Welcome!
Welcome dans ce blog infâme, un des endroits du web ou je poste les choses que je ne poste pas ailleurs (i.e: Geeks3D...) et où je teste les thèmes WP que je trouve sympa.

RSS Feed

  • English
  • French

This demo uses instancing techniques (simple instancing, pseudo-instancing and geometry instancing(or GI)) to render a ring made of 10,000 small spheres. The demo is delivered in 5 versions:

  • each sphere is made of 1,800 triangles (18 millions triangles for the whole ring)
  • each sphere is made of 800 triangles (8 millions triangles for the whole ring)
  • each sphere is made of 200 triangles (2 millions triangles for the whole ring)
  • each sphere is made of 72 triangles (720,000 triangles for the whole ring)
  • each sphere is made of 18 triangles (180,000 triangles for the whole ring)

I added in the last moment a bonus: a 20,000 instances version, each instance made of 5,000 triangles. We get the monstruous count of 100 millions triangles (file Demo_Instancing_100MTriangles_20kInstances.exe).





DOWNLOAD

Several instancing techniques are used and you can select them with F1 to F6 keys.

  • F1: simple instancing with camera frustum culling: there is one source for geometry (a mesh) and it’s rendered for each instance. The tranformation matrix calculation is done on the CPU as well as the camera frustum test. OpenGL rendering uses the glDrawElements() function.
  • F2: simple instancing without camera frustum culling: there is one source for geometry (a mesh) and it’s rendered for each instance. The tranformation matrix calculation is done on the CPU but there is no longer camera frustum test. OpenGL rendering uses the glDrawElements() function.
  • F3: slow pseudo-instancing: there is one source for geometry (a mesh) and it’s rendered for each instance. Now the tranformation matrix calculation is done on the GPU and per-instance data are passed via uniform variables. There is no camera frustum test. OpenGL rendering uses the glDrawElements() function.
  • F4: pseudo-instancing: there is one source for geometry (a mesh) and it’s rendered for each instance. The tranformation matrix calculation is done on the GPU and per-instance data are passed via persistent vertex attributes (like texture coordinates or color). This technique has been shown by NVIDIA in the following whitepaper: GLSL Pseudo-Instancing. There is no camera frustum test. OpenGL rendering uses the glDrawElements() function.
  • F5: geometry instancing: it’s the real hardware instancing. There is one source for geometry (a mesh) and rendering is done by batchs of 400 instances per draw-call. The whole rendering of the ring requires 25 draw-calls instead of 10,000. The tranformation matrix calculation is done on the GPU and per-batch data is passed via uniform arrays. There is no camera frustum test. OpenGL rendering uses the glDrawElementsInstancedEXT() function. Currently, only NVIDIA GeForce 8 (and higher) support this function.
  • F6: geometry instancing with persistant vertex attributes: it’s the hardware instancing coupled with the transmission of parameters is done via the persistent vertex attributes. But the number of persistent vertex attributes is very limited. The best I did is to render 4 instances per draw-call. But oddly, I got the best results with 2 instances per draw-call. In that case, the rendering of whole ring requires 5000 draw-calls. The tranformation matrix calculation is done on the GPU and per-batch data is passed via uniform arrays. There is no camera frustum test. OpenGL rendering uses the glDrawElementsInstancedEXT() function. Currently, only NVIDIA GeForce 8 (and higher) support this function.

Ok now, let’s see some results with a NVIDIA GeForce 8800 GTX and an ATI Radeon HD 3870. Both cards have been tested with an AMD 64 3800+.

18 millions triangles – 1800 tri/instance

NVIDIA GeForce 8800 GTX – Forceware 169.38 XP32

  • F1: 223MTris/sec – 13FPS
  • F2: 223MTris/sec – 13FPS
  • F3: 223MTris/sec – 13FPS
  • F4: 223MTris/sec – 13FPS
  • F5: 223MTris/sec – 13FPS
  • F6: 171MTris/sec – 10FPS

ATI Radeon HD 3870 – Catalyst 8.2 XP32

  • F1: 429MTris/sec – 25FPS
  • F2: 463MTris/sec – 27FPS
  • F3: 446MTris/sec – 26FPS
  • F4: 274MTris/sec – 16FPS
  • F5: mode not available
  • F6: mode not available

8 millions de triangles – 800 tri/instance

NVIDIA GeForce 8800 GTX – Forceware 169.38 XP32

  • F1: 190MTri/sec – 25FPS
  • F2: 190MTri/sec – 25FPS
  • F3: 205MTri/sec – 27FPS
  • F4: 213MTri/sec – 28FPS
  • F5: 205MTri/sec – 27FPS
  • F6: 152MTri/sec – 20FPS

ATI Radeon HD 3870 – Catalyst 8.2 XP32

  • F1: 251MTris/sec – 33FPS
  • F2: 236MTris/sec – 31FPS
  • F3: 297MTris/sec – 39FPS
  • F4: 251MTris/sec – 33FPS
  • F5: mode not available
  • F6: mode not available

2 millions de triangles – 200 tri/instance

NVIDIA GeForce 8800 GTX – Forceware 169.38 XP32

  • F1: 47MTri/sec – 25FPS
  • F2: 47MTri/sec – 25FPS
  • F3: 57MTri/sec – 30FPS
  • F4: 131MTri/sec – 69FPS
  • F5: 167MTri/sec – 88FPS
  • F6: 148MTri/sec – 78FPS

ATI Radeon HD 3870 – Catalyst 8.2 XP32

  • F1: 47MTris/sec – 25FPS
  • F2: 59MTris/sec – 31FPS
  • F3: 74MTris/sec – 39FPS
  • F4: 112MTris/sec – 59FPS
  • F5: mode not available
  • F6: mode not available

720,000 triangles – 72 tri/instance

NVIDIA GeForce 8800 GTX – Forceware 169.38 XP32

  • F1: 17MTri/sec – 25FPS
  • F2: 17MTri/sec – 25FPS
  • F3: 20MTri/sec – 30FPS
  • F4: 47MTri/sec – 69FPS
  • F5: 60MTri/sec – 88FPS
  • F6: 53MTri/sec – 78FPS

ATI Radeon HD 3870 – Catalyst 8.2 XP32

  • F1: 17MTris/sec – 25FPS
  • F2: 21MTris/sec – 31FPS
  • F3: 26MTris/sec – 39FPS
  • F4: 40MTris/sec – 59FPS
  • F5: mode not available
  • F6: mode not available

180000 triangles – 18 tri/instance

NVIDIA GeForce 8800 GTX – Forceware 169.38 XP32

  • F1: 4MTri/sec – 25FPS
  • F2: 4MTri/sec – 25FPS
  • F3: 5MTri/sec – 30FPS
  • F4: 11MTri/sec – 69FPS
  • F5: 15MTri/sec – 89FPS
  • F6: 13MTri/sec – 79FPS

ATI Radeon HD 3870 – Catalyst 8.2 XP32

  • F1: 4MTris/sec – 25FPS
  • F2: 5MTris/sec – 31FPS
  • F3: 6MTris/sec – 39FPS
  • F4: 10MTris/sec – 59FPS
  • F5: mode not available
  • F6: mode not available

Quick results analysis:

  • we now understand why NVIDIA has called the technique using persistent vertex attributes “Pseudo-Instancing” (key F4). OpenGL glDrawElements() function is extremly fastand persistent vertex attributes require less overhead than uniforms to be passed to vertex shader. Both coupled together give this performance boost.
  • benefit of real hardware geometry instancing is mostly visible with few triangles per instance.
  • when there are many triangles per instance (1,800), the hardware implementation of glDrawElements() seems to be more efficient (twice!) on RV670 GPU than on G80.

Conclusion

From the results, hardware geometry instancing isn’t the kill-feature I expected. I find that very weird since the différence between 10000 render-calls with glDrawElements and 25 render-calls with glDrawElementsInstancedEXT is not verx important. Seems the instancing management (gl_InstanceID variable in the vertex shader) is a GPU-cycle eater!What a pity ATI hasn’t implemented yet geometry instancing in the Catalyst drivers. I’d be very curious to test hardware GI with a RV670.

7 Responses to “OpenGL Geometry Instancing”

  • Aw says:

    I have a GTX 280 and the F5,F6 options throw up the dialog about needing Gforce 8+ hardware.
    just FYI..

  • JeGX says:

    yes, the way the capability is checked is based on old NVIDIA drivers. Start GPU Caps Viewer GI demo to see the difference ;)

  • Delta says:

    Thanks for the information! The tests you have performed give important insight about the relationship between the batch sizes and the instancing techniques.

  • Joh says:

    the download link of the DemoPack seems broken… could you fix/reupload it? thnx :-)

  • Alex says:

    Hi.

    You wrote:
    ”I find that very weird since the différence between 10000 render-calls with glDrawElements and 25 render-calls with glDrawElementsInstancedEXT is not verx important.”

    But have you also considered the CPU-Overhead. The impact might not be very big due to GPU work, but it should definitely on CPU side! Have you checked this too?
    I mean in terms of CPU/GPU boundness.

    Anyway, good work, this is quite interesting.

    Btw, I got similar results with texture buffer objects.

  • Alessandro says:

    Hi

    The download file isn’t available anymore.

Leave a Reply