ENH: Add CPU feature detection for VSX4 (Power10)

4 files changed, 42 insertions, 1 deletions

diff --git a/numpy/distutils/ccompiler_opt.py b/numpy/distutils/ccompiler_opt.py
index f1d024b94..794d88433 100644
--- a/numpy/distutils/ccompiler_opt.py
+++ b/numpy/distutils/ccompiler_opt.py
@@ -294,6 +294,8 @@ class _Config:
         VSX2 = dict(interest=2, implies="VSX", implies_detect=False),
         ## Power9/ISA 3.00
         VSX3 = dict(interest=3, implies="VSX2", implies_detect=False),
+        ## Power10/ISA 3.1
+        VSX4 = dict(interest=4, implies="VSX3", extra_checks="VSX4_MMA", implies_detect=False),
         # IBM/Z
         ## VX(z13) support
         VX = dict(interest=1, headers="vecintrin.h"),
@@ -471,12 +473,16 @@ class _Config:
                 ),
                 VSX3 = dict(
                     flags="-mcpu=power9 -mtune=power9", implies_detect=False
+                ),
+                VSX4 = dict(
+                    flags="-mcpu=power10 -mtune=power10", implies_detect=False
                 )
             )
             if self.cc_is_clang:
                 partial["VSX"]["flags"]  = "-maltivec -mvsx"
                 partial["VSX2"]["flags"] = "-mpower8-vector"
                 partial["VSX3"]["flags"] = "-mpower9-vector"
+                partial["VSX4"]["flags"] = "-mpower10-vector"
 
             return partial
 
diff --git a/numpy/distutils/checks/cpu_vsx4.c b/numpy/distutils/checks/cpu_vsx4.c
new file mode 100644
index 000000000..a6acc7384
--- /dev/null
+++ b/numpy/distutils/checks/cpu_vsx4.c
@@ -0,0 +1,14 @@
+#ifndef __VSX__
+    #error "VSX is not supported"
+#endif
+#include <altivec.h>
+
+typedef __vector unsigned int v_uint32x4;
+
+int main(void)
+{
+    v_uint32x4 v1 = (v_uint32x4){2, 4, 8, 16};
+    v_uint32x4 v2 = (v_uint32x4){2, 2, 2, 2};
+    v_uint32x4 v3 = vec_mod(v1, v2);
+    return (int)vec_extractm(v3);
+}
diff --git a/numpy/distutils/checks/extra_vsx4_mma.c b/numpy/distutils/checks/extra_vsx4_mma.c
new file mode 100644
index 000000000..a70b2a9f6
--- /dev/null
+++ b/numpy/distutils/checks/extra_vsx4_mma.c
@@ -0,0 +1,21 @@
+#ifndef __VSX__
+    #error "VSX is not supported"
+#endif
+#include <altivec.h>
+
+typedef __vector float fv4sf_t;
+typedef __vector unsigned char vec_t;
+
+int main(void)
+{
+    __vector_quad acc0;
+    float a[4] = {0,1,2,3};
+    float b[4] = {0,1,2,3};
+    vec_t *va = (vec_t *) a;
+    vec_t *vb = (vec_t *) b;
+    __builtin_mma_xvf32ger(&acc0, va[0], vb[0]);
+    fv4sf_t result[4];
+    __builtin_mma_disassemble_acc((void *)result, &acc0);
+    fv4sf_t c0 = result[0];
+    return (int)((float*)&c0)[0];
+}
diff --git a/numpy/distutils/tests/test_ccompiler_opt.py b/numpy/distutils/tests/test_ccompiler_opt.py
index 6f9970c75..1f2408dc1 100644
--- a/numpy/distutils/tests/test_ccompiler_opt.py
+++ b/numpy/distutils/tests/test_ccompiler_opt.py
@@ -405,7 +405,7 @@ class _Test_CCompilerOpt:
                 # in msvc, avx512_knl avx512_knm aren't supported
                 x86_msvc=".* xop fma4 .* avx512f .* avx512_skx .*",
                 armhf=".* asimd asimdhp asimddp .*",
-                ppc64="vsx vsx2 vsx3.*",
+                ppc64="vsx vsx2 vsx3 vsx4.*",
                 s390x="vx vxe vxe2.*"
             )
         # min