runtime dispatching with MIPP library, papi for performance measurement

* pf_conv_* libraries for naive convolution with architecture specific optimizations - utilizing MIPP library for SIMD and benchmark utilitzing papi for measurement * MIPP with [un]install patch from https://github.com/hayguen/MIPP should be installed * bench_mixers also uses papi (if available) * fixed address sanitizer (asan) * removed cmake options DISABLE_SIMD_AVX, USE_SIMD_NEON * target_optimizations.cmake: - added optional additional(extra) options to target_set_c[xx]_arch_flags() - added gcc_clang_fpu options to target_set_cxx_arch_option() macro * with 'neon' variants for dispatching on cmake processor 'armv7l' papi doesn't work on this 32 bit Raspbian on the Raspberry Pi 4B hardware, but linux 'time' shows a total speedup of nearly factor 2 comparing '-a 0' (none) agains '-a <n>' (neon) Signed-off-by: hayati ayguen <[email protected]>
marton78 · Feb 5, 2022 · dc74fd4 · dc74fd4
1 parent 08bd199
commit dc74fd4
Show file tree

Hide file tree

Showing 16 changed files with 1,734 additions and 228 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -7,8 +7,6 @@ option(USE_TYPE_DOUBLE "activate 'double' precision float?" ON)
 
 # architecture/optimization options
 option(USE_SIMD        "use SIMD (SSE/AVX/NEON/ALTIVEC) CPU features? - " ON)
-option(DISABLE_SIMD_AVX "disable AVX CPU features? - " OFF)
-option(USE_SIMD_NEON   "force using NEON on ARM? (requires USE_SIMD)" OFF)
 option(USE_SCALAR_VECT "use 4-element vector scalar operations (if no other SIMD)" ON)
 
 # what to install?
@@ -46,6 +44,18 @@ if ( (NOT USE_TYPE_FLOAT) AND (NOT USE_TYPE_DOUBLE) )
   message(FATAL_ERROR "activate at least one of USE_TYPE_FLOAT or USE_TYPE_DOUBLE")
 endif()
 
+list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
+include(cmake/target_optimizations.cmake)
+include(cmake/compiler_warnings.cmake)
+find_package(PAPI)
+find_package(MIPP)
+if (MIPP_FOUND)
+# if (TARGET MIPP)
+    message(STATUS "found MIPP")
+else()
+    message(STATUS "NOT found MIPP")
+endif()
+
 
 if (USE_DEBUG_ASAN)
   set(ASANLIB "asan")
@@ -147,40 +157,18 @@ endif()
 add_library(PFFFT STATIC ${FLOAT_SOURCES} ${DOUBLE_SOURCES} pffft_common.c pffft_priv_impl.h pffft.hpp )
 set_target_properties(PFFFT PROPERTIES OUTPUT_NAME "pffft")
 target_compile_definitions(PFFFT PRIVATE _USE_MATH_DEFINES)
+target_activate_c_compiler_warnings(PFFFT)
 if (USE_SCALAR_VECT)
   target_compile_definitions(PFFFT PRIVATE PFFFT_SCALVEC_ENABLED=1)
 endif()
 if (USE_DEBUG_ASAN)
   target_compile_options(PFFFT PRIVATE "-fsanitize=address")
 endif()
+target_set_c_arch_flags(PFFFT)
 if (NOT USE_SIMD)
   target_compile_definitions(PFFFT PRIVATE PFFFT_SIMD_DISABLE=1)
 endif()
-if (USE_SIMD AND USE_SIMD_NEON)
-  target_compile_definitions(PFFFT PRIVATE PFFFT_ENABLE_NEON=1)
-  target_compile_options(PFFFT PRIVATE "-mfpu=neon")
-endif()
-if (USE_SIMD AND USE_TYPE_DOUBLE)
-  if(MSVC)
-    if(DISABLE_SIMD_AVX)
-      set_property(SOURCE pffft_double.c PROPERTY COMPILE_FLAGS "/arch:SSE2")
-    else()
-      set_property(SOURCE pffft_double.c PROPERTY COMPILE_FLAGS "/arch:AVX")
-    endif()
-  elseif(CMAKE_COMPILER_IS_GNUCC)
-    if(DISABLE_SIMD_AVX)
-      set_property(SOURCE pffft_double.c PROPERTY COMPILE_FLAGS "-msse2")
-    else()
-      set_property(SOURCE pffft_double.c PROPERTY COMPILE_FLAGS "-mavx")
-    endif()
-  else()
-    set_property(SOURCE pffft_double.c PROPERTY COMPILE_FLAGS "-march=native")
-  endif()
-  if(DISABLE_SIMD_AVX)
-    target_compile_definitions(PFFFT PRIVATE PFFFT_AVX_DISABLE=1)
-  endif()
-endif()
-target_link_libraries( PFFFT ${MATHLIB} )
+target_link_libraries( PFFFT ${ASANLIB} ${MATHLIB} )
 set_property(TARGET PFFFT APPEND PROPERTY INTERFACE_INCLUDE_DIRECTORIES
   $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}>
 )
@@ -193,23 +181,25 @@ endif()
 
 if (USE_TYPE_FLOAT)
 
-  add_library(PFDSP STATIC pf_mixer.cpp pf_mixer.h pf_carrier.cpp pf_carrier.h pf_cic.cpp pf_cic.h fmv.h )
+  add_library(PFDSP STATIC pf_mixer.cpp pf_mixer.h pf_cplx.h pf_carrier.cpp pf_carrier.h pf_cic.cpp pf_cic.h fmv.h )
+  set_property(TARGET PFDSP PROPERTY CXX_STANDARD 11)
+  set_property(TARGET PFDSP PROPERTY CXX_STANDARD_REQUIRED ON)
   set_target_properties(PFDSP PROPERTIES OUTPUT_NAME "pfdsp")
   target_compile_definitions(PFDSP PRIVATE _USE_MATH_DEFINES)
+  target_activate_cxx_compiler_warnings(PFDSP)
   if (USE_DEBUG_ASAN)
     target_compile_options(PFDSP PRIVATE "-fsanitize=address")
   endif()
-  if (USE_SIMD AND USE_SIMD_NEON)
-    target_compile_definitions(PFDSP PRIVATE PFFFT_ENABLE_NEON=1)
-    target_compile_options(PFDSP PRIVATE "-march=armv7-a" "-mfpu=neon")
+  if (USE_SIMD)
+    target_set_cxx_arch_flags(PFDSP)
   endif()
   target_link_libraries( PFDSP ${MATHLIB} )
   set_property(TARGET PFDSP APPEND PROPERTY INTERFACE_INCLUDE_DIRECTORIES
     $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}>
   )
   if (INSTALL_PFDSP)
     set(INSTALL_TARGETS ${INSTALL_TARGETS} PFDSP)
-    set(INSTALL_HEADERS ${INSTALL_HEADERS} pf_mixer.h pf_carrier.h pf_cic.h)
+    set(INSTALL_HEADERS ${INSTALL_HEADERS} pf_mixer.h pf_cplx.h pf_carrier.h pf_cic.h)
   endif()
 endif()
 
@@ -220,7 +210,7 @@ if (USE_FFTPACK)
   # float / single precision
   add_library(FFTPACK_FLOAT STATIC fftpack.c fftpack.h)
   target_compile_definitions(FFTPACK_FLOAT PRIVATE _USE_MATH_DEFINES)
-  target_compile_options(FFTPACK_FLOAT PRIVATE $<$<C_COMPILER_ID:GNU>:-Wall -Wextra -pedantic>)
+  target_activate_c_compiler_warnings(FFTPACK_FLOAT)
   target_link_libraries( FFTPACK_FLOAT ${MATHLIB} )
   set_property(TARGET FFTPACK_FLOAT APPEND PROPERTY INTERFACE_INCLUDE_DIRECTORIES
     $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}>
@@ -230,7 +220,7 @@ if (USE_FFTPACK)
   add_library(FFTPACK_DOUBLE STATIC fftpack.c fftpack.h)
   target_compile_definitions(FFTPACK_DOUBLE PRIVATE _USE_MATH_DEFINES)
   target_compile_definitions(FFTPACK_DOUBLE PUBLIC FFTPACK_DOUBLE_PRECISION)
-  target_compile_options(FFTPACK_DOUBLE PRIVATE $<$<C_COMPILER_ID:GNU>:-Wall -Wextra -pedantic>)
+  target_activate_c_compiler_warnings(FFTPACK_DOUBLE)
   target_link_libraries( FFTPACK_DOUBLE ${MATHLIB} )
   set_property(TARGET FFTPACK_DOUBLE APPEND PROPERTY INTERFACE_INCLUDE_DIRECTORIES
     $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}>
@@ -254,6 +244,7 @@ if (USE_TYPE_FLOAT)
   add_library(PFFASTCONV STATIC pffastconv.c pffastconv.h pffft.h )
   set_target_properties(PFFASTCONV PROPERTIES OUTPUT_NAME "pffastconv")
   target_compile_definitions(PFFASTCONV PRIVATE _USE_MATH_DEFINES)
+  target_activate_c_compiler_warnings(PFFASTCONV)
   if (USE_DEBUG_ASAN)
     target_compile_options(PFFASTCONV PRIVATE "-fsanitize=address")
   endif()
@@ -343,14 +334,12 @@ if (USE_TYPE_FLOAT)
   if (USE_DEBUG_ASAN)
     target_compile_options(test_pffastconv PRIVATE "-fsanitize=address")
   endif()
+  target_set_c_arch_flags(test_pffastconv)
   if (NOT USE_SIMD)
     target_compile_definitions(test_pffastconv PRIVATE PFFFT_SIMD_DISABLE=1)
   endif()
-  if (USE_SIMD AND USE_SIMD_NEON)
-    target_compile_definitions(test_pffastconv PRIVATE PFFFT_ENABLE_NEON=1)
-    target_compile_options(test_pffastconv PRIVATE "-mfpu=neon")
-  endif()
   target_link_libraries( test_pffastconv  PFFASTCONV ${ASANLIB} ${MATHLIB} )
+
 endif()
 
 ######################################################
@@ -359,6 +348,9 @@ if (USE_TYPE_FLOAT)
   add_executable(bench_pffft_float   bench_pffft.c pffft.h)
   target_compile_definitions(bench_pffft_float PRIVATE _USE_MATH_DEFINES)
   target_compile_definitions(bench_pffft_float PRIVATE PFFFT_ENABLE_FLOAT)
+  if (USE_DEBUG_ASAN)
+    target_compile_options(bench_pffft_float PRIVATE "-fsanitize=address")
+  endif()
 
   target_link_libraries( bench_pffft_float  PFFFT ${ASANLIB} )
 
@@ -391,7 +383,7 @@ if (USE_TYPE_FLOAT)
     if ( (CMAKE_SYSTEM_PROCESSOR STREQUAL "i686") OR (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64") )
       # has chances to work
     else()
-      # other PROCESSORs could be "ppc", "ppc64",  "arm", "aarch64" - or something else?!
+      # other PROCESSORs could be "ppc", "ppc64",  "arm", "aarch64", "armv7l" - or something else?!
       message(WARNING "using Intel MKL on '${CMAKE_SYSTEM_PROCESSOR}' might fail.")
     endif()
     message(STATUS "In case compiling/linking with Intel MKL fails, check CMakeLists.txt or deactivate USE_BENCH_MKL")
@@ -404,6 +396,9 @@ if (USE_TYPE_DOUBLE)
   add_executable(bench_pffft_double   bench_pffft.c pffft.h)
   target_compile_definitions(bench_pffft_double PRIVATE _USE_MATH_DEFINES)
   target_compile_definitions(bench_pffft_double PRIVATE PFFFT_ENABLE_DOUBLE)
+  if (USE_DEBUG_ASAN)
+    target_compile_options(bench_pffft_double PRIVATE "-fsanitize=address")
+  endif()
   target_link_libraries( bench_pffft_double  PFFFT ${ASANLIB} )
 
   if (USE_FFTPACK)
@@ -425,7 +420,7 @@ if (USE_TYPE_DOUBLE)
     if ( (CMAKE_SYSTEM_PROCESSOR STREQUAL "i686") OR (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64") )
       # has chances to work
     else()
-      # other PROCESSORs could be "ppc", "ppc64",  "arm", "aarch64" - or something else?!
+      # other PROCESSORs could be "ppc", "ppc64",  "arm", "aarch64", "armv7l" - or something else?!
       message(WARNING "using Intel MKL on '${CMAKE_SYSTEM_PROCESSOR}' might fail.")
     endif()
     message(STATUS "In case compiling/linking with Intel MKL fails, check CMakeLists.txt or deactivate USE_BENCH_MKL")
@@ -437,11 +432,153 @@ endif()
 ######################################################
 
 if (USE_TYPE_FLOAT)
-  add_executable(bench_pf_mixer_float   bench_mixers.c)
+
+  add_executable(bench_pf_mixer_float   bench_mixers.cpp papi_perf_counter.h)
   target_compile_definitions(bench_pf_mixer_float PRIVATE _USE_MATH_DEFINES)
   target_compile_definitions(bench_pf_mixer_float PRIVATE PFFFT_ENABLE_FLOAT)
+  target_link_libraries( bench_pf_mixer_float  ${ASANLIB} )
+  if (USE_DEBUG_ASAN)
+    target_compile_options(bench_pf_mixer_float PRIVATE "-fsanitize=address")
+  endif()
+  if (PAPI_FOUND)
+      target_compile_definitions(bench_pf_mixer_float PRIVATE HAVE_PAPI=1)
+      target_link_libraries(bench_pf_mixer_float ${PAPI_LIBRARIES})
+  endif()
+  target_link_libraries( bench_pf_mixer_float  PFDSP )
+
+
+  ############################################################################
+
+  add_library(pf_conv_arch_none pf_conv.cpp pf_conv.h pf_cplx.h)
+  target_compile_definitions(pf_conv_arch_none PRIVATE CONV_ARCH_POST=none MIPP_NO_INTRINSICS=1)
+  set_property(TARGET pf_conv_arch_none PROPERTY CXX_STANDARD 11)
+  set_property(TARGET pf_conv_arch_none PROPERTY CXX_STANDARD_REQUIRED ON)
+  target_activate_cxx_compiler_warnings(pf_conv_arch_none)
+  add_library(pf_conv_dispatcher  pf_conv_dispatcher.cpp pf_conv_dispatcher.h pf_conv.h pf_cplx.h)
+  set_property(TARGET pf_conv_dispatcher PROPERTY CXX_STANDARD 11)
+  set_property(TARGET pf_conv_dispatcher PROPERTY CXX_STANDARD_REQUIRED ON)
+  target_activate_cxx_compiler_warnings(pf_conv_dispatcher)
+
+  add_library(pf_conv_arch_dflt pf_conv.cpp pf_conv.h pf_cplx.h)
+  target_compile_definitions(pf_conv_arch_dflt PRIVATE CONV_ARCH_POST=dflt)
+  set_property(TARGET pf_conv_arch_dflt PROPERTY CXX_STANDARD 11)
+  set_property(TARGET pf_conv_arch_dflt PROPERTY CXX_STANDARD_REQUIRED ON)
+  target_activate_cxx_compiler_warnings(pf_conv_arch_dflt)
+  target_set_cxx_arch_flags(pf_conv_arch_dflt)
+
+  target_link_libraries(pf_conv_dispatcher pf_conv_arch_none pf_conv_arch_dflt)
+
+  if ((CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64") OR (CMAKE_SYSTEM_PROCESSOR MATCHES "AMD64"))
+
+      if ((CMAKE_CXX_COMPILER_ID STREQUAL "GNU") OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+          set(PF_CONV_ARCHES "sse3;sse4;avx;avx2")
+          set(PF_CONV_OPT_sse3 "core2")  # emulate a map
+          set(PF_CONV_OPT_sse4 "nehalem")
+          set(PF_CONV_OPT_avx  "sandybridge")
+          set(PF_CONV_OPT_avx2 "haswell")
+          target_compile_definitions(pf_conv_dispatcher PRIVATE CONV_ARCH_GCC_AMD64)
+
+      elseif (CMAKE_CXX_COMPILER_ID MATCHES "MSVC")
+          set(PF_CONV_ARCHES "sse2;avx;avx2")
+          set(PF_CONV_OPT_sse2 "SSE2")  # emulate a map
+          set(PF_CONV_OPT_avx  "AVX")
+          set(PF_CONV_OPT_avx2 "AVX2")
+          target_compile_definitions(pf_conv_dispatcher PRIVATE CONV_ARCH_MSVC_AMD64)
+
+      else()
+          set(PF_CONV_ARCHES "")
+          message(WARNING "unknown compiler ${CMAKE_CXX_COMPILER_ID} on CMAKE_SYSTEM_PROCESSOR ${CMAKE_SYSTEM_PROCESSOR}: can't do architecture specific compilation")
+      endif()
+
+  elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64")
+
+      if ((CMAKE_CXX_COMPILER_ID STREQUAL "GNU") OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+          set(PF_CONV_ARCHES "armv8a")
+          set(PF_CONV_OPT_armv8a   "armv8-a")  # emulate a map for arch
+
+          target_compile_definitions(pf_conv_dispatcher PRIVATE CONV_ARCH_GCC_AARCH64)
+      else()
+          set(PF_CONV_ARCHES "")
+          message(WARNING "unknown compiler ${CMAKE_CXX_COMPILER_ID} on CMAKE_SYSTEM_PROCESSOR ${CMAKE_SYSTEM_PROCESSOR}: can't do architecture specific compilation")
+      endif()
+
+  elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "armv7l")
+
+      if ((CMAKE_CXX_COMPILER_ID STREQUAL "GNU") OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+          set(PF_CONV_ARCHES "neon_vfpv4;neon_rpi3_a53;neon_rpi4_a72")
+          set(PF_CONV_OPT_neon_vfpv4        "armv7-a")    # emulate a map for arch
+          set(PF_CONV_EXTRA_neon_vfpv4      "neon_vfpv4") # emulate a map for additional options (EXTRA)
+          set(PF_CONV_OPT_neon_rpi3_a53     "armv7-a")
+          set(PF_CONV_EXTRA_neon_rpi3_a53   "neon_rpi3_a53")
+          set(PF_CONV_OPT_neon_rpi4_a72     "armv7-a")
+          set(PF_CONV_EXTRA_neon_rpi4_a72   "neon_rpi4_a72")
+
+          target_compile_definitions(pf_conv_dispatcher PRIVATE CONV_ARCH_GCC_ARM32NEON)
+      else()
+          set(PF_CONV_ARCHES "")
+          message(WARNING "unknown compiler ${CMAKE_CXX_COMPILER_ID} on CMAKE_SYSTEM_PROCESSOR ${CMAKE_SYSTEM_PROCESSOR}: can't do architecture specific compilation")
+      endif()
+
+  else()
+      message(WARNING "this is unforseen CMAKE_SYSTEM_PROCESSOR ${CMAKE_SYSTEM_PROCESSOR}: can't do architecture specific compilation")
+  endif()
+
+  foreach (arch_opt ${PF_CONV_ARCHES})
+      add_library(pf_conv_arch_${arch_opt} pf_conv.cpp pf_conv.h pf_cplx.h)
+      set_property(TARGET pf_conv_arch_${arch_opt} PROPERTY CXX_STANDARD 11)
+      set_property(TARGET pf_conv_arch_${arch_opt} PROPERTY CXX_STANDARD_REQUIRED ON)
+      target_activate_cxx_compiler_warnings(pf_conv_arch_${arch_opt})
+      target_compile_definitions(pf_conv_arch_${arch_opt} PRIVATE CONV_ARCH_POST=${arch_opt})
+
+      target_set_cxx_arch_option(pf_conv_arch_${arch_opt} "${PF_CONV_OPT_${arch_opt}}" "${PF_CONV_EXTRA_${arch_opt}}"  "${PF_CONV_OPT_${arch_opt}}")
+      target_link_libraries(pf_conv_dispatcher  pf_conv_arch_${arch_opt})
+      message(STATUS "added library pf_conv_arch_${arch_opt}  with CONV_ARCH_POST=${arch_opt}")
+  endforeach()
+
+  if (USE_DEBUG_ASAN)
+      foreach (arch_opt ${PF_CONV_ARCHES})
+          target_compile_options(pf_conv_arch_${arch_opt} PRIVATE "-fsanitize=address")
+          target_link_libraries( pf_conv_arch_${arch_opt} ${ASANLIB})
+      endforeach()
+
+      target_compile_options(pf_conv_arch_none  PRIVATE "-fsanitize=address")
+      target_link_libraries( pf_conv_arch_none  ${ASANLIB})
+
+      target_compile_options(pf_conv_dispatcher  PRIVATE "-fsanitize=address")
+      target_link_libraries(pf_conv_dispatcher ${ASANLIB})
+  endif()
+
+  if(MIPP_FOUND)
+      foreach (arch_opt ${PF_CONV_ARCHES})
+          message(STATUS "link pf_conv_arch_${arch_opt} against MIPP")
+          target_link_libraries(pf_conv_arch_${arch_opt} MIPP)
+      endforeach()
+
+      message(STATUS "link pf_conv_arch_none against MIPP")
+      target_link_libraries(pf_conv_arch_none MIPP)
+  endif()
+
+  ############################################################################
+
+
+  add_executable(bench_pf_conv_float   bench_conv.cpp papi_perf_counter.h)
+  set_property(TARGET bench_pf_conv_float PROPERTY CXX_STANDARD 11)
+  set_property(TARGET bench_pf_conv_float PROPERTY CXX_STANDARD_REQUIRED ON)
+  target_compile_definitions(bench_pf_conv_float PRIVATE _USE_MATH_DEFINES)
+  target_compile_definitions(bench_pf_conv_float PRIVATE PFFFT_ENABLE_FLOAT)
+  if (USE_DEBUG_ASAN)
+      target_compile_options(bench_pf_conv_float PRIVATE "-fsanitize=address")
+  endif()
+  target_link_libraries( bench_pf_conv_float  ${ASANLIB} )
+  if (PAPI_FOUND)
+      target_compile_definitions(bench_pf_conv_float PRIVATE HAVE_PAPI=1)
+      target_link_libraries(bench_pf_conv_float ${PAPI_LIBRARIES})
+  endif()
+  if(MIPP_FOUND)
+      target_link_libraries(bench_pf_conv_float MIPP)
+  endif()
 
-  target_link_libraries( bench_pf_mixer_float  PFDSP ${ASANLIB} )
+  target_link_libraries( bench_pf_conv_float  pf_conv_dispatcher PFDSP )
 
 endif()