From 83338a2f753ebbb7ba7b154cc60b2f427b6121b0 Mon Sep 17 00:00:00 2001
From: Sebastiano Tronto <sebastiano@tronto.net>
Date: Wed, 26 Mar 2025 19:22:50 +0100
Subject: Added inverse coordinate for CO

this is necessary for DR coordinate
---
 src/arch/avx2.h     | 49 ++++++++++++++++++++++++++++++++++---------------
 src/arch/common.h   |  3 ++-
 src/arch/neon.h     | 19 +++++++++++++++++++
 src/arch/portable.h | 42 +++++++++++++++++++++++++++++-------------
 4 files changed, 84 insertions(+), 29 deletions(-)

(limited to 'src/arch')

diff --git a/src/arch/avx2.h b/src/arch/avx2.h
index ebcf84f..4682e33 100644
--- a/src/arch/avx2.h
+++ b/src/arch/avx2.h
@@ -188,6 +188,25 @@ coord_co(cube_t c)
 	return ret;
 }
 
+STATIC_INLINE cube_t
+invcoord_co(int64_t coord)
+{
+	int64_t i, c, p, co, mem[4] = {0};
+	cube_t cube, cc;
+
+	for (i = 0, p = 0, c = coord; i < 8; i++, c /= 3) {
+		co = i == 7 ? ((3 - (p % 3)) % 3) : (c % 3);
+		p += co;
+		mem[0] |= (int64_t)(i + (co << COSHIFT)) << (int64_t)(8 * i);
+	}
+
+	cc = _mm256_loadu_si256((const __m256i *)mem);
+	cube = SOLVED_CUBE;
+	copy_corners(&cube, cc);
+
+	return cube;
+}
+
 STATIC_INLINE int64_t
 coord_csep(cube_t c)
 {
@@ -251,6 +270,21 @@ coord_esep(cube_t c)
 	return ret1 * 70 + ret2;
 }
 
+STATIC_INLINE cube_t
+invcoord_esep(int64_t esep)
+{
+	cube_t eee, ret;
+	uint8_t mem[32] = {0};
+
+	invcoord_esep_array(esep % 70, esep / 70, mem+16);
+
+	ret = SOLVED_CUBE;
+	eee = _mm256_loadu_si256((__m256i_u *)&mem);
+	copy_edges(&ret, eee);
+
+	return ret;
+}
+
 STATIC_INLINE void
 copy_corners(cube_t dest[static 1], cube_t src)
 {
@@ -287,18 +321,3 @@ set_eo(cube_t cube[static 1], int64_t eo)
 	*cube = _mm256_andnot_si256(EO_AVX2, *cube);
 	*cube = _mm256_or_si256(*cube, veo);
 }
-
-STATIC_INLINE cube_t
-invcoord_esep(int64_t esep)
-{
-	cube_t eee, ret;
-	uint8_t mem[32] = {0};
-
-	invcoord_esep_array(esep % 70, esep / 70, mem+16);
-
-	ret = SOLVED_CUBE;
-	eee = _mm256_loadu_si256((__m256i_u *)&mem);
-	copy_edges(&ret, eee);
-
-	return ret;
-}
diff --git a/src/arch/common.h b/src/arch/common.h
index 5c1efde..e57fba9 100644
--- a/src/arch/common.h
+++ b/src/arch/common.h
@@ -9,15 +9,16 @@ STATIC_INLINE cube_t compose(cube_t, cube_t);
 STATIC_INLINE cube_t inverse(cube_t);
 
 STATIC_INLINE int64_t coord_co(cube_t);
+STATIC_INLINE cube_t invcoord_co(int64_t);
 STATIC_INLINE int64_t coord_csep(cube_t);
 STATIC_INLINE int64_t coord_cocsep(cube_t);
 STATIC_INLINE int64_t coord_eo(cube_t);
 STATIC_INLINE int64_t coord_esep(cube_t);
+STATIC_INLINE cube_t invcoord_esep(int64_t);
 
 STATIC_INLINE void copy_corners(cube_t [static 1], cube_t);
 STATIC_INLINE void copy_edges(cube_t [static 1], cube_t);
 STATIC_INLINE void set_eo(cube_t [static 1], int64_t);
-STATIC_INLINE cube_t invcoord_esep(int64_t);
 
 STATIC_INLINE void invcoord_esep_array(int64_t, int64_t, uint8_t[static 12]);
 STATIC_INLINE cube_t invcoord_eoesep(int64_t);
diff --git a/src/arch/neon.h b/src/arch/neon.h
index d5f5f8b..fc0e612 100644
--- a/src/arch/neon.h
+++ b/src/arch/neon.h
@@ -219,6 +219,25 @@ coord_co(cube_t c)
 	return ret;
 }
 
+STATIC_INLINE cube_t
+invcoord_co(int64_t coord)
+{
+	int64_t co, c, i, p;
+	uint8_t mem[8];
+	cube_t cube;
+
+	for (i = 0, p = 0, c = coord; i < 8; i++, c /= 3) {
+		co = i == 7 ? ((3 - (p % 3)) % 3) : (c % 3);
+		p += co;
+		mem[i] = i + (co << COSHIFT);
+	}
+
+	cube.corner = vld1_u8(mem);
+	cube.edge = SOLVED_CUBE.edge;
+
+	return cube;
+}
+
 STATIC_INLINE int64_t
 coord_csep(cube_t c)
 {
diff --git a/src/arch/portable.h b/src/arch/portable.h
index 4ac5fe7..56d3074 100644
--- a/src/arch/portable.h
+++ b/src/arch/portable.h
@@ -144,8 +144,7 @@ inverse(cube_t cube)
 STATIC_INLINE int64_t
 coord_co(cube_t c)
 {
-	int i, p;
-	int64_t ret;
+	int i, p, ret;
 
 	for (ret = 0, i = 0, p = 1; i < 7; i++, p *= 3)
 		ret += p * (c.corner[i] >> COSHIFT);
@@ -153,6 +152,23 @@ coord_co(cube_t c)
 	return ret;
 }
 
+STATIC_INLINE cube_t
+invcoord_co(int64_t coord)
+{
+	int64_t i, c, p;
+	cube_t cube;
+
+	cube = SOLVED_CUBE;
+	for (i = 0, p = 0, c = coord; i < 7; i++, c /= 3) {
+		p += c % 3;
+		cube.corner[i] |= (c % 3) << COSHIFT;
+	}
+
+	cube.corner[7] |= ((3 - (p % 3)) % 3) << COSHIFT;
+
+	return cube;
+}
+
 /*
 For corner separation, we consider the axis (a.k.a. tetrad) each
 corner belongs to as 0 or 1 and we translate this sequence into binary.
@@ -226,6 +242,17 @@ coord_esep(cube_t c)
 	return ret1 * 70 + ret2;
 }
 
+STATIC_INLINE cube_t
+invcoord_esep(int64_t esep)
+{
+	cube_t ret;
+
+	ret = SOLVED_CUBE;
+	invcoord_esep_array(esep % 70, esep / 70, ret.edge);
+
+	return ret;
+}
+
 STATIC_INLINE void
 copy_corners(cube_t dest[static 1], cube_t src)
 {
@@ -250,14 +277,3 @@ set_eo(cube_t cube[static 1], int64_t eo)
 	}
 	cube->edge[0] = (cube->edge[0] & ~EOBIT) | (EOBIT * (sum % 2));
 }
-
-STATIC_INLINE cube_t
-invcoord_esep(int64_t esep)
-{
-	cube_t ret;
-
-	ret = SOLVED_CUBE;
-	invcoord_esep_array(esep % 70, esep / 70, ret.edge);
-
-	return ret;
-}
-- 
cgit v1.3