From 0ab50f720809e91c1d4f1edc5849ed0fc920d871 Mon Sep 17 00:00:00 2001
From: Recep Aslantas <info@recp.me>
Date: Sun, 25 Apr 2021 03:41:39 +0300
Subject: [PATCH] arm, neon: update mat4_mul to use FMA

---
 include/cglm/simd/arm.h       |  2 +-
 include/cglm/simd/neon/mat4.h | 57 +++++++++++++++++------------------
 2 files changed, 28 insertions(+), 31 deletions(-)

diff --git a/include/cglm/simd/arm.h b/include/cglm/simd/arm.h
index a7009e5..e943ce3 100644
--- a/include/cglm/simd/arm.h
+++ b/include/cglm/simd/arm.h
@@ -123,7 +123,7 @@ static inline
 float32x4_t
 glmm_fmadd(float32x4_t a, float32x4_t b, float32x4_t c) {
 #if CGLM_ARM64
-  return vfmaq_f32(c, a, b);
+  return vfmaq_f32(c, a, b); /* why vfmaq_f32 is slower than vmlaq_f32 ??? */
 #else
   return vmlaq_f32(c, a, b);
 #endif
diff --git a/include/cglm/simd/neon/mat4.h b/include/cglm/simd/neon/mat4.h
index 2fecebe..f1b3b60 100644
--- a/include/cglm/simd/neon/mat4.h
+++ b/include/cglm/simd/neon/mat4.h
@@ -41,42 +41,39 @@ glm_mat4_transp_neon(mat4 m, mat4 dest) {
 CGLM_INLINE
 void
 glm_mat4_mul_neon(mat4 m1, mat4 m2, mat4 dest) {
-  /* D = R * L (Column-Major) */
-  float32x4_t l0, l1, l2, l3, r, d0, d1, d2, d3;
+   /* D = R * L (Column-Major) */
 
-  l0 = vld1q_f32(m2[0]);
-  l1 = vld1q_f32(m2[1]);
-  l2 = vld1q_f32(m2[2]);
-  l3 = vld1q_f32(m2[3]);
+  glmm_128 l0, l1, l2, l3, r0, r1, r2, r3, v0, v1, v2, v3;
 
-  r  = vld1q_f32(m1[0]);
-  d0 = vmulq_lane_f32(r, vget_low_f32(l0), 0);
-  d1 = vmulq_lane_f32(r, vget_low_f32(l1), 0);
-  d2 = vmulq_lane_f32(r, vget_low_f32(l2), 0);
-  d3 = vmulq_lane_f32(r, vget_low_f32(l3), 0);
+  l0 = glmm_load(m1[0]);    r0 = glmm_load(m2[0]);
+  l1 = glmm_load(m1[1]);    r1 = glmm_load(m2[1]);
+  l2 = glmm_load(m1[2]);    r2 = glmm_load(m2[2]);
+  l3 = glmm_load(m1[3]);    r3 = glmm_load(m2[3]);
 
-  r  = vld1q_f32(m1[1]);
-  d0 = vmlaq_lane_f32(d0, r, vget_low_f32(l0), 1);
-  d1 = vmlaq_lane_f32(d1, r, vget_low_f32(l1), 1);
-  d2 = vmlaq_lane_f32(d2, r, vget_low_f32(l2), 1);
-  d3 = vmlaq_lane_f32(d3, r, vget_low_f32(l3), 1);
+  v0 = vmulq_f32(glmm_splat_x(r0), l0);
+  v1 = vmulq_f32(glmm_splat_x(r1), l0);
+  v2 = vmulq_f32(glmm_splat_x(r2), l0);
+  v3 = vmulq_f32(glmm_splat_x(r3), l0);
 
-  r  = vld1q_f32(m1[2]);
-  d0 = vmlaq_lane_f32(d0, r, vget_high_f32(l0), 0);
-  d1 = vmlaq_lane_f32(d1, r, vget_high_f32(l1), 0);
-  d2 = vmlaq_lane_f32(d2, r, vget_high_f32(l2), 0);
-  d3 = vmlaq_lane_f32(d3, r, vget_high_f32(l3), 0);
+  v0 = glmm_fmadd(glmm_splat_y(r0), l1, v0);
+  v1 = glmm_fmadd(glmm_splat_y(r1), l1, v1);
+  v2 = glmm_fmadd(glmm_splat_y(r2), l1, v2);
+  v3 = glmm_fmadd(glmm_splat_y(r3), l1, v3);
 
-  r  = vld1q_f32(m1[3]);
-  d0 = vmlaq_lane_f32(d0, r, vget_high_f32(l0), 1);
-  d1 = vmlaq_lane_f32(d1, r, vget_high_f32(l1), 1);
-  d2 = vmlaq_lane_f32(d2, r, vget_high_f32(l2), 1);
-  d3 = vmlaq_lane_f32(d3, r, vget_high_f32(l3), 1);
+  v0 = glmm_fmadd(glmm_splat_z(r0), l2, v0);
+  v1 = glmm_fmadd(glmm_splat_z(r1), l2, v1);
+  v2 = glmm_fmadd(glmm_splat_z(r2), l2, v2);
+  v3 = glmm_fmadd(glmm_splat_z(r3), l2, v3);
 
-  vst1q_f32(dest[0], d0);
-  vst1q_f32(dest[1], d1);
-  vst1q_f32(dest[2], d2);
-  vst1q_f32(dest[3], d3);
+  v0 = glmm_fmadd(glmm_splat_w(r0), l3, v0);
+  v1 = glmm_fmadd(glmm_splat_w(r1), l3, v1);
+  v2 = glmm_fmadd(glmm_splat_w(r2), l3, v2);
+  v3 = glmm_fmadd(glmm_splat_w(r3), l3, v3);
+
+  glmm_store(dest[0], v0);
+  glmm_store(dest[1], v1);
+  glmm_store(dest[2], v2);
+  glmm_store(dest[3], v3);
 }
 
 CGLM_INLINE