#

.gitignore

@@ -1,2 +1 @@
 pg.txt
-ref.txt

CMakeLists.txt

@@ -10,8 +10,7 @@ set(CMAKE_CXX_EXTENSIONS ON)
 foreach(target IN ITEMS "basic" "pb_ds" "lt" "oaph")
     add_executable("${target}")
     target_sources("${target}" PRIVATE "${target}.cpp" "io.hpp" "timer.hpp" "helpers.hpp")
-    target_compile_options("${target}" PRIVATE $<$<CXX_COMPILER_ID:Clang>:-march=native>)
-    target_compile_options("${target}" PRIVATE $<$<CXX_COMPILER_ID:GNU>:-march=native>)
+    target_compile_options("${target}" PRIVATE $<$<OR:$<CXX_COMPILER_ID:Clang>,$<CXX_COMPILER_ID:GNU>>:-march=native>)
     if(target STREQUAL "pb_ds")
         target_compile_options("${target}" PRIVATE $<$<CXX_COMPILER_ID:Clang>:-stdlib=libstdc++>)
         target_link_options("${target}" PRIVATE $<$<CXX_COMPILER_ID:Clang>:-stdlib=libstdc++>)

basic.cpp

@@ -1,5 +1,4 @@
 #include "helpers.hpp"
-#include "io.hpp"
 #include "timer.hpp"
 
 #include <algorithm>
@@ -50,13 +49,13 @@ int main(int argc, char * argv[])
     std::unordered_map<std::string_view, std::size_t> wordCounts;
 
     auto isAlpha = [](char c) {
-        return std::isalpha(std::make_unsigned_t<char>(c));
+        return bool(std::isalpha(std::make_unsigned_t<char>(c)));
     };
-    auto end = std::end(input);
-    auto beg = std::find_if(std::begin(input), end, isAlpha);
+    auto end = input.data() + input.size();
+    auto beg = std::find_if(input.data(), end, isAlpha);
     while (beg != end) {
         auto it = std::find_if(beg, end, std::not_fn(isAlpha));
-        ++wordCounts[std::string_view(&*beg, std::distance(beg, it))];
+        ++wordCounts[{beg, std::size_t(std::distance(beg, it))}];
         beg = std::find_if(it, end, isAlpha);
     }
 
@@ -83,7 +82,6 @@ int main(int argc, char * argv[])
     for (auto wordCount : output) {
         o << wordCount->second << ' ' << wordCount->first << '\n';
     }
-
     timer.report("write output");
 
     return EXIT_SUCCESS;

helpers.hpp

@@ -23,6 +23,18 @@
 #error "!"
 #endif
 
+#if defined(__linux__) || defined(__APPLE__)
+#define RED(s) "\e[3;31m" s "\e[0m"
+#define GREEN(s) "\e[3;32m" s "\e[0m"
+#define YELLOW(s) "\e[3;33m" s "\e[0m"
+#define BLUE(s) "\e[3;34m" s "\e[0m"
+#else
+#define RED(s) s
+#define GREEN(s) s
+#define YELLOW(s) s
+#define BLUE(s) s
+#endif
+
 #ifdef __cpp_lib_hardware_interference_size
 inline constexpr std::size_t kHardwareConstructiveInterferenceSize =
     std::hardware_constructive_interference_size;
@@ -33,33 +45,20 @@ inline constexpr std::size_t kHardwareConstructiveInterferenceSize = 64;
 inline constexpr std::size_t kHardwareDestructiveInterferenceSize = 64;
 #endif
 
-inline void toLower(char * s, char * const end)
+inline void toLower(char * beg, char * const end)
 {
     assert(s <= end);
-    for (; s < end; s += sizeof(__m128i)) {
-        __m128i string = _mm_load_si128(reinterpret_cast<const __m128i *>(s));
+    assert((reinterpret_cast<std::uintptr_t>(beg) % sizeof(__m128i)) == 0);
+    assert((reinterpret_cast<std::uintptr_t>(end) % sizeof(__m128i)) == 0);
+    for (; beg < end; beg += sizeof(__m128i)) {
+        __m128i string = _mm_load_si128(reinterpret_cast<const __m128i *>(beg));
         __m128i lowercase = _mm_add_epi8(
             string, _mm_and_si128(_mm_cmplt_epi8(string, _mm_set1_epi8('a')),
                                   _mm_set1_epi8('a' - 'A')));
         __m128i mask =
             _mm_or_si128(_mm_cmplt_epi8(lowercase, _mm_set1_epi8('a')),
                          _mm_cmpgt_epi8(lowercase, _mm_set1_epi8('z')));
-        _mm_store_si128(reinterpret_cast<__m128i *>(s),
+        _mm_store_si128(reinterpret_cast<__m128i *>(beg),
                         _mm_andnot_si128(mask, lowercase));
     }
-    if (s < end) {
-        __m128i string{};
-        auto c = reinterpret_cast<char *>(&string);
-        std::copy(s, end, c);
-        __m128i lowercase = _mm_add_epi8(
-            string, _mm_and_si128(_mm_cmplt_epi8(string, _mm_set1_epi8('a')),
-                                  _mm_set1_epi8('a' - 'A')));
-        __m128i mask =
-            _mm_or_si128(_mm_cmplt_epi8(lowercase, _mm_set1_epi8('a')),
-                         _mm_cmpgt_epi8(lowercase, _mm_set1_epi8('z')));
-        string = _mm_andnot_si128(mask, lowercase);
-        for (; s < end; ++s) {
-            *s++ = *c++;
-        }
-    }
 }

io.hpp

@@ -8,17 +8,26 @@
 #include <cstdio>
 #include <cstdlib>
 
-#if defined(__linux__) || defined(__APPLE__)
-#define RED(s) "\e[3;31m" s "\e[0m"
-#define GREEN(s) "\e[3;32m" s "\e[0m"
-#define YELLOW(s) "\e[3;33m" s "\e[0m"
-#define BLUE(s) "\e[3;34m" s "\e[0m"
-#else
-#define RED(s) s
-#define GREEN(s) s
-#define YELLOW(s) s
-#define BLUE(s) s
-#endif
+inline std::size_t readInput(char * inputBegin, std::size_t inputSize,
+                             std::FILE * inputFile)
+{
+    std::size_t readSize = std::fread(inputBegin, 1, inputSize, inputFile);
+    std::fprintf(stderr, "input size = %zu bytes\n", readSize);
+    if (!(readSize < inputSize)) {
+        std::fprintf(stderr, "input is too large\n");
+        return 0;
+    }
+
+    inputBegin += readSize;
+    while ((readSize % sizeof(__m128i)) != 0) {
+        *inputBegin++ = '\0';
+        if (++readSize > inputSize) {
+            std::fprintf(stderr, "input is too large\n");
+            return 0;
+        }
+    }
+    return readSize;
+}
 
 template<std::size_t bufferSize = 131072>
 class OutputStream

lt.cpp

@@ -21,7 +21,7 @@ namespace
 alignas(__m128i) char input[1 << 29];
 auto inputEnd = input;
 
-struct alignas(kHardwareDestructiveInterferenceSize) TrieNode
+struct TrieNode
 {
     uint32_t count = 0;
     uint32_t children['z' - 'a' + 1] = {};
@@ -56,20 +56,12 @@ int main(int argc, char * argv[])
 
     timer.report("open files");
 
-    {
-        std::size_t size =
-            std::fread(input, sizeof *input, std::extent_v<decltype(input)>,
-                       inputFile.get());
-        if (!(size < std::extent_v<decltype(input)>)) {
-            std::fprintf(stderr, "input is too large\n");
-            return EXIT_FAILURE;
-        }
-        std::fprintf(stderr, "input size = %zu bytes\n", size);
-
-        inputEnd +=
-            ((size + sizeof(__m128i) - 1) / sizeof(__m128i)) * sizeof(__m128i);
-        std::fill(input + size, inputEnd, '\0');
+    std::size_t readSize =
+        readInput(std::begin(input), std::size(input), inputFile.get());
+    if (readSize == 0) {
+        return EXIT_FAILURE;
     }
+    inputEnd += readSize;
     timer.report("read input");
 
     toLower(input, inputEnd);
@@ -132,7 +124,6 @@ int main(int argc, char * argv[])
     timer.report(YELLOW("sort words"));
 
     OutputStream<> outputStream{outputFile.get()};
-
     for (const auto & [count, word] : rank) {
         if (!outputStream.print(count)) {
             std::fprintf(stderr, "output failure");
@@ -151,8 +142,7 @@ int main(int argc, char * argv[])
             return EXIT_FAILURE;
         }
     }
-
-    timer.report("output");
+    timer.report("write output");
 
     return EXIT_SUCCESS;
 }

oaph.cpp

@@ -275,17 +275,12 @@ int main(int argc, char * argv[])
         return EXIT_FAILURE;
     }
 
-    {
-        std::size_t size =
-            std::fread(input, sizeof *input, std::extent_v<decltype(input)>,
-                       inputFile.get());
-        assert(size < std::extent_v<decltype(input)>);
-        std::fprintf(stderr, "input size = %zu bytes\n", size);
-
-        inputEnd +=
-            ((size + sizeof(__m128i) - 1) / sizeof(__m128i)) * sizeof(__m128i);
-        std::fill(input + size, inputEnd, '\0');
+    std::size_t readSize =
+        readInput(std::begin(input), std::size(input), inputFile.get());
+    if (readSize == 0) {
+        return EXIT_FAILURE;
     }
+    inputEnd += readSize;
     timer.report("read input");
 
     for (Chunk & chunk : hashTable) {
@@ -358,7 +353,7 @@ int main(int argc, char * argv[])
             return EXIT_FAILURE;
         }
     }
-    timer.report("output");
+    timer.report("write output");
 
     return EXIT_SUCCESS;
 }

pb_ds.cpp

@@ -1,5 +1,4 @@
 #include "helpers.hpp"
-#include "io.hpp"
 #include "timer.hpp"
 
 #include <algorithm>
@@ -57,7 +56,7 @@ int main(int argc, char * argv[])
     Trie trie;
 
     auto isAlpha = [](char c) {
-        return std::isalpha(std::make_unsigned_t<char>(c));
+        return bool(std::isalpha(std::make_unsigned_t<char>(c)));
     };
     auto end = std::end(input);
     auto beg = std::find_if(std::begin(input), end, isAlpha);
@@ -89,7 +88,6 @@ int main(int argc, char * argv[])
     for (auto wordCount : output) {
         o << wordCount->second << ' ' << wordCount->first << '\n';
     }
-
     timer.report("write output");
 
     return EXIT_SUCCESS;

run.bash

@@ -1,34 +1,80 @@
+#! /usr/bin/bash
+
 set -e
-if ! [[ -x $1 ]]
+
+if [[ ! -f pg.txt ]]
+then
+    >&2 echo "No pg.txt in current directory"
+    exit 2
+fi
+
+if [[ ! -x $1 ]]
 then
-    >&2 echo "Usage: bash run.bash <executable> <N>"
-    exit 1
+    >&2 echo "Usage: bash run.bash ABSOLUTE_PATH_TO_EXECUTABLE [N]"
+    exit 3
 fi
 
-if [[ $2 =~ [[:digit:]]+ ]]
+if [[ ! $1 = /* ]]
 then
-    N="$2"
+    >&2 echo "Value '$1' of the first parameter is not an absolute path"
+    exit 4
+fi
+
+if [[ $2 ]]
+then
+    if [[ ! $2 =~ ^[[:digit:]]+$ ]]
+    then
+        >&2 echo "Value '$2' of the second parameter is not a number"
+        exit 5
+    fi
+    N=$2
 else
     N=1
 fi
 
-rsync pg.txt /tmp/
-rsync ref.txt /tmp/out.txt
-md5sum /tmp/out.txt >/tmp/out.txt.md5
+if [[ ! -f pg.txt ]]
+then
+    >&2 echo "File pg.txt is not found in current directory"
+fi
+
+if ! WORKSPACE="$( mktemp -d --tmpdir 'freq.XXXXXX' )"
+then
+    >&2 echo "Unable to create temporary directory"
+    exit 7
+fi
 
-sudo find /sys/devices/system/cpu -name scaling_governor -exec sh -c 'echo performance >{}' ';'
-sudo sh -c 'echo off >/sys/devices/system/cpu/smt/control ; tuna --cpus=1-7 --isolate'
+NPROC="$( nproc )"
 
 function on_exit {
-    sudo sh -c 'echo on >/sys/devices/system/cpu/smt/control ; tuna --cpus=0-15 --include'
-    sudo find /sys/devices/system/cpu -name scaling_governor -exec sh -c 'echo powersave >{}' ';'
+    set +e
+    if [[ $( cat /sys/devices/system/cpu/smt/control ) == "off" ]]
+    then
+        sudo sh -c 'echo on >/sys/devices/system/cpu/smt/control'
+    fi
+    sudo sh -c "tuna --cpus=0-$NPROC --include"
+    echo powersave | sudo tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
+    rm -r "$WORKSPACE"
 }
 trap on_exit EXIT
 
-for (( i = 0 ; i < N; ++i ))
+cp -a pg.txt "$WORKSPACE"
+echo performance | sudo tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
+if [[ NPROC > 1 ]]
+then
+    if [[ $( cat /sys/devices/system/cpu/smt/control ) == "on" ]]
+    then
+        sudo sh -c 'echo off >/sys/devices/system/cpu/smt/control'
+    fi
+    sudo sh -c "tuna --cpus=1-$NPROC --isolate"
+fi
+
+pushd "$WORKSPACE"
+echo '850944413ba9fd1dbf2b9694abaa930d  -' >out.txt.md5sum
+for (( i = 0 ; i < N ; ++i ))
 do
-    rm /tmp/out.txt
-    time LC_ALL="C" taskset --cpu-list 1-7 "$1" /tmp/pg.txt /tmp/out.txt
-    md5sum -c /tmp/out.txt.md5
+    time LC_ALL=C taskset --cpu-list 1-$NPROC "$1" pg.txt out.txt
+    >&2 echo -n
+    cat out.txt | md5sum --check out.txt.md5sum
+    rm out.txt
 done
-
+popd