guezoloic/t3dsr
1
0
Fork 0
mirror of https://github.com/guezoloic/t3dsr.git synced 2026-01-25 04:34:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix(mv4.c): fix few sse x64 functions

Browse Source
This commit is contained in:
Loïc GUEZO
2025-09-04 18:58:21 +02:00
parent 9ae586e0ba
commit 8443ad9372
1 changed files with 8 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
src/math/mv4.c
View File

@@ -15,17 +15,18 @@ Vec4f_t mat4f_mul_vec4f(const Mat4f_t mat, Vec4f_t v)
for (int i = 0; i < 4; i++) for (int i = 0; i < 4; i++)
{ {
__m128 row = _mm_load_ps(&mat.m[4*i]); __m128 row = _mm_load_ps(&mat.m[i * 4]);
__m128 mul = _mm_mul_ps(row, vec); __m128 mul = _mm_mul_ps(row, vec);
__m128 shuf = _mm_movehl_ps(mul, mul); __m128 shuf = _mm_shuffle_ps(mul, mul, _MM_SHUFFLE(2, 3, 0, 1)); // [y z y z]
__m128 sum = _mm_add_ps(mul, shuf); __m128 sum = _mm_add_ps(mul, shuf);
shuf = _mm_movehl_ps(shuf, sum); shuf = _mm_shuffle_ps(sum, sum, _MM_SHUFFLE(1, 0, 3, 2));
sum = _mm_add_ss(sum, shuf); sum = _mm_add_ss(sum, shuf);
out.data[i] = _mm_cvtss_f32(sum); out.data[i] = _mm_cvtss_f32(sum);
} }
#elif defined (SIMD_ARCH) #elif defined (SIMD_ARCH)
float32x4_t vec = vld1q_f32(v.data); float32x4_t vec = vld1q_f32(v.data);