Merge branch 'avx2'

.travis.yml

@@ -11,11 +11,12 @@ env:
     - OMP_NUM_THREADS=4
 
 before_install:
+  - sudo add-apt-repository ppa:george-edison55/precise-backports -y
   - sudo apt-get update -qq
 
 # Install cmake, OpenMPI, cppcheck
 install:
-  - sudo apt-get install -y cmake openmpi-bin libopenmpi-dev cppcheck
+  - sudo apt-get install -y cmake cmake-data openmpi-bin libopenmpi-dev cppcheck
 
 # Test build.sh script
 before_script:

CMakeLists.txt

@@ -1,20 +1,21 @@
-cmake_minimum_required(VERSION 2.8.4)
+cmake_minimum_required(VERSION 3.0)
 project(primecount)
 set(PRIMECOUNT_VERSION_MAJOR 3)
-set(PRIMECOUNT_VERSION_MINOR 5)
+set(PRIMECOUNT_VERSION_MINOR 6)
 set(PRIMECOUNT_VERSION "${PRIMECOUNT_VERSION_MAJOR}.${PRIMECOUNT_VERSION_MINOR}")
 include_directories(include src/primesieve/include)
 set(CMAKE_BUILD_TYPE Release)
 
 # Build options ######################################################
 
 option(ENABLE_POPCNT "Enable POPCNT instruction" ON)
-option(ENABLE_MPI "Enable MPI (Message Passing Interface)" OFF)
+option(ENABLE_MPI    "Enable MPI (Message Passing Interface)" OFF)
 
 # Include Check* #####################################################
 
-include(CheckCXXCompilerFlag)
+include(CheckCXXSourceCompiles)
 include(CheckCXXSourceRuns)
+include(CheckCXXCompilerFlag)
 include(CheckTypeSize)
 
 # primecount binary source files #####################################
@@ -26,6 +27,7 @@ set(BIN_SRC src/app/cmdoptions.cpp
 # primecount library source files ####################################
 
 set(LIB_SRC src/BitSieve.cpp
+            src/BitSieve-popcnt.cpp
             src/FactorTable.cpp
             src/Li.cpp
             src/P2.cpp
@@ -103,19 +105,48 @@ endif()
 # Check for OpenMP ###################################################
 
 find_package(OpenMP QUIET)
-if (OPENMP_FOUND)
+if(OPENMP_FOUND)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
 endif()
 
-# Check for -mpopcnt compiler option #################################
+# Check -mpopcnt compiler flag #######################################
 
 if(ENABLE_POPCNT)
-    CHECK_CXX_COMPILER_FLAG(-mpopcnt MPOPCNT_FLAG)
-    if(MPOPCNT_FLAG)
+    set(COPY_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
+    set(CMAKE_CXX_FLAGS -Werror)
+    check_cxx_compiler_flag(-mpopcnt mpopcnt)
+    set(CMAKE_CXX_FLAGS "${COPY_CXX_FLAGS}")
+    if(mpopcnt)
         set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mpopcnt")
     endif()
 endif()
 
+# Check -mpopcntd compiler flag ######################################
+
+if(ENABLE_POPCNT)
+    set(COPY_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
+    set(CMAKE_CXX_FLAGS -Werror)
+    check_cxx_compiler_flag(-mpopcntd mpopcntd)
+    set(CMAKE_CXX_FLAGS "${COPY_CXX_FLAGS}")
+    if(mpopcntd)
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mpopcntd")
+    endif()
+endif()
+
+# Check AVX2 (x86_64) compiler support  ##############################
+
+check_cxx_source_compiles("
+    #include <immintrin.h>
+    __attribute__ ((target (\"avx2\")))
+    void func() { __m256i v = _mm256_setzero_si256(); }
+    int x = __builtin_cpu_supports(\"avx2\");
+    int main() { return 0; }"
+    AVX2)
+
+if(AVX2)
+    add_definitions(-DHAVE_AVX2)
+endif()
+
 # Check if int128_t type exists ######################################
 
 check_type_size("int128_t" INT128_T)
@@ -164,11 +195,11 @@ check_cxx_source_runs("
 
         return 0;
     }"
-    HAVE_LIBDIVIDE)
+    use_libdivide)
 
 # Include S2_easy* source files ######################################
 
-if(HAVE_LIBDIVIDE)
+if(use_libdivide)
     set(LIB_SRC ${LIB_SRC} src/deleglise-rivat/S2_easy_libdivide.cpp)
     if(ENABLE_MPI)
         set(LIB_SRC ${LIB_SRC} src/mpi/deleglise-rivat/S2_easy_mpi_libdivide.cpp)

COPYING

@@ -1,6 +1,6 @@
 BSD 2-Clause License
 
-Copyright (c) 2013 - 2016, Kim Walisch.
+Copyright (c) 2013 - 2017, Kim Walisch.
 All rights reserved.
 
 Redistribution and use in source and binary forms, with or without

ChangeLog

@@ -1,3 +1,16 @@
+2017-03-04 Kim Walisch  <[email protected]>
+
+  Version 3.6 released.
+
+  This version features a new AVX2 popcount algorithm which
+  computes the hard special leaves up to 15% faster on x86 CPUs
+  with AVX2 support (2013 or later).
+
+  * BitSieve-popcnt.cpp: New AVX2 popcount algorithm.
+  * popcnt.hpp: Fix clang performance bug.
+  * primecount.cpp: Fix clang time measuring.
+  * CMakeLists.txt: Add AVX2 check.
+
 2016-12-16 Kim Walisch  <[email protected]>
 
   Version 3.5 released.

README.md

@@ -26,13 +26,14 @@ used to compute several world records e.g.
 
 Binaries
 --------
-Below are the latest precompiled binaries for Windows 64-bit, Linux and OS X.
+Below are the latest primecount binaries for Windows 64-bit, Linux and macOS.
 These binaries are statically linked and require a CPU which supports the POPCNT
-instruction (2010 or later).
+instruction (2010 or later). primecount also uses the AVX2 instruction set (if
+available) to speed up the computation of the hard special leaves. 
 
-* [primecount-3.5-win64.zip](https://dl.bintray.com/kimwalisch/primecount/primecount-3.5-win64.zip), 404K
-* [primecount-3.5-linux-x64.tar.gz](https://dl.bintray.com/kimwalisch/primecount/primecount-3.5-linux-x64.tar.gz), 915K
-* [primecount-3.5-macOS-x64.tar.gz](https://dl.bintray.com/kimwalisch/primecount/primecount-3.5-macOS-x64.tar.gz), 688K
+* [primecount-3.6-win64.zip](https://dl.bintray.com/kimwalisch/primecount/primecount-3.6-win64.zip), 400 KB
+* [primecount-3.6-linux-x64.tar.gz](https://dl.bintray.com/kimwalisch/primecount/primecount-3.6-linux-x64.tar.gz), 1 MB
+* [primecount-3.6-macOS-x64.zip](https://dl.bintray.com/kimwalisch/primecount/primecount-3.6-macOS-x64.zip), 899 KB
 * Binaries with backup functionality are available [here](https://github.com/kimwalisch/primecount/tree/backup#primecount-backup)
 
 Usage examples
@@ -207,64 +208,64 @@ Benchmarks
     <td>3,204,941,750,802</td>
     <td>4.01s</td>
     <td>1.96s</td>
-    <td>0.43s</td>
+    <td>0.38s</td>
     <td>0.22s</td>
   </tr>
   <tr align="right">
     <td>10<sup>15</sup></td>
     <td>29,844,570,422,669</td>
     <td>27.75s</td>
     <td>12.08s</td>
-    <td>1.65s</td>
-    <td>0.77s</td>
+    <td>1.52s</td>
+    <td>0.76s</td>
   </tr>
   <tr align="right">
     <td>10<sup>16</sup></td>
     <td>279,238,341,033,925</td>
     <td>232.30s</td>
     <td>92.09s</td>
-    <td>7.31s</td>
-    <td>2.83s</td>
+    <td>6.87s</td>
+    <td>2.67s</td>
   </tr>
   <tr align="right">
     <td>10<sup>17</sup></td>
     <td>2,623,557,157,654,233</td>
     <td>1,836.73s</td>
     <td>731.35s</td>
-    <td>33.66s</td>
-    <td>11.20s</td>
+    <td>31.63s</td>
+    <td>10.66s</td>
   </tr>
   <tr align="right">
     <td>10<sup>18</sup></td>
     <td>24,739,954,287,740,860</td>
     <td>14,949.16s</td>
     <td>6,631.73s</td>
-    <td>157.15s</td>
-    <td>47.07s</td>
+    <td>146.55s</td>
+    <td>44.54s</td>
   </tr>
   <tr align="right">
     <td>10<sup>19</sup></td>
     <td>234,057,667,276,344,607</td>
     <td>NaN</td>
     <td>NaN</td>
     <td>NaN</td>
-    <td>225.08s</td>
+    <td>209.57s</td>
   </tr>
   <tr align="right">
     <td>10<sup>20</sup></td>
     <td>2,220,819,602,560,918,840</td>
     <td>NaN</td>
     <td>NaN</td>
     <td>NaN</td>
-    <td>1,029.41s</td>
+    <td>939.88s</td>
   </tr>
   <tr align="right">
     <td>10<sup>21</sup></td>
     <td>21,127,269,486,018,731,928</td>
     <td>NaN</td>
     <td>NaN</td>
     <td>NaN</td>
-    <td>4,867.45s</td>
+    <td>4,536.14s</td>
   </tr>
   <tr align="right">
     <td>10<sup>22</sup></td>
@@ -278,14 +279,14 @@ Benchmarks
 
 The benchmarks above were run on an Intel Core i7-6700 CPU (4 x 3.4 GHz) from
 2015 using a Linux x64 operating system and primecount was compiled using
-GCC 5.2.
+GCC 5.4.
 
 Build instructions
 ------------------
 You need to have installed a C++ compiler, cmake and make to build primecount.
 
 Download
-[primecount-3.5.zip](https://github.com/kimwalisch/primecount/archive/v3.5.zip)
+[primecount-3.6.zip](https://github.com/kimwalisch/primecount/archive/v3.6.zip)
 and build it using:
 
 ```sh

appveyor.yml

@@ -3,10 +3,6 @@
 
 version: 1.0.{build}
 
-branches:
-  only:
-  - master
-
 platform:
   - x86
   - x64

doc/primecount-MPI.md

@@ -25,7 +25,7 @@ sudo apt-get install g++ make cmake libopenmpi-dev openmpi-bin
 ```
 
 Then download
-[primecount-3.5.zip](https://github.com/kimwalisch/primecount/archive/v3.5.zip)
+[primecount-3.6.zip](https://github.com/kimwalisch/primecount/archive/v3.6.zip)
 and build it using:
 ```sh
 cmake -DENABLE_MPI=ON .

include/PhiTiny.hpp

@@ -1,7 +1,7 @@
 ///
 /// @file  PhiTiny.hpp
 ///
-/// Copyright (C) 2016 Kim Walisch, <[email protected]>
+/// Copyright (C) 2017 Kim Walisch, <[email protected]>
 ///
 /// This file is distributed under the BSD License. See the COPYING
 /// file in the top level directory.
@@ -56,6 +56,8 @@ class PhiTiny {
   static const int totients[7];
 };
 
+extern const PhiTiny phiTiny;
+
 inline bool is_phi_tiny(int64_t a)
 {
   return PhiTiny::is_tiny(a);
@@ -66,7 +68,6 @@ inline bool is_phi_tiny(int64_t a)
 template <typename X, typename A>
 typename prt::make_signed<X>::type phi_tiny(X x, A a)
 {
-  extern const PhiTiny phiTiny;
   return phiTiny.phi(x, a);
 }
 
@@ -75,7 +76,6 @@ typename prt::make_signed<X>::type phi_tiny(X x, A a)
 template <typename X, typename A>
 X phi_tiny(X x, A a)
 {
-  extern const PhiTiny phiTiny;
   return phiTiny.phi(x, a);
 }