adding in which monomial the timing is measured

[Marcogarofalo]

I introduce a global char * which is set to the working monomial in the update

[kostrzewa]

I would prefer not to have yet another global variable to take care of. I think it would be better to add arguments to tm_stopwatch_push, such that it can also print at some debug level (let's say >= 3).

[Marcogarofalo]

Then how tm_stopwatch_push should know the monomial? The other option that I can think of is to pass the monomial name to all the functions like sw_spinor_eo, sw_all, deriv_Sb, ... Do you have something better?

[kostrzewa]

It doesn't need to. The only thing that's required is that inside the monomial functions, one calls:

tm_stopwatch_push(&g_timers, 0, 3, mnl->name, __func__);

which then prints something like

# monomial_name: Entering __func__ level: %d

In the analysis of the output, the timings can then be associated to the correct context by reading monomial_name and the level and the final output of:

tm_stopwatch_pop(&g_timers, 0, 1, mnl->name, __func__);

Anything at levels higher than these two outputs belongs to that context.

[kostrzewa]

Another option is to add an explicit "context" information for tm_stopwatch_push. In other words, the g_timers struct contains another array of strings (with some good length each, up to 500 or so).

tm_stopwatch_push(&g_timers, "context");

which then stores "context" in g_timers.context[level]. This context could be, for example, mnl->name and can even be used recursively in the logic (if the argument is empty, use the context of the previous level).

tm_stopwatch_push(&g_timers, "")

would thus result (internally) in

g_timers.context[level] = g_timers.context[level-1];

And then you can even have the context output at every call.

[Marcogarofalo]

ok, then in a function like

void deriv_Sb(const int ieo, spinor * const l, spinor * const k,
              hamiltonian_field_t * const hf, const double factor) 

I have to add

void deriv_Sb(const int ieo, spinor * const l, spinor * const k,
              hamiltonian_field_t * const hf, const double factor, const char* mnl_name ) 

[kostrzewa]

I have to add

No, it's not necessary. When you have:

Entering xxx in yyy, level: n

in the output, anything above this level is automatically within that context. In other words: the deriv_Sb used within the monomial function will automatically have level: n+1 and can be associated correctly at the analysis stage.

In addition, doing what I wrote above in https://github.com/etmc/tmLQCD/pull/505#issuecomment-918173628 would even give you the context locally.

[kostrzewa]

To be more concrete, you could have:

tm_stopwatch_push(&g_timers, mnl->name);

at the top of det_derivative, for example. Internally, this sets

timers->context[level] = "name of the monomial";

Then, when tm_stopwatch_push(&g_timers, "") is called with an empty context (for example, inside deriv_Sb), this instead sets

timers->context[level] = timers->context[level-1];

which can then be printed as part of the tm_stopwatch_pop inside of the function is question. This way all pops would be instrumented and one would only provide context when this makes sense from a design and/or logical pov.

[kostrzewa]

@Marcogarofalo can you push this branch to the etmc repo please?