ENH: Add CPU feature detection for VSX4 (Power10)

2 files changed, 35 insertions, 0 deletions

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];
+}