Merge pull request #72 from recp/simd-update

SIMD update (NEON, SSE3, SSE4) + Features

.gitignore

@@ -69,3 +69,4 @@ win/cglm_test_*
 win/x64
 win/x85
 win/Debug
+cglm-test-ios*

CREDITS

@@ -52,3 +52,12 @@ https://gamedev.stackexchange.com/questions/28395/rotating-vector3-by-a-quaterni
 
 9. Sphere AABB intersect
 https://github.com/erich666/GraphicsGems/blob/master/gems/BoxSphere.c
+
+10. Horizontal add
+https://stackoverflow.com/questions/6996764/fastest-way-to-do-horizontal-float-vector-sum-on-x86
+
+11. de casteljau implementation and comments
+https://forums.khronos.org/showthread.php/10264-Animations-in-1-4-1-release-notes-revision-A/page2?highlight=bezier
+https://forums.khronos.org/showthread.php/10644-Animation-Bezier-interpolation
+https://forums.khronos.org/showthread.php/10387-2D-Tangents-in-Bezier-Splines?p=34164&viewfull=1#post34164
+https://forums.khronos.org/showthread.php/10651-Animation-TCB-Spline-Interpolation-in-COLLADA?highlight=bezier

README.md

@@ -82,7 +82,11 @@ Currently *cglm* uses default clip space configuration (-1, 1) for camera functi
 - inline or pre-compiled function call
 - frustum (extract view frustum planes, corners...)
 - bounding box  (AABB in Frustum (culling), crop, merge...)
+- bounding sphere
 - project, unproject
+- easing functions
+- curves
+- curve interpolation helpers (S*M*C, deCasteljau...)
 - and other...
 
 <hr />

cglm.podspec

@@ -2,7 +2,7 @@ Pod::Spec.new do |s|
 
   # Description
   s.name         = "cglm"
-  s.version      = "0.4.6"
+  s.version      = "0.5.1"
   s.summary      = "📽 Optimized OpenGL/Graphics Math (glm) for C"
   s.description  = <<-DESC
 cglm is math library for graphics programming for C. It is similar to original glm but it is written for C instead of C++ (you can use here too). See the documentation or README for all features.

configure.ac

@@ -29,6 +29,7 @@ LT_INIT
 # Checks for libraries.
 AC_CHECK_LIB([m], [floor])
 
+m4_ifdef([AM_SILENT_RULES], [AM_SILENT_RULES([yes])])
 AC_SYS_LARGEFILE
 
 # Checks for header files.

docs/source/api.rst

@@ -46,3 +46,5 @@ Follow the :doc:`build` documentation for this
    io
    call
    sphere
+   curve
+   bezier

docs/source/bezier.rst

@@ -0,0 +1,89 @@
+.. default-domain:: C
+
+Bezier
+================================================================================
+
+Header: cglm/bezier.h
+
+Common helpers for cubic bezier and similar curves.
+
+Table of contents (click to go):
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Functions:
+
+1. :c:func:`glm_bezier`
+2. :c:func:`glm_hermite`
+3. :c:func:`glm_decasteljau`
+
+Functions documentation
+~~~~~~~~~~~~~~~~~~~~~~~
+
+.. c:function:: float glm_bezier(float s, float p0, float c0, float c1, float p1)
+
+    | cubic bezier interpolation
+    | formula:
+
+    .. code-block:: text
+
+      B(s) = P0*(1-s)^3 + 3*C0*s*(1-s)^2 + 3*C1*s^2*(1-s) + P1*s^3
+
+    | similar result using matrix:
+
+    .. code-block:: text
+
+      B(s) = glm_smc(t, GLM_BEZIER_MAT, (vec4){p0, c0, c1, p1})
+
+    | glm_eq(glm_smc(...), glm_bezier(...)) should return TRUE
+
+    Parameters:
+      | *[in]*  **s**   parameter between 0 and 1
+      | *[in]*  **p0**  begin point
+      | *[in]*  **c0**  control point 1
+      | *[in]*  **c1**  control point 2
+      | *[in]*  **p1**  end point
+
+    Returns:
+        B(s)
+
+.. c:function:: float glm_hermite(float s, float p0, float t0, float t1, float p1)
+
+    | cubic hermite interpolation
+    | formula:
+
+    .. code-block:: text
+
+      H(s) = P0*(2*s^3 - 3*s^2 + 1) + T0*(s^3 - 2*s^2 + s) + P1*(-2*s^3 + 3*s^2) + T1*(s^3 - s^2)
+
+    | similar result using matrix:
+
+    .. code-block:: text
+
+      H(s) = glm_smc(t, GLM_HERMITE_MAT, (vec4){p0, p1, c0, c1})
+
+    | glm_eq(glm_smc(...), glm_hermite(...)) should return TRUE
+
+
+    Parameters:
+      | *[in]*  **s**   parameter between 0 and 1
+      | *[in]*  **p0**  begin point
+      | *[in]*  **t0**  tangent 1
+      | *[in]*  **t1**  tangent 2
+      | *[in]*  **p1**  end point
+
+    Returns:
+        B(s)
+
+.. c:function:: float glm_decasteljau(float prm, float p0, float c0, float c1, float p1)
+
+    | iterative way to solve cubic equation
+
+    Parameters:
+      | *[in]*  **prm** parameter between 0 and 1
+      | *[in]*  **p0**  begin point
+      | *[in]*  **c0**  control point 1
+      | *[in]*  **c1**  control point 2
+      | *[in]*  **p1**  end point
+
+    Returns:
+        parameter to use in cubic equation

docs/source/curve.rst

@@ -0,0 +1,41 @@
+.. default-domain:: C
+
+Curve
+================================================================================
+
+Header: cglm/curve.h
+
+Common helpers for common curves. For specific curve see its header/doc
+e.g bezier
+
+Table of contents (click to go):
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Functions:
+
+1. :c:func:`glm_smc`
+
+Functions documentation
+~~~~~~~~~~~~~~~~~~~~~~~
+
+.. c:function:: float  glm_smc(float s, mat4 m, vec4 c)
+
+    | helper function to calculate **S** * **M** * **C** multiplication for curves
+
+    | this function does not encourage you to use SMC, instead it is a helper if you use SMC.
+
+    | if you want to specify S as vector then use more generic glm_mat4_rmc() func.
+
+    | Example usage:
+
+    .. code-block:: c
+
+       Bs = glm_smc(s, GLM_BEZIER_MAT, (vec4){p0, c0, c1, p1})
+
+    Parameters:
+      | *[in]*  **s**  parameter between 0 and 1 (this will be [s3, s2, s, 1])
+      | *[in]*  **m**  basis matrix
+      | *[out]* **c**  position/control vector
+
+    Returns:
+        scalar value e.g. Bs

docs/source/mat4.rst

@@ -45,6 +45,7 @@ Functions:
 #. :c:func:`glm_mat4_inv_fast`
 #. :c:func:`glm_mat4_swap_col`
 #. :c:func:`glm_mat4_swap_row`
+#. :c:func:`glm_mat4_rmc`
 
 Functions documentation
 ~~~~~~~~~~~~~~~~~~~~~~~
@@ -270,3 +271,20 @@ Functions documentation
       | *[in, out]*  **mat**   matrix
       | *[in]*       **row1**  row1
       | *[in]*       **row2**  row2
+
+.. c:function:: float  glm_mat4_rmc(vec4 r, mat4 m, vec4 c)
+
+    | **rmc** stands for **Row** * **Matrix** * **Column**
+
+    | helper for  R (row vector) * M (matrix) * C (column vector)
+
+    | the result is scalar because S * M = Matrix1x4 (row vector),
+    | then Matrix1x4 * Vec4 (column vector) = Matrix1x1 (Scalar)
+
+    Parameters:
+      | *[in]*  **r**  row vector or matrix1x4
+      | *[in]*  **m**  matrix4x4
+      | *[in]*  **c**  column vector or matrix4x1
+
+    Returns:
+        scalar value e.g. Matrix1x1

docs/source/vec4.rst

@@ -58,11 +58,7 @@ Functions:
 #. :c:func:`glm_vec4_minv`
 #. :c:func:`glm_vec4_clamp`
 #. :c:func:`glm_vec4_lerp`
-#. :c:func:`glm_vec4_isnan`
+#. :c:func:`glm_vec4_cubic`
-#. :c:func:`glm_vec4_isinf`
-#. :c:func:`glm_vec4_isvalid`
-#. :c:func:`glm_vec4_sign`
-#. :c:func:`glm_vec4_sqrt`
 
 Functions documentation
 ~~~~~~~~~~~~~~~~~~~~~~~
@@ -401,3 +397,11 @@ Functions documentation
       | *[in]*  **to**     to value
       | *[in]*  **t**      interpolant (amount) clamped between 0 and 1
       | *[out]* **dest**   destination
+
+.. c:function:: void  glm_vec4_cubic(float s, vec4 dest)
+
+    helper to fill vec4 as [S^3, S^2, S, 1]
+
+    Parameters:
+      | *[in]*  **s**      parameter
+      | *[out]* **dest**   destination

include/cglm/bezier.h

@@ -0,0 +1,152 @@
+/*
+ * Copyright (c), Recep Aslantas.
+ *
+ * MIT License (MIT), http://opensource.org/licenses/MIT
+ * Full license can be found in the LICENSE file
+ */
+
+#ifndef cglm_bezier_h
+#define cglm_bezier_h
+
+#define GLM_BEZIER_MAT_INIT  {{-1.0f,  3.0f, -3.0f,  1.0f},                   \
+                              { 3.0f, -6.0f,  3.0f,  0.0f},                   \
+                              {-3.0f,  3.0f,  0.0f,  0.0f},                   \
+                              { 1.0f,  0.0f,  0.0f,  0.0f}}
+#define GLM_HERMITE_MAT_INIT {{ 2.0f, -3.0f,  0.0f,  1.0f},                   \
+                              {-2.0f,  3.0f,  0.0f,  0.0f},                   \
+                              { 1.0f, -2.0f,  1.0f,  0.0f},                   \
+                              { 1.0f, -1.0f,  0.0f,  0.0f}}
+/* for C only */
+#define GLM_BEZIER_MAT  ((mat4)GLM_BEZIER_MAT_INIT)
+#define GLM_HERMITE_MAT ((mat4)GLM_HERMITE_MAT_INIT)
+
+#define CGLM_DECASTEL_EPS   1e-9
+#define CGLM_DECASTEL_MAX   1000
+#define CGLM_DECASTEL_SMALL 1e-20
+
+/*!
+ * @brief cubic bezier interpolation
+ *
+ * Formula:
+ *  B(s) = P0*(1-s)^3 + 3*C0*s*(1-s)^2 + 3*C1*s^2*(1-s) + P1*s^3
+ *
+ * similar result using matrix:
+ *  B(s) = glm_smc(t, GLM_BEZIER_MAT, (vec4){p0, c0, c1, p1})
+ *
+ * glm_eq(glm_smc(...), glm_bezier(...)) should return TRUE
+ *
+ * @param[in]  s    parameter between 0 and 1
+ * @param[in]  p0   begin point
+ * @param[in]  c0   control point 1
+ * @param[in]  c1   control point 2
+ * @param[in]  p1   end point
+ *
+ * @return B(s)
+ */
+CGLM_INLINE
+float
+glm_bezier(float s, float p0, float c0, float c1, float p1) {
+  float x, xx, ss, xs3, a;
+
+  x   = 1.0f - s;
+  xx  = x * x;
+  ss  = s * s;
+  xs3 = (s - ss) * 3.0f;
+  a   = p0 * xx + c0 * xs3;
+
+  return a + s * (c1 * xs3 + p1 * ss - a);
+}
+
+/*!
+ * @brief cubic hermite interpolation
+ *
+ * Formula:
+ *  H(s) = P0*(2*s^3 - 3*s^2 + 1) + T0*(s^3 - 2*s^2 + s)
+ *            + P1*(-2*s^3 + 3*s^2) + T1*(s^3 - s^2)
+ *
+ * similar result using matrix:
+ *  H(s) = glm_smc(t, GLM_HERMITE_MAT, (vec4){p0, p1, c0, c1})
+ *
+ * glm_eq(glm_smc(...), glm_hermite(...)) should return TRUE
+ *
+ * @param[in]  s    parameter between 0 and 1
+ * @param[in]  p0   begin point
+ * @param[in]  t0   tangent 1
+ * @param[in]  t1   tangent 2
+ * @param[in]  p1   end point
+ *
+ * @return H(s)
+ */
+CGLM_INLINE
+float
+glm_hermite(float s, float p0, float t0, float t1, float p1) {
+  float ss, d, a, b, c, e, f;
+
+  ss = s  * s;
+  a  = ss + ss;
+  c  = a  + ss;
+  b  = a  * s;
+  d  = s  * ss;
+  f  = d  - ss;
+  e  = b  - c;
+
+  return p0 * (e + 1.0f) + t0 * (f - ss + s) + t1 * f - p1 * e;
+}
+
+/*!
+ * @brief iterative way to solve cubic equation
+ *
+ * @param[in]  prm  parameter between 0 and 1
+ * @param[in]  p0   begin point
+ * @param[in]  c0   control point 1
+ * @param[in]  c1   control point 2
+ * @param[in]  p1   end point
+ *
+ * @return parameter to use in cubic equation
+ */
+CGLM_INLINE
+float
+glm_decasteljau(float prm, float p0, float c0, float c1, float p1) {
+  float u, v, a, b, c, d, e, f;
+  int   i;
+
+  if (prm - p0 < CGLM_DECASTEL_SMALL)
+    return 0.0f;
+
+  if (p1 - prm < CGLM_DECASTEL_SMALL)
+    return 1.0f;
+
+  u  = 0.0f;
+  v  = 1.0f;
+
+  for (i = 0; i < CGLM_DECASTEL_MAX; i++) {
+    /* de Casteljau Subdivision */
+    a  = (p0 + c0) * 0.5f;
+    b  = (c0 + c1) * 0.5f;
+    c  = (c1 + p1) * 0.5f;
+    d  = (a  + b)  * 0.5f;
+    e  = (b  + c)  * 0.5f;
+    f  = (d  + e)  * 0.5f; /* this one is on the curve! */
+
+    /* The curve point is close enough to our wanted t */
+    if (fabsf(f - prm) < CGLM_DECASTEL_EPS)
+      return glm_clamp_zo((u  + v) * 0.5f);
+
+    /* dichotomy */
+    if (f < prm) {
+      p0 = f;
+      c0 = e;
+      c1 = c;
+      u  = (u  + v) * 0.5f;
+    } else {
+      c0 = a;
+      c1 = d;
+      p1 = f;
+      v  = (u  + v) * 0.5f;
+    }
+  }
+
+  return glm_clamp_zo((u  + v) * 0.5f);
+}
+
+#endif /* cglm_bezier_h */

include/cglm/call.h

@@ -27,6 +27,8 @@ extern "C" {
 #include "call/project.h"
 #include "call/sphere.h"
 #include "call/ease.h"
+#include "call/curve.h"
+#include "call/bezier.h"
 
 #ifdef __cplusplus
 }

include/cglm/call/bezier.h

@@ -0,0 +1,31 @@
+/*
+ * Copyright (c), Recep Aslantas.
+ *
+ * MIT License (MIT), http://opensource.org/licenses/MIT
+ * Full license can be found in the LICENSE file
+ */
+
+#ifndef cglmc_bezier_h
+#define cglmc_bezier_h
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "../cglm.h"
+
+CGLM_EXPORT
+float
+glmc_bezier(float s, float p0, float c0, float c1, float p1);
+
+CGLM_EXPORT
+float
+glmc_hermite(float s, float p0, float t0, float t1, float p1);
+
+CGLM_EXPORT
+float
+glmc_decasteljau(float prm, float p0, float c0, float c1, float p1);
+
+#ifdef __cplusplus
+}
+#endif
+#endif /* cglmc_bezier_h */

include/cglm/call/curve.h

@@ -0,0 +1,23 @@
+/*
+ * Copyright (c), Recep Aslantas.
+ *
+ * MIT License (MIT), http://opensource.org/licenses/MIT
+ * Full license can be found in the LICENSE file
+ */
+
+#ifndef cglmc_curve_h
+#define cglmc_curve_h
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "../cglm.h"
+
+CGLM_EXPORT
+float
+glmc_smc(float s, mat4 m, vec4 c);
+
+#ifdef __cplusplus
+}
+#endif
+#endif /* cglmc_curve_h */

include/cglm/call/ease.h

@@ -137,4 +137,7 @@ CGLM_EXPORT
 float
 glmc_ease_bounce_inout(float t);
 
+#ifdef __cplusplus
+}
+#endif
 #endif /* cglmc_ease_h */

include/cglm/call/mat4.h

@@ -113,6 +113,10 @@ CGLM_EXPORT
 void
 glmc_mat4_swap_row(mat4 mat, int row1, int row2);
 
+CGLM_EXPORT
+float
+glmc_mat4_rmc(vec4 r, mat4 m, vec4 c);
+
 #ifdef __cplusplus
 }
 #endif

include/cglm/call/sphere.h

@@ -33,4 +33,7 @@ CGLM_EXPORT
 bool
 glmc_sphere_point(vec4 s, vec3 point);
 
+#ifdef __cplusplus
+}
+#endif
 #endif /* cglmc_sphere_h */

include/cglm/call/vec4.h

@@ -153,6 +153,10 @@ CGLM_EXPORT
 void
 glmc_vec4_lerp(vec4 from, vec4 to, float t, vec4 dest);
 
+CGLM_EXPORT
+void
+glmc_vec4_cubic(float s, vec4 dest);
+
 /* ext */
 
 CGLM_EXPORT

include/cglm/cglm.h

@@ -26,5 +26,7 @@
 #include "project.h"
 #include "sphere.h"
 #include "ease.h"
+#include "curve.h"
+#include "bezier.h"
 
 #endif /* cglm_h */

include/cglm/curve.h

@@ -0,0 +1,40 @@
+/*
+ * Copyright (c), Recep Aslantas.
+ *
+ * MIT License (MIT), http://opensource.org/licenses/MIT
+ * Full license can be found in the LICENSE file
+ */
+
+#ifndef cglm_curve_h
+#define cglm_curve_h
+
+#include "common.h"
+#include "vec4.h"
+#include "mat4.h"
+
+/*!
+ * @brief helper function to calculate S*M*C multiplication for curves
+ *
+ * This function does not encourage you to use SMC,
+ * instead it is a helper if you use SMC.
+ *
+ * if you want to specify S as vector then use more generic glm_mat4_rmc() func.
+ *
+ * Example usage:
+ *  B(s) = glm_smc(s, GLM_BEZIER_MAT, (vec4){p0, c0, c1, p1})
+ *
+ * @param[in]  s  parameter between 0 and 1 (this will be [s3, s2, s, 1])
+ * @param[in]  m  basis matrix
+ * @param[in]  c  position/control vector
+ *
+ * @return B(s)
+ */
+CGLM_INLINE
+float
+glm_smc(float s, mat4 m, vec4 c) {
+  vec4 vs;
+  glm_vec4_cubic(s, vs);
+  return glm_mat4_rmc(vs, m, c);
+}
+
+#endif /* cglm_curve_h */

include/cglm/mat4.h

@@ -118,6 +118,11 @@ glm_mat4_copy(mat4 mat, mat4 dest) {
   glmm_store(dest[1], glmm_load(mat[1]));
   glmm_store(dest[2], glmm_load(mat[2]));
   glmm_store(dest[3], glmm_load(mat[3]));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest[0], vld1q_f32(mat[0]));
+  vst1q_f32(dest[1], vld1q_f32(mat[1]));
+  vst1q_f32(dest[2], vld1q_f32(mat[2]));
+  vst1q_f32(dest[3], vld1q_f32(mat[3]));
 #else
   glm_mat4_ucopy(mat, dest);
 #endif
@@ -252,7 +257,7 @@ 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 )
+#elif defined(CGLM_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],
@@ -504,10 +509,15 @@ glm_mat4_scale_p(mat4 m, float s) {
 CGLM_INLINE
 void
 glm_mat4_scale(mat4 m, float s) {
-#ifdef __AVX__
+#if defined( __SSE__ ) || defined( __SSE2__ )
-  glm_mat4_scale_avx(m, s);
-#elif defined( __SSE__ ) || defined( __SSE2__ )
   glm_mat4_scale_sse2(m, s);
+#elif defined(CGLM_NEON_FP)
+  float32x4_t v0;
+  v0 = vdupq_n_f32(s);
+  vst1q_f32(m[0], vmulq_f32(vld1q_f32(m[0]), v0));
+  vst1q_f32(m[1], vmulq_f32(vld1q_f32(m[1]), v0));
+  vst1q_f32(m[2], vmulq_f32(vld1q_f32(m[2]), v0));
+  vst1q_f32(m[3], vmulq_f32(vld1q_f32(m[3]), v0));
 #else
   glm_mat4_scale_p(m, s);
 #endif
@@ -556,9 +566,7 @@ glm_mat4_det(mat4 mat) {
 CGLM_INLINE
 void
 glm_mat4_inv(mat4 mat, mat4 dest) {
-#ifdef __AVX__
+#if defined( __SSE__ ) || defined( __SSE2__ )
-  glm_mat4_inv_avx(mat, dest);
-#elif defined( __SSE__ ) || defined( __SSE2__ )
   glm_mat4_inv_sse2(mat, dest);
 #else
   float t[6];
@@ -619,9 +627,7 @@ glm_mat4_inv(mat4 mat, mat4 dest) {
 CGLM_INLINE
 void
 glm_mat4_inv_fast(mat4 mat, mat4 dest) {
-#ifdef __AVX__
+#if defined( __SSE__ ) || defined( __SSE2__ )
-  glm_mat4_inv_fast_avx(mat, dest);
-#elif defined( __SSE__ ) || defined( __SSE2__ )
   glm_mat4_inv_fast_sse2(mat, dest);
 #else
   glm_mat4_inv(mat, dest);
@@ -671,4 +677,26 @@ glm_mat4_swap_row(mat4 mat, int row1, int row2) {
   mat[3][row2] = tmp[3];
 }
 
+/*!
+ * @brief helper for  R (row vector) * M (matrix) * C (column vector)
+ *
+ * rmc stands for Row * Matrix * Column
+ *
+ * the result is scalar because S * M = Matrix1x4 (row vector),
+ * then Matrix1x4 * Vec4 (column vector) = Matrix1x1 (Scalar)
+ *
+ * @param[in]  r   row vector or matrix1x4
+ * @param[in]  m   matrix4x4
+ * @param[in]  c   column vector or matrix4x1
+ *
+ * @return scalar value e.g. B(s)
+ */
+CGLM_INLINE
+float
+glm_mat4_rmc(vec4 r, mat4 m, vec4 c) {
+  vec4 tmp;
+  glm_mat4_mulv(m, c, tmp);
+  return glm_vec4_dot(r, tmp);
+}
+
 #endif /* cglm_mat_h */

include/cglm/quat.h

@@ -218,7 +218,7 @@ glm_quat_normalize_to(versor q, versor dest) {
   float  dot;
 
   x0   = glmm_load(q);
-  xdot = glmm_dot(x0, x0);
+  xdot = glmm_vdot(x0, x0);
   dot  = _mm_cvtss_f32(xdot);
 
   if (dot <= 0.0f) {

include/cglm/simd/arm.h

@@ -0,0 +1,41 @@
+/*
+ * Copyright (c), Recep Aslantas.
+ *
+ * MIT License (MIT), http://opensource.org/licenses/MIT
+ * Full license can be found in the LICENSE file
+ */
+
+#ifndef cglm_simd_arm_h
+#define cglm_simd_arm_h
+#include "intrin.h"
+#ifdef CGLM_SIMD_ARM
+
+#define glmm_load(p)      vld1q_f32(p)
+#define glmm_store(p, a)  vst1q_f32(p, a)
+
+static inline
+float
+glmm_hadd(float32x4_t v) {
+#if defined(__aarch64__)
+  return vaddvq_f32(v);
+#else
+  v = vaddq_f32(v, vrev64q_f32(v));
+  v = vaddq_f32(v, vcombine_f32(vget_high_f32(v), vget_low_f32(v)));
+  return vgetq_lane_f32(v, 0);
+#endif
+}
+
+static inline
+float
+glmm_dot(float32x4_t a, float32x4_t b) {
+  return glmm_hadd(vmulq_f32(a, b));
+}
+
+static inline
+float
+glmm_norm(float32x4_t a) {
+  return sqrtf(glmm_dot(a, a));
+}
+
+#endif
+#endif /* cglm_simd_arm_h */

include/cglm/simd/avx/mat4.h

@@ -14,23 +14,12 @@
 
 #include <immintrin.h>
 
-CGLM_INLINE
-void
-glm_mat4_scale_avx(mat4 m, float s) {
-  __m256 y0;
-  y0 = _mm256_set1_ps(s);
-
-  glmm_store256(m[0], _mm256_mul_ps(y0, glmm_load256(m[0])));
-  glmm_store256(m[2], _mm256_mul_ps(y0, glmm_load256(m[2])));
-}
-
 CGLM_INLINE
 void
 glm_mat4_mul_avx(mat4 m1, mat4 m2, mat4 dest) {
   /* D = R * L (Column-Major) */
 
   __m256 y0, y1, y2, y3, y4, y5, y6, y7, y8, y9;
-  __m256i yi0, yi1, yi2, yi3;
 
   y0 = glmm_load256(m2[0]); /* h g f e d c b a */
   y1 = glmm_load256(m2[2]); /* p o n m l k j i */
@@ -42,19 +31,14 @@ glm_mat4_mul_avx(mat4 m1, mat4 m2, mat4 dest) {
   y4 = _mm256_permute2f128_ps(y2, y2, 0x03); /* d c b a h g f e */
   y5 = _mm256_permute2f128_ps(y3, y3, 0x03); /* l k j i p o n m */
 
-  yi0 = _mm256_set_epi32(1, 1, 1, 1, 0, 0, 0, 0);
-  yi1 = _mm256_set_epi32(3, 3, 3, 3, 2, 2, 2, 2);
-  yi2 = _mm256_set_epi32(0, 0, 0, 0, 1, 1, 1, 1);
-  yi3 = _mm256_set_epi32(2, 2, 2, 2, 3, 3, 3, 3);
-
   /* f f f f a a a a */
   /* h h h h c c c c */
   /* e e e e b b b b */
   /* g g g g d d d d */
-  y6 = _mm256_permutevar_ps(y0, yi0);
+  y6 = _mm256_permutevar_ps(y0, _mm256_set_epi32(1, 1, 1, 1, 0, 0, 0, 0));
-  y7 = _mm256_permutevar_ps(y0, yi1);
+  y7 = _mm256_permutevar_ps(y0, _mm256_set_epi32(3, 3, 3, 3, 2, 2, 2, 2));
-  y8 = _mm256_permutevar_ps(y0, yi2);
+  y8 = _mm256_permutevar_ps(y0, _mm256_set_epi32(0, 0, 0, 0, 1, 1, 1, 1));
-  y9 = _mm256_permutevar_ps(y0, yi3);
+  y9 = _mm256_permutevar_ps(y0, _mm256_set_epi32(2, 2, 2, 2, 3, 3, 3, 3));
 
   glmm_store256(dest[0],
                 _mm256_add_ps(_mm256_add_ps(_mm256_mul_ps(y2, y6),
@@ -66,10 +50,10 @@ glm_mat4_mul_avx(mat4 m1, mat4 m2, mat4 dest) {
   /* p p p p k k k k */
   /* m m m m j j j j */
   /* o o o o l l l l */
-  y6 = _mm256_permutevar_ps(y1, yi0);
+  y6 = _mm256_permutevar_ps(y1, _mm256_set_epi32(1, 1, 1, 1, 0, 0, 0, 0));
-  y7 = _mm256_permutevar_ps(y1, yi1);
+  y7 = _mm256_permutevar_ps(y1, _mm256_set_epi32(3, 3, 3, 3, 2, 2, 2, 2));
-  y8 = _mm256_permutevar_ps(y1, yi2);
+  y8 = _mm256_permutevar_ps(y1, _mm256_set_epi32(0, 0, 0, 0, 1, 1, 1, 1));
-  y9 = _mm256_permutevar_ps(y1, yi3);
+  y9 = _mm256_permutevar_ps(y1, _mm256_set_epi32(2, 2, 2, 2, 3, 3, 3, 3));
 
   glmm_store256(dest[2],
                 _mm256_add_ps(_mm256_add_ps(_mm256_mul_ps(y2, y6),
@@ -78,365 +62,5 @@ glm_mat4_mul_avx(mat4 m1, mat4 m2, mat4 dest) {
                                             _mm256_mul_ps(y5, y9))));
 }
 
-CGLM_INLINE
-void
-glm_mat4_inv_avx(mat4 mat, mat4 dest) {
-  __m256  y0, y1, y2, y3, y4, y5, y6, y7, y8, y9, y10, y11, y12, y13;
-  __m256  yt0, yt1, yt2;
-  __m256  t0, t1, t2;
-  __m256  r1, r2;
-  __m256  flpsign;
-  __m256i yi1, yi2, yi3;
-
-  y0  = glmm_load256(mat[0]); /* h g f e d c b a */
-  y1  = glmm_load256(mat[2]); /* p o n m l k j i */
-
-  y2  = _mm256_permute2f128_ps(y1, y1, 0x00); /* l k j i l k j i */
-  y3  = _mm256_permute2f128_ps(y1, y1, 0x11); /* p o n m p o n m */
-  y4  = _mm256_permute2f128_ps(y0, y0, 0x03); /* d c b a h g f e */
-  y13 = _mm256_permute2f128_ps(y4, y4, 0x00); /* h g f e h g f e */
-
-  yi1     = _mm256_set_epi32(0, 0, 0, 0, 0, 1, 1, 2);
-  yi2     = _mm256_set_epi32(1, 1, 1, 2, 3, 2, 3, 3);
-  flpsign = _mm256_set_ps(0.f, -0.f, 0.f, -0.f, -0.f, 0.f, -0.f, 0.f);
-
-  /* i i i i i j j k */
-  /* n n n o p o p p */
-  /* m m m m m n n o */
-  /* j j j k l k l l */
-  /* e e e e e f f g */
-  /* f f f g h g h h */
-  y5  = _mm256_permutevar_ps(y2, yi1);
-  y6  = _mm256_permutevar_ps(y3, yi2);
-  y7  = _mm256_permutevar_ps(y3, yi1);
-  y8  = _mm256_permutevar_ps(y2, yi2);
-  y2  = _mm256_permutevar_ps(y13, yi1);
-  y3  = _mm256_permutevar_ps(y13, yi2);
-
-  yi1 = _mm256_set_epi32(2, 1, 0, 0, 2, 1, 0, 0);
-  yi2 = _mm256_set_epi32(2, 1, 1, 0, 2, 1, 1, 0);
-  yi3 = _mm256_set_epi32(3, 3, 2, 0, 3, 3, 2, 0);
-
-  /*
-   t0[0] = k * p - o * l;    t1[0] = g * p - o * h;    t2[0] = g * l - k * h;
-   t0[1] = j * p - n * l;    t1[1] = f * p - n * h;    t2[1] = f * l - j * h;
-   t0[2] = j * o - n * k;    t1[2] = f * o - n * g;    t2[2] = f * k - j * g;
-   t0[3] = i * p - m * l;    t1[3] = e * p - m * h;    t2[3] = e * l - i * h;
-   t0[4] = i * o - m * k;    t1[4] = e * o - m * g;    t2[4] = e * k - i * g;
-   t0[5] = i * n - m * j;    t1[5] = e * n - m * f;    t2[5] = e * j - i * f;
-   */
-  yt0 = _mm256_sub_ps(_mm256_mul_ps(y5, y6), _mm256_mul_ps(y7, y8));
-  yt1 = _mm256_sub_ps(_mm256_mul_ps(y2, y6), _mm256_mul_ps(y7, y3));
-  yt2 = _mm256_sub_ps(_mm256_mul_ps(y2, y8), _mm256_mul_ps(y5, y3));
-
-  /* t3 t2 t1 t0 t3 t2 t1 t0 */
-  /* t5 t5 t5 t4 t5 t5 t5 t4 */
-  y9  = _mm256_permute2f128_ps(yt0, yt0, 0x00);
-  y10 = _mm256_permute2f128_ps(yt0, yt0, 0x11);
-//
-  /* t2 t1 t0 t0 t2 t1 t0 t0 */
-  t0  = _mm256_permutevar_ps(y9, yi1);
-
-  /* t4 t3 t3 t1 t4 t3 t3 t1 */
-  y11 = _mm256_shuffle_ps(y9, y10, 0x4D);
-  y12 = _mm256_permutevar_ps(y11, yi2);
-  t1  = _mm256_permute2f128_ps(y12, y9, 0x00);
-
-  /* t5 t5 t4 t2 t5 t5 t4 t2 */
-  y11 = _mm256_shuffle_ps(y9, y10, 0x4A);
-  y12 = _mm256_permutevar_ps(y11, yi3);
-  t2  = _mm256_permute2f128_ps(y12, y12, 0x00);
-
-  /* a a a b e e e f */
-  /* b b c c f f g g */
-  /* c d d d g h h h */
-  y9  = _mm256_permute_ps(y4, 0x01);
-  y10 = _mm256_permute_ps(y4, 0x5A);
-  y11 = _mm256_permute_ps(y4, 0xBF);
-
-  /*
-   dest[0][0] =  f * t[0] - g * t[1] + h * t[2];
-   dest[1][0] =-(e * t[0] - g * t[3] + h * t[4]);
-   dest[2][0] =  e * t[1] - f * t[3] + h * t[5];
-   dest[3][0] =-(e * t[2] - f * t[4] + g * t[5]);
-
-   dest[0][1] =-(b * t[0] - c * t[1] + d * t[2]);
-   dest[1][1] =  a * t[0] - c * t[3] + d * t[4];
-   dest[2][1] =-(a * t[1] - b * t[3] + d * t[5]);
-   dest[3][1] =  a * t[2] - b * t[4] + c * t[5];
-   */
-  r1 = _mm256_xor_ps(_mm256_add_ps(_mm256_sub_ps(_mm256_mul_ps(y9, t0),
-                                                 _mm256_mul_ps(y10, t1)),
-                                   _mm256_mul_ps(y11, t2)),
-                     flpsign);
-
-  /* d c b a d c b a */
-  y2 = _mm256_permute2f128_ps(y0, y0, 0x0);
-
-  /* a a a b a a a b */
-  /* b b c c b b c c */
-  /* c d d d c d d d */
-  y3 = _mm256_permutevar_ps(y2, _mm256_set_epi32(0, 0, 0, 1, 0, 0, 0, 1));
-  y4 = _mm256_permutevar_ps(y2, _mm256_set_epi32(1, 1, 2, 2, 1, 1, 2, 2));
-  y5 = _mm256_permutevar_ps(y2, _mm256_set_epi32(2, 3, 3, 3, 2, 3, 3, 3));
-
-  /* t2[3] t2[2] t2[1] t2[0] t1[3] t1[2] t1[1] t1[0] */
-  /* t2[5] t2[5] t2[5] t2[4] t1[5] t1[5] t1[5] t1[4] */
-  y6 = _mm256_permute2f128_ps(yt1, yt2, 0x20);
-  y7 = _mm256_permute2f128_ps(yt1, yt2, 0x31);
-
-  /* t2[2] t2[1] t2[0] t2[0] t1[2] t1[1] t1[0] t1[0] */
-  t0 = _mm256_permutevar_ps(y6, yi1);
-
-  /* t1[4] t1[3] t1[3] t1[1] t1[4] t1[3] t1[3] t1[1] */
-
-  /* t1[4] t1[3] t1[3] t1[1] t1[4] t1[3] t1[3] t1[1] */
-  y11 = _mm256_shuffle_ps(y6, y7, 0x4D);
-  t1  = _mm256_permutevar_ps(y11, yi2);
-
-
-  /* t2[5] t2[5] t2[4] t2[2] t1[5] t1[5] t1[4] t1[2] */
-  y11 = _mm256_shuffle_ps(y6, y7, 0x4A);
-  t2  = _mm256_permutevar_ps(y11, yi3);
-
-  /*
-   dest[0][2] =  b * t1[0] - c * t1[1] + d * t1[2];
-   dest[1][2] =-(a * t1[0] - c * t1[3] + d * t1[4]);
-   dest[2][2] =  a * t1[1] - b * t1[3] + d * t1[5];
-   dest[3][2] =-(a * t1[2] - b * t1[4] + c * t1[5]);
-
-   dest[0][3] =-(b * t2[0] - c * t2[1] + d * t2[2]);
-   dest[1][3] =  a * t2[0] - c * t2[3] + d * t2[4];
-   dest[2][3] =-(a * t2[1] - b * t2[3] + d * t2[5]);
-   dest[3][3] =  a * t2[2] - b * t2[4] + c * t2[5];
-   */
-  r2 = _mm256_xor_ps(_mm256_add_ps(_mm256_sub_ps(_mm256_mul_ps(y3, t0),
-                                                 _mm256_mul_ps(y4, t1)),
-                                   _mm256_mul_ps(y5, t2)),
-                     flpsign);
-
-  /* determinant */
-
-  y4 = _mm256_mul_ps(y0, r1);
-  y4 = _mm256_permute2f128_ps(y4, y4, 0x30);
-  y4 = _mm256_dp_ps(y0, r1, 0xff);
-
-  y5 = _mm256_div_ps(_mm256_set1_ps(1.0f), y4);
-  r1 = _mm256_mul_ps(r1, y5);
-  r2 = _mm256_mul_ps(r2, y5);
-
-  /* transpose */
-
-  /* d c b a h g f e */
-  /* l k j i p o n m */
-  y0 = _mm256_permute2f128_ps(r1, r1, 0x03);
-  y1 = _mm256_permute2f128_ps(r2, r2, 0x03);
-
-  /* b a f e f e b a */
-  /* j i n m n m j i */
-  /* i m a e m i e a */
-  /* j n b f n j f b */
-  /* n j f b m i e a */
-  y2 = _mm256_shuffle_ps(r1, y0, 0x44);
-  y3 = _mm256_shuffle_ps(r2, y1, 0x44);
-  y4 = _mm256_shuffle_ps(y2, y3, 0x88);
-  y5 = _mm256_shuffle_ps(y2, y3, 0xDD);
-  y6 = _mm256_permute2f128_ps(y4, y5, 0x20);
-
-  /* d c h g h g d c */
-  /* l k p o p o l k */
-  /* k o c g o k g c */
-  /* l p d h p l h d */
-  /* p l h d o k g c */
-  y2 = _mm256_shuffle_ps(r1, y0, 0xEE);
-  y3 = _mm256_shuffle_ps(r2, y1, 0xEE);
-  y4 = _mm256_shuffle_ps(y2, y3, 0x88);
-  y5 = _mm256_shuffle_ps(y2, y3, 0xDD);
-  y7 = _mm256_permute2f128_ps(y4, y5, 0x20);
-
-  glmm_store256(dest[0], y6);
-  glmm_store256(dest[2], y7);
-}
-
-CGLM_INLINE
-void
-glm_mat4_inv_fast_avx(mat4 mat, mat4 dest) {
-  __m256  y0, y1, y2, y3, y4, y5, y6, y7, y8, y9, y10, y11, y12, y13;
-  __m256  yt0, yt1, yt2;
-  __m256  t0, t1, t2;
-  __m256  r1, r2;
-  __m256  flpsign;
-  __m256i yi1, yi2, yi3;
-
-  y0  = glmm_load256(mat[0]); /* h g f e d c b a */
-  y1  = glmm_load256(mat[2]); /* p o n m l k j i */
-
-  y2  = _mm256_permute2f128_ps(y1, y1, 0x00); /* l k j i l k j i */
-  y3  = _mm256_permute2f128_ps(y1, y1, 0x11); /* p o n m p o n m */
-  y4  = _mm256_permute2f128_ps(y0, y0, 0x03); /* d c b a h g f e */
-  y13 = _mm256_permute2f128_ps(y4, y4, 0x00); /* h g f e h g f e */
-
-  yi1     = _mm256_set_epi32(0, 0, 0, 0, 0, 1, 1, 2);
-  yi2     = _mm256_set_epi32(1, 1, 1, 2, 3, 2, 3, 3);
-  flpsign = _mm256_set_ps(0.f, -0.f, 0.f, -0.f, -0.f, 0.f, -0.f, 0.f);
-
-  /* i i i i i j j k */
-  /* n n n o p o p p */
-  /* m m m m m n n o */
-  /* j j j k l k l l */
-  /* e e e e e f f g */
-  /* f f f g h g h h */
-  y5  = _mm256_permutevar_ps(y2, yi1);
-  y6  = _mm256_permutevar_ps(y3, yi2);
-  y7  = _mm256_permutevar_ps(y3, yi1);
-  y8  = _mm256_permutevar_ps(y2, yi2);
-  y2  = _mm256_permutevar_ps(y13, yi1);
-  y3  = _mm256_permutevar_ps(y13, yi2);
-
-  yi1 = _mm256_set_epi32(2, 1, 0, 0, 2, 1, 0, 0);
-  yi2 = _mm256_set_epi32(2, 1, 1, 0, 2, 1, 1, 0);
-  yi3 = _mm256_set_epi32(3, 3, 2, 0, 3, 3, 2, 0);
-
-  /*
-   t0[0] = k * p - o * l;    t1[0] = g * p - o * h;    t2[0] = g * l - k * h;
-   t0[1] = j * p - n * l;    t1[1] = f * p - n * h;    t2[1] = f * l - j * h;
-   t0[2] = j * o - n * k;    t1[2] = f * o - n * g;    t2[2] = f * k - j * g;
-   t0[3] = i * p - m * l;    t1[3] = e * p - m * h;    t2[3] = e * l - i * h;
-   t0[4] = i * o - m * k;    t1[4] = e * o - m * g;    t2[4] = e * k - i * g;
-   t0[5] = i * n - m * j;    t1[5] = e * n - m * f;    t2[5] = e * j - i * f;
-   */
-  yt0 = _mm256_sub_ps(_mm256_mul_ps(y5, y6), _mm256_mul_ps(y7, y8));
-  yt1 = _mm256_sub_ps(_mm256_mul_ps(y2, y6), _mm256_mul_ps(y7, y3));
-  yt2 = _mm256_sub_ps(_mm256_mul_ps(y2, y8), _mm256_mul_ps(y5, y3));
-
-  /* t3 t2 t1 t0 t3 t2 t1 t0 */
-  /* t5 t5 t5 t4 t5 t5 t5 t4 */
-  y9  = _mm256_permute2f128_ps(yt0, yt0, 0x00);
-  y10 = _mm256_permute2f128_ps(yt0, yt0, 0x11);
-
-  /* t2 t1 t0 t0 t2 t1 t0 t0 */
-  t0  = _mm256_permutevar_ps(y9, yi1);
-
-  /* t4 t3 t3 t1 t4 t3 t3 t1 */
-  y11 = _mm256_shuffle_ps(y9, y10, 0x4D);
-  y12 = _mm256_permutevar_ps(y11, yi2);
-  t1  = _mm256_permute2f128_ps(y12, y9, 0x00);
-
-  /* t5 t5 t4 t2 t5 t5 t4 t2 */
-  y11 = _mm256_shuffle_ps(y9, y10, 0x4A);
-  y12 = _mm256_permutevar_ps(y11, yi3);
-  t2  = _mm256_permute2f128_ps(y12, y12, 0x00);
-
-  /* a a a b e e e f */
-  /* b b c c f f g g */
-  /* c d d d g h h h */
-  y9  = _mm256_permute_ps(y4, 0x01);
-  y10 = _mm256_permute_ps(y4, 0x5A);
-  y11 = _mm256_permute_ps(y4, 0xBF);
-
-  /*
-   dest[0][0] =  f * t[0] - g * t[1] + h * t[2];
-   dest[1][0] =-(e * t[0] - g * t[3] + h * t[4]);
-   dest[2][0] =  e * t[1] - f * t[3] + h * t[5];
-   dest[3][0] =-(e * t[2] - f * t[4] + g * t[5]);
-
-   dest[0][1] =-(b * t[0] - c * t[1] + d * t[2]);
-   dest[1][1] =  a * t[0] - c * t[3] + d * t[4];
-   dest[2][1] =-(a * t[1] - b * t[3] + d * t[5]);
-   dest[3][1] =  a * t[2] - b * t[4] + c * t[5];
-   */
-  r1 = _mm256_xor_ps(_mm256_add_ps(_mm256_sub_ps(_mm256_mul_ps(y9, t0),
-                                                 _mm256_mul_ps(y10, t1)),
-                                   _mm256_mul_ps(y11, t2)),
-                     flpsign);
-
-  /* d c b a d c b a */
-  y2 = _mm256_permute2f128_ps(y0, y0, 0x0);
-
-  /* a a a b a a a b */
-  /* b b c c b b c c */
-  /* c d d d c d d d */
-  y3 = _mm256_permutevar_ps(y2, _mm256_set_epi32(0, 0, 0, 1, 0, 0, 0, 1));
-  y4 = _mm256_permutevar_ps(y2, _mm256_set_epi32(1, 1, 2, 2, 1, 1, 2, 2));
-  y5 = _mm256_permutevar_ps(y2, _mm256_set_epi32(2, 3, 3, 3, 2, 3, 3, 3));
-
-  /* t2[3] t2[2] t2[1] t2[0] t1[3] t1[2] t1[1] t1[0] */
-  /* t2[5] t2[5] t2[5] t2[4] t1[5] t1[5] t1[5] t1[4] */
-  y6 = _mm256_permute2f128_ps(yt1, yt2, 0x20);
-  y7 = _mm256_permute2f128_ps(yt1, yt2, 0x31);
-
-  /* t2[2] t2[1] t2[0] t2[0] t1[2] t1[1] t1[0] t1[0] */
-  t0 = _mm256_permutevar_ps(y6, yi1);
-
-  /* t1[4] t1[3] t1[3] t1[1] t1[4] t1[3] t1[3] t1[1] */
-
-  /* t1[4] t1[3] t1[3] t1[1] t1[4] t1[3] t1[3] t1[1] */
-  y11 = _mm256_shuffle_ps(y6, y7, 0x4D);
-  t1  = _mm256_permutevar_ps(y11, yi2);
-
-
-  /* t2[5] t2[5] t2[4] t2[2] t1[5] t1[5] t1[4] t1[2] */
-  y11 = _mm256_shuffle_ps(y6, y7, 0x4A);
-  t2  = _mm256_permutevar_ps(y11, yi3);
-
-  /*
-   dest[0][2] =  b * t1[0] - c * t1[1] + d * t1[2];
-   dest[1][2] =-(a * t1[0] - c * t1[3] + d * t1[4]);
-   dest[2][2] =  a * t1[1] - b * t1[3] + d * t1[5];
-   dest[3][2] =-(a * t1[2] - b * t1[4] + c * t1[5]);
-
-   dest[0][3] =-(b * t2[0] - c * t2[1] + d * t2[2]);
-   dest[1][3] =  a * t2[0] - c * t2[3] + d * t2[4];
-   dest[2][3] =-(a * t2[1] - b * t2[3] + d * t2[5]);
-   dest[3][3] =  a * t2[2] - b * t2[4] + c * t2[5];
-   */
-  r2 = _mm256_xor_ps(_mm256_add_ps(_mm256_sub_ps(_mm256_mul_ps(y3, t0),
-                                                 _mm256_mul_ps(y4, t1)),
-                                   _mm256_mul_ps(y5, t2)),
-                     flpsign);
-
-  /* determinant */
-
-  y4 = _mm256_mul_ps(y0, r1);
-  y4 = _mm256_permute2f128_ps(y4, y4, 0x30);
-  y4 = _mm256_dp_ps(y0, r1, 0xff);
-
-  y5 = _mm256_rcp_ps(y4);
-  r1 = _mm256_mul_ps(r1, y5);
-  r2 = _mm256_mul_ps(r2, y5);
-
-  /* transpose */
-
-  /* d c b a h g f e */
-  /* l k j i p o n m */
-  y0 = _mm256_permute2f128_ps(r1, r1, 0x03);
-  y1 = _mm256_permute2f128_ps(r2, r2, 0x03);
-
-  /* b a f e f e b a */
-  /* j i n m n m j i */
-  /* i m a e m i e a */
-  /* j n b f n j f b */
-  /* n j f b m i e a */
-  y2 = _mm256_shuffle_ps(r1, y0, 0x44);
-  y3 = _mm256_shuffle_ps(r2, y1, 0x44);
-  y4 = _mm256_shuffle_ps(y2, y3, 0x88);
-  y5 = _mm256_shuffle_ps(y2, y3, 0xDD);
-  y6 = _mm256_permute2f128_ps(y4, y5, 0x20);
-
-  /* d c h g h g d c */
-  /* l k p o p o l k */
-  /* k o c g o k g c */
-  /* l p d h p l h d */
-  /* p l h d o k g c */
-  y2 = _mm256_shuffle_ps(r1, y0, 0xEE);
-  y3 = _mm256_shuffle_ps(r2, y1, 0xEE);
-  y4 = _mm256_shuffle_ps(y2, y3, 0x88);
-  y5 = _mm256_shuffle_ps(y2, y3, 0xDD);
-  y7 = _mm256_permute2f128_ps(y4, y5, 0x20);
-
-  glmm_store256(dest[0], y6);
-  glmm_store256(dest[2], y7);
-}
-
 #endif
 #endif /* cglm_mat_simd_avx_h */

include/cglm/simd/intrin.h

@@ -27,90 +27,64 @@
 #if defined( __SSE__ ) || defined( __SSE2__ )
 #  include <xmmintrin.h>
 #  include <emmintrin.h>
-
-/* OPTIONAL: You may save some instructions but latency (not sure) */
-#ifdef CGLM_USE_INT_DOMAIN
-#  define glmm_shuff1(xmm, z, y, x, w)                                        \
-     _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(xmm),                \
-                                        _MM_SHUFFLE(z, y, x, w)))
-#else
-#  define glmm_shuff1(xmm, z, y, x, w)                                        \
-     _mm_shuffle_ps(xmm, xmm, _MM_SHUFFLE(z, y, x, w))
-#endif
-
-#define glmm_shuff1x(xmm, x) glmm_shuff1(xmm, x, x, x, x)
-#define glmm_shuff2(a, b, z0, y0, x0, w0, z1, y1, x1, w1)                     \
-     glmm_shuff1(_mm_shuffle_ps(a, b, _MM_SHUFFLE(z0, y0, x0, w0)),           \
-                 z1, y1, x1, w1)
-
-static inline
-__m128
-glmm_dot(__m128 a, __m128 b) {
-  __m128 x0;
-  x0 = _mm_mul_ps(a, b);
-  x0 = _mm_add_ps(x0, glmm_shuff1(x0, 1, 0, 3, 2));
-  return _mm_add_ps(x0, glmm_shuff1(x0, 0, 1, 0, 1));
-}
-
-static inline
-__m128
-glmm_norm(__m128 a) {
-  return _mm_sqrt_ps(glmm_dot(a, a));
-}
-
-static inline
-__m128
-glmm_load3(float v[3]) {
-  __m128i xy;
-  __m128  z;
-
-  xy = _mm_loadl_epi64((const __m128i *)v);
-  z  = _mm_load_ss(&v[2]);
-
-  return _mm_movelh_ps(_mm_castsi128_ps(xy), z);
-}
-
-static inline
-void
-glmm_store3(__m128 vx, float v[3]) {
-  _mm_storel_pi((__m64 *)&v[0], vx);
-  _mm_store_ss(&v[2], glmm_shuff1(vx, 2, 2, 2, 2));
-}
-
-#ifdef CGLM_ALL_UNALIGNED
-#  define glmm_load(p)      _mm_loadu_ps(p)
-#  define glmm_store(p, a)  _mm_storeu_ps(p, a)
-#else
-#  define glmm_load(p)      _mm_load_ps(p)
-#  define glmm_store(p, a)  _mm_store_ps(p, a)
-#endif
-
-#endif
-
-/* x86, x64 */
-#if defined( __SSE__ ) || defined( __SSE2__ )
 #  define CGLM_SSE_FP 1
+#  ifndef CGLM_SIMD_x86
+#    define CGLM_SIMD_x86
+#  endif
+#endif
+
+#if defined(__SSE3__)
+#  include <x86intrin.h>
+#  ifndef CGLM_SIMD_x86
+#    define CGLM_SIMD_x86
+#  endif
+#endif
+
+#if defined(__SSE4_1__)
+#  include <smmintrin.h>
+#  ifndef CGLM_SIMD_x86
+#    define CGLM_SIMD_x86
+#  endif
+#endif
+
+#if defined(__SSE4_2__)
+#  include <nmmintrin.h>
+#  ifndef CGLM_SIMD_x86
+#    define CGLM_SIMD_x86
+#  endif
 #endif
 
 #ifdef __AVX__
+#  include <immintrin.h>
 #  define CGLM_AVX_FP 1
-
+#  ifndef CGLM_SIMD_x86
-#ifdef CGLM_ALL_UNALIGNED
+#    define CGLM_SIMD_x86
-#  define glmm_load256(p)      _mm256_loadu_ps(p)
+#  endif
-#  define glmm_store256(p, a)  _mm256_storeu_ps(p, a)
-#else
-#  define glmm_load256(p)      _mm256_load_ps(p)
-#  define glmm_store256(p, a)  _mm256_store_ps(p, a)
-#endif
-
 #endif
 
 /* ARM Neon */
-#if defined(__ARM_NEON) && defined(__ARM_NEON_FP)
+#if defined(__ARM_NEON)
 #  include <arm_neon.h>
+#  if defined(__ARM_NEON_FP)
 #    define CGLM_NEON_FP 1
-#else
+#    ifndef CGLM_SIMD_ARM
-#  undef  CGLM_NEON_FP
+#      define CGLM_SIMD_ARM
+#    endif
+#  endif
+#endif
+
+#if defined(CGLM_SIMD_x86) || defined(CGLM_NEON_FP)
+#  ifndef CGLM_SIMD
+#    define CGLM_SIMD
+#  endif
+#endif
+
+#if defined(CGLM_SIMD_x86)
+#  include "x86.h"
+#endif
+
+#if defined(CGLM_SIMD_ARM)
+#  include "arm.h"
 #endif
 
 #endif /* cglm_intrin_h */

include/cglm/simd/x86.h

@@ -0,0 +1,136 @@
+/*
+ * Copyright (c), Recep Aslantas.
+ *
+ * MIT License (MIT), http://opensource.org/licenses/MIT
+ * Full license can be found in the LICENSE file
+ */
+
+#ifndef cglm_simd_x86_h
+#define cglm_simd_x86_h
+#include "intrin.h"
+#ifdef CGLM_SIMD_x86
+
+#ifdef CGLM_ALL_UNALIGNED
+#  define glmm_load(p)      _mm_loadu_ps(p)
+#  define glmm_store(p, a)  _mm_storeu_ps(p, a)
+#else
+#  define glmm_load(p)      _mm_load_ps(p)
+#  define glmm_store(p, a)  _mm_store_ps(p, a)
+#endif
+
+#ifdef CGLM_USE_INT_DOMAIN
+#  define glmm_shuff1(xmm, z, y, x, w)                                        \
+     _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(xmm),                \
+                                        _MM_SHUFFLE(z, y, x, w)))
+#else
+#  define glmm_shuff1(xmm, z, y, x, w)                                        \
+       _mm_shuffle_ps(xmm, xmm, _MM_SHUFFLE(z, y, x, w))
+#endif
+
+#define glmm_shuff1x(xmm, x) glmm_shuff1(xmm, x, x, x, x)
+#define glmm_shuff2(a, b, z0, y0, x0, w0, z1, y1, x1, w1)                     \
+     glmm_shuff1(_mm_shuffle_ps(a, b, _MM_SHUFFLE(z0, y0, x0, w0)),           \
+                 z1, y1, x1, w1)
+
+#ifdef __AVX__
+#  ifdef CGLM_ALL_UNALIGNED
+#    define glmm_load256(p)      _mm256_loadu_ps(p)
+#    define glmm_store256(p, a)  _mm256_storeu_ps(p, a)
+#  else
+#    define glmm_load256(p)      _mm256_load_ps(p)
+#    define glmm_store256(p, a)  _mm256_store_ps(p, a)
+#  endif
+#endif
+
+static inline
+__m128
+glmm_vhadds(__m128 v) {
+#if defined(__SSE3__)
+  __m128 shuf, sums;
+  shuf = _mm_movehdup_ps(v);
+  sums = _mm_add_ps(v, shuf);
+  shuf = _mm_movehl_ps(shuf, sums);
+  sums = _mm_add_ss(sums, shuf);
+  return sums;
+#else
+  __m128 shuf, sums;
+  shuf = glmm_shuff1(v, 2, 3, 0, 1);
+  sums = _mm_add_ps(v, shuf);
+  shuf = _mm_movehl_ps(shuf, sums);
+  sums = _mm_add_ss(sums, shuf);
+  return sums;
+#endif
+}
+
+static inline
+float
+glmm_hadd(__m128 v) {
+  return _mm_cvtss_f32(glmm_vhadds(v));
+}
+
+static inline
+__m128
+glmm_vdots(__m128 a, __m128 b) {
+#if (defined(__SSE4_1__) || defined(__SSE4_2__)) && defined(CGLM_SSE4_DOT)
+  return _mm_dp_ps(a, b, 0xFF);
+#elif defined(__SSE3__) && defined(CGLM_SSE3_DOT)
+  __m128 x0, x1;
+  x0 = _mm_mul_ps(a, b);
+  x1 = _mm_hadd_ps(x0, x0);
+  return _mm_hadd_ps(x1, x1);
+#else
+  return glmm_vhadds(_mm_mul_ps(a, b));
+#endif
+}
+
+static inline
+__m128
+glmm_vdot(__m128 a, __m128 b) {
+#if (defined(__SSE4_1__) || defined(__SSE4_2__)) && defined(CGLM_SSE4_DOT)
+  return _mm_dp_ps(a, b, 0xFF);
+#elif defined(__SSE3__) && defined(CGLM_SSE3_DOT)
+  __m128 x0, x1;
+  x0 = _mm_mul_ps(a, b);
+  x1 = _mm_hadd_ps(x0, x0);
+  return _mm_hadd_ps(x1, x1);
+#else
+  __m128 x0;
+  x0 = _mm_mul_ps(a, b);
+  x0 = _mm_add_ps(x0, glmm_shuff1(x0, 1, 0, 3, 2));
+  return _mm_add_ps(x0, glmm_shuff1(x0, 0, 1, 0, 1));
+#endif
+}
+
+static inline
+float
+glmm_dot(__m128 a, __m128 b) {
+  return _mm_cvtss_f32(glmm_vdots(a, b));
+}
+
+static inline
+float
+glmm_norm(__m128 a) {
+  return _mm_cvtss_f32(_mm_sqrt_ss(glmm_vhadds(_mm_mul_ps(a, a))));
+}
+
+static inline
+__m128
+glmm_load3(float v[3]) {
+  __m128i xy;
+  __m128  z;
+
+  xy = _mm_loadl_epi64((const __m128i *)v);
+  z  = _mm_load_ss(&v[2]);
+
+  return _mm_movelh_ps(_mm_castsi128_ps(xy), z);
+}
+
+static inline
+void
+glmm_store3(__m128 vx, float v[3]) {
+  _mm_storel_pi((__m64 *)&v[0], vx);
+  _mm_store_ss(&v[2], glmm_shuff1(vx, 2, 2, 2, 2));
+}
+
+#endif
+#endif /* cglm_simd_x86_h */

include/cglm/vec4.h

@@ -122,6 +122,8 @@ void
 glm_vec4_copy(vec4 v, vec4 dest) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(dest, glmm_load(v));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vld1q_f32(v));
 #else
   dest[0] = v[0];
   dest[1] = v[1];
@@ -157,6 +159,8 @@ void
 glm_vec4_zero(vec4 v) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(v, _mm_setzero_ps());
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(v, vdupq_n_f32(0.0f));
 #else
   v[0] = 0.0f;
   v[1] = 0.0f;
@@ -175,6 +179,8 @@ void
 glm_vec4_one(vec4 v) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(v, _mm_set1_ps(1.0f));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(v, vdupq_n_f32(1.0f));
 #else
   v[0] = 1.0f;
   v[1] = 1.0f;
@@ -194,11 +200,8 @@ glm_vec4_one(vec4 v) {
 CGLM_INLINE
 float
 glm_vec4_dot(vec4 a, vec4 b) {
-#if defined( __SSE__ ) || defined( __SSE2__ )
+#if defined(CGLM_SIMD)
-  __m128 x0;
+  return glmm_dot(glmm_load(a), glmm_load(b));
-  x0 = _mm_mul_ps(glmm_load(a), glmm_load(b));
-  x0 = _mm_add_ps(x0, glmm_shuff1(x0, 1, 0, 3, 2));
-  return _mm_cvtss_f32(_mm_add_ss(x0, glmm_shuff1(x0, 0, 1, 0, 1)));
 #else
   return a[0] * b[0] + a[1] * b[1] + a[2] * b[2] + a[3] * b[3];
 #endif
@@ -218,15 +221,7 @@ glm_vec4_dot(vec4 a, vec4 b) {
 CGLM_INLINE
 float
 glm_vec4_norm2(vec4 v) {
-#if defined( __SSE__ ) || defined( __SSE2__ )
+  return glm_vec4_dot(v, v);
-  __m128 x0;
-  x0 = glmm_load(v);
-  x0 = _mm_mul_ps(x0, x0);
-  x0 = _mm_add_ps(x0, glmm_shuff1(x0, 1, 0, 3, 2));
-  return _mm_cvtss_f32(_mm_add_ss(x0, glmm_shuff1(x0, 0, 1, 0, 1)));
-#else
-  return v[0] * v[0] + v[1] * v[1] + v[2] * v[2] + v[3] * v[3];
-#endif
 }
 
 /*!
@@ -239,12 +234,10 @@ glm_vec4_norm2(vec4 v) {
 CGLM_INLINE
 float
 glm_vec4_norm(vec4 v) {
-#if defined( __SSE__ ) || defined( __SSE2__ )
+#if defined(CGLM_SIMD)
-  __m128 x0;
+  return glmm_norm(glmm_load(v));
-  x0 = glmm_load(v);
-  return _mm_cvtss_f32(_mm_sqrt_ss(glmm_dot(x0, x0)));
 #else
-  return sqrtf(glm_vec4_norm2(v));
+  return sqrtf(glm_vec4_dot(v, v));
 #endif
 }
 
@@ -260,6 +253,8 @@ void
 glm_vec4_add(vec4 a, vec4 b, vec4 dest) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(dest, _mm_add_ps(glmm_load(a), glmm_load(b)));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vaddq_f32(vld1q_f32(a), vld1q_f32(b)));
 #else
   dest[0] = a[0] + b[0];
   dest[1] = a[1] + b[1];
@@ -280,6 +275,8 @@ void
 glm_vec4_adds(vec4 v, float s, vec4 dest) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(dest, _mm_add_ps(glmm_load(v), _mm_set1_ps(s)));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vaddq_f32(vld1q_f32(v), vdupq_n_f32(s)));
 #else
   dest[0] = v[0] + s;
   dest[1] = v[1] + s;
@@ -300,6 +297,8 @@ void
 glm_vec4_sub(vec4 a, vec4 b, vec4 dest) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(dest, _mm_sub_ps(glmm_load(a), glmm_load(b)));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vsubq_f32(vld1q_f32(a), vld1q_f32(b)));
 #else
   dest[0] = a[0] - b[0];
   dest[1] = a[1] - b[1];
@@ -320,6 +319,8 @@ void
 glm_vec4_subs(vec4 v, float s, vec4 dest) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(dest, _mm_sub_ps(glmm_load(v), _mm_set1_ps(s)));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vsubq_f32(vld1q_f32(v), vdupq_n_f32(s)));
 #else
   dest[0] = v[0] - s;
   dest[1] = v[1] - s;
@@ -340,6 +341,8 @@ void
 glm_vec4_mul(vec4 a, vec4 b, vec4 dest) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(dest, _mm_mul_ps(glmm_load(a), glmm_load(b)));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vmulq_f32(vld1q_f32(a), vld1q_f32(b)));
 #else
   dest[0] = a[0] * b[0];
   dest[1] = a[1] * b[1];
@@ -360,6 +363,8 @@ void
 glm_vec4_scale(vec4 v, float s, vec4 dest) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(dest, _mm_mul_ps(glmm_load(v), _mm_set1_ps(s)));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vmulq_f32(vld1q_f32(v), vdupq_n_f32(s)));
 #else
   dest[0] = v[0] * s;
   dest[1] = v[1] * s;
@@ -442,6 +447,10 @@ glm_vec4_addadd(vec4 a, vec4 b, vec4 dest) {
   glmm_store(dest, _mm_add_ps(glmm_load(dest),
                               _mm_add_ps(glmm_load(a),
                                          glmm_load(b))));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vaddq_f32(vld1q_f32(dest),
+                            vaddq_f32(vld1q_f32(a),
+                                      vld1q_f32(b))));
 #else
   dest[0] += a[0] + b[0];
   dest[1] += a[1] + b[1];
@@ -466,6 +475,10 @@ glm_vec4_subadd(vec4 a, vec4 b, vec4 dest) {
   glmm_store(dest, _mm_add_ps(glmm_load(dest),
                               _mm_sub_ps(glmm_load(a),
                                          glmm_load(b))));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vaddq_f32(vld1q_f32(dest),
+                            vsubq_f32(vld1q_f32(a),
+                                      vld1q_f32(b))));
 #else
   dest[0] += a[0] - b[0];
   dest[1] += a[1] - b[1];
@@ -490,6 +503,10 @@ glm_vec4_muladd(vec4 a, vec4 b, vec4 dest) {
   glmm_store(dest, _mm_add_ps(glmm_load(dest),
                               _mm_mul_ps(glmm_load(a),
                                          glmm_load(b))));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vaddq_f32(vld1q_f32(dest),
+                            vmulq_f32(vld1q_f32(a),
+                                      vld1q_f32(b))));
 #else
   dest[0] += a[0] * b[0];
   dest[1] += a[1] * b[1];
@@ -514,6 +531,10 @@ glm_vec4_muladds(vec4 a, float s, vec4 dest) {
   glmm_store(dest, _mm_add_ps(glmm_load(dest),
                               _mm_mul_ps(glmm_load(a),
                                          _mm_set1_ps(s))));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vaddq_f32(vld1q_f32(dest),
+                            vsubq_f32(vld1q_f32(a),
+                                      vdupq_n_f32(s))));
 #else
   dest[0] += a[0] * s;
   dest[1] += a[1] * s;
@@ -538,6 +559,10 @@ glm_vec4_maxadd(vec4 a, vec4 b, vec4 dest) {
   glmm_store(dest, _mm_add_ps(glmm_load(dest),
                               _mm_max_ps(glmm_load(a),
                                          glmm_load(b))));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vaddq_f32(vld1q_f32(dest),
+                            vmaxq_f32(vld1q_f32(a),
+                                      vld1q_f32(b))));
 #else
   dest[0] += glm_max(a[0], b[0]);
   dest[1] += glm_max(a[1], b[1]);
@@ -562,6 +587,10 @@ glm_vec4_minadd(vec4 a, vec4 b, vec4 dest) {
   glmm_store(dest, _mm_add_ps(glmm_load(dest),
                               _mm_min_ps(glmm_load(a),
                                          glmm_load(b))));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vaddq_f32(vld1q_f32(dest),
+                            vminq_f32(vld1q_f32(a),
+                                      vld1q_f32(b))));
 #else
   dest[0] += glm_min(a[0], b[0]);
   dest[1] += glm_min(a[1], b[1]);
@@ -581,6 +610,8 @@ void
 glm_vec4_negate_to(vec4 v, vec4 dest) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(dest, _mm_xor_ps(glmm_load(v), _mm_set1_ps(-0.0f)));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, veorq_s32(vld1q_f32(v), vdupq_n_f32(-0.0f)));
 #else
   dest[0] = -v[0];
   dest[1] = -v[1];
@@ -614,7 +645,7 @@ glm_vec4_normalize_to(vec4 v, vec4 dest) {
   float  dot;
 
   x0   = glmm_load(v);
-  xdot = glmm_dot(x0, x0);
+  xdot = glmm_vdot(x0, x0);
   dot  = _mm_cvtss_f32(xdot);
 
   if (dot == 0.0f) {
@@ -658,10 +689,25 @@ 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
 }
 
 /*!
@@ -676,6 +722,8 @@ void
 glm_vec4_maxv(vec4 a, vec4 b, vec4 dest) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(dest, _mm_max_ps(glmm_load(a), glmm_load(b)));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vmaxq_f32(vld1q_f32(a), vld1q_f32(b)));
 #else
   dest[0] = glm_max(a[0], b[0]);
   dest[1] = glm_max(a[1], b[1]);
@@ -696,6 +744,8 @@ void
 glm_vec4_minv(vec4 a, vec4 b, vec4 dest) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(dest, _mm_min_ps(glmm_load(a), glmm_load(b)));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(dest, vminq_f32(vld1q_f32(a), vld1q_f32(b)));
 #else
   dest[0] = glm_min(a[0], b[0]);
   dest[1] = glm_min(a[1], b[1]);
@@ -717,6 +767,9 @@ glm_vec4_clamp(vec4 v, float minVal, float maxVal) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glmm_store(v, _mm_min_ps(_mm_max_ps(glmm_load(v), _mm_set1_ps(minVal)),
                            _mm_set1_ps(maxVal)));
+#elif defined(CGLM_NEON_FP)
+  vst1q_f32(v, vminq_f32(vmaxq_f32(vld1q_f32(v), vdupq_n_f32(minVal)),
+                         vdupq_n_f32(maxVal)));
 #else
   v[0] = glm_clamp(v[0], minVal, maxVal);
   v[1] = glm_clamp(v[1], minVal, maxVal);
@@ -747,4 +800,23 @@ glm_vec4_lerp(vec4 from, vec4 to, float t, vec4 dest) {
   glm_vec4_add(from, v, dest);
 }
 
+/*!
+ * @brief helper to fill vec4 as [S^3, S^2, S, 1]
+ *
+ * @param[in]   s    parameter
+ * @param[out]  dest destination
+ */
+CGLM_INLINE
+void
+glm_vec4_cubic(float s, vec4 dest) {
+  float ss;
+
+  ss = s * s;
+
+  dest[0] = ss * s;
+  dest[1] = ss;
+  dest[2] = s;
+  dest[3] = 1.0f;
+}
+
 #endif /* cglm_vec4_h */

makefile.am

@@ -57,7 +57,9 @@ cglm_HEADERS = include/cglm/version.h \
                include/cglm/color.h \
                include/cglm/project.h \
                include/cglm/sphere.h \
-                  include/cglm/ease.h
+               include/cglm/ease.h \
+               include/cglm/curve.h \
+               include/cglm/bezier.h
 
 cglm_calldir=$(includedir)/cglm/call
 cglm_call_HEADERS = include/cglm/call/mat4.h \
@@ -74,10 +76,14 @@ cglm_call_HEADERS = include/cglm/call/mat4.h \
                     include/cglm/call/box.h \
                     include/cglm/call/project.h \
                     include/cglm/call/sphere.h \
-                    include/cglm/call/ease.h
+                    include/cglm/call/ease.h \
+                    include/cglm/call/curve.h \
+                    include/cglm/call/bezier.h
 
 cglm_simddir=$(includedir)/cglm/simd
-cglm_simd_HEADERS = include/cglm/simd/intrin.h
+cglm_simd_HEADERS = include/cglm/simd/intrin.h \
+                    include/cglm/simd/x86.h \
+                    include/cglm/simd/arm.h
 
 cglm_simd_sse2dir=$(includedir)/cglm/simd/sse2
 cglm_simd_sse2_HEADERS = include/cglm/simd/sse2/affine.h \
@@ -107,7 +113,9 @@ libcglm_la_SOURCES=\
     src/box.c \
     src/project.c \
     src/sphere.c \
-    src/ease.c
+    src/ease.c \
+    src/curve.c \
+    src/bezier.c
 
 test_tests_SOURCES=\
     test/src/test_common.c \
@@ -121,7 +129,8 @@ test_tests_SOURCES=\
     test/src/test_vec4.c \
     test/src/test_vec3.c \
     test/src/test_mat3.c \
-    test/src/test_affine.c
+    test/src/test_affine.c \
+    test/src/test_bezier.c
 
 all-local:
 	sh ./post-build.sh

src/bezier.c

@@ -0,0 +1,27 @@
+/*
+ * Copyright (c), Recep Aslantas.
+ *
+ * MIT License (MIT), http://opensource.org/licenses/MIT
+ * Full license can be found in the LICENSE file
+ */
+
+#include "../include/cglm/cglm.h"
+#include "../include/cglm/call.h"
+
+CGLM_EXPORT
+float
+glmc_bezier(float s, float p0, float c0, float c1, float p1) {
+  return glm_bezier(s, p0, c0, c1, p1);
+}
+
+CGLM_EXPORT
+float
+glmc_hermite(float s, float p0, float t0, float t1, float p1) {
+  return glm_hermite(s, p0, t0, t1, p1);
+}
+
+CGLM_EXPORT
+float
+glmc_decasteljau(float prm, float p0, float c0, float c1, float p1) {
+  return glm_decasteljau(prm, p0, c0, c1, p1);
+}

src/curve.c

@@ -0,0 +1,15 @@
+/*
+ * Copyright (c), Recep Aslantas.
+ *
+ * MIT License (MIT), http://opensource.org/licenses/MIT
+ * Full license can be found in the LICENSE file
+ */
+
+#include "../include/cglm/cglm.h"
+#include "../include/cglm/call.h"
+
+CGLM_EXPORT
+float
+glmc_smc(float s, mat4 m, vec4 c) {
+  return glm_smc(s, m, c);
+}

src/mat4.c

@@ -151,3 +151,9 @@ void
 glmc_mat4_swap_row(mat4 mat, int row1, int row2) {
   glm_mat4_swap_row(mat, row1, row2);
 }
+
+CGLM_EXPORT
+float
+glmc_mat4_rmc(vec4 r, mat4 m, vec4 c) {
+  return glm_mat4_rmc(r, m, c);
+}

src/vec4.c

@@ -206,6 +206,12 @@ glmc_vec4_lerp(vec4 from, vec4 to, float t, vec4 dest) {
   glm_vec4_lerp(from, to, t, dest);
 }
 
+CGLM_EXPORT
+void
+glmc_vec4_cubic(float s, vec4 dest) {
+  glm_vec4_cubic(s, dest);
+}
+
 /* ext */
 
 CGLM_EXPORT

test/src/test_bezier.c

@@ -0,0 +1,65 @@
+/*
+ * Copyright (c), Recep Aslantas.
+ *
+ * MIT License (MIT), http://opensource.org/licenses/MIT
+ * Full license can be found in the LICENSE file
+ */
+
+#include "test_common.h"
+
+CGLM_INLINE
+float
+test_bezier_plain(float s, float p0, float c0, float c1, float p1) {
+  float x, xx, xxx, ss, sss;
+
+  x   = 1.0f - s;
+  xx  = x * x;
+  xxx = xx * x;
+  ss  = s * s;
+  sss = ss * s;
+
+  return p0 * xxx + 3.0f * (c0 * s * xx + c1 * ss * x) + p1 * sss;
+}
+
+CGLM_INLINE
+float
+test_hermite_plain(float s, float p0, float t0, float t1, float p1) {
+  float ss, sss;
+
+  ss  = s  * s;
+  sss = ss * s;
+
+  return p0 * (2.0f * sss - 3.0f * ss + 1.0f)
+       + t0 * (sss - 2.0f * ss + s)
+       + p1 * (-2.0f * sss + 3.0f * ss)
+       + t1 * (sss - ss);
+}
+
+void
+test_bezier(void **state) {
+  float s, p0, p1, c0, c1, smc, Bs, Bs_plain;
+
+  s        = test_rand();
+  p0       = test_rand();
+  p1       = test_rand();
+  c0       = test_rand();
+  c1       = test_rand();
+
+  /* test cubic bezier */
+  smc      = glm_smc(s, GLM_BEZIER_MAT, (vec4){p0, c0, c1, p1});
+  Bs       = glm_bezier(s, p0, c0, c1, p1);
+  Bs_plain = test_bezier_plain(s, p0, c0, c1, p1);
+
+  assert_true(glm_eq(Bs,  Bs_plain));
+  assert_true(glm_eq(smc, Bs_plain));
+  assert_true(glm_eq(Bs,  smc));
+
+  /* test cubic hermite */
+  smc      = glm_smc(s, GLM_HERMITE_MAT, (vec4){p0, p1, c0, c1});
+  Bs       = glm_hermite(s, p0, c0, c1, p1);
+  Bs_plain = test_hermite_plain(s, p0, c0, c1, p1);
+
+  assert_true(glm_eq(Bs,  Bs_plain));
+  assert_true(glm_eq(smc, Bs_plain));
+  assert_true(glm_eq(Bs,  smc));
+}

test/src/test_common.c

@@ -58,7 +58,7 @@ test_rand_vec4(vec4 dest) {
 }
 
 float
-test_rand_angle(void) {
+test_rand(void) {
   srand((unsigned int)time(NULL));
   return drand48();
 }

test/src/test_common.h

@@ -59,7 +59,7 @@ void
 test_rand_vec4(vec4 dest) ;
 
 float
-test_rand_angle(void);
+test_rand(void);
 
 void
 test_rand_quat(versor q);

test/src/test_main.c

@@ -38,7 +38,10 @@ main(int argc, const char * argv[]) {
     cmocka_unit_test(test_vec3),
 
     /* affine */
-    cmocka_unit_test(test_affine)
+    cmocka_unit_test(test_affine),
+
+    /* bezier */
+    cmocka_unit_test(test_bezier)
   };
 
   return cmocka_run_group_tests(tests, NULL, NULL);

test/src/test_tests.h

@@ -40,4 +40,7 @@ test_vec3(void **state);
 void
 test_affine(void **state);
 
+void
+test_bezier(void **state);
+
 #endif /* test_tests_h */

win/cglm.vcxproj

@@ -20,8 +20,10 @@
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="..\src\affine.c" />
+    <ClCompile Include="..\src\bezier.c" />
     <ClCompile Include="..\src\box.c" />
     <ClCompile Include="..\src\cam.c" />
+    <ClCompile Include="..\src\curve.c" />
     <ClCompile Include="..\src\dllmain.c" />
     <ClCompile Include="..\src\ease.c" />
     <ClCompile Include="..\src\euler.c" />
@@ -39,11 +41,14 @@
   <ItemGroup>
     <ClInclude Include="..\include\cglm\affine-mat.h" />
     <ClInclude Include="..\include\cglm\affine.h" />
+    <ClInclude Include="..\include\cglm\bezier.h" />
     <ClInclude Include="..\include\cglm\box.h" />
     <ClInclude Include="..\include\cglm\call.h" />
     <ClInclude Include="..\include\cglm\call\affine.h" />
+    <ClInclude Include="..\include\cglm\call\bezier.h" />
     <ClInclude Include="..\include\cglm\call\box.h" />
     <ClInclude Include="..\include\cglm\call\cam.h" />
+    <ClInclude Include="..\include\cglm\call\curve.h" />
     <ClInclude Include="..\include\cglm\call\ease.h" />
     <ClInclude Include="..\include\cglm\call\euler.h" />
     <ClInclude Include="..\include\cglm\call\frustum.h" />
@@ -60,6 +65,7 @@
     <ClInclude Include="..\include\cglm\cglm.h" />
     <ClInclude Include="..\include\cglm\color.h" />
     <ClInclude Include="..\include\cglm\common.h" />
+    <ClInclude Include="..\include\cglm\curve.h" />
     <ClInclude Include="..\include\cglm\ease.h" />
     <ClInclude Include="..\include\cglm\euler.h" />
     <ClInclude Include="..\include\cglm\frustum.h" />
@@ -69,6 +75,7 @@
     <ClInclude Include="..\include\cglm\plane.h" />
     <ClInclude Include="..\include\cglm\project.h" />
     <ClInclude Include="..\include\cglm\quat.h" />
+    <ClInclude Include="..\include\cglm\simd\arm.h" />
     <ClInclude Include="..\include\cglm\simd\avx\affine.h" />
     <ClInclude Include="..\include\cglm\simd\avx\mat4.h" />
     <ClInclude Include="..\include\cglm\simd\intrin.h" />
@@ -77,6 +84,7 @@
     <ClInclude Include="..\include\cglm\simd\sse2\mat3.h" />
     <ClInclude Include="..\include\cglm\simd\sse2\mat4.h" />
     <ClInclude Include="..\include\cglm\simd\sse2\quat.h" />
+    <ClInclude Include="..\include\cglm\simd\x86.h" />
     <ClInclude Include="..\include\cglm\sphere.h" />
     <ClInclude Include="..\include\cglm\types.h" />
     <ClInclude Include="..\include\cglm\util.h" />

win/cglm.vcxproj.filters

@@ -84,6 +84,12 @@
     <ClCompile Include="..\src\ease.c">
       <Filter>src</Filter>
     </ClCompile>
+    <ClCompile Include="..\src\curve.c">
+      <Filter>src</Filter>
+    </ClCompile>
+    <ClCompile Include="..\src\bezier.c">
+      <Filter>src</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="..\src\config.h">
@@ -233,5 +239,23 @@
     <ClInclude Include="..\include\cglm\ease.h">
       <Filter>include\cglm</Filter>
     </ClInclude>
+    <ClInclude Include="..\include\cglm\simd\arm.h">
+      <Filter>include\cglm\simd</Filter>
+    </ClInclude>
+    <ClInclude Include="..\include\cglm\simd\x86.h">
+      <Filter>include\cglm\simd</Filter>
+    </ClInclude>
+    <ClInclude Include="..\include\cglm\call\curve.h">
+      <Filter>include\cglm\call</Filter>
+    </ClInclude>
+    <ClInclude Include="..\include\cglm\curve.h">
+      <Filter>include\cglm</Filter>
+    </ClInclude>
+    <ClInclude Include="..\include\cglm\bezier.h">
+      <Filter>include\cglm</Filter>
+    </ClInclude>
+    <ClInclude Include="..\include\cglm\call\bezier.h">
+      <Filter>include\cglm\call</Filter>
+    </ClInclude>
   </ItemGroup>
 </Project>