Hardware performance for real-time reconstruction

Comparative performances of reconstruction times with different graphic cards.

Reconstruction times

One of the largest process in Cyber is dedicated to reconstruction. This task uses massively graphic card features and is done in following steps:

[A] transformation from silhouettes to texture
[B] drawing quads along voxel space with texture projection activated
[C] readback of framebuffer
[D] processing frame values to usable data structure

Percentage of time usage on each reconstruction step:

sgi IR3 (1RM)

sgi IR3 (2RM)

sgi Tezro V12

GeForce4800tiSE

ReadTextures [A]

<1ms

DrawQuads [B]

1	2	3	4
1ms	2ms	3.2ms	5ms

1	2	3	4
<1ms	1ms	2ms	4ms

1	2	3	4
11ms	23ms	35ms	46ms

1	2	3	4
0.2ms	0.3ms	0.5ms	0.6ms

ReadPixels [C]

5ms (RED)

<3ms (RED)

6ms (RGBA)

3 ms (RGB332^*)

CreateVoxels [D] (CPU)

4ms

3ms

1ms

Total time for reconstruction
with 4 cameras [A,B,C,D]

14ms

10-11ms

55ms

< 5ms

(configs: IR3=R12,G12,B12,A12,Depth31 - TezroV12=R4,G4,B4,A4,Depth24)
^* RGB332 mode allow reading 3 voxels (drawn in R,G and B channels) in a single byte. Without this optimisation, this step would take twice time.

The time for renderings quads on V12 platform is very important due to 4bits per composant framebuffer(!!!!!!, it’s the “best” mode available with a double buffer, depth and alpha..., while the IR3 has 12bits per composant!).

Details on glReadPixels performances

Following table gives times (in milliseconds) for reading back a block of 512x512 pixels freom GPU to RAM (corresponding to a voxelic space of size 64x64x64):

BGR ou BGRA —>X—>X —>35 —>X —>X —>X —>

config	R, G, B ou A	LUMINANCE	RGB	RGBA
Onyx R12000+IR3 (1RM)	4.2	<script language="javascript"> 6.9	<script language="javascript"> 7.1	<script language="javascript"> 7.0
Onyx R12000+IR3 (2RM)	<script language="javascript"> 2.4	<script language="javascript"> 6.3	<script language="javascript"> 6.3	<script language="javascript"> 6.3
Tezro R16000+V12	<script language="javascript"> 27.0	<script language="javascript"> 28.0	<script language="javascript"> 31.0	<script language="javascript"> 7.2
? titre comparatif
PIV 1.6Ghz+Gf2MX	<script language="javascript"> 5.7	<script language="javascript"> 55.0	<script language="javascript"> 6.0	<script language="javascript"> 5.5
Bi-PIII 933Mhz+Radeon9700	<script language="javascript"> 6.9	<script language="javascript"> 134.0	<script language="javascript"> 22.2	<script language="javascript"> 23.0
Bi-Xeon2.4Ghz+GF4800tiSE	<script language="javascript"> 5.0	<script language="javascript"> 35.0	<script language="javascript"> 5.0	<script language="javascript"> 5.0