Skip to content
GitLab
Select Git revision
  • master default
Find file BlameHistoryPermalink
Forked from Juni May / proj105-yatuner
Source project has a limited visibility.
CHANGELOG 6.92 KiB
1 
-------------------
2 
* Changes in 4.2.1:
3 
-------------------
4 
  - Fixed an issue with PAPI support.
5 
    Reported by Keyur Joshi <cs13b1014@iith.ac.in>
6 
  - Added support for PAPI 5.4.x.
7
8 
-----------------
9 
* Changes in 4.2:
10 
-----------------
11 
  - Fixed an issue with OpenMP pragma for flush_cache.
12 
    Reported by Sudheer Kumar <sudheer3@gmail.com>.
13 
  - Fixed a bug in Makefile generation (utilities/makefile-gen.pl).
14 
    Reported by Willy Wolff <willy.mh.wolff@gmail.com>.
15 
  - Fixed a bug in syr2k that caused the kernel to perform twice the work.
16 
    It was computing the full symmetric output whereas the BLAS version
17 
    only computes one side of the symmetry.
18 
  - Input generation for some of the kernels has been updated to produce
19 
    fewer zeros in output.
20 
  - Add support for inter-array padding when using heap allocation,
21 
    and adjust the allocation alignment with posix_memalign to 4096.
22
23 
-----------------
24 
* Changes in 4.1:
25 
-----------------
26 
  - Added LICENSE.txt
27 
  - Fixed a bug in documentation of cholesky. (Reported by François Gindraud)
28 
  - Removed two statements from cholesky, which were useless with inplace
29 
    memory allocation. (Reported by François Gindraud)
30 
  - Updated polybench.R in utilities to match the change in input generation
31 
    of lu/ludcmp/cholesky.
32 
  - Simplified the macros for switching between data types.
33 
      - Now users may specify DATA_TYPE_IS_XXX where XXX is one of
34 
        FLOAT/DOUBLE/INT to change all macros associated with data types.
35 
  - Fixed a typo in Jacobi-1d loop iterators.
36 
  - Fixed issues with SCALAR_VAL macro in some kernels.
37 
    (Reported by Sven Verdoolaege)
38 
  - Fixed a typo in POLYBENCH_GFLOPS system.
39
40 
-----------------
41 
* Changes in 4.0:
42 
-----------------
43
44 
This is a detailed ChangeLog for PolyBench/C 4.0 to indicate all
45 
changes made from version 3.2. Due to the high number of changes,
46 
users should be warned the baseline performance of kernels available
47 
in both PolyBench/C 3.2 and 4.0 may differ.
48
49
50 
= General
51 
  - tuned the initialization functions to reduce the possibility to have 'inf'
52 
    in outputs. Specifically, the initial values are much closer to 1 for most
53 
    linear algebra kernels now (was up to N in prior versions).
54 
  - changed the name of predefined problem sizes STANDARD to MEDIUM.
55 
  - changed the default dataset to LARGE.
56 
  - added a few perl scripts to utilities.
57 
  - replaced create_cpped_version.sh with a perl script.
58 
  - fixed a bug in polybench.h when 4D arrays were allocated. (extra comma)
59 
    (Reported by Amarin Phaosawasdi)
60 
  - added polybench.pdf as a documentation of all kernels and its
61 
    underlying algorithms
62 
  - added POLYBENCH_USE_RESTRICT to allow compilers to assume alias-free
63 
    (Patch by Tobias Grosser)
64 
  - added SCALAR_VAL, SQRT_FUN, EXP_FUN, and POW_FUN macros to switch
65 
    float/double versions of the math functions depending on the data type.
66 
  - changed naming of a variable in polybench.c/xmalloc to avoid
67 
    issues when compiled as C++ (Reported by Sven Verdoolaege)
68 
  - changed the outputs produced by POLYBENCH_DUMP_ARRAYS option to
69 
    separately print out each array.
70 
  - added polybench.R in utilities which defines datamining and
7172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140
linear-algebra kernels using R script for testing. = Datamining - fixed a bug in covariance where array size and loop bounds did not match each other, causing out-of-bounds for certain parameters. (Reported by Tobias Grosser) - fixed a bug in covariance where division by N at the end was missing. - fixed the initialization of float_n in covariance and correlation. The value is supposed to be the input size N casted as floats to be used in division, but was initialized to 1.2. - changed the name of array 'symmat' in covariance to 'cov'. - changed the name of array 'symmat' in correlation to 'corr'. - changed the loop indices in covariance and correlation to match the documentation. - removed sqrt_of_array_cell macro from correlation = Linear Algebra - fixed a bug in 2mm where array sizes did not match each other, causing out-of-bounds for non-square inputs. (Reported by Tobias Grosser) - fixed a bug in syrk where the loop bounds were incorrect, having rectangular loop instead of triangular. - fixed a bug in trmm where loop bounds were incorrect, using the wrong triangular region in the multiplication. - reduced the size of 'sum' array in doitgen from 3D to 1D to avoid obvious waste in memory. - removed A from print_arrays in gramschmidt; A is updated, but is not an output. - removed dynprog; replaced by nussinov in medley. - moved BLAS kernels (including updated BLAS) to its own sub-category. - BLAS kernels are now commented with the parameters in original BLAS which corresponds to the implementation in PolyBench. - BLAS kernels now closely matches the original version. However, gemver and gesummv remains the same as it is not part of current BLAS. - moved cholesky and trisolv to solvers. - re-implemented cholesky so that it computes the full L inplace. Previously the diagonal was stored in a separate vector. - re-implemented durbin to match Durbin's algorithm from a book. There were off-by-one errors and an excessive memory allocation (due to expanding accumulation of a summation). - re-implemented lu. The original implementation was missing a inner most loop for computing the U matrix. - re-implemented ludcmp. There were off-by-one errors leading to incorrect outputs. - lu and ludcmp now use same input as cholesky to ensure it works. - input of cholesky/lu/ludcmp now uses L.L^T instead of L^T.L to create inputs. - loop bounds of symm/syrk/syr2k/trmm are changed to match the documentation. = Medley - changed default datatype of floyd_warshall from double to int. - removed reg_detect; replaced by deriche. - added nussinov; a dynamic programming algorithm for sequence alignment. (Code by Dave Wonnacott and his students) - added deriche; edge detector filter. (Code by Gael Deest) = Stencils - re-implemented adi based on a figure in "Automatic Data and Computation Decomposition on Distributed Memory Parallel Computers" by Peizong Lee and Zvi Meir Kedem, TOPLAS, 2002 - removed fdtd-apml; replaced by heat-3d - added heat-3d; Heat equation over 3D data domain (4D iteration space) (Original specification from Pochoir compiler test case) - changed jacobi-1d-imper and jacobi-2d-imper to jacobi-1d and jacobi-2d, respectively. - jacobi-1d, jacobi-2d, and heat-3d performs two time steps using alternating arrays per an iteration of the outermost loop. This is to avoid the copy loop in the old xxx-imper versions. The number of stencil iterations are now restricted to be even numbers (2x of
141142143144145146147148
the parameter TSTEPS). = Known Issues - output of correlation will always be mostly 1.0 due to how the input is generated. It is difficult to avoid this case when the input is some function of the indices, and will be addressed with more fundamental update in the future.