Merge pull request #1628 from albinahlback/fast_flint

Add -O3 and -march=native to default CFLAGS
flintlib · Nov 16, 2023 · f09196d · f09196d
2 parents 4bbb654 + 43da347
commit f09196d
Show file tree

Hide file tree

Showing 4 changed files with 33 additions and 6 deletions.
diff --git a/.github/workflows/CI.yml b/.github/workflows/CI.yml
@@ -16,10 +16,10 @@ concurrency:
 
 jobs:
   ##############################################################################
-  # ubuntu, gcc, avx2, code coverage
+  # ubuntu, gcc, code coverage
   ##############################################################################
-  ubuntu-avx2-codecoverage:
-    name: Ubuntu GCC, AVX2, Code Coverage (x10)
+  ubuntu-codecoverage:
+    name: Ubuntu GCC, Code Coverage (x10)
 
     runs-on: ubuntu-latest
 

diff --git a/configure.ac b/configure.ac
@@ -958,7 +958,11 @@ AX_CHECK_COMPILE_FLAG([-Wall],[DEFAULT_CFLAGS="-Wall $DEFAULT_CFLAGS"])
 AX_CHECK_COMPILE_FLAG([-Wno-stringop-overread],[DEFAULT_CFLAGS="-Wno-stringop-overread $DEFAULT_CFLAGS"])
 AX_CHECK_COMPILE_FLAG([-Wno-stringop-overflow],[DEFAULT_CFLAGS="-Wno-stringop-overflow $DEFAULT_CFLAGS"])
 AX_CHECK_COMPILE_FLAG([-Werror=implicit-function-declaration],[DEFAULT_CFLAGS="-Werror=implicit-function-declaration $DEFAULT_CFLAGS"])
-AX_CHECK_COMPILE_FLAG([-O2],[DEFAULT_CFLAGS="-O2 $DEFAULT_CFLAGS"])
+AX_CHECK_COMPILE_FLAG([-O3],[DEFAULT_CFLAGS="-O3 $DEFAULT_CFLAGS"])
+if test "$host" = "$build";
+then
+    AX_CHECK_COMPILE_FLAG([-march=native],[DEFAULT_CFLAGS="-march=native $DEFAULT_CFLAGS"])
+fi
 AX_CHECK_COMPILE_FLAG([-std=c11],[DEFAULT_CFLAGS="-std=c11 $DEFAULT_CFLAGS"])
 AX_CHECK_COMPILE_FLAG([-pedantic],[DEFAULT_CFLAGS="-pedantic $DEFAULT_CFLAGS"])
 
@@ -1028,7 +1032,15 @@ error
 error
 #endif
              ])],
-            [enable_fft_small="yes"]
+            [AC_COMPILE_IFELSE(
+                [AC_LANG_PROGRAM([#include <gmp.h>],[
+#if GMP_LIMB_BITS == 32
+#error Dead man
+error
+#endif
+                 ])],
+                [enable_fft_small="yes"]
+            )]
         )]
     )
 fi

diff --git a/src/fft_small/sd_fft_ctx.c b/src/fft_small/sd_fft_ctx.c
@@ -13,6 +13,12 @@
 #include "fft_small.h"
 #include "nmod.h"
 
+#if defined(_MSC_VER) || defined(__MINGW32__) || defined(__MINGW64__)
+# include <malloc.h>
+# define aligned_alloc(alignment, size) _aligned_malloc(size, alignment)
+# define free _aligned_free
+#endif
+
 void * flint_aligned_alloc(ulong alignment, ulong size)
 {
     void * p;
@@ -32,6 +38,11 @@ void flint_aligned_free(void * p)
     free(p);
 }
 
+#if defined(_MSC_VER) || defined(__MINGW32__) || defined(__MINGW64__)
+# undef aligned_alloc
+# undef free
+#endif
+
 void sd_fft_ctx_clear(sd_fft_ctx_t Q)
 {
     ulong k;

diff --git a/src/flint.h.in b/src/flint.h.in
@@ -132,7 +132,11 @@ typedef struct __FLINT_FILE FLINT_FILE;
 # define FLINT_OPTIMIZE(x) FLINT_OPTIMIZE_NESTED_1(x)
 #else
 # define __attribute__(x)
-# define FLINT_FORCE_INLINE static inline
+# if defined(_MSC_VER)
+#  define FLINT_FORCE_INLINE static __forceinline
+# else
+#  define FLINT_FORCE_INLINE static inline
+# endif
 # define FLINT_STATIC_NOINLINE static
 # define FLINT_UNUSED(x) x
 # define FLINT_SET_BUT_UNUSED(x) x