Merge pull request fast-pack#108 from seb711/simde

 Add ARM Support for FastPFOR using SIMDe

.github/workflows/vs17-arm-ci.yml

@@ -0,0 +1,31 @@
+name: VS17-ARM-CI
+
+on: [push, pull_request]
+
+jobs:
+  ci:
+    name: windows-vs17
+    runs-on: windows-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - {gen: Visual Studio 17 2022, arch: ARM64}
+    steps:
+      - name: checkout
+        uses: actions/checkout@v2
+      - name: Configure
+        run: |
+          cmake -B build
+      - name: Build Debug
+        run: cmake --build build --config Debug --verbose
+      - name: Build Release
+        run: cmake --build build --config Release --verbose
+      - name: Run Release tests
+        run: |
+          cd build
+          ctest -C Release  -LE explicitonly  --output-on-failure
+      - name: Run Debug tests
+        run: |
+          cd build
+          ctest -C Debug  -LE explicitonly  --output-on-failure

CMakeLists.txt

@@ -23,7 +23,11 @@ set(PROJECT_DESCRIPTION "The FastPFOR C++ library: Fast integer compression")
 set(PROJECT_VERSION "0.1.9")
 
 include(DetectCPUFeatures)
-#
+include("${CMAKE_MODULE_PATH}/environment.cmake")
+
+
+message("Building for architecture: ${CMAKE_SYSTEM_PROCESSOR}")
+
 # Runs compiler with "-dumpversion" and parses major/minor
 # version with a regex.
 #
@@ -57,7 +61,12 @@ MESSAGE( STATUS "CXX_COMPILER_VERSION: " ${CXX_COMPILER_VERSION} )
 if( SUPPORT_SSE42 )
     MESSAGE( STATUS "SSE 4.2 support detected" )
 else()
-    MESSAGE( STATUS "SSE 4.2 support not detected" )
+    if (SUPPORT_NEON)
+        include("${CMAKE_MODULE_PATH}/simde.cmake")
+        MESSAGE(STATUS "USING SIMDE FOR SIMD OPERATIONS")
+    else ()
+        MESSAGE(STATUS "SIMDE and SSE 4.2 support not detected")
+    endif ()
 endif()
 
 if(${CMAKE_CXX_COMPILER_ID} STREQUAL "GNU")
@@ -85,10 +94,10 @@ elseif(${CMAKE_CXX_COMPILER_ID} STREQUAL "Clang" OR ${CMAKE_CXX_COMPILER_ID} STR
     if (CXX_COMPILER_VERSION VERSION_LESS 4.2.1)
         message(STATUS  "Clang version must be at least 4.2.1!" )
     endif()
-    set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wcast-align -O3 -DNDEBUG -std=c++11 -DHAVE_CXX0X -msse4.1 -march=native")
-    set (CMAKE_CXX_FLAGS_DEBUG   "-Wall -Wcast-align -ggdb  -std=c++11 -DHAVE_CXX0X -msse4.1 -march=native")
-    set (CMAKE_C_FLAGS_RELEASE "-Wall -Wcast-align -O3 -DNDEBUG -std=c99 -msse4.1 -march=native")
-    set (CMAKE_C_FLAGS_DEBUG   "-Wall -Wcast-align -ggdb  -std=c99 -msse4.1 -march=native")
+    set (CMAKE_CXX_FLAGS_RELEASE "-Wall -Wcast-align -O3 -DNDEBUG -std=c++11 -DHAVE_CXX0X -march=native")
+    set (CMAKE_CXX_FLAGS_DEBUG   "-Wall -Wcast-align -ggdb  -std=c++11 -DHAVE_CXX0X -march=native")
+    set (CMAKE_C_FLAGS_RELEASE "-Wall -Wcast-align -O3 -DNDEBUG -std=c99 -march=native")
+    set (CMAKE_C_FLAGS_DEBUG   "-Wall -Wcast-align -ggdb  -std=c99 -march=native")
 elseif(WIN32)
     # TODO add support for later versions?
     if(NOT MSVC12)
@@ -126,6 +135,13 @@ add_executable(gapstats src/gapstats.cpp)
 add_executable(partitionbylength src/partitionbylength.cpp)
 add_executable(csv2maropu src/csv2maropu.cpp)
 
+if (SUPPORT_NEON)
+    target_link_libraries(FastPFOR PUBLIC simde)
+    target_link_libraries(gapstats PUBLIC simde)
+    target_link_libraries(partitionbylength PUBLIC simde)
+    target_link_libraries(csv2maropu PUBLIC simde)
+endif()
+
 add_executable(entropy src/entropy.cpp)
 target_link_libraries(entropy FastPFOR)
 

cmake_modules/environment.cmake

@@ -0,0 +1,10 @@
+include(CheckCXXCompilerFlag)
+
+if (CMAKE_SYSTEM_PROCESSOR MATCHES "arm64" OR CMAKE_SYSTEM_PROCESSOR MATCHES "arm" OR CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64")
+    set(SUPPORT_NEON ON)
+endif ()
+
+# Check if the Visual Studio build is targeting ARM
+if (CMAKE_GENERATOR_PLATFORM MATCHES "ARM64" OR CMAKE_GENERATOR_PLATFORM MATCHES "ARM")
+    set(SUPPORT_NEON ON)
+endif ()

cmake_modules/simde.cmake

@@ -0,0 +1,13 @@
+include(FetchContent)
+FetchContent_Declare(
+        simde
+        GIT_REPOSITORY https://github.com/simd-everywhere/simde.git
+        GIT_TAG master
+)
+FetchContent_MakeAvailable(simde)
+
+add_library(simde INTERFACE IMPORTED GLOBAL)
+target_include_directories(simde INTERFACE "${simde_SOURCE_DIR}")
+
+# Enables native aliases. Not ideal but makes it easier to convert old code.
+target_compile_definitions(simde INTERFACE SIMDE_ENABLE_NATIVE_ALIASES)

headers/VarIntG8IU.h

@@ -2,7 +2,7 @@
  * This code is released under the
  * Apache License Version 2.0 http://www.apache.org/licenses/.
  */
-#if !defined(__SSSE3__) && !(defined(_MSC_VER) && defined(__AVX__))
+#if (!defined(__SSSE3__) && !(defined(_MSC_VER) && defined(__AVX__))) && !(defined(__ARM_NEON) || defined(__aarch64__))
 #ifndef _MSC_VER
 #pragma message                                                                \
     "Disabling varintg8iu due to lack of SSSE3 support, try adding -mssse3 or the equivalent on your compiler"
@@ -12,7 +12,12 @@
 #else
 #ifndef VARINTG8IU_H__
 #define VARINTG8IU_H__
+#if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
 #include <emmintrin.h>
+#elif defined(__aarch64__)
+/* GCC-compatible compiler, targeting ARM with NEON */
+#include <simde/x86/sse3.h>
+#endif
 #include "codecs.h"
 #ifdef __GNUC__
 #define PREDICT_FALSE(x) (__builtin_expect(x, 0))

headers/common.h

@@ -10,7 +10,12 @@
 // C headers (sorted)
 #include <errno.h>
 #include <fcntl.h>
+#if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
 #include <immintrin.h>
+#elif defined(__GNUC__) && defined(__aarch64__)
+#include <simde/x86/sse4.1.h>
+#endif
+
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -44,7 +49,12 @@
 #ifdef _MSC_VER
 #include <iso646.h>
 #include <stdint.h>
+
+#if (defined(_M_X64) || defined(_M_AMD64))
 #include <intrin.h>
+#elif defined(_M_ARM64)
+#include <simde/x86/sse4.1.h>
+#endif
 
 #define __attribute__(n)
 #define __restrict__ __restrict

headers/cpubenchmark.h

@@ -42,11 +42,34 @@ static __inline__ unsigned long long stopRDTSCP(void) {
                  "%rdx");
   return (static_cast<unsigned long long>(cycles_high) << 32) | cycles_low;
 }
-#elif defined(_MSC_VER)
+#elif  (defined(_MSC_VER) && (defined(_M_X64) || defined(_M_AMD64)))
 
 static inline unsigned long long startRDTSC(void) { return __rdtsc(); }
 
 static inline unsigned long long stopRDTSCP(void) { return __rdtsc(); }
+#elif defined(_MSC_VER) && defined(_M_ARM64)
+  // oriented by zeromq implementation for msc arm/arm64
+  // https://github.com/zeromq/libzmq/blob/master/src/clock.cpp
+  inline unsigned long long rdtsc() {
+    const int64_t pmccntr_el0 = (((3 & 1) << 14) |  // op0
+                                 ((3 & 7) << 11) |  // op1
+                                 ((9 & 15) << 7) |  // crn
+                                 ((13 & 15) << 3) | // crm
+                                 ((0 & 7) << 0));   // op2
+
+    return _ReadStatusReg (pmccntr_el0);
+  }
+
+  static inline unsigned long long startRDTSC(void) { return rdtsc(); }
+
+  static inline unsigned long long stopRDTSCP(void) { return rdtsc(); }
+#elif  (defined(_MSC_VER) && (defined(_M_ARM64)))
+// Taken from microsoft documentation (see
+// https://learn.microsoft.com/en-us/cpp/build/overview-of-arm-abi-conventions?view=msvc-170
+
+static inline unsigned long long startRDTSC(void) { return __rdpmccntr64(); }
+
+static inline unsigned long long stopRDTSCP(void) { return __rdpmccntr64(); }
 
 #elif defined(__i386__) || defined(__x86_64__)
 
@@ -66,15 +89,25 @@ inline unsigned long long rdtsc() {
 static __inline__ unsigned long long startRDTSC(void) { return rdtsc(); }
 
 static __inline__ unsigned long long stopRDTSCP(void) { return rdtsc(); }
+#elif (defined(__GNUC__) && (defined(__arch64__)))
+    inline uint64_t rdtsc() {
+        uint64_t cycles;
+        asm volatile("mrs %0, cntvct_el0"
+                : "=r"(cycles)); /* output */
+        return cycles;
+    }
+
+    static __inline__ uint64_t startRDTSC(void) { return rdtsc(); }
 
-#elif(defined(__arm__) || defined(__ppc__) || defined(__ppc64__))
+    static __inline__ uint64_t stopRDTSCP(void) { return rdtsc(); }
+#elif(defined(__arm__) || defined(__ppc__) || defined(__ppc64__)) || (defined(_MSC_VER) && defined(_M_ARM64))
 
 // for PPC we should be able to use tbl, but I could not find
 // an equivalent to rdtsc for ARM.
 
-inline uint64 rdtsc() { return 0; }
-static __inline__ ticks startRDTSC(void) { return 0; }
-static __inline__ ticks stopRDTSCP(void) { return 0; }
+inline uint64_t rdtsc() { return 0; }
+static __inline__ uint64_t startRDTSC(void) { return 0; }
+static __inline__ uint64_t stopRDTSCP(void) { return 0; }
 #else
 #error Unknown architecture
 #endif

headers/horizontalbitpacking.h

@@ -21,8 +21,8 @@
 #ifndef HORIZONTALBITPACKING_H_
 #define HORIZONTALBITPACKING_H_
 
-#if !defined(__SSE4_1__) && !(defined(_MSC_VER) && defined(__AVX__))
 
+#if (!defined(__SSE4_1__) && !(defined(_MSC_VER) && defined(__AVX__))) && (!(defined(__ARM_NEON) || defined(__aarch64__)))
 #ifndef _MSC_VER
 #pragma message "No SSSE4.1 support? try adding -msse4.1 or the equivalent on your compiler"
 #else