BLIS tree barrier is broken

[Jeff Hammond]

I have an 80-core AArch64 CPU (Ampere Altra) and think it prudent to use the more scalable tree barrier designed for Xeon Phi on it.

Sadly, this code appears to have rotted.

Compiling obj/thunderx2/frame/thread/bli_thrcomm_openmp.o ('thunderx2' CFLAGS for framework code)
frame/thread/bli_thrcomm_openmp.c: In function ‘bli_thrcomm_init’:
frame/thread/bli_thrcomm_openmp.c:119:8: error: ‘thrcomm_t’ {aka ‘struct thrcomm_s’} has no member named ‘barriers’; did you mean ‘barrier_sense’?
  119 |  comm->barriers = bli_malloc_intl( sizeof( barrier_t* ) * n_threads, &r_val );
      |        ^~~~~~~~
      |        barrier_sense
compilation terminated due to -Wfatal-errors.

You can reproduce by making this change:

--- a/config/thunderx2/bli_family_thunderx2.h
+++ b/config/thunderx2/bli_family_thunderx2.h
@@ -36,6 +36,12 @@
 //#define BLIS_FAMILY_H
 
 
+// -- THREADING PARAMTERS ------------------------------------------------------
+
+#define BLIS_TREE_BARRIER
+#define BLIS_TREE_BARRIER_ARITY        4
+
+

[Jeff Hammond]

Ah, it seems that BLIS_TREE_BARRIER does not get defined in frame/thread/bli_thrcomm_openmp.h when it is defined in the above config file.

[Devin Matthews]

@jeffhammond I'm not sure what you mean by this? I gather there is still a bug? but I'm not sure what

[Jeff Hammond]

This is a trivial bug where struct members need to be updated to track the usage, when the preprocessor switch is on.

frame/thread/bli_thrcomm_openmp.c:119:8: error: ‘thrcomm_t’ {aka ‘struct thrcomm_s’} has no member named ‘barriers’; did you mean ‘barrier_sense’?
  119 |  comm->barriers = bli_malloc_intl( sizeof( barrier_t* ) * n_threads, &r_val );
      |        ^~~~~~~~
      |        barrier_sense

Whether there is a preprocessor/configuration bug here is TBD.