Enable CPU optimizations for x86 and arm64.

Some libFLAC functions can be optimized by using architecture-specific
functions in place of the generic (standard C) implementations.
Enable these optimizations if supported by the compiler.

Author: Benjamin Moody

setup.py

@@ -110,6 +110,30 @@ def try_compile(src_name, extra_compile_args=[], trap_errors=True):
 
     if try_compile('conftest_fseeko.c'):
         define_macros += [('PLIBFLAC_HAVE_FSEEKO', '1')]
+    if try_compile('conftest_cpuid_h.c'):
+        define_macros += [('PLIBFLAC_HAVE_CPUID_H', '1')]
+
+    machine = platform.machine().lower()
+    if re.match('i[3-6]86', machine):
+        define_macros += [('FLAC__CPU_IA32', '1'),
+                          ('FLAC__USE_AVX', '1'),
+                          ('FLAC__ALIGN_MALLOC_DATA', '1')]
+        if try_compile('conftest_x86intrin.c'):
+            define_macros += [('FLAC__HAS_X86INTRIN', '1')]
+    if re.match('x86_64|amd64', machine):
+        define_macros += [('FLAC__CPU_X86_64', '1'),
+                          ('FLAC__HAS_X86INTRIN', '1'),
+                          ('FLAC__USE_AVX', '1'),
+                          ('FLAC__ALIGN_MALLOC_DATA', '1')]
+        if try_compile('conftest_x86intrin.c'):
+            define_macros += [('FLAC__HAS_X86INTRIN', '1')]
+    if re.match('arm64|aarch64', machine):
+        define_macros += [('FLAC__CPU_ARM64', '1')]
+        if try_compile('conftest_a64neonintrin.c'):
+            define_macros += [('FLAC__HAS_NEONINTRIN', '1'),
+                              ('FLAC__HAS_A64NEONINTRIN', '1')]
+        if try_compile('conftest_neonintrin.c'):
+            define_macros += [('FLAC__HAS_NEONINTRIN', '1')]
 
     return {
         'sources': sources,

src/conf/conftest_a64neonintrin.c

@@ -0,0 +1,8 @@
+/* Test for FLAC__HAS_A64NEONINTRIN */
+#include <arm_neon.h>
+int main()
+{
+    float64x2_t tmp;
+    tmp = vdupq_n_f64(0.0f);
+    return 0;
+}

src/conf/conftest_cpuid_h.c

@@ -0,0 +1,6 @@
+/* Test for HAVE_CPUID_H */
+#include <cpuid.h>
+int main()
+{
+    return 0;
+}

src/conf/conftest_neonintrin.c

@@ -0,0 +1,6 @@
+/* Test for FLAC__HAS_NEONINTRIN */
+#include <arm_neon.h>
+int main()
+{
+    return 0;
+}

src/conf/conftest_x86intrin.c

@@ -0,0 +1,10 @@
+/* Test for FLAC__HAS_X86INTRIN */
+#ifdef _MSC_VER
+# include <intrin.h>
+#else
+# include <x86intrin.h>
+#endif
+int main()
+{
+    return 0;
+}

src/flac/config.h

@@ -65,7 +65,9 @@
 //#cmakedefine HAVE_CLOCK_GETTIME
 
 /* Define to 1 if you have the <cpuid.h> header file. */
-//#cmakedefine HAVE_CPUID_H
+#ifdef PLIBFLAC_HAVE_CPUID_H
+# define HAVE_CPUID_H 1
+#endif
 
 /* Define to 1 if fseeko (and presumably ftello) exists and is declared. */
 #ifdef PLIBFLAC_HAVE_FSEEKO