What problem does this solve or what need does it fill?
Currently, using the Mesh API incurs a lot of unnecessary memcpy. If you have a custom mesh representation, it's even worse because you have to build a Mesh copying all your data in, and then go through all the normal Mesh overhead.
To solve 2:
introduce an api which both Mesh and custom mesh representations uses, that permits writing directly into a u8 buffer thats already mapped. this can be via a builder style api potentially.
To solve 3:
a. length can be computed without copying, b. write directly into the slab instead of creating a vec and then copying
To solve 4:
create_buffer_init() is a convenience function. If you want to minimize copies, then call create_buffer() with mapped_at_creation: true, then write into the mapping. - kpreid
What alternative(s) have you considered?
Not changing it
Additional context
i mostly want to expose an api that allows alternative mesh representations to exist easily within the renderer, and make Mesh use that api.
atlv24
commented
2 months ago
What problem does this solve or what need does it fill?
Currently, using the Mesh API incurs a lot of unnecessary memcpy. If you have a custom mesh representation, it's even worse because you have to build a Mesh copying all your data in, and then go through all the normal Mesh overhead.
Current state of Mesh copies:
Mesh gets extracted which may involve a clone in
extract_render_assetextracted_assets.push((id, asset.clone()));then it gets prepared into a GpuMesh, which interleave-copies the vertex data into a
Vec<u8>inMesh::get_vertex_buffer_data:let mut attributes_interleaved_buffer = vec![0; vertex_count * vertex_size];the interleave copy actually happens twice, the first time just to get the length and pre-allocate a slab 😬
mesh.get_vertex_buffer_data().len() as u64,and then to fill in the buffer
let vertex_data = mesh.get_vertex_buffer_data();create_buffer_with_datawhich immediately copies it yet again inDeviceExt::create_buffer_init:buffer.slice(..).get_mapped_range_mut()[..unpadded_size as usize].copy_from_slice(descriptor.contents);which then actually gets uploaded to the gpuWhat solution would you like?
To solve 1: Retained render world will help here
To solve 2: introduce an api which both Mesh and custom mesh representations uses, that permits writing directly into a u8 buffer thats already mapped. this can be via a builder style api potentially.
To solve 3: a. length can be computed without copying, b. write directly into the slab instead of creating a vec and then copying
To solve 4:
create_buffer_init()is a convenience function. If you want to minimize copies, then callcreate_buffer()withmapped_at_creation: true, then write into the mapping. - kpreidWhat alternative(s) have you considered?
Not changing it
Additional context
i mostly want to expose an api that allows alternative mesh representations to exist easily within the renderer, and make Mesh use that api.