Sync BLAKE3

CMakeLists.txt

@@ -3,6 +3,8 @@ cmake_minimum_required(VERSION 3.27)
 
 project(bela CXX C ASM)
 
+include(FeatureSummary)
+
 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_BINARY_DIR AND NOT MSVC_IDE)
   message(
     FATAL_ERROR

src/belahash/CMakeLists.txt

@@ -1,70 +1,167 @@
 # bela::hash https://docs.microsoft.com/en-us/cpp/intrinsics/arm64-intrinsics
 # https://static.docs.arm.com/ihi0073/c/IHI0073C_arm_neon_intrinsics_ref.pdf
 
-set(BELA_BLAKE3_SOURCES blake3/blake3.c blake3/blake3_dispatch.c blake3/blake3_portable.c)
+# architecture lists for which to enable assembly / SIMD sources
+set(BLAKE3_AMD64_NAMES amd64 AMD64 x86_64)
+set(BLAKE3_X86_NAMES i686 x86 X86)
+set(BLAKE3_ARMv8_NAMES aarch64 AArch64 arm64 ARM64 armv8 armv8a)
+# default SIMD compiler flag configuration (can be overriden by toolchains or CLI)
+if(MSVC)
+  set(BLAKE3_CFLAGS_SSE2 "/arch:SSE2" CACHE STRING "the compiler flags to enable SSE2")
+  # MSVC has no dedicated sse4.1 flag (see https://learn.microsoft.com/en-us/cpp/build/reference/arch-x86?view=msvc-170)
+  set(BLAKE3_CFLAGS_SSE4.1 "/arch:AVX" CACHE STRING "the compiler flags to enable SSE4.1")
+  set(BLAKE3_CFLAGS_AVX2 "/arch:AVX2" CACHE STRING "the compiler flags to enable AVX2")
+  set(BLAKE3_CFLAGS_AVX512 "/arch:AVX512" CACHE STRING "the compiler flags to enable AVX512")
 
-if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
-  set(BELA_ENABLE_ASSEMBLY_FILES ON)
+  set(BLAKE3_AMD64_ASM_SOURCES
+    blake3/blake3_avx2_x86-64_windows_msvc.asm
+    blake3/blake3_avx512_x86-64_windows_msvc.asm
+    blake3/blake3_sse2_x86-64_windows_msvc.asm
+    blake3/blake3_sse41_x86-64_windows_msvc.asm
+  )
+
+elseif(CMAKE_C_COMPILER_ID STREQUAL "GNU"
+       OR CMAKE_C_COMPILER_ID STREQUAL "Clang"
+       OR CMAKE_C_COMPILER_ID STREQUAL "AppleClang")
+  set(BLAKE3_CFLAGS_SSE2 "-msse2" CACHE STRING "the compiler flags to enable SSE2")
+  set(BLAKE3_CFLAGS_SSE4.1 "-msse4.1" CACHE STRING "the compiler flags to enable SSE4.1")
+  set(BLAKE3_CFLAGS_AVX2 "-mavx2" CACHE STRING "the compiler flags to enable AVX2")
+  set(BLAKE3_CFLAGS_AVX512 "-mavx512f -mavx512vl" CACHE STRING "the compiler flags to enable AVX512")
+
+  if (WIN32)
+    set(BLAKE3_AMD64_ASM_SOURCES
+      blake3/blake3_avx2_x86-64_windows_gnu.S
+      blake3/blake3_avx512_x86-64_windows_gnu.S
+      blake3/blake3_sse2_x86-64_windows_gnu.S
+      blake3/blake3_sse41_x86-64_windows_gnu.S
+    )
+
+  elseif(UNIX)
+    set(BLAKE3_AMD64_ASM_SOURCES
+      blake3/blake3_avx2_x86-64_unix.S
+      blake3/blake3_avx512_x86-64_unix.S
+      blake3/blake3_sse2_x86-64_unix.S
+      blake3/blake3_sse41_x86-64_unix.S
+    )
+  endif()
+
+  if (CMAKE_SYSTEM_PROCESSOR IN_LIST BLAKE3_ARMv8_NAMES
+      AND NOT CMAKE_SIZEOF_VOID_P EQUAL 8)
+    # 32-bit ARMv8 needs NEON to be enabled explicitly
+    set(BLAKE3_CFLAGS_NEON "-mfpu=neon" CACHE STRING "the compiler flags to enable NEON")
+  endif()
 endif()
 
+mark_as_advanced(BLAKE3_CFLAGS_SSE2 BLAKE3_CFLAGS_SSE4.1 BLAKE3_CFLAGS_AVX2 BLAKE3_CFLAGS_AVX512 BLAKE3_CFLAGS_NEON)
+mark_as_advanced(BLAKE3_AMD64_ASM_SOURCES)
+
+message(STATUS "BLAKE3 SIMD configuration: ${CMAKE_C_COMPILER_ARCHITECTURE_ID}")
+if(MSVC AND DEFINED CMAKE_C_COMPILER_ARCHITECTURE_ID)
+  if(CMAKE_C_COMPILER_ARCHITECTURE_ID MATCHES "[Xx]86")
+    set(BLAKE3_SIMD_TYPE "x86-intrinsics" CACHE STRING "the SIMD acceleration type to use")
+
+  elseif(CMAKE_C_COMPILER_ARCHITECTURE_ID MATCHES "[Xx]64")
+    set(BLAKE3_SIMD_TYPE "amd64-asm" CACHE STRING "the SIMD acceleration type to use")
+
+  elseif(CMAKE_C_COMPILER_ARCHITECTURE_ID MATCHES "[Aa][Rr][Mm]64")
+    set(BLAKE3_SIMD_TYPE "neon-intrinsics" CACHE STRING "the SIMD acceleration type to use")
 
-# FIXME CMAKE CMAKE_MSVC_RUNTIME_LIBRARY_DEFAULT
-if(BELA_ARCHITECTURE_64BIT)
-  if(BELA_ENABLE_ASSEMBLY_FILES)
-    if(MSVC)
-      enable_language(ASM_MASM)
-      list(
-        APPEND
-        BELA_BLAKE3_SOURCES
-        blake3/blake3_sse2_x86-64_windows_msvc.asm
-        blake3/blake3_sse41_x86-64_windows_msvc.asm
-        blake3/blake3_avx2_x86-64_windows_msvc.asm
-        blake3/blake3_avx512_x86-64_windows_msvc.asm)
-    elseif(WIN32 OR CYGWIN)
-      list(
-        APPEND
-        BELA_BLAKE3_SOURCES
-        blake3/blake3_sse2_x86-64_windows_gnu.S
-        blake3/blake3_sse41_x86-64_windows_gnu.S
-        blake3/blake3_avx2_x86-64_windows_gnu.S
-        blake3/blake3_avx512_x86-64_windows_gnu.S)
-    else()
-      list(
-        APPEND
-        BELA_BLAKE3_SOURCES
-        blake3/blake3_sse2_x86-64_unix.S
-        blake3/blake3_sse41_x86-64_unix.S
-        blake3/blake3_avx2_x86-64_unix.S
-        blake3/blake3_avx512_x86-64_unix.S)
-    endif()
   else()
-    list(
-      APPEND
-      BELA_BLAKE3_SOURCES
-      blake3/blake3_sse2.c
-      blake3/blake3_sse41.c
-      blake3/blake3_avx2.c
-      blake3/blake3_avx512.c)
-  endif(BELA_ENABLE_ASSEMBLY_FILES)
-elseif(BELA_ARCHITECTURE_ARM64)
-  list(APPEND BELA_BLAKE3_SOURCES blake3/blake3_neon.c)
-elseif(BELA_ARCHITECTURE_32BIT)
-  list(
-    APPEND
-    BELA_BLAKE3_SOURCES
-    blake3/blake3_sse2.c
-    blake3/blake3_sse41.c
-    blake3/blake3_avx2.c
-    blake3/blake3_avx512.c)
+    set(BLAKE3_SIMD_TYPE "none" CACHE STRING "the SIMD acceleration type to use")
+  endif()
+
+elseif(CMAKE_SYSTEM_PROCESSOR IN_LIST BLAKE3_AMD64_NAMES)
+  set(BLAKE3_SIMD_TYPE "amd64-asm" CACHE STRING "the SIMD acceleration type to use")
+
+elseif(CMAKE_SYSTEM_PROCESSOR IN_LIST BLAKE3_X86_NAMES
+       AND DEFINED BLAKE3_CFLAGS_SSE2
+       AND DEFINED BLAKE3_CFLAGS_SSE4.1
+       AND DEFINED BLAKE3_CFLAGS_AVX2
+       AND DEFINED BLAKE3_CFLAGS_AVX512)
+  set(BLAKE3_SIMD_TYPE "x86-intrinsics" CACHE STRING "the SIMD acceleration type to use")
+
+elseif((CMAKE_SYSTEM_PROCESSOR IN_LIST BLAKE3_ARMv8_NAMES
+          OR ANDROID_ABI STREQUAL "armeabi-v7a"
+          OR BLAKE3_USE_NEON_INTRINSICS)
+        AND (DEFINED BLAKE3_CFLAGS_NEON
+          OR CMAKE_SIZEOF_VOID_P EQUAL 8))
+  set(BLAKE3_SIMD_TYPE "neon-intrinsics" CACHE STRING "the SIMD acceleration type to use")
+
+else()
+  set(BLAKE3_SIMD_TYPE "none" CACHE STRING "the SIMD acceleration type to use")
 endif()
 
+mark_as_advanced(BLAKE3_SIMD_TYPE)
+
 add_library(
   belahash STATIC
   sha256.cc
   sha512.cc
   sha3.cc
   sm3.cc
-  ${BELA_BLAKE3_SOURCES})
+  blake3/blake3.c
+  blake3/blake3_dispatch.c
+  blake3/blake3_portable.c)
+
+# optional SIMD sources
+if(BLAKE3_SIMD_TYPE STREQUAL "amd64-asm")
+  if (NOT DEFINED BLAKE3_AMD64_ASM_SOURCES)
+    message(FATAL_ERROR "BLAKE3_SIMD_TYPE is set to 'amd64-asm' but no assembly sources are available for the target architecture.")
+  endif()
+  set(BLAKE3_SIMD_AMD64_ASM ON)
+
+  if(MSVC)
+    enable_language(ASM_MASM)
+  endif()
+
+  target_sources(belahash PRIVATE ${BLAKE3_AMD64_ASM_SOURCES})
+
+elseif(BLAKE3_SIMD_TYPE STREQUAL "x86-intrinsics")
+  if (NOT DEFINED BLAKE3_CFLAGS_SSE2
+      OR NOT DEFINED BLAKE3_CFLAGS_SSE4.1
+      OR NOT DEFINED BLAKE3_CFLAGS_AVX2
+      OR NOT DEFINED BLAKE3_CFLAGS_AVX512)
+    message(FATAL_ERROR "BLAKE3_SIMD_TYPE is set to 'x86-intrinsics' but no compiler flags are available for the target architecture.")
+  endif()
+  set(BLAKE3_SIMD_X86_INTRINSICS ON)
+
+  target_sources(belahash PRIVATE
+    blake3/blake3_avx2.c
+    blake3/blake3_avx512.c
+    blake3/blake3_sse2.c
+    blake3/blake3_sse41.c
+  )
+  set_source_files_properties(blake3/blake3_avx2.c PROPERTIES COMPILE_FLAGS "${BLAKE3_CFLAGS_AVX2}")
+  set_source_files_properties(blake3/blake3_avx512.c PROPERTIES COMPILE_FLAGS "${BLAKE3_CFLAGS_AVX512}")
+  set_source_files_properties(blake3/blake3_sse2.c PROPERTIES COMPILE_FLAGS "${BLAKE3_CFLAGS_SSE2}")
+  set_source_files_properties(blake3/blake3_sse41.c PROPERTIES COMPILE_FLAGS "${BLAKE3_CFLAGS_SSE4.1}")
+
+elseif(BLAKE3_SIMD_TYPE STREQUAL "neon-intrinsics")
+  set(BLAKE3_SIMD_NEON_INTRINSICS ON)
+
+  target_sources(belahash PRIVATE
+    blake3/blake3_neon.c
+  )
+  target_compile_definitions(belahash PRIVATE
+    BLAKE3_USE_NEON=1
+  )
+
+  if (DEFINED BLAKE3_CFLAGS_NEON)
+    set_source_files_properties(blake3/blake3_neon.c PROPERTIES COMPILE_FLAGS "${BLAKE3_CFLAGS_NEON}")
+  endif()
+
+elseif(BLAKE3_SIMD_TYPE STREQUAL "none")
+  target_compile_definitions(belahash PRIVATE
+    BLAKE3_USE_NEON=0
+    BLAKE3_NO_SSE2
+    BLAKE3_NO_SSE41
+    BLAKE3_NO_AVX2
+    BLAKE3_NO_AVX512
+  )
+
+else()
+  message(FATAL_ERROR "BLAKE3_SIMD_TYPE is set to an unknown value: '${BLAKE3_SIMD_TYPE}'")
+endif()
 
 target_link_libraries(belahash bela)
 

src/belahash/blake3.lock

@@ -1,2 +1,2 @@
 https://github.com/BLAKE3-team/BLAKE3
-fc2f7e4206f016b0cac0593f23a7d5976ce066e6
+e81557689b0c80abf312772a1a5f89d1881a9878

src/belahash/blake3/CMakeLists.txt

@@ -10,7 +10,7 @@ if (POLICY CMP0102)
 endif()
 
 project(libblake3
-  VERSION 1.5.2
+  VERSION 1.5.4
   DESCRIPTION "BLAKE3 C implementation"
   LANGUAGES C ASM
 )

src/belahash/blake3/CMakePresets.json

@@ -0,0 +1,73 @@
+{
+    "version": 3,
+    "cmakeMinimumRequired": {
+        "major": 3,
+        "minor": 22,
+        "patch": 0
+    },
+    "configurePresets": [
+        {
+            "name": "base",
+            "hidden": true,
+            "binaryDir": "${sourceDir}/build/${presetName}"
+        },
+        {
+            "name": "msvc",
+            "hidden": true,
+            "generator": "Visual Studio 17 2022",
+            "vendor": {
+                "microsoft.com/VisualStudioSettings/CMake/1.0": {
+                    "hostOS": [
+                        "Windows"
+                    ]
+                }
+            }
+        },
+        {
+            "name": "x64-windows-msvc",
+            "inherits": [
+                "msvc",
+                "base"
+            ],
+            "architecture": "x64"
+        },
+        {
+            "name": "x86-windows-msvc",
+            "inherits": [
+                "msvc",
+                "base"
+            ],
+            "architecture": "Win32"
+        },
+        {
+            "name": "arm64-windows-msvc",
+            "inherits": [
+                "msvc",
+                "base"
+            ],
+            "architecture": "ARM64"
+        }
+    ],
+    "buildPresets": [
+        {
+            "name": "x64-windows-msvc-debug",
+            "configurePreset": "x64-windows-msvc",
+            "configuration": "Debug"
+        },
+        {
+            "name": "x64-windows-msvc-release",
+            "configurePreset": "x64-windows-msvc",
+            "configuration": "RelWithDebInfo"
+        },
+        {
+            "name": "x86-windows-msvc-debug",
+            "configurePreset": "x86-windows-msvc",
+            "configuration": "Debug"
+        },
+        {
+            "name": "x86-windows-msvc-release",
+            "configurePreset": "x86-windows-msvc",
+            "configuration": "RelWithDebInfo"
+        }
+    ]
+}

src/belahash/blake3/blake3.c

@@ -88,24 +88,30 @@ INLINE void output_chaining_value(const output_t *self, uint8_t cv[32]) {
 
 INLINE void output_root_bytes(const output_t *self, uint64_t seek, uint8_t *out,
                               size_t out_len) {
+  if (out_len == 0) {
+      return;
+  }
   uint64_t output_block_counter = seek / 64;
   size_t offset_within_block = seek % 64;
   uint8_t wide_buf[64];
-  while (out_len > 0) {
-    blake3_compress_xof(self->input_cv, self->block, self->block_len,
-                        output_block_counter, self->flags | ROOT, wide_buf);
-    size_t available_bytes = 64 - offset_within_block;
-    size_t memcpy_len;
-    if (out_len > available_bytes) {
-      memcpy_len = available_bytes;
-    } else {
-      memcpy_len = out_len;
-    }
-    memcpy(out, wide_buf + offset_within_block, memcpy_len);
-    out += memcpy_len;
-    out_len -= memcpy_len;
+  if(offset_within_block) {
+    blake3_compress_xof(self->input_cv, self->block, self->block_len, output_block_counter, self->flags | ROOT, wide_buf);
+    const size_t available_bytes = 64 - offset_within_block;
+    const size_t bytes = out_len > available_bytes ? available_bytes : out_len;
+    memcpy(out, wide_buf + offset_within_block, bytes);
+    out += bytes;
+    out_len -= bytes;
     output_block_counter += 1;
-    offset_within_block = 0;
+  }
+  if(out_len / 64) {
+    blake3_xof_many(self->input_cv, self->block, self->block_len, output_block_counter, self->flags | ROOT, out, out_len / 64);
+  }
+  output_block_counter += out_len / 64;
+  out += out_len & -64;
+  out_len -= out_len & -64;
+  if(out_len) {
+    blake3_compress_xof(self->input_cv, self->block, self->block_len, output_block_counter, self->flags | ROOT, wide_buf);
+    memcpy(out, wide_buf, out_len);
   }
 }
 

src/belahash/blake3/blake3.h

@@ -30,7 +30,7 @@
 extern "C" {
 #endif
 
-#define BLAKE3_VERSION_STRING "1.5.2"
+#define BLAKE3_VERSION_STRING "1.5.4"
 #define BLAKE3_KEY_LEN 32
 #define BLAKE3_OUT_LEN 32
 #define BLAKE3_BLOCK_LEN 64