SSE & AVX Vectorization

Automatic Vectorization

GCC Autovectorization flags

GCC is an advanced compiler, and with the optimization flags -O3 or -ftree-vectorize the compiler will search for loop vectorizations (remember to specify the -mavx flag too). The source code remains the same, but the compiled code by GCC is completely different.

GCC won't log anything about automatic vectorization unless some flags are enabled. If you need details of autovectorization results you can use the compiler flags:

• -fopt-info-vec or -fopt-info-vec-optimized: The compiler will log which loops (by line N°) are being vector optimized.
• -fopt-info-vec-missed: Detailed info about loops not being vectorized, and a lot of other detailed information.
• -fopt-info-vec-note: Detailed info about all loops and optimizations being done.
• -fopt-info-vec-all: All previous options together.

NOTE: There are similar -fopt-info-[options]-optimized flags for other compiler optimizations, like inline: -fopt-info-inline-optimized

In this example we will use GCC Autovectorization, enabled with -O3 and with verbose mode flag -fopt-info-vec-optimized. You can change the compiler flags to see the different logging options:

If everything is correct, you should see the compiler Test Results:

/project/target/autovector/autovector.cpp:22:21: note: loop vectorized
/project/target/autovector/autovector.cpp:15:23: note: loop vectorized

With the same code you can test two different things:

• In the compilerflags.txt file, change the log options to -fopt-info-vec-all and rerun the Example.
• On line 1 of autovector.cpp, change optimize("O3", to optimize("O2", and rerun. Non-vectorized compilations will be slower than vectorized ones.

Criteria for loop vectorization

Not all loops can be vectorized. In order for vectorization to be performed, there are some strict requirements for the loop.

• The loop count can't change once the loop starts. That means that the end of the loop can be a dynamic variable, increasing or decreasing its value at will, but once the loop starts, it must be constant.
• There are limitations to using break or continue sentences. Sometimes the compiler is clever enough to make it work, but in some cases the loop won't be vectorized. Try to add the line if (j==27) continue; right before the result[j] assignment on line 25, and rerun.
• There are some limits on calling external functions inside a loop.
• There shouldn't be data dependencies with other indexes of the loop. I.E. the loop: for (int i=1; i<N; ++i) x[i]=x[i-1]*2; is traversed with a variable i, and data x[i] depends on the previous x[i-1] value. Since AVX registers are loaded as 8 floats, the compiler can't do these calculations with a vector.
• Conditionals sentences (if/else) can be used if they don't change the control flow, and are only used to conditionally load A or B values into a C variable. Selecting A or B is done with a mask in the compiler, so it calculates both branches A and B, and C will store either one value or another:

 if ( s >= 0 ) {
   x[i] = (-b[i]+s)/(2.0f*a[i]);
   y[i] = (-b[i]-s)/(2.0f*a[i]);
 }
 else {
   x[i] = 0.0f;
   y[i] = 0.0f;
 }

This is a vectorizable loop. The control flow never changes, and the x[i] and y[i] values are always set to either one or the other value.

Read Vectorization with Intel C++ Compilers for additional info on autovectorization. This document is aimed at Intel Compilers, but it gives interesting and complete information about autovectorization. The GCC compiler has one webpage with some information on Auto-vectorization in GCC but it's pretty outdated.

The good thing about autovectorization is that it comes for free. The developer doesn't need to change anything, and maybe the loop will be vectorized. But sometimes (especially in high performance computing applications) loops and vectorization need to be fine tuned, ensuring maximum throughput by using manual AVX vectorization. We will see that in the next lessons.