From f65f1d491bec03b6de58fb12918e7083f7f5a1ce Mon Sep 17 00:00:00 2001
From: Recep Aslantas <info@recp.me>
Date: Tue, 22 Jan 2019 09:23:51 +0300
Subject: [PATCH] simd: optimize vec4_distance with sse and neon

---
 include/cglm/vec4.h | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/include/cglm/vec4.h b/include/cglm/vec4.h
index 5d0e466..03dc405 100644
--- a/include/cglm/vec4.h
+++ b/include/cglm/vec4.h
@@ -707,10 +707,25 @@ glm_vec4_normalize(vec4 v) {
 CGLM_INLINE
 float
 glm_vec4_distance(vec4 a, vec4 b) {
+#if defined( __SSE__ ) || defined( __SSE2__ )
+  __m128 x0;
+  x0 = _mm_sub_ps(glmm_load(b), glmm_load(a));
+  x0 = _mm_mul_ps(x0, x0);
+  x0 = _mm_add_ps(x0, glmm_shuff1(x0, 1, 0, 3, 2));
+  return _mm_cvtss_f32(_mm_sqrt_ss(_mm_add_ss(x0,
+                                              glmm_shuff1(x0, 0, 1, 0, 1))));
+#elif defined(CGLM_NEON_FP)
+  float32x4_t v0;
+  float32_t   r;
+  v0 = vsubq_f32(vld1q_f32(a), vld1q_f32(b));
+  r  = vaddvq_f32(vmulq_f32(v0, v0));
+  return sqrtf(r);
+#else
   return sqrtf(glm_pow2(b[0] - a[0])
              + glm_pow2(b[1] - a[1])
              + glm_pow2(b[2] - a[2])
              + glm_pow2(b[3] - a[3]));
+#endif
 }
 
 /*!