fix output

[mohanchen]

fix output

[dzzz2001]

In K_Vectors::set_kup_and_kdw (source_cell/klist.cpp:651-664), nkstot is only doubled for nspin == 2; nspin == 4 falls through without modifying nkstot. So the new formula will display nkstot / 4 for noncollinear / SOC runs — a 4× underreport.
Suggest either: "const int real_nk = (kv.get_nspin() == 2) ? kv.get_nkstot() / 2 : kv.get_nkstot();" or, better, expose an accessor on K_Vectors that returns the spin-independent k-point count and use it everywhere this distinction matters.

[dzzz2001]

The old name documented the operation: take a 2D-block-cyclic DM and reshape it into the gint serial layout. transfer_dm2d reads as "transfer a 2D dm" which doesn't say where to. Since the rename ripples through 9 files plus timer/TITLE labels (which downstream profiling scripts may grep), the win has to clear a high bar. I'd push back on this unless there's a follow-up reason.

If you do want a shorter name, “dm_2d_to_gint” is a better landing spot: drop the verb transfer (the X_to_Y form already conveys "convert/reshape" by convention) but keep both the source layout (2d) and the destination (gint). It goes from 22 → 13 chars without losing information, and the timer/TITLE labels can migrate with a single grep.

[dzzz2001]

The first added line is indented with a tab; the second with spaces. The surrounding file uses spaces — please normalize both to 12 spaces (matching the surrounding code).

[dzzz2001]

A pre-existing bug: print(find) is called from both record and record_gpu and unconditionally prints consume[find] — never consume_gpu[find]. A GPU allocation that crosses the new 20 MB line will print whatever happens to be in the CPU slot (often 0). Worth fixing while you're here.

Recommended fix: pass the data, drop the index
Decouple print from the global table layout entirely. It only needs a name and a size:

// memory.h
static void print(const std::string& name, double size_mb);

// memory.cpp
void Memory::print(const std::string& mem_name, double size_mb)
{
    GlobalV::ofs_running << "\n *** Memory Allocation Warning *** "
        << " " << mem_name << " " << size_mb << " MB" << std::endl;
}

Then at every call site, replace print(find); with the explicit form:

// CPU path
if (consume[find] > memory_warning_threshold_mb) {
    print(name[find], consume[find]);
}

// GPU path
if (consume_gpu[find] > memory_warning_threshold_mb) {
    print(name_gpu[find], consume_gpu[find]);
}

This is the cleanest fix — print no longer has a hidden coupling to "which table am I being called from", and it becomes trivially testable.