Changeset cdbdcb1c for general/genlib/fftw.xml

Timestamp:

02/12/2023 05:46:15 AM (17 months ago)

Author:

Xi Ruoyao <xry111@…>

Branches:

11.3, 12.0, 12.1, kea, ken/TL2024, ken/inkscape-core-mods, ken/tuningfonts, lazarus, lxqt, plabs/newcss, python3.11, qt5new, rahul/power-profiles-daemon, renodr/vulkan-addition, trunk, xry111/llvm18, xry111/xf86-video-removal

Children:

f848bc0

Parents:

90354d5

git-author:

Xi Ruoyao <xry111@…> (02/12/2023 05:26:44 AM)

git-committer:

Xi Ruoyao <xry111@…> (02/12/2023 05:46:15 AM)

Message:

fftw: Add --enable-avx2, and document --enable-{sse2,avx,avx2,avx512}

The first AVX2-capable CPUs are released in 2013 (ten years ago)
and now all recent Intel Core and AMD Ryzen models support it. So
it makes sense to enable it now. AVX512F is much uncommon now for
non-server CPU models (only Rocket Lake, Alder Lake with E-cores
disabled, and Zen 4), so leave it as <option>.

File:

• general/genlib/fftw.xml (modified) (4 diffs)

general/genlib/fftw.xml

@@ -98 +98 @@
             --enable-threads \
             --enable-sse2    \
-            --enable-avx     &&
+            --enable-avx     \
+            --enable-avx2    &&
 make</userinput></screen>
 
@@ -125 +126 @@
             --enable-sse2    \
             --enable-avx     \
+            --enable-avx2    \
             --enable-float   &amp;&amp;
 make</userinput></screen>
@@ -171 +173 @@
 
       <para>
+        <parameter>--enable-{sse2,avx,avx2}</parameter>: These enables building
+        the optimized routines using SSE2, AVX, and AVX2 instructions.  FFTW
+        will check if these routines can be really used on the current CPU when
+        the FFTW library is loaded, so a FFTW build with these routines enabled
+        can still run on a CPU without SSE2, AVX, or AVX512.  These options
+        are not compatible with <parameter>--enable-long-double</parameter>.
+      </para>
+
+      <para>
         <parameter>--enable-float</parameter>: This enables building the library that
         uses single precision floating point arithmetic.  It is faster but less
@@ -185 +196 @@
       </para>
 
+      <para>
+        <option>--enable-avx512</option>: This enables building the optimized
+        routines using AVX512F instructions.  FFTW will check if these routines
+        can be really used on the current CPU when the FFTW library is loaded,
+        so a FFTW build with these routines enabled can still run on a CPU
+        without AVX512F.  Use this option if the FFTW build will be used on
+        a CPU with AVX512F.  This option is not compatible with
+        <option>--enable-long-double</option>.
+      </para>
   </sect2>