use cpucycles

CMakeLists.txt

@@ -13,6 +13,22 @@ endif()
 
 project(SMHasher C CXX ASM)
 
+# cat https://cpucycles.cr.yp.to/libcpucycles-latest-version.txt
+set(libcpucycle_version 20240318)
+include(ExternalProject)
+# libcpucycle-prefix/src/libcpucycle/build/amd64/package/lib/libcpucycles.a
+# => lib/libcpucycles.a
+ExternalProject_Add(
+  libcpucycle
+  URL https://cpucycles.cr.yp.to/libcpucycles-${libcpucycle_version}.tar.gz
+  DOWNLOAD_EXTRACT_TIMESTAMP false
+  #PREFIX ""
+  BINARY_DIR libcpucycle-prefix/src/libcpucycle
+  CONFIGURE_COMMAND cd ../libcpucycle && ./configure --prefix=${CMAKE_BINARY_DIR}
+  BUILD_COMMAND cd ../libcpucycle && make
+  INSTALL_COMMAND cd ../libcpucycle && make install
+  )
+
 include(CheckCCompilerFlag)
 # Check if the same compile family is used for both C and CXX
 if(NOT (CMAKE_C_COMPILER_ID STREQUAL CMAKE_CXX_COMPILER_ID))
@@ -669,6 +685,10 @@ if((CMAKE_MAJOR_VERSION EQUAL 3 AND CMAKE_MINOR_VERSION GREATER_EQUAL 9)
   check_ipo_supported(RESULT ipo_supported OUTPUT error)
 endif()
 
+include_directories(
+  ${CMAKE_BINARY_DIR}/include
+  )
+
 add_library(
   SMHasherSupport STATIC
   AvalancheTest.cpp
@@ -770,7 +790,8 @@ endif()
 
 target_link_libraries(SMHasher SMHasherSupport
                       ${HIGHWAY_LIB} ${BLAKE3_LIB} ${AHASH_C_LIB}
-                      ${CMAKE_THREAD_LIBS_INIT})
+                      ${CMAKE_THREAD_LIBS_INIT}
+                      ${CMAKE_BINARY_DIR}/lib/libcpucycles.a)
 
 SET(exectargets ${exectargets} SMHasher)
 # add_executable( bittest bittest.cpp )

SpeedTest.cpp

@@ -14,6 +14,8 @@
 #include <parallel_hashmap/phmap.h>
 #include <functional>
 
+#include "cpucycles.h"
+
 #define COUNT_OF(x) ((sizeof(x)/sizeof(0[x])) / ((size_t)(!(sizeof(x) % sizeof(0[x])))))
 #define ARRAY_END(x) (&(x)[COUNT_OF(x)])
 
@@ -275,23 +277,27 @@ void BulkSpeedTest ( pfHash hash, uint32_t seed )
   const int trials = 2999;
   const int blocksize = 256 * 1024;
 
+  printf("Using libcpucycles-%s\n", cpucycles_version());
   printf("Bulk speed test - %d-byte keys\n",blocksize);
   double sumbpc = 0.0;
 
   volatile double warmup_cycles = SpeedTest(hash,seed,trials,blocksize,0);
+  long long cpu_cycles = cpucycles_persecond();
 
   for(int align = 7; align >= 0; align--)
   {
     double cycles = SpeedTest(hash,seed,trials,blocksize,align);
 
     double bestbpc = double(blocksize)/cycles;
 
-    double bestbps = (bestbpc * 3000000000.0 / 1048576.0);
-    printf("Alignment %2d - %6.3f bytes/cycle - %7.2f MiB/sec @ 3 ghz\n",align,bestbpc,bestbps);
+    double bestbps = (bestbpc * cpu_cycles / 1048576.0) / 100000.0;
+    printf("Alignment %2d - %6.3f bytes/cycle - %7.2f MiB/sec @ %0.1f GHz\n",
+           align, bestbpc, bestbps);
     sumbpc += bestbpc;
   }
   sumbpc = sumbpc / 8.0;
-  printf("Average      - %6.3f bytes/cycle - %7.2f MiB/sec @ 3 ghz\n",sumbpc,(sumbpc * 3000000000.0 / 1048576.0));
+  printf("Average      - %6.3f bytes/cycle - %7.2f MiB/sec @ %0.1f GHz\n",
+         sumbpc, (sumbpc * cpu_cycles / 1048576.0) / 100000.0);
   fflush(NULL);
 }