merge master -> libedax_sensuikan1973

.github/workflows/build.yaml

@@ -21,7 +21,7 @@ jobs:
           - os: windows-latest
             build_command: make build ARCH=x64 COMP=gcc OS=windows
           - os: macos-latest
-            build_command: make build ARCH=x64-modern COMP=gcc OS=osx
+            build_command: make build ARCH=arm COMP=gcc OS=osx
 
     steps:
     - uses: actions/checkout@v2

.gitignore

@@ -1,7 +1,44 @@
+<<<<<<< HEAD
+<<<<<<< HEAD
 bin/
 data/
 doc/
 
 *.7z
 
 .DS_Store
+=======
+*.o
+*.s
+*.exe
+*~
+pgopti*
+*.dyn
+all.gc*
+generate_flip.*
+generate_count_flip.*
+bcnttest.*
+*.utf8.c
+bin/
+doc/
+<<<<<<< HEAD
+>>>>>>> dd6b636 (Bcc32 friendly and minor improvement on Flip_32.)
+=======
+problem/
+>>>>>>> 48873fa (calc opponent_feature once in eval_open)
+=======
+*.o
+*.s
+*.exe
+*~
+pgopti*
+*.dyn
+all.gc*
+generate_flip.*
+generate_count_flip.*
+bcnttest.*
+*.utf8.c
+bin/
+doc/
+problem/
+>>>>>>> 3e1ed4f (fix cr/lf in repository to lf)

README.md

@@ -51,3 +51,111 @@ cd src
 doxygen
 open ../doc/html/index.html
 ```
+=======
+# edax-reversi-AVX
+Automatically exported from code.google.com/p/okuharaandroid-edax-reversi
+
+=======
+# edax-reversi-AVX
+Automatically exported from code.google.com/p/okuharaandroid-edax-reversi
+
+Edax is a strong othello program. Its main features are:
+
+    fast bitboard based & multithreaded engine.
+    accurate midgame-evaluation function.
+    opening book learning capability.
+    text based rich interface.
+    multi-protocol support to connect to graphical interfaces or play on Internet (GGS).
+    multi-OS support to run under MS-Windows, Linux and Mac OS X.
+
+>>>>>>> 81dec96 (Kindergarten last flip for arm32; MSVC arm Windows build (not tested))
+This is SSE/AVX optimized version of Edax 4.4.0. Functionally equivalent to the parent project, provided no bugs are introduced.
+
+Thanks to AVX2, x64-modern build solves fforum-40-59.obf 60% faster than official edax-4.4 on Haswell, and runs level 30 autoplay 80% faster.
+
+See http://www.amy.hi-ho.ne.jp/okuhara/bitboard.htm and http://www.amy.hi-ho.ne.jp/okuhara/edaxopt.htm for optimization details in Japanese.
+
+## 1. Mobility (board_sse.c, board_mmx.c)
+
+### 1.1 new SSE2 version of get_moves
+Diagonals are SIMD'd using vertical mirroring by bswap.
+
+    Athlon -get_moves_sse
+    problem\fforum-20-39.obf: 111349635 nodes in 0:07.998 (13922185 nodes/s).
+    mobility: 81.10 < 81.28 +/- 0.17 < 82.03
+    Athlon +get_moves_sse
+    problem\fforum-20-39.obf: 111349635 nodes in 0:07.889 (14114544 nodes/s).
+    mobility: 71.08 < 71.72 +/- 0.34 < 73.53
+    Core2 -get_moves_sse
+    problem/fforum-20-39.obf: 111349635 nodes in 0:10.180 (10938078 nodes/s).
+    mobility: 78.06 < 78.18 +/- 0.08 < 78.41
+    Core2 +get_moves_sse
+    problem/fforum-20-39.obf: 111349635 nodes in 0:09.978 (11159514 nodes/s).
+    mobility: 60.84 < 61.19 +/- 0.13 < 61.47
+
+### 1.2 can_move
+Now calls SIMD'd get_moves for x86/x64 build.
+
+## 2. Stability (board.c, board_sse.c, board_mmx.c)
+
+### 2.1 get_full_lines_h, get_full_lines_v
+get_full_lines for horizontal and vertical are simplified. The latter is compiled into rotation instrunction.
+
+### 2.2 rearranged loop
+The last while loop is rearranged not to call bit_count in case stable == 0.
+
+### 2.3 new SSE2 version with bswap and pcmpeqb
+    Athlon -get_stability_sse
+    stability: 90.10 < 90.28 +/- 0.24 < 91.20
+    Athlon +get_stability_sse
+    stability: 81.59 < 81.93 +/- 0.73 < 86.25
+    Core2 -get_stability_sse
+    stability: 79.24 < 79.39 +/- 0.15 < 79.93
+    Core2 +get_stability_sse
+    stability: 71.80 < 71.85 +/- 0.06 < 72.07
+
+### 2.4 get_corner_stability
+Kindergarten version eliminates bit_count call.
+
+### 2.5 find_edge_stable
+Loop optimization and flip using carry propagation. One time execution but affect total solving time.
+
+## 3. eval.c (4.4.5)
+Eval feature calculation using SSE2 / AVX2 (now in eval_sse.c) improves midgame by 15-30% and endgame by 8-12%.
+Restoring eval from backup instead of rewinding.
+eval_open (one time execution) is also optimized.
+
+## 4. hash.c
+I think hash->data.move[0] on line 677 should be hash->data.move[1].
+
+## 5. board_symetry, board_unique (board.c, board_sse.c)
+SSE optimization and mirroring reduction. (Not used in solving game)
+
+## 6. endgame_sse.c (4.4.7)
+Keep more variables in SSE registers.  SSE optimized count_last_flip.  Parity sort by shuffle.
+
+## 7. board_get_hash_code (4.5.0)
+Changed to use CRC32c.  This enables hardware acceleration on modern build.
+
+## 8. AVX2 versions (x64-modern build only)
+In many cases AVX2 version is simplest, thanks to variable shift instructions (although they are 3 micro-op instructions).
+
+Benchmarks are on Core i5-4260U (Haswell) 1.4GHz (TB 2.7GHz) single thread.
+
+    4.4.0 original x64-modern clang
+    problem/fforum-20-39.obf: 111349635 nodes in 0:05.726 (19446321 nodes/s).
+    +optimizations 1-5 above, no-avx2
+    problem/fforum-20-39.obf: 111349635 nodes in 0:05.342 (20844185 nodes/s).
+    +get_moves (board_sse.c)
+    problem/fforum-20-39.obf: 111349635 nodes in 0:05.142 (21654927 nodes/s).
+    +flip_avx.c
+    problem/fforum-20-39.obf: 111349635 nodes in 0:04.946 (22513068 nodes/s).
+    +count_last_flip_sse.c
+    problem/fforum-20-39.obf: 111349635 nodes in 0:04.906 (22696624 nodes/s).
+
+## 9. makefile
+gcc-old, x86 build should be -m32, not -m64. Some flags and defines added for optimization.
+<<<<<<< HEAD
+>>>>>>> b9d48c1 (Create README.md)
+=======
+>>>>>>> 81dec96 (Kindergarten last flip for arm32; MSVC arm Windows build (not tested))

src/Android.mk

@@ -0,0 +1,8 @@
+LOCAL_PATH := $(call my-dir)
+include $(CLEAR_VARS)
+LOCAL_MODULE := aEdax # should be renamed to lib..aEdax..so afterwords
+LOCAL_CFLAGS += -DUNICODE
+LOCAL_SRC_FILES := all.c board_sse.c.neon eval_sse.c.neon flip_neon_bitscan.c.neon android/cpu-features.c
+LOCAL_ARM_NEON := false
+# cmd-strip :=
+include $(BUILD_EXECUTABLE)

src/Application.mk

@@ -0,0 +1,3 @@
+APP_ABI := armeabi-v7a arm64-v8a x86 x86_64
+APP_PLATFORM := android-14
+APP_BUILD_SCRIPT := Android.mk

src/Doxyfile

@@ -687,7 +687,7 @@ RECURSIVE              = NO
 # Note that relative paths are relative to the directory from which doxygen is
 # run.
 
-EXCLUDE                =
+EXCLUDE                = _*.c
 
 # The EXCLUDE_SYMLINKS tag can be used to select whether or not files or
 # directories that are symbolic links (a Unix file system feature) are excluded
@@ -1599,7 +1599,7 @@ HIDE_UNDOC_RELATIONS   = YES
 # toolkit from AT&T and Lucent Bell Labs. The other options in this section
 # have no effect if this option is set to NO (the default)
 
-HAVE_DOT               = YES
+HAVE_DOT               = NO
 
 # The DOT_NUM_THREADS specifies the number of dot invocations doxygen is
 # allowed to run in parallel. When set to 0 (the default) doxygen will
@@ -1688,15 +1688,15 @@ INCLUDED_BY_GRAPH      = YES
 # the time of a run. So in most cases it will be better to enable call graphs
 # for selected functions only using the \callgraph command.
 
-CALL_GRAPH             = YES
+CALL_GRAPH             = NO
 
 # If the CALLER_GRAPH and HAVE_DOT tags are set to YES then
 # doxygen will generate a caller dependency graph for every global function
 # or class method. Note that enabling this option will significantly increase
 # the time of a run. So in most cases it will be better to enable caller
 # graphs for selected functions only using the \callergraph command.
 
-CALLER_GRAPH           = YES
+CALLER_GRAPH           = NO
 
 # If the GRAPHICAL_HIERARCHY and HAVE_DOT tags are set to YES then doxygen
 # will generate a graphical hierarchy of all classes instead of a textual one.