h48

commit 2d66a0ffea4b03cd27b0f4363f8c4b8b0a1242e4
parent 18c9a8b8905304cf5f8fc15825769046a3144866
Author: Sebastiano Tronto <sebastiano@tronto.net>
Date:   Sun, 18 Aug 2024 15:31:06 +0200

Removed duplication

Diffstat:
M
src/arch/avx2.h
 | 
25
++-----------------------
M
src/arch/common.h
 | 
28
++++++++++++++++++++++++++++
M
src/arch/neon.h
 | 
29
++++-------------------------
M
src/arch/portable.h
 | 
23
+----------------------

4 files changed, 35 insertions(+), 70 deletions(-)
diff --git a/src/arch/avx2.h b/src/arch/avx2.h
@@ -286,30 +286,9 @@ _static_inline cube_t
 invcoord_esep(int64_t esep)
 {
 	cube_t eee, ret;
-	int64_t bit1, bit2, i, j, jj, k, l, s, v, w, is1, set1, set2;
-	uint8_t mem[32];
-	uint8_t slice[3] = {0};
+	uint8_t mem[32] = {0};
 
-	set1 = esep % 70;
-	set2 = esep / 70;
-
-	for (i = 0, j = 0, k = 4, l = 4; i < 12; i++) {
-		v = binomial[11-i][k];
-		jj = j < 8;
-		w = jj * binomial[7-(j*jj)][l];
-		bit2 = set2 >= v;
-		bit1 = set1 >= w;
-		is1 = (1 - bit2) * bit1;
-
-		set2 -= bit2 * v;
-		k -= bit2;
-		set1 -= is1 * w;
-		l -= is1;
-		j += (1-bit2);
-		s = 2*bit2 + (1-bit2)*bit1;
-
-		mem[i+16] = (slice[s]++) | (uint8_t)(s << 2);
-	}
+	invcoord_esep_array(esep % 70, esep / 70, mem+16);
 
 	ret = solved;
 	eee = _mm256_loadu_si256((__m256i_u *)&mem);
diff --git a/src/arch/common.h b/src/arch/common.h
@@ -17,3 +17,31 @@ _static_inline void copy_corners(cube_t *, cube_t);
 _static_inline void copy_edges(cube_t *, cube_t);
 _static_inline void set_eo(cube_t *, 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 void
+invcoord_esep_array(int64_t set1, int64_t set2, uint8_t mem[static 12])
+{
+	int64_t bit1, bit2, i, j, jj, k, l, s, v, w, is1;
+	uint8_t slice[3] = {0};
+
+	for (i = 0, j = 0, k = 4, l = 4; i < 12; i++)
+	{
+		v = binomial[11 - i][k];
+		jj = j < 8;
+		w = jj * binomial[7 - (j * jj)][l];
+		bit2 = set2 >= v;
+		bit1 = set1 >= w;
+		is1 = (1 - bit2) * bit1;
+
+		set2 -= bit2 * v;
+		k -= bit2;
+		set1 -= is1 * w;
+		l -= is1;
+		j += (1 - bit2);
+		s = 2 * bit2 + (1 - bit2) * bit1;
+
+		mem[i] = (slice[s]++) | (uint8_t)(s << 2);
+	}
+}
diff --git a/src/arch/neon.h b/src/arch/neon.h
@@ -316,33 +316,12 @@ _static_inline cube_t
 invcoord_esep(int64_t esep)
 {
 	cube_t ret;
-	int64_t bit1, bit2, i, j, jj, k, l, s, v, w, is1, set1, set2;
-	uint8_t slice[3] = {0};
+	uint8_t mem[16] = {0};
 
-	ret = solved;
-	uint8_t mem[16];
-	set1 = esep % 70;
-	set2 = esep / 70;
-
-	for (i = 0, j = 0, k = 4, l = 4; i < 12; i++)
-	{
-		v = binomial[11 - i][k];
-		jj = j < 8;
-		w = jj * binomial[7 - (j * jj)][l];
-		bit2 = set2 >= v;
-		bit1 = set1 >= w;
-		is1 = (1 - bit2) * bit1;
-
-		set2 -= bit2 * v;
-		k -= bit2;
-		set1 -= is1 * w;
-		l -= is1;
-		j += (1 - bit2);
-		s = 2 * bit2 + (1 - bit2) * bit1;
-
-		mem[i] = (slice[s]++) | (uint8_t)(s << 2);
-	}
+	invcoord_esep_array(esep % 70, esep / 70, mem);
 
+	ret = solved;
 	ret.edge = vld1q_u8(mem);
+
 	return ret;
 }
diff --git a/src/arch/portable.h b/src/arch/portable.h
@@ -243,30 +243,9 @@ _static_inline cube_t
 invcoord_esep(int64_t esep)
 {
 	cube_t ret;
-	int64_t bit1, bit2, i, j, jj, k, l, s, v, w, is1, set1, set2;
-	uint8_t slice[3] = {0};
 
 	ret = solved;
-	set1 = esep % 70;
-	set2 = esep / 70;
-
-	for (i = 0, j = 0, k = 4, l = 4; i < 12; i++) {
-		v = binomial[11-i][k];
-		jj = j < 8;
-		w = jj * binomial[7-(j*jj)][l];
-		bit2 = set2 >= v;
-		bit1 = set1 >= w;
-		is1 = (1 - bit2) * bit1;
-
-		set2 -= bit2 * v;
-		k -= bit2;
-		set1 -= is1 * w;
-		l -= is1;
-		j += (1-bit2);
-		s = 2*bit2 + (1-bit2)*bit1;
-
-		ret.edge[i] = (slice[s]++) | (uint8_t)(s << 2);
-	}
+	invcoord_esep_array(esep % 70, esep / 70, ret.edge);
 
 	return ret;
 }