AuroraMiddleware/DirectXMath
1
0
Fork 0
mirror of https://github.com/microsoft/DirectXMath synced 2024-09-19 14:49:54 +00:00
Code Issues Releases Wiki Activity

Using broadcast instead of load/shuffle is faster when it is available. (#47)

Browse Source 
Using broadcast instead of load/shuffle is faster when it is available.
This commit is contained in:
Nicholas Frechette 2017-04-18 18:45:53 -04:00 committed by Chuck Walbourn
parent 67cffde166
commit d1aa003720
1 changed files with 62 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

68
Inc/DirectXMathMatrix.inl
View File

@ -297,13 +297,20 @@ inline XMMATRIX XM_CALLCONV XMMatrixMultiply
return mResult;
#elif defined(_XM_SSE_INTRINSICS_)
XMMATRIX mResult;
// Splat the component X,Y,Z then W
#if defined(_XM_AVX_INTRINSICS_)
XMVECTOR vX = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[0]) + 0);
XMVECTOR vY = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[0]) + 1);
XMVECTOR vZ = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[0]) + 2);
XMVECTOR vW = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[0]) + 3);
#else
// Use vW to hold the original row
XMVECTOR vW = M1.r[0];
// Splat the component X,Y,Z then W
XMVECTOR vX = XM_PERMUTE_PS(vW,_MM_SHUFFLE(0,0,0,0));
XMVECTOR vY = XM_PERMUTE_PS(vW,_MM_SHUFFLE(1,1,1,1));
XMVECTOR vZ = XM_PERMUTE_PS(vW,_MM_SHUFFLE(2,2,2,2));
vW = XM_PERMUTE_PS(vW,_MM_SHUFFLE(3,3,3,3));
#endif
// Perform the operation on the first row
vX = _mm_mul_ps(vX,M2.r[0]);
vY = _mm_mul_ps(vY,M2.r[1]);
@ -315,11 +322,18 @@ inline XMMATRIX XM_CALLCONV XMMatrixMultiply
vX = _mm_add_ps(vX,vY);
mResult.r[0] = vX;
// Repeat for the other 3 rows
#if defined(_XM_AVX_INTRINSICS_)
vX = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[1]) + 0);
vY = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[1]) + 1);
vZ = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[1]) + 2);
vW = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[1]) + 3);
#else
vW = M1.r[1];
vX = XM_PERMUTE_PS(vW,_MM_SHUFFLE(0,0,0,0));
vY = XM_PERMUTE_PS(vW,_MM_SHUFFLE(1,1,1,1));
vZ = XM_PERMUTE_PS(vW,_MM_SHUFFLE(2,2,2,2));
vW = XM_PERMUTE_PS(vW,_MM_SHUFFLE(3,3,3,3));
#endif
vX = _mm_mul_ps(vX,M2.r[0]);
vY = _mm_mul_ps(vY,M2.r[1]);
vZ = _mm_mul_ps(vZ,M2.r[2]);
@ -328,11 +342,18 @@ inline XMMATRIX XM_CALLCONV XMMatrixMultiply
vY = _mm_add_ps(vY,vW);
vX = _mm_add_ps(vX,vY);
mResult.r[1] = vX;
#if defined(_XM_AVX_INTRINSICS_)
vX = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[2]) + 0);
vY = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[2]) + 1);
vZ = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[2]) + 2);
vW = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[2]) + 3);
#else
vW = M1.r[2];
vX = XM_PERMUTE_PS(vW,_MM_SHUFFLE(0,0,0,0));
vY = XM_PERMUTE_PS(vW,_MM_SHUFFLE(1,1,1,1));
vZ = XM_PERMUTE_PS(vW,_MM_SHUFFLE(2,2,2,2));
vW = XM_PERMUTE_PS(vW,_MM_SHUFFLE(3,3,3,3));
#endif
vX = _mm_mul_ps(vX,M2.r[0]);
vY = _mm_mul_ps(vY,M2.r[1]);
vZ = _mm_mul_ps(vZ,M2.r[2]);
@ -341,11 +362,18 @@ inline XMMATRIX XM_CALLCONV XMMatrixMultiply
vY = _mm_add_ps(vY,vW);
vX = _mm_add_ps(vX,vY);
mResult.r[2] = vX;
#if defined(_XM_AVX_INTRINSICS_)
vX = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[3]) + 0);
vY = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[3]) + 1);
vZ = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[3]) + 2);
vW = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[3]) + 3);
#else
vW = M1.r[3];
vX = XM_PERMUTE_PS(vW,_MM_SHUFFLE(0,0,0,0));
vY = XM_PERMUTE_PS(vW,_MM_SHUFFLE(1,1,1,1));
vZ = XM_PERMUTE_PS(vW,_MM_SHUFFLE(2,2,2,2));
vW = XM_PERMUTE_PS(vW,_MM_SHUFFLE(3,3,3,3));
#endif
vX = _mm_mul_ps(vX,M2.r[0]);
vY = _mm_mul_ps(vY,M2.r[1]);
vZ = _mm_mul_ps(vZ,M2.r[2]);
@ -450,13 +478,20 @@ inline XMMATRIX XM_CALLCONV XMMatrixMultiplyTranspose
mResult.r[3] = T1.val[1];
return mResult;
#elif defined(_XM_SSE_INTRINSICS_)
// Splat the component X,Y,Z then W
#if defined(_XM_AVX_INTRINSICS_)
XMVECTOR vX = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[0]) + 0);
XMVECTOR vY = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[0]) + 1);
XMVECTOR vZ = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[0]) + 2);
XMVECTOR vW = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[0]) + 3);
#else
// Use vW to hold the original row
XMVECTOR vW = M1.r[0];
// Splat the component X,Y,Z then W
XMVECTOR vX = XM_PERMUTE_PS(vW,_MM_SHUFFLE(0,0,0,0));
XMVECTOR vY = XM_PERMUTE_PS(vW,_MM_SHUFFLE(1,1,1,1));
XMVECTOR vZ = XM_PERMUTE_PS(vW,_MM_SHUFFLE(2,2,2,2));
vW = XM_PERMUTE_PS(vW,_MM_SHUFFLE(3,3,3,3));
#endif
// Perform the operation on the first row
vX = _mm_mul_ps(vX,M2.r[0]);
vY = _mm_mul_ps(vY,M2.r[1]);
@ -466,13 +501,20 @@ inline XMMATRIX XM_CALLCONV XMMatrixMultiplyTranspose
vX = _mm_add_ps(vX,vZ);
vY = _mm_add_ps(vY,vW);
vX = _mm_add_ps(vX,vY);
__m128 r0 = vX;
XMVECTOR r0 = vX;
// Repeat for the other 3 rows
#if defined(_XM_AVX_INTRINSICS_)
vX = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[1]) + 0);
vY = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[1]) + 1);
vZ = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[1]) + 2);
vW = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[1]) + 3);
#else
vW = M1.r[1];
vX = XM_PERMUTE_PS(vW,_MM_SHUFFLE(0,0,0,0));
vY = XM_PERMUTE_PS(vW,_MM_SHUFFLE(1,1,1,1));
vZ = XM_PERMUTE_PS(vW,_MM_SHUFFLE(2,2,2,2));
vW = XM_PERMUTE_PS(vW,_MM_SHUFFLE(3,3,3,3));
#endif
vX = _mm_mul_ps(vX,M2.r[0]);
vY = _mm_mul_ps(vY,M2.r[1]);
vZ = _mm_mul_ps(vZ,M2.r[2]);
@ -480,12 +522,19 @@ inline XMMATRIX XM_CALLCONV XMMatrixMultiplyTranspose
vX = _mm_add_ps(vX,vZ);
vY = _mm_add_ps(vY,vW);
vX = _mm_add_ps(vX,vY);
__m128 r1 = vX;
XMVECTOR r1 = vX;
#if defined(_XM_AVX_INTRINSICS_)
vX = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[2]) + 0);
vY = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[2]) + 1);
vZ = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[2]) + 2);
vW = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[2]) + 3);
#else
vW = M1.r[2];
vX = XM_PERMUTE_PS(vW,_MM_SHUFFLE(0,0,0,0));
vY = XM_PERMUTE_PS(vW,_MM_SHUFFLE(1,1,1,1));
vZ = XM_PERMUTE_PS(vW,_MM_SHUFFLE(2,2,2,2));
vW = XM_PERMUTE_PS(vW,_MM_SHUFFLE(3,3,3,3));
#endif
vX = _mm_mul_ps(vX,M2.r[0]);
vY = _mm_mul_ps(vY,M2.r[1]);
vZ = _mm_mul_ps(vZ,M2.r[2]);
@ -493,12 +542,19 @@ inline XMMATRIX XM_CALLCONV XMMatrixMultiplyTranspose
vX = _mm_add_ps(vX,vZ);
vY = _mm_add_ps(vY,vW);
vX = _mm_add_ps(vX,vY);
__m128 r2 = vX;
XMVECTOR r2 = vX;
#if defined(_XM_AVX_INTRINSICS_)
vX = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[3]) + 0);
vY = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[3]) + 1);
vZ = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[3]) + 2);
vW = _mm_broadcast_ss(reinterpret_cast<const float*>(&M1.r[3]) + 3);
#else
vW = M1.r[3];
vX = XM_PERMUTE_PS(vW,_MM_SHUFFLE(0,0,0,0));
vY = XM_PERMUTE_PS(vW,_MM_SHUFFLE(1,1,1,1));
vZ = XM_PERMUTE_PS(vW,_MM_SHUFFLE(2,2,2,2));
vW = XM_PERMUTE_PS(vW,_MM_SHUFFLE(3,3,3,3));
#endif
vX = _mm_mul_ps(vX,M2.r[0]);
vY = _mm_mul_ps(vY,M2.r[1]);
vZ = _mm_mul_ps(vZ,M2.r[2]);
@ -506,7 +562,7 @@ inline XMMATRIX XM_CALLCONV XMMatrixMultiplyTranspose
vX = _mm_add_ps(vX,vZ);
vY = _mm_add_ps(vY,vW);
vX = _mm_add_ps(vX,vY);
__m128 r3 = vX;
XMVECTOR r3 = vX;
// x.x,x.y,y.x,y.y
XMVECTOR vTemp1 = _mm_shuffle_ps(r0,r1,_MM_SHUFFLE(1,0,1,0));