Provided code: lab11.zip.

As usual, you have been provided with a file lab11.h with headers for the functions described, lab11c.cpp with analogous C implementations, tests.cpp with reasonably thorough tests of the functions as described.

We're switching to C++ this week so we can use vectorclass.

SIMD Assembly

Now that we know about the x86 vector/SIMD instructions, let's use them with problems we're familiar with.

In lab11.S, copy your dot product and polynomial evaluation implementations from last week: dot_double, dot_single, map_poly_double, map_poly_single.

Write functions that produce the same results, but using the SIMD instructions and %ymm registers: dot_double_vec, dot_single_vec, map_poly_double_vec, map_poly_single_vec. You can assume the array length is divisible by the 4 (for double-precision) or 8 (for single-precision).

You will need the vbroadcastsd and vbroadcastss instructions (which weren't mentioned in lecture) which broadcast one double-/single-precision value to all of the fields in a vector register.

SIMD with Vectorclass

Create a C++ file lab11_vc.cpp. In it, repeat the above, creating C++ implementations dot_double_vc, dot_single_vc, map_poly_double_vc, map_poly_single_vc that use the vectorclass library to implement this logic using SIMD instructions accessed from C++.

Time It

The provided timing.cpp provides some timing tests on reasonably-sized arrays. Have a look. How does your code compare to what the compiler wrote? (Use -O3 to give the compiler its best chance.) ❓

Mini-Project

This lab exercise is a little shorter than usual, to leave some time for you to get some of the Mini-Project done. Do that.

Questions

1. Relative to your assembly code last week, how much did the "dot product" and "map polynomial" implementations speed up when using the vector instructions?
2. On the two problems, what was the relative speedup of vectorized implementations on single-precision floating point values, over double-precision?
3. When timing your assembly (and vectorclass) implementations and the implementations created by the compiler, you likely saw that for the "dot product" problem, the C implementation performed more like the non-vectorized assembly. For the "map polynomial" problem, the C implementation performed more like the vectorized assembly. Why was the compiler able to vectorize one but not the other?

Submit

Submit your work to Lab 11 in CourSys.