Intel® VTune™ Profiler User Guide

User Guide

Contents

View Comparison Data

Intel® VTune™
Profiler
 compares analysis results and displays difference in a separate result tab
<result1>-<result2>
 in the following windows:
  • Summary
     window provides top-level difference for the analysis run.
  • Bottom-up
     window displays difference for functions and their callers per metric.
  • Top-down Tree
     window displays difference for functions and their callees per metric.
  • Caller/Callee
     window displays difference for a selected function, their callers and callees per metric.

Comparing Recompiled Binary Files

By default, the
VTune
Profiler
 displays compared functions grouped by the
Call Stack
 granularity, which is based on function instances. But you may want to switch to the
Source Function Stack
 grouping to get more accurate comparison results in the following cases:
  • You slightly changed the source and recompiled the code.
  • You changed compilation options and recompiled the code.
  • You are comparing results compiled and collected for different Intel microarchitectures
For example, your binary with a
my_f
function was modified with adding a new function
my_f1
 and new calls of this function. As a result,
my_f
 address has changed. If you compare the results before and after the modification using the default
Call Stack
 grouping, the
VTune
Profiler
 treats the same functions with different addresses as separate instances and does not compare them:
When the data is aggregated by
Source Function Stack
, the
VTune
Profiler
 ignores start addresses and compares functions by source file objects:

Bar Data Respresentation

If you chose the
Bar
 format to display the performance data in the
Bottom-up
 or
Top-down Tree
 window, the
VTune
Profiler
 calculates the bar size as follows:
Result Data Column
Difference Column
cell_data_value/absolute_max_value_in_result_column
cell_data_value/max(absolute_max_value_in__1st_result_column, absolute_max_value_in_corresponding_2nd_result_column)
Example: Calculation of the Bar Size
The table below provides an example on how the
VTune
Profiler
 calculates the bar size in the compare mode based on the absolute max CPU time value and performance data per column:
CPU Time:r001
CPU Time:r002
CPU Time:Difference
Absolute max value (calculated by the
VTune
Profiler
 internally but not exposed in the grid)
10s
20s
20s
Performance data
1s
3s
2s
Bar size
1s/max(10s,20s)
3s/max(10s,20s)
2s/max(10s,20s)

Product and Performance Information

1

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.