Android build with Cross Compilation for BLAS and OpenCL back-ends #848

lealgo · 2019-05-10T12:22:27Z

Provides the files and tweaks to meson that enable the cross-compilation to Android, and also to other platforms like ARM Linux. This PR depends on the work done by @borg323's on the subprojects: protobuf and opencl, allowing the build to be performed from source instead of using the build machine's shared libraries.

You could test this PR as it is right now, but the file subprojects/protobuf-3.5.1/meson.build should be replaced with this one: meson.build . The problem with cross-compiling protobuf is better explained here: #848 (comment)

Note: This PR started as a method for compiling lc0 on Termux, but then it evolved into cross-compilation. I kept the previous comments for reference. If you want instructions on Termux then follow the wiki

meson.build

lealgo · 2019-05-11T14:30:49Z

I'm starting to look at this PR as a proof of concept. I realize the real solution here is to use static versions of the libraries and ultimately to create a statically-linked executable, possibly with a separate meson file for cross-compiling with NDK. That's the only way if we want to distribute binaries. Some links:

http://dev-smart.com/cross-compiling-with-android/
http://mesonbuild.com/Cross-compilation.html
https://clang.llvm.org/docs/CrossCompilation.html

OpenBLAS for Android

https://github.com/xianyi/OpenBLAS/wiki/How-to-build-OpenBLAS-for-Android

OpenCL for Mali & Adreno

https://arrayfire.com/getting-started-with-opencl-on-android/
https://github.com/ARM-software/ComputeLibrary
https://github.com/ARM-software/armnn/blob/branches/armnn_19_02/BuildGuideAndroidNDK.md
https://developer.qualcomm.com/software/adreno-gpu-sdk

lealgo · 2019-05-14T06:09:20Z

Back to the real problem. I made some good progress today, achieving a successful cross-compilation with meson from my Linux workstation. For now the build was carried out without the back-ends. The final executable was created with its proper architecture.

The cross compilation file, cross-files/arm-linux-gnueabi.txt:

[host_machine]
system = 'linux'
cpu_family = 'arm'
cpu = 'armv7'
endian = 'little'

[binaries]
c = 'arm-linux-gnueabi-gcc'
cpp = 'arm-linux-gnueabi-g++'
ar = 'arm-linux-gnueabi-ar'
strip = 'arm-linux-gnueabi-strip'
ld = 'arm-linux-gnueabi-ld'
ranlib = 'arm-linux-gnueabi-ranlib'
as = 'arm-linux-gnueabi-as'
pkgconfig = 'arm-linux-gnueabi-pkg-config'
exe_wrapper = 'qemu-arm-static'

the commands:

meson --cross-file cross-files/arm-linux-gnueabi.txt build/ -Dbuild_backends=false
cd build
ninja

gonzalezjo · 2019-05-14T11:26:07Z

If the Tensorflow backend was brought up to date, would that mean that we could use the NN accelerators on modern Android devices?

mooskagh · 2019-05-14T19:33:03Z

Mobile devices use tensorflow-lite, not tensorflow. I think it's totally incompatible API-wise, but still pretty easy to write a backend for tensorflow-lite..

lealgo · 2019-05-15T20:03:43Z

For now the GNU's toolchain is able to build the project without the back-ends, but the resulting executable is for a generic linux system (different linker, libs, etc.) .

I'd prefer to use the NDK toolchain as it produces binaries for an Android platform. So I'm trying really hard to make the cross-compilation work with the NDK toolchain

@borg323 was helping me fix a problem with exe_wrapper in the meson cross-file. Apparently the protobuf build could not continue without an exe_wrapper. It seemed strange to me. Then I stumbled upon this:

From https://mesonbuild.com/Cross-compilation.html

Mixing host and build targets

Sometimes you need to build a tool which is used to generate source files. These are then compiled for the actual target. For this you would want to build some targets with the system's native compiler. This requires only one extra keyword argument.

native_exe = executable('mygen', 'mygen.c', native : true)

You can then take native_exe and use it as part of a generator rule or anything else you might want.

Now the previous meson error seems clearer:

ERROR: Can not use target protoc as a generator because it is cross-built
and no exe wrapper is defined. You might want to set it to native instead.

Meson has no protoc to call, only its source. Now I just need to tell it to use the local protoc. Right?

lealgo · 2019-05-16T03:51:30Z

Cross compilation finally achieved with the NDK toolchain!

After a real struggle and some magic tricks from @borg323 we could make protobuf pass the build. Here's the output of the first cross-compiled binary running on my phone through adb:

HWHWI:/data/local/tmp/new $ ./lc0 benchmark -b random -w /sdcard/DroidFish/lib/11258-48x5-se.pb.gz                                                                                                          
       _
|   _ | |
|_ |_ |_| v0.22.0-dev built May 15 2019
Loading weights file from: /sdcard/DroidFish/lib/11258-48x5-se.pb.gz
Creating backend [random]...
Benchmark time 8ms, 18 nodes, 2250 nps, move e2e3
Benchmark time 13ms, 32 nodes, 2461 nps, move e2e3
Benchmark time 20ms, 56 nodes, 2800 nps, move e2e3
...
Benchmark time 7304ms, 72391 nodes, 9911 nps, move b2b3
Benchmark time 8014ms, 79777 nodes, 9954 nps, move b2b3
Benchmark time 9746ms, 96296 nodes, 9880 nps, move b2b3
bestmove b2b3
Benchmark final time 9.77028s calculating 9888.25 nodes per second.

The random back-end is blazing fast! :D

Here's the meson cross-file:

[host_machine]
system = 'android'
cpu_family = 'arm'
cpu = 'aarch64'
endian = 'little'

[properties]
needs_exe_wrapper = true
cpp_args = []
cpp_link_args = ['-llog', '-latomic']

[binaries]
c = 'aarch64-linux-android28-clang'
cpp = 'aarch64-linux-android28-clang++'
ar = 'aarch64-linux-android-ar'
strip = 'aarch64-linux-android-strip'
ld = 'aarch64-linux-android-ld'
ranlib = 'aarch64-linux-android-ranlib'
as = 'aarch64-linux-android-as'

Now the proper back-ends need to be tackled. Expect a good fight!

lealgo · 2019-05-17T02:25:37Z

Guys,

@borg323 is making this look easy. Now his new branch opencl_sub, passed the cross-compilation on my android gnu toolchain. This means that apparently we've got the first back-end working on ARM! Look at this:

$ file lc0
lc0: ELF 32-bit LSB executable, ARM, EABI5 version 1 (GNU/Linux), dynamically linked, interpreter /lib/ld-, for GNU/Linux 3.2.0, BuildID[sha1]=28711f228713334199d6267c5726474c38f4a2f8, not stripped

$ qemu-arm-static lc0
       _
|   _ | |
|_ |_ |_| v0.22.0-dev built May 16 2019
uci
id name Lc0 v0.22.0-dev
id author The LCZero Authors.
option name WeightsFile type string default <autodiscover>
option name Backend type combo default opencl var opencl var check var random var roundrobin var multiplexing var demux
...

$ qemu-arm-static lc0 benchmark -w ../../../../weights/dkappe/32930-112x9-se.pb.gz 
       _
|   _ | |
|_ |_ |_| v0.22.0-dev built May 16 2019
Loading weights file from: ../../../../weights/dkappe/32930-112x9-se.pb.gz
Creating backend [opencl]...
OpenCL, maximum batch size set to 16.
Initializing OpenCL.
OpenCL: clGetPlatformIDs
terminate called after throwing an instance of 'cl::Error'
  what():  clGetPlatformIDs
qemu: uncaught target signal 6 (Aborted) - core dumped

Now I just gotta massage the NDK toolchain so that it follows suit...
This is great!

Some notes:

libzma linking is not needed and not asked for.
lpthread is not recognized but it's not needed, neither in the project nor in the sub. So we'll need more conditions in the block "if host_machine.system()..."

lealgo · 2019-05-17T04:43:58Z

Lc0 cross-built for android with NDK, opencl loader statically linked. adb output:

$ file lc0
lc0: ELF shared object, 64-bit LSB arm64, dynamic (/system/bin/linker64), for Android 28, built by NDK r19c (5345600), not stripped

$ ./lc0 benchmark -w /sdcard/DroidFish/11258-48x5-se.pb.gz                                                                                                                        
       _
|   _ | |
|_ |_ |_| v0.22.0-dev built May 17 2019
Loading weights file from: /sdcard/DroidFish/11258-48x5-se.pb.gz
Creating backend [opencl]...
OpenCL, maximum batch size set to 16.
Initializing OpenCL.
OpenCL: clGetPlatformIDs
terminating with uncaught exception of type cl::Error: clGetPlatformIDs
Aborted

Now, on to investigate why opencl is not being detected...
Maybe a mismatch of CL_TARGET_OPENCL_VERSION?

https://www.khronos.org/registry/OpenCL/specs/2.2/html/OpenCL_ICD_Installation.html#_android_sdkndk

lealgo · 2019-05-17T13:13:17Z

dynamically linked opencl on android:

$ ./lc0 benchmark -w /sdcard/DroidFish/11258-48x5-se.pb.gz                                                                                                                        
       _
|   _ | |
|_ |_ |_| v0.22.0-dev built May 17 2019
Loading weights file from: /sdcard/DroidFish/11258-48x5-se.pb.gz
Creating backend [opencl]...
OpenCL, maximum batch size set to 16.
Initializing OpenCL.
Detected 1 OpenCL platforms.
Platform version: OpenCL 2.0 v1.r14p0-00cet0.b33ebadea0b6ef5b967c7f21064b122d
Platform profile: FULL_PROFILE
Platform name:    ARM Platform
Platform vendor:  ARM
Device ID:      0
Device name:    Mali-G71
Device type:    GPU
Device vendor:  ARM
Device driver:  2.0
Device speed:   5 MHZ
Device cores:   8 CU
Device score:   120
Selected platform: ARM Platform
Selected device: Mali-G71
with OpenCL 2.0 capability.
Started OpenCL SGEMM tuner with batch size 16.
Will try 578 valid configurations.
(2/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=16 SA=0 SB=0 STRM=0 STRN=0 VWM=1 VWN=1 2741.6 us (6.9 GFLOPS)
(3/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=64 NDIMB=8 NDIMC=8 NWG=16 SA=0 SB=0 STRM=0 STRN=0 VWM=1 VWN=1 2644.4 us (7.1 GFLOPS)
(5/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=32 SA=0 SB=0 STRM=0 STRN=0 VWM=1 VWN=1 1655.1 us (11.4 GFLOPS)
(17/578) KWG=32 KWI=2 MDIMA=16 MDIMC=16 MWG=32 NDIMB=8 NDIMC=8 NWG=64 SA=0 SB=0 STRM=0 STRN=0 VWM=1 VWN=1 1415.2 us (13.3 GFLOPS)
(24/578) KWG=32 KWI=2 MDIMA=32 MDIMC=32 MWG=64 NDIMB=8 NDIMC=8 NWG=64 SA=0 SB=0 STRM=0 STRN=0 VWM=1 VWN=1 1411.8 us (13.4 GFLOPS)
(69/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=32 SA=0 SB=0 STRM=0 STRN=0 VWM=2 VWN=1 1381.3 us (13.7 GFLOPS)
(77/578) KWG=32 KWI=2 MDIMA=16 MDIMC=16 MWG=64 NDIMB=8 NDIMC=8 NWG=32 SA=0 SB=0 STRM=0 STRN=0 VWM=2 VWN=1 1235.2 us (15.3 GFLOPS)
(90/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=16 NDIMC=16 NWG=64 SA=0 SB=0 STRM=0 STRN=0 VWM=2 VWN=1 1062.2 us (17.8 GFLOPS)
(115/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=32 SA=0 SB=0 STRM=0 STRN=0 VWM=4 VWN=1 892.4 us (21.1 GFLOPS)
(198/578) KWG=32 KWI=2 MDIMA=16 MDIMC=16 MWG=32 NDIMB=8 NDIMC=8 NWG=64 SA=0 SB=0 STRM=0 STRN=0 VWM=2 VWN=2 871.3 us (21.7 GFLOPS)
(223/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=32 SA=0 SB=0 STRM=0 STRN=0 VWM=4 VWN=2 732.8 us (25.8 GFLOPS)
(281/578) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=32 SA=0 SB=0 STRM=0 STRN=0 VWM=4 VWN=4 699.7 us (27.0 GFLOPS)
Wavefront/Warp size: 4

Max workgroup size: 384
Max workgroup dimensions: 384 384 384
Benchmark time 1038ms, 7 nodes, 6 nps, move d2d4
Benchmark time 1406ms, 16 nodes, 11 nps, move g2g3
Benchmark time 1896ms, 35 nodes, 18 nps, move g2g3
Benchmark time 2526ms, 56 nodes, 22 nps, move g2g3
Benchmark time 2685ms, 74 nodes, 27 nps, move g2g3
Benchmark time 3160ms, 98 nodes, 31 nps, move g2g3
Benchmark time 3688ms, 130 nodes, 35 nps, move g2g3
Benchmark time 4275ms, 170 nodes, 39 nps, move g2g3
Benchmark time 4427ms, 203 nodes, 45 nps, move c2c4
Benchmark time 4907ms, 240 nodes, 48 nps, move g1f3
Benchmark time 4916ms, 245 nodes, 49 nps, move c2c4
Benchmark time 5067ms, 269 nodes, 53 nps, move g1f3
Benchmark time 5784ms, 353 nodes, 61 nps, move g1f3
Benchmark time 6565ms, 446 nodes, 67 nps, move c2c4
Benchmark time 7041ms, 480 nodes, 68 nps, move c2c4
Benchmark time 7476ms, 533 nodes, 71 nps, move g1f3
Benchmark time 7521ms, 542 nodes, 72 nps, move c2c4
Benchmark time 7801ms, 575 nodes, 73 nps, move g1f3
Benchmark time 9224ms, 745 nodes, 80 nps, move c2c4
Benchmark time 9521ms, 797 nodes, 83 nps, move g1f3
Benchmark time 10000ms, 814 nodes, 81 nps, move g1f3
bestmove g1f3
Benchmark final time 10.7405s calculating 88.4505 nodes per second.

lealgo · 2019-05-17T15:29:42Z

Now we've got a beautiful 8 MB binary engine that should work with regular Android chess GUIs!

With OpenCL back-end.

This is bliss!!!

lealgo · 2019-05-19T15:54:11Z

Attached meson detection when cross compiling
meson detection.zip

lealgo · 2019-05-19T15:58:40Z

@borg323

Some comments about the most recent commit:

The build will fail unless you apply the following subprojects changes. Attached the modified meson.build files, not the diffs.

opencl meson.zip
protobuf meson.zip

I tested the changes to work for a regular non-cross compile build.
There is at least one change (maybe more) that is not ideal, but I pushed just it to be able to comment on them and seek advise.

lealgo · 2019-05-19T16:33:45Z

meson_options.txt

@@ -118,11 +118,6 @@ option('pext',
       value: false,
       description: 'Use the pext instruction')

-option('android',


This is no longer needed as one can detect android on meson cross build with host_machine.system() == 'android'

lealgo · 2019-05-19T16:46:06Z

meson_options.txt

@@ -15,7 +15,7 @@ option('openblas_include',

 option('opencl_include',
       type: 'array',
-       value: ['/usr/include/'],
+#       value: ['/usr/include/'],


I had to comment this path as it seems to confuse the Android NDK, it mixes the headers and fail compilation. The GNU toolchain doesn't have that problem. Any idea on how to solve this?

baltazar388 · 2019-05-20T01:00:51Z

lealgo sir you can make a small movie for see how install weithgs good my phone is a huawei p smart but doesnt work lc0 help me please , thansa for your time :) regards

flither · 2019-05-20T09:05:59Z

Not working on my side (Xiaomi Mi 6, Android 9, opencl 2.0, Snapdragon 835).
Engine "thinks" for 2 minutes, then terminates with "engine error" message.
Termux method works fine though.

ankan-ban · 2019-05-20T13:40:30Z

Works perfectly on my one plus 6 (Snapdragon 845). With a recent T40 network:

$ ls
lc0                       weights_run1_42372.pb.gz
$ ./lc0 benchmark
       _
|   _ | |
|_ |_ |_| v0.22.0-dev built May 20 2019
Found pb network file: ./weights_run1_42372.pb.gz
Creating backend [blas]...
BLAS, maximum batch size set to 256
BLAS vendor: OpenBlas.
OpenBlas [OpenBLAS 0.3.6 NO_LAPACK NO_LAPACKE NO_AFFINITY ARMV8 MAX_THREADS=8].
OpenBlas found 8 ARMV8 core(s).
OpenBLAS using 1 core(s) for this backend.
BLAS max batch size is 256.
Benchmark time 2313ms, 4 nodes, 1 nps, move e2e4
Benchmark time 5318ms, 8 nodes, 1 nps, move g1f3
Benchmark time 8763ms, 10 nodes, 1 nps, move g1f3
Benchmark time 9010ms, 19 nodes, 2 nps, move g1f3
Benchmark time 10000ms, 21 nodes, 2 nps, move g1f3
bestmove g1f3
Benchmark final time 12.791s calculating 3.04901 nodes per second.
$

This reverts commit c933b1a.

bicho2 · 2019-07-03T19:30:50Z

Hi, I compared openblas with opencl engines on samsungS8+. I set the uci depth to 2 moves. Openblas needeed 10 seconds against 7 for opencl. I am quite deceived by opencl. It need 3 minutes to start and then it is just a bit quicker than blas. The engine is stronger than stockfish for a given depth but because it is slow stockfish beats it because it can go deeper thanks to its fast evaluation speed. I thank you a lot to allow me to play chess against an IA. It is more human like. Better evaluation of position and less depth in move calculation.

lealgo · 2019-08-12T23:28:39Z

New builds for 0.22 release:

aarch64 builds for Android 5.0 and up:

lc0-0.22-eigen-aarch64.zip
lc0-0.22-blas-aarch64.zip
lc0-0.22-opencl-aarch64.zip

with Little Demon 2 embedded:
lc0-0.22-LD2-eigen-aarch64.zip
lc0-0.22-LD2-blas-aarch64.zip
lc0-0.22-LD2-opencl-aarch64.zip

armv7a builds for Android 4.1 and up:

lc0-0.22-eigen-armv7a.zip
lc0-0.22-blas-armv7a.zip

with Little Demon 2 embedded:
lc0-0.22-LD2-eigen-armv7a.zip
lc0-0.22-LD2-blas-armv7a.zip

UA2425 · 2019-11-22T09:00:25Z

Hey there lealgo! Great to see that you're still active and working on this project.

…

On Fri, 22 Nov 2019, 10:31 Leandro Álvarez González, < ***@***.***> wrote: New builds for 0.23-rc1: - aarch64 build for Android 5.0 and up: lc0-eigen-opencl-aarch64.zip <https://github.com/LeelaChessZero/lc0/files/3878016/lc0-eigen-opencl-aarch64.zip> - armv7a build for Android 4.1 and up: lc0-eigen-opencl-armv7a.zip <https://github.com/LeelaChessZero/lc0/files/3878018/lc0-eigen-opencl-armv7a.zip> Notice they differ from previous builds in that they now have both the EIgen and OpenCL back-ends compiled into the same binary, defaulting to the Eigen one. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#848?email_source=notifications&email_token=AI4GAHTM7GB6K4ALP6AOMUDQU5RRDA5CNFSM4HMCTILKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEE4RQUQ#issuecomment-557389906>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AI4GAHUT7U4U4BC36ANGMDTQU5RRDANCNFSM4HMCTILA> .

lealgo · 2019-11-22T22:51:43Z

Hi,

@UA2425 Thanks.
I'm not really active again but at least I try to keep the builds alive for the main releases. I'm going to merge the current release back into this branch to keep it current, and from what I've seen there are no breaking changes, that's very good. This branch really has very few diffs. against main and most are for the build system. The only road block is the protobuf dependency for cross-compiling, that it's working thanks to some magic @borg323 did a while ago. But as far as I know there's no clean way to do it.

Best regards,
Leandro

lealgo · 2019-12-02T04:40:46Z

Builds for the new release 0.23:

aarch64 for Android 5.0 and up:
lc0-0.23-blas-aarch64.zip
lc0-0.23-eigen-aarch64.zip
lc0-0.23-opencl-aarch64.zip

armv7a for Android 4.1 and up:
lc0-0.23-blas-armv7.zip
lc0-0.23-eigen-armv7.zip
lc0-0.23-opencl-armv7.zip

lealgo · 2020-02-14T18:18:06Z

After the change to c++17, the master branch is building fine with the Android NDK r21.

There's an (old) issue though when targeting armv7a and trying to build for EABI < 24 (Android 7). The problem is that recent NDK's apparently have trouble targeting 32-bit Android for older API levels. The issue is explained here:

https://android.googlesource.com/platform/bionic/+/master/docs/32-bit-abi.md

Previously I was using an older NDK as a workaround, but such toolchains don't support c++17 so now we're out of luck.

TLDR From now on the armv7a builds will require Android 7 and above. aarch64 builds are unaffected and will continue to support Android 5.

lealgo · 2020-02-24T03:46:15Z

Big news!

First appveyor TEST build successful, by @borg323 :

https://ci.appveyor.com/project/borg323/lc0/builds/30998575

HWHWI:/data/local/tmp $ ./lc0 benchmark -w 11258-32x4-se --threads=8 --max-prefetch=0 --minibatch-size=8                                                                                                    
WARNING: linker: "/data/local/tmp/lc0" unused DT entry: type 0xf arg 0x629
       _
|   _ | |
|_ |_ |_| v0.25.0-dev+git.9be6552 built Feb 23 2020
Loading weights file from: 11258-32x4-se
Creating backend [blas]...
Using Eigen version 3.3.5
BLAS max batch size is 256.
Benchmark time 49ms, 4 nodes, 117 nps, move d2d4
Benchmark time 77ms, 6 nodes, 96 nps, move e2e4
Benchmark time 91ms, 10 nodes, 131 nps, move d2d4
...
Benchmark time 4792ms, 4349 nodes, 910 nps, move d2d4
Benchmark time 7834ms, 7769 nodes, 993 nps, move d2d4
Benchmark time 10000ms, 10354 nodes, 1036 nps, move d2d4
bestmove d2d4
Benchmark final time 10.0874s calculating 1032.77 nodes per second.

Note a small linker warning on the first line, it's produced because of the fairly old NDK r17 toolchain. Newer NDK's produce clean binaries.

lealgo · 2020-03-02T01:27:00Z

The Android builds are official now, I think I can close this PR. Thank you.

lealgo · 2020-03-02T17:42:50Z

Test builds for comparing OpenBLAS 0.3.9 vs 0.3.8

lc0-blas-aarch64.zip

Android build

c954e33

borg323 reviewed May 10, 2019

View reviewed changes

meson.build Outdated Show resolved Hide resolved

borg323 added the wip Work in progress label May 11, 2019

UA2425 approved these changes May 12, 2019

View reviewed changes

This comment has been minimized.

Sign in to view

Move lzma link argument after opencl detection

6afb880

This comment has been minimized.

Sign in to view

borg323 added the help wanted Extra attention is needed label May 14, 2019

Initial support for cross-compilation

c9039f3

lealgo changed the title ~~Android build with Termux for BLAS and OpenCL back-ends~~ Android build with Cross Compilation for BLAS and OpenCL back-ends May 19, 2019

lealgo commented May 19, 2019

View reviewed changes

This comment has been minimized.

Sign in to view

lealgo added 3 commits June 10, 2019 15:50

Revert "Add alternate paths for the tuning file on Android"

3d55004

This reverts commit c933b1a.

Revert to master's tuning file location

241ed4a

add link to zlib issue with older NDK's

090e68c

lealgo mentioned this pull request Jun 25, 2019

Update meson.build #891

Closed

This comment has been minimized.

Sign in to view

l1t1 mentioned this pull request Sep 3, 2019

lc0 android version supports opencl leela-zero/leela-zero#2484

Open

v0.23.0-dev -> v0.23.0-rc1

5535d75

This comment has been minimized.

Sign in to view

mooskagh and others added 8 commits November 27, 2019 21:08

Merge remote-tracking branch 'origin/master' into release/0.23

12bc3f0

v0.23.0-rc2

65e13d8

Merge remote-tracking branch 'upstream/release/0.23' into android-build

e4d50c4

Fix meson opencl vars messed up after merge

b966995

make eigen the last blas option (LeelaChessZero#1025)

89a9ec3

v0.23.0-rc2 -> v0.23.0

02fc8e0

Merge remote-tracking branch 'upstream/release/0.23' into android-build

2498564

Add PIE switches to armv7 cross-file

120566e

borg323 mentioned this pull request Dec 5, 2019

downloading problem #1029

Closed

lealgo closed this Mar 2, 2020

This comment has been minimized.

Sign in to view

Android build with Cross Compilation for BLAS and OpenCL back-ends #848

Android build with Cross Compilation for BLAS and OpenCL back-ends #848

Conversation

lealgo commented May 10, 2019 • edited Loading

lealgo commented May 11, 2019 • edited Loading

OpenBLAS for Android

OpenCL for Mali & Adreno

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

lealgo commented May 14, 2019

gonzalezjo commented May 14, 2019 • edited Loading

mooskagh commented May 14, 2019

lealgo commented May 15, 2019 • edited Loading

lealgo commented May 16, 2019

lealgo commented May 17, 2019 • edited Loading

lealgo commented May 17, 2019 • edited Loading

lealgo commented May 17, 2019

lealgo commented May 17, 2019 • edited Loading

lealgo commented May 19, 2019 • edited Loading

lealgo commented May 19, 2019 • edited Loading

lealgo May 19, 2019

Choose a reason for hiding this comment

lealgo May 19, 2019 • edited Loading

Choose a reason for hiding this comment

This comment has been minimized.

baltazar388 commented May 20, 2019

flither commented May 20, 2019

This comment has been minimized.

ankan-ban commented May 20, 2019

bicho2 commented Jul 3, 2019

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

lealgo commented Aug 12, 2019 • edited Loading

This comment has been minimized.

UA2425 commented Nov 22, 2019 via email

lealgo commented Nov 22, 2019 • edited Loading

lealgo commented Dec 2, 2019

lealgo commented Feb 14, 2020 • edited Loading

lealgo commented Feb 24, 2020 • edited Loading

lealgo commented Mar 2, 2020

lealgo commented Mar 2, 2020

This comment has been minimized.

lealgo commented May 10, 2019 •

edited

Loading

lealgo commented May 11, 2019 •

edited

Loading

gonzalezjo commented May 14, 2019 •

edited

Loading

lealgo commented May 15, 2019 •

edited

Loading

lealgo commented May 17, 2019 •

edited

Loading

lealgo commented May 17, 2019 •

edited

Loading

lealgo commented May 17, 2019 •

edited

Loading

lealgo commented May 19, 2019 •

edited

Loading

lealgo commented May 19, 2019 •

edited

Loading

lealgo May 19, 2019 •

edited

Loading

lealgo commented Aug 12, 2019 •

edited

Loading

lealgo commented Nov 22, 2019 •

edited

Loading

lealgo commented Feb 14, 2020 •

edited

Loading

lealgo commented Feb 24, 2020 •

edited

Loading