1 |
On Tue, 2019-05-28 at 01:37 -0700, Mo Zhou wrote: |
2 |
> Different BLAS/LAPACK implementations are expected to be compatible |
3 |
> to each other in both the API and ABI level. They can be used as |
4 |
> drop-in replacement to the others. This sounds nice, but the difference |
5 |
> in SONAME hampered the gentoo integration of well-optimized ones. |
6 |
|
7 |
If SONAMEs are different, then they are not compatible by definition. |
8 |
|
9 |
> Assume a Gentoo user compiled a pile of packages on top of the reference |
10 |
> BLAS and LAPACK, namely these reverse dependencies are linked against |
11 |
> libblas.so.3 and liblapack.so.3 . When the user discovered that |
12 |
> OpenBLAS provides much better performance, they'll have to recompile |
13 |
> the whole reverse dependency tree in order to take advantage from |
14 |
> OpenBLAS, |
15 |
> because the SONAME of OpenBLAS is libopenblas.so.0 . When the user |
16 |
> wants to try MKL (libmkl_rt.so), they'll have to recompile the whole |
17 |
> reverse dependency tree again. |
18 |
> |
19 |
> This is not friendly to our earth. |
20 |
> |
21 |
> Goal |
22 |
> ---- |
23 |
> |
24 |
> * When a program is linked against libblas.so or liblapack.so |
25 |
> provided by any BLAS/LAPACK provider, the eselect-based solution |
26 |
> will allow user to switch the underlying library without recompiling |
27 |
> anything. |
28 |
> |
29 |
> * When a program is linked against a specific implementation, e.g. |
30 |
> libmkl_rt.so, the solution doesn't break anything. |
31 |
> |
32 |
> Solution |
33 |
> -------- |
34 |
> |
35 |
> Similar to Debian's update-alternatives mechanism, Gentoo's eselect |
36 |
> is good at dealing with drop-in replacements as well. My preliminary |
37 |
> investigation suggests that eselect is enough for enabling BLAS/LAPACK |
38 |
> runtime switching. Hence, the proposed solution is eselect-based: |
39 |
> |
40 |
> * Every BLAS/LAPACK implementation should provide generic library |
41 |
> and eselect candidate libraries at the same time. Taking netlib, |
42 |
> BLIS and OpenBLAS as examples: |
43 |
> |
44 |
> reference: |
45 |
> |
46 |
> usr/lib64/blas/reference/libblas.so.3 (SONAME=libblas.so.3) |
47 |
> -- default BLAS provider |
48 |
> -- candidate of the eselect "blas" unit |
49 |
> -- will be symlinked to usr/lib64/libblas.so.3 by eselect |
50 |
|
51 |
/usr/lib64 is not supposed to be modified by eselect, it's package |
52 |
manager area. Yes, I know a lot of modules still do that but that's no |
53 |
reason to make things worse when people are putting significant effort |
54 |
to actually improve things. |
55 |
|
56 |
> usr/lib64/lapack/reference/liblapack.so.3 (SONAME=liblapack.so.3) |
57 |
> -- default LAPACK provider |
58 |
> -- candidate of the eselect "lapack" unit |
59 |
> -- will be symlinked to usr/lib64/liblapack.so.3 by eselect |
60 |
> |
61 |
> blis (doesn't provide LAPACK): |
62 |
> |
63 |
> usr/lib64/libblis.so.2 (SONAME=libblis.so.2) |
64 |
> -- general purpose |
65 |
> |
66 |
> usr/lib64/blas/blis/libblas.so.3 (SONAME=libblas.so.3) |
67 |
> -- candidate of the eselect "blas" unit |
68 |
> -- will be symlinked to usr/lib64/libblas.so.3 by eselect |
69 |
> -- compiled from the same set of object files as libblis.so.2 |
70 |
> |
71 |
> openblas: |
72 |
> |
73 |
> usr/lib64/libopenblas.so.0 (SONAME=libopenblas.so.0) |
74 |
> -- general purpose |
75 |
> |
76 |
> usr/lib64/blas/openblas/libblas.so.3 (SONAME=libblas.so.3) |
77 |
> -- candidate of the eselect "blas" unit |
78 |
> -- will be symlinked to usr/lib64/libblas.so.3 by eselect |
79 |
> -- compiled from the same set of object files as |
80 |
> libopenblas.so.0 |
81 |
> |
82 |
> usr/lib64/lapack/openblas/liblapack.so.3 (SONAME=liblapack.so.3) |
83 |
> -- candidate of the eselect "lapack" unit |
84 |
> -- will be symlinked to usr/lib64/liblapack.so.3 by eselect |
85 |
> -- compiled from the same set of object files as |
86 |
> libopenblas.so.0 |
87 |
> |
88 |
> This solution is similar to Debian's[3]. This solution achieves our |
89 |
> goal, |
90 |
> and it requires us to patch upstream build systems (same to Debian). |
91 |
> Preliminary demonstration for this solution is available, see below. |
92 |
|
93 |
So basically the three walls of text say in round-about way that you're |
94 |
going to introduce custom hacks to recompile libraries with different |
95 |
SONAME. Ok. |
96 |
|
97 |
> |
98 |
> Is this solution reliable? |
99 |
> -------------------------- |
100 |
> |
101 |
> * A similar solution has been used by Debian for many years. |
102 |
> * Many projects call BLAS/LAPACK libraries through FFI, including Julia. |
103 |
> (See Julia's standard library: LinearAlgebra) |
104 |
> |
105 |
> Proposed Changes |
106 |
> ---------------- |
107 |
> |
108 |
> 1. Deprecate sci-libs/{blas,cblas,lapack,lapacke}-reference from gentoo |
109 |
> main repo. They use exactly the same source tarball. It's not quite |
110 |
> helpful to package these components in a fine-grained manner. A |
111 |
> single |
112 |
> sci-libs/lapack package is enough. |
113 |
|
114 |
Where's the gain in that? |
115 |
|
116 |
> 2. Merge the "cblas" eselect unit into "blas" unit. It is potentially |
117 |
> harmful when "blas" and "cblas" point to different implementations. |
118 |
> That means "app-eselect/eselect-cblas" should be deprecated. |
119 |
> |
120 |
> 3. Update virtual/{blas,cblas,lapack,lapacke}. BLAS/LAPACK providers |
121 |
> will be registered in their dependency information. |
122 |
> |
123 |
> Note, ebuilds for BLAS/LAPACK reverse dependencies are expected to work |
124 |
> with these changes correctly without change. For example, my local |
125 |
> numpy-1.16.1 compilation was successful without change. |
126 |
> |
127 |
> Preliminary Demonstration |
128 |
> ------------------------- |
129 |
> |
130 |
> The preliminary implementation is available in my personal overlay[4]. |
131 |
> A simple sanity test script `check-cpp.sh` is provided to illustrate |
132 |
> the effectiveness of the proposed solution. |
133 |
> |
134 |
> The script `check-cpp.sh` compiles two C++ programs -- one calls general |
135 |
> matrix-matrix multiplication from BLAS, while another one calls general |
136 |
> singular value decomposition from LAPACK. Once compiled, this script |
137 |
> will switch different BLAS/LAPACK implementations and run the C++ |
138 |
> programs |
139 |
> without recompilation. |
140 |
> |
141 |
> The preliminary result is avaiable here[5]. (CPU=Power9, ARCH=ppc64le) |
142 |
> From the experimental results, we find that |
143 |
> |
144 |
> For (512x512) single precision matrix multiplication: |
145 |
> * reference BLAS takes ~360 ms |
146 |
> * BLIS takes ~70 ms |
147 |
> * OpenBLAS takes ~10 ms |
148 |
> |
149 |
> For (512x512) single precision singular value decomposition: |
150 |
> * reference LAPACK takes ~1900 ms |
151 |
> * BLIS (+reference LAPACK) takes ~1500 ms |
152 |
> * OpenBLAS takes ~1100 ms |
153 |
> |
154 |
> The difference in computation speed illustrates the effectiveness of |
155 |
> the proposed solution. Theoretically, any other package could take |
156 |
> advantage from this solution without any recompilation as long as |
157 |
> it's linked against a library with SONAME. |
158 |
|
159 |
An actual ABI compliance test, e.g. done using abi-compliance-checker |
160 |
would be more interesting. |
161 |
|
162 |
> |
163 |
> Acknowledgement |
164 |
> --------------- |
165 |
> This is an on-going GSoC-2019 Porject: |
166 |
> https://summerofcode.withgoogle.com/projects/?sp-page=2#6268942782300160 |
167 |
|
168 |
It would probably have been better if the project was discussed before |
169 |
GSoC. I'm really against pushing a bad idea forward just because |
170 |
someone set it for GSoC without discussing it first. |
171 |
|
172 |
-- |
173 |
Best regards, |
174 |
Michał Górny |