User:Ankitm/GSoC 2020/Daily Reports

Daily Reports

Week 12

Weekly Summary User:Ankitm/GSoC_2020_Weekly_Reports#Week_12_August_17_-_21

August 23

August 22

  • Use blender::StringRef instead of std::string_view. rB73454388
  • Use remove_prefix for easy transition to StringRef. rBae122533

August 21

  • For polygons with holes:
    • Commit as it is ngon-tessellate. rBefab0bc7
    • Fix tessellate crash due to wrong Vector size. rB5b0cb5bb
    • Add newly creates faces to imported faces and dissolve edges. rBe76ab124
    • Edit FaceElem list outside polygon creation code. rB10e3f23b
  • Move object creation code from constructor. rB4777a6a5
  • Remove unused uv vertex index offset. rBedd5307e

August 19-20

  • Kept working on ngon_tessellate and adding the new polygons to the mesh.

August 18

  • Fix overlapping nodes in the nodetree. rB0c7801fa
  • Silence normal being zero warnings in mesh validate's verbose mode. rB3f8a8d17
  • Cleanup: use const, remove single use variables. rB69270c78

Working on writing ngon_tessellate and tessellate_polygon in C++ for ngon polygons with holes in them.

August 17

  • Parse vertex normals: vn lines. rBc583afd6
  • Decouple getting smooth groups from export parameters rB0479dc41
  • Export vertex normals if smooth shading is enabled. rB4923087d
  • Move file open errors to constructors. rB491bd8b8
  • Add a public member function in MaterialWrap. rBc815ba0a

Week 11

Weekly summary: User:Ankitm/GSoC_2020_Weekly_Reports#Week_11_August_10_-_15

14 August

  • Apply axes transform to exportable Nurbs. rBd882a63e
  • Work on separating material data collection and MTL writers. Fix bugs in MTL import too.

13 August

  • Fix usemtl line not being written. rB69f70829
  • Use function pointer; remove redundant variables. rBba0d376e
  • Export packed image file names as if they're in current directory. rBd9cdfba2
  • Use const; move constructors to implementation file. rBa725b6f7
  • Fix several material import bugs, remove wrong assert. rB606b0d7d

12 August

  • Fix several material import bugs, remove wrong assert. rB606b0d7d.

11 August

  • Fix loose edge export & import, and offsets. rBc010cf58.
  • Fix vertex offsets for MEdges and MLoops. rBe0c0b8ff.

10 August

  • Catch out of range exceptions for stoi/stof. rBbb2eca07.
  • Add vertex offset due to Curves' vertices also. rBb378f84f.
  • Fix new Geometry creation with empty initial object. rB009b3771.
  • Use const& instead of pointers in Mesh class. rB82eff7a0.
  • Correct geometry creation code with Curves. rBd062c712.
  • Accept 3D UV vertex coordinates rB6ee696e5.
  • Create new Geometry instance at the start. rB9b37f943.

Week 10

Weekly Summary User:Ankitm/GSoC_2020_Weekly_Reports#Week_10_August_4_-_7

August 7

Minor cleanups

  • Remove deleted header from CMakeLists.txt. rB6e419e95.
  • Use pragma once as per style guide. rB0a339bb5.
  • Fix crash due to OBJ files with no o name line. rB3cff0de4.

August 6

August 5

  • Cleanup: clang-tidy warnings rBb8e4d4c6
  • Load image and add it to image texture node. rBf9348be8
  • Add created material to the object. rB9b8f2042
  • Replace enum and its functions with Map for maps rB928736b1

August 4

Committed material creation prototype & fixed several issues later on.

  • Cleanup: Use MutableSpan, add asserts rB48d35821.
  • Initial material creation support. rBc401d8a0.
  • MTL: release shader_output_ node instead of bsdf_. rB8e58bd09.
  • MTL: use mtllib for importing MTL files. rB722a793b.
  • Use Map's lookup instead of direct pointer to MTLMaterial rB64ff38a7.
  • Fix bsdf_ being null; asserts for SOCK_RGBA rBd5ad01ed.

August 3

Holiday

Week 9

Weekly Summary: User:Ankitm/GSoC_2020_Weekly_Reports#Week_9_July_27_-_31

July 30

  • Add MTL parser. rB19145856
  • Off-topic (Master) Enable header auto-complete suggestions in Xcode. rB27d50d6f This turned into a standardised way of dealing with Xcode's include file paths.
  • Off-topic Respond to a new developer on a bug fix and the devtalk thread

July 29

  • Add fallback value in string to number conversions. rBc5bd1631
  • Tested some Xcode attributes used for header search paths with a final goal of enabling auto-complete suggestions,

July 28

July 27

  • Discussed naming conventions with mentors.
  • Worked on MTL parser.
  • Off-topic: checked out ClangBuildAnalyzer on Blender (lite build).

Week 8

Weekly Summary: User:Ankitm/GSoC_2020_Weekly_Reports#Week_8:_July_20_-_25

July 24

  • Support importing vertex deform groups. rB25526821.
  • Keep vertex deform group names and def_nr in sync rB37467ec5. To handle the same deform groups appearing in the file multiple times, a deduplication is required and order of occurrence of deform groups is also important. So use std::vector<std::string> instead of Map<std::string, int>.
  • Use VectorSet instead of vector for def_group. rBb33a4592. The commit rB37467ec5 used std::vector<std::string> which is slow due to linear searching. Jacques Lucke suggested using VectorSet for faster lookups. Also StringRef is used to avoid copying.
  • Fix curve not being visible in Object mode. rBaf278ce5. Thanks to Howard Trickey for debugging this!
  • Cleanup: silence clang-tidy warnings; comments. rBffaa1df4.

July 23

  • Move Object creation code in Mesh creation class. rB2bb56c83. This helps adding deform groups and materials to the Object in one place and away from the code that adds that Object to a Collection.
  • Rename exportable_meshes to exportable_as_mesh rB47166605. (This was long due.)
  • Refactor: Conform to style guide, add documentation. rB97aa9d44.

July 22

  • Worked on Refactor: Conform to style guide, add documentation, which got committed on July 23rd. 97aa9d44fa.
  • Set active NURBS of a curve; set object group. cab598f2b2.
  • The NURBS bug mentioned yesterday is still there.

July 21

  • Support NURBS curve import rB6e21f8c2. BUT the vertices are visible in edit mode, but the black spline is not visible in edit mode. Committed the rest of the code since it was working to some extent. Will try to fix it next day.
  • Fix build errors: avoid uint with Vector's size() function. rB7bd38c27. Style guide was updated to discourage uint usage and {} initialisation doesn't allow narrowing conversions. There's still a lot of code in OBJ exporter that uses uint, that's for a Friday. This commit was only for removing uint with Vector.size().

July 20

  • Move vertex (and UV vertex) list outside OBJRawObject so that objects that are defined (o curvename) after the list of vertices is written can access them using relative indices. rB71eadb4b
  • Kept working on importing NURBS curve.

Week 7

Weekly Summary: User:Ankitm/GSoC_2020_Weekly_Reports#Week_7:_July_13_-_17

July 17

  • Use MutableSpan instead of raw pointers in string to float converter 65fd3be1fa
  • Support meshes with edges with or without polygons. 5a9b983263
  • Cleanup: Rename variables, use {} initialisation. 37d59dbc8c

July 16

  • Importer: Support smooth shading for polygons. rB031c4732
  • Importer: Fix misaligned UV vertices of a mesh rBfe4c5350
  • Move stoi into a function; Add exception handling rBef5941f3
  • Fix UV crashes and redundant loops in v1//vn1 case. rB92be92be
  • Exporter: fix vertex normals not being exported when smooth shading is enabled. Calculating smooth groups before exporting normals was the solution. rB501ead4b
  • Exporter: Fix freeing invalid pointer in ~OBJMesh rBc275b278
  • Exporter: Fix UV indices being edited by the writer. rB151e8825

July 15

  • Add UV vertices to the mesh. a4a1184ece. This needs some corrections though: the coordinates are correct, but UV edges are not matched to edges in the mesh. Also spent some time trying to find the problem of the assert: (uniform), function GPU_batch_uniform_4fv, it was present in rBc8a62e4f has been fixed before rB065a00ee.
  • Exporter: Fix UV vertex indices off by one. d68899e99a
  • Cleanup single use variable, comments, formatting aa9e4b23e5
  • Responded to some build error and a bug fix help post on Devtalk.

July 14

  • Use stoi and stof instead of >> for faster parsing. aacb1f4756. Time taken in string to float conversion is less now as compared to >> operator. But the cost is in splitting the string and allocations made for several substrings. Overall this is faster.
    OBJ import the same file as yesterday using stoi. cube subd 8; 23.3MB file.
  • Parse UV coordinates and normals. 5e9196ed11
  • Directly create Mesh without using intermediate BMesh ec04edfe5c. It lowered the peak memory usage by half. Also made mesh creation faster.
  • Refactor: Keep mesh_from_raw_object in a class 3cfcf37cea

July 13

  • Refactor: Keep mesh_from_raw_object in a class rB3cfcf37c
  • Rename files: exporter -> ex, importer -> im. rB582bf439. It will be reverted and files moved into folders though.
  • Break code into mesh, objects, and reader files. rB7582bbd5

Week 6

Weekly Summary: User:Ankitm/GSoC_2020_Weekly_Reports#Week6:_July_6_-_10

July 8, 9, 10

  • Parsed and stored the file contents in a temporary, OBJ-local OBJRawObject structure. It is not the Blender object, rather an OBJ object with the data from the OBJ file.
  • Successfully created a mesh in the viewport and added it to an Blender object. Also added the new Blender object to a collection in viewport.
    • So far the mesh has vertex data only. Texture coordinates and normals are not there yet.
  • Initial profiling shows ~70% time being spent in parsing and storing the data in OBJRawObject. Rest ~30% is spent in making a Blender object from a Raw one. So there's a huge room for optimization. See attached image.
Imported a level 8 subdivided cube back using new C++ importer. OBJ File size: 23.3 MB. Corresponds to the commit rBf8d64b39
  • Also set up clang-11 and its leak detector since clang-10 in Xcode and in /Library/Developer/Command Line Tools doesn't have the lsan library. Needed to edit and enable @rpath.

July 7

July 6

  • Set up clang-tidy on macOS. Had to build llvm from source, and remove extra executables from PATH to avoid conflicts with other compilers.
  • Got carried away a little bit and experimented a little bit to get clang-tidy to work with Xcode, using a compilation database (JSON) with clang-tidy etc. Finally settled with using Ninja as the generator in a new folder, solely for checking clang-tidy warnings.

Week 5

Weekly Summary: User:Ankitm/GSoC_2020_Weekly_Reports#Week_5:_June_29_-_July_4

July 3

  • Lots of tiny changes: rBc2eb16f6
    • Adhering to recently updated style guide Style_Guide/C_Cpp#C.2B.2B_Namespaces. Specifically, _ in the end of private class members, combined namespace, reorder class members, use const wherever suitable.
    • Remove documentation comments from headers and keep only at implementation. Reduces duplication, and reading a class easier.
    • Break functions doing different things into smaller ones.
    • Better comments.

(It was like a code quality day for me.)

July 2

  • Fix T78533: Skip writing image texture in MTL (map_Kd ) if image filepath is empty.
  • Support exporting smooth groups for smooth shaded meshes rBd85c2620.
  • Worked on feedback received by mentors for evaluation. Google Doc. Since the comments were on commits not posted for review, and I don't receive comments on my branch commits, the doc was made.

July 1

  • Support vertex group to which a face element belongs. rB012175f8. It took some followup commits to make it working properly.

June 30

  • Build error: Replace nodes_with_idname with nodes_by_type 71c6d384c1, namespace BKE > bke e3e5ae5da8
  • Cleanup: simplified conditionals, remove warnings on linux 2585882973
  • Use unique_ptr to avoid early destruction of new meshes. 9279377f4b
  • Remove axis transform defines and use existing enums c2e43ca99a
  • Working on exporting vertex groups.

June 29

  • Null checks

Week 4

Weekly Summary: User:Ankitm/GSoC_2020_Weekly_Reports#Week_4:_June_22-27

June 25

  • Refactored to separate OBJMesh and OBJNurbs classes in separate files. rB6c98925d
  • Trying to keep MTL writer in C++ itself: wrote object's material name and material library filename in OBJ file, create MTL file.
    • Need to extract required properties from the Principled-BSDF node. If it is not present, default hardcoded values are being written.

June 24

  • Clean up OBJNurbs to be consistent with OBJMesh. Also, Nurbs surfaces are exported after being converted to mesh. Nurbs curves are exported in form of control points if "export curves as nurbs" is checked. rB31d48017
  • Fixed a bug: duplicated UV vertices and indices. rBef0eff0b
  • To reuse the MTL writer in export_obj.py : Testing with registering an operator in Python and calling it from C++, and also sending some strings to be operator upon.

June 23

  • Update D8089 in response to the review. Goal was to move more variables to private access in OBJMesh & OBJWriter and expose functions only.
  • Finished Nurbs support. Now just need to refactor it like OBJMesh.

June 22

  • Cleaned up and put D8089 up for review.
  • Worked on exporting curves as nurbs.
  • Replaced std::vector with blender::Vector proposed in D7931

Week 3

Weekly Summary: User:Ankitm/GSoC_2020_Weekly_Reports#Week_3:_June_15-19

June 19

  • Worked on exporting curves as meshes.
  • Started working on refactor to reduce memory allocation and copying the mesh data into it. Rather at the time of writing, relevant info should be obtained on-the-fly.

June 18

  • Committed mesh triangulation feature. rBc3f8e955
  • Renamed occurrences of objects with ob_mesh or similar to indicate that mesh objects are being handled. It is necessary to differentiate it from curve objects.
  • Made changes in response to the review in D7959

June 17

  • Made writing UV coordinates optional (default true). By extension, the faces are now written as f v1//vn1 ...
  • Face normals are optional too (default true). Faces are written as f v1/vt1 ... rB41f47cd8
  • Added scaling factor. (default 1.000) rBd47318fb
  • Worked on triangulation of meshes. Need to fix a memory leak though.

June 16

  • Slow day today. Worked on axes transform (up/forward) and exposing it in the exporter preferences UI.

June 15

  • Allowed all frames to be exported, instead of 0-1000. Added animation checkbox to make single frame export intuitive.
  • Destination filename is edited only when multiple frames are being exported. rB45461012
  • Found and installed XcodeClangFormat for easy formatting inside the IDE itself. (Off-topic: Plugins were deprecated in Xcode 8 and extensions are now the way to go.)

Week 2

Weekly Summary: User:Ankitm/GSoC_2020_Weekly_Reports#Week_2:_June_8-13

June 12

  • Finished work on multiple objects to be exported in the same file.
  • Need to test its accuracy. Will commit tomorrow.

June 11

  • Not a lot of work today unfortunately.
  • Tested {fmt::fprintf} library as compared to fprintf. Saw only 10-15% speed up.
  • Pushed work upto UV coordinates. rBf01a7d50.
  • Fixed a out of bounds crash on windows in rBc2fbfb42.

June 10

  • Finished exporting non-duplicate and exhaustive list of UV coordinates and also stored their correct indices in Polygons.
  • Tested it by exporting with C++ exporter and importing it back with python importer and comparing the texture map with that of original object.
  • Minor review update in D7959.

Todo

  • Export multiple objects in the same file, multiple frames (different files).
  • Axes transform.

June 9

  • My initial attempt at UV coords was wrong. The list was incomplete, since it was not iterating over the whole me_eval->mloopuv list; there were duplicates since a mesh vertex can be in multiple places in the texture map.
  • Asked about it in chat and found BKE_mesh_uv_vert_map_create. Took me a lot of time to adapt it to my requirements.
  • Ended the day at merely listing non-duplicated list of vertices.

Todo Next day

  • Now that we have a list of all UV vertex coordinates, we need to match every 3D vertex of all polygons with that list. So todo is to add correct indices, vt1, vt2.. and store it in every Polygon member to export it like:

vt1 u v
vt2 u v
f */vt1/* */vt2/*

June 8

  • Updated soc-2020-io-performance to improve readability inside loops using iteration variables. Minor namespace cleanup. 7af1ec516f and 86845ea907
  • Merged those changes in review branch & created D7959.
  • Searched for averaging function, to replace face_normal_axis_component, but only found some functions in cycles utilities. However calculating face normals using vertex position can be done using cross product of two edges. That function is there. Downside is wasted memory in data_to_export->mvert[i].no.
  • Worked on UV coordinates. Testing for correctness at the end of the day.

Week 1

Weekly summary: User:Ankitm/GSoC_2020_Weekly_Reports#Week_1:_June_1_to_June_6

June 5

  • Correctly transformed vertex normals according to object's transformation matrix. Still need to incorporate axes transformation matrix.
  • Also calculated face normal from its vertices' normals inlined while writing them to the file. It's a dirty averaging function.
    • This removed the need of BLI::Vector<uint> face_normal_index;. Since vn and f lines are written by looping over the same polygons, their indices would be the same.
    • This also ensures that the memory allocated in MVert *mvert is used better. We were discarding all data_to_export->mvert[i].no arrays. Now we don't. They're used to calculate face normals on the fly.
    • It'll be more clear when I'll send it for review.
  • Added object's name in the file.
  • Encountered non-freed memory while testing both of these exporters^. Since that has been fixed in master, most probably with the commits between 25th May & 28th May, I'll merge master once.
  • Did a comparison between the old & new exporters. Release configuration, lite build + opensubdiv, default cube with 6+3 subsurf. See images: F8586987 F8586984

In the table below, we're checking the time taken by wm_handler_fileselect_do function. Release build.

Object File Size (note that vt are not there in C++ yet so C++ file size is smaller.)
(C++/ Python + MTL)
Export Time (fstream/fprintf)
(s)
Export time Python
(s)
Default cube + 8 subdivision surface levels 48 MB / 63 MB 2.4 / 1.47 s 20 s
Default cube + 9 subdivision surface levels 186 MB / 262 MB + 229 bytes MTL 7.6 / 4.6 s 82 s
Ico-sphere + 8 subdivision surface levels 478 MB / 674 MB + 124 bytes MTL 22 / 13 s 210 s


Todo on Weekend:

  • Texture coordinates.
  • Try to add multiple objects. An object can be a made a class instance with data_to_export its struct. But we'll discuss.

June 4

  • Review update on D7918 and D7917:
    • Renamed obj to wavefront_obj in filenames to remove confusion with blender's data types. Also renamed C++ headers to .hh. moved the code to wavefront_obj/intern/ leaving only wavefront_obj/IO_wavefront_obj.h externally.
    • Improved comments in structs & functions.
    • arc diff wasn't the best tool for getting two sequential commits reviewed. So there was some branch juggling, diff files creation & resets. Now onwards, I'll stick to getting one change reviewed at a time → Merge it to soc branch & repeat.
  • Got familiar with Xcode's Address sanitizer. (also fixed a small heap buffer overflow).
  • Tomorrow: I'll be working on transforming vertex normals, _obtaining_ face normals from vertex normals, write object name to the file. _Writing_ face normals to file is a on halt waiting for D7931

June 3

  • Finished major changes in io/obj/*.cc files.
  • Uploaded D7918 and D7917 which will be committed to the soc-2020-io-performance branch after review.
  • Experimenting with "multiplying obmat and mvert.co while writing to the file" vs "multiplying earlier & keeping its copies" as per Howard's comment D7917#191581. But first I need to find appropriate method to transform normals. After data structure for polygons is finalised, I'd be able to fix its issue with wrong vertex_index.
  • Also updated project plan to give more time and spread to such refactoring and bug fixing. Squeezing PLY and STL.

June 2

  • Received review from mentors on the commit rB485cc433.
  • Worked on resolving the straightforward bits, renaming files (.h → .hh), writing comments & proper naming.
  • Will finish the changes on the next day for review before committing it. That involves finding how to transform face normals according to global transform, using BLI_* instead of STL, breaking up functions into smaller, specific ones.

June 1

  • Cleaned and prepared a commit for the work done in the past 4 days. See rB485cc433.
  • The UI blocks and dummy files were added in week 0. After that, OBJ Exporter now exports a single, selected object's, vertices, vertex normals, and faces. Need to verify it with complex geometries. The axes transformation (Z_up, Y_forward) is not there yet.
  • Tried to get ASan to work on macOS, but failed. It's not supported, see the docs. gperftools' heap-profiler & heap-checker work on simple standalone scripts, not in Blender. Instruments.app's leak detector crashes after detecting a leak that happens without even doing anything.