Test for SSE4 instructions on the host before enabling them in the python extension module.

dlib/cmake_utils/check_if_sse4_instructions_executable_on_host.cmake

@@ -0,0 +1,19 @@
+# This script checks if your compiler and host processor can generate and then run programs with SSE4 instructions.
+
+cmake_minimum_required(VERSION 2.8.12)
+
+# Don't rerun this script if its already been executed.
+if (DEFINED SSE4_IS_AVAILABLE_ON_HOST)
+   return()
+endif()
+
+# Set to false unless we find out otherwise in the code below.
+set(SSE4_IS_AVAILABLE_ON_HOST 0)
+
+try_compile(test_for_sse4_worked ${PROJECT_BINARY_DIR}/sse4_test_build ${CMAKE_CURRENT_LIST_DIR}/test_for_sse4
+	sse4_test)
+
+if(test_for_sse4_worked)
+	message (STATUS "SSE4 instructions can be executed by the host processor.")
+	set(SSE4_IS_AVAILABLE_ON_HOST 1)
+endif()

dlib/cmake_utils/set_compiler_specific_options.cmake

@@ -9,8 +9,11 @@ endif()
 
 # Check if we are being built as part of a pybind11 module. 
 if (COMMAND pybind11_add_module)
-   # For python users, assume they have SSE4 at least and then if the host machine has AVX use that too.
-   set(USE_SSE4_INSTRUCTIONS ON CACHE BOOL "Use SSE4 instructions")
+   # For python users, enable SSE4 and AVX if they have these instructions.
+   include(${CMAKE_CURRENT_LIST_DIR}/check_if_sse4_instructions_executable_on_host.cmake)
+   if (SSE4_IS_AVAILABLE_ON_HOST)
+      set(USE_SSE4_INSTRUCTIONS ON CACHE BOOL "Use SSE4 instructions")
+   endif()
    include(${CMAKE_CURRENT_LIST_DIR}/check_if_avx_instructions_executable_on_host.cmake)
    if (AVX_IS_AVAILABLE_ON_HOST)
       set(USE_AVX_INSTRUCTIONS ON CACHE BOOL "Use AVX instructions")

dlib/cmake_utils/test_for_sse4/CMakeLists.txt

@@ -0,0 +1,23 @@
+
+cmake_minimum_required(VERSION 2.8.12)
+project(sse4_test)
+
+set(USE_SSE4_INSTRUCTIONS ON CACHE BOOL "Use SSE4 instructions")
+
+# Pull this in since it sets the SSE4 compile options by putting that kind of stuff into the active_compile_opts list.
+include(../set_compiler_specific_options.cmake)
+
+
+try_run(run_result compile_result ${PROJECT_BINARY_DIR}/sse4_test_try_run_build ${CMAKE_CURRENT_LIST_DIR}/sse4_test.cpp
+   COMPILE_DEFINITIONS ${active_compile_opts})
+
+message(STATUS "run_result = ${run_result}")
+message(STATUS "compile_result = ${compile_result}")
+
+if ("${run_result}" EQUAL 0 AND compile_result)
+   message(STATUS "Ran SSE4 test program successfully, you have SSE4 available.")
+else()
+   message(STATUS "Unable to run SSE4 test program, you don't seem to have SSE4 instructions available.")
+   # make this build fail so that calling try_compile statements will error in this case.
+   add_library(make_this_build_fail ${CMAKE_CURRENT_LIST_DIR}/this_file_doesnt_compile.cpp)
+endif()

dlib/cmake_utils/test_for_sse4/sse4_test.cpp

@@ -0,0 +1,18 @@
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+#include <mmintrin.h>
+#include <pmmintrin.h> // SSE3
+#include <tmmintrin.h>
+#include <smmintrin.h> // SSE4
+
+int main()
+{
+    __m128 x;
+    x = _mm_set1_ps(1.23);
+    x = _mm_ceil_ps(x);
+    return 0;
+}
+
+// ------------------------------------------------------------------------------------
+

dlib/cmake_utils/test_for_sse4/this_file_doesnt_compile.cpp

@@ -0,0 +1,3 @@
+
+#error "This file doesn't compile!"
+