SSE2 double precision transforms

[ulf.schroeter]

Proposal

 

Usage of SS2 for increased double precision 3D transform performance. This INTEL article has details and sample code.

[frustum]

In progress.

[ulf.schroeter]

That's quick :)

[ulf.schroeter]

Any update on this one ?

[frustum]

It was implemented in the previous SDK.

[ulf.schroeter]

really, hasn't been mentioned in to realease notes. Anyway good news !

[ulf.schroeter]

Found double precision SSE2 code, but just a question: no SSE2 acceleration for matrix * vector multiplication e.g.

 

dvec4 &mul(dvec4 &ret,const dmat4 &m,const dvec4 &v) {
   ret.x = m.m00 * v.x + m.m01 * v.y + m.m02 * v.z + m.m03 * v.w;
   ret.y = m.m10 * v.x + m.m11 * v.y + m.m12 * v.z + m.m13 * v.w;
   ret.z = m.m20 * v.x + m.m21 * v.y + m.m22 * v.z + m.m23 * v.w;
   ret.w = v.w;
   return ret;
}

 

Besides matrix * matrix multiplication this should be the most common operation where optimization would make sense (at least for float precision it is optimized)

[ulf.schroeter]

Think I found the answer. Mass double precision matrix * vector multiplications are optimized via

 

void SimdSSE2::mulMat3Vec3()
void SimdSSE2::mulMat4Vec4()

[frustum]

Memory layout of dmat4 is not effective for single vector multiplication. dmat4 is rearranged for multiple vector multiplications.

[ulf.schroeter]

Do you have any rough numbers how much SSE2 improved double precision transform performance ?

[frustum]

The CPU is "Intel® Core™ i5 CPU 650 @ 3.20GHz"

 

Set Generic simd processor

minMaxVec3f Time: 0.063 FPS: 15.9834 Mem: 2557.34 Mb/s

minMaxVec4f Time: 0.080 FPS: 12.4738 Mem: 1995.81 Mb/s

minMaxVec3d Time: 0.063 FPS: 15.9289 Mem: 5097.25 Mb/s

minMaxVec4d Time: 0.081 FPS: 12.2850 Mem: 3931.20 Mb/s

mulMat3Vec3f Time: 0.054 FPS: 18.6081 Mem: 2977.30 Mb/s

mulMat4Vec3f Time: 0.054 FPS: 18.4176 Mem: 2946.81 Mb/s

mulMat4Vec4f Time: 0.056 FPS: 17.7939 Mem: 2847.03 Mb/s

mulMat3Vec3d Time: 0.113 FPS: 8.8857 Mem: 2843.41 Mb/s

mulMat4Vec3d Time: 0.115 FPS: 8.6618 Mem: 2771.76 Mb/s

 

Set SSE simd processor

minMaxVec3f Time: 0.015 FPS: 68.6153 Mem: 10978.45 Mb/s

minMaxVec4f Time: 0.014 FPS: 69.0894 Mem: 11054.30 Mb/s

minMaxVec3d Time: 0.061 FPS: 16.3983 Mem: 5247.45 Mb/s

minMaxVec4d Time: 0.079 FPS: 12.6208 Mem: 4038.67 Mb/s

mulMat3Vec3f Time: 0.023 FPS: 43.8059 Mem: 7008.94 Mb/s

mulMat4Vec3f Time: 0.026 FPS: 38.7357 Mem: 6197.71 Mb/s

mulMat4Vec4f Time: 0.025 FPS: 39.6589 Mem: 6345.43 Mb/s

mulMat3Vec3d Time: 0.107 FPS: 9.3884 Mem: 3004.30 Mb/s

mulMat4Vec3d Time: 0.115 FPS: 8.6634 Mem: 2772.29 Mb/s

 

Set SSE2 simd processor

minMaxVec3f Time: 0.015 FPS: 68.6059 Mem: 10976.95 Mb/s

minMaxVec4f Time: 0.014 FPS: 69.1085 Mem: 11057.36 Mb/s

minMaxVec3d Time: 0.016 FPS: 63.1592 Mem: 20210.95 Mb/s

minMaxVec4d Time: 0.016 FPS: 63.4961 Mem: 20318.75 Mb/s

mulMat3Vec3f Time: 0.023 FPS: 43.8020 Mem: 7008.32 Mb/s

mulMat4Vec3f Time: 0.026 FPS: 38.7012 Mem: 6192.19 Mb/s

mulMat4Vec4f Time: 0.025 FPS: 39.6873 Mem: 6349.96 Mb/s

mulMat3Vec3d Time: 0.045 FPS: 22.2119 Mem: 7107.79 Mb/s

mulMat4Vec3d Time: 0.045 FPS: 22.2040 Mem: 7105.27 Mb/s

[ulf.schroeter]

Thanks frustum, quite nice improvement !

[alexei.garbuzenko]

As I see double operations are faster than float with SSE2. Is it reasonable to switch from floats to doubles on SSE2-enabled cores to have math improvements?

[frustum]

Nope, double operation isn't faster than the float ones. The memory bandwidth is up to two times higher but overall FPS is slower.

[alexei.garbuzenko]

Oh, my bad. I thought that 'Mem' column is proportional to 'FPS'

Thanks for clarifying!