Merge pull request #39 from okuhara/neon_vaddvq

Resurrect neon vaddvq version of count_last_flip
abulmo · Dec 7, 2024 · 8dbf3b6 · 8dbf3b6
2 parents dbcac52 + 9635daf
commit 8dbf3b6
Show file tree

Hide file tree

Showing 3 changed files with 24 additions and 10 deletions.
diff --git a/.github/workflows/build.yaml b/.github/workflows/build.yaml
@@ -32,7 +32,7 @@ jobs:
         cd src
         ${{ matrix.build_command }}
 
-    - uses: actions/upload-artifact@v2
+    - uses: actions/upload-artifact@v4
       with:
         name: artifact_${{ runner.os }}
         path: bin
diff --git a/src/board.c b/src/board.c
@@ -412,7 +412,7 @@ static uint64x2_t transpose_neon(const uint64x2_t board)
  */
 void board_horizontal_mirror(const Board *board, Board *sym)
 {
-#if defined(__ARM_NEON)
+#if USE_SIMD & defined(__ARM_NEON)
 	vst1q_u64((uint64_t *) sym, horizontal_mirror_neon(vld1q_u64((uint64_t *) board)));
 #else
 	sym->player = horizontal_mirror(board->player);
@@ -428,7 +428,7 @@ void board_horizontal_mirror(const Board *board, Board *sym)
  */
 void board_vertical_mirror(const Board *board, Board *sym)
 {
-#if defined(__ARM_NEON)
+#if USE_SIMD & defined(__ARM_NEON)
 	vst1q_u64((uint64_t *) sym, vertical_mirror_neon(vld1q_u64((uint64_t *) board)));
 #else
 	sym->player = vertical_mirror(board->player);
@@ -716,7 +716,7 @@ static inline uint64_t get_some_moves(const uint64_t P, const uint64_t mask, con
 	const uint64x2_t PP = vdupq_n_u64(P);
 	const uint64x2_t MM = vdupq_n_u64(mask);
 	const uint64x2_t moves = get_some_moves_neon(PP, MM, dir);
-	return 	vget_low_u64(moves) | vget_high_u64(moves);
+	return 	vgetq_lane_u64(moves, 0) | vgetq_lane_u64(moves, 1);
 
 #elif PARALLEL_PREFIX & 1
 
@@ -831,7 +831,7 @@ uint64_t get_moves(const uint64_t P, const uint64_t O)
 	moves = vorrq_u64(moves, get_some_moves_neon(PP, OO, 8));    // vertical
 	moves = vorrq_u64(moves, get_some_moves_neon(PP, MM, 7)); // diagonals
 	moves = vorrq_u64(moves, get_some_moves_neon(PP, MM, 9));
-	return (vget_low_u64(moves) | vget_high_u64(moves)) & E; // mask with empties
+	return (vgetq_lane_u64(moves, 0) | vgetq_lane_u64(moves, 1)) & E; // mask with empties
 
 #else
 

diff --git a/src/count_last_flip_neon.c b/src/count_last_flip_neon.c
@@ -17,16 +17,18 @@
  * For optimization purpose, the value returned is twice the number of flipped
  * disc, to facilitate the computation of disc difference.
  *
- * @date 1998 - 2023
+ * @date 1998 - 2024
  * @author Richard Delorme
  * @author Toshihiko Okuhara
- * @version 4.5
+ * @version 4.6
  * 
  */
 
 #include <arm_neon.h>
 #include <stdio.h>
 
+#define	COUNT_LAST_FLIP_NEON_VADDVQ
+
 /** precomputed count flip array */
 const unsigned char COUNT_FLIP[8][256] = {
 	{
@@ -111,7 +113,7 @@ const unsigned char COUNT_FLIP[8][256] = {
 	},
 };
 
-#ifdef HAS_CPU_64
+#ifdef COUNT_LAST_FLIP_NEON_VADDVQ
 /* bit masks for diagonal lines (interleaved) */
 const uint64x2_t mask_dvhd[64][2] = {
 	{{ 0x000000000000ff01, 0x0000000000000000 }, { 0x0801040102010101, 0x8001400120011001 }},
@@ -264,9 +266,21 @@ int count_last_flip(int pos, uint64_t P)
 	const unsigned char *COUNT_FLIP_Y = COUNT_FLIP[pos >> 3];
 	uint64x2_t	PP = vdupq_n_u64(P);
 	uint64x2_t	II;
+#ifdef COUNT_LAST_FLIP_NEON_VADDVQ	// vaddvq
+	unsigned int t;
+	const uint64x2_t dmask = { 0x0808040402020101, 0x8080404020201010 };
 
-// removed the buggy vaddvq code
+	PP = vreinterpretq_u64_u8(vzip1q_u8(vreinterpretq_u8_u64(PP), vreinterpretq_u8_u64(PP)));
+	II = vandq_u64(PP, mask_dvhd[pos][0]);	// 2 dirs interleaved
+	t = vaddvq_u16(vreinterpretq_u16_u64(II));
+	n_flips  = COUNT_FLIP_X[t >> 8];
+	n_flips += COUNT_FLIP_X[t & 0xFF];
+	II = vandq_u64(vreinterpretq_u64_u8(vtstq_u8(vreinterpretq_u8_u64(PP), vreinterpretq_u8_u64(mask_dvhd[pos][1]))), dmask);
+	t = vaddvq_u16(vreinterpretq_u16_u64(II));
+	n_flips += COUNT_FLIP_Y[t >> 8];
+	n_flips += COUNT_FLIP_Y[t & 0xFF];
 
+#else // Neon kindergarten
 	const uint64x2_t dmask = { 0x1020408001020408, 0x1020408001020408 };
 	uint64x2_t	PP = vdupq_n_u64(P);
 	n_flips = 0;
@@ -277,6 +291,6 @@ int count_last_flip(int pos, uint64_t P)
 	II = vpaddlq_u32(vmulq_u32(vreinterpretq_u32_u64(dmask), vreinterpretq_u32_u64(II)));
 	n_flips += COUNT_FLIP_Y[vgetq_lane_u8(vreinterpretq_u8_u64(II), 11)];
 	n_flips += COUNT_FLIP_Y[vgetq_lane_u8(vreinterpretq_u8_u64(II), 3)];
-
+#endif
 	return n_flips;
 }