wleoncio
commented
1 year ago Using Exact_cond_midP_unspecific_ordering_rx2(t(fontanella_2008), "decreasing") as an example, this is the gold standard profile, achieved when calc_prob.ExactCond_unspecific() is called directly:
$by.self
self.time self.pct total.time total.pct
"test_statistic" 33.92 65.36 41.96 80.85
"calc_prob.ExactCond_unspecific" 3.34 6.44 3.46 6.67
"calc_Pvalue_4x2.ExactCond_unspecific" 3.10 5.97 51.88 99.96
"rbind" 2.62 5.05 3.22 6.20
"rowSums" 1.96 3.78 2.82 5.43
"matrix" 1.64 3.16 1.70 3.28
"c" 0.74 1.43 0.74 1.43
"==" 0.62 1.19 0.62 1.19
"%in%" 0.54 1.04 0.72 1.39
":" 0.52 1.00 0.52 1.00
"is.data.frame" 0.44 0.85 0.44 0.85
"^" 0.42 0.81 0.42 0.81
"/" 0.32 0.62 0.32 0.62
"+" 0.32 0.62 0.32 0.62
"prod" 0.30 0.58 0.30 0.58
"(" 0.26 0.50 0.26 0.50
"-" 0.22 0.42 0.22 0.42
">" 0.22 0.42 0.22 0.42
"*" 0.16 0.31 0.16 0.31
"is.array" 0.06 0.12 0.06 0.12
"is.atomic" 0.06 0.12 0.06 0.12
"<" 0.04 0.08 0.04 0.08
"length" 0.04 0.08 0.04 0.08
"Exact_cond_midP_unspecific_ordering_rx2" 0.02 0.04 51.90 100.00
"dimnames" 0.02 0.04 0.02 0.04
$by.total
total.time total.pct self.time self.pct
"Exact_cond_midP_unspecific_ordering_rx2" 51.90 100.00 0.02 0.04
"calc_Pvalue_4x2.ExactCond_unspecific" 51.88 99.96 3.10 5.97
"test_statistic" 41.96 80.85 33.92 65.36
"calc_prob.ExactCond_unspecific" 3.46 6.67 3.34 6.44
"rbind" 3.22 6.20 2.62 5.05
"rowSums" 2.82 5.43 1.96 3.78
"matrix" 1.70 3.28 1.64 3.16
"c" 0.74 1.43 0.74 1.43
"%in%" 0.72 1.39 0.54 1.04
"==" 0.62 1.19 0.62 1.19
":" 0.52 1.00 0.52 1.00
"is.data.frame" 0.44 0.85 0.44 0.85
"^" 0.42 0.81 0.42 0.81
"/" 0.32 0.62 0.32 0.62
"+" 0.32 0.62 0.32 0.62
"prod" 0.30 0.58 0.30 0.58
"(" 0.26 0.50 0.26 0.50
"-" 0.22 0.42 0.22 0.42
">" 0.22 0.42 0.22 0.42
"*" 0.16 0.31 0.16 0.31
"is.array" 0.06 0.12 0.06 0.12
"is.atomic" 0.06 0.12 0.06 0.12
"<" 0.04 0.08 0.04 0.08
"length" 0.04 0.08 0.04 0.08
"dimnames" 0.02 0.04 0.02 0.04
$sample.interval
[1] 0.02
$sampling.time
[1] 51.9As a matter of fact, we can measure the overhead of convertFunName2Method() by adding a breakpoint on calc_Pvalue_4x2.ExactCond_unspecific() and running the following:
Browse[1]> microbenchmark::microbenchmark(
"method" = calc_prob.ExactCond_unspecific(x[, 1], 4, N_choose_np1, nip_choose_xi1),
"generic" = calc_prob(x[, 1], 4, N_choose_np1, nip_choose_xi1)
)
Unit: microseconds
expr min lq mean median uq max neval cld
method 4.017 4.3655 4.84964 4.5275 4.776 16.109 100 a
generic 76.810 78.1670 81.29589 80.0300 81.728 165.326 100 bSo the dispatching adds a horrifying 20x time penalty to calc_prob() in this example.
Looks like the callstack can get really long, which makes dispatching using this function very slow. One idea would be to replace the if-else statements with
switch(). Perhaps simplifying thecallstackobject (from a list to a vector, for example) also helps.Example: