From 80c5e2c3f63cd0872c299cb2153e0531898bee3a Mon Sep 17 00:00:00 2001
From: Bill Currie <bill@taniwha.org>
Date: Thu, 6 Jan 2022 18:06:56 +0900
Subject: [PATCH] [simd] Remove requirements for AVX2 for vec4d

It seems gcc-11 does a pretty good job of emulating the instructions (it
no longer requires avx2 for 256-bit wide vectors).
---
 include/QF/simd/types.h    |  2 --
 include/QF/simd/vec4d.h    | 18 +++++++-----------
 libs/util/test/test-simd.c | 10 ----------
 3 files changed, 7 insertions(+), 23 deletions(-)

diff --git a/include/QF/simd/types.h b/include/QF/simd/types.h
index ad8345f44..1f2a511c1 100644
--- a/include/QF/simd/types.h
+++ b/include/QF/simd/types.h
@@ -43,7 +43,6 @@ typedef double vec3d_t[3];
 VEC_TYPE (double, vec2d_t, 2);
 VEC_TYPE (int64_t, vec2l_t, 2);
 
-#ifdef __AVX2__
 /** Four element vector type for horizontal (AOS) vector data.
  *
  * This is used for both vectors (3D and 4D) and quaternions. 3D vectors
@@ -58,7 +57,6 @@ VEC_TYPE (double, vec4d_t, 4);
 /** Used mostly for __builtin_shuffle.
  */
 VEC_TYPE (int64_t, vec4l_t, 4);
-#endif
 
 /** Three element vector type for interfacing with compact data.
  *
diff --git a/include/QF/simd/vec4d.h b/include/QF/simd/vec4d.h
index d0e3f19bc..1e1d7bb89 100644
--- a/include/QF/simd/vec4d.h
+++ b/include/QF/simd/vec4d.h
@@ -28,7 +28,6 @@
 #ifndef __QF_simd_vec4d_h
 #define __QF_simd_vec4d_h
 
-#ifdef __AVX2__
 #include <immintrin.h>
 
 #include "QF/simd/types.h"
@@ -96,9 +95,9 @@ GNU89INLINE inline vec4d_t qrotd (vec4d_t a, vec4d_t b) __attribute__((const));
 GNU89INLINE inline vec4d_t qconjd (vec4d_t q) __attribute__((const));
 GNU89INLINE inline vec4d_t loadvec3d (const double v3[]) __attribute__((pure));
 GNU89INLINE inline void storevec3d (double v3[3], vec4d_t v4);
-GNU89INLINE inline vec4l_t loadvec3l (const long *v3) __attribute__((pure));
-GNU89INLINE inline vec4l_t loadvec3l1 (const long *v3) __attribute__((pure));
-GNU89INLINE inline void storevec3l (long *v3, vec4l_t v4);
+GNU89INLINE inline vec4l_t loadvec3l (const int64_t *v3) __attribute__((pure));
+GNU89INLINE inline vec4l_t loadvec3l1 (const int64_t *v3) __attribute__((pure));
+GNU89INLINE inline void storevec3l (int64_t *v3, vec4l_t v4);
 
 #ifndef IMPLEMENT_VEC4D_Funcs
 GNU89INLINE inline
@@ -187,8 +186,7 @@ qmuld (vec4d_t a, vec4d_t b)
 	vec4d_t c = crossd (a, b) + a * b[3] + a[3] * b;
 	vec4d_t d = dotd (a, b);
 	// zero out the vector component of dot product so only the scalar remains
-	d = _mm256_permute2f128_pd (d, d, 0x18);
-	d = _mm256_permute4x64_pd (d, 0xc0);
+	d = (vec4d_t) { 0, 0, 0, d[3] };
 	return c - d;
 }
 
@@ -302,7 +300,7 @@ GNU89INLINE inline
 VISIBLE
 #endif
 vec4l_t
-loadvec3l (const long *v3)
+loadvec3l (const int64_t *v3)
 {
 	vec4l_t v4 = { v3[0], v3[1], v3[2], 0 };
 	return v4;
@@ -314,7 +312,7 @@ GNU89INLINE inline
 VISIBLE
 #endif
 vec4l_t
-loadvec3l1 (const long *v3)
+loadvec3l1 (const int64_t *v3)
 {
 	vec4l_t v4 = { v3[0], v3[1], v3[2], 1 };
 	return v4;
@@ -326,13 +324,11 @@ GNU89INLINE inline
 VISIBLE
 #endif
 void
-storevec3l (long *v3, vec4l_t v4)
+storevec3l (int64_t *v3, vec4l_t v4)
 {
 	v3[0] = v4[0];
 	v3[1] = v4[1];
 	v3[2] = v4[2];
 }
 
-#endif
-
 #endif//__QF_simd_vec4d_h
diff --git a/libs/util/test/test-simd.c b/libs/util/test/test-simd.c
index d1b4bf286..7cba48f3b 100644
--- a/libs/util/test/test-simd.c
+++ b/libs/util/test/test-simd.c
@@ -48,7 +48,6 @@
 
 #define s05 0.70710678118654757
 
-#ifdef __AVX2__
 typedef  struct {
 	int         line;
 	vec4d_t   (*op) (vec4d_t a, vec4d_t b);
@@ -57,7 +56,6 @@ typedef  struct {
 	vec4d_t     expect;
 	vec4d_t     ulp_errors;
 } vec4d_test_t;
-#endif
 
 typedef  struct {
 	int         line;
@@ -94,7 +92,6 @@ typedef  struct {
 	mat4f_t     ulp_errors;
 } mq4f_test_t;
 
-#ifdef __AVX2__
 static vec4d_t tvtruncd (vec4d_t v, vec4d_t ignore)
 {
 	return vtrunc4d (v);
@@ -114,7 +111,6 @@ static vec4d_t tqconjd (vec4d_t v, vec4d_t ignore)
 {
 	return qconjd (v);
 }
-#endif
 
 static vec4f_t tvtruncf (vec4f_t v, vec4f_t ignore)
 {
@@ -158,7 +154,6 @@ static vec4f_t tmagnitude3f (vec4f_t v, vec4f_t ignore)
 
 #define T(t...) { __LINE__, t }
 
-#ifdef __AVX2__
 static vec4d_test_t vec4d_tests[] = {
 	// 3D dot products
 	T(dotd, right,   right,   one  ),
@@ -285,7 +280,6 @@ static vec4d_test_t vec4d_tests[] = {
 	T(tqconjd,  one, {}, { -1, -1, -1, 1 } ),
 };
 #define num_vec4d_tests (sizeof (vec4d_tests) / (sizeof (vec4d_tests[0])))
-#endif
 
 static vec4f_test_t vec4f_tests[] = {
 	// 3D dot products
@@ -487,7 +481,6 @@ static mq4f_test_t mq4f_tests[] = {
 };
 #define num_mq4f_tests (sizeof (mq4f_tests) / (sizeof (mq4f_tests[0])))
 
-#ifdef __AVX2__
 static int
 run_vec4d_tests (void)
 {
@@ -512,7 +505,6 @@ run_vec4d_tests (void)
 	}
 	return ret;
 }
-#endif
 
 static int
 run_vec4f_tests (void)
@@ -684,9 +676,7 @@ int
 main (void)
 {
 	int         ret = 0;
-#ifdef __AVX2__
 	ret |= run_vec4d_tests ();
-#endif
 	ret |= run_vec4f_tests ();
 	ret |= run_mat4f_tests ();
 	ret |= run_mv4f_tests ();