=========================================
Fine-tuning the Unrolled Assembler Kernel                                 [TOC]
=========================================

In our previous assembler kernel we stored the pointer to the panel of $A$ in
register `%rax` and the pointer to the $B$ panel in `%rbx`.  Inside the unrolled
loop we have copy and pasted the original loop body four times.  So we also have
copied the pointer increments four times.  These are the lines with

---- CODE(type=c) --------------------------------------------------------------
    "addq      $32,     %%rax    \n\t"  // A += 4;
    "addq      $32,     %%rbx    \n\t"  // B += 4;
--------------------------------------------------------------------------------

We replace that with a single increment, i.e.

---- CODE(type=c) --------------------------------------------------------------
    "addq      $4*32,     %%rax    \n\t"  // A += 16;
    "addq      $4*32,     %%rbx    \n\t"  // B += 16;
--------------------------------------------------------------------------------

at the end of the loop.  As the value of `%rax` remains constant in the loop
pipelining can be improved by the CPU.  But as a consequently we have to use
offsets in the loop, e.g. when we do the second update step we write

---- CODE(type=c) --------------------------------------------------------------
    "movaps  48(%%rbx), %%xmm3   \n\t"  // tmp3     = _mm_load_pd(B+6)
--------------------------------------------------------------------------------

instead of

---- CODE(type=c) --------------------------------------------------------------
    "movaps  16(%%rbx), %%xmm3   \n\t"  // tmp3     = _mm_load_pd(B+6)
--------------------------------------------------------------------------------


Select the demo-sse-asm-unrolled-v2 Branch
==========================================
Again, we do a `make clean` before switching a branch:

    *--[SHELL(height=6)]------------------------------------------------*
    |                                                                   |
    |  cd ulmBLAS                                                       |
    |  make clean                                                       |
    |                                                                   |
    *-------------------------------------------------------------------*

Then we are checking out the `demo-sse-asm-unrolled-v2` branch:

    *--[SHELL(path=ulmBLAS)]--------------------------------------------*
    |                                                                   |
    |  git branch -a                                                    |
    |  git checkout -B demo-sse-asm-unrolled-v2                      +++|
    |                  remotes/origin/demo-sse-asm-unrolled-v2          |
    |                                                                   |
    *-------------------------------------------------------------------*

Then we compile the project

    *--[SHELL(path=ulmBLAS,height=15)]----------------------------------*
    |                                                                   |
    |  make                                                             |
    |                                                                   |
    *-------------------------------------------------------------------*


The dgemm_nn Code
=================
:import: ulmBLAS/src/level3/dgemm_nn.c [linenumbers]


Benchmark Results
=================
We run the benchmarks

    *--[SHELL(path=ulmBLAS)]--------------------------------------------*
    |                                                                   |
    |  cd bench                                                         |
    |  ./xdl3blastst > report                                           |
    |                                                                   |
    *-------------------------------------------------------------------*

and filter out the results for the `demo-sse-asm-unrolled-v2`
branch:

    *--[SHELL(path=ulmBLAS/bench)]--------------------------------------*
    |                                                                   |
    |  grep PASS report > demo-sse-asm-unrolled-v2                      |
    |                                                                   |
    *-------------------------------------------------------------------*

With the gnuplot script

:import: ulmBLAS/bench/bench11.gps

we feed gnuplot

    *--[SHELL(path=ulmBLAS/bench)]--------------------------------------*
    |                                                                   |
    |  gnuplot bench11.gps                                              |
    |                                                                   |
    *-------------------------------------------------------------------*

and get

    ---- IMAGE --------------
    ulmBLAS/bench/bench11.svg
    -------------------------

:navigate: __up__    -> doc:index
           __back__  -> doc:page09/index
           __next__  -> doc:page11/index