Merge pull request #1813 from albinahlback/fft_small_threshold_arm

More thresholds for fft_small
flintlib · Mar 5, 2024 · 3e5b15d · 3e5b15d
2 parents 7b5dccf + be3cc3a
commit 3e5b15d
Show file tree

Hide file tree

Showing 3 changed files with 25 additions and 9 deletions.
diff --git a/CMake/cmake_config.h.in b/CMake/cmake_config.h.in
@@ -16,12 +16,15 @@
 
 #cmakedefine01 FLINT_USES_FENV
 
-#cmakedefine FLINT_HAVE_FFT_SMALL
+#cmakedefine01 FLINT_HAVE_FFT_SMALL
 
 #cmakedefine01 FLINT_KNOW_STRONG_ORDER
 
+#if FLINT_HAVE_FFT_SMALL
 /* Just set to some reasonable threshold */
-#define FLINT_FFT_SMALL_THRESHOLD 600
+# define FLINT_FFT_SMALL_MUL_THRESHOLD 800
+# define FLINT_FFT_SMALL_SQR_THRESHOLD 1400
+#endif
 
 #ifdef _MSC_VER
 # if defined(FLINT_BUILD_DLL)

diff --git a/configure.ac b/configure.ac
@@ -1003,18 +1003,31 @@ FLINT_CHECK_FFT_SMALL(
     [AC_SUBST(FFT_SMALL, [fft_small\ \ \ ])
      AC_DEFINE(FLINT_HAVE_FFT_SMALL, 1, [Define to use the fft_small module])
 
+     dnl FIXME: Push different configuration files instead of hardcoding them here.
      case $flint_cv_arch in
          FAST_VROUNDPD_PATTERN)
-             fft_small_threshold="400"
+             dnl Zen 3
+             fft_small_mul_threshold="400"
+             fft_small_sqr_threshold="800"
              ;;
          SLOW_VROUNDPD_PATTERN)
-             fft_small_threshold="1540"
+             dnl Skylake
+             fft_small_mul_threshold="1540"
+             fft_small_sqr_threshold="3080"
+             ;;
+         ARM64_PATTERN)
+             dnl Apple M1
+             fft_small_threshold="810"
+             fft_small_threshold="865"
              ;;
          *)
-             fft_small_threshold="500"
+             dnl Be conservative here
+             fft_small_mul_threshold="800"
+             fft_small_sqr_threshold="1400"
              ;;
      esac
-     AC_DEFINE_UNQUOTED(FLINT_FFT_SMALL_THRESHOLD,[$fft_small_threshold],[Define to set threshold for when to use fft_small module])],
+     AC_DEFINE_UNQUOTED(FLINT_FFT_SMALL_MUL_THRESHOLD,[$fft_small_mul_threshold],[Define to set threshold for when to use multiplication with fft_small module])
+     AC_DEFINE_UNQUOTED(FLINT_FFT_SMALL_SQR_THRESHOLD,[$fft_small_sqr_threshold],[Define to set threshold for when to use squaring with fft_small module])],
     [AC_SUBST(FFT_SMALL, [\ \ \ \ \ \ \ \ \ \ \ \ ])])
 
 ################################################################################

diff --git a/src/mpn_extras.h b/src/mpn_extras.h
@@ -130,9 +130,9 @@ mp_limb_t flint_mpn_2add_n_inplace(mp_ptr, mp_srcptr, mp_srcptr, mp_size_t);
 
 /* General multiplication ****************************************************/
 
-#ifdef FLINT_HAVE_FFT_SMALL
-# define FLINT_FFT_MUL_THRESHOLD FLINT_FFT_SMALL_THRESHOLD
-# define FLINT_FFT_SQR_THRESHOLD (2 * FLINT_FFT_SMALL_THRESHOLD)
+#if FLINT_HAVE_FFT_SMALL
+# define FLINT_FFT_MUL_THRESHOLD FLINT_FFT_SMALL_MUL_THRESHOLD
+# define FLINT_FFT_SQR_THRESHOLD FLINT_FFT_SMALL_SQR_THRESHOLD
 #else
 /* FLINT's FFT can beat GMP below this threshold but apparently
    not consistently. Something needs retuning? */