mirrors/HandmadeMath
1
0
Fork 0
mirror of https://github.com/HandmadeMath/HandmadeMath.git synced 2025-12-28 15:44:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

Better Neon tranpose

Browse Source 
vld can de-interleave up to 4-float element structures in one go into multiple registers. This allows us to load each column of the matrix and put each element into a corresponding vector, essentially doing the entire transpose in 1 instruction

Fixed it
This commit is contained in:
Jack Punter
2023-12-11 23:53:04 +00:00
parent 5f20d693c9
commit 133a595b6f
1 changed files with 7 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
HandmadeMath.h
View File

@@ -1517,21 +1517,16 @@ static inline HMM_Mat4 HMM_TransposeM4(HMM_Mat4 Matrix)
{
ASSERT_COVERED(HMM_TransposeM4);
HMM_Mat4 Result = Matrix;
HMM_Mat4 Result;
#ifdef HANDMADE_MATH__USE_SSE
Result = Matrix;
_MM_TRANSPOSE4_PS(Result.Columns[0].SSE, Result.Columns[1].SSE, Result.Columns[2].SSE, Result.Columns[3].SSE);
#elif HANDMADE_MATH__USE_NEON
// Based on Fabian's article on SIMD Transposes, although he says that the interleave32 is
// a VUZP instruction, its actually a VZIP instruction
// https://fgiesen.wordpress.com/2013/07/09/simd-transposes-1/
float32x4x2_t B0B2 = vzipq_f32(Result.Columns[0].NEON, Result.Columns[2].NEON);
float32x4x2_t B1B3 = vzipq_f32(Result.Columns[1].NEON, Result.Columns[3].NEON);
float32x4x2_t XY = vzipq_f32(B0B2.val[0], B1B3.val[0]);
float32x4x2_t ZW = vzipq_f32(B0B2.val[1], B1B3.val[1]);
Result.Columns[0].NEON = XY.val[0];
Result.Columns[1].NEON = XY.val[1];
Result.Columns[2].NEON = ZW.val[0];
Result.Columns[3].NEON = ZW.val[1];
float32x4x4_t Transposed = vld4q_f32((float*)Matrix.Columns);
Result.Columns[0].NEON = Transposed.val[0];
Result.Columns[1].NEON = Transposed.val[1];
Result.Columns[2].NEON = Transposed.val[2];
Result.Columns[3].NEON = Transposed.val[3];
#else
Result.Elements[0][1] = Matrix.Elements[1][0];
Result.Elements[0][2] = Matrix.Elements[2][0];