From 3b159cdf021f9e77ba3edc202ffa9e57615ff76a Mon Sep 17 00:00:00 2001
From: Recep Aslantas <info@recp.me>
Date: Tue, 11 Jul 2017 16:33:41 +0300
Subject: [PATCH] neon: implement matrix multiplication for ARM Neon

---
 .gitignore                    |  3 ++
 include/arch/simd/neon/mat4.h | 57 +++++++++++++++++++++++++++++++++++
 include/cglm-mat.h            | 16 ++++++++--
 3 files changed, 74 insertions(+), 2 deletions(-)
 create mode 100644 include/arch/simd/neon/mat4.h

diff --git a/.gitignore b/.gitignore
index 96858dc..09c8746 100644
--- a/.gitignore
+++ b/.gitignore
@@ -54,3 +54,6 @@ test/test_*
 test-*
 test/.libs/*
 test/tests
+cglm_arm/*
+cglm_test_ios/*
+cglm_test_iosTests/*
\ No newline at end of file
diff --git a/include/arch/simd/neon/mat4.h b/include/arch/simd/neon/mat4.h
new file mode 100644
index 0000000..8694ea2
--- /dev/null
+++ b/include/arch/simd/neon/mat4.h
@@ -0,0 +1,57 @@
+/*
+ * Copyright (c), Recep Aslantas.
+ *
+ * MIT License (MIT), http://opensource.org/licenses/MIT
+ * Full license can be found in the LICENSE file
+ */
+
+#ifndef cglm_mat4_neon_h
+#define cglm_mat4_neon_h
+#if defined(__ARM_NEON_FP)
+
+#include "../../../cglm-common.h"
+#include <arm_neon.h>
+
+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;
+
+  l0 = vld1q_f32(m2[0]);
+  l1 = vld1q_f32(m2[1]);
+  l2 = vld1q_f32(m2[2]);
+  l3 = vld1q_f32(m2[3]);
+
+  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);
+
+  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);
+
+  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);
+
+  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);
+
+  vst1q_f32(dest[0], d0);
+  vst1q_f32(dest[1], d1);
+  vst1q_f32(dest[2], d2);
+  vst1q_f32(dest[3], d3);
+}
+
+#endif
+#endif /* cglm_mat4_neon_h */
diff --git a/include/cglm-mat.h b/include/cglm-mat.h
index 6ced3a4..c5dad6a 100644
--- a/include/cglm-mat.h
+++ b/include/cglm-mat.h
@@ -45,8 +45,18 @@
 #define cglm_mat_h
 
 #include "cglm-common.h"
-#include "arch/simd/cglm-mat-simd-sse2.h"
-#include "arch/simd/cglm-mat-simd-avx.h"
+
+#ifdef CGLM_SSE_FP
+#  include "arch/simd/cglm-mat-simd-sse2.h"
+#endif
+
+#ifdef CGLM_AVX_FP
+#  include "arch/simd/cglm-mat-simd-avx.h"
+#endif
+
+#ifdef CGLM_NEON_FP
+#  include "arch/simd/neon/mat4.h"
+#endif
 
 #include <assert.h>
 
@@ -215,6 +225,8 @@ glm_mat4_mul(mat4 m1, mat4 m2, mat4 dest) {
   glm_mat4_mul_avx(m1, m2, dest);
 #elif defined( __SSE__ ) || defined( __SSE2__ )
   glm_mat4_mul_sse2(m1, m2, dest);
+#elif defined( __ARM_NEON_FP )
+  glm_mat4_mul_neon(m1, m2, dest);
 #else
   float a00 = m1[0][0], a01 = m1[0][1], a02 = m1[0][2], a03 = m1[0][3],
         a10 = m1[1][0], a11 = m1[1][1], a12 = m1[1][2], a13 = m1[1][3],