arm, neon: neon/fma support for glm_mul_rot()

* use glmm_set1() for each for now

CMakeLists.txt

@@ -1,5 +1,5 @@
 cmake_minimum_required(VERSION 3.8.2)
-project(cglm VERSION 0.8.0 LANGUAGES C)
+project(cglm VERSION 0.8.2 LANGUAGES C)
 
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED YES)
@@ -93,6 +93,11 @@ target_include_directories(${PROJECT_NAME}
         ${CMAKE_CURRENT_SOURCE_DIR}/src
 )
 
+# Target for header-only usage
+add_library(${PROJECT_NAME}_headers INTERFACE)
+target_include_directories(${PROJECT_NAME}_headers INTERFACE
+  ${CMAKE_CURRENT_SOURCE_DIR}/include)
+
 # Test Configuration
 if(CGLM_USE_TEST)
   include(CTest)
@@ -121,3 +126,12 @@ install(EXPORT      ${PROJECT_NAME}
         NAMESPACE   ${PROJECT_NAME}::
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/${PROJECT_NAME})
 
+set(PACKAGE_NAME ${PROJECT_NAME})
+set(prefix ${CMAKE_INSTALL_PREFIX})
+set(exec_prefix ${CMAKE_INSTALL_PREFIX})
+set(includedir "\${prefix}/${CMAKE_INSTALL_INCLUDEDIR}")
+set(libdir "\${prefix}/${CMAKE_INSTALL_LIBDIR}")
+configure_file(${CMAKE_CURRENT_LIST_DIR}/cglm.pc.in ${CMAKE_BINARY_DIR}/cglm.pc @ONLY)
+
+install(FILES ${CMAKE_BINARY_DIR}/cglm.pc
+  DESTINATION ${CMAKE_INSTALL_LIBDIR}/pkgconfig)

CREDITS

@@ -74,3 +74,7 @@ Link to paper: http://webserver2.tecgraf.puc-rio.br/~mgattass/cg/trbRR/Fast%20Mi
 
 14. ARM NEON: Matrix Vector Multiplication
 https://stackoverflow.com/a/57793352/2676533
+
+16. ARM NEON Div
+
+http://github.com/microsoft/DirectXMath

README.md

@@ -168,6 +168,24 @@ option(CGLM_USE_C99 "" OFF) # C11
 option(CGLM_USE_TEST "Enable Tests" OFF) # for make check - make test
 ```
 
+#### Use as header-only library with your CMake project
+
+This requires no building or installation of cglm.
+
+* Example:
+
+``` cmake
+cmake_minimum_required(VERSION 3.8.2)
+
+project(<Your Project Name>)
+
+add_executable(${PROJECT_NAME} src/main.c)
+target_link_libraries(${LIBRARY_NAME} PRIVATE
+  cglm_headers)
+
+add_subdirectory(external/cglm/ EXCLUDE_FROM_ALL)
+```
+
 #### Use with your CMake project
 * Example:
 ```cmake

cglm.podspec

@@ -2,7 +2,7 @@ Pod::Spec.new do |s|
 
   # Description
   s.name         = "cglm"
-  s.version      = "0.7.9"
+  s.version      = "0.8.1"
   s.summary      = "📽 Highly Optimized Graphics Math (glm) for C"
   s.description  = <<-DESC
 cglm is math library for graphics programming for C. See the documentation or README for all features.

configure.ac

@@ -7,7 +7,7 @@
 #*****************************************************************************
 
 AC_PREREQ([2.69])
-AC_INIT([cglm], [0.8.0], [info@recp.me])
+AC_INIT([cglm], [0.8.2], [info@recp.me])
 AM_INIT_AUTOMAKE([-Wall -Werror foreign subdir-objects serial-tests])
 
 # Don't use the default cflags (-O2 -g), we set ours manually in Makefile.am.

docs/source/build.rst

@@ -32,6 +32,22 @@ If you don't want to install **cglm** to your system's folder you can get static
   option(CGLM_USE_C99 "" OFF) # C11 
   option(CGLM_USE_TEST "Enable Tests" OFF) # for make check - make test
 
+**Use as header-only library with your CMake project example**
+This requires no building or installation of cglm.
+
+.. code-block:: CMake
+  :linenos:
+
+  cmake_minimum_required(VERSION 3.8.2)
+  
+  project(<Your Project Name>)
+  
+  add_executable(${PROJECT_NAME} src/main.c)
+  target_link_libraries(${LIBRARY_NAME} PRIVATE
+    cglm_headers)
+  
+  add_subdirectory(external/cglm/ EXCLUDE_FROM_ALL)
+
 **Use with your CMake project example**
 
 .. code-block:: CMake

docs/source/conf.py

@@ -62,9 +62,9 @@ author = u'Recep Aslantas'
 # built documents.
 #
 # The short X.Y version.
-version = u'0.8.0'
+version = u'0.8.2'
 # The full version, including alpha/beta/rc tags.
-release = u'0.8.0'
+release = u'0.8.2'
 
 # The language for content autogenerated by Sphinx. Refer to documentation
 # for a list of supported languages.

include/cglm/affine-mat.h

@@ -26,6 +26,10 @@
 #  include "simd/avx/affine.h"
 #endif
 
+#ifdef CGLM_NEON_FP
+#  include "simd/neon/affine.h"
+#endif
+
 /*!
  * @brief this is similar to glm_mat4_mul but specialized to affine transform
  *
@@ -49,6 +53,8 @@ glm_mul(mat4 m1, mat4 m2, mat4 dest) {
   glm_mul_avx(m1, m2, dest);
 #elif defined( __SSE__ ) || defined( __SSE2__ )
   glm_mul_sse2(m1, m2, dest);
+#elif defined(CGLM_NEON_FP)
+  glm_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],
@@ -103,6 +109,8 @@ void
 glm_mul_rot(mat4 m1, mat4 m2, mat4 dest) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glm_mul_rot_sse2(m1, m2, dest);
+#elif defined(CGLM_NEON_FP)
+  glm_mul_rot_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],

include/cglm/affine.h

@@ -50,26 +50,22 @@
 CGLM_INLINE
 void
 glm_translate(mat4 m, vec3 v) {
-#if defined( __SSE__ ) || defined( __SSE2__ )
+#if defined(CGLM_SIMD)
+  glmm_128 m0, m1, m2, m3;
+
+  m0 = glmm_load(m[0]);
+  m1 = glmm_load(m[1]);
+  m2 = glmm_load(m[2]);
+  m3 = glmm_load(m[3]);
+
   glmm_store(m[3],
-             _mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_load(m[0]),
-                                              _mm_set1_ps(v[0])),
-                                   _mm_mul_ps(glmm_load(m[1]),
-                                              _mm_set1_ps(v[1]))),
-                        _mm_add_ps(_mm_mul_ps(glmm_load(m[2]),
-                                              _mm_set1_ps(v[2])),
-                                   glmm_load(m[3]))))
-  ;
+             glmm_fmadd(m0, glmm_set1(v[0]),
+                        glmm_fmadd(m1, glmm_set1(v[1]),
+                                   glmm_fmadd(m2, glmm_set1(v[2]), m3))));
 #else
-  vec4 v1, v2, v3;
-
-  glm_vec4_scale(m[0], v[0], v1);
-  glm_vec4_scale(m[1], v[1], v2);
-  glm_vec4_scale(m[2], v[2], v3);
-
-  glm_vec4_add(v1, m[3], m[3]);
-  glm_vec4_add(v2, m[3], m[3]);
-  glm_vec4_add(v3, m[3], m[3]);
+  glm_vec4_muladds(m[0], v[0], m[3]);
+  glm_vec4_muladds(m[1], v[1], m[3]);
+  glm_vec4_muladds(m[2], v[2], m[3]);
 #endif
 }
 
@@ -99,12 +95,8 @@ glm_translate_to(mat4 m, vec3 v, mat4 dest) {
 CGLM_INLINE
 void
 glm_translate_x(mat4 m, float x) {
-#if defined( __SSE__ ) || defined( __SSE2__ )
-  glmm_store(m[3],
-             _mm_add_ps(_mm_mul_ps(glmm_load(m[0]),
-                                   _mm_set1_ps(x)),
-                        glmm_load(m[3])))
-  ;
+#if defined(CGLM_SIMD)
+  glmm_store(m[3], glmm_fmadd(glmm_load(m[0]), glmm_set1(x), glmm_load(m[3])));
 #else
   vec4 v1;
   glm_vec4_scale(m[0], x, v1);
@@ -121,12 +113,8 @@ glm_translate_x(mat4 m, float x) {
 CGLM_INLINE
 void
 glm_translate_y(mat4 m, float y) {
-#if defined( __SSE__ ) || defined( __SSE2__ )
-  glmm_store(m[3],
-             _mm_add_ps(_mm_mul_ps(glmm_load(m[1]),
-                                   _mm_set1_ps(y)),
-                        glmm_load(m[3])))
-  ;
+#if defined(CGLM_SIMD)
+  glmm_store(m[3], glmm_fmadd(glmm_load(m[1]), glmm_set1(y), glmm_load(m[3])));
 #else
   vec4 v1;
   glm_vec4_scale(m[1], y, v1);
@@ -143,12 +131,8 @@ glm_translate_y(mat4 m, float y) {
 CGLM_INLINE
 void
 glm_translate_z(mat4 m, float z) {
-#if defined( __SSE__ ) || defined( __SSE2__ )
-  glmm_store(m[3],
-             _mm_add_ps(_mm_mul_ps(glmm_load(m[2]),
-                                   _mm_set1_ps(z)),
-                        glmm_load(m[3])))
-  ;
+#if defined(CGLM_SIMD)
+  glmm_store(m[3], glmm_fmadd(glmm_load(m[2]), glmm_set1(z), glmm_load(m[3])));
 #else
   vec4 v1;
   glm_vec4_scale(m[2], z, v1);

include/cglm/mat4.h

@@ -539,7 +539,9 @@ glm_mat4_scale_p(mat4 m, float s) {
 CGLM_INLINE
 void
 glm_mat4_scale(mat4 m, float s) {
-#if defined( __SSE__ ) || defined( __SSE2__ )
+#ifdef __AVX__
+  glm_mat4_scale_avx(m, s);
+#elif defined( __SSE__ ) || defined( __SSE2__ )
   glm_mat4_scale_sse2(m, s);
 #elif defined(CGLM_NEON_FP)
   glm_mat4_scale_neon(m, s);
@@ -560,6 +562,8 @@ float
 glm_mat4_det(mat4 mat) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   return glm_mat4_det_sse2(mat);
+#elif defined(CGLM_NEON_FP)
+  return glm_mat4_det_neon(mat);
 #else
   /* [square] det(A) = det(At) */
   float t[6];
@@ -593,6 +597,8 @@ void
 glm_mat4_inv(mat4 mat, mat4 dest) {
 #if defined( __SSE__ ) || defined( __SSE2__ )
   glm_mat4_inv_sse2(mat, dest);
+#elif defined(CGLM_NEON_FP)
+  glm_mat4_inv_neon(mat, dest);
 #else
   float t[6];
   float det;

include/cglm/simd/arm.h

@@ -10,19 +10,42 @@
 #include "intrin.h"
 #ifdef CGLM_SIMD_ARM
 
+#if defined(_M_ARM64) || defined(_M_HYBRID_X86_ARM64) || defined(_M_ARM64EC) || defined(__aarch64__)
+# define CGLM_ARM64 1
+#endif
+
 #define glmm_load(p)      vld1q_f32(p)
 #define glmm_store(p, a)  vst1q_f32(p, a)
 
+#define glmm_set1(x) vdupq_n_f32(x)
+#define glmm_128     float32x4_t
+
+#define glmm_splat_x(x) vdupq_lane_f32(vget_low_f32(x),  0)
+#define glmm_splat_y(x) vdupq_lane_f32(vget_low_f32(x),  1)
+#define glmm_splat_z(x) vdupq_lane_f32(vget_high_f32(x), 0)
+#define glmm_splat_w(x) vdupq_lane_f32(vget_high_f32(x), 1)
+
+#define glmm_xor(a, b)                                                        \
+  vreinterpretq_f32_s32(veorq_s32(vreinterpretq_s32_f32(a),                   \
+                                  vreinterpretq_s32_f32(b)))
+
 static inline
 float32x4_t
 glmm_abs(float32x4_t v) {
   return vabsq_f32(v);
 }
 
+static inline
+float32x4_t
+glmm_vhadd(float32x4_t v) {
+  v = vaddq_f32(v, vrev64q_f32(v));
+  return vaddq_f32(v, vcombine_f32(vget_high_f32(v), vget_low_f32(v)));
+}
+
 static inline
 float
 glmm_hadd(float32x4_t v) {
-#if defined(__aarch64__)
+#if CGLM_ARM64
   return vaddvq_f32(v);
 #else
   v = vaddq_f32(v, vrev64q_f32(v));
@@ -79,5 +102,58 @@ glmm_norm_inf(float32x4_t a) {
   return glmm_hmax(glmm_abs(a));
 }
 
+static inline
+float32x4_t
+glmm_div(float32x4_t a, float32x4_t b) {
+#if CGLM_ARM64
+  return vdivq_f32(a, b);
+#else
+  /* 2 iterations of Newton-Raphson refinement of reciprocal */
+  float32x4_t r0, r1;
+  r0 = vrecpeq_f32(b);
+  r1 = vrecpsq_f32(r0, b);
+  r0 = vmulq_f32(r1, r0);
+  r1 = vrecpsq_f32(r0, b);
+  r0 = vmulq_f32(r1, r0);
+  return vmulq_f32(a, r0);
+#endif
+}
+
+static inline
+float32x4_t
+glmm_fmadd(float32x4_t a, float32x4_t b, float32x4_t c) {
+#if CGLM_ARM64
+  return vfmaq_f32(c, a, b); /* why vfmaq_f32 is slower than vmlaq_f32 ??? */
+#else
+  return vmlaq_f32(c, a, b);
+#endif
+}
+
+static inline
+float32x4_t
+glmm_fnmadd(float32x4_t a, float32x4_t b, float32x4_t c) {
+#if CGLM_ARM64
+  return vfmsq_f32(c, a, b);
+#else
+  return vmlsq_f32(c, a, b);
+#endif
+}
+
+static inline
+float32x4_t
+glmm_fmsub(float32x4_t a, float32x4_t b, float32x4_t c) {
+#if CGLM_ARM64
+  return vfmsq_f32(c, a, b);
+#else
+  return vmlsq_f32(c, a, b);
+#endif
+}
+
+static inline
+float32x4_t
+glmm_fnmsub(float32x4_t a, float32x4_t b, float32x4_t c) {
+  return vsubq_f32(vdupq_n_f32(0.0f), glmm_fmadd(a, b, c));
+}
+
 #endif
 #endif /* cglm_simd_arm_h */

include/cglm/simd/avx/mat4.h

@@ -14,6 +14,16 @@
 
 #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) {

include/cglm/simd/neon/affine.h

@@ -0,0 +1,80 @@
+/*
+ * Copyright (c), Recep Aslantas.
+ *
+ * MIT License (MIT), http://opensource.org/licenses/MIT
+ * Full license can be found in the LICENSE file
+ */
+
+#ifndef cglm_affine_neon_h
+#define cglm_affine_neon_h
+#if defined(__ARM_NEON_FP)
+
+#include "../../common.h"
+#include "../intrin.h"
+
+CGLM_INLINE
+void
+glm_mul_neon(mat4 m1, mat4 m2, mat4 dest) {
+  /* D = R * L (Column-Major) */
+
+  glmm_128 l0, l1, l2, l3, r0, r1, r2, r3, v0, v1, v2, v3;
+
+  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]);
+
+  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);
+
+  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);
+
+  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);
+
+  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
+void
+glm_mul_rot_neon(mat4 m1, mat4 m2, mat4 dest) {
+  /* D = R * L (Column-Major) */
+
+  glmm_128 l0, l1, l2, r0, r1, r2, v0, v1, v2;
+
+  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]);
+
+  v0 = vmulq_f32(glmm_splat_x(r0), l0);
+  v1 = vmulq_f32(glmm_splat_x(r1), l0);
+  v2 = vmulq_f32(glmm_splat_x(r2), l0);
+
+  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);
+
+  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);
+
+  glmm_store(dest[0], v0);
+  glmm_store(dest[1], v1);
+  glmm_store(dest[2], v2);
+  glmm_store(dest[3], glmm_load(m1[3]));
+}
+
+#endif
+#endif /* cglm_affine_neon_h */

include/cglm/simd/neon/mat4.h

@@ -42,41 +42,38 @@ 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]);
+  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
@@ -101,5 +98,216 @@ glm_mat4_mulv_neon(mat4 m, vec4 v, vec4 dest) {
   vst1q_f32(dest, l0);
 }
 
+CGLM_INLINE
+float
+glm_mat4_det_neon(mat4 mat) {
+  float32x4_t   r0, r1, r2, r3, x0, x1, x2;
+  float32x2_t   ij, op, mn, kl, nn, mm, jj, ii, gh, ef, t12, t34;
+  float32x4x2_t a1;
+  float32x4_t   x3 = { 0.f, -0.f, 0.f, -0.f };
+
+  /* 127 <- 0, [square] det(A) = det(At) */
+  r0 = glmm_load(mat[0]);              /* d c b a */
+  r1 = vrev64q_f32(glmm_load(mat[1])); /* g h e f */
+  r2 = vrev64q_f32(glmm_load(mat[2])); /* l k i j */
+  r3 = vrev64q_f32(glmm_load(mat[3])); /* o p m n */
+
+  gh = vget_high_f32(r1);
+  ef = vget_low_f32(r1);
+  kl = vget_high_f32(r2);
+  ij = vget_low_f32(r2);
+  op = vget_high_f32(r3);
+  mn = vget_low_f32(r3);
+  mm = vdup_lane_f32(mn, 1);
+  nn = vdup_lane_f32(mn, 0);
+  ii = vdup_lane_f32(ij, 1);
+  jj = vdup_lane_f32(ij, 0);
+  
+  /*
+   t[1] = j * p - n * l;
+   t[2] = j * o - n * k;
+   t[3] = i * p - m * l;
+   t[4] = i * o - m * k;
+   */
+  x0 = glmm_fnmadd(vcombine_f32(kl, kl), vcombine_f32(nn, mm),
+                   vmulq_f32(vcombine_f32(op, op), vcombine_f32(jj, ii)));
+
+  t12 = vget_low_f32(x0);
+  t34 = vget_high_f32(x0);
+  
+  /* 1 3 1 3 2 4 2 4 */
+  a1 = vuzpq_f32(x0, x0);
+  
+  /*
+   t[0] = k * p - o * l;
+   t[0] = k * p - o * l;
+   t[5] = i * n - m * j;
+   t[5] = i * n - m * j;
+   */
+  x1 = glmm_fnmadd(vcombine_f32(vdup_lane_f32(kl, 0), jj),
+                   vcombine_f32(vdup_lane_f32(op, 1), mm),
+                   vmulq_f32(vcombine_f32(vdup_lane_f32(op, 0), nn),
+                             vcombine_f32(vdup_lane_f32(kl, 1), ii)));
+
+  /*
+     a * (f * t[0] - g * t[1] + h * t[2])
+   - b * (e * t[0] - g * t[3] + h * t[4])
+   + c * (e * t[1] - f * t[3] + h * t[5])
+   - d * (e * t[2] - f * t[4] + g * t[5])
+   */
+  x2 = glmm_fnmadd(vcombine_f32(vdup_lane_f32(gh, 1), vdup_lane_f32(ef, 0)),
+                   vcombine_f32(vget_low_f32(a1.val[0]), t34),
+                   vmulq_f32(vcombine_f32(ef, vdup_lane_f32(ef, 1)),
+                             vcombine_f32(vget_low_f32(x1), t12)));
+
+  x2 = glmm_fmadd(vcombine_f32(vdup_lane_f32(gh, 0), gh),
+                  vcombine_f32(vget_low_f32(a1.val[1]), vget_high_f32(x1)), x2);
+
+  x2 = glmm_xor(x2, x3);
+
+  return glmm_hadd(vmulq_f32(x2, r0));
+}
+
+CGLM_INLINE
+void
+glm_mat4_inv_neon(mat4 mat, mat4 dest) {
+  float32x4_t   r0, r1, r2, r3,
+                v0, v1, v2, v3,
+                t0, t1, t2, t3, t4, t5,
+                x0, x1, x2, x3, x4, x5, x6, x7, x8;
+  float32x4x2_t a1;
+  float32x2_t   lp, ko, hg, jn, im, fe, ae, bf, cg, dh;
+  float32x4_t   x9 = { -0.f, 0.f, -0.f, 0.f };
+
+  x8 = vrev64q_f32(x9);
+
+  /* 127 <- 0 */
+  r0 = glmm_load(mat[0]); /* d c b a */
+  r1 = glmm_load(mat[1]); /* h g f e */
+  r2 = glmm_load(mat[2]); /* l k j i */
+  r3 = glmm_load(mat[3]); /* p o n m */
+  
+  /* l p k o, j n i m */
+  a1  = vzipq_f32(r3, r2);
+  
+  jn  = vget_high_f32(a1.val[0]);
+  im  = vget_low_f32(a1.val[0]);
+  lp  = vget_high_f32(a1.val[1]);
+  ko  = vget_low_f32(a1.val[1]);
+  hg  = vget_high_f32(r1);
+
+  x1  = vcombine_f32(vdup_lane_f32(lp, 0), lp);                   /* l p p p */
+  x2  = vcombine_f32(vdup_lane_f32(ko, 0), ko);                   /* k o o o */
+  x0  = vcombine_f32(vdup_lane_f32(lp, 1), vdup_lane_f32(hg, 1)); /* h h l l */
+  x3  = vcombine_f32(vdup_lane_f32(ko, 1), vdup_lane_f32(hg, 0)); /* g g k k */
+  
+  /* t1[0] = k * p - o * l;
+     t1[0] = k * p - o * l;
+     t2[0] = g * p - o * h;
+     t3[0] = g * l - k * h; */
+  t0 = glmm_fnmadd(x2, x0, vmulq_f32(x3, x1));
+
+  fe = vget_low_f32(r1);
+  x4 = vcombine_f32(vdup_lane_f32(jn, 0), jn);                   /* j n n n */
+  x5 = vcombine_f32(vdup_lane_f32(jn, 1), vdup_lane_f32(fe, 1)); /* f f j j */
+  
+  /* t1[1] = j * p - n * l;
+     t1[1] = j * p - n * l;
+     t2[1] = f * p - n * h;
+     t3[1] = f * l - j * h; */
+   t1 = glmm_fnmadd(x4, x0, vmulq_f32(x5, x1));
+  
+  /* t1[2] = j * o - n * k
+     t1[2] = j * o - n * k;
+     t2[2] = f * o - n * g;
+     t3[2] = f * k - j * g; */
+  t2 = glmm_fnmadd(x4, x3, vmulq_f32(x5, x2));
+  
+  x6 = vcombine_f32(vdup_lane_f32(im, 1), vdup_lane_f32(fe, 0)); /* e e i i */
+  x7 = vcombine_f32(vdup_lane_f32(im, 0), im);                   /* i m m m */
+  
+  /* t1[3] = i * p - m * l;
+     t1[3] = i * p - m * l;
+     t2[3] = e * p - m * h;
+     t3[3] = e * l - i * h; */
+  t3 = glmm_fnmadd(x7, x0, vmulq_f32(x6, x1));
+  
+  /* t1[4] = i * o - m * k;
+     t1[4] = i * o - m * k;
+     t2[4] = e * o - m * g;
+     t3[4] = e * k - i * g; */
+  t4 = glmm_fnmadd(x7, x3, vmulq_f32(x6, x2));
+  
+  /* t1[5] = i * n - m * j;
+     t1[5] = i * n - m * j;
+     t2[5] = e * n - m * f;
+     t3[5] = e * j - i * f; */
+  t5 = glmm_fnmadd(x7, x5, vmulq_f32(x6, x4));
+  
+  /* h d f b, g c e a */
+  a1 = vtrnq_f32(r0, r1);
+  
+  x4 = vrev64q_f32(a1.val[0]); /* c g a e */
+  x5 = vrev64q_f32(a1.val[1]); /* d h b f */
+
+  ae = vget_low_f32(x4);
+  cg = vget_high_f32(x4);
+  bf = vget_low_f32(x5);
+  dh = vget_high_f32(x5);
+  
+  x0 = vcombine_f32(ae, vdup_lane_f32(ae, 1)); /* a a a e */
+  x1 = vcombine_f32(bf, vdup_lane_f32(bf, 1)); /* b b b f */
+  x2 = vcombine_f32(cg, vdup_lane_f32(cg, 1)); /* c c c g */
+  x3 = vcombine_f32(dh, vdup_lane_f32(dh, 1)); /* d d d h */
+  
+  /*
+   dest[0][0] =  f * t1[0] - g * t1[1] + h * t1[2];
+   dest[0][1] =-(b * t1[0] - c * t1[1] + d * t1[2]);
+   dest[0][2] =  b * t2[0] - c * t2[1] + d * t2[2];
+   dest[0][3] =-(b * t3[0] - c * t3[1] + d * t3[2]); */
+  v0 = glmm_xor(glmm_fmadd(x3, t2, glmm_fnmadd(x2, t1, vmulq_f32(x1, t0))), x8);
+  
+  /*
+   dest[2][0] =  e * t1[1] - f * t1[3] + h * t1[5];
+   dest[2][1] =-(a * t1[1] - b * t1[3] + d * t1[5]);
+   dest[2][2] =  a * t2[1] - b * t2[3] + d * t2[5];
+   dest[2][3] =-(a * t3[1] - b * t3[3] + d * t3[5]);*/
+  v2 = glmm_xor(glmm_fmadd(x3, t5, glmm_fnmadd(x1, t3, vmulq_f32(x0, t1))), x8);
+
+  /*
+   dest[1][0] =-(e * t1[0] - g * t1[3] + h * t1[4]);
+   dest[1][1] =  a * t1[0] - c * t1[3] + d * t1[4];
+   dest[1][2] =-(a * t2[0] - c * t2[3] + d * t2[4]);
+   dest[1][3] =  a * t3[0] - c * t3[3] + d * t3[4]; */
+  v1 = glmm_xor(glmm_fmadd(x3, t4, glmm_fnmadd(x2, t3, vmulq_f32(x0, t0))), x9);
+  
+  /*
+   dest[3][0] =-(e * t1[2] - f * t1[4] + g * t1[5]);
+   dest[3][1] =  a * t1[2] - b * t1[4] + c * t1[5];
+   dest[3][2] =-(a * t2[2] - b * t2[4] + c * t2[5]);
+   dest[3][3] =  a * t3[2] - b * t3[4] + c * t3[5]; */
+  v3 = glmm_xor(glmm_fmadd(x2, t5, glmm_fnmadd(x1, t4, vmulq_f32(x0, t2))), x9);
+
+  /* determinant */
+  x0 = vcombine_f32(vget_low_f32(vzipq_f32(v0, v1).val[0]),
+                    vget_low_f32(vzipq_f32(v2, v3).val[0]));
+
+  /*
+  x0 = glmm_div(glmm_set1(1.0f), glmm_vhadd(vmulq_f32(x0, r0)));
+
+  glmm_store(dest[0], vmulq_f32(v0, x0));
+  glmm_store(dest[1], vmulq_f32(v1, x0));
+  glmm_store(dest[2], vmulq_f32(v2, x0));
+  glmm_store(dest[3], vmulq_f32(v3, x0));
+  */
+
+  x0 = glmm_vhadd(vmulq_f32(x0, r0));
+
+  glmm_store(dest[0], glmm_div(v0, x0));
+  glmm_store(dest[1], glmm_div(v1, x0));
+  glmm_store(dest[2], glmm_div(v2, x0));
+  glmm_store(dest[3], glmm_div(v3, x0));
+}
+
 #endif
 #endif /* cglm_mat4_neon_h */

include/cglm/simd/sse2/affine.h

@@ -22,31 +22,31 @@ glm_mul_sse2(mat4 m1, mat4 m2, mat4 dest) {
   l1 = glmm_load(m1[1]);
   l2 = glmm_load(m1[2]);
   l3 = glmm_load(m1[3]);
-
+  
   r = glmm_load(m2[0]);
   glmm_store(dest[0],
-             _mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
-                                   _mm_mul_ps(glmm_shuff1x(r, 1), l1)),
-                        _mm_mul_ps(glmm_shuff1x(r, 2), l2)));
-
+             glmm_fmadd(glmm_splat(r, 0), l0,
+                        glmm_fmadd(glmm_splat(r, 1), l1,
+                                   _mm_mul_ps(glmm_splat(r, 2), l2))));
+  
   r = glmm_load(m2[1]);
   glmm_store(dest[1],
-             _mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
-                                   _mm_mul_ps(glmm_shuff1x(r, 1), l1)),
-                        _mm_mul_ps(glmm_shuff1x(r, 2), l2)));
+             glmm_fmadd(glmm_splat(r, 0), l0,
+                        glmm_fmadd(glmm_splat(r, 1), l1,
+                                   _mm_mul_ps(glmm_splat(r, 2), l2))));
 
   r = glmm_load(m2[2]);
   glmm_store(dest[2],
-             _mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
-                                   _mm_mul_ps(glmm_shuff1x(r, 1), l1)),
-                        _mm_mul_ps(glmm_shuff1x(r, 2), l2)));
+             glmm_fmadd(glmm_splat(r, 0), l0,
+                        glmm_fmadd(glmm_splat(r, 1), l1,
+                                   _mm_mul_ps(glmm_splat(r, 2), l2))));
 
   r = glmm_load(m2[3]);
   glmm_store(dest[3],
-             _mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
-                                   _mm_mul_ps(glmm_shuff1x(r, 1), l1)),
-                        _mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 2), l2),
-                                   _mm_mul_ps(glmm_shuff1x(r, 3), l3))));
+       glmm_fmadd(glmm_splat(r, 0), l0,
+                  glmm_fmadd(glmm_splat(r, 1), l1,
+                             glmm_fmadd(glmm_splat(r, 2), l2,
+                                        _mm_mul_ps(glmm_splat(r, 3), l3)))));
 }
 
 CGLM_INLINE
@@ -62,21 +62,22 @@ glm_mul_rot_sse2(mat4 m1, mat4 m2, mat4 dest) {
 
   r = glmm_load(m2[0]);
   glmm_store(dest[0],
-             _mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
-                                   _mm_mul_ps(glmm_shuff1x(r, 1), l1)),
-                        _mm_mul_ps(glmm_shuff1x(r, 2), l2)));
-
+             glmm_fmadd(glmm_splat(r, 0), l0,
+                        glmm_fmadd(glmm_splat(r, 1), l1,
+                                   _mm_mul_ps(glmm_splat(r, 2), l2))));
+  
   r = glmm_load(m2[1]);
   glmm_store(dest[1],
-             _mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
-                                   _mm_mul_ps(glmm_shuff1x(r, 1), l1)),
-                        _mm_mul_ps(glmm_shuff1x(r, 2), l2)));
-
+             glmm_fmadd(glmm_splat(r, 0), l0,
+                        glmm_fmadd(glmm_splat(r, 1), l1,
+                                   _mm_mul_ps(glmm_splat(r, 2), l2))));
+  
+  
   r = glmm_load(m2[2]);
   glmm_store(dest[2],
-             _mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
-                                   _mm_mul_ps(glmm_shuff1x(r, 1), l1)),
-                        _mm_mul_ps(glmm_shuff1x(r, 2), l2)));
+             glmm_fmadd(glmm_splat(r, 0), l0,
+                        glmm_fmadd(glmm_splat(r, 1), l1,
+                                   _mm_mul_ps(glmm_splat(r, 2), l2))));
 
   glmm_store(dest[3], l3);
 }
@@ -94,9 +95,9 @@ glm_inv_tr_sse2(mat4 mat) {
 
   _MM_TRANSPOSE4_PS(r0, r1, r2, x1);
 
-  x0 = _mm_add_ps(_mm_mul_ps(r0, glmm_shuff1(r3, 0, 0, 0, 0)),
-                  _mm_mul_ps(r1, glmm_shuff1(r3, 1, 1, 1, 1)));
-  x0 = _mm_add_ps(x0, _mm_mul_ps(r2, glmm_shuff1(r3, 2, 2, 2, 2)));
+  x0 = glmm_fmadd(r0, glmm_shuff1(r3, 0, 0, 0, 0),
+             glmm_fmadd(r1, glmm_shuff1(r3, 1, 1, 1, 1),
+                        _mm_mul_ps(r2, glmm_shuff1(r3, 2, 2, 2, 2))));
   x0 = _mm_xor_ps(x0, _mm_set1_ps(-0.f));
 
   x0 = _mm_add_ps(x0, x1);

include/cglm/simd/sse2/mat2.h

@@ -26,11 +26,11 @@ glm_mat2_mul_sse2(mat2 m1, mat2 m2, mat2 dest) {
    dest[1][0] = a * g + c * h;
    dest[1][1] = b * g + d * h;
    */
-  x0 = _mm_mul_ps(_mm_movelh_ps(x1, x1), glmm_shuff1(x2, 2, 2, 0, 0));
-  x1 = _mm_mul_ps(_mm_movehl_ps(x1, x1), glmm_shuff1(x2, 3, 3, 1, 1));
-  x1 = _mm_add_ps(x0, x1);
+  x0 = glmm_fmadd(_mm_movelh_ps(x1, x1), glmm_shuff1(x2, 2, 2, 0, 0),
+                  _mm_mul_ps(_mm_movehl_ps(x1, x1),
+                             glmm_shuff1(x2, 3, 3, 1, 1)));
 
-  glmm_store(dest[0], x1);
+  glmm_store(dest[0], x0);
 }
 
 CGLM_INLINE

include/cglm/simd/sse2/mat3.h

@@ -30,24 +30,17 @@ glm_mat3_mul_sse2(mat3 m1, mat3 m2, mat3 dest) {
   x1 = glmm_shuff2(l0, l1, 1, 0, 3, 3, 0, 3, 2, 0);
   x2 = glmm_shuff2(l1, l2, 0, 0, 3, 2, 0, 2, 1, 0);
 
-  x0 = _mm_add_ps(_mm_mul_ps(glmm_shuff1(l0, 0, 2, 1, 0),
-                             glmm_shuff1(r0, 3, 0, 0, 0)),
-                  _mm_mul_ps(x1, glmm_shuff2(r0, r1, 0, 0, 1, 1, 2, 0, 0, 0)));
-
-  x0 = _mm_add_ps(x0,
-                  _mm_mul_ps(x2, glmm_shuff2(r0, r1, 1, 1, 2, 2, 2, 0, 0, 0)));
+  x0 = glmm_fmadd(glmm_shuff1(l0, 0, 2, 1, 0), glmm_shuff1(r0, 3, 0, 0, 0),
+                  glmm_fmadd(x1, glmm_shuff2(r0, r1, 0, 0, 1, 1, 2, 0, 0, 0),
+                             _mm_mul_ps(x2, glmm_shuff2(r0, r1, 1, 1, 2, 2, 2, 0, 0, 0))));
 
   _mm_storeu_ps(dest[0], x0);
 
-  x0 = _mm_add_ps(_mm_mul_ps(glmm_shuff1(l0, 1, 0, 2, 1),
-                             _mm_shuffle_ps(r0, r1, _MM_SHUFFLE(2, 2, 3, 3))),
-                  _mm_mul_ps(glmm_shuff1(x1, 1, 0, 2, 1),
-                             glmm_shuff1(r1, 3, 3, 0, 0)));
-
-  x0 = _mm_add_ps(x0,
-                  _mm_mul_ps(glmm_shuff1(x2, 1, 0, 2, 1),
-                             _mm_shuffle_ps(r1, r2, _MM_SHUFFLE(0, 0, 1, 1))));
-
+  x0 = glmm_fmadd(glmm_shuff1(l0, 1, 0, 2, 1), _mm_shuffle_ps(r0, r1, _MM_SHUFFLE(2, 2, 3, 3)),
+                  glmm_fmadd(glmm_shuff1(x1, 1, 0, 2, 1), glmm_shuff1(r1, 3, 3, 0, 0),
+                             _mm_mul_ps(glmm_shuff1(x2, 1, 0, 2, 1),
+                                        _mm_shuffle_ps(r1, r2, _MM_SHUFFLE(0, 0, 1, 1)))));
+  
   _mm_storeu_ps(&dest[1][1], x0);
 
   dest[2][2] = m1[0][2] * m2[2][0]

include/cglm/simd/sse2/mat4.h

@@ -55,47 +55,37 @@ glm_mat4_mul_sse2(mat4 m1, mat4 m2, mat4 dest) {
   l1 = glmm_load(m1[1]);
   l2 = glmm_load(m1[2]);
   l3 = glmm_load(m1[3]);
+  
+#define XX(C)                                                                 \
+                                                                              \
+  r = glmm_load(m2[C]);                                                       \
+  glmm_store(dest[C],                                                         \
+       glmm_fmadd(glmm_splat(r, 0), l0,                                       \
+                   glmm_fmadd(glmm_splat(r, 1), l1,                           \
+                             glmm_fmadd(glmm_splat(r, 2), l2,                 \
+                                         _mm_mul_ps(glmm_splat(r, 3), l3)))));
 
-  r = glmm_load(m2[0]);
-  glmm_store(dest[0],
-             _mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
-                                   _mm_mul_ps(glmm_shuff1x(r, 1), l1)),
-                        _mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 2), l2),
-                                   _mm_mul_ps(glmm_shuff1x(r, 3), l3))));
-  r = glmm_load(m2[1]);
-  glmm_store(dest[1],
-             _mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
-                                   _mm_mul_ps(glmm_shuff1x(r, 1), l1)),
-                        _mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 2), l2),
-                                   _mm_mul_ps(glmm_shuff1x(r, 3), l3))));
-  r = glmm_load(m2[2]);
-  glmm_store(dest[2],
-             _mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
-                                   _mm_mul_ps(glmm_shuff1x(r, 1), l1)),
-                        _mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 2), l2),
-                                   _mm_mul_ps(glmm_shuff1x(r, 3), l3))));
+  XX(0);
+  XX(1);
+  XX(2);
+  XX(3);
 
-  r = glmm_load(m2[3]);
-  glmm_store(dest[3],
-             _mm_add_ps(_mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 0), l0),
-                                   _mm_mul_ps(glmm_shuff1x(r, 1), l1)),
-                        _mm_add_ps(_mm_mul_ps(glmm_shuff1x(r, 2), l2),
-                                   _mm_mul_ps(glmm_shuff1x(r, 3), l3))));
+#undef XX
 }
 
 CGLM_INLINE
 void
 glm_mat4_mulv_sse2(mat4 m, vec4 v, vec4 dest) {
-  __m128 x0, x1, x2;
-
+  __m128 x0, x1;
+  
   x0 = glmm_load(v);
-  x1 = _mm_add_ps(_mm_mul_ps(glmm_load(m[0]), glmm_shuff1x(x0, 0)),
-                  _mm_mul_ps(glmm_load(m[1]), glmm_shuff1x(x0, 1)));
-
-  x2 = _mm_add_ps(_mm_mul_ps(glmm_load(m[2]), glmm_shuff1x(x0, 2)),
-                  _mm_mul_ps(glmm_load(m[3]), glmm_shuff1x(x0, 3)));
-
-  glmm_store(dest, _mm_add_ps(x1, x2));
+  x1 = glmm_fmadd(glmm_load(m[0]), glmm_splat(x0, 0),
+                  glmm_fmadd(glmm_load(m[1]), glmm_splat(x0, 1),
+                             glmm_fmadd(glmm_load(m[2]), glmm_splat(x0, 2),
+                                        _mm_mul_ps(glmm_load(m[3]),
+                                                   glmm_splat(x0, 3)))));
+  
+  glmm_store(dest, x1);
 }
 
 CGLM_INLINE
@@ -115,20 +105,18 @@ glm_mat4_det_sse2(mat4 mat) {
    t[3] = i * p - m * l;
    t[4] = i * o - m * k;
    */
-  x0 = _mm_sub_ps(_mm_mul_ps(glmm_shuff1(r2, 0, 0, 1, 1),
-                             glmm_shuff1(r3, 2, 3, 2, 3)),
-                  _mm_mul_ps(glmm_shuff1(r3, 0, 0, 1, 1),
-                             glmm_shuff1(r2, 2, 3, 2, 3)));
+  x0 = glmm_fnmadd(glmm_shuff1(r3, 0, 0, 1, 1), glmm_shuff1(r2, 2, 3, 2, 3),
+                   _mm_mul_ps(glmm_shuff1(r2, 0, 0, 1, 1),
+                              glmm_shuff1(r3, 2, 3, 2, 3)));
   /*
    t[0] = k * p - o * l;
    t[0] = k * p - o * l;
    t[5] = i * n - m * j;
    t[5] = i * n - m * j;
    */
-  x1 = _mm_sub_ps(_mm_mul_ps(glmm_shuff1(r2, 0, 0, 2, 2),
-                             glmm_shuff1(r3, 1, 1, 3, 3)),
-                  _mm_mul_ps(glmm_shuff1(r3, 0, 0, 2, 2),
-                             glmm_shuff1(r2, 1, 1, 3, 3)));
+  x1 = glmm_fnmadd(glmm_shuff1(r3, 0, 0, 2, 2), glmm_shuff1(r2, 1, 1, 3, 3),
+                   _mm_mul_ps(glmm_shuff1(r2, 0, 0, 2, 2),
+                              glmm_shuff1(r3, 1, 1, 3, 3)));
 
   /*
      a * (f * t[0] - g * t[1] + h * t[2])
@@ -136,21 +124,16 @@ glm_mat4_det_sse2(mat4 mat) {
    + c * (e * t[1] - f * t[3] + h * t[5])
    - d * (e * t[2] - f * t[4] + g * t[5])
    */
-  x2 = _mm_sub_ps(_mm_mul_ps(glmm_shuff1(r1, 0, 0, 0, 1),
-                             _mm_shuffle_ps(x1, x0, _MM_SHUFFLE(1, 0, 0, 0))),
-                  _mm_mul_ps(glmm_shuff1(r1, 1, 1, 2, 2),
-                             glmm_shuff1(x0, 3, 2, 2, 0)));
-
-  x2 = _mm_add_ps(x2,
-                  _mm_mul_ps(glmm_shuff1(r1, 2, 3, 3, 3),
-                             _mm_shuffle_ps(x0, x1, _MM_SHUFFLE(2, 2, 3, 1))));
+  x2 = glmm_fnmadd(glmm_shuff1(r1, 1, 1, 2, 2), glmm_shuff1(x0, 3, 2, 2, 0),
+                   _mm_mul_ps(glmm_shuff1(r1, 0, 0, 0, 1),
+                              _mm_shuffle_ps(x1, x0, _MM_SHUFFLE(1, 0, 0, 0))));
+  x2 = glmm_fmadd(glmm_shuff1(r1, 2, 3, 3, 3),
+                  _mm_shuffle_ps(x0, x1, _MM_SHUFFLE(2, 2, 3, 1)),
+                  x2);
+  
   x2 = _mm_xor_ps(x2, _mm_set_ps(-0.f, 0.f, -0.f, 0.f));
-
-  x0 = _mm_mul_ps(r0, x2);
-  x0 = _mm_add_ps(x0, glmm_shuff1(x0, 0, 1, 2, 3));
-  x0 = _mm_add_ps(x0, glmm_shuff1(x0, 1, 3, 3, 1));
-
-  return _mm_cvtss_f32(x0);
+  
+  return glmm_hadd(_mm_mul_ps(x2, r0));
 }
 
 CGLM_INLINE
@@ -159,8 +142,11 @@ glm_mat4_inv_fast_sse2(mat4 mat, mat4 dest) {
   __m128 r0, r1, r2, r3,
          v0, v1, v2, v3,
          t0, t1, t2, t3, t4, t5,
-         x0, x1, x2, x3, x4, x5, x6, x7;
-
+         x0, x1, x2, x3, x4, x5, x6, x7, x8, x9;
+  
+  x8 = _mm_set_ps(-0.f, 0.f, -0.f, 0.f);
+  x9 = glmm_shuff1(x8, 2, 1, 2, 1);
+  
   /* 127 <- 0 */
   r0 = glmm_load(mat[0]); /* d c b a */
   r1 = glmm_load(mat[1]); /* h g f e */
@@ -177,8 +163,8 @@ glm_mat4_inv_fast_sse2(mat4 mat, mat4 dest) {
      t1[0] = k * p - o * l;
      t2[0] = g * p - o * h;
      t3[0] = g * l - k * h; */
-  t0 = _mm_sub_ps(_mm_mul_ps(x3, x1), _mm_mul_ps(x2, x0));
-
+  t0 = glmm_fnmadd(x2, x0, _mm_mul_ps(x3, x1));
+  
   x4 = _mm_shuffle_ps(r2, r3, _MM_SHUFFLE(2, 1, 2, 1)); /* o n k j */
   x4 = glmm_shuff1(x4, 0, 2, 2, 2);                     /* j n n n */
   x5 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(1, 1, 1, 1)); /* f f j j */
@@ -187,14 +173,14 @@ glm_mat4_inv_fast_sse2(mat4 mat, mat4 dest) {
      t1[1] = j * p - n * l;
      t2[1] = f * p - n * h;
      t3[1] = f * l - j * h; */
-  t1 = _mm_sub_ps(_mm_mul_ps(x5, x1), _mm_mul_ps(x4, x0));
-
+   t1 = glmm_fnmadd(x4, x0, _mm_mul_ps(x5, x1));
+  
   /* t1[2] = j * o - n * k
      t1[2] = j * o - n * k;
      t2[2] = f * o - n * g;
      t3[2] = f * k - j * g; */
-  t2 = _mm_sub_ps(_mm_mul_ps(x5, x2), _mm_mul_ps(x4, x3));
-
+  t2 = glmm_fnmadd(x4, x3, _mm_mul_ps(x5, x2));
+  
   x6 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(0, 0, 0, 0)); /* e e i i */
   x7 = glmm_shuff2(r3, r2, 0, 0, 0, 0, 2, 0, 0, 0);     /* i m m m */
 
@@ -202,20 +188,20 @@ glm_mat4_inv_fast_sse2(mat4 mat, mat4 dest) {
      t1[3] = i * p - m * l;
      t2[3] = e * p - m * h;
      t3[3] = e * l - i * h; */
-  t3 = _mm_sub_ps(_mm_mul_ps(x6, x1), _mm_mul_ps(x7, x0));
-
+  t3 = glmm_fnmadd(x7, x0, _mm_mul_ps(x6, x1));
+  
   /* t1[4] = i * o - m * k;
      t1[4] = i * o - m * k;
      t2[4] = e * o - m * g;
      t3[4] = e * k - i * g; */
-  t4 = _mm_sub_ps(_mm_mul_ps(x6, x2), _mm_mul_ps(x7, x3));
-
+  t4 = glmm_fnmadd(x7, x3, _mm_mul_ps(x6, x2));
+  
   /* t1[5] = i * n - m * j;
      t1[5] = i * n - m * j;
      t2[5] = e * n - m * f;
      t3[5] = e * j - i * f; */
-  t5 = _mm_sub_ps(_mm_mul_ps(x6, x4), _mm_mul_ps(x7, x5));
-
+  t5 = glmm_fnmadd(x7, x5, _mm_mul_ps(x6, x4));
+  
   x0 = glmm_shuff2(r1, r0, 0, 0, 0, 0, 2, 2, 2, 0); /* a a a e */
   x1 = glmm_shuff2(r1, r0, 1, 1, 1, 1, 2, 2, 2, 0); /* b b b f */
   x2 = glmm_shuff2(r1, r0, 2, 2, 2, 2, 2, 2, 2, 0); /* c c c g */
@@ -226,50 +212,35 @@ glm_mat4_inv_fast_sse2(mat4 mat, mat4 dest) {
    dest[0][1] =-(b * t1[0] - c * t1[1] + d * t1[2]);
    dest[0][2] =  b * t2[0] - c * t2[1] + d * t2[2];
    dest[0][3] =-(b * t3[0] - c * t3[1] + d * t3[2]); */
-  v0 = _mm_add_ps(_mm_mul_ps(x3, t2),
-                  _mm_sub_ps(_mm_mul_ps(x1, t0),
-                             _mm_mul_ps(x2, t1)));
-  v0 = _mm_xor_ps(v0, _mm_set_ps(-0.f, 0.f, -0.f, 0.f));
+  v0 = _mm_xor_ps(glmm_fmadd(x3, t2, glmm_fnmadd(x2, t1, _mm_mul_ps(x1, t0))), x8);
+  
+  /*
+   dest[2][0] =  e * t1[1] - f * t1[3] + h * t1[5];
+   dest[2][1] =-(a * t1[1] - b * t1[3] + d * t1[5]);
+   dest[2][2] =  a * t2[1] - b * t2[3] + d * t2[5];
+   dest[2][3] =-(a * t3[1] - b * t3[3] + d * t3[5]);*/
+  v2 = _mm_xor_ps(glmm_fmadd(x3, t5, glmm_fnmadd(x1, t3, _mm_mul_ps(x0, t1))), x8);
 
   /*
    dest[1][0] =-(e * t1[0] - g * t1[3] + h * t1[4]);
    dest[1][1] =  a * t1[0] - c * t1[3] + d * t1[4];
    dest[1][2] =-(a * t2[0] - c * t2[3] + d * t2[4]);
    dest[1][3] =  a * t3[0] - c * t3[3] + d * t3[4]; */
-  v1 = _mm_add_ps(_mm_mul_ps(x3, t4),
-                  _mm_sub_ps(_mm_mul_ps(x0, t0),
-                             _mm_mul_ps(x2, t3)));
-  v1 = _mm_xor_ps(v1, _mm_set_ps(0.f, -0.f, 0.f, -0.f));
-
-  /*
-   dest[2][0] =  e * t1[1] - f * t1[3] + h * t1[5];
-   dest[2][1] =-(a * t1[1] - b * t1[3] + d * t1[5]);
-   dest[2][2] =  a * t2[1] - b * t2[3] + d * t2[5];
-   dest[2][3] =-(a * t3[1] - b * t3[3] + d * t3[5]);*/
-  v2 = _mm_add_ps(_mm_mul_ps(x3, t5),
-                  _mm_sub_ps(_mm_mul_ps(x0, t1),
-                             _mm_mul_ps(x1, t3)));
-  v2 = _mm_xor_ps(v2, _mm_set_ps(-0.f, 0.f, -0.f, 0.f));
+  v1 = _mm_xor_ps(glmm_fmadd(x3, t4, glmm_fnmadd(x2, t3, _mm_mul_ps(x0, t0))), x9);
 
   /*
    dest[3][0] =-(e * t1[2] - f * t1[4] + g * t1[5]);
    dest[3][1] =  a * t1[2] - b * t1[4] + c * t1[5];
    dest[3][2] =-(a * t2[2] - b * t2[4] + c * t2[5]);
    dest[3][3] =  a * t3[2] - b * t3[4] + c * t3[5]; */
-  v3 = _mm_add_ps(_mm_mul_ps(x2, t5),
-                  _mm_sub_ps(_mm_mul_ps(x0, t2),
-                             _mm_mul_ps(x1, t4)));
-  v3 = _mm_xor_ps(v3, _mm_set_ps(0.f, -0.f, 0.f, -0.f));
+  v3 = _mm_xor_ps(glmm_fmadd(x2, t5, glmm_fnmadd(x1, t4, _mm_mul_ps(x0, t2))), x9);
 
   /* determinant */
   x0 = _mm_shuffle_ps(v0, v1, _MM_SHUFFLE(0, 0, 0, 0));
   x1 = _mm_shuffle_ps(v2, v3, _MM_SHUFFLE(0, 0, 0, 0));
   x0 = _mm_shuffle_ps(x0, x1, _MM_SHUFFLE(2, 0, 2, 0));
 
-  x0 = _mm_mul_ps(x0, r0);
-  x0 = _mm_add_ps(x0, glmm_shuff1(x0, 0, 1, 2, 3));
-  x0 = _mm_add_ps(x0, glmm_shuff1(x0, 1, 0, 0, 1));
-  x0 = _mm_rcp_ps(x0);
+  x0 = _mm_rcp_ps(glmm_vhadd(_mm_mul_ps(x0, r0)));
 
   glmm_store(dest[0], _mm_mul_ps(v0, x0));
   glmm_store(dest[1], _mm_mul_ps(v1, x0));
@@ -283,8 +254,11 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
   __m128 r0, r1, r2, r3,
          v0, v1, v2, v3,
          t0, t1, t2, t3, t4, t5,
-         x0, x1, x2, x3, x4, x5, x6, x7;
-
+         x0, x1, x2, x3, x4, x5, x6, x7, x8, x9;
+  
+  x8 = _mm_set_ps(-0.f, 0.f, -0.f, 0.f);
+  x9 = glmm_shuff1(x8, 2, 1, 2, 1);
+  
   /* 127 <- 0 */
   r0 = glmm_load(mat[0]); /* d c b a */
   r1 = glmm_load(mat[1]); /* h g f e */
@@ -301,8 +275,8 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
      t1[0] = k * p - o * l;
      t2[0] = g * p - o * h;
      t3[0] = g * l - k * h; */
-  t0 = _mm_sub_ps(_mm_mul_ps(x3, x1), _mm_mul_ps(x2, x0));
-
+  t0 = glmm_fnmadd(x2, x0, _mm_mul_ps(x3, x1));
+  
   x4 = _mm_shuffle_ps(r2, r3, _MM_SHUFFLE(2, 1, 2, 1)); /* o n k j */
   x4 = glmm_shuff1(x4, 0, 2, 2, 2);                     /* j n n n */
   x5 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(1, 1, 1, 1)); /* f f j j */
@@ -311,14 +285,14 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
      t1[1] = j * p - n * l;
      t2[1] = f * p - n * h;
      t3[1] = f * l - j * h; */
-  t1 = _mm_sub_ps(_mm_mul_ps(x5, x1), _mm_mul_ps(x4, x0));
-
+   t1 = glmm_fnmadd(x4, x0, _mm_mul_ps(x5, x1));
+  
   /* t1[2] = j * o - n * k
      t1[2] = j * o - n * k;
      t2[2] = f * o - n * g;
      t3[2] = f * k - j * g; */
-  t2 = _mm_sub_ps(_mm_mul_ps(x5, x2), _mm_mul_ps(x4, x3));
-
+  t2 = glmm_fnmadd(x4, x3, _mm_mul_ps(x5, x2));
+  
   x6 = _mm_shuffle_ps(r2, r1, _MM_SHUFFLE(0, 0, 0, 0)); /* e e i i */
   x7 = glmm_shuff2(r3, r2, 0, 0, 0, 0, 2, 0, 0, 0);     /* i m m m */
 
@@ -326,20 +300,20 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
      t1[3] = i * p - m * l;
      t2[3] = e * p - m * h;
      t3[3] = e * l - i * h; */
-  t3 = _mm_sub_ps(_mm_mul_ps(x6, x1), _mm_mul_ps(x7, x0));
-
+  t3 = glmm_fnmadd(x7, x0, _mm_mul_ps(x6, x1));
+  
   /* t1[4] = i * o - m * k;
      t1[4] = i * o - m * k;
      t2[4] = e * o - m * g;
      t3[4] = e * k - i * g; */
-  t4 = _mm_sub_ps(_mm_mul_ps(x6, x2), _mm_mul_ps(x7, x3));
-
+  t4 = glmm_fnmadd(x7, x3, _mm_mul_ps(x6, x2));
+  
   /* t1[5] = i * n - m * j;
      t1[5] = i * n - m * j;
      t2[5] = e * n - m * f;
      t3[5] = e * j - i * f; */
-  t5 = _mm_sub_ps(_mm_mul_ps(x6, x4), _mm_mul_ps(x7, x5));
-
+  t5 = glmm_fnmadd(x7, x5, _mm_mul_ps(x6, x4));
+  
   x0 = glmm_shuff2(r1, r0, 0, 0, 0, 0, 2, 2, 2, 0); /* a a a e */
   x1 = glmm_shuff2(r1, r0, 1, 1, 1, 1, 2, 2, 2, 0); /* b b b f */
   x2 = glmm_shuff2(r1, r0, 2, 2, 2, 2, 2, 2, 2, 0); /* c c c g */
@@ -350,50 +324,35 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
    dest[0][1] =-(b * t1[0] - c * t1[1] + d * t1[2]);
    dest[0][2] =  b * t2[0] - c * t2[1] + d * t2[2];
    dest[0][3] =-(b * t3[0] - c * t3[1] + d * t3[2]); */
-  v0 = _mm_add_ps(_mm_mul_ps(x3, t2),
-                  _mm_sub_ps(_mm_mul_ps(x1, t0),
-                             _mm_mul_ps(x2, t1)));
-  v0 = _mm_xor_ps(v0, _mm_set_ps(-0.f, 0.f, -0.f, 0.f));
+  v0 = _mm_xor_ps(glmm_fmadd(x3, t2, glmm_fnmadd(x2, t1, _mm_mul_ps(x1, t0))), x8);
+  
+  /*
+   dest[2][0] =  e * t1[1] - f * t1[3] + h * t1[5];
+   dest[2][1] =-(a * t1[1] - b * t1[3] + d * t1[5]);
+   dest[2][2] =  a * t2[1] - b * t2[3] + d * t2[5];
+   dest[2][3] =-(a * t3[1] - b * t3[3] + d * t3[5]);*/
+  v2 = _mm_xor_ps(glmm_fmadd(x3, t5, glmm_fnmadd(x1, t3, _mm_mul_ps(x0, t1))), x8);
 
   /*
    dest[1][0] =-(e * t1[0] - g * t1[3] + h * t1[4]);
    dest[1][1] =  a * t1[0] - c * t1[3] + d * t1[4];
    dest[1][2] =-(a * t2[0] - c * t2[3] + d * t2[4]);
    dest[1][3] =  a * t3[0] - c * t3[3] + d * t3[4]; */
-  v1 = _mm_add_ps(_mm_mul_ps(x3, t4),
-                  _mm_sub_ps(_mm_mul_ps(x0, t0),
-                             _mm_mul_ps(x2, t3)));
-  v1 = _mm_xor_ps(v1, _mm_set_ps(0.f, -0.f, 0.f, -0.f));
-
-  /*
-   dest[2][0] =  e * t1[1] - f * t1[3] + h * t1[5];
-   dest[2][1] =-(a * t1[1] - b * t1[3] + d * t1[5]);
-   dest[2][2] =  a * t2[1] - b * t2[3] + d * t2[5];
-   dest[2][3] =-(a * t3[1] - b * t3[3] + d * t3[5]);*/
-  v2 = _mm_add_ps(_mm_mul_ps(x3, t5),
-                  _mm_sub_ps(_mm_mul_ps(x0, t1),
-                             _mm_mul_ps(x1, t3)));
-  v2 = _mm_xor_ps(v2, _mm_set_ps(-0.f, 0.f, -0.f, 0.f));
+  v1 = _mm_xor_ps(glmm_fmadd(x3, t4, glmm_fnmadd(x2, t3, _mm_mul_ps(x0, t0))), x9);
 
   /*
    dest[3][0] =-(e * t1[2] - f * t1[4] + g * t1[5]);
    dest[3][1] =  a * t1[2] - b * t1[4] + c * t1[5];
    dest[3][2] =-(a * t2[2] - b * t2[4] + c * t2[5]);
    dest[3][3] =  a * t3[2] - b * t3[4] + c * t3[5]; */
-  v3 = _mm_add_ps(_mm_mul_ps(x2, t5),
-                  _mm_sub_ps(_mm_mul_ps(x0, t2),
-                             _mm_mul_ps(x1, t4)));
-  v3 = _mm_xor_ps(v3, _mm_set_ps(0.f, -0.f, 0.f, -0.f));
+  v3 = _mm_xor_ps(glmm_fmadd(x2, t5, glmm_fnmadd(x1, t4, _mm_mul_ps(x0, t2))), x9);
 
   /* determinant */
   x0 = _mm_shuffle_ps(v0, v1, _MM_SHUFFLE(0, 0, 0, 0));
   x1 = _mm_shuffle_ps(v2, v3, _MM_SHUFFLE(0, 0, 0, 0));
   x0 = _mm_shuffle_ps(x0, x1, _MM_SHUFFLE(2, 0, 2, 0));
 
-  x0 = _mm_mul_ps(x0, r0);
-  x0 = _mm_add_ps(x0, glmm_shuff1(x0, 0, 1, 2, 3));
-  x0 = _mm_add_ps(x0, glmm_shuff1(x0, 1, 0, 0, 1));
-  x0 = _mm_div_ps(_mm_set1_ps(1.0f), x0);
+  x0 = _mm_div_ps(_mm_set1_ps(1.0f), glmm_vhadd(_mm_mul_ps(x0, r0)));
 
   glmm_store(dest[0], _mm_mul_ps(v0, x0));
   glmm_store(dest[1], _mm_mul_ps(v1, x0));

include/cglm/simd/sse2/quat.h

@@ -27,15 +27,15 @@ glm_quat_mul_sse2(versor p, versor q, versor dest) {
   xp = glmm_load(p); /* 3 2 1 0 */
   xq = glmm_load(q);
 
-  r  = _mm_mul_ps(glmm_shuff1x(xp, 3), xq);
+  r  = _mm_mul_ps(glmm_splat(xp, 3), xq);
 
-  x0 = _mm_xor_ps(glmm_shuff1x(xp, 0), _mm_set_ps(-0.f, 0.f, -0.f, 0.f));
+  x0 = _mm_xor_ps(glmm_splat(xp, 0), _mm_set_ps(-0.f, 0.f, -0.f, 0.f));
   r  = _mm_add_ps(r, _mm_mul_ps(x0, glmm_shuff1(xq, 0, 1, 2, 3)));
 
-  x0 = _mm_xor_ps(glmm_shuff1x(xp, 1), _mm_set_ps(-0.f, -0.f, 0.f, 0.f));
+  x0 = _mm_xor_ps(glmm_splat(xp, 1), _mm_set_ps(-0.f, -0.f, 0.f, 0.f));
   r  = _mm_add_ps(r, _mm_mul_ps(x0, glmm_shuff1(xq, 1, 0, 3, 2)));
 
-  x0 = _mm_xor_ps(glmm_shuff1x(xp, 2), _mm_set_ps(-0.f, 0.f, 0.f, -0.f));
+  x0 = _mm_xor_ps(glmm_splat(xp, 2), _mm_set_ps(-0.f, 0.f, 0.f, -0.f));
   r  = _mm_add_ps(r, _mm_mul_ps(x0, glmm_shuff1(xq, 2, 3, 0, 1)));
 
   glmm_store(dest, r);

include/cglm/simd/x86.h

@@ -18,6 +18,9 @@
 #  define glmm_store(p, a)  _mm_store_ps(p, a)
 #endif
 
+#define glmm_set1(x) _mm_set1_ps(x)
+#define glmm_128     __m128
+
 #ifdef CGLM_USE_INT_DOMAIN
 #  define glmm_shuff1(xmm, z, y, x, w)                                        \
      _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(xmm),                \
@@ -27,7 +30,16 @@
        _mm_shuffle_ps(xmm, xmm, _MM_SHUFFLE(z, y, x, w))
 #endif
 
+#define glmm_splat(x, lane) glmm_shuff1(x, lane, lane, lane, lane)
+
+#define glmm_splat_x(x) glmm_splat(x, 0)
+#define glmm_splat_y(x) glmm_splat(x, 1)
+#define glmm_splat_z(x) glmm_splat(x, 2)
+#define glmm_splat_w(x) glmm_splat(x, 3)
+
+/* glmm_shuff1x() is DEPRECATED!, use glmm_splat() */
 #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)
@@ -48,6 +60,15 @@ glmm_abs(__m128 x) {
   return _mm_andnot_ps(_mm_set1_ps(-0.0f), x);
 }
 
+static inline
+__m128
+glmm_vhadd(__m128 v) {
+  __m128 x0;
+  x0 = _mm_add_ps(v,  glmm_shuff1(v, 0, 1, 2, 3));
+  x0 = _mm_add_ps(x0, glmm_shuff1(x0, 1, 0, 0, 1));
+  return x0;
+}
+
 static inline
 __m128
 glmm_vhadds(__m128 v) {
@@ -80,7 +101,7 @@ glmm_vhmin(__m128 v) {
   __m128 x0, x1, x2;
   x0 = _mm_movehl_ps(v, v);     /* [2, 3, 2, 3] */
   x1 = _mm_min_ps(x0, v);       /* [0|2, 1|3, 2|2, 3|3] */
-  x2 = glmm_shuff1x(x1, 1);     /* [1|3, 1|3, 1|3, 1|3] */
+  x2 = glmm_splat(x1, 1);       /* [1|3, 1|3, 1|3, 1|3] */
   return _mm_min_ss(x1, x2);
 }
 
@@ -96,7 +117,7 @@ glmm_vhmax(__m128 v) {
   __m128 x0, x1, x2;
   x0 = _mm_movehl_ps(v, v);     /* [2, 3, 2, 3] */
   x1 = _mm_max_ps(x0, v);       /* [0|2, 1|3, 2|2, 3|3] */
-  x2 = glmm_shuff1x(x1, 1);     /* [1|3, 1|3, 1|3, 1|3] */
+  x2 = glmm_splat(x1, 1);       /* [1|3, 1|3, 1|3, 1|3] */
   return _mm_max_ss(x1, x2);
 }
 
@@ -188,5 +209,99 @@ glmm_store3(float v[3], __m128 vx) {
   _mm_store_ss(&v[2], glmm_shuff1(vx, 2, 2, 2, 2));
 }
 
+static inline
+__m128
+glmm_div(__m128 a, __m128 b) {
+  return _mm_div_ps(a, b);
+}
+
+/* enable FMA macro for MSVC? */
+#if defined(_MSC_VER) && !defined(__FMA__) && defined(__AVX2__)
+#  define __FMA__ 1
+#endif
+
+static inline
+__m128
+glmm_fmadd(__m128 a, __m128 b, __m128 c) {
+#ifdef __FMA__
+  return _mm_fmadd_ps(a, b, c);
+#else
+  return _mm_add_ps(c, _mm_mul_ps(a, b));
+#endif
+}
+
+static inline
+__m128
+glmm_fnmadd(__m128 a, __m128 b, __m128 c) {
+#ifdef __FMA__
+  return _mm_fnmadd_ps(a, b, c);
+#else
+  return _mm_sub_ps(c, _mm_mul_ps(a, b));
+#endif
+}
+
+static inline
+__m128
+glmm_fmsub(__m128 a, __m128 b, __m128 c) {
+#ifdef __FMA__
+  return _mm_fmsub_ps(a, b, c);
+#else
+  return _mm_sub_ps(_mm_mul_ps(a, b), c);
+#endif
+}
+
+static inline
+__m128
+glmm_fnmsub(__m128 a, __m128 b, __m128 c) {
+#ifdef __FMA__
+  return _mm_fnmsub_ps(a, b, c);
+#else
+  return _mm_xor_ps(_mm_add_ps(_mm_mul_ps(a, b), c), _mm_set1_ps(-0.0f));
+#endif
+}
+
+#if defined(__AVX__)
+static inline
+__m256
+glmm256_fmadd(__m256 a, __m256 b, __m256 c) {
+#ifdef __FMA__
+  return _mm256_fmadd_ps(a, b, c);
+#else
+  return _mm256_add_ps(c, _mm256_mul_ps(a, b));
+#endif
+}
+
+static inline
+__m256
+glmm256_fnmadd(__m256 a, __m256 b, __m256 c) {
+#ifdef __FMA__
+  return _mm256_fnmadd_ps(a, b, c);
+#else
+  return _mm256_sub_ps(c, _mm256_mul_ps(a, b));
+#endif
+}
+
+static inline
+__m256
+glmm256_fmsub(__m256 a, __m256 b, __m256 c) {
+#ifdef __FMA__
+  return _mm256_fmsub_ps(a, b, c);
+#else
+  return _mm256_sub_ps(_mm256_mul_ps(a, b), c);
+#endif
+}
+
+static inline
+__m256
+glmm256_fnmsub(__m256 a, __m256 b, __m256 c) {
+#ifdef __FMA__
+  return _mm256_fmsub_ps(a, b, c);
+#else
+  return _mm256_xor_ps(_mm256_sub_ps(_mm256_mul_ps(a, b), c),
+                       _mm256_set1_ps(-0.0f));
+#endif
+}
+#endif
+
 #endif
 #endif /* cglm_simd_x86_h */

include/cglm/vec4-ext.h

@@ -224,10 +224,10 @@ glm_vec4_sign(vec4 v, vec4 dest) {
 
   x0 = glmm_load(v);
   x1 = _mm_set_ps(0.0f, 0.0f, 1.0f, -1.0f);
-  x2 = glmm_shuff1x(x1, 2);
+  x2 = glmm_splat(x1, 2);
 
-  x3 = _mm_and_ps(_mm_cmpgt_ps(x0, x2), glmm_shuff1x(x1, 1));
-  x4 = _mm_and_ps(_mm_cmplt_ps(x0, x2), glmm_shuff1x(x1, 0));
+  x3 = _mm_and_ps(_mm_cmpgt_ps(x0, x2), glmm_splat(x1, 1));
+  x4 = _mm_and_ps(_mm_cmplt_ps(x0, x2), glmm_splat(x1, 0));
 
   glmm_store(dest, _mm_or_ps(x3, x4));
 #else

include/cglm/vec4.h

@@ -473,8 +473,8 @@ glm_vec4_scale_as(vec4 v, float s, vec4 dest) {
 CGLM_INLINE
 void
 glm_vec4_div(vec4 a, vec4 b, vec4 dest) {
-#if defined( __SSE__ ) || defined( __SSE2__ )
-  glmm_store(dest, _mm_div_ps(glmm_load(a), glmm_load(b)));
+#if defined(CGLM_SIMD)
+  glmm_store(dest, glmm_div(glmm_load(a), glmm_load(b)));
 #else
   dest[0] = a[0] / b[0];
   dest[1] = a[1] / b[1];
@@ -568,14 +568,8 @@ glm_vec4_subadd(vec4 a, vec4 b, vec4 dest) {
 CGLM_INLINE
 void
 glm_vec4_muladd(vec4 a, vec4 b, vec4 dest) {
-#if defined( __SSE__ ) || defined( __SSE2__ )
-  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))));
+#if defined(CGLM_SIMD)
+  glmm_store(dest, glmm_fmadd(glmm_load(a), glmm_load(b), glmm_load(dest)));
 #else
   dest[0] += a[0] * b[0];
   dest[1] += a[1] * b[1];
@@ -596,14 +590,8 @@ glm_vec4_muladd(vec4 a, vec4 b, vec4 dest) {
 CGLM_INLINE
 void
 glm_vec4_muladds(vec4 a, float s, vec4 dest) {
-#if defined( __SSE__ ) || defined( __SSE2__ )
-  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),
-                            vmulq_f32(vld1q_f32(a),
-                                      vdupq_n_f32(s))));
+#if defined(CGLM_SIMD)
+  glmm_store(dest, glmm_fmadd(glmm_load(a), glmm_set1(s), glmm_load(dest)));
 #else
   dest[0] += a[0] * s;
   dest[1] += a[1] * s;

include/cglm/version.h

@@ -10,6 +10,6 @@
 
 #define CGLM_VERSION_MAJOR 0
 #define CGLM_VERSION_MINOR 8
-#define CGLM_VERSION_PATCH 0
+#define CGLM_VERSION_PATCH 2
 
 #endif /* cglm_version_h */

meson.build

@@ -1,5 +1,5 @@
 project('cglm', 'c',
-    version : '0.8.0', 
+    version : '0.8.2', 
     license : 'mit',
     default_options : [
         'c_std=c11',