FLAME Central

Code generation related links

Spark: FLAME code-skeleton generator

Operation to BLAS translator

Downloads

PFLAME

Note: For those who have accounts on the CS machines at UT-Austin, we suggest you link to the libraries in /u/rvdg/PFLAME rather than building your own.

• FLAME@lab (for MATLAB and Octave),
• FLAME@C, and
• PLAPACK

To set this up on a Linux machine:

• gunzip PFLAME.tar.gz
• tar -xf PFLAME.tar
  (this will create a subdirectory PFLAME)
• cd PFLAME
• You will have to edit the Make.machine files and fix some of the Make.include files in the subdirectories.
• in subdirectory PFLAME type
  • make FLAME
    to make the FLAME library
  • make PLAPACK
    to make the PLAPACK library
• You may wish to download BLAS libraries from http://www.cs.utexas.edu/users/kgoto/

FLAME WebSite

Note: You probably want to just use the FLAME Website itself at http://www.cs.utexas.edu/users/flame/

To set this up on a Linux machine:

• gunzip flame.tar.gz
• tar -xf flame.tar
  (this will create a subdirectory flame)

Examples

FLaTeX

• First, download the file flatex.tex into the same directory where the file is with the worksheet.
• A reference for the simple example discussed below is can be downloaded:
  • sum.pdf (PDF)
  • sum.ps (postscript)
  • Source:
    • sum.tex
    • flatex.tex
    • blank_ws.tex
    • sum_ws.tex
• To create a worksheet for this example, which describes how to add the elements of a vector,
  • Go to the Spark WebPage.
  • Fill out the menu on the left:
    • Name of the function: Sum
    • Type of function: unblocked
    • Variant Name: 1
    • Number of operands: 2
    • (click on "Generate Code and/or Update Form")
    • Properties of operands:

      1:	A	scalar	none	both
      2:	X	vector	T->B	input

    • Output language: FLaTeX
    • (click on "Generate Code and/or Update Form")
  • Cut an paste the contents in the frame on the right into a file, say sum_unb_var1.tex.
  • In the directory where sum_unb_var1.tex exists, type
    • latex sum_unb_var1.tex
latex sum_unb_var1.tex
      (twice!) to format the document
    • xdvi sum_unb_var1.dvi
      to view the result
    • dvips -f sum_unb_var1.dvi -t letter > sum_unb_var1.ps
      to create postscript, and/or
    • dvipdf sum_unb_var1.dvi
      to create PDF (in file sum_unb_var1.pdf )
  • Modify the file sum_unb_var1.tex to create the worksheet as desired.

FLAME@lab

• If you are not on the UT-Austin CS machines, download and install PFLAME.
  Notice that in order to use FLAME@lab, you need not change Makefile, Make.machine, or Make.include files.
• Create a directory for your project, e.g. named MYFLAME , and a subdirectory, e.g. named MYFLAME/FLAME@lab.
• Download Trsm_blk_var1.m into a file with that name.
• If you will use
  • Octave: (skip to MATLAB)

    octave --traditional
    

    -- stuff is printed --
    Cut and paste the following. Note that anything that starts with "%" is interpreted as a comment, so you can cut and paste it as well, or skip it.

    Set the path to where the basic FLAME@lab functions exist:

    path( path, "/u/rvdg/PFLAME/FLAME@lab/Octave" )
    

    Naturally, if you downloaded your own copy of PFLAME, you would want to adjust the path to where you installed PFLAME.

    % Create a 5x4 random matrix B
    
    B = rand( 5, 4 )
    
    % Create a 5x5 random matrix L, make it lower triangular and
    % add something big to the diagonal to ensure it is not almost singular
    
    L = tril( rand( 5, 5 ) ) + 5 * eye( 5 )
    
    % Use the FLAME@lab code in Trsm_blk_var1.m to solve L X = B
    % Here the "2" is the block size chosen for the blocked algorithm
    
    X = Trsm_blk_var1( L, B, 2 )
    
    % Compute the same, but using Mscript intrinsic functions
    
    Xref = L \ B
    
    % Compare X and Xref
    
    X - Xref
    

    Notice that the result is something like

    1.0e-17 *
     
    -- some matrix printed here --
    

    which is very small, due to the fact that 1.0e-17 multiplies all entries of the matrix.

• MATLAB:

  matlab
  

  -- stuff is printed -- and/or a new window pops up
  Cut and paste the following. Note that anything that starts with "%" is interpreted as a comment, so you can cut and paste it as well, or skip it.

  Set the path to where the basic FLAME@lab functions exist:

  path( path, '/u/rvdg/PFLAME/FLAME@lab/MATLAB' )
  

  Naturally, if you downloaded your own copy of PFLAME, you would want to adjust the path to where you installed PFLAME.

  % Create a 5x4 random matrix B
  
  B = rand( 5, 4 )
  
  % Create a 5x5 random matrix L, make it lower triangular and
  % add something big to the diagonal to ensure it is not almost singular
  
  L = tril( rand( 5, 5 ) ) + 5 * eye( 5 )
  
  % Use the FLAME@lab code in Trsm_blk_var1.m to solve L X = B
  % Here the "2" is the block size chosen for the blocked algorithm
  
  X = Trsm_blk_var1( L, B, 2 )
  
  % Compute the same, but using MATLAB intrinsic functions
  
  Xref = L \ B
  
  % Compare X and Xref
  
  X - Xref
  

  Notice that the result is something like

  1.0e-17 *
   
  -- some matrix printed here --
  

  which is very small, due to the fact that 1.0e-17 multiplies all entries of the matrix.

FLAMEC

• If you are not on the UTCS workstations, install PFLAME.
• Create a directory for your project, e.g. named MYFLAME , and a subdirectory, e.g. named MYFLAME/FLAMEC.
• Download Trsm_example.tar.gz into a file with that name.
• gunzip and untar the file:

  > gunzip Trsm_example.tar.gz
  > tar -xf Trsm_example.tar
  

• Enter the directory:

  > cd Trsm_example
  

• If you are not on the UTCS workstations, modify the makefile: Makefile, changing the path in the first line to a path to where you installed PFLAME.
• Make and execute:

  > make
  
  ---- Lots of output ----
  
  > ./test_llnn.x
  % number of repeats:% 3
  % Enter blocking size:% 104
  % enter nfirst, nlast, ninc:% 100 400 100
  % enter m: (-1 means m=n)% -1
  data_REF( 1, 1:2 ) = [ 100  1181.22 ]; 
  data_var1( 1, 1:7 ) = [ 100  861.24 3.47e-18 830.54 3.47e-18 1130.66 0.00e+00  ]; 
  data_var2( 1, 1:7 ) = [ 100  425.85 1.21e-17 435.38 1.21e-17 1129.89 0.00e+00  ]; 
  
     ----------  Lots of output ---------
  data_var4( 4, 1:7 ) = [ 400  409.15 8.67e-19 1572.46 0.00e+00 1576.98 0.00e+00  ]; 
  
  close all
  plot( data_REF( :,1 ), data_REF( :, 2 ), '-' ); 
  hold on
  plot( data_var1( :,1 ), data_var1( :, 2 ), 'b:o' ); 
  plot( data_var1( :,1 ), data_var1( :, 4 ), 'b-.o' ); 
  plot( data_var1( :,1 ), data_var1( :, 6 ), 'b--o' ); 
  plot( data_var2( :,1 ), data_var2( :, 2 ), 'r:+' ); 
  plot( data_var2( :,1 ), data_var2( :, 4 ), 'r-.+' ); 
  plot( data_var2( :,1 ), data_var2( :, 6 ), 'r--+' ); 
  plot( data_var3( :,1 ), data_var3( :, 2 ), 'k:*' ); 
  plot( data_var3( :,1 ), data_var3( :, 4 ), 'k-.*' ); 
  plot( data_var3( :,1 ), data_var3( :, 6 ), 'k--*' ); 
  plot( data_var4( :,1 ), data_var4( :, 2 ), 'g:x' ); 
  plot( data_var4( :,1 ), data_var4( :, 4 ), 'g-.x' ); 
  plot( data_var4( :,1 ), data_var4( :, 6 ), 'g--x' ); 
  legend( 'Reference', 'unb\_var1', 'blk\_var1', 'opt\_var1', ... 
  'unb\_var2', 'blk\_var2', 'opt\_var2', ... 
  'unb\_var3', 'blk\_var3', 'opt\_var3', ... 
  'unb\_var4', 'blk\_var4', 'opt\_var4', 2 ); 
  xlabel( 'matrix dimension n' );
  ylabel( 'MFLOPS/sec.' );
  axis( [ 0 400 0 3600 ] ); 
  print -depsc graphs.eps
  hold off
  

• The reason for all the output is that you can run the output through Matlab to create a graph. Try

  echo "4 104 100 1000 100 -1 " | ./test_llnn.x > output
  matlab < output
  gv graph.eps
  

• Notice: There is a small change you will want to make to the file output: In the line axis( [ 0 400 0 3600 ] ); change the 3600 to the peak (in MFLOPS/sec) of the machine on which you ran the experiment. You can get this information from the file /proc/cpuinfo. by multiplying the cpu MHz number by the number of floating point operations that your architecture can do per clock cycle (e.g., 2 for a Pentium 4. Multiply by 4 if you have a 2 CPU Pentium 4 based SMP and you are linking to a multi-threaded BLAS library. Etc.).