From b6ab832ed4aee2c55f5b66a0227159c763a0f353 Mon Sep 17 00:00:00 2001
From: Bill Currie <bill@taniwha.org>
Date: Sun, 28 Mar 2021 19:49:43 +0900
Subject: [PATCH] [simd] Add vabsf and some more tests

---
 include/QF/simd/vec4f.h    | 14 +++++++++++++
 libs/util/test/test-simd.c | 42 ++++++++++++++++++++++++++++++++++++++
 2 files changed, 56 insertions(+)

diff --git a/include/QF/simd/vec4f.h b/include/QF/simd/vec4f.h
index 04ee24c94..8ae4cbd5e 100644
--- a/include/QF/simd/vec4f.h
+++ b/include/QF/simd/vec4f.h
@@ -32,6 +32,7 @@
 
 #include "QF/simd/types.h"
 
+GNU89INLINE inline vec4f_t vabsf (vec4f_t v) __attribute__((const));
 GNU89INLINE inline vec4f_t vsqrtf (vec4f_t v) __attribute__((const));
 GNU89INLINE inline vec4f_t vceilf (vec4f_t v) __attribute__((const));
 GNU89INLINE inline vec4f_t vfloorf (vec4f_t v) __attribute__((const));
@@ -97,6 +98,19 @@ GNU89INLINE inline vec4f_t normalf (vec4f_t v) __attribute__((pure));
 GNU89INLINE inline vec4f_t magnitudef (vec4f_t v) __attribute__((pure));
 GNU89INLINE inline vec4f_t magnitude3f (vec4f_t v) __attribute__((pure));
 
+#ifndef IMPLEMENT_VEC4F_Funcs
+GNU89INLINE inline
+#else
+VISIBLE
+#endif
+vec4f_t
+vabsf (vec4f_t v)
+{
+	const uint32_t  nan = ~0u >> 1;
+	const vec4i_t   abs = { nan, nan, nan, nan };
+	return _mm_and_ps (v, (__m128) abs);
+}
+
 #ifndef IMPLEMENT_VEC4F_Funcs
 GNU89INLINE inline
 #else
diff --git a/libs/util/test/test-simd.c b/libs/util/test/test-simd.c
index b14a1be9c..ebaab413f 100644
--- a/libs/util/test/test-simd.c
+++ b/libs/util/test/test-simd.c
@@ -27,6 +27,9 @@
 #define none     { -1, -1, -1, -1 }
 #define nqident  {  0,  0,  0, -1 }
 
+#define pi       {  M_PI, M_PI,  M_PI, M_PI } // lots of bits
+#define pmpi     { -M_PI, M_PI, -M_PI, M_PI } // lots of bits
+
 #define identity \
 	{	{ 1, 0, 0, 0 }, \
 		{ 0, 1, 0, 0 }, \
@@ -124,6 +127,26 @@ static vec4f_t tqconjf (vec4f_t v, vec4f_t ignore)
 	return qconjf (v);
 }
 
+static vec4f_t tvabsf (vec4f_t v, vec4f_t ignore)
+{
+	return vabsf (v);
+}
+
+static vec4f_t tvsqrtf (vec4f_t v, vec4f_t ignore)
+{
+	return vsqrtf (v);
+}
+
+static vec4f_t tmagnitudef (vec4f_t v, vec4f_t ignore)
+{
+	return magnitudef (v);
+}
+
+static vec4f_t tmagnitude3f (vec4f_t v, vec4f_t ignore)
+{
+	return magnitude3f (v);
+}
+
 static vec4d_test_t vec4d_tests[] = {
 	// 3D dot products
 	{ dotd, right,   right,   one  },
@@ -348,10 +371,29 @@ static vec4f_test_t vec4f_tests[] = {
 	{ qrotf, up,      forward,  { -s05,    0,    0,  s05 } },
 	{ qrotf, up,      up,       qident },
 
+	{ tvabsf,    pmpi, {}, pi },
+	{ tvsqrtf,  { 1, 4, 9, 16}, {}, {1, 2, 3, 4} },
 	{ tvtruncf, { 1.1, 2.9, -1.1, -2.9 }, {}, { 1, 2, -1, -2 } },
 	{ tvceilf,  { 1.1, 2.9, -1.1, -2.9 }, {}, { 2, 3, -1, -2 } },
 	{ tvfloorf, { 1.1, 2.9, -1.1, -2.9 }, {}, { 1, 2, -2, -3 } },
 	{ tqconjf,  one, {}, { -1, -1, -1, 1 } },
+	{ tmagnitudef,  {  3,  4,  12,  84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  {  3,  4,  12, -84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  {  3,  4, -12,  84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  {  3,  4, -12, -84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  {  3, -4,  12,  84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  {  3, -4,  12, -84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  {  3, -4, -12,  84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  {  3, -4, -12, -84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  { -3,  4,  12,  84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  { -3,  4,  12, -84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  { -3,  4, -12,  84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  { -3,  4, -12, -84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  { -3, -4,  12,  84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  { -3, -4,  12, -84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  { -3, -4, -12,  84}, {}, {85, 85, 85, 85} },
+	{ tmagnitudef,  { -3, -4, -12, -84}, {}, {85, 85, 85, 85} },
+	{ tmagnitude3f, { -3, -4, -12, -84}, {}, {13, 13, 13, 13} },
 };
 #define num_vec4f_tests (sizeof (vec4f_tests) / (sizeof (vec4f_tests[0])))