Differences

This shows you the differences between two versions of the page.

--- technical:generic:openmpi-4-ucx-issue [2024-12-05 10:37] – frey
+++ technical:generic:openmpi-4-ucx-issue [2024-12-05 10:47] (current) – [How is it fixed] frey
@@ Line 75: / Line 75: @@
 </code>
-The GCC compiler implements a ''%%__builtin_ctz()%%'' function that may directly produce machine-level assembly code (very fast) when the target ISA supports it.  The surrounding code in the UCX PML ensures the integer value is a power of 2 and non-zero, so the fix simplifies to
+Adding the precondition that ''v ≠ 0'' removes the leading conditional
+<code c>
+int ctz(unsigned int v)
+{
+    int     l = 0;
+    while ( (v & 1) == 0 ) l++, v >>= 1;
+    return l;
+}
+</code>
+and if ''v'' is guaranteed to be a power of two -- implying a single bit is set -- the code becomes
+<code c>
+int ctz(unsigned int v)
+{
+    int     l = -1;
+    do { l++, v >>= 1; } while (v);
+    return l;
+}
+</code>
+The GCC compiler implements a ''%%__builtin_ctz()%%'' function that may directly produce machine-level assembly code (very fast) when the target ISA supports it.  The surrounding code in the UCX PML ensures the integer value is a power of 2 and non-zero, so the final form above is permissible:
 <code c>