Merge #13386: SHA256 implementations based on Intel SHA Extensions

66b2cf1 Use immintrin.h everywhere for intrinsics (Pieter Wuille)
4c935e2 Add SHA256 implementation using using Intel SHA intrinsics (Pieter Wuille)
268400d [Refactor] CPU feature detection logic for SHA256 (Pieter Wuille)

Pull request description:

  Based on #13191.

  This adds SHA256 implementations that use Intel's SHA Extension instructions (using intrinsics). This needs GCC 4.9 or Clang 3.4.

  In addition to #13191, two extra implementations are provided:
  * (a) A variable-length SHA256 implementation using SHA extensions.
  * (b) A 2-way 64-byte input double-SHA256 implementation using SHA extensions.

  Benchmarks for 9001-element Merkle tree root computation on an AMD Ryzen 1800X system:
  * Using generic C++ code (pre-#10821): 6.1ms
  * Using SSE4 (master, #10821): 4.6ms
  * Using 4-way SSE4 specialized for 64-byte inputs (#13191): 2.8ms
  * Using 8-way AVX2 specialized for 64-byte inputs (#13191): 2.1ms
  * Using 2-way SHA-NI specialized for 64-byte inputs (this PR): 0.56ms

  Benchmarks for 32-byte SHA256 on the same system:
  * Using SSE4 (master, #10821): 190ns
  * Using SHA-NI (this PR): 53ns

  Benchmarks for 1000000-byte SHA256 on the same system:
  * Using SSE4 (master, #10821): 2.5ms
  * Using SHA-NI (this PR): 0.51ms

Tree-SHA512: 2b319e33b22579f815d91f9daf7994a5e1e799c4f73c13e15070dd54ba71f3f6438ccf77ae9cbd1ce76f972d9cbeb5f0edfea3d86f101bbc1055db70e42743b7

Author: laanwj

configure.ac

@@ -320,6 +320,7 @@ fi
 AX_CHECK_COMPILE_FLAG([-msse4.2],[[SSE42_CXXFLAGS="-msse4.2"]],,[[$CXXFLAG_WERROR]])
 AX_CHECK_COMPILE_FLAG([-msse4.1],[[SSE41_CXXFLAGS="-msse4.1"]],,[[$CXXFLAG_WERROR]])
 AX_CHECK_COMPILE_FLAG([-mavx -mavx2],[[AVX2_CXXFLAGS="-mavx -mavx2"]],,[[$CXXFLAG_WERROR]])
+AX_CHECK_COMPILE_FLAG([-msse4 -msha],[[SHANI_CXXFLAGS="-msse4 -msha"]],,[[$CXXFLAG_WERROR]])
 
 TEMP_CXXFLAGS="$CXXFLAGS"
 CXXFLAGS="$CXXFLAGS $SSE42_CXXFLAGS"
@@ -348,11 +349,7 @@ CXXFLAGS="$CXXFLAGS $SSE41_CXXFLAGS"
 AC_MSG_CHECKING(for SSE4.1 intrinsics)
 AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
     #include <stdint.h>
-    #if defined(_MSC_VER)
     #include <immintrin.h>
-    #elif defined(__GNUC__)
-    #include <x86intrin.h>
-    #endif
   ]],[[
     __m128i l = _mm_set1_epi32(0);
     return _mm_extract_epi32(l, 3);
@@ -367,11 +364,7 @@ CXXFLAGS="$CXXFLAGS $AVX2_CXXFLAGS"
 AC_MSG_CHECKING(for AVX2 intrinsics)
 AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
     #include <stdint.h>
-    #if defined(_MSC_VER)
     #include <immintrin.h>
-    #elif defined(__GNUC__) && defined(__AVX2__)
-    #include <x86intrin.h>
-    #endif
   ]],[[
     __m256i l = _mm256_set1_epi32(0);
     return _mm256_extract_epi32(l, 7);
@@ -381,6 +374,23 @@ AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
 )
 CXXFLAGS="$TEMP_CXXFLAGS"
 
+TEMP_CXXFLAGS="$CXXFLAGS"
+CXXFLAGS="$CXXFLAGS $SHANI_CXXFLAGS"
+AC_MSG_CHECKING(for SHA-NI intrinsics)
+AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[
+    #include <stdint.h>
+    #include <immintrin.h>
+  ]],[[
+    __m128i i = _mm_set1_epi32(0);
+    __m128i j = _mm_set1_epi32(1);
+    __m128i k = _mm_set1_epi32(2);
+    return _mm_extract_epi32(_mm_sha256rnds2_epu32(i, i, k), 0);
+  ]])],
+ [ AC_MSG_RESULT(yes); enable_shani=yes; AC_DEFINE(ENABLE_SHANI, 1, [Define this symbol to build code that uses SHA-NI intrinsics]) ],
+ [ AC_MSG_RESULT(no)]
+)
+CXXFLAGS="$TEMP_CXXFLAGS"
+
 CPPFLAGS="$CPPFLAGS -DHAVE_BUILD_INFO -D__STDC_FORMAT_MACROS"
 
 AC_ARG_WITH([utils],
@@ -1309,6 +1319,7 @@ AM_CONDITIONAL([HARDEN],[test x$use_hardening = xyes])
 AM_CONDITIONAL([ENABLE_HWCRC32],[test x$enable_hwcrc32 = xyes])
 AM_CONDITIONAL([ENABLE_SSE41],[test x$enable_sse41 = xyes])
 AM_CONDITIONAL([ENABLE_AVX2],[test x$enable_avx2 = xyes])
+AM_CONDITIONAL([ENABLE_SHANI],[test x$enable_shani = xyes])
 AM_CONDITIONAL([USE_ASM],[test x$use_asm = xyes])
 
 AC_DEFINE(CLIENT_VERSION_MAJOR, _CLIENT_VERSION_MAJOR, [Major version])
@@ -1353,6 +1364,7 @@ AC_SUBST(SANITIZER_LDFLAGS)
 AC_SUBST(SSE42_CXXFLAGS)
 AC_SUBST(SSE41_CXXFLAGS)
 AC_SUBST(AVX2_CXXFLAGS)
+AC_SUBST(SHANI_CXXFLAGS)
 AC_SUBST(LIBTOOL_APP_LDFLAGS)
 AC_SUBST(USE_UPNP)
 AC_SUBST(USE_QRCODE)

src/Makefile.am

@@ -52,6 +52,10 @@ if ENABLE_AVX2
 LIBBITCOIN_CRYPTO_AVX2 = crypto/libbitcoin_crypto_avx2.a
 LIBBITCOIN_CRYPTO += $(LIBBITCOIN_CRYPTO_AVX2)
 endif
+if ENABLE_SHANI
+LIBBITCOIN_CRYPTO_SHANI = crypto/libbitcoin_crypto_shani.a
+LIBBITCOIN_CRYPTO += $(LIBBITCOIN_CRYPTO_SHANI)
+endif
 
 $(LIBSECP256K1): $(wildcard secp256k1/src/*.h) $(wildcard secp256k1/src/*.c) $(wildcard secp256k1/include/*)
 	$(AM_V_at)$(MAKE) $(AM_MAKEFLAGS) -C $(@D) $(@F)
@@ -318,6 +322,12 @@ crypto_libbitcoin_crypto_avx2_a_CXXFLAGS += $(AVX2_CXXFLAGS)
 crypto_libbitcoin_crypto_avx2_a_CPPFLAGS += -DENABLE_AVX2
 crypto_libbitcoin_crypto_avx2_a_SOURCES = crypto/sha256_avx2.cpp
 
+crypto_libbitcoin_crypto_shani_a_CXXFLAGS = $(AM_CXXFLAGS) $(PIE_FLAGS)
+crypto_libbitcoin_crypto_shani_a_CPPFLAGS = $(AM_CPPFLAGS)
+crypto_libbitcoin_crypto_shani_a_CXXFLAGS += $(SHANI_CXXFLAGS)
+crypto_libbitcoin_crypto_shani_a_CPPFLAGS += -DENABLE_SHANI
+crypto_libbitcoin_crypto_shani_a_SOURCES = crypto/sha256_shani.cpp
+
 # consensus: shared between all executables that validate any consensus rules.
 libbitcoin_consensus_a_CPPFLAGS = $(AM_CPPFLAGS) $(BITCOIN_INCLUDES)
 libbitcoin_consensus_a_CXXFLAGS = $(AM_CXXFLAGS) $(PIE_FLAGS)

src/Makefile.test.include

@@ -137,6 +137,7 @@ test_test_bitcoin_fuzzy_LDADD = \
   $(LIBBITCOIN_CRYPTO) \
   $(LIBBITCOIN_CRYPTO_SSE41) \
   $(LIBBITCOIN_CRYPTO_AVX2) \
+  $(LIBBITCOIN_CRYPTO_SHANI) \
   $(LIBSECP256K1)
 
 test_test_bitcoin_fuzzy_LDADD += $(BOOST_LIBS) $(CRYPTO_LIBS)

src/crypto/sha256.cpp

@@ -29,6 +29,16 @@ namespace sha256d64_avx2
 void Transform_8way(unsigned char* out, const unsigned char* in);
 }
 
+namespace sha256d64_shani
+{
+void Transform_2way(unsigned char* out, const unsigned char* in);
+}
+
+namespace sha256_shani
+{
+void Transform(uint32_t* s, const unsigned char* chunk, size_t blocks);
+}
+
 // Internal implementation code.
 namespace
 {
@@ -448,6 +458,7 @@ void TransformD64Wrapper(unsigned char* out, const unsigned char* in)
 
 TransformType Transform = sha256::Transform;
 TransformD64Type TransformD64 = sha256::TransformD64;
+TransformD64Type TransformD64_2way = nullptr;
 TransformD64Type TransformD64_4way = nullptr;
 TransformD64Type TransformD64_8way = nullptr;
 
@@ -512,6 +523,13 @@ bool SelfTest() {
     TransformD64(out, data + 1);
     if (!std::equal(out, out + 32, result_d64)) return false;
 
+    // Test TransformD64_2way, if available.
+    if (TransformD64_2way) {
+        unsigned char out[64];
+        TransformD64_2way(out, data + 1);
+        if (!std::equal(out, out + 64, result_d64)) return false;
+    }
+
     // Test TransformD64_4way, if available.
     if (TransformD64_4way) {
         unsigned char out[128];
@@ -556,32 +574,64 @@ std::string SHA256AutoDetect()
 {
     std::string ret = "standard";
 #if defined(USE_ASM) && (defined(__x86_64__) || defined(__amd64__) || defined(__i386__))
-    (void)AVXEnabled; // Silence unused warning (in case ENABLE_AVX2 is not defined)
+    bool have_sse4 = false;
+    bool have_xsave = false;
+    bool have_avx = false;
+    bool have_avx2 = false;
+    bool have_shani = false;
+    bool enabled_avx = false;
+
+    (void)AVXEnabled;
+    (void)have_sse4;
+    (void)have_avx;
+    (void)have_xsave;
+    (void)have_avx2;
+    (void)have_shani;
+    (void)enabled_avx;
+
     uint32_t eax, ebx, ecx, edx;
     cpuid(1, 0, eax, ebx, ecx, edx);
-    if ((ecx >> 19) & 1) {
+    have_sse4 = (ecx >> 19) & 1;
+    have_xsave = (ecx >> 27) & 1;
+    have_avx = (ecx >> 28) & 1;
+    if (have_xsave && have_avx) {
+        enabled_avx = AVXEnabled();
+    }
+    if (have_sse4) {
+        cpuid(7, 0, eax, ebx, ecx, edx);
+        have_avx2 = (ebx >> 5) & 1;
+        have_shani = (ebx >> 29) & 1;
+    }
+
+#if defined(ENABLE_SHANI) && !defined(BUILD_BITCOIN_INTERNAL)
+    if (have_shani) {
+        Transform = sha256_shani::Transform;
+        TransformD64 = TransformD64Wrapper<sha256_shani::Transform>;
+        TransformD64_2way = sha256d64_shani::Transform_2way;
+        ret = "shani(1way,2way)";
+        have_sse4 = false; // Disable SSE4/AVX2;
+        have_avx2 = false;
+    }
+#endif
+
+    if (have_sse4) {
 #if defined(__x86_64__) || defined(__amd64__)
         Transform = sha256_sse4::Transform;
         TransformD64 = TransformD64Wrapper<sha256_sse4::Transform>;
+        ret = "sse4(1way)";
 #endif
 #if defined(ENABLE_SSE41) && !defined(BUILD_BITCOIN_INTERNAL)
         TransformD64_4way = sha256d64_sse41::Transform_4way;
-        ret = "sse4(1way+4way)";
-#if defined(ENABLE_AVX2) && !defined(BUILD_BITCOIN_INTERNAL)
-        if (((ecx >> 27) & 1) && ((ecx >> 28) & 1)) { // XSAVE and AVX
-            cpuid(7, 0, eax, ebx, ecx, edx);
-            if ((ebx >> 5) & 1) { // AVX2 flag
-                if (AVXEnabled()) { // OS has enabled AVX registers
-                    TransformD64_8way = sha256d64_avx2::Transform_8way;
-                    ret += ",avx2(8way)";
-                }
-            }
-        }
-#endif
-#else
-        ret = "sse4";
+        ret += ",sse41(4way)";
 #endif
     }
+
+#if defined(ENABLE_AVX2) && !defined(BUILD_BITCOIN_INTERNAL)
+    if (have_avx2 && have_avx && enabled_avx) {
+        TransformD64_8way = sha256d64_avx2::Transform_8way;
+        ret += ",avx2(8way)";
+    }
+#endif
 #endif
 
     assert(SelfTest());
@@ -663,6 +713,14 @@ void SHA256D64(unsigned char* out, const unsigned char* in, size_t blocks)
             blocks -= 4;
         }
     }
+    if (TransformD64_2way) {
+        while (blocks >= 2) {
+            TransformD64_2way(out, in);
+            out += 64;
+            in += 128;
+            blocks -= 2;
+        }
+    }
     while (blocks) {
         TransformD64(out, in);
         out += 32;

src/crypto/sha256_avx2.cpp

@@ -1,11 +1,7 @@
 #ifdef ENABLE_AVX2
 
 #include <stdint.h>
-#if defined(_MSC_VER)
 #include <immintrin.h>
-#elif defined(__GNUC__)
-#include <x86intrin.h>
-#endif
 
 #include <crypto/sha256.h>
 #include <crypto/common.h>