OpenACC Tutorial - Profiling

Learning objectives

• Understand what a profiler is
• Understand how to use the NVPROF profiler
• Understand how the code is performing
• Understand where to focus your time and rewrite most time consuming routines

Code profiling

Why would one need to profile code? Because it's the only way to understand:

• Where time is being spent (hotspots)
• How the code is performing
• Where to focus your development time

What is so important about hotspots in the code? Amdahl's law says that "Parallelizing the most time-consuming routines (i.e. the hotspots) will have the most impact".

Build the Sample Code

For this example we use code from this Git repository. Download the package and go to the cpp or the f90 directory. The object of this exercise is to compile and link the code, obtain an executable, and then profile it.

Which compiler ?

Being pushed by NVIDIA through its Portland Group division until 2020 and now through its HPC SDK, as well as by Cray, these two lines of compilers offer the most advanced OpenACC support.

As for the GNU Compiler, since GCC version 6, the support for OpenACC 2.x kept improving. As of July 2022, GCC versions 10, 11 and 12 support OpenACC version 2.6.

For the purpose of this tutorial, we use version 22.7 of the NVIDIA HPC SDK. We note that NVIDIA compilers are free for academic usage.

[name@server ~]$ module load nvhpc/22.7
Lmod is automatically replacing "intel/2020.1.217" with "nvhpc/22.7".

The following have been reloaded with a version change:
  1) gcccore/.9.3.0 => gcccore/.11.3.0        3) openmpi/4.0.3 => openmpi/4.1.4
  2) libfabric/1.10.1 => libfabric/1.15.1     4) ucx/1.8.0 => ucx/1.12.1

[name@server ~]$ make 
nvc++    -c -o main.o main.cpp
nvc++ main.o -o cg.x

After the executable is created, we are going to profile that code. Important: this executable uses about 3GB of memory and one CPU core at near 100%. Therefore, a proper test environment should have at least 4GB of available memory and at least two (2) CPU cores.

Which profiler ?

For the purpose of this tutorial, we use two profilers:

• NVPROF - a command line text-based profiler that can analyze non-GPU codes.
• NVIDIA Visual Profiler NVVP - a graphical cross-platform analyzing tool for the codes written with OpenACC and CUDA C/C++ instructions.

Since our previously built cg.x is not yet using the GPU, we will start the analysis with the nvprof profiler.

NVIDIA NVPROF Command Line Profiler

NVIDIA also provides a command line version called NVPROF, similar to GPU prof

[name@server ~]$ module load cuda/11.7

[name@server ~]$ nvprof --cpu-profiling on ./cg.x 
...
<Program output >
...
======== CPU profiling result (bottom up):
Time(%)      Time  Name
 83.54%  90.6757s  matvec(matrix const &, vector const &, vector const &)
 83.54%  90.6757s  | main
  7.94%  8.62146s  waxpby(double, vector const &, double, vector const &, vector const &)
  7.94%  8.62146s  | main
  5.86%  6.36584s  dot(vector const &, vector const &)
  5.86%  6.36584s  | main
  2.47%  2.67666s  allocate_3d_poisson_matrix(matrix&, int)
  2.47%  2.67666s  | main
  0.13%  140.35ms  initialize_vector(vector&, double)
  0.13%  140.35ms  | main
...
======== Data collected at 100Hz frequency

NVIDIA Visual Profiler - (to be moved to another page)

NVVP profiler

Browse for the executable you want to profile

One graphical profiler available for OpenACC applications is the NVIDIA Visual Profiler (NVVP). It's a cross-platform analyzing tool for codes written with OpenACC and CUDA C/C++ instructions. Consequently, if the executable is not using the GPU, you will get no result from this profiler.

When X11 is forwarded to an X-Server, or when using a Linux desktop environment (also via JupyterHub with two (2) CPU cores, 5000M of memory and one (1) GPU), it is possible to launch the NVVP from a terminal:

[name@server ~]$ module load cuda/11.7 java/1.8

[name@server ~]$ nvvp

1. After the NVVP startup window, you get prompted for a Workspace directory, which will be used for temporary files. Replace home with scratch in the suggested path. Then click OK.
2. Select File > New Session, or click on the corresponding button in the toolbar.
3. Click on the Browse button at the right of the File path editor.
   1. Change directory if needed.
   2. Select an executable built from codes written with OpenACC and CUDA C/C++ instructions.
4. Below the Arguments editor, select the profiling option Profile current process only.
5. Click Next > to review additional profiling options.
6. Click Finish to start profiling the executable.

Compiler Feedback

Before working on the routine, we need to understand what the compiler is actually doing by asking ourselves the following questions:

• What optimizations were applied?
• What prevented further optimizations?
• Can very minor modifications of the code affect performance?

The PGI compiler offers you a -Minfo flag with the following options:

• accel – Print compiler operations related to the accelerator
• all – Print all compiler output
• intensity – Print loop intensity information
• ccff–Add information to the object files for use by tools

How to Enable Compiler Feedback

• Edit the Makefile

 CXX=nvc++
 CXXFLAGS=-fast -Minfo=all,intensity,ccff
 LDFLAGS=${CXXFLAGS}

• Rebuild

[name@server ~]$ make clean; make
nvc++ -fast -Minfo=all,intensity,ccff   -c -o main.o main.cpp
initialize_vector(vector &, double):
     20, include "vector.h"
          36, Intensity = 0.0
              Memory set idiom, loop replaced by call to __c_mset8
dot(const vector &, const vector &):
     21, include "vector_functions.h"
          27, Intensity = 1.00
              Generated vector simd code for the loop containing reductions
              FMA (fused multiply-add) instruction(s) generated
waxpby(double, const vector &, double, const vector &, const vector &):
     21, include "vector_functions.h"
          39, Intensity = 1.00
              Loop not vectorized: data dependency
              Generated vector simd code for the loop
              Loop unrolled 2 times
              FMA (fused multiply-add) instruction(s) generated
allocate_3d_poisson_matrix(matrix &, int):
     22, include "matrix.h"
          43, Intensity = 0.0
              Loop not fused: different loop trip count
          44, Intensity = 0.0
              Loop not vectorized/parallelized: loop count too small
          45, Intensity = 0.0
          57, Intensity = 0.0
          59, Intensity = 0.0
              Loop not vectorized: data dependency
matvec(const matrix &, const vector &, const vector &):
     23, include "matrix_functions.h"
          29, Intensity = (num_rows*((row_end-row_start)*         2))/(num_rows+(num_rows+(num_rows+((row_end-row_start)+(row_end-row_start)))))
              FMA (fused multiply-add) instruction(s) generated
          33, Intensity = 1.00
              Loop not vectorized: non-stride-1 array reference
              Loop not vectorized: mixed data types
              Loop unrolled 2 times
              FMA (fused multiply-add) instruction(s) generated
main:
     38, allocate_3d_poisson_matrix(matrix &, int) inlined, size=41 (inline) file main.cpp (29)
          43, Intensity = 0.0
              Loop not fused: different loop trip count
          44, Intensity = 0.0
              Loop not vectorized/parallelized: loop count too small
          45, Intensity = 0.0
          57, Intensity = 0.0
              Loop not fused: function call before adjacent loop
          59, Intensity = 0.0
              Loop not vectorized: data dependency
     42, allocate_vector(vector &, unsigned int) inlined, size=3 (inline) file main.cpp (24)
     43, allocate_vector(vector &, unsigned int) inlined, size=3 (inline) file main.cpp (24)
     44, allocate_vector(vector &, unsigned int) inlined, size=3 (inline) file main.cpp (24)
     45, allocate_vector(vector &, unsigned int) inlined, size=3 (inline) file main.cpp (24)
     46, allocate_vector(vector &, unsigned int) inlined, size=3 (inline) file main.cpp (24)
     48, initialize_vector(vector &, double) inlined, size=5 (inline) file main.cpp (34)
          36, Intensity = 0.0
              Loop not vectorized/parallelized: not countable
     49, initialize_vector(vector &, double) inlined, size=5 (inline) file main.cpp (34)
          36, Intensity = 0.0
              Loop not vectorized/parallelized: not countable
     52, waxpby(double, const vector &, double, const vector &, const vector &) inlined, size=10 (inline) file main.cpp (33)
          39, Intensity = 0.0
              Memory copy idiom, loop replaced by call to __c_mcopy8
     53, matvec(const matrix &, const vector &, const vector &) inlined, size=19 (inline) file main.cpp (20)
          29, Intensity = [symbolic], and not printable, try the -Mpfi -Mpfo options
              Loop not fused: different loop trip count
          33, Intensity = 1.00
              Loop not vectorized: non-stride-1 array reference
              Loop not vectorized: mixed data types
              Loop unrolled 2 times
     54, waxpby(double, const vector &, double, const vector &, const vector &) inlined, size=10 (inline) file main.cpp (33)
          27, FMA (fused multiply-add) instruction(s) generated
          29, FMA (fused multiply-add) instruction(s) generated
          33, FMA (fused multiply-add) instruction(s) generated
          39, Intensity = 0.67
              Loop not fused: different loop trip count
              Loop not vectorized: data dependency
              Generated vector simd code for the loop
              Loop unrolled 4 times
              FMA (fused multiply-add) instruction(s) generated
     56, dot(const vector &, const vector &) inlined, size=9 (inline) file main.cpp (21)
          27, Intensity = 1.00
              Loop not fused: function call before adjacent loop
              Generated vector simd code for the loop containing reductions
     61, Intensity = 0.0
     62, waxpby(double, const vector &, double, const vector &, const vector &) inlined, size=10 (inline) file main.cpp (33)
          39, Intensity = 0.0
              Memory copy idiom, loop replaced by call to __c_mcopy8
     65, dot(const vector &, const vector &) inlined, size=9 (inline) file main.cpp (21)
          27, Intensity = 1.00
              Loop not fused: different loop trip count
              Generated vector simd code for the loop containing reductions
     67, waxpby(double, const vector &, double, const vector &, const vector &) inlined, size=10 (inline) file main.cpp (33)
          39, Intensity = 0.67
              Loop not fused: different loop trip count
              Loop not vectorized: data dependency
              Generated vector simd code for the loop
              Loop unrolled 4 times
     72, matvec(const matrix &, const vector &, const vector &) inlined, size=19 (inline) file main.cpp (20)
          29, Intensity = [symbolic], and not printable, try the -Mpfi -Mpfo options
              Loop not fused: different loop trip count
          33, Intensity = 1.00
              Loop not vectorized: non-stride-1 array reference
              Loop not vectorized: mixed data types
              Loop unrolled 2 times
     73, dot(const vector &, const vector &) inlined, size=9 (inline) file main.cpp (21)
          27, Intensity = 1.00
              Loop not fused: different loop trip count
              Generated vector simd code for the loop containing reductions
     77, waxpby(double, const vector &, double, const vector &, const vector &) inlined, size=10 (inline) file main.cpp (33)
          39, Intensity = 0.67
              Loop not fused: different loop trip count
              Loop not vectorized: data dependency
              Generated vector simd code for the loop
              Loop unrolled 4 times
     78, waxpby(double, const vector &, double, const vector &, const vector &) inlined, size=10 (inline) file main.cpp (33)
          39, Intensity = 0.67
              Loop not fused: function call before adjacent loop
              Loop not vectorized: data dependency
              Generated vector simd code for the loop
              Loop unrolled 4 times
     88, free_vector(vector &) inlined, size=2 (inline) file main.cpp (29)
     89, free_vector(vector &) inlined, size=2 (inline) file main.cpp (29)
     90, free_vector(vector &) inlined, size=2 (inline) file main.cpp (29)
     91, free_vector(vector &) inlined, size=2 (inline) file main.cpp (29)
     92, free_matrix(matrix &) inlined, size=5 (inline) file main.cpp (73)
nvc++ main.o -o cg.x -fast -Minfo=all,intensity,ccff

Computational Intensity

Computational Intensity of a loop is a measure of how much work is being done compared to memory operations.

Computation Intensity = Compute Operations / Memory Operations

Computational Intensity of 1.0 or greater suggests that the loop might run well on a GPU.

Understanding the code

Let's look closely at the following code from matrix_functions.h:

for(int i=0;i<num_rows;i++) {
  double sum=0;
  int row_start=row_offsets[i];
  int row_end=row_offsets[i+1];
  for(int j=row_start; j<row_end;j++) {
    unsigned int Acol=cols[j];
    double Acoef=Acoefs[j]; 
    double xcoef=xcoefs[Acol]; 
    sum+=Acoef*xcoef;
  }
  ycoefs[i]=sum;
}

Given the code above, we search for data dependencies:

• Does one loop iteration affect other loop iterations?
• Do loop iterations read from and write to different places in the same array?
• Is sum a data dependency? No, it’s a reduction.

Now that the code analysis is done, we are ready to add directives to the compiler.

<- Previous unit: Introduction | ^- Back to the lesson plan | Onward to the next unit: Adding directives ->

GNU Compiler Collection, an open source compiler collection