The Intel R @Software Autotuning Tool (ISAT) User Manual (version
Contents
1. with the default name isat report txt in the top level of the output directory From this report you can find out the performance of all trial runs 10 1 Plotting the Tuning Results You can plot the results in isat report txt using the xgraph tool in the ISAT package as follows isat Source Python xgraph py i isat report txt o graphs The graphs generated are put in the directory graphs A 2D graph like Figure 1 a will be generated for an independently tuned variable A 3D temperature graph like Figure 1 b will be generated for a pair of dependent variables if there are more than two dependent variables a 3D graph will be generated for each distinct pair of dependent variables 11 ISAT Command line Switches You can do isat h to see all the command line switches of ISAT Table 2 lists these switches and their meanings Switch Description i Input directory that contains the untuned version o Output directory that contains the tuned version nt Number of test runs per code variant default 1 rp Tuning report name default isat report txt mainfile Name of the file that contains main clean_command The command that ISAT calls to clean old executables default isat clean build_command The command that ISAT calls to build the executable default isat build test_command The command that ISAT calls to run and time the executable default isat t2. In order to use ISAT to tune your program you need to have its source code You also need to add three scripts to the directory where the source is located isat clean isat build isat test you can change the names of these scripts using the isat command line see Section 11 for the switch names isat clean will be invoked by ISAT to clean up the build environment before the build isat build will be invoked by ISAT to build the executable from the source isat test will be invoked by ISAT to run the executable and measure its performance In many cases each of these scripts contains simply an one line command like make clean make build or make run Examples of these scripts can be found in isat Examples BlockedMatrixMultiply Src 6 How to Invoke ISAT The simplest way to invoke ISAT is as follows isat Source Python isat py i Src o Dst where Src is the directory that contains the source files of the program that you want to tune The tuned version will be generated in Dst Please look at the examples in isat Examples The first example to look at should be BlockedMatrixMultiply Follow the directions in BlockedMatrixMultiply README file to learn how to apply ISAT to this example 7 ISAT Pragmas New pragmas are introduced by ISAT to allow a programmer to specify where in the program needs tuning and how the tuning should be done These pragmas are of the form pragma isat There are two types of ISAT
3. The Intel Software Autotuning Tool ISAT User Manual version 1 0 0 Chi Keung Luk Technology Pathfinding and Innovation DPD SSG Intel Corporation November 30 2010 1 Introduction Autotuning is 3 2 1 is an approach for producing efficient and portable codes It works by generating many different variants of the same code and then empirically finding the best performing variant on the target machine The Intel Software Autotuning Tool ISAT is a tool being developed to facilitate autotun ing Version 1 0 supports program parameter tuning of C C programs Future versions would support additional kinds of tuning This manual describes how to use this tool 2 System Requirements ISAT currently runs on Linux both 32 and 64 bit and it requires the following packages pre installed 1 Python version gt 2 4 but lt 3 0 2 Gnuplot optional only needed if you want to plot your tuning results 3 Compiler Compatibility ISAT is implemented in Python using source to source translation As a result ISAT can be used on top of any native C C compilers 4 How to Install ISAT After your download the ISAT tarball isat tar gz simply untar it into a directory of your choice tar xzvf isat tar gz Then invoke config sh in the top level directory and follow the directions as prompted In the rest of this manual we assume that ISAT is installed right under your home directory at isat 5 Preparing Your Program for ISAT
4. ading library Specific Tuning ISAT has built in support to tune certain parameters in the OpenMP and TBB APIs as described below 9 1 OpenMP ISAT supports the tuning of the scheduling parameters in OpenMP parallel loops parallel do and parallel for The parameters are the scheduling type and the chunk size which can be tuned inde pendently or dependently These parameters can be explicitly or implicitly defined in parallel statements When they are implicitly defined you can refer them as omp_schedule_type and omp_schedule_chunk in a ISAT tuning pragma See isat Examples Omp Src for an example 9 2 TBB ISAT supports the tuning of the grain size in TBB parallel_for parallel_reduce and parallel _scan The grain size is the last parameter in a blocked_range It can be explicitly or implicitly defined When it is implicitly defined you can refer it as tbb_grain_size in a ISAT tuning pragma See isat Examples Tbb Src for an example 10 Using the Output of ISAT ISAT puts the tuning results into the output directory that you specified via the o option when invoking ISAT The output directory is essentially a copy of the input directory the one you specified via the i option except that source files in the output directory use the optimal values of all parameters tuned by ISAT So if you simply do a build in the output directory the resulting executable will use the tuned values ISAT also generates a tuning report
5. est ignore_buggy_testrun If a particular code variant results into errors ignore it and continue trying other code variants no_interactive Proceed the tuning without waiting for user s response Table 2 ISAT command line switches Followings are more explanations on when you may need to use some of these switches Tuned region tr2 Tuned region tr0 2 sec PRR ROMA mNmi briwsion Time sec Time sec BRE goon wutah niwa tno x_blksize oO 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 y_blksize view 60 0000 30 0000 scale 1 00000 1 00000 a 2D tuning graph b 3D tuning graph Figure 1 Examples of plotting ISAT results mainfile ISAT needs to know which source file contains the main function It will try to figure this out itself by looking for the unique file that contains main in the given source directory However if there are multiple files that contains main only one of them is actually linked in ISAT needs you to specify the main file ignore_buggy_testrun It is possible that some values tried for a particular variable being tuned may be invalid for that variable resulting into errors like segmentation faults By default ISAT will stop tuning when such errors occur However you can use the ignore_buggy_testrun option to force ISAT to continue the tuning after encountering an error 12 Debugging ISAT yourself To debug ISAT the first thing to do is to look a
6. feedback out generated after running this code variant which contains the timing of this run References 1 K Datta M Murphy V Volkov S Williams J Carter L Oliker D Patterson J Shalf and K Yelick Stencil Computation Optimization and Auto tuning on State of the Art Multicore Architectures In Proceedings of the 2008 ACM IEEE Conference on Supercomputing 2008 2 M Puschel J Moura J Johnson D Pauda M Veloso B Singer J Xiong F Franchetti A Gacic Y Voronenko K Chen R Johnson and N Rizzolo SPIRAL Code Generation for DSP Transforms Proceedings of the IEEE special issue on Program Generation Optimization and Adaption 93 2 232 275 2005 3 R C Whaley A Petitet and J J Dongarra Automated Empirical Optimization of Software and the ATLAS Project Parallel Computing 27 1 2 3 35 2001
7. he compiler 8 Searching Scheme When there are multiple parameters being tuned in a tuning pragma the programmer can select the scheme used for searching the parameter space Currently ISAT supports two schemes independent search and dependent search The search scheme can be specified in a tuning by adding a search clause For instance you can choose to use independent search in the following pragma pragma isat tuning scope Ml_begin M1_end measure M2_begin M2_end variable x 10 20 30 variable y 70 80 90 100 search independent For the above pragma ISAT will first search for x from 10 20 30 that results in the minimum time measured between M2_begin and M2_end Then it will search for y from 70 80 90 100 And ISAT will use the best value of x it just found while searching y So the total number of code variants tried is the sum of the code variants of each variable which is 3 4 in this example If you change the search scheme from independent to dependent in the above pragma pragma isat tuning scope Mi_begin Mi_end measure M2_begin M2_end variable x 10 20 30 variable y 70 80 90 100 search dependent Then ISAT will search for the best performing x y from all possible combinations 10 70 10 80 10 90 10 100 20 70 20 80 20 90 20 100 30 70 30 80 30 90 30 100 If no searching scheme is specified in a tuning pragma independent search will be used 9 Thre
8. pragmas marker and tuning A marker pragma marks a location in the program while a tuning pragma specifies the tuning details As an example let s look at isat Examples BlockedMatrixMultiply Src blocked_mm cpp In this example we use ISAT to find the best block sizes in a blocked matrix multiply program static void BlockedMatrixMultiply FLOAT A FLOAT B FLOAT C const int n pragma isat tuning name tune_mm scope Ml_begin Mi_end measure M2_begin M2_end variable blk_k ran L1_CACHE_LINESIZE 2 variable blk_j range L1_CACHE_LINESIZE 16 L1_CACHE_LINESIZE L1_CACHE_LIN NESIZE 16 L1_CACHE_LINESIZE L1_CACHE_LINESIZE 2 pragma isat marker Mi_begin const int blk_i 64 const int blk_j 64 const int blk_k 64 for int i 0 i lt n i blk_i for int j 0 j lt n j blk_j for int k 0 k lt n k blk_k for int ii i ii lt MIN i blk_i n iit for int jj j jj lt MIN jtblk_j n jj for int kk k kk lt MIN ktblk_k n kk Cliixnt jj Clii n jj A ii n kk B kk n jj pragma isat marker Mi_end main int argc char argv pragma isat marker M2_begin BlockedMatrixMultiply A B C N pragma isat marker M2_end We use marker pragmas to define two regions in the source code Mi_begin Mi_end and M2_begin M2_end These regions are referenced in the tuning pragma which contains a name clause scope clause a measure clause and three variable clauses A name clause
9. specifies the name of region being tuned which shall be referenced in the tuning report A scope clause specifies the lexical scope that contains the program parameters being tuned A measure clause specifies where we should time a code variant A scope clause usually has two arguments to specify the starting and ending points of the region being tuned The two points must be in the same source file Nevertheless one can use scope to specify that the lexical scope includes every source file A measure clause must have two arguments to specify the starting and ending points of the region being measured A variable clause has two arguments The first argument is the variable that requires tuning The second argument is the list of values that should be tried for the variable A list can be in the enumerated form or in the formula form The enumerated form explicitly lists the values of a variable that should be tried such as 2 4 6 and red blue yellow green Note that an identifier like red or blue can be used in an enumerated list only if it is valid to substitute for the variable in the original program A list in the formula form is constructed by the range method defined in ISAT which accepts two to four arguments range beginExpr endExpr strideExpr fun The first argument beginExpr is an expression that calculates the beginning value of the list The second argument endExpr is an expression that calculates the ending value of
10. t logging warning error messages generated by ISAT These messages are with the prefix ISAT and are output to the console The ultimate way to debug ISAT is to inspect the code variants that it generates When you run ISAT it will print the name of the working directory on the console such as ISAT log Generate top working directory tmp_proj cluk1 ISAT Src f7ZOdy Now go to that working directory There should be a list of subdirectories v0 v1 v99 etc under the top working directory Each of these subdirectories contains a code variant that ISAT generated Go to one of them or the one where ISAT s error happened as reported by the ISAT warning error message That subdirectory is a copy of the original source directory which you specified with the i switch with a parameter being tuned set to a particular value Now type isat clean or the command you specified with clean_command and then type isat build or the command you specified with build_command You should be able to build the executable If there is any compilation or linking error look at the error message to understand what ISAT did wrongly in generating the code variant that caused the build error If the build is successful then type isat test or the command you specified with test_command Then check if the execution is successful and generates the expected result If everything is correct there should be a file named isat
11. the list Following C notation the beginning value is inclusive while the ending value is exclusive The third argument strideExpr is optional which is an expression that calculates the step value If it is not given a step value of one will be used The fourth argument fun is also optional which is the name of the function that will be applied to each list element to get the actual value to be substituted for the variable Currently fun must be identical pow2 or pow10 it is default to identical if not specified Followings are some examples range 1 6 generates 1 2 3 4 5 range 1 6 2 generates 1 3 5 range 1 6 1 pow2 generates 2 4 8 16 32 range 1 6 1 powl0 generates 10 100 1000 10000 100000 range L1_CACHE_LINESIZE 4 L1_CACHE_LINESIZE 1 L1_CACHE_LINESIZE 2 Assuming that L1_CACHE_LINESIZE is 64 the above range expression generates 64 96 128 160 192 224 256 As the last example illustrates ISAT predefined values those started with can be used in range expressions to represent system dependent parameters Table 1 lists all predefined values with their meanings Name Meaning NUM_CPU_THREADS Total number of hardware threads on the CPU L1_CACHE_LINESIZE L1 cache line size in bytes Table 1 ISAT pre defined values Finally note that a program with ISAT pragmas can be built by any native compiler the ISAT pragmas will be simply ignored by t
Download Pdf Manuals
Related Search