum_mul_8_16_neon);
   timingtest(bmi,"bignum_mul_p25519",call_bignum_mul_p25519);
   timingtest(all,"bignum_mul_p25519_alt",call_bignum_mul_p25519_alt);
   timingtest(bmi,"bignum_mul_p256k1",call_bignum_mul_p256k1);
@@ -1105,7 +1112,7 @@ int main(int argc, char *argv[])
   timingtest(all,"bignum_sqr_6_12_alt",call_bignum_sqr_6_12_alt);
   timingtest(bmi,"bignum_sqr_8_16",call_bignum_sqr_8_16);
   timingtest(all,"bignum_sqr_8_16_alt",call_bignum_sqr_8_16_alt);
-  timingtest(neon,"bignum_sqr_8_16_neon",call_bignum_sqr_8_16_neon);
+  timingtest(arm, "bignum_sqr_8_16_neon", call_bignum_sqr_8_16_neon);
   timingtest(bmi,"bignum_sqr_p25519",call_bignum_sqr_p25519);
   timingtest(all,"bignum_sqr_p25519_alt",call_bignum_sqr_p25519_alt);
   timingtest(bmi,"bignum_sqr_p256k1",call_bignum_sqr_p256k1);

tests/arch.h

@@ -18,11 +18,6 @@ int supports_bmi2_and_adx(void)
   return (c & (1ul<<8)) && (c & (1ul<<19));
 }
 
-int supports_neon(void)
-{ // X86_64 does not support NEON.
-  return 0;
-}
-
 enum arch_name get_arch_name()
 { return ARCH_X86_64;
 }
@@ -34,15 +29,6 @@ int supports_bmi2_and_adx(void)
   return 0;
 }
 
-int supports_neon(void)
-{
-#ifdef __ARM_NEON
-  return 1;
-#else
-  return 0;
-#endif
-}
-
 enum arch_name get_arch_name()
 { return ARCH_AARCH64;
 }

tests/test.c

@@ -2846,18 +2846,25 @@ int test_bignum_copy_row_from_table(void)
 
 int test_bignum_copy_row_from_table_8n_neon(void)
 {
-#ifdef __ARM_NEON
-  return test_bignum_copy_row_from_table_specific("bignum_copy_row_from_table_8n_neon",
-      0, 1, bignum_copy_row_from_table_8n_neon);
-#else
+#ifdef __x86_64__
   return 1;
+#else
+  return test_bignum_copy_row_from_table_specific(
+      "bignum_copy_row_from_table_8n_neon", 0, 1,
+      bignum_copy_row_from_table_8n_neon);
 #endif
 }
 
-#ifdef __ARM_NEON
+#ifdef __x86_64__
+int test_bignum_copy_row_from_table_16_neon(void)
+{ return 1;
+}
+#else
 void _bignum_copy_row_from_table_16_neon_wrapper(uint64_t *z, uint64_t *table,
-    uint64_t height, uint64_t width, uint64_t index)
-{ assert(width == 16);
+                                                 uint64_t height,
+                                                 uint64_t width,
+                                                 uint64_t index) {
+  assert(width == 16);
   bignum_copy_row_from_table_16_neon(z, table, height, index);
 }
 
@@ -2866,16 +2873,18 @@ int test_bignum_copy_row_from_table_16_neon(void)
       "bignum_copy_row_from_table_16_neon", 16, 0,
       _bignum_copy_row_from_table_16_neon_wrapper);
 }
-#else
-int test_bignum_copy_row_from_table_16_neon(void)
-{ return 1;
-}
 #endif
 
-#ifdef __ARM_NEON
+#ifdef __x86_64__
+int test_bignum_copy_row_from_table_32_neon(void)
+{ return 1;
+}
+#else
 void _bignum_copy_row_from_table_32_neon_wrapper(uint64_t *z, uint64_t *table,
-    uint64_t height, uint64_t width, uint64_t index)
-{ assert(width == 32);
+                                                 uint64_t height,
+                                                 uint64_t width,
+                                                 uint64_t index) {
+  assert(width == 32);
   bignum_copy_row_from_table_32_neon(z, table, height, index);
 }
 
@@ -2884,10 +2893,6 @@ int test_bignum_copy_row_from_table_32_neon(void)
       "bignum_copy_row_from_table_32_neon", 32, 0,
       _bignum_copy_row_from_table_32_neon_wrapper);
 }
-#else
-int test_bignum_copy_row_from_table_32_neon(void)
-{ return 1;
-}
 #endif
 
 int test_bignum_ctd(void)
@@ -3762,12 +3767,12 @@ int test_bignum_emontredc_8n(void)
 
 int test_bignum_emontredc_8n_neon(void)
 {
-#ifdef __ARM_NEON
-  return test_bignum_emontredc_specific("bignum_emontredc_8n_neon", 1,
-                                        bignum_emontredc_8n_neon);
-#else
+#ifdef __x86_64__
   // Do not call the neon function to avoid a linking failure error.
   return 1;
+#else
+  return test_bignum_emontredc_specific("bignum_emontredc_8n_neon", 1,
+                                        bignum_emontredc_8n_neon);
 #endif
 }
 
@@ -4451,12 +4456,12 @@ int test_bignum_kmul_16_32(void)
 
 int test_bignum_kmul_16_32_neon(void)
 {
-#ifdef __ARM_NEON
-  return test_bignum_kmul_specific(32,16,16,"bignum_kmul_16_32_neon",
-                                   bignum_kmul_16_32_neon);
-#else
+#ifdef __x86_64__
   // Do not call the neon function to avoid a linking failure error.
   return 1;
+#else
+  return test_bignum_kmul_specific(32, 16, 16, "bignum_kmul_16_32_neon",
+                                   bignum_kmul_16_32_neon);
 #endif
 }
 
@@ -4466,12 +4471,12 @@ int test_bignum_kmul_32_64(void)
 
 int test_bignum_kmul_32_64_neon(void)
 {
-#ifdef __ARM_NEON
-  return test_bignum_kmul_specific(64,32,32,"bignum_kmul_32_64_neon",
-                                   bignum_kmul_32_64_neon);
-#else
+#ifdef __x86_64__
   // Do not call the neon function to avoid a linking failure error.
   return 1;
+#else
+  return test_bignum_kmul_specific(64, 32, 32, "bignum_kmul_32_64_neon",
+                                   bignum_kmul_32_64_neon);
 #endif
 }
 
@@ -4511,12 +4516,12 @@ int test_bignum_ksqr_16_32(void)
 
 int test_bignum_ksqr_16_32_neon(void)
 {
-#ifdef __ARM_NEON
-  return test_bignum_ksqr_specific(32,16,"bignum_ksqr_16_32_neon",
-                                   bignum_ksqr_16_32_neon);
-#else
+#ifdef __x86_64__
   // Do not call the neon function to avoid a linking failure error.
   return 1;
+#else
+  return test_bignum_ksqr_specific(32, 16, "bignum_ksqr_16_32_neon",
+                                   bignum_ksqr_16_32_neon);
 #endif
 }
 
@@ -4526,12 +4531,12 @@ int test_bignum_ksqr_32_64(void)
 
 int test_bignum_ksqr_32_64_neon(void)
 {
-#ifdef __ARM_NEON
-  return test_bignum_ksqr_specific(64,32,"bignum_ksqr_32_64_neon",
-                                   bignum_ksqr_32_64_neon);
-#else
+#ifdef __x86_64__
   // Do not call the neon function to avoid a linking failure error.
   return 1;
+#else
+  return test_bignum_ksqr_specific(64, 32, "bignum_ksqr_32_64_neon",
+                                   bignum_ksqr_32_64_neon);
 #endif
 }
 
@@ -6698,12 +6703,12 @@ int test_bignum_mul_8_16_alt(void)
 
 int test_bignum_mul_8_16_neon(void)
 {
-#ifdef __ARM_NEON
-  return test_bignum_mul_specific(16, 8, 8, "bignum_mul_8_16_neon",
-                                  bignum_mul_8_16_neon);
-#else
+#ifdef __x86_64__
   // Do not call the neon function to avoid a linking failure error.
   return 1;
+#else
+  return test_bignum_mul_specific(16, 8, 8, "bignum_mul_8_16_neon",
+                                  bignum_mul_8_16_neon);
 #endif
 }
 
@@ -8024,11 +8029,12 @@ int test_bignum_sqr_8_16_alt(void)
 
 int test_bignum_sqr_8_16_neon(void)
 {
-#ifdef __ARM_NEON
-  return test_bignum_sqr_specific(16,8,"bignum_sqr_8_16_neon",bignum_sqr_8_16_neon);
-#else
+#ifdef __x86_64__
   // Do not call the neon function to avoid a linking failure error.
   return 1;
+#else
+  return test_bignum_sqr_specific(16, 8, "bignum_sqr_8_16_neon",
+                                  bignum_sqr_8_16_neon);
 #endif
 }
 
@@ -11987,17 +11993,16 @@ int main(int argc, char *argv[])
   functionaltest(all,"word_recip",test_word_recip);
 
   if (get_arch_name() == ARCH_AARCH64) {
-    int neon = supports_neon();
-    functionaltest(neon,"bignum_copy_row_from_table_8n_neon",test_bignum_copy_row_from_table_8n_neon);
-    functionaltest(neon,"bignum_copy_row_from_table_16_neon",test_bignum_copy_row_from_table_16_neon);
-    functionaltest(neon,"bignum_copy_row_from_table_32_neon",test_bignum_copy_row_from_table_32_neon);
-    functionaltest(neon,"bignum_emontredc_8n_neon",test_bignum_emontredc_8n_neon);
-    functionaltest(neon,"bignum_kmul_16_32_neon", test_bignum_kmul_16_32_neon);
-    functionaltest(neon,"bignum_kmul_32_64_neon", test_bignum_kmul_32_64_neon);
-    functionaltest(neon,"bignum_ksqr_16_32_neon",test_bignum_ksqr_16_32_neon);
-    functionaltest(neon,"bignum_ksqr_32_64_neon",test_bignum_ksqr_32_64_neon);
-    functionaltest(neon,"bignum_mul_8_16_neon",test_bignum_mul_8_16_neon);
-    functionaltest(neon,"bignum_sqr_8_16_neon",test_bignum_sqr_8_16_neon);
+    functionaltest(all,"bignum_copy_row_from_table_8n_neon",test_bignum_copy_row_from_table_8n_neon);
+    functionaltest(all,"bignum_copy_row_from_table_16_neon",test_bignum_copy_row_from_table_16_neon);
+    functionaltest(all,"bignum_copy_row_from_table_32_neon",test_bignum_copy_row_from_table_32_neon);
+    functionaltest(all,"bignum_emontredc_8n_neon",test_bignum_emontredc_8n_neon);
+    functionaltest(all,"bignum_kmul_16_32_neon", test_bignum_kmul_16_32_neon);
+    functionaltest(all,"bignum_kmul_32_64_neon", test_bignum_kmul_32_64_neon);
+    functionaltest(all,"bignum_ksqr_16_32_neon",test_bignum_ksqr_16_32_neon);
+    functionaltest(all,"bignum_ksqr_32_64_neon",test_bignum_ksqr_32_64_neon);
+    functionaltest(all,"bignum_mul_8_16_neon",test_bignum_mul_8_16_neon);
+    functionaltest(all,"bignum_sqr_8_16_neon",test_bignum_sqr_8_16_neon);
   }
 
   if (extrastrigger) function_to_test = "_";