Re: [PATCH v2 3/5] perf report: Sort by sampled cycles percent per block for stdio

From: Jin, Yao
Date: Wed Oct 16 2019 - 06:51:12 EST




On 10/16/2019 6:15 PM, Jiri Olsa wrote:
On Tue, Oct 15, 2019 at 10:53:18PM +0800, Jin, Yao wrote:

SNIP

+static struct block_header_column{
+ const char *name;
+ int width;
+} block_columns[PERF_HPP_REPORT__BLOCK_MAX_INDEX] = {
+ [PERF_HPP_REPORT__BLOCK_TOTAL_CYCLES_COV] = {
+ .name = "Sampled Cycles%",
+ .width = 15,
+ },
+ [PERF_HPP_REPORT__BLOCK_LBR_CYCLES] = {
+ .name = "Sampled Cycles",
+ .width = 14,
+ },
+ [PERF_HPP_REPORT__BLOCK_CYCLES_PCT] = {
+ .name = "Avg Cycles%",
+ .width = 11,
+ },
+ [PERF_HPP_REPORT__BLOCK_AVG_CYCLES] = {
+ .name = "Avg Cycles",
+ .width = 10,
+ },
+ [PERF_HPP_REPORT__BLOCK_RANGE] = {
+ .name = "[Program Block Range]",
+ .width = 70,
+ },
+ [PERF_HPP_REPORT__BLOCK_DSO] = {
+ .name = "Shared Object",
+ .width = 20,
+ }
};

so we already have support for multiple columns,
why don't you add those as 'struct sort_entry' objects?


For 'struct sort_entry' objects, do you mean I should reuse the "sort_dso"
which has been implemented yet in util/sort.c?

For other columns, it looks we can't reuse the existing sort_entry objects.

I did not mean reuse, just add new sort entries
to current sort framework


Does it seem like what the c2c does?

For example, my new update is like:

struct block_dimension {
const char *header;
int idx;
int width;
struct sort_entry *se;
int64_t (*cmp)(struct perf_hpp_fmt *,
struct hist_entry *, struct hist_entry *);
int64_t (*sort)(struct perf_hpp_fmt *,
struct hist_entry *, struct hist_entry *);
int (*entry)(struct perf_hpp_fmt *, struct perf_hpp *,
struct hist_entry *);
};

struct block_fmt {
struct perf_hpp_fmt fmt;
struct report *rep;
struct block_dimension *dim;
};

static struct block_dimension dim_total_cycles_pct {
.header = "Sampled Cycles%",
.idx = PERF_HPP_REPORT__BLOCK_TOTAL_CYCLES_PCT,
.width = 15,
.cmp = block_info__cmp;
.sort = block_cycles_cov_sort;
.entry = block_cycles_cov_entry;
};

......

Is above new update correct?


SNIP

+{
+ struct block_hist *bh = &rep->block_hist;
+
+ get_block_hists(hists, bh, rep);
+ symbol_conf.report_individual_block = true;
+ hists__fprintf(&bh->block_hists, true, 0, 0, 0,
+ stdout, true);
+ hists__delete_entries(&bh->block_hists);
+ return 0;
+}
+
static int perf_evlist__tty_browse_hists(struct evlist *evlist,
struct report *rep,
const char *help)
@@ -500,6 +900,12 @@ static int perf_evlist__tty_browse_hists(struct evlist *evlist,
continue;
hists__fprintf_nr_sample_events(hists, rep, evname, stdout);
+
+ if (rep->total_cycles) {
+ hists__fprintf_all_blocks(hists, rep);

so this call kicks all the block info setup/count/print, right?


Yes, all in this call.

I thingk it shouldn't be in the output code, but in the code before..
from what I see you could count block_info counts during the sample
processing, no?


In sample processing, we just get all symbols and account the cycles per
symbol. We need to create/count the block_info at some points after the
sample processing.

understand, but it needs to be outside display function


OK, I will move the code for block_info collection outside of display function.

also, can't you gather the block_info data gradually
during the sample processing?


Looks we have to gather the block_info after the sample processing. That's because in each sample processing, we will call hist__account_cycles(). Then finally __symbol__account_cycles() gets called for accounting one basic block in this symbol.

ch[offset].num_aggr++;
ch[offset].cycles_aggr += cycles;

So actually, after the sample processing, we will get the num_aggr and cycles_aggr for this basic block and compute the average cycles for this block (cycles_aggr / num_aggr).

That's why I think we should gather the block_info after the sample processing.

Thanks
Jin Yao

jirka


Maybe it's not very good to put block info setup/count/print in a call, but
it's really not easy to process the block_info during the sample processing.

Thanks
Jin Yao

jirka