sculpt mode performance regression in 2.83. #77641

New Issue

Jean Da Costa · 2020-06-08T19:32:39+02:00

Jean Da Costa commented

2020-06-08 19:32:39 +02:00

System Information
Operating system: Linux-5.3.0-53-generic-x86_64-with-debian-buster-sid 64 Bits
Graphics card: Radeon RX 570 Series (POLARIS10, DRM 3.36.0, 5.3.0-53-generic, LLVM 9.0.0) X.Org 4.5 (Core Profile) Mesa 19.3.4

Blender Version
Broken: version: 2.83.0, branch: master, commit date: 2020-06-03 14:38, hash: 211b6c29f7
Worked: version 2.82 before 56d6666689 (commit that introduced the issue)

Short description of error
Performance regression when navigating while sculpting dense meshes

Exact steps for others to reproduce the error
load a highpoly mesh 1.000.000 faces or more, enter sculpt mode and pan/orbit

**System Information** Operating system: Linux-5.3.0-53-generic-x86_64-with-debian-buster-sid 64 Bits Graphics card: Radeon RX 570 Series (POLARIS10, DRM 3.36.0, 5.3.0-53-generic, LLVM 9.0.0) X.Org 4.5 (Core Profile) Mesa 19.3.4 **Blender Version** Broken: version: 2.83.0, branch: master, commit date: 2020-06-03 14:38, hash: `211b6c29f7` Worked: version 2.82 before 56d6666689 (commit that introduced the issue) **Short description of error** Performance regression when navigating while sculpting dense meshes **Exact steps for others to reproduce the error** load a highpoly mesh 1.000.000 faces or more, enter sculpt mode and pan/orbit

Jean Da Costa commented

2020-06-08 19:32:39 +02:00

Added subscriber: @jeacom

Nelson commented

2020-06-08 19:50:35 +02:00

Added subscriber: @NelsonNAS

Nelson commented

2020-06-08 19:50:35 +02:00

Yes, after #74692 (Sculpt Mode significantly laggy after the addition of Face Sets?) things never became as smooth as before.
Maybe there's no solution to this.

Yes, after #74692 (Sculpt Mode significantly laggy after the addition of Face Sets?) things never became as smooth as before. Maybe there's no solution to this.

Ankit Meel commented

2020-06-08 19:56:12 +02:00

Added subscriber: @ankitm

Ankit Meel commented

2020-06-08 19:56:12 +02:00

Does the performance improve if you turn off Dyntopo ?
https://developer.blender.org/T76109

Does the performance improve if you turn off Dyntopo ? https://developer.blender.org/T76109

Jean Da Costa commented

2020-06-08 20:25:26 +02:00

In #77641#948988, @ankitm wrote:
Does the performance improve if you turn off Dyntopo ?
https://developer.blender.org/T76109

The performance is bad with or without dyntopo. but with dyntopo its worse, almost unbearable.

> In #77641#948988, @ankitm wrote: > Does the performance improve if you turn off Dyntopo ? > https://developer.blender.org/T76109 The performance is bad with or without dyntopo. but with dyntopo its worse, almost unbearable.

Jean Da Costa commented

2020-06-08 22:00:44 +02:00

@ankitm
I also noticed a weird behaviour.

It seems like as I orbit around the model, the performance slowly increases, it takes anywhere between 1-2 mins but it does get a usable performance eventually, in 2.82 this also happens but it takes seconds to stabilize instead of minutes.
it's really weird.

@ankitm I also noticed a weird behaviour. It seems like as I orbit around the model, the performance slowly increases, it takes anywhere between 1-2 mins but it does get a usable performance eventually, in 2.82 this also happens but it takes seconds to stabilize instead of minutes. it's really weird.

Jean Da Costa commented

2020-06-09 01:02:28 +02:00

Added subscriber: @fclem

Jean Da Costa commented

2020-06-09 01:02:28 +02:00

After bisecting between v2.82a and v2.83 I found this commit to be the first with the issue.

56d6666689

To be clear, even before this commit the performance wasnt good but after this it got much worse.

I dont see any relation besides using an amd GPU.
@fclem

After bisecting between v2.82a and v2.83 I found this commit to be the first with the issue. 56d6666689 To be clear, even before this commit the performance wasnt good but after this it got much worse. I dont see any relation besides using an amd GPU. @fclem

Ankit Meel commented

2020-06-09 08:29:17 +02:00

https://mesa3d.org
Can you update the mesa version ?

https://mesa3d.org Can you update the mesa version ?

Jean Da Costa commented

2020-06-09 18:11:29 +02:00

In #77641#949405, @ankitm wrote:
https://mesa3d.org
Can you update the mesa version ?

I tried a later version and nothing changed.

> In #77641#949405, @ankitm wrote: > https://mesa3d.org > Can you update the mesa version ? I tried a later version and nothing changed.

Jean Da Costa commented

2020-06-09 22:08:26 +02:00

ok, I was playing with perf and got these record results. Maybe it could shine some light Idk.

Samples: 139K of event 'cycles', Event count (approx.): 97085600534
Overhead  Command          Shared Object           Symbol
  46,64%  blender          blender                 [.] GPU_draw_list_submit
   8,11%  blender          blender                 [.] BLI_mempool_iterstep
   6,46%  blender          blender                 [.] test_frustum_aabb
   2,83%  blender          blender                 [.] dot_v3v3
   2,56%  blender          blender                 [.] pbvh_iter_next
   1,29%  blender          blender                 [.] draw_call_batching_do
   1,22%  blender          blender                 [.] BLI_ghash_lookup 
   1,14%  blender          blender                 [.] bmesh_elem_check
   1,05%  blender          radeonsi_dri.so         [.] 0x000000000056a7a1
   1,00%  blender          blender                 [.] BM_mesh_elem_index_validate
   0,98%  blender          blender                 [.] update_search_cb
   0,98%  blender          blender                 [.] pbvh_stack_push
   0,92%  blender          radeonsi_dri.so         [.] 0x000000000013d56b
   0,90%  blender          blender                 [.] BLI_mempool_len
   0,78%  blender          radeonsi_dri.so         [.] 0x000000000073fe8f
   0,75%  blender          blender                 [.] BM_mesh_bm_to_me
   0,75%  blender          blender                 [.] draw_shgroup
   0,74%  blender          blender                 [.] BKE_pbvh_draw_cb
   0,68%  blender          blender                 [.] BLI_ghash_haskey
   0,60%  blender          blender                 [.] bmesh_radial_length
   0,55%  blender          radeonsi_dri.so         [.] 0x000000000073fe6e
   0,45%  blender          blender                 [.] BKE_pbvh_search_gather
   0,44%  blender:gdrv0    radeonsi_dri.so         [.] 0x00000000002a287a
   0,43%  blender:gdrv0    radeonsi_dri.so         [.] 0x0000000000314944
   0,39%  blender          blender                 [.] bmiter__elem_of_mesh_step
   0,38%  blender          blender                 [.] pbvh_draw_search_cb
   0,37%  blender          blender                 [.] CustomData_from_bmesh_block
   0,34%  blender          blender                 [.] BKE_pbvh_node_frustum_contain_AABB
   0,30%  blender          libc-2.27.so            [.] __memcpy_ssse3
   0,27%  blender          radeonsi_dri.so         [.] 0x000000000073fe9f
   0,27%  blender          blender                 [.] BLI_memblock_elem_get
   0,25%  blender:gdrv0    radeonsi_dri.so         [.] 0x00000000003147d5
...

ok, I was playing with perf and got these record results. Maybe it could shine some light Idk. ``` Samples: 139K of event 'cycles', Event count (approx.): 97085600534 Overhead Command Shared Object Symbol 46,64% blender blender [.] GPU_draw_list_submit 8,11% blender blender [.] BLI_mempool_iterstep 6,46% blender blender [.] test_frustum_aabb 2,83% blender blender [.] dot_v3v3 2,56% blender blender [.] pbvh_iter_next 1,29% blender blender [.] draw_call_batching_do 1,22% blender blender [.] BLI_ghash_lookup 1,14% blender blender [.] bmesh_elem_check 1,05% blender radeonsi_dri.so [.] 0x000000000056a7a1 1,00% blender blender [.] BM_mesh_elem_index_validate 0,98% blender blender [.] update_search_cb 0,98% blender blender [.] pbvh_stack_push 0,92% blender radeonsi_dri.so [.] 0x000000000013d56b 0,90% blender blender [.] BLI_mempool_len 0,78% blender radeonsi_dri.so [.] 0x000000000073fe8f 0,75% blender blender [.] BM_mesh_bm_to_me 0,75% blender blender [.] draw_shgroup 0,74% blender blender [.] BKE_pbvh_draw_cb 0,68% blender blender [.] BLI_ghash_haskey 0,60% blender blender [.] bmesh_radial_length 0,55% blender radeonsi_dri.so [.] 0x000000000073fe6e 0,45% blender blender [.] BKE_pbvh_search_gather 0,44% blender:gdrv0 radeonsi_dri.so [.] 0x00000000002a287a 0,43% blender:gdrv0 radeonsi_dri.so [.] 0x0000000000314944 0,39% blender blender [.] bmiter__elem_of_mesh_step 0,38% blender blender [.] pbvh_draw_search_cb 0,37% blender blender [.] CustomData_from_bmesh_block 0,34% blender blender [.] BKE_pbvh_node_frustum_contain_AABB 0,30% blender libc-2.27.so [.] __memcpy_ssse3 0,27% blender radeonsi_dri.so [.] 0x000000000073fe9f 0,27% blender blender [.] BLI_memblock_elem_get 0,25% blender:gdrv0 radeonsi_dri.so [.] 0x00000000003147d5 ... ```

Jean Da Costa commented

2020-06-09 23:24:09 +02:00

Ok, I am sort of sure that GPU_draw_list_submit is from where the lag comes.

Ok, I am sort of sure that GPU_draw_list_submit is from where the lag comes. ![flamegraph.svg](https://archive.blender.org/developer/F8604574/flamegraph.svg) ![image.png](https://archive.blender.org/developer/F8604576/image.png)

Ankit Meel commented

2020-06-10 08:12:06 +02:00

Changed status from 'Needs Triage' to: 'Needs Developer To Reproduce'

Ankit Meel commented

2020-06-10 08:12:06 +02:00

Thank you for the bisect and perf graph!

Clément Foucault commented

2020-06-10 11:17:43 +02:00

@jeacom Can you test this?

diff --git a/source/blender/gpu/intern/gpu_extensions.c b/source/blender/gpu/intern/gpu_extensions.c
index 8dd468b5414..277d23d19a0 100644
--- a/source/blender/gpu/intern/gpu_extensions.c
+++ b/source/blender/gpu/intern/gpu_extensions.c
@@ -227,7 +227,7 @@ bool GPU_context_local_shaders_workaround(void)
 bool GPU_crappy_amd_driver(void)
 {
   /* Currently are the same drivers with the `unused_fb_slot` problem. */
-  return GG.broken_amd_driver;
+  return true;
 }
 
 void gpu_extensions_init(void)```

@jeacom Can you test this? ``` diff --git a/source/blender/gpu/intern/gpu_extensions.c b/source/blender/gpu/intern/gpu_extensions.c index 8dd468b5414..277d23d19a0 100644 --- a/source/blender/gpu/intern/gpu_extensions.c +++ b/source/blender/gpu/intern/gpu_extensions.c @@ -227,7 +227,7 @@ bool GPU_context_local_shaders_workaround(void) bool GPU_crappy_amd_driver(void) { /* Currently are the same drivers with the `unused_fb_slot` problem. */ - return GG.broken_amd_driver; + return true; } void gpu_extensions_init(void)```

Jean Da Costa commented

2020-06-10 11:59:03 +02:00

In #77641#950519, @fclem wrote:
@jeacom Can you test this?

diff --git a/source/blender/gpu/intern/gpu_extensions.c b/source/blender/gpu/intern/gpu_extensions.c
index 8dd468b5414..277d23d19a0 100644
--- a/source/blender/gpu/intern/gpu_extensions.c
+++ b/source/blender/gpu/intern/gpu_extensions.c
@@ -227,7 +227,7 @@ bool GPU_context_local_shaders_workaround(void)
 bool GPU_crappy_amd_driver(void)
 {
   /* Currently are the same drivers with the `unused_fb_slot` problem. */
-  return GG.broken_amd_driver;
+  return true;
 }
 
 void gpu_extensions_init(void)```

I tried it and the performance is back at how it was before the commit, much better when dyntopo is enabled.
It still makes me wonder why object mode drawing feels orders of magnitude faster than sculpt mode at high polycounts.
2020-06-10 06-53-28.mp4

> In #77641#950519, @fclem wrote: > @jeacom Can you test this? > ``` > diff --git a/source/blender/gpu/intern/gpu_extensions.c b/source/blender/gpu/intern/gpu_extensions.c > index 8dd468b5414..277d23d19a0 100644 > --- a/source/blender/gpu/intern/gpu_extensions.c > +++ b/source/blender/gpu/intern/gpu_extensions.c > @@ -227,7 +227,7 @@ bool GPU_context_local_shaders_workaround(void) > bool GPU_crappy_amd_driver(void) > { > /* Currently are the same drivers with the `unused_fb_slot` problem. */ > - return GG.broken_amd_driver; > + return true; > } > > void gpu_extensions_init(void)``` I tried it and the performance is back at how it was before the commit, much better when dyntopo is enabled. It still makes me wonder why object mode drawing feels orders of magnitude faster than sculpt mode at high polycounts. [2020-06-10 06-53-28.mp4](https://archive.blender.org/developer/F8605617/2020-06-10_06-53-28.mp4)

blender-admin commented

2020-07-01 02:10:10 +02:00

This issue was referenced by 2087b87807

This issue was referenced by 2087b878078bf16f212e9bce48c05b4456b90d96

blender-admin commented

2020-07-01 02:10:10 +02:00

This issue was referenced by aeb4792044

This issue was referenced by aeb47920447da116b776f08379d5029c052fbc9c

Clément Foucault commented

2020-07-01 02:12:09 +02:00

Changed status from 'Needs Developer To Reproduce' to: 'Resolved'

Clément Foucault closed this issue

2020-07-01 02:12:09 +02:00

Clément Foucault self-assigned this 2020-07-01 02:12:09 +02:00

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

sculpt mode performance regression in 2.83. #77641