External/cglm
1
0
Fork 0
mirror of https://github.com/recp/cglm.git synced 2026-01-01 21:18:12 +00:00
Code Issues Packages Projects Releases Wiki Activity

simd: use glmm_set1() to optimize broadcasting single float

Browse Source
This commit is contained in:
Recep Aslantas
2024-04-10 22:52:53 +03:00
parent 62c0448e25
commit 68bdec4510
6 changed files with 24 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
include/cglm/mat2.h
View File

@@ -235,7 +235,7 @@ glm_mat2_scale(mat2 m, float s) {
glmm_store(m[0], wasm_f32x4_mul(wasm_v128_load(m[0]),
wasm_f32x4_splat(s)));
#elif defined( __SSE__ ) || defined( __SSE2__ )
glmm_store(m[0], _mm_mul_ps(_mm_loadu_ps(m[0]), _mm_set1_ps(s)));
glmm_store(m[0], _mm_mul_ps(_mm_loadu_ps(m[0]), glmm_set1(s)));
#elif defined(CGLM_NEON_FP)
vst1q_f32(m[0], vmulq_f32(vld1q_f32(m[0]), vdupq_n_f32(s)));
#else

4
include/cglm/simd/avx/mat4.h
View File

@@ -18,8 +18,8 @@ CGLM_INLINE
void
glm_mat4_scale_avx(mat4 m, float s) {
__m256 y0;
y0 = _mm256_set1_ps(s);
y0 = _mm256_broadcast_ss(&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])));
}

6
include/cglm/simd/sse2/mat4.h
View File

@@ -18,7 +18,7 @@ CGLM_INLINE
void
glm_mat4_scale_sse2(mat4 m, float s) {
__m128 x0;
x0 = _mm_set1_ps(s);
x0 = glmm_set1(s);
glmm_store(m[0], _mm_mul_ps(glmm_load(m[0]), x0));
glmm_store(m[1], _mm_mul_ps(glmm_load(m[1]), x0));
@@ -426,7 +426,7 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
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_div_ps(_mm_set1_ps(1.0f), glmm_vhadd(_mm_mul_ps(x0, r0)));
x0 = _mm_div_ps(glmm_set1(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));
@@ -494,7 +494,7 @@ glm_mat4_inv_sse2(mat4 mat, mat4 dest) {
/* v0: c3 * c10 + c4 * c9 + c1 * c8 + c2 * c7 */
/* v1: c5 * c12 + c6 * c11 */
v5 = _mm_set1_ps(1.0f);
v5 = glmm_set1(1.0f);
v0 = glmm_shuff1(t2, 2, 3, 0, 1);
v1 = glmm_shuff1(t1, 0, 1, 2, 3);
v0 = _mm_mul_ps(t0, v0);

11
include/cglm/simd/x86.h
View File

@@ -18,9 +18,16 @@
# define glmm_store(p, a) _mm_store_ps(p, a)
#endif
#define glmm_set1(x) _mm_set1_ps(x)
#define glmm_128 __m128
#ifdef __AVX__
# define glmm_set1(x) _mm_broadcast_ss(&x)
# define glmm_set1_ptr(x) _mm_broadcast_ss(x)
#else
# define glmm_set1(x) _mm_set1_ps(x)
# define glmm_set1_ptr(x) _mm_set1_ps(*x)
#endif
#if defined(CGLM_USE_INT_DOMAIN) && defined(__SSE2__)
# define glmm_shuff1(xmm, z, y, x, w) \
_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(xmm), \
@@ -86,7 +93,7 @@
#if defined(__SSE2__)
# define glmm_float32x4_SIGNMASK_NEG _mm_castsi128_ps(_mm_set1_epi32(GLMM_NEGZEROf)) /* _mm_set1_ps(-0.0f) */
#else
# define glmm_float32x4_SIGNMASK_NEG _mm_set1_ps(GLMM_NEGZEROf)
# define glmm_float32x4_SIGNMASK_NEG glmm_set1(GLMM_NEGZEROf)
#endif
#define glmm_float32x8_SIGNMASK_NEG _mm256_castsi256_ps(_mm256_set1_epi32(GLMM_NEGZEROf))

4
include/cglm/vec4-ext.h
View File

@@ -48,7 +48,7 @@ glm_vec4_broadcast(float val, vec4 d) {
#if defined(__wasm__) && defined(__wasm_simd128__)
glmm_store(d, wasm_f32x4_splat(val));
#elif defined( __SSE__ ) || defined( __SSE2__ )
glmm_store(d, _mm_set1_ps(val));
glmm_store(d, glmm_set1(val));
#else
d[0] = d[1] = d[2] = d[3] = val;
#endif
@@ -66,7 +66,7 @@ glm_vec4_fill(vec4 v, float val) {
#if defined(__wasm__) && defined(__wasm_simd128__)
glmm_store(v, wasm_f32x4_splat(val));
#elif defined( __SSE__ ) || defined( __SSE2__ )
glmm_store(v, _mm_set1_ps(val));
glmm_store(v, glmm_set1(val));
#else
v[0] = v[1] = v[2] = v[3] = val;
#endif

14
include/cglm/vec4.h
View File

@@ -215,7 +215,7 @@ glm_vec4_one(vec4 v) {
#if defined(__wasm__) && defined(__wasm_simd128__)
glmm_store(v, wasm_f32x4_const_splat(1.0f));
#elif defined( __SSE__ ) || defined( __SSE2__ )
glmm_store(v, _mm_set1_ps(1.0f));
glmm_store(v, glmm_set1(1.0f));
#elif defined(CGLM_NEON_FP)
vst1q_f32(v, vdupq_n_f32(1.0f));
#else
@@ -367,7 +367,7 @@ glm_vec4_adds(vec4 v, float s, vec4 dest) {
#if defined(__wasm__) && defined(__wasm_simd128__)
glmm_store(dest, wasm_f32x4_add(glmm_load(v), wasm_f32x4_splat(s)));
#elif defined( __SSE__ ) || defined( __SSE2__ )
glmm_store(dest, _mm_add_ps(glmm_load(v), _mm_set1_ps(s)));
glmm_store(dest, _mm_add_ps(glmm_load(v), glmm_set1(s)));
#elif defined(CGLM_NEON_FP)
vst1q_f32(dest, vaddq_f32(vld1q_f32(v), vdupq_n_f32(s)));
#else
@@ -415,7 +415,7 @@ glm_vec4_subs(vec4 v, float s, vec4 dest) {
#if defined(__wasm__) && defined(__wasm_simd128__)
glmm_store(dest, wasm_f32x4_sub(glmm_load(v), wasm_f32x4_splat(s)));
#elif defined( __SSE__ ) || defined( __SSE2__ )
glmm_store(dest, _mm_sub_ps(glmm_load(v), _mm_set1_ps(s)));
glmm_store(dest, _mm_sub_ps(glmm_load(v), glmm_set1(s)));
#elif defined(CGLM_NEON_FP)
vst1q_f32(dest, vsubq_f32(vld1q_f32(v), vdupq_n_f32(s)));
#else
@@ -463,7 +463,7 @@ glm_vec4_scale(vec4 v, float s, vec4 dest) {
#if defined(__wasm__) && defined(__wasm_simd128__)
glmm_store(dest, wasm_f32x4_mul(glmm_load(v), wasm_f32x4_splat(s)));
#elif defined( __SSE__ ) || defined( __SSE2__ )
glmm_store(dest, _mm_mul_ps(glmm_load(v), _mm_set1_ps(s)));
glmm_store(dest, _mm_mul_ps(glmm_load(v), glmm_set1(s)));
#elif defined(CGLM_NEON_FP)
vst1q_f32(dest, vmulq_f32(vld1q_f32(v), vdupq_n_f32(s)));
#else
@@ -528,7 +528,7 @@ glm_vec4_divs(vec4 v, float s, vec4 dest) {
#if defined(__wasm__) && defined(__wasm_simd128__)
glmm_store(dest, wasm_f32x4_div(glmm_load(v), wasm_f32x4_splat(s)));
#elif defined( __SSE__ ) || defined( __SSE2__ )
glmm_store(dest, _mm_div_ps(glmm_load(v), _mm_set1_ps(s)));
glmm_store(dest, _mm_div_ps(glmm_load(v), glmm_set1(s)));
#else
glm_vec4_scale(v, 1.0f / s, dest);
#endif
@@ -1065,8 +1065,8 @@ glm_vec4_clamp(vec4 v, float minVal, float maxVal) {
glmm_store(v, glmm_min(glmm_max(glmm_load(v), wasm_f32x4_splat(minVal)),
wasm_f32x4_splat(maxVal)));
#elif defined( __SSE__ ) || defined( __SSE2__ )
glmm_store(v, glmm_min(glmm_max(glmm_load(v), _mm_set1_ps(minVal)),
_mm_set1_ps(maxVal)));
glmm_store(v, glmm_min(glmm_max(glmm_load(v), glmm_set1(minVal)),
glmm_set1(maxVal)));
#elif defined(CGLM_NEON_FP)
glmm_store(v, glmm_min(glmm_max(vld1q_f32(v), vdupq_n_f32(minVal)),
vdupq_n_f32(maxVal)));