me
/
Univerxel
1
0
Fork 0
Code Releases Activity

queue lock handle, mesh_opt

This commit is contained in:
May B. 2020-08-08 13:01:07 +02:00
parent 05c69190d2
commit 2f3335fe88
30 changed files with 15555 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
CMakeLists.txt
View File

@ -39,7 +39,7 @@ add_definitions(
)
file(GLOB_RECURSE SOURCES "src/*/*.cpp")
file(GLOB INCLUDE_SOURCES "include/imgui-1.76/*.cpp" "include/FastNoiseSIMD/*.cpp" "include/tracy/TracyClient.cpp")
file(GLOB INCLUDE_SOURCES "include/imgui-1.76/*.cpp" "include/FastNoiseSIMD/*.cpp" "include/tracy/TracyClient.cpp" "include/meshoptimizer/*.cpp")
set(INCLUDE_LIBS
"include/imgui-1.76"
"include/FastNoiseSIMD"
@ -47,6 +47,7 @@ set(INCLUDE_LIBS
"include/robin_hood"
"include/libguarded"
"include/tracy"
"include/meshoptimizer"
)
add_executable(univerxel "src/main.cpp" ${SOURCES} ${INCLUDE_SOURCES})
@ -75,7 +76,7 @@ file(GLOB SMP_SOURCES "src/zstd_sampler.cpp" "src/world/Chunk.cpp" "include/Fast
add_executable(zstd_sampler EXCLUDE_FROM_ALL ${SMP_SOURCES})
target_compile_features(zstd_sampler PUBLIC cxx_std_17)
target_link_libraries(zstd_sampler ${LINKED_LIBS})
target_include_directories(zstd_sampler PRIVATE "include/FastNoiseSIMD")
target_include_directories(zstd_sampler PRIVATE "include/FastNoiseSIMD" "include/robin_hood")
# target_compile_definitions(zstd_sampler PRIVATE HN_USE_FILESYSTEM=1)
add_custom_command(OUTPUT "${CMAKE_BINARY_DIR}/content/zstd.dict"

54
include/meshoptimizer/CONTRIBUTING.md Normal file
View File

@ -0,0 +1,54 @@
Thanks for deciding to contribute to meshoptimizer! These guidelines will try to help make the process painless and efficient.
## Questions
If you have a question regarding the library usage, please [open a GitHub issue](https://github.com/zeux/meshoptimizer/issues/new).
Some questions just need answers, but it's nice to keep them for future reference in case other people want to know the same thing.
Some questions help improve the library interface or documentation by inspiring future changes.
## Bugs
If the library doesn't compile on your system, compiles with warnings, doesn't seem to run correctly for your input data or if anything else is amiss, please [open a GitHub issue](https://github.com/zeux/meshoptimizer/issues/new).
It helps if you note the version of the library this issue happens in, the version of your compiler for compilation issues, and a reproduction case for runtime bugs.
Of course, feel free to [create a pull request](https://help.github.com/articles/about-pull-requests/) to fix the bug yourself.
## Features
New algorithms and improvements to existing algorithms are always welcome; you can open an issue or make the change yourself and submit a pull request.
For major features, consider opening an issue describing an improvement you'd like to see or make before opening a pull request.
This will give us a chance to discuss the idea before implementing it - some algorithms may not be easy to integrate into existing programs, may not be robust to arbitrary meshes or may be expensive to run or implement/maintain, so a discussion helps make sure these don't block the algorithm development.
## Code style
Contributions to this project are expected to follow the existing code style.
`.clang-format` file mostly defines syntactic styling rules (you can run `make format` to format the code accordingly).
As for naming conventions, this library uses `snake_case` for variables, `lowerCamelCase` for functions, `UpperCamelCase` for types, `kCamelCase` for global constants and `SCARY_CASE` for macros. All public functions/types must additionally have an extra `meshopt_` prefix to avoid symbol conflicts.
## Dependencies
Please note that this library uses C89 interface for all APIs and a C++98 implementation - C++11 features can not be used.
This choice is made to maximize compatibility to make sure that any toolchain, including legacy proprietary gaming console toolchains, can compile this code.
Additionally, the library code has zero external dependencies, does not depend on STL and does not use RTTI or exceptions.
This, again, maximizes compatibility and makes sure the library can be used in environments where STL use is discouraged or prohibited, as well as maximizing runtime performance and minimizing compilation times.
The demo program uses STL since it serves as an example of usage and as a test harness, not as production-ready code.
## Testing
All pull requests will run through a continuous integration pipeline using GitHub Actions that will run the built-in unit tests and integration tests on Windows, macOS and Linux with gcc, clang and msvc compilers.
You can run the tests yourself using `make test` or building the demo program with `cmake -DBUILD_DEMO=ON` and running it.
Unit tests can be found in `demo/tests.cpp` and functional tests - in `demo/main.cpp`; when making code changes please try to make sure they are covered by an existing test or add a new test accordingly.
## Documentation
Documentation for this library resides in the `meshoptimizer.h` header, with examples as part of a usage manual available in `README.md`.
Changes to documentation are always welcome and should use issues/pull requests as outlined above; please note that `README.md` only contains documentation for stable algorithms, as experimental algorithms may change the interface without concern for backwards compatibility.
## Sensitive communication
If you prefer to not disclose the issues or information relevant to the issue such as reproduction case to the public, you can always contact the author via e-mail (arseny.kapoulkine@gmail.com).

21
include/meshoptimizer/LICENSE.md Normal file
View File

@ -0,0 +1,21 @@
MIT License
Copyright (c) 2016-2020 Arseny Kapoulkine
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

274
include/meshoptimizer/README.md Normal file
View File

@ -0,0 +1,274 @@
# 🐇 meshoptimizer [![Actions Status](https://github.com/zeux/meshoptimizer/workflows/build/badge.svg)](https://github.com/zeux/meshoptimizer/actions) [![codecov.io](https://codecov.io/github/zeux/meshoptimizer/coverage.svg?branch=master)](https://codecov.io/github/zeux/meshoptimizer?branch=master) ![MIT](https://img.shields.io/badge/license-MIT-blue.svg) [![GitHub](https://img.shields.io/badge/repo-github-green.svg)](https://github.com/zeux/meshoptimizer)
## Purpose
When a GPU renders triangle meshes, various stages of the GPU pipeline have to process vertex and index data. The efficiency of these stages depends on the data you feed to them; this library provides algorithms to help optimize meshes for these stages, as well as algorithms to reduce the mesh complexity and storage overhead.
The library provides a C and C++ interface for all algorithms; you can use it from C/C++ or from other languages via FFI (such as P/Invoke). If you want to use this library from Rust, you should use [meshopt crate](https://crates.io/crates/meshopt).
[gltfpack](gltf), which is a tool that can automatically optimize glTF files, is developed and distributed alongside the library.
## Installing
meshoptimizer is hosted on GitHub; you can download the latest release using git:
```
git clone -b v0.14 https://github.com/zeux/meshoptimizer.git
```
Alternatively you can [download the .zip archive from GitHub](https://github.com/zeux/meshoptimizer/archive/v0.14.zip).
## Building
meshoptimizer is distributed as a set of C++ source files. To include it into your project, you can use one of the two options:
* Use CMake to build the library (either as a standalone project or as part of your project)
* Add source files to your project's build system
The source files are organized in such a way that you don't need to change your build-system settings, and you only need to add the files for the algorithms you use.
## Pipeline
When optimizing a mesh, you should typically feed it through a set of optimizations (the order is important!):
1. Indexing
2. Vertex cache optimization
3. Overdraw optimization
4. Vertex fetch optimization
5. Vertex quantization
6. (optional) Vertex/index buffer compression
## Indexing
Most algorithms in this library assume that a mesh has a vertex buffer and an index buffer. For algorithms to work well and also for GPU to render your mesh efficiently, the vertex buffer has to have no redundant vertices; you can generate an index buffer from an unindexed vertex buffer or reindex an existing (potentially redundant) index buffer as follows:
First, generate a remap table from your existing vertex (and, optionally, index) data:
```c++
size_t index_count = face_count * 3;
std::vector<unsigned int> remap(index_count); // allocate temporary memory for the remap table
size_t vertex_count = meshopt_generateVertexRemap(&remap[0], NULL, index_count, &unindexed_vertices[0], index_count, sizeof(Vertex));
```
Note that in this case we only have an unindexed vertex buffer; the remap table is generated based on binary equivalence of the input vertices, so the resulting mesh will render the same way.
After generating the remap table, you can allocate space for the target vertex buffer (`vertex_count` elements) and index buffer (`index_count` elements) and generate them:
```c++
meshopt_remapIndexBuffer(indices, NULL, index_count, &remap[0]);
meshopt_remapVertexBuffer(vertices, &unindexed_vertices[0], index_count, sizeof(Vertex), &remap[0]);
```
You can then further optimize the resulting buffers by calling the other functions on them in-place.
## Vertex cache optimization
When the GPU renders the mesh, it has to run the vertex shader for each vertex; usually GPUs have a built-in fixed size cache that stores the transformed vertices (the result of running the vertex shader), and uses this cache to reduce the number of vertex shader invocations. This cache is usually small, 16-32 vertices, and can have different replacement policies; to use this cache efficiently, you have to reorder your triangles to maximize the locality of reused vertex references like so:
```c++
meshopt_optimizeVertexCache(indices, indices, index_count, vertex_count);
```
## Overdraw optimization
After transforming the vertices, GPU sends the triangles for rasterization which results in generating pixels that are usually first ran through the depth test, and pixels that pass it get the pixel shader executed to generate the final color. As pixel shaders get more expensive, it becomes more and more important to reduce overdraw. While in general improving overdraw requires view-dependent operations, this library provides an algorithm to reorder triangles to minimize the overdraw from all directions, which you should run after vertex cache optimization like this:
```c++
meshopt_optimizeOverdraw(indices, indices, index_count, &vertices[0].x, vertex_count, sizeof(Vertex), 1.05f);
```
The overdraw optimizer needs to read vertex positions as a float3 from the vertex; the code snippet above assumes that the vertex stores position as `float x, y, z`.
When performing the overdraw optimization you have to specify a floating-point threshold parameter. The algorithm tries to maintain a balance between vertex cache efficiency and overdraw; the threshold determines how much the algorithm can compromise the vertex cache hit ratio, with 1.05 meaning that the resulting ratio should be at most 5% worse than before the optimization.
## Vertex fetch optimization
After the final triangle order has been established, we still can optimize the vertex buffer for memory efficiency. Before running the vertex shader GPU has to fetch the vertex attributes from the vertex buffer; the fetch is usually backed by a memory cache, and as such optimizing the data for the locality of memory access is important. You can do this by running this code:
To optimize the index/vertex buffers for vertex fetch efficiency, call:
```c++
meshopt_optimizeVertexFetch(vertices, indices, index_count, vertices, vertex_count, sizeof(Vertex));
```
This will reorder the vertices in the vertex buffer to try to improve the locality of reference, and rewrite the indices in place to match; if the vertex data is stored using multiple streams, you should use `meshopt_optimizeVertexFetchRemap` instead. This optimization has to be performed on the final index buffer since the optimal vertex order depends on the triangle order.
Note that the algorithm does not try to model cache replacement precisely and instead just orders vertices in the order of use, which generally produces results that are close to optimal.
## Vertex quantization
To optimize memory bandwidth when fetching the vertex data even further, and to reduce the amount of memory required to store the mesh, it is often beneficial to quantize the vertex attributes to smaller types. While this optimization can technically run at any part of the pipeline (and sometimes doing quantization as the first step can improve indexing by merging almost identical vertices), it generally is easier to run this after all other optimizations since some of them require access to float3 positions.
Quantization is usually domain specific; it's common to quantize normals using 3 8-bit integers but you can use higher-precision quantization (for example using 10 bits per component in a 10_10_10_2 format), or a different encoding to use just 2 components. For positions and texture coordinate data the two most common storage formats are half precision floats, and 16-bit normalized integers that encode the position relative to the AABB of the mesh or the UV bounding rectangle.
The number of possible combinations here is very large but this library does provide the building blocks, specifically functions to quantize floating point values to normalized integers, as well as half-precision floats. For example, here's how you can quantize a normal:
```c++
unsigned int normal =
(meshopt_quantizeUnorm(v.nx, 10) << 20) |
(meshopt_quantizeUnorm(v.ny, 10) << 10) |
meshopt_quantizeUnorm(v.nz, 10);
```
and here's how you can quantize a position:
```c++
unsigned short px = meshopt_quantizeHalf(v.x);
unsigned short py = meshopt_quantizeHalf(v.y);
unsigned short pz = meshopt_quantizeHalf(v.z);
```
## Vertex/index buffer compression
In case storage size or transmission bandwidth is of importance, you might want to additionally compress vertex and index data. While several mesh compression libraries, like Google Draco, are available, they typically are designed to maximize the compression ratio at the cost of disturbing the vertex/index order (which makes the meshes inefficient to render on GPU) or decompression performance. They also frequently don't support custom game-ready quantized vertex formats and thus require to re-quantize the data after loading it, introducing extra quantization errors and making decoding slower.
Alternatively you can use general purpose compression libraries like zstd or Oodle to compress vertex/index data - however these compressors aren't designed to exploit redundancies in vertex/index data and as such compression rates can be unsatisfactory.
To that end, this library provides algorithms to "encode" vertex and index data. The result of the encoding is generally significantly smaller than initial data, and remains compressible with general purpose compressors - so you can either store encoded data directly (for modest compression ratios and maximum decoding performance), or further compress it with zstd/Oodle to maximize compression ratio.
To encode, you need to allocate target buffers (preferably using the worst case bound) and call encoding functions:
```c++
std::vector<unsigned char> vbuf(meshopt_encodeVertexBufferBound(vertex_count, sizeof(Vertex)));
vbuf.resize(meshopt_encodeVertexBuffer(&vbuf[0], vbuf.size(), vertices, vertex_count, sizeof(Vertex)));
std::vector<unsigned char> ibuf(meshopt_encodeIndexBufferBound(index_count, vertex_count));
ibuf.resize(meshopt_encodeIndexBuffer(&ibuf[0], ibuf.size(), indices, index_count));
```
You can then either serialize `vbuf`/`ibuf` as is, or compress them further. To decode the data at runtime, call decoding functions:
```c++
int resvb = meshopt_decodeVertexBuffer(vertices, vertex_count, sizeof(Vertex), &vbuf[0], vbuf.size());
int resib = meshopt_decodeIndexBuffer(indices, index_count, &buffer[0], buffer.size());
assert(resvb == 0 && resib == 0);
```
Note that vertex encoding assumes that vertex buffer was optimized for vertex fetch, and that vertices are quantized; index encoding assumes that the vertex/index buffers were optimized for vertex cache and vertex fetch. Feeding unoptimized data into the encoders will produce poor compression ratios. Both codecs are lossless - the only lossy step is quantization that happens before encoding.
To reduce the data size further, it's recommended to use `meshopt_optimizeVertexCacheStrip` instead of `meshopt_optimizeVertexCache` when optimizing for vertex cache, and to use new index codec version (`meshopt_encodeIndexVersion(1)`). This trades off some efficiency in vertex transform for smaller vertex and index data.
Decoding functions are heavily optimized and can directly target write-combined memory; you can expect both decoders to run at 1-3 GB/s on modern desktop CPUs. Compression ratios depend on the data; vertex data compression ratio is typically around 2-4x (compared to already quantized data), index data compression ratio is around 5-6x (compared to raw 16-bit index data). General purpose lossless compressors can further improve on these results.
Index buffer codec only supports triangle list topology; when encoding triangle strips or line lists, use `meshopt_encodeIndexSequence`/`meshopt_decodeIndexSequence` instead. This codec typically encodes indices into ~1 byte per index, but compressing the results further with a general purpose compressor can improve the results to 1-3 bits per index.
Due to a very high decoding performance and compatibility with general purpose lossless compressors, the compression is a good fit for the use on the web. To that end, meshoptimizer provides both vertex and index decoders compiled into WebAssembly and wrapped into a module with JavaScript-friendly interface, `js/meshopt_decoder.js`, that you can use to decode meshes that were encoded offline:
```js
// ready is a Promise that is resolved when (asynchronous) WebAssembly compilation finishes
await MeshoptDecoder.ready;
// decode from *Data (Uint8Array) into *Buffer (Uint8Array)
MeshoptDecoder.decodeVertexBuffer(vertexBuffer, vertexCount, vertexSize, vertexData);
MeshoptDecoder.decodeIndexBuffer(indexBuffer, indexCount, indexSize, indexData);
```
[Usage example](https://meshoptimizer.org/demo/) is available, with source in `demo/index.html`; this example uses .GLB files encoded using `gltfpack`.
## Triangle strip conversion
On most hardware, indexed triangle lists are the most efficient way to drive the GPU. However, in some cases triangle strips might prove beneficial:
- On some older GPUs, triangle strips may be a bit more efficient to render
- On extremely memory constrained systems, index buffers for triangle strips could save a bit of memory
This library provides an algorithm for converting a vertex cache optimized triangle list to a triangle strip:
```c++
std::vector<unsigned int> strip(meshopt_stripifyBound(index_count));
unsigned int restart_index = ~0u;
size_t strip_size = meshopt_stripify(&strip[0], indices, index_count, vertex_count, restart_index);
```
Typically you should expect triangle strips to have ~50-60% of indices compared to triangle lists (~1.5-1.8 indices per triangle) and have ~5% worse ACMR.
Note that triangle strips can be stitched with or without restart index support. Using restart indices can result in ~10% smaller index buffers, but on some GPUs restart indices may result in decreased performance.
To reduce the triangle strip size further, it's recommended to use `meshopt_optimizeVertexCacheStrip` instead of `meshopt_optimizeVertexCache` when optimizing for vertex cache. This trades off some efficiency in vertex transform for smaller index buffers.
## Deinterleaved geometry
All of the examples above assume that geometry is represented as a single vertex buffer and a single index buffer. This requires storing all vertex attributes - position, normal, texture coordinate, skinning weights etc. - in a single contiguous struct. However, in some cases using multiple vertex streams may be preferable. In particular, if some passes require only positional data - such as depth pre-pass or shadow map - then it may be beneficial to split it from the rest of the vertex attributes to make sure the bandwidth use during these passes is optimal. On some mobile GPUs a position-only attribute stream also improves efficiency of tiling algorithms.
Most of the functions in this library either only need the index buffer (such as vertex cache optimization) or only need positional information (such as overdraw optimization). However, several tasks require knowledge about all vertex attributes.
For indexing, `meshopt_generateVertexRemap` assumes that there's just one vertex stream; when multiple vertex streams are used, it's necessary to use `meshopt_generateVertexRemapMulti` as follows:
```c++
meshopt_Stream streams[] = {
{&unindexed_pos[0], sizeof(float) * 3, sizeof(float) * 3},
{&unindexed_nrm[0], sizeof(float) * 3, sizeof(float) * 3},
{&unindexed_uv[0], sizeof(float) * 2, sizeof(float) * 2},
};
std::vector<unsigned int> remap(index_count);
size_t vertex_count = meshopt_generateVertexRemapMulti(&remap[0], NULL, index_count, index_count, streams, sizeof(streams) / sizeof(streams[0]));
```
After this `meshopt_remapVertexBuffer` needs to be called once for each vertex stream to produce the correctly reindexed stream.
Instead of calling `meshopt_optimizeVertexFetch` for reordering vertices in a single vertex buffer for efficiency, calling `meshopt_optimizeVertexFetchRemap` and then calling `meshopt_remapVertexBuffer` for each stream again is recommended.
Finally, when compressing vertex data, `meshopt_encodeVertexBuffer` should be used on each vertex stream separately - this allows the encoder to best utilize corellation between attribute values for different vertices.
## Simplification
All algorithms presented so far don't affect visual appearance at all, with the exception of quantization that has minimal controlled impact. However, fundamentally the most effective way at reducing the rendering or transmission cost of a mesh is to make the mesh simpler.
This library provides two simplification algorithms that reduce the number of triangles in the mesh. Given a vertex and an index buffer, they generate a second index buffer that uses existing vertices in the vertex buffer. This index buffer can be used directly for rendering with the original vertex buffer (preferably after vertex cache optimization), or a new compact vertex/index buffer can be generated using `meshopt_optimizeVertexFetch` that uses the optimal number and order of vertices.
The first simplification algorithm, `meshopt_simplify`, follows the topology of the original mesh in an attempt to preserve attribute seams, borders and overall appearance. For meshes with inconsistent topology or many seams, such as faceted meshes, it can result in simplifier getting "stuck" and not being able to simplify the mesh fully; it's recommended to preprocess the index buffer with `meshopt_generateShadowIndexBuffer` to discard any vertex attributes that aren't critical and can be rebuilt later such as normals.
```c++
float threshold = 0.2f;
size_t target_index_count = size_t(index_count * threshold);
float target_error = 1e-2f;
std::vector<unsigned int> lod(index_count);
lod.resize(meshopt_simplify(&lod[0], indices, index_count, &vertices[0].x, vertex_count, sizeof(Vertex), target_index_count, target_error));
```
Target error is an approximate measure of the deviation from the original mesh using distance normalized to 0..1 (so 1e-2f means that simplifier will try to maintain the error to be below 1% of the mesh extents). Note that because of topological restrictions and error bounds simplifier isn't guaranteed to reach the target index count and can stop earlier.
The second simplification algorithm, `meshopt_simplifySloppy`, doesn't follow the topology of the original mesh. This means that it doesn't preserve attribute seams or borders, but it can collapse internal details that are too small to matter better because it can merge mesh features that are topologically disjoint but spatially close.
```c++
float threshold = 0.2f;
size_t target_index_count = size_t(index_count * threshold);
std::vector<unsigned int> lod(target_index_count);
lod.resize(meshopt_simplifySloppy(&lod[0], indices, index_count, &vertices[0].x, vertex_count, sizeof(Vertex), target_index_count));
```
This algorithm is guaranteed to return a result at or below the target index count. It is 5-6x faster than `meshopt_simplify` when simplification ratio is large, and is able to reach ~20M triangles/sec on a desktop CPU (`meshopt_simplify` works at ~3M triangles/sec).
When a sequence of LOD meshes is generated that all use the original vertex buffer, care must be taken to order vertices optimally to not penalize mobile GPU architectures that are only capable of transforming a sequential vertex buffer range. It's recommended in this case to first optimize each LOD for vertex cache, then assemble all LODs in one large index buffer starting from the coarsest LOD (the one with fewest triangles), and call `meshopt_optimizeVertexFetch` on the final large index buffer. This will make sure that coarser LODs require a smaller vertex range and are efficient wrt vertex fetch and transform.
## Efficiency analyzers
While the only way to get precise performance data is to measure performance on the target GPU, it can be valuable to measure the impact of these optimization in a GPU-independent manner. To this end, the library provides analyzers for all three major optimization routines. For each optimization there is a corresponding analyze function, like `meshopt_analyzeOverdraw`, that returns a struct with statistics.
`meshopt_analyzeVertexCache` returns vertex cache statistics. The common metric to use is ACMR - average cache miss ratio, which is the ratio of the total number of vertex invocations to the triangle count. The worst-case ACMR is 3 (GPU has to process 3 vertices for each triangle); on regular grids the optimal ACMR approaches 0.5. On real meshes it usually is in [0.5..1.5] range depending on the amount of vertex splits. One other useful metric is ATVR - average transformed vertex ratio - which represents the ratio of vertex shader invocations to the total vertices, and has the best case of 1.0 regardless of mesh topology (each vertex is transformed once).
`meshopt_analyzeVertexFetch` returns vertex fetch statistics. The main metric it uses is overfetch - the ratio between the number of bytes read from the vertex buffer to the total number of bytes in the vertex buffer. Assuming non-redundant vertex buffers, the best case is 1.0 - each byte is fetched once.
`meshopt_analyzeOverdraw` returns overdraw statistics. The main metric it uses is overdraw - the ratio between the number of pixel shader invocations to the total number of covered pixels, as measured from several different orthographic cameras. The best case for overdraw is 1.0 - each pixel is shaded once.
Note that all analyzers use approximate models for the relevant GPU units, so the numbers you will get as the result are only a rough approximation of the actual performance.
## Memory management
Many algorithms allocate temporary memory to store intermediate results or accelerate processing. The amount of memory allocated is a function of various input parameters such as vertex count and index count. By default memory is allocated using `operator new` and `operator delete`; if these operators are overloaded by the application, the overloads will be used instead. Alternatively it's possible to specify custom allocation/deallocation functions using `meshopt_setAllocator`, e.g.
```c++
meshopt_setAllocator(malloc, free);
```
> Note that the library expects the allocation function to either throw in case of out-of-memory (in which case the exception will propagate to the caller) or abort, so technically the use of `malloc` above isn't safe. If you want to handle out-of-memory errors without using C++ exceptions, you can use `setjmp`/`longjmp` instead.
Vertex and index decoders (`meshopt_decodeVertexBuffer` and `meshopt_decodeIndexBuffer`) do not allocate memory and work completely within the buffer space provided via arguments.
All functions have bounded stack usage that does not exceed 32 KB for any algorithms.
## License
This library is available to anybody free of charge, under the terms of MIT License (see LICENSE.md).

8
include/meshoptimizer/allocator.cpp Normal file
View File

@ -0,0 +1,8 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
void meshopt_setAllocator(void* (*allocate)(size_t), void (*deallocate)(void*))
{
meshopt_Allocator::Storage::allocate = allocate;
meshopt_Allocator::Storage::deallocate = deallocate;
}

351
include/meshoptimizer/clusterizer.cpp Normal file
View File

@ -0,0 +1,351 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
#include <assert.h>
#include <math.h>
#include <string.h>
// This work is based on:
// Graham Wihlidal. Optimizing the Graphics Pipeline with Compute. 2016
// Matthaeus Chajdas. GeometryFX 1.2 - Cluster Culling. 2016
// Jack Ritter. An Efficient Bounding Sphere. 1990
namespace meshopt
{
static void computeBoundingSphere(float result[4], const float points[][3], size_t count)
{
assert(count > 0);
// find extremum points along all 3 axes; for each axis we get a pair of points with min/max coordinates
size_t pmin[3] = {0, 0, 0};
size_t pmax[3] = {0, 0, 0};
for (size_t i = 0; i < count; ++i)
{
const float* p = points[i];
for (int axis = 0; axis < 3; ++axis)
{
pmin[axis] = (p[axis] < points[pmin[axis]][axis]) ? i : pmin[axis];
pmax[axis] = (p[axis] > points[pmax[axis]][axis]) ? i : pmax[axis];
}
}
// find the pair of points with largest distance
float paxisd2 = 0;
int paxis = 0;
for (int axis = 0; axis < 3; ++axis)
{
const float* p1 = points[pmin[axis]];
const float* p2 = points[pmax[axis]];
float d2 = (p2[0] - p1[0]) * (p2[0] - p1[0]) + (p2[1] - p1[1]) * (p2[1] - p1[1]) + (p2[2] - p1[2]) * (p2[2] - p1[2]);
if (d2 > paxisd2)
{
paxisd2 = d2;
paxis = axis;
}
}
// use the longest segment as the initial sphere diameter
const float* p1 = points[pmin[paxis]];
const float* p2 = points[pmax[paxis]];
float center[3] = {(p1[0] + p2[0]) / 2, (p1[1] + p2[1]) / 2, (p1[2] + p2[2]) / 2};
float radius = sqrtf(paxisd2) / 2;
// iteratively adjust the sphere up until all points fit
for (size_t i = 0; i < count; ++i)
{
const float* p = points[i];
float d2 = (p[0] - center[0]) * (p[0] - center[0]) + (p[1] - center[1]) * (p[1] - center[1]) + (p[2] - center[2]) * (p[2] - center[2]);
if (d2 > radius * radius)
{
float d = sqrtf(d2);
assert(d > 0);
float k = 0.5f + (radius / d) / 2;
center[0] = center[0] * k + p[0] * (1 - k);
center[1] = center[1] * k + p[1] * (1 - k);
center[2] = center[2] * k + p[2] * (1 - k);
radius = (radius + d) / 2;
}
}
result[0] = center[0];
result[1] = center[1];
result[2] = center[2];
result[3] = radius;
}
} // namespace meshopt
size_t meshopt_buildMeshletsBound(size_t index_count, size_t max_vertices, size_t max_triangles)
{
assert(index_count % 3 == 0);
assert(max_vertices >= 3);
assert(max_triangles >= 1);
// meshlet construction is limited by max vertices and max triangles per meshlet
// the worst case is that the input is an unindexed stream since this equally stresses both limits
// note that we assume that in the worst case, we leave 2 vertices unpacked in each meshlet - if we have space for 3 we can pack any triangle
size_t max_vertices_conservative = max_vertices - 2;
size_t meshlet_limit_vertices = (index_count + max_vertices_conservative - 1) / max_vertices_conservative;
size_t meshlet_limit_triangles = (index_count / 3 + max_triangles - 1) / max_triangles;
return meshlet_limit_vertices > meshlet_limit_triangles ? meshlet_limit_vertices : meshlet_limit_triangles;
}
size_t meshopt_buildMeshlets(meshopt_Meshlet* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, size_t max_vertices, size_t max_triangles)
{
assert(index_count % 3 == 0);
assert(max_vertices >= 3);
assert(max_triangles >= 1);
meshopt_Allocator allocator;
meshopt_Meshlet meshlet;
memset(&meshlet, 0, sizeof(meshlet));
assert(max_vertices <= sizeof(meshlet.vertices) / sizeof(meshlet.vertices[0]));
assert(max_triangles <= sizeof(meshlet.indices) / 3);
// index of the vertex in the meshlet, 0xff if the vertex isn't used
unsigned char* used = allocator.allocate<unsigned char>(vertex_count);
memset(used, -1, vertex_count);
size_t offset = 0;
for (size_t i = 0; i < index_count; i += 3)
{
unsigned int a = indices[i + 0], b = indices[i + 1], c = indices[i + 2];
assert(a < vertex_count && b < vertex_count && c < vertex_count);
unsigned char& av = used[a];
unsigned char& bv = used[b];
unsigned char& cv = used[c];
unsigned int used_extra = (av == 0xff) + (bv == 0xff) + (cv == 0xff);
if (meshlet.vertex_count + used_extra > max_vertices || meshlet.triangle_count >= max_triangles)
{
destination[offset++] = meshlet;
for (size_t j = 0; j < meshlet.vertex_count; ++j)
used[meshlet.vertices[j]] = 0xff;
memset(&meshlet, 0, sizeof(meshlet));
}
if (av == 0xff)
{
av = meshlet.vertex_count;
meshlet.vertices[meshlet.vertex_count++] = a;
}
if (bv == 0xff)
{
bv = meshlet.vertex_count;
meshlet.vertices[meshlet.vertex_count++] = b;
}
if (cv == 0xff)
{
cv = meshlet.vertex_count;
meshlet.vertices[meshlet.vertex_count++] = c;
}
meshlet.indices[meshlet.triangle_count][0] = av;
meshlet.indices[meshlet.triangle_count][1] = bv;
meshlet.indices[meshlet.triangle_count][2] = cv;
meshlet.triangle_count++;
}
if (meshlet.triangle_count)
destination[offset++] = meshlet;
assert(offset <= meshopt_buildMeshletsBound(index_count, max_vertices, max_triangles));
return offset;
}
meshopt_Bounds meshopt_computeClusterBounds(const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
{
using namespace meshopt;
assert(index_count % 3 == 0);
assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
assert(vertex_positions_stride % sizeof(float) == 0);
assert(index_count / 3 <= 256);
(void)vertex_count;
size_t vertex_stride_float = vertex_positions_stride / sizeof(float);
// compute triangle normals and gather triangle corners
float normals[256][3];
float corners[256][3][3];
size_t triangles = 0;
for (size_t i = 0; i < index_count; i += 3)
{
unsigned int a = indices[i + 0], b = indices[i + 1], c = indices[i + 2];
assert(a < vertex_count && b < vertex_count && c < vertex_count);
const float* p0 = vertex_positions + vertex_stride_float * a;
const float* p1 = vertex_positions + vertex_stride_float * b;
const float* p2 = vertex_positions + vertex_stride_float * c;
float p10[3] = {p1[0] - p0[0], p1[1] - p0[1], p1[2] - p0[2]};
float p20[3] = {p2[0] - p0[0], p2[1] - p0[1], p2[2] - p0[2]};
float normalx = p10[1] * p20[2] - p10[2] * p20[1];
float normaly = p10[2] * p20[0] - p10[0] * p20[2];
float normalz = p10[0] * p20[1] - p10[1] * p20[0];
float area = sqrtf(normalx * normalx + normaly * normaly + normalz * normalz);
// no need to include degenerate triangles - they will be invisible anyway
if (area == 0.f)
continue;
// record triangle normals & corners for future use; normal and corner 0 define a plane equation
normals[triangles][0] = normalx / area;
normals[triangles][1] = normaly / area;
normals[triangles][2] = normalz / area;
memcpy(corners[triangles][0], p0, 3 * sizeof(float));
memcpy(corners[triangles][1], p1, 3 * sizeof(float));
memcpy(corners[triangles][2], p2, 3 * sizeof(float));
triangles++;
}
meshopt_Bounds bounds = {};
// degenerate cluster, no valid triangles => trivial reject (cone data is 0)
if (triangles == 0)
return bounds;
// compute cluster bounding sphere; we'll use the center to determine normal cone apex as well
float psphere[4] = {};
computeBoundingSphere(psphere, corners[0], triangles * 3);
float center[3] = {psphere[0], psphere[1], psphere[2]};
// treating triangle normals as points, find the bounding sphere - the sphere center determines the optimal cone axis
float nsphere[4] = {};
computeBoundingSphere(nsphere, normals, triangles);
float axis[3] = {nsphere[0], nsphere[1], nsphere[2]};
float axislength = sqrtf(axis[0] * axis[0] + axis[1] * axis[1] + axis[2] * axis[2]);
float invaxislength = axislength == 0.f ? 0.f : 1.f / axislength;
axis[0] *= invaxislength;
axis[1] *= invaxislength;
axis[2] *= invaxislength;
// compute a tight cone around all normals, mindp = cos(angle/2)
float mindp = 1.f;
for (size_t i = 0; i < triangles; ++i)
{
float dp = normals[i][0] * axis[0] + normals[i][1] * axis[1] + normals[i][2] * axis[2];
mindp = (dp < mindp) ? dp : mindp;
}
// fill bounding sphere info; note that below we can return bounds without cone information for degenerate cones
bounds.center[0] = center[0];
bounds.center[1] = center[1];
bounds.center[2] = center[2];
bounds.radius = psphere[3];
// degenerate cluster, normal cone is larger than a hemisphere => trivial accept
// note that if mindp is positive but close to 0, the triangle intersection code below gets less stable
// we arbitrarily decide that if a normal cone is ~168 degrees wide or more, the cone isn't useful
if (mindp <= 0.1f)
{
bounds.cone_cutoff = 1;
bounds.cone_cutoff_s8 = 127;
return bounds;
}
float maxt = 0;
// we need to find the point on center-t*axis ray that lies in negative half-space of all triangles
for (size_t i = 0; i < triangles; ++i)
{
// dot(center-t*axis-corner, trinormal) = 0
// dot(center-corner, trinormal) - t * dot(axis, trinormal) = 0
float cx = center[0] - corners[i][0][0];
float cy = center[1] - corners[i][0][1];
float cz = center[2] - corners[i][0][2];
float dc = cx * normals[i][0] + cy * normals[i][1] + cz * normals[i][2];
float dn = axis[0] * normals[i][0] + axis[1] * normals[i][1] + axis[2] * normals[i][2];
// dn should be larger than mindp cutoff above
assert(dn > 0.f);
float t = dc / dn;
maxt = (t > maxt) ? t : maxt;
}
// cone apex should be in the negative half-space of all cluster triangles by construction
bounds.cone_apex[0] = center[0] - axis[0] * maxt;
bounds.cone_apex[1] = center[1] - axis[1] * maxt;
bounds.cone_apex[2] = center[2] - axis[2] * maxt;
// note: this axis is the axis of the normal cone, but our test for perspective camera effectively negates the axis
bounds.cone_axis[0] = axis[0];
bounds.cone_axis[1] = axis[1];
bounds.cone_axis[2] = axis[2];
// cos(a) for normal cone is mindp; we need to add 90 degrees on both sides and invert the cone
// which gives us -cos(a+90) = -(-sin(a)) = sin(a) = sqrt(1 - cos^2(a))
bounds.cone_cutoff = sqrtf(1 - mindp * mindp);
// quantize axis & cutoff to 8-bit SNORM format
bounds.cone_axis_s8[0] = (signed char)(meshopt_quantizeSnorm(bounds.cone_axis[0], 8));
bounds.cone_axis_s8[1] = (signed char)(meshopt_quantizeSnorm(bounds.cone_axis[1], 8));
bounds.cone_axis_s8[2] = (signed char)(meshopt_quantizeSnorm(bounds.cone_axis[2], 8));
// for the 8-bit test to be conservative, we need to adjust the cutoff by measuring the max. error
float cone_axis_s8_e0 = fabsf(bounds.cone_axis_s8[0] / 127.f - bounds.cone_axis[0]);
float cone_axis_s8_e1 = fabsf(bounds.cone_axis_s8[1] / 127.f - bounds.cone_axis[1]);
float cone_axis_s8_e2 = fabsf(bounds.cone_axis_s8[2] / 127.f - bounds.cone_axis[2]);
// note that we need to round this up instead of rounding to nearest, hence +1
int cone_cutoff_s8 = int(127 * (bounds.cone_cutoff + cone_axis_s8_e0 + cone_axis_s8_e1 + cone_axis_s8_e2) + 1);
bounds.cone_cutoff_s8 = (cone_cutoff_s8 > 127) ? 127 : (signed char)(cone_cutoff_s8);
return bounds;
}
meshopt_Bounds meshopt_computeMeshletBounds(const meshopt_Meshlet* meshlet, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
{
assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
assert(vertex_positions_stride % sizeof(float) == 0);
unsigned int indices[sizeof(meshlet->indices) / sizeof(meshlet->indices[0][0])];
for (size_t i = 0; i < meshlet->triangle_count; ++i)
{
unsigned int a = meshlet->vertices[meshlet->indices[i][0]];
unsigned int b = meshlet->vertices[meshlet->indices[i][1]];
unsigned int c = meshlet->vertices[meshlet->indices[i][2]];
assert(a < vertex_count && b < vertex_count && c < vertex_count);
indices[i * 3 + 0] = a;
indices[i * 3 + 1] = b;
indices[i * 3 + 2] = c;
}
return meshopt_computeClusterBounds(indices, meshlet->triangle_count * 3, vertex_positions, vertex_count, vertex_positions_stride);
}

144
include/meshoptimizer/extern/basisu_format.h vendored Normal file
View File

@ -0,0 +1,144 @@
// basis_file_headers.h + basisu.h
// Copyright (C) 2019-2020 Binomial LLC. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
namespace basisu
{
// Always little endian 2-4 byte unsigned int
template<uint32_t NumBytes>
struct packed_uint
{
uint8_t m_bytes[NumBytes];
operator uint32_t() const
{
uint32_t result = 0;
for (uint32_t i = 0; i < NumBytes; i++)
result |= m_bytes[i] << (8 * i);
return result;
}
};
}
namespace basist
{
// Slice desc header flags
enum basis_slice_desc_flags
{
cSliceDescFlagsHasAlpha = 1,
cSliceDescFlagsFrameIsIFrame = 2 // Video only: Frame doesn't refer to previous frame (no usage of conditional replenishment pred symbols)
};
#pragma pack(push)
#pragma pack(1)
struct basis_slice_desc
{
basisu::packed_uint<3> m_image_index; // The index of the source image provided to the encoder (will always appear in order from first to last, first image index is 0, no skipping allowed)
basisu::packed_uint<1> m_level_index; // The mipmap level index (mipmaps will always appear from largest to smallest)
basisu::packed_uint<1> m_flags; // enum basis_slice_desc_flags
basisu::packed_uint<2> m_orig_width; // The original image width (may not be a multiple of 4 pixels)
basisu::packed_uint<2> m_orig_height; // The original image height (may not be a multiple of 4 pixels)
basisu::packed_uint<2> m_num_blocks_x; // The slice's block X dimensions. Each block is 4x4 pixels. The slice's pixel resolution may or may not be a power of 2.
basisu::packed_uint<2> m_num_blocks_y; // The slice's block Y dimensions.
basisu::packed_uint<4> m_file_ofs; // Offset from the header to the start of the slice's data
basisu::packed_uint<4> m_file_size; // The size of the compressed slice data in bytes
basisu::packed_uint<2> m_slice_data_crc16; // The CRC16 of the compressed slice data, for extra-paranoid use cases
};
// File header files
enum basis_header_flags
{
cBASISHeaderFlagETC1S = 1, // Always set for ETC1S files. Not set for UASTC files.
cBASISHeaderFlagYFlipped = 2, // Set if the texture had to be Y flipped before encoding
cBASISHeaderFlagHasAlphaSlices = 4 // True if any slices contain alpha (for ETC1S, if the odd slices contain alpha data)
};
// The image type field attempts to describe how to interpret the image data in a Basis file.
// The encoder library doesn't really do anything special or different with these texture types, this is mostly here for the benefit of the user.
// We do make sure the various constraints are followed (2DArray/cubemap/videoframes/volume implies that each image has the same resolution and # of mipmap levels, etc., cubemap implies that the # of image slices is a multiple of 6)
enum basis_texture_type
{
cBASISTexType2D = 0, // An arbitrary array of 2D RGB or RGBA images with optional mipmaps, array size = # images, each image may have a different resolution and # of mipmap levels
cBASISTexType2DArray = 1, // An array of 2D RGB or RGBA images with optional mipmaps, array size = # images, each image has the same resolution and mipmap levels
cBASISTexTypeCubemapArray = 2, // an array of cubemap levels, total # of images must be divisable by 6, in X+, X-, Y+, Y-, Z+, Z- order, with optional mipmaps
cBASISTexTypeVideoFrames = 3, // An array of 2D video frames, with optional mipmaps, # frames = # images, each image has the same resolution and # of mipmap levels
cBASISTexTypeVolume = 4, // A 3D texture with optional mipmaps, Z dimension = # images, each image has the same resolution and # of mipmap levels
cBASISTexTypeTotal
};
enum
{
cBASISMaxUSPerFrame = 0xFFFFFF
};
enum basis_tex_format
{
cETC1S = 0,
cUASTC4x4 = 1
};
struct basis_file_header
{
enum
{
cBASISSigValue = ('B' << 8) | 's',
cBASISFirstVersion = 0x10
};
basisu::packed_uint<2> m_sig; // 2 byte file signature
basisu::packed_uint<2> m_ver; // Baseline file version
basisu::packed_uint<2> m_header_size; // Header size in bytes, sizeof(basis_file_header)
basisu::packed_uint<2> m_header_crc16; // crc16 of the remaining header data
basisu::packed_uint<4> m_data_size; // The total size of all data after the header
basisu::packed_uint<2> m_data_crc16; // The CRC16 of all data after the header
basisu::packed_uint<3> m_total_slices; // The total # of compressed slices (1 slice per image, or 2 for alpha basis files)
basisu::packed_uint<3> m_total_images; // The total # of images
basisu::packed_uint<1> m_tex_format; // enum basis_tex_format
basisu::packed_uint<2> m_flags; // enum basist::header_flags
basisu::packed_uint<1> m_tex_type; // enum basist::basis_texture_type
basisu::packed_uint<3> m_us_per_frame; // Framerate of video, in microseconds per frame
basisu::packed_uint<4> m_reserved; // For future use
basisu::packed_uint<4> m_userdata0; // For client use
basisu::packed_uint<4> m_userdata1; // For client use
basisu::packed_uint<2> m_total_endpoints; // The number of endpoints in the endpoint codebook
basisu::packed_uint<4> m_endpoint_cb_file_ofs; // The compressed endpoint codebook's file offset relative to the header
basisu::packed_uint<3> m_endpoint_cb_file_size; // The compressed endpoint codebook's size in bytes
basisu::packed_uint<2> m_total_selectors; // The number of selectors in the endpoint codebook
basisu::packed_uint<4> m_selector_cb_file_ofs; // The compressed selectors codebook's file offset relative to the header
basisu::packed_uint<3> m_selector_cb_file_size; // The compressed selector codebook's size in bytes
basisu::packed_uint<4> m_tables_file_ofs; // The file offset of the compressed Huffman codelength tables, for decompressing slices
basisu::packed_uint<4> m_tables_file_size; // The file size in bytes of the compressed huffman codelength tables
basisu::packed_uint<4> m_slice_desc_file_ofs; // The file offset to the slice description array, usually follows the header
basisu::packed_uint<4> m_extended_file_ofs; // The file offset of the "extended" header and compressed data, for future use
basisu::packed_uint<4> m_extended_file_size; // The file size in bytes of the "extended" header and compressed data, for future use
};
#pragma pack (pop)
} // namespace basist

5205
include/meshoptimizer/extern/cgltf.h vendored Normal file
View File

File diff suppressed because it is too large Load Diff

1400
include/meshoptimizer/extern/fast_obj.h vendored Normal file
View File

File diff suppressed because it is too large Load Diff

629
include/meshoptimizer/extern/khr_df.h vendored Normal file
View File

@ -0,0 +1,629 @@
/* The Khronos Data Format Specification (version 1.3) */
/*
** Copyright (c) 2015-19 The Khronos Group Inc.
**
** Permission is hereby granted, free of charge, to any person obtaining a
** copy of this software and/or associated documentation files (the
** "Materials"), to deal in the Materials without restriction, including
** without limitation the rights to use, copy, modify, merge, publish,
** distribute, sublicense, and/or sell copies of the Materials, and to
** permit persons to whom the Materials are furnished to do so, subject to
** the following conditions:
**
** The above copyright notice and this permission notice shall be included
** in all copies or substantial portions of the Materials.
**
** THE MATERIALS ARE PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
** EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
** MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
** IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
** CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
** TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
** MATERIALS OR THE USE OR OTHER DEALINGS IN THE MATERIALS.
*/
/* This header defines a structure that can describe the layout of image
formats in memory. This means that the data format is transparent to
the application, and the expectation is that this should be used when
the layout is defined external to the API. Many Khronos APIs deliberately
keep the internal layout of images opaque, to allow proprietary layouts
and optimisations. This structure is not appropriate for describing
opaque layouts. */
/* We stick to standard C89 constructs for simplicity and portability. */
#ifndef _KHR_DATA_FORMAT_H_
#define _KHR_DATA_FORMAT_H_
/* Accessors */
typedef enum _khr_word_e {
KHR_DF_WORD_VENDORID = 0U,
KHR_DF_WORD_DESCRIPTORTYPE = 0U,
KHR_DF_WORD_VERSIONNUMBER = 1U,
KHR_DF_WORD_DESCRIPTORBLOCKSIZE = 1U,
KHR_DF_WORD_MODEL = 2U,
KHR_DF_WORD_PRIMARIES = 2U,
KHR_DF_WORD_TRANSFER = 2U,
KHR_DF_WORD_FLAGS = 2U,
KHR_DF_WORD_TEXELBLOCKDIMENSION0 = 3U,
KHR_DF_WORD_TEXELBLOCKDIMENSION1 = 3U,
KHR_DF_WORD_TEXELBLOCKDIMENSION2 = 3U,
KHR_DF_WORD_TEXELBLOCKDIMENSION3 = 3U,
KHR_DF_WORD_BYTESPLANE0 = 4U,
KHR_DF_WORD_BYTESPLANE1 = 4U,
KHR_DF_WORD_BYTESPLANE2 = 4U,
KHR_DF_WORD_BYTESPLANE3 = 4U,
KHR_DF_WORD_BYTESPLANE4 = 5U,
KHR_DF_WORD_BYTESPLANE5 = 5U,
KHR_DF_WORD_BYTESPLANE6 = 5U,
KHR_DF_WORD_BYTESPLANE7 = 5U,
KHR_DF_WORD_SAMPLESTART = 6U,
KHR_DF_WORD_SAMPLEWORDS = 4U
} khr_df_word_e;
typedef enum _khr_df_shift_e {
KHR_DF_SHIFT_VENDORID = 0U,
KHR_DF_SHIFT_DESCRIPTORTYPE = 17U,
KHR_DF_SHIFT_VERSIONNUMBER = 0U,
KHR_DF_SHIFT_DESCRIPTORBLOCKSIZE = 16U,
KHR_DF_SHIFT_MODEL = 0U,
KHR_DF_SHIFT_PRIMARIES = 8U,
KHR_DF_SHIFT_TRANSFER = 16U,
KHR_DF_SHIFT_FLAGS = 24U,
KHR_DF_SHIFT_TEXELBLOCKDIMENSION0 = 0U,
KHR_DF_SHIFT_TEXELBLOCKDIMENSION1 = 8U,
KHR_DF_SHIFT_TEXELBLOCKDIMENSION2 = 16U,
KHR_DF_SHIFT_TEXELBLOCKDIMENSION3 = 24U,
KHR_DF_SHIFT_BYTESPLANE0 = 0U,
KHR_DF_SHIFT_BYTESPLANE1 = 8U,
KHR_DF_SHIFT_BYTESPLANE2 = 16U,
KHR_DF_SHIFT_BYTESPLANE3 = 24U,
KHR_DF_SHIFT_BYTESPLANE4 = 0U,
KHR_DF_SHIFT_BYTESPLANE5 = 8U,
KHR_DF_SHIFT_BYTESPLANE6 = 16U,
KHR_DF_SHIFT_BYTESPLANE7 = 24U
} khr_df_shift_e;
typedef enum _khr_df_mask_e {
KHR_DF_MASK_VENDORID = 0x1FFFFU,
KHR_DF_MASK_DESCRIPTORTYPE = 0x7FFFU,
KHR_DF_MASK_VERSIONNUMBER = 0xFFFFU,
KHR_DF_MASK_DESCRIPTORBLOCKSIZE = 0xFFFFU,
KHR_DF_MASK_MODEL = 0xFFU,
KHR_DF_MASK_PRIMARIES = 0xFFU,
KHR_DF_MASK_TRANSFER = 0xFFU,
KHR_DF_MASK_FLAGS = 0xFFU,
KHR_DF_MASK_TEXELBLOCKDIMENSION0 = 0xFFU,
KHR_DF_MASK_TEXELBLOCKDIMENSION1 = 0xFFU,
KHR_DF_MASK_TEXELBLOCKDIMENSION2 = 0xFFU,
KHR_DF_MASK_TEXELBLOCKDIMENSION3 = 0xFFU,
KHR_DF_MASK_BYTESPLANE0 = 0xFFU,
KHR_DF_MASK_BYTESPLANE1 = 0xFFU,
KHR_DF_MASK_BYTESPLANE2 = 0xFFU,
KHR_DF_MASK_BYTESPLANE3 = 0xFFU,
KHR_DF_MASK_BYTESPLANE4 = 0xFFU,
KHR_DF_MASK_BYTESPLANE5 = 0xFFU,
KHR_DF_MASK_BYTESPLANE6 = 0xFFU,
KHR_DF_MASK_BYTESPLANE7 = 0xFFU
} khr_df_mask_e;
/* Helper macro:
Extract field X from basic descriptor block BDB */
#define KHR_DFDVAL(BDB, X) \
(((BDB)[KHR_DF_WORD_ ## X] >> (KHR_DF_SHIFT_ ## X)) \
& (KHR_DF_MASK_ ## X))
/* Helper macro:
Set field X of basic descriptor block BDB */
#define KHR_DFDSETVAL(BDB, X, val) \
((BDB)[KHR_DF_WORD_ ## X] = \
((BDB)[KHR_DF_WORD_ ## X] & \
~((KHR_DF_MASK_ ## X) << (KHR_DF_SHIFT_ ## X))) | \
(((val) & (KHR_DF_MASK_ ## X)) << (KHR_DF_SHIFT_ ## X)))
/* Offsets relative to the start of a sample */
typedef enum _khr_df_sampleword_e {
KHR_DF_SAMPLEWORD_BITOFFSET = 0U,
KHR_DF_SAMPLEWORD_BITLENGTH = 0U,
KHR_DF_SAMPLEWORD_CHANNELID = 0U,
KHR_DF_SAMPLEWORD_QUALIFIERS = 0U,
KHR_DF_SAMPLEWORD_SAMPLEPOSITION0 = 1U,
KHR_DF_SAMPLEWORD_SAMPLEPOSITION1 = 1U,
KHR_DF_SAMPLEWORD_SAMPLEPOSITION2 = 1U,
KHR_DF_SAMPLEWORD_SAMPLEPOSITION3 = 1U,
KHR_DF_SAMPLEWORD_SAMPLEPOSITION_ALL = 1U,
KHR_DF_SAMPLEWORD_SAMPLELOWER = 2U,
KHR_DF_SAMPLEWORD_SAMPLEUPPER = 3U
} khr_df_sampleword_e;
typedef enum _khr_df_sampleshift_e {
KHR_DF_SAMPLESHIFT_BITOFFSET = 0U,
KHR_DF_SAMPLESHIFT_BITLENGTH = 16U,
KHR_DF_SAMPLESHIFT_CHANNELID = 24U,
/* N.B. Qualifiers are defined as an offset into a byte */
KHR_DF_SAMPLESHIFT_QUALIFIERS = 24U,
KHR_DF_SAMPLESHIFT_SAMPLEPOSITION0 = 0U,
KHR_DF_SAMPLESHIFT_SAMPLEPOSITION1 = 8U,
KHR_DF_SAMPLESHIFT_SAMPLEPOSITION2 = 16U,
KHR_DF_SAMPLESHIFT_SAMPLEPOSITION3 = 24U,
KHR_DF_SAMPLESHIFT_SAMPLEPOSITION_ALL = 0U,
KHR_DF_SAMPLESHIFT_SAMPLELOWER = 0U,
KHR_DF_SAMPLESHIFT_SAMPLEUPPER = 0U
} khr_df_sampleshift_e;
typedef enum _khr_df_samplemask_e {
KHR_DF_SAMPLEMASK_BITOFFSET = 0xFFFFU,
KHR_DF_SAMPLEMASK_BITLENGTH = 0xFF,
KHR_DF_SAMPLEMASK_CHANNELID = 0xF,
/* N.B. Qualifiers are defined as an offset into a byte */
KHR_DF_SAMPLEMASK_QUALIFIERS = 0xF0,
KHR_DF_SAMPLEMASK_SAMPLEPOSITION0 = 0xFF,
KHR_DF_SAMPLEMASK_SAMPLEPOSITION1 = 0xFF,
KHR_DF_SAMPLEMASK_SAMPLEPOSITION2 = 0xFF,
KHR_DF_SAMPLEMASK_SAMPLEPOSITION3 = 0xFF,
/* ISO C restricts enum values to range of int hence the
cast. We do it verbosely instead of using -1 to ensure
it is a 32-bit value even if int is 64 bits. */
KHR_DF_SAMPLEMASK_SAMPLEPOSITION_ALL = (int) 0xFFFFFFFFU,
KHR_DF_SAMPLEMASK_SAMPLELOWER = (int) 0xFFFFFFFFU,
KHR_DF_SAMPLEMASK_SAMPLEUPPER = (int) 0xFFFFFFFFU
} khr_df_samplemask_e;
/* Helper macro:
Extract field X of sample S from basic descriptor block BDB */
#define KHR_DFDSVAL(BDB, S, X) \
(((BDB)[KHR_DF_WORD_SAMPLESTART + \
((S) * KHR_DF_WORD_SAMPLEWORDS) + \
KHR_DF_SAMPLEWORD_ ## X] >> (KHR_DF_SAMPLESHIFT_ ## X)) \
& (KHR_DF_SAMPLEMASK_ ## X))
/* Helper macro:
Set field X of sample S of basic descriptor block BDB */
#define KHR_DFDSETSVAL(BDB, S, X, val) \
((BDB)[KHR_DF_WORD_SAMPLESTART + \
((S) * KHR_DF_WORD_SAMPLEWORDS) + \
KHR_DF_SAMPLEWORD_ ## X] = \
((BDB)[KHR_DF_WORD_SAMPLESTART + \
((S) * KHR_DF_WORD_SAMPLEWORDS) + \
KHR_DF_SAMPLEWORD_ ## X] & \
~((uint32_t)(KHR_DF_SAMPLEMASK_ ## X) << (KHR_DF_SAMPLESHIFT_ ## X))) | \
(((val) & (uint32_t)(KHR_DF_SAMPLEMASK_ ## X)) << (KHR_DF_SAMPLESHIFT_ ## X)))
/* Helper macro:
Number of samples in basic descriptor block BDB */
#define KHR_DFDSAMPLECOUNT(BDB) \
(((KHR_DFDVAL(BDB, DESCRIPTORBLOCKSIZE) >> 2) - \
KHR_DF_WORD_SAMPLESTART) \
/ KHR_DF_WORD_SAMPLEWORDS)
/* Helper macro:
Size in words of basic descriptor block for S samples */
#define KHR_DFDSIZEWORDS(S) \
(KHR_DF_WORD_SAMPLESTART + \
(S) * KHR_DF_WORD_SAMPLEWORDS)
/* Vendor ids */
typedef enum _khr_df_vendorid_e {
/* Standard Khronos descriptor */
KHR_DF_VENDORID_KHRONOS = 0U,
KHR_DF_VENDORID_MAX = 0x1FFFFU
} khr_df_vendorid_e;
/* Descriptor types */
typedef enum _khr_df_khr_descriptortype_e {
/* Default Khronos basic descriptor block */
KHR_DF_KHR_DESCRIPTORTYPE_BASICFORMAT = 0U,
/* Extension descriptor block for additional planes */
KHR_DF_KHR_DESCRIPTORTYPE_ADDITIONAL_PLANES = 0x6001U,
/* Extension descriptor block for additional dimensions */
KHR_DF_KHR_DESCRIPTORTYPE_ADDITIONAL_DIMENSIONS = 0x6002U,
/* Bit indicates modifying requires understanding this extension */
KHR_DF_KHR_DESCRIPTORTYPE_NEEDED_FOR_WRITE_BIT = 0x2000U,
/* Bit indicates processing requires understanding this extension */
KHR_DF_KHR_DESCRIPTORTYPE_NEEDED_FOR_DECODE_BIT = 0x4000U,
KHR_DF_KHR_DESCRIPTORTYPE_MAX = 0x7FFFU
} khr_df_khr_descriptortype_e;
/* Descriptor block version */
typedef enum _khr_df_versionnumber_e {
/* Standard Khronos descriptor */
KHR_DF_VERSIONNUMBER_1_0 = 0U, /* Version 1.0 of the specification */
KHR_DF_VERSIONNUMBER_1_1 = 0U, /* Version 1.1 did not bump the version number */
KHR_DF_VERSIONNUMBER_1_2 = 1U, /* Version 1.2 increased the version number */
KHR_DF_VERSIONNUMBER_1_3 = 2U, /* Version 1.3 increased the version number */
KHR_DF_VERSIONNUMBER_LATEST = KHR_DF_VERSIONNUMBER_1_3,
KHR_DF_VERSIONNUMBER_MAX = 0xFFFFU
} khr_df_versionnumber_e;
/* Model in which the color coordinate space is defined.
There is no requirement that a color format use all the
channel types that are defined in the color model. */
typedef enum _khr_df_model_e {
/* No interpretation of color channels defined */
KHR_DF_MODEL_UNSPECIFIED = 0U,
/* Color primaries (red, green, blue) + alpha, depth and stencil */
KHR_DF_MODEL_RGBSDA = 1U,
/* Color differences (Y', Cb, Cr) + alpha, depth and stencil */
KHR_DF_MODEL_YUVSDA = 2U,
/* Color differences (Y', I, Q) + alpha, depth and stencil */
KHR_DF_MODEL_YIQSDA = 3U,
/* Perceptual color (CIE L*a*b*) + alpha, depth and stencil */
KHR_DF_MODEL_LABSDA = 4U,
/* Subtractive colors (cyan, magenta, yellow, black) + alpha */
KHR_DF_MODEL_CMYKA = 5U,
/* Non-color coordinate data (X, Y, Z, W) */
KHR_DF_MODEL_XYZW = 6U,
/* Hue, saturation, value, hue angle on color circle, plus alpha */
KHR_DF_MODEL_HSVA_ANG = 7U,
/* Hue, saturation, lightness, hue angle on color circle, plus alpha */
KHR_DF_MODEL_HSLA_ANG = 8U,
/* Hue, saturation, value, hue on color hexagon, plus alpha */
KHR_DF_MODEL_HSVA_HEX = 9U,
/* Hue, saturation, lightness, hue on color hexagon, plus alpha */
KHR_DF_MODEL_HSLA_HEX = 10U,
/* Lightweight approximate color difference (luma, orange, green) */
KHR_DF_MODEL_YCGCOA = 11U,
/* ITU BT.2020 constant luminance YcCbcCrc */
KHR_DF_MODEL_YCCBCCRC = 12U,
/* ITU BT.2100 constant intensity ICtCp */
KHR_DF_MODEL_ICTCP = 13U,
/* CIE 1931 XYZ color coordinates (X, Y, Z) */
KHR_DF_MODEL_CIEXYZ = 14U,
/* CIE 1931 xyY color coordinates (X, Y, Y) */
KHR_DF_MODEL_CIEXYY = 15U,
/* Compressed formats start at 128. */
/* These compressed formats should generally have a single sample,
sited at the 0,0 position of the texel block. Where multiple
channels are used to distinguish formats, these should be cosited. */
/* Direct3D (and S3) compressed formats */
/* Note that premultiplied status is recorded separately */
/* DXT1 "channels" are RGB (0), Alpha (1) */
/* DXT1/BC1 with one channel is opaque */
/* DXT1/BC1 with a cosited alpha sample is transparent */
KHR_DF_MODEL_DXT1A = 128U,
KHR_DF_MODEL_BC1A = 128U,
/* DXT2/DXT3/BC2, with explicit 4-bit alpha */
KHR_DF_MODEL_DXT2 = 129U,
KHR_DF_MODEL_DXT3 = 129U,
KHR_DF_MODEL_BC2 = 129U,
/* DXT4/DXT5/BC3, with interpolated alpha */
KHR_DF_MODEL_DXT4 = 130U,
KHR_DF_MODEL_DXT5 = 130U,
KHR_DF_MODEL_BC3 = 130U,
/* BC4 - single channel interpolated 8-bit data */
/* (The UNORM/SNORM variation is recorded in the channel data) */
KHR_DF_MODEL_BC4 = 131U,
/* BC5 - two channel interpolated 8-bit data */
/* (The UNORM/SNORM variation is recorded in the channel data) */
KHR_DF_MODEL_BC5 = 132U,
/* BC6H - DX11 format for 16-bit float channels */
KHR_DF_MODEL_BC6H = 133U,
/* BC7 - DX11 format */
KHR_DF_MODEL_BC7 = 134U,
/* Gap left for future desktop expansion */
/* Mobile compressed formats follow */
/* A format of ETC1 indicates that the format shall be decodable
by an ETC1-compliant decoder and not rely on ETC2 features */
KHR_DF_MODEL_ETC1 = 160U,
/* A format of ETC2 is permitted to use ETC2 encodings on top of
the baseline ETC1 specification */
/* The ETC2 format has channels "red", "green", "RGB" and "alpha",
which should be cosited samples */
/* Punch-through alpha can be distinguished from full alpha by
the plane size in bytes required for the texel block */
KHR_DF_MODEL_ETC2 = 161U,
/* Adaptive Scalable Texture Compression */
/* ASTC HDR vs LDR is determined by the float flag in the channel */
/* ASTC block size can be distinguished by texel block size */
KHR_DF_MODEL_ASTC = 162U,
/* ETC1S is a simplified subset of ETC1 */
KHR_DF_MODEL_ETC1S = 163U,
/* PowerVR Texture Compression */
KHR_DF_MODEL_PVRTC = 164U,
KHR_DF_MODEL_PVRTC2 = 165U,
/* Universal Adaptive Scalable Texture Compression */
KHR_DF_MODEL_UASTC = 166U,
/* Proprietary formats (ATITC, etc.) should follow */
KHR_DF_MODEL_MAX = 0xFFU
} khr_df_model_e;
/* Definition of channel names for each color model */
typedef enum _khr_df_model_channels_e {
/* Unspecified format with nominal channel numbering */
KHR_DF_CHANNEL_UNSPECIFIED_0 = 0U,
KHR_DF_CHANNEL_UNSPECIFIED_1 = 1U,
KHR_DF_CHANNEL_UNSPECIFIED_2 = 2U,
KHR_DF_CHANNEL_UNSPECIFIED_3 = 3U,
KHR_DF_CHANNEL_UNSPECIFIED_4 = 4U,
KHR_DF_CHANNEL_UNSPECIFIED_5 = 5U,
KHR_DF_CHANNEL_UNSPECIFIED_6 = 6U,
KHR_DF_CHANNEL_UNSPECIFIED_7 = 7U,
KHR_DF_CHANNEL_UNSPECIFIED_8 = 8U,
KHR_DF_CHANNEL_UNSPECIFIED_9 = 9U,
KHR_DF_CHANNEL_UNSPECIFIED_10 = 10U,
KHR_DF_CHANNEL_UNSPECIFIED_11 = 11U,
KHR_DF_CHANNEL_UNSPECIFIED_12 = 12U,
KHR_DF_CHANNEL_UNSPECIFIED_13 = 13U,
KHR_DF_CHANNEL_UNSPECIFIED_14 = 14U,
KHR_DF_CHANNEL_UNSPECIFIED_15 = 15U,
/* MODEL_RGBSDA - red, green, blue, stencil, depth, alpha */
KHR_DF_CHANNEL_RGBSDA_RED = 0U,
KHR_DF_CHANNEL_RGBSDA_R = 0U,
KHR_DF_CHANNEL_RGBSDA_GREEN = 1U,
KHR_DF_CHANNEL_RGBSDA_G = 1U,
KHR_DF_CHANNEL_RGBSDA_BLUE = 2U,
KHR_DF_CHANNEL_RGBSDA_B = 2U,
KHR_DF_CHANNEL_RGBSDA_STENCIL = 13U,
KHR_DF_CHANNEL_RGBSDA_S = 13U,
KHR_DF_CHANNEL_RGBSDA_DEPTH = 14U,
KHR_DF_CHANNEL_RGBSDA_D = 14U,
KHR_DF_CHANNEL_RGBSDA_ALPHA = 15U,
KHR_DF_CHANNEL_RGBSDA_A = 15U,
/* MODEL_YUVSDA - luma, Cb, Cr, stencil, depth, alpha */
KHR_DF_CHANNEL_YUVSDA_Y = 0U,
KHR_DF_CHANNEL_YUVSDA_CB = 1U,
KHR_DF_CHANNEL_YUVSDA_U = 1U,
KHR_DF_CHANNEL_YUVSDA_CR = 2U,
KHR_DF_CHANNEL_YUVSDA_V = 2U,
KHR_DF_CHANNEL_YUVSDA_STENCIL = 13U,
KHR_DF_CHANNEL_YUVSDA_S = 13U,
KHR_DF_CHANNEL_YUVSDA_DEPTH = 14U,
KHR_DF_CHANNEL_YUVSDA_D = 14U,
KHR_DF_CHANNEL_YUVSDA_ALPHA = 15U,
KHR_DF_CHANNEL_YUVSDA_A = 15U,
/* MODEL_YIQSDA - luma, in-phase, quadrature, stencil, depth, alpha */
KHR_DF_CHANNEL_YIQSDA_Y = 0U,
KHR_DF_CHANNEL_YIQSDA_I = 1U,
KHR_DF_CHANNEL_YIQSDA_Q = 2U,
KHR_DF_CHANNEL_YIQSDA_STENCIL = 13U,
KHR_DF_CHANNEL_YIQSDA_S = 13U,
KHR_DF_CHANNEL_YIQSDA_DEPTH = 14U,
KHR_DF_CHANNEL_YIQSDA_D = 14U,
KHR_DF_CHANNEL_YIQSDA_ALPHA = 15U,
KHR_DF_CHANNEL_YIQSDA_A = 15U,
/* MODEL_LABSDA - CIELAB/L*a*b* luma, red-green, blue-yellow, stencil, depth, alpha */
KHR_DF_CHANNEL_LABSDA_L = 0U,
KHR_DF_CHANNEL_LABSDA_A = 1U,
KHR_DF_CHANNEL_LABSDA_B = 2U,
KHR_DF_CHANNEL_LABSDA_STENCIL = 13U,
KHR_DF_CHANNEL_LABSDA_S = 13U,
KHR_DF_CHANNEL_LABSDA_DEPTH = 14U,
KHR_DF_CHANNEL_LABSDA_D = 14U,
KHR_DF_CHANNEL_LABSDA_ALPHA = 15U,
/* NOTE: KHR_DF_CHANNEL_LABSDA_A is not a synonym for alpha! */
/* MODEL_CMYKA - cyan, magenta, yellow, key/blacK, alpha */
KHR_DF_CHANNEL_CMYKSDA_CYAN = 0U,
KHR_DF_CHANNEL_CMYKSDA_C = 0U,
KHR_DF_CHANNEL_CMYKSDA_MAGENTA = 1U,
KHR_DF_CHANNEL_CMYKSDA_M = 1U,
KHR_DF_CHANNEL_CMYKSDA_YELLOW = 2U,
KHR_DF_CHANNEL_CMYKSDA_Y = 2U,
KHR_DF_CHANNEL_CMYKSDA_KEY = 3U,
KHR_DF_CHANNEL_CMYKSDA_BLACK = 3U,
KHR_DF_CHANNEL_CMYKSDA_K = 3U,
KHR_DF_CHANNEL_CMYKSDA_ALPHA = 15U,
KHR_DF_CHANNEL_CMYKSDA_A = 15U,
/* MODEL_XYZW - coordinates x, y, z, w */
KHR_DF_CHANNEL_XYZW_X = 0U,
KHR_DF_CHANNEL_XYZW_Y = 1U,
KHR_DF_CHANNEL_XYZW_Z = 2U,
KHR_DF_CHANNEL_XYZW_W = 3U,
/* MODEL_HSVA_ANG - value (luma), saturation, hue, alpha, angular projection, conical space */
KHR_DF_CHANNEL_HSVA_ANG_VALUE = 0U,
KHR_DF_CHANNEL_HSVA_ANG_V = 0U,
KHR_DF_CHANNEL_HSVA_ANG_SATURATION = 1U,
KHR_DF_CHANNEL_HSVA_ANG_S = 1U,
KHR_DF_CHANNEL_HSVA_ANG_HUE = 2U,
KHR_DF_CHANNEL_HSVA_ANG_H = 2U,
KHR_DF_CHANNEL_HSVA_ANG_ALPHA = 15U,
KHR_DF_CHANNEL_HSVA_ANG_A = 15U,
/* MODEL_HSLA_ANG - lightness (luma), saturation, hue, alpha, angular projection, double conical space */
KHR_DF_CHANNEL_HSLA_ANG_LIGHTNESS = 0U,
KHR_DF_CHANNEL_HSLA_ANG_L = 0U,
KHR_DF_CHANNEL_HSLA_ANG_SATURATION = 1U,
KHR_DF_CHANNEL_HSLA_ANG_S = 1U,
KHR_DF_CHANNEL_HSLA_ANG_HUE = 2U,
KHR_DF_CHANNEL_HSLA_ANG_H = 2U,
KHR_DF_CHANNEL_HSLA_ANG_ALPHA = 15U,
KHR_DF_CHANNEL_HSLA_ANG_A = 15U,
/* MODEL_HSVA_HEX - value (luma), saturation, hue, alpha, hexagonal projection, conical space */
KHR_DF_CHANNEL_HSVA_HEX_VALUE = 0U,
KHR_DF_CHANNEL_HSVA_HEX_V = 0U,
KHR_DF_CHANNEL_HSVA_HEX_SATURATION = 1U,
KHR_DF_CHANNEL_HSVA_HEX_S = 1U,
KHR_DF_CHANNEL_HSVA_HEX_HUE = 2U,
KHR_DF_CHANNEL_HSVA_HEX_H = 2U,
KHR_DF_CHANNEL_HSVA_HEX_ALPHA = 15U,
KHR_DF_CHANNEL_HSVA_HEX_A = 15U,
/* MODEL_HSLA_HEX - lightness (luma), saturation, hue, alpha, hexagonal projection, double conical space */
KHR_DF_CHANNEL_HSLA_HEX_LIGHTNESS = 0U,
KHR_DF_CHANNEL_HSLA_HEX_L = 0U,
KHR_DF_CHANNEL_HSLA_HEX_SATURATION = 1U,
KHR_DF_CHANNEL_HSLA_HEX_S = 1U,
KHR_DF_CHANNEL_HSLA_HEX_HUE = 2U,
KHR_DF_CHANNEL_HSLA_HEX_H = 2U,
KHR_DF_CHANNEL_HSLA_HEX_ALPHA = 15U,
KHR_DF_CHANNEL_HSLA_HEX_A = 15U,
/* MODEL_YCGCOA - luma, green delta, orange delta, alpha */
KHR_DF_CHANNEL_YCGCOA_Y = 0U,
KHR_DF_CHANNEL_YCGCOA_CG = 1U,
KHR_DF_CHANNEL_YCGCOA_CO = 2U,
KHR_DF_CHANNEL_YCGCOA_ALPHA = 15U,
KHR_DF_CHANNEL_YCGCOA_A = 15U,
/* MODEL_CIEXYZ - CIE 1931 X, Y, Z */
KHR_DF_CHANNEL_CIEXYZ_X = 0U,
KHR_DF_CHANNEL_CIEXYZ_Y = 1U,
KHR_DF_CHANNEL_CIEXYZ_Z = 2U,
/* MODEL_CIEXYY - CIE 1931 x, y, Y */
KHR_DF_CHANNEL_CIEXYY_X = 0U,
KHR_DF_CHANNEL_CIEXYY_YCHROMA = 1U,
KHR_DF_CHANNEL_CIEXYY_YLUMA = 2U,
/* Compressed formats */
/* MODEL_DXT1A/MODEL_BC1A */
KHR_DF_CHANNEL_DXT1A_COLOR = 0U,
KHR_DF_CHANNEL_BC1A_COLOR = 0U,
KHR_DF_CHANNEL_DXT1A_ALPHAPRESENT = 1U,
KHR_DF_CHANNEL_DXT1A_ALPHA = 1U,
KHR_DF_CHANNEL_BC1A_ALPHAPRESENT = 1U,
KHR_DF_CHANNEL_BC1A_ALPHA = 1U,
/* MODEL_DXT2/3/MODEL_BC2 */
KHR_DF_CHANNEL_DXT2_COLOR = 0U,
KHR_DF_CHANNEL_DXT3_COLOR = 0U,
KHR_DF_CHANNEL_BC2_COLOR = 0U,
KHR_DF_CHANNEL_DXT2_ALPHA = 15U,
KHR_DF_CHANNEL_DXT3_ALPHA = 15U,
KHR_DF_CHANNEL_BC2_ALPHA = 15U,
/* MODEL_DXT4/5/MODEL_BC3 */
KHR_DF_CHANNEL_DXT4_COLOR = 0U,
KHR_DF_CHANNEL_DXT5_COLOR = 0U,
KHR_DF_CHANNEL_BC3_COLOR = 0U,
KHR_DF_CHANNEL_DXT4_ALPHA = 15U,
KHR_DF_CHANNEL_DXT5_ALPHA = 15U,
KHR_DF_CHANNEL_BC3_ALPHA = 15U,
/* MODEL_BC4 */
KHR_DF_CHANNEL_BC4_DATA = 0U,
/* MODEL_BC5 */
KHR_DF_CHANNEL_BC5_RED = 0U,
KHR_DF_CHANNEL_BC5_R = 0U,
KHR_DF_CHANNEL_BC5_GREEN = 1U,
KHR_DF_CHANNEL_BC5_G = 1U,
/* MODEL_BC6H */
KHR_DF_CHANNEL_BC6H_COLOR = 0U,
KHR_DF_CHANNEL_BC6H_DATA = 0U,
/* MODEL_BC7 */
KHR_DF_CHANNEL_BC7_DATA = 0U,
KHR_DF_CHANNEL_BC7_COLOR = 0U,
/* MODEL_ETC1 */
KHR_DF_CHANNEL_ETC1_DATA = 0U,
KHR_DF_CHANNEL_ETC1_COLOR = 0U,
/* MODEL_ETC2 */
KHR_DF_CHANNEL_ETC2_RED = 0U,
KHR_DF_CHANNEL_ETC2_R = 0U,
KHR_DF_CHANNEL_ETC2_GREEN = 1U,
KHR_DF_CHANNEL_ETC2_G = 1U,
KHR_DF_CHANNEL_ETC2_COLOR = 2U,
KHR_DF_CHANNEL_ETC2_ALPHA = 15U,
KHR_DF_CHANNEL_ETC2_A = 15U,
/* MODEL_ASTC */
KHR_DF_CHANNEL_ASTC_DATA = 0U,
/* MODEL_ETC1S */
KHR_DF_CHANNEL_ETC1S_DATA = 0U,
KHR_DF_CHANNEL_ETC1S_COLOR = 0U,
/* MODEL_PVRTC */
KHR_DF_CHANNEL_PVRTC_DATA = 0U,
KHR_DF_CHANNEL_PVRTC_COLOR = 0U,
/* MODEL_PVRTC2 */
KHR_DF_CHANNEL_PVRTC2_DATA = 0U,
KHR_DF_CHANNEL_PVRTC2_COLOR = 0U,
/* Common channel names shared by multiple formats */
KHR_DF_CHANNEL_COMMON_LUMA = 0U,
KHR_DF_CHANNEL_COMMON_L = 0U,
KHR_DF_CHANNEL_COMMON_STENCIL = 13U,
KHR_DF_CHANNEL_COMMON_S = 13U,
KHR_DF_CHANNEL_COMMON_DEPTH = 14U,
KHR_DF_CHANNEL_COMMON_D = 14U,
KHR_DF_CHANNEL_COMMON_ALPHA = 15U,
KHR_DF_CHANNEL_COMMON_A = 15U
} khr_df_model_channels_e;
/* Definition of the primary colors in color coordinates.
This is implicitly responsible for defining the conversion
between RGB an YUV color spaces.
LAB and related absolute color models should use
KHR_DF_PRIMARIES_CIEXYZ. */
typedef enum _khr_df_primaries_e {
/* No color primaries defined */
KHR_DF_PRIMARIES_UNSPECIFIED = 0U,
/* Color primaries of ITU-R BT.709 and sRGB */
KHR_DF_PRIMARIES_BT709 = 1U,
/* Synonym for KHR_DF_PRIMARIES_BT709 */
KHR_DF_PRIMARIES_SRGB = 1U,
/* Color primaries of ITU-R BT.601 (625-line EBU variant) */
KHR_DF_PRIMARIES_BT601_EBU = 2U,
/* Color primaries of ITU-R BT.601 (525-line SMPTE C variant) */
KHR_DF_PRIMARIES_BT601_SMPTE = 3U,
/* Color primaries of ITU-R BT.2020 */
KHR_DF_PRIMARIES_BT2020 = 4U,
/* CIE theoretical color coordinate space */
KHR_DF_PRIMARIES_CIEXYZ = 5U,
/* Academy Color Encoding System primaries */
KHR_DF_PRIMARIES_ACES = 6U,
/* Color primaries of ACEScc */
KHR_DF_PRIMARIES_ACESCC = 7U,
/* Legacy NTSC 1953 primaries */
KHR_DF_PRIMARIES_NTSC1953 = 8U,
/* Legacy PAL 525-line primaries */
KHR_DF_PRIMARIES_PAL525 = 9U,
/* Color primaries of Display P3 */
KHR_DF_PRIMARIES_DISPLAYP3 = 10U,
/* Color primaries of Adobe RGB (1998) */
KHR_DF_PRIMARIES_ADOBERGB = 11U,
KHR_DF_PRIMARIES_MAX = 0xFFU
} khr_df_primaries_e;
/* Definition of the optical to digital transfer function
("gamma correction"). Most transfer functions are not a pure
power function and also include a linear element.
LAB and related absolute color representations should use
KHR_DF_TRANSFER_UNSPECIFIED. */
typedef enum _khr_df_transfer_e {
/* No transfer function defined */
KHR_DF_TRANSFER_UNSPECIFIED = 0U,
/* Linear transfer function (value proportional to intensity) */
KHR_DF_TRANSFER_LINEAR = 1U,
/* Perceptually-linear transfer function of sRGH (~2.4) */
KHR_DF_TRANSFER_SRGB = 2U,
/* Perceptually-linear transfer function of ITU non-HDR specifications (~1/.45) */
KHR_DF_TRANSFER_ITU = 3U,
/* SMTPE170M (digital NTSC) defines an alias for the ITU transfer function (~1/.45) */
KHR_DF_TRANSFER_SMTPE170M = 3U,
/* Perceptually-linear gamma function of original NTSC (simple 2.2 gamma) */
KHR_DF_TRANSFER_NTSC = 4U,
/* Sony S-log used by Sony video cameras */
KHR_DF_TRANSFER_SLOG = 5U,
/* Sony S-log 2 used by Sony video cameras */
KHR_DF_TRANSFER_SLOG2 = 6U,
/* ITU BT.1886 EOTF */
KHR_DF_TRANSFER_BT1886 = 7U,
/* ITU BT.2100 HLG OETF */
KHR_DF_TRANSFER_HLG_OETF = 8U,
/* ITU BT.2100 HLG EOTF */
KHR_DF_TRANSFER_HLG_EOTF = 9U,
/* ITU BT.2100 PQ EOTF */
KHR_DF_TRANSFER_PQ_EOTF = 10U,
/* ITU BT.2100 PQ OETF */
KHR_DF_TRANSFER_PQ_OETF = 11U,
/* DCI P3 transfer function */
KHR_DF_TRANSFER_DCIP3 = 12U,
/* Legacy PAL OETF */
KHR_DF_TRANSFER_PAL_OETF = 13U,
/* Legacy PAL 625-line EOTF */
KHR_DF_TRANSFER_PAL625_EOTF = 14U,
/* Legacy ST240 transfer function */
KHR_DF_TRANSFER_ST240 = 15U,
/* ACEScc transfer function */
KHR_DF_TRANSFER_ACESCC = 16U,
/* ACEScct transfer function */
KHR_DF_TRANSFER_ACESCCT = 17U,
/* Adobe RGB (1998) transfer function */
KHR_DF_TRANSFER_ADOBERGB = 18U,
KHR_DF_TRANSFER_MAX = 0xFFU
} khr_df_transfer_e;
typedef enum _khr_df_flags_e {
KHR_DF_FLAG_ALPHA_STRAIGHT = 0U,
KHR_DF_FLAG_ALPHA_PREMULTIPLIED = 1U
} khr_df_flags_e;
typedef enum _khr_df_sample_datatype_qualifiers_e {
KHR_DF_SAMPLE_DATATYPE_LINEAR = 1U << 4U,
KHR_DF_SAMPLE_DATATYPE_EXPONENT = 1U << 5U,
KHR_DF_SAMPLE_DATATYPE_SIGNED = 1U << 6U,
KHR_DF_SAMPLE_DATATYPE_FLOAT = 1U << 7U
} khr_df_sample_datatype_qualifiers_e;
#endif

752
include/meshoptimizer/indexcodec.cpp Normal file
View File

@ -0,0 +1,752 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
#include <assert.h>
#include <string.h>
#ifndef TRACE
#define TRACE 0
#endif
#if TRACE
#include <stdio.h>
#endif
// This work is based on:
// Fabian Giesen. Simple lossless index buffer compression & follow-up. 2013
// Conor Stokes. Vertex Cache Optimised Index Buffer Compression. 2014
namespace meshopt
{
const unsigned char kIndexHeader = 0xe0;
const unsigned char kSequenceHeader = 0xd0;
static int gEncodeIndexVersion = 0;
typedef unsigned int VertexFifo[16];
typedef unsigned int EdgeFifo[16][2];
static const unsigned int kTriangleIndexOrder[3][3] = {
{0, 1, 2},
{1, 2, 0},
{2, 0, 1},
};
static const unsigned char kCodeAuxEncodingTable[16] = {
0x00, 0x76, 0x87, 0x56, 0x67, 0x78, 0xa9, 0x86, 0x65, 0x89, 0x68, 0x98, 0x01, 0x69,
0, 0, // last two entries aren't used for encoding
};
static int rotateTriangle(unsigned int a, unsigned int b, unsigned int c, unsigned int next)
{
(void)a;
return (b == next) ? 1 : (c == next) ? 2 : 0;
}
static int getEdgeFifo(EdgeFifo fifo, unsigned int a, unsigned int b, unsigned int c, size_t offset)
{
for (int i = 0; i < 16; ++i)
{
size_t index = (offset - 1 - i) & 15;
unsigned int e0 = fifo[index][0];
unsigned int e1 = fifo[index][1];
if (e0 == a && e1 == b)
return (i << 2) | 0;
if (e0 == b && e1 == c)
return (i << 2) | 1;
if (e0 == c && e1 == a)
return (i << 2) | 2;
}
return -1;
}
static void pushEdgeFifo(EdgeFifo fifo, unsigned int a, unsigned int b, size_t& offset)
{
fifo[offset][0] = a;
fifo[offset][1] = b;
offset = (offset + 1) & 15;
}
static int getVertexFifo(VertexFifo fifo, unsigned int v, size_t offset)
{
for (int i = 0; i < 16; ++i)
{
size_t index = (offset - 1 - i) & 15;
if (fifo[index] == v)
return i;
}
return -1;
}
static void pushVertexFifo(VertexFifo fifo, unsigned int v, size_t& offset, int cond = 1)
{
fifo[offset] = v;
offset = (offset + cond) & 15;
}
static void encodeVByte(unsigned char*& data, unsigned int v)
{
// encode 32-bit value in up to 5 7-bit groups
do
{
*data++ = (v & 127) | (v > 127 ? 128 : 0);
v >>= 7;
} while (v);
}
static unsigned int decodeVByte(const unsigned char*& data)
{
unsigned char lead = *data++;
// fast path: single byte
if (lead < 128)
return lead;
// slow path: up to 4 extra bytes
// note that this loop always terminates, which is important for malformed data
unsigned int result = lead & 127;
unsigned int shift = 7;
for (int i = 0; i < 4; ++i)
{
unsigned char group = *data++;
result |= (group & 127) << shift;
shift += 7;
if (group < 128)
break;
}
return result;
}
static void encodeIndex(unsigned char*& data, unsigned int index, unsigned int last)
{
unsigned int d = index - last;
unsigned int v = (d << 1) ^ (int(d) >> 31);
encodeVByte(data, v);
}
static unsigned int decodeIndex(const unsigned char*& data, unsigned int last)
{
unsigned int v = decodeVByte(data);
unsigned int d = (v >> 1) ^ -int(v & 1);
return last + d;
}
static int getCodeAuxIndex(unsigned char v, const unsigned char* table)
{
for (int i = 0; i < 16; ++i)
if (table[i] == v)
return i;
return -1;
}
static void writeTriangle(void* destination, size_t offset, size_t index_size, unsigned int a, unsigned int b, unsigned int c)
{
if (index_size == 2)
{
static_cast<unsigned short*>(destination)[offset + 0] = (unsigned short)(a);
static_cast<unsigned short*>(destination)[offset + 1] = (unsigned short)(b);
static_cast<unsigned short*>(destination)[offset + 2] = (unsigned short)(c);
}
else
{
static_cast<unsigned int*>(destination)[offset + 0] = a;
static_cast<unsigned int*>(destination)[offset + 1] = b;
static_cast<unsigned int*>(destination)[offset + 2] = c;
}
}
#if TRACE
static size_t sortTop16(unsigned char dest[16], size_t stats[256])
{
size_t destsize = 0;
for (size_t i = 0; i < 256; ++i)
{
size_t j = 0;
for (; j < destsize; ++j)
{
if (stats[i] >= stats[dest[j]])
{
if (destsize < 16)
destsize++;
memmove(&dest[j + 1], &dest[j], destsize - 1 - j);
dest[j] = (unsigned char)i;
break;
}
}
if (j == destsize && destsize < 16)
{
dest[destsize] = (unsigned char)i;
destsize++;
}
}
return destsize;
}
#endif
} // namespace meshopt
size_t meshopt_encodeIndexBuffer(unsigned char* buffer, size_t buffer_size, const unsigned int* indices, size_t index_count)
{
using namespace meshopt;
assert(index_count % 3 == 0);
#if TRACE
size_t codestats[256] = {};
size_t codeauxstats[256] = {};
#endif
// the minimum valid encoding is header, 1 byte per triangle and a 16-byte codeaux table
if (buffer_size < 1 + index_count / 3 + 16)
return 0;
int version = gEncodeIndexVersion;
buffer[0] = (unsigned char)(kIndexHeader | version);
EdgeFifo edgefifo;
memset(edgefifo, -1, sizeof(edgefifo));
VertexFifo vertexfifo;
memset(vertexfifo, -1, sizeof(vertexfifo));
size_t edgefifooffset = 0;
size_t vertexfifooffset = 0;
unsigned int next = 0;
unsigned int last = 0;
unsigned char* code = buffer + 1;
unsigned char* data = code + index_count / 3;
unsigned char* data_safe_end = buffer + buffer_size - 16;
int fecmax = version >= 1 ? 13 : 15;
// use static encoding table; it's possible to pack the result and then build an optimal table and repack
// for now we keep it simple and use the table that has been generated based on symbol frequency on a training mesh set
const unsigned char* codeaux_table = kCodeAuxEncodingTable;
for (size_t i = 0; i < index_count; i += 3)
{
// make sure we have enough space to write a triangle
// each triangle writes at most 16 bytes: 1b for codeaux and 5b for each free index
// after this we can be sure we can write without extra bounds checks
if (data > data_safe_end)
return 0;
int fer = getEdgeFifo(edgefifo, indices[i + 0], indices[i + 1], indices[i + 2], edgefifooffset);
if (fer >= 0 && (fer >> 2) < 15)
{
const unsigned int* order = kTriangleIndexOrder[fer & 3];
unsigned int a = indices[i + order[0]], b = indices[i + order[1]], c = indices[i + order[2]];
// encode edge index and vertex fifo index, next or free index
int fe = fer >> 2;
int fc = getVertexFifo(vertexfifo, c, vertexfifooffset);
int fec = (fc >= 1 && fc < fecmax) ? fc : (c == next) ? (next++, 0) : 15;
if (fec == 15 && version >= 1)
{
// encode last-1 and last+1 to optimize strip-like sequences
if (c + 1 == last)
fec = 13, last = c;
if (c == last + 1)
fec = 14, last = c;
}
*code++ = (unsigned char)((fe << 4) | fec);
#if TRACE
codestats[code[-1]]++;
#endif
// note that we need to update the last index since free indices are delta-encoded
if (fec == 15)
encodeIndex(data, c, last), last = c;
// we only need to push third vertex since first two are likely already in the vertex fifo
if (fec == 0 || fec >= fecmax)
pushVertexFifo(vertexfifo, c, vertexfifooffset);
// we only need to push two new edges to edge fifo since the third one is already there
pushEdgeFifo(edgefifo, c, b, edgefifooffset);
pushEdgeFifo(edgefifo, a, c, edgefifooffset);
}
else
{
int rotation = rotateTriangle(indices[i + 0], indices[i + 1], indices[i + 2], next);
const unsigned int* order = kTriangleIndexOrder[rotation];
unsigned int a = indices[i + order[0]], b = indices[i + order[1]], c = indices[i + order[2]];
// if a/b/c are 0/1/2, we emit a reset code
bool reset = false;
if (a == 0 && b == 1 && c == 2 && next > 0 && version >= 1)
{
reset = true;
next = 0;
// reset vertex fifo to make sure we don't accidentally reference vertices from that in the future
// this makes sure next continues to get incremented instead of being stuck
memset(vertexfifo, -1, sizeof(vertexfifo));
}
int fb = getVertexFifo(vertexfifo, b, vertexfifooffset);
int fc = getVertexFifo(vertexfifo, c, vertexfifooffset);
// after rotation, a is almost always equal to next, so we don't waste bits on FIFO encoding for a
int fea = (a == next) ? (next++, 0) : 15;
int feb = (fb >= 0 && fb < 14) ? (fb + 1) : (b == next) ? (next++, 0) : 15;
int fec = (fc >= 0 && fc < 14) ? (fc + 1) : (c == next) ? (next++, 0) : 15;
// we encode feb & fec in 4 bits using a table if possible, and as a full byte otherwise
unsigned char codeaux = (unsigned char)((feb << 4) | fec);
int codeauxindex = getCodeAuxIndex(codeaux, codeaux_table);
// <14 encodes an index into codeaux table, 14 encodes fea=0, 15 encodes fea=15
if (fea == 0 && codeauxindex >= 0 && codeauxindex < 14 && !reset)
{
*code++ = (unsigned char)((15 << 4) | codeauxindex);
}
else
{
*code++ = (unsigned char)((15 << 4) | 14 | fea);
*data++ = codeaux;
}
#if TRACE
codestats[code[-1]]++;
codeauxstats[codeaux]++;
#endif
// note that we need to update the last index since free indices are delta-encoded
if (fea == 15)
encodeIndex(data, a, last), last = a;
if (feb == 15)
encodeIndex(data, b, last), last = b;
if (fec == 15)
encodeIndex(data, c, last), last = c;
// only push vertices that weren't already in fifo
if (fea == 0 || fea == 15)
pushVertexFifo(vertexfifo, a, vertexfifooffset);
if (feb == 0 || feb == 15)
pushVertexFifo(vertexfifo, b, vertexfifooffset);
if (fec == 0 || fec == 15)
pushVertexFifo(vertexfifo, c, vertexfifooffset);
// all three edges aren't in the fifo; pushing all of them is important so that we can match them for later triangles
pushEdgeFifo(edgefifo, b, a, edgefifooffset);
pushEdgeFifo(edgefifo, c, b, edgefifooffset);
pushEdgeFifo(edgefifo, a, c, edgefifooffset);
}
}
// make sure we have enough space to write codeaux table
if (data > data_safe_end)
return 0;
// add codeaux encoding table to the end of the stream; this is used for decoding codeaux *and* as padding
// we need padding for decoding to be able to assume that each triangle is encoded as <= 16 bytes of extra data
// this is enough space for aux byte + 5 bytes per varint index which is the absolute worst case for any input
for (size_t i = 0; i < 16; ++i)
{
// decoder assumes that table entries never refer to separately encoded indices
assert((codeaux_table[i] & 0xf) != 0xf && (codeaux_table[i] >> 4) != 0xf);
*data++ = codeaux_table[i];
}
// since we encode restarts as codeaux without a table reference, we need to make sure 00 is encoded as a table reference
assert(codeaux_table[0] == 0);
assert(data >= buffer + index_count / 3 + 16);
assert(data <= buffer + buffer_size);
#if TRACE
unsigned char codetop[16], codeauxtop[16];
size_t codetopsize = sortTop16(codetop, codestats);
size_t codeauxtopsize = sortTop16(codeauxtop, codeauxstats);
size_t sumcode = 0, sumcodeaux = 0;
for (size_t i = 0; i < 256; ++i)
sumcode += codestats[i], sumcodeaux += codeauxstats[i];
size_t acccode = 0, acccodeaux = 0;
printf("code\t\t\t\t\tcodeaux\n");
for (size_t i = 0; i < codetopsize && i < codeauxtopsize; ++i)
{
acccode += codestats[codetop[i]];
acccodeaux += codeauxstats[codeauxtop[i]];
printf("%2d: %02x = %d (%.1f%% ..%.1f%%)\t\t%2d: %02x = %d (%.1f%% ..%.1f%%)\n",
int(i), codetop[i], int(codestats[codetop[i]]), double(codestats[codetop[i]]) / double(sumcode) * 100, double(acccode) / double(sumcode) * 100,
int(i), codeauxtop[i], int(codeauxstats[codeauxtop[i]]), double(codeauxstats[codeauxtop[i]]) / double(sumcodeaux) * 100, double(acccodeaux) / double(sumcodeaux) * 100);
}
#endif
return data - buffer;
}
size_t meshopt_encodeIndexBufferBound(size_t index_count, size_t vertex_count)
{
assert(index_count % 3 == 0);
// compute number of bits required for each index
unsigned int vertex_bits = 1;
while (vertex_bits < 32 && vertex_count > size_t(1) << vertex_bits)
vertex_bits++;
// worst-case encoding is 2 header bytes + 3 varint-7 encoded index deltas
unsigned int vertex_groups = (vertex_bits + 1 + 6) / 7;
return 1 + (index_count / 3) * (2 + 3 * vertex_groups) + 16;
}
void meshopt_encodeIndexVersion(int version)
{
assert(unsigned(version) <= 1);
meshopt::gEncodeIndexVersion = version;
}
int meshopt_decodeIndexBuffer(void* destination, size_t index_count, size_t index_size, const unsigned char* buffer, size_t buffer_size)
{
using namespace meshopt;
assert(index_count % 3 == 0);
assert(index_size == 2 || index_size == 4);
// the minimum valid encoding is header, 1 byte per triangle and a 16-byte codeaux table
if (buffer_size < 1 + index_count / 3 + 16)
return -2;
if ((buffer[0] & 0xf0) != kIndexHeader)
return -1;
int version = buffer[0] & 0x0f;
if (version > 1)
return -1;
EdgeFifo edgefifo;
memset(edgefifo, -1, sizeof(edgefifo));
VertexFifo vertexfifo;
memset(vertexfifo, -1, sizeof(vertexfifo));
size_t edgefifooffset = 0;
size_t vertexfifooffset = 0;
unsigned int next = 0;
unsigned int last = 0;
int fecmax = version >= 1 ? 13 : 15;
// since we store 16-byte codeaux table at the end, triangle data has to begin before data_safe_end
const unsigned char* code = buffer + 1;
const unsigned char* data = code + index_count / 3;
const unsigned char* data_safe_end = buffer + buffer_size - 16;
const unsigned char* codeaux_table = data_safe_end;
for (size_t i = 0; i < index_count; i += 3)
{
// make sure we have enough data to read for a triangle
// each triangle reads at most 16 bytes of data: 1b for codeaux and 5b for each free index
// after this we can be sure we can read without extra bounds checks
if (data > data_safe_end)
return -2;
unsigned char codetri = *code++;
if (codetri < 0xf0)
{
int fe = codetri >> 4;
// fifo reads are wrapped around 16 entry buffer
unsigned int a = edgefifo[(edgefifooffset - 1 - fe) & 15][0];
unsigned int b = edgefifo[(edgefifooffset - 1 - fe) & 15][1];
int fec = codetri & 15;
// note: this is the most common path in the entire decoder
// inside this if we try to stay branchless (by using cmov/etc.) since these aren't predictable
if (fec < fecmax)
{
// fifo reads are wrapped around 16 entry buffer
unsigned int cf = vertexfifo[(vertexfifooffset - 1 - fec) & 15];
unsigned int c = (fec == 0) ? next : cf;
int fec0 = fec == 0;
next += fec0;
// output triangle
writeTriangle(destination, i, index_size, a, b, c);
// push vertex/edge fifo must match the encoding step *exactly* otherwise the data will not be decoded correctly
pushVertexFifo(vertexfifo, c, vertexfifooffset, fec0);
pushEdgeFifo(edgefifo, c, b, edgefifooffset);
pushEdgeFifo(edgefifo, a, c, edgefifooffset);
}
else
{
unsigned int c = 0;
// fec - (fec ^ 3) decodes 13, 14 into -1, 1
// note that we need to update the last index since free indices are delta-encoded
last = c = (fec != 15) ? last + (fec - (fec ^ 3)) : decodeIndex(data, last);
// output triangle
writeTriangle(destination, i, index_size, a, b, c);
// push vertex/edge fifo must match the encoding step *exactly* otherwise the data will not be decoded correctly
pushVertexFifo(vertexfifo, c, vertexfifooffset);
pushEdgeFifo(edgefifo, c, b, edgefifooffset);
pushEdgeFifo(edgefifo, a, c, edgefifooffset);
}
}
else
{
// fast path: read codeaux from the table
if (codetri < 0xfe)
{
unsigned char codeaux = codeaux_table[codetri & 15];
// note: table can't contain feb/fec=15
int feb = codeaux >> 4;
int fec = codeaux & 15;
// fifo reads are wrapped around 16 entry buffer
// also note that we increment next for all three vertices before decoding indices - this matches encoder behavior
unsigned int a = next++;
unsigned int bf = vertexfifo[(vertexfifooffset - feb) & 15];
unsigned int b = (feb == 0) ? next : bf;
int feb0 = feb == 0;
next += feb0;
unsigned int cf = vertexfifo[(vertexfifooffset - fec) & 15];
unsigned int c = (fec == 0) ? next : cf;
int fec0 = fec == 0;
next += fec0;
// output triangle
writeTriangle(destination, i, index_size, a, b, c);
// push vertex/edge fifo must match the encoding step *exactly* otherwise the data will not be decoded correctly
pushVertexFifo(vertexfifo, a, vertexfifooffset);
pushVertexFifo(vertexfifo, b, vertexfifooffset, feb0);
pushVertexFifo(vertexfifo, c, vertexfifooffset, fec0);
pushEdgeFifo(edgefifo, b, a, edgefifooffset);
pushEdgeFifo(edgefifo, c, b, edgefifooffset);
pushEdgeFifo(edgefifo, a, c, edgefifooffset);
}
else
{
// slow path: read a full byte for codeaux instead of using a table lookup
unsigned char codeaux = *data++;
int fea = codetri == 0xfe ? 0 : 15;
int feb = codeaux >> 4;
int fec = codeaux & 15;
// reset: codeaux is 0 but encoded as not-a-table
if (codeaux == 0)
next = 0;
// fifo reads are wrapped around 16 entry buffer
// also note that we increment next for all three vertices before decoding indices - this matches encoder behavior
unsigned int a = (fea == 0) ? next++ : 0;
unsigned int b = (feb == 0) ? next++ : vertexfifo[(vertexfifooffset - feb) & 15];
unsigned int c = (fec == 0) ? next++ : vertexfifo[(vertexfifooffset - fec) & 15];
// note that we need to update the last index since free indices are delta-encoded
if (fea == 15)
last = a = decodeIndex(data, last);
if (feb == 15)
last = b = decodeIndex(data, last);
if (fec == 15)
last = c = decodeIndex(data, last);
// output triangle
writeTriangle(destination, i, index_size, a, b, c);
// push vertex/edge fifo must match the encoding step *exactly* otherwise the data will not be decoded correctly
pushVertexFifo(vertexfifo, a, vertexfifooffset);
pushVertexFifo(vertexfifo, b, vertexfifooffset, (feb == 0) | (feb == 15));
pushVertexFifo(vertexfifo, c, vertexfifooffset, (fec == 0) | (fec == 15));
pushEdgeFifo(edgefifo, b, a, edgefifooffset);
pushEdgeFifo(edgefifo, c, b, edgefifooffset);
pushEdgeFifo(edgefifo, a, c, edgefifooffset);
}
}
}
// we should've read all data bytes and stopped at the boundary between data and codeaux table
if (data != data_safe_end)
return -3;
return 0;
}
size_t meshopt_encodeIndexSequence(unsigned char* buffer, size_t buffer_size, const unsigned int* indices, size_t index_count)
{
using namespace meshopt;
// the minimum valid encoding is header, 1 byte per index and a 4-byte tail
if (buffer_size < 1 + index_count + 4)
return 0;
int version = gEncodeIndexVersion;
buffer[0] = (unsigned char)(kSequenceHeader | version);
unsigned int last[2] = {};
unsigned int current = 0;
unsigned char* data = buffer + 1;
unsigned char* data_safe_end = buffer + buffer_size - 4;
for (size_t i = 0; i < index_count; ++i)
{
// make sure we have enough data to write
// each index writes at most 5 bytes of data; there's a 4 byte tail after data_safe_end
// after this we can be sure we can write without extra bounds checks
if (data >= data_safe_end)
return 0;
unsigned int index = indices[i];
// this is a heuristic that switches between baselines when the delta grows too large
// we want the encoded delta to fit into one byte (7 bits), but 2 bits are used for sign and baseline index
// for now we immediately switch the baseline when delta grows too large - this can be adjusted arbitrarily
int cd = int(index - last[current]);
current ^= ((cd < 0 ? -cd : cd) >= 30);
// encode delta from the last index
unsigned int d = index - last[current];
unsigned int v = (d << 1) ^ (int(d) >> 31);
// note: low bit encodes the index of the last baseline which will be used for reconstruction
encodeVByte(data, (v << 1) | current);
// update last for the next iteration that uses it
last[current] = index;
}
// make sure we have enough space to write tail
if (data > data_safe_end)
return 0;
for (int k = 0; k < 4; ++k)
*data++ = 0;
return data - buffer;
}
size_t meshopt_encodeIndexSequenceBound(size_t index_count, size_t vertex_count)
{
// compute number of bits required for each index
unsigned int vertex_bits = 1;
while (vertex_bits < 32 && vertex_count > size_t(1) << vertex_bits)
vertex_bits++;
// worst-case encoding is 1 varint-7 encoded index delta for a K bit value and an extra bit
unsigned int vertex_groups = (vertex_bits + 1 + 1 + 6) / 7;
return 1 + index_count * vertex_groups + 4;
}
int meshopt_decodeIndexSequence(void* destination, size_t index_count, size_t index_size, const unsigned char* buffer, size_t buffer_size)
{
using namespace meshopt;
// the minimum valid encoding is header, 1 byte per index and a 4-byte tail
if (buffer_size < 1 + index_count + 4)
return -2;
if ((buffer[0] & 0xf0) != kSequenceHeader)
return -1;
int version = buffer[0] & 0x0f;
if (version > 1)
return -1;
const unsigned char* data = buffer + 1;
const unsigned char* data_safe_end = buffer + buffer_size - 4;
unsigned int last[2] = {};
for (size_t i = 0; i < index_count; ++i)
{
// make sure we have enough data to read
// each index reads at most 5 bytes of data; there's a 4 byte tail after data_safe_end
// after this we can be sure we can read without extra bounds checks
if (data >= data_safe_end)
return -2;
unsigned int v = decodeVByte(data);
// decode the index of the last baseline
unsigned int current = v & 1;
v >>= 1;
// reconstruct index as a delta
unsigned int d = (v >> 1) ^ -int(v & 1);
unsigned int index = last[current] + d;
// update last for the next iteration that uses it
last[current] = index;
if (index_size == 2)
{
static_cast<unsigned short*>(destination)[i] = (unsigned short)(index);
}
else
{
static_cast<unsigned int*>(destination)[i] = index;
}
}
// we should've read all data bytes and stopped at the boundary between data and tail
if (data != data_safe_end)
return -3;
return 0;
}

347
include/meshoptimizer/indexgenerator.cpp Normal file
View File

@ -0,0 +1,347 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
#include <assert.h>
#include <string.h>
namespace meshopt
{
static unsigned int hashUpdate4(unsigned int h, const unsigned char* key, size_t len)
{
// MurmurHash2
const unsigned int m = 0x5bd1e995;
const int r = 24;
while (len >= 4)
{
unsigned int k = *reinterpret_cast<const unsigned int*>(key);
k *= m;
k ^= k >> r;
k *= m;
h *= m;
h ^= k;
key += 4;
len -= 4;
}
return h;
}
struct VertexHasher
{
const unsigned char* vertices;
size_t vertex_size;
size_t vertex_stride;
size_t hash(unsigned int index) const
{
return hashUpdate4(0, vertices + index * vertex_stride, vertex_size);
}
bool equal(unsigned int lhs, unsigned int rhs) const
{
return memcmp(vertices + lhs * vertex_stride, vertices + rhs * vertex_stride, vertex_size) == 0;
}
};
struct VertexStreamHasher
{
const meshopt_Stream* streams;
size_t stream_count;
size_t hash(unsigned int index) const
{
unsigned int h = 0;
for (size_t i = 0; i < stream_count; ++i)
{
const meshopt_Stream& s = streams[i];
const unsigned char* data = static_cast<const unsigned char*>(s.data);
h = hashUpdate4(h, data + index * s.stride, s.size);
}
return h;
}
bool equal(unsigned int lhs, unsigned int rhs) const
{
for (size_t i = 0; i < stream_count; ++i)
{
const meshopt_Stream& s = streams[i];
const unsigned char* data = static_cast<const unsigned char*>(s.data);
if (memcmp(data + lhs * s.stride, data + rhs * s.stride, s.size) != 0)
return false;
}
return true;
}
};
static size_t hashBuckets(size_t count)
{
size_t buckets = 1;
while (buckets < count)
buckets *= 2;
return buckets;
}
template <typename T, typename Hash>
static T* hashLookup(T* table, size_t buckets, const Hash& hash, const T& key, const T& empty)
{
assert(buckets > 0);
assert((buckets & (buckets - 1)) == 0);
size_t hashmod = buckets - 1;
size_t bucket = hash.hash(key) & hashmod;
for (size_t probe = 0; probe <= hashmod; ++probe)
{
T& item = table[bucket];
if (item == empty)
return &item;
if (hash.equal(item, key))
return &item;
// hash collision, quadratic probing
bucket = (bucket + probe + 1) & hashmod;
}
assert(false && "Hash table is full"); // unreachable
return 0;
}
} // namespace meshopt
size_t meshopt_generateVertexRemap(unsigned int* destination, const unsigned int* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size)
{
using namespace meshopt;
assert(indices || index_count == vertex_count);
assert(index_count % 3 == 0);
assert(vertex_size > 0 && vertex_size <= 256);
meshopt_Allocator allocator;
memset(destination, -1, vertex_count * sizeof(unsigned int));
VertexHasher hasher = {static_cast<const unsigned char*>(vertices), vertex_size, vertex_size};
size_t table_size = hashBuckets(vertex_count);
unsigned int* table = allocator.allocate<unsigned int>(table_size);
memset(table, -1, table_size * sizeof(unsigned int));
unsigned int next_vertex = 0;
for (size_t i = 0; i < index_count; ++i)
{
unsigned int index = indices ? indices[i] : unsigned(i);
assert(index < vertex_count);
if (destination[index] == ~0u)
{
unsigned int* entry = hashLookup(table, table_size, hasher, index, ~0u);
if (*entry == ~0u)
{
*entry = index;
destination[index] = next_vertex++;
}
else
{
assert(destination[*entry] != ~0u);
destination[index] = destination[*entry];
}
}
}
assert(next_vertex <= vertex_count);
return next_vertex;
}
size_t meshopt_generateVertexRemapMulti(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, const struct meshopt_Stream* streams, size_t stream_count)
{
using namespace meshopt;
assert(indices || index_count == vertex_count);
assert(index_count % 3 == 0);
assert(stream_count > 0 && stream_count <= 16);
for (size_t i = 0; i < stream_count; ++i)
{
assert(streams[i].size > 0 && streams[i].size <= 256);
assert(streams[i].size <= streams[i].stride);
}
meshopt_Allocator allocator;
memset(destination, -1, vertex_count * sizeof(unsigned int));
VertexStreamHasher hasher = {streams, stream_count};
size_t table_size = hashBuckets(vertex_count);
unsigned int* table = allocator.allocate<unsigned int>(table_size);
memset(table, -1, table_size * sizeof(unsigned int));
unsigned int next_vertex = 0;
for (size_t i = 0; i < index_count; ++i)
{
unsigned int index = indices ? indices[i] : unsigned(i);
assert(index < vertex_count);
if (destination[index] == ~0u)
{
unsigned int* entry = hashLookup(table, table_size, hasher, index, ~0u);
if (*entry == ~0u)
{
*entry = index;
destination[index] = next_vertex++;
}
else
{
assert(destination[*entry] != ~0u);
destination[index] = destination[*entry];
}
}
}
assert(next_vertex <= vertex_count);
return next_vertex;
}
void meshopt_remapVertexBuffer(void* destination, const void* vertices, size_t vertex_count, size_t vertex_size, const unsigned int* remap)
{
assert(vertex_size > 0 && vertex_size <= 256);
meshopt_Allocator allocator;
// support in-place remap
if (destination == vertices)
{
unsigned char* vertices_copy = allocator.allocate<unsigned char>(vertex_count * vertex_size);
memcpy(vertices_copy, vertices, vertex_count * vertex_size);
vertices = vertices_copy;
}
for (size_t i = 0; i < vertex_count; ++i)
{
if (remap[i] != ~0u)
{
assert(remap[i] < vertex_count);
memcpy(static_cast<unsigned char*>(destination) + remap[i] * vertex_size, static_cast<const unsigned char*>(vertices) + i * vertex_size, vertex_size);
}
}
}
void meshopt_remapIndexBuffer(unsigned int* destination, const unsigned int* indices, size_t index_count, const unsigned int* remap)
{
assert(index_count % 3 == 0);
for (size_t i = 0; i < index_count; ++i)
{
unsigned int index = indices ? indices[i] : unsigned(i);
assert(remap[index] != ~0u);
destination[i] = remap[index];
}
}
void meshopt_generateShadowIndexBuffer(unsigned int* destination, const unsigned int* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size, size_t vertex_stride)
{
using namespace meshopt;
assert(indices);
assert(index_count % 3 == 0);
assert(vertex_size > 0 && vertex_size <= 256);
assert(vertex_size <= vertex_stride);
meshopt_Allocator allocator;
unsigned int* remap = allocator.allocate<unsigned int>(vertex_count);
memset(remap, -1, vertex_count * sizeof(unsigned int));
VertexHasher hasher = {static_cast<const unsigned char*>(vertices), vertex_size, vertex_stride};
size_t table_size = hashBuckets(vertex_count);
unsigned int* table = allocator.allocate<unsigned int>(table_size);
memset(table, -1, table_size * sizeof(unsigned int));
for (size_t i = 0; i < index_count; ++i)
{
unsigned int index = indices[i];
assert(index < vertex_count);
if (remap[index] == ~0u)
{
unsigned int* entry = hashLookup(table, table_size, hasher, index, ~0u);
if (*entry == ~0u)
*entry = index;
remap[index] = *entry;
}
destination[i] = remap[index];
}
}
void meshopt_generateShadowIndexBufferMulti(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, const struct meshopt_Stream* streams, size_t stream_count)
{
using namespace meshopt;
assert(indices);
assert(index_count % 3 == 0);
assert(stream_count > 0 && stream_count <= 16);
for (size_t i = 0; i < stream_count; ++i)
{
assert(streams[i].size > 0 && streams[i].size <= 256);
assert(streams[i].size <= streams[i].stride);
}
meshopt_Allocator allocator;
unsigned int* remap = allocator.allocate<unsigned int>(vertex_count);
memset(remap, -1, vertex_count * sizeof(unsigned int));
VertexStreamHasher hasher = {streams, stream_count};
size_t table_size = hashBuckets(vertex_count);
unsigned int* table = allocator.allocate<unsigned int>(table_size);
memset(table, -1, table_size * sizeof(unsigned int));
for (size_t i = 0; i < index_count; ++i)
{
unsigned int index = indices[i];
assert(index < vertex_count);
if (remap[index] == ~0u)
{
unsigned int* entry = hashLookup(table, table_size, hasher, index, ~0u);
if (*entry == ~0u)
*entry = index;
remap[index] = *entry;
}
destination[i] = remap[index];
}
}

943
include/meshoptimizer/meshoptimizer.h Normal file
View File

@ -0,0 +1,943 @@
/**
* meshoptimizer - version 0.14
*
* Copyright (C) 2016-2020, by Arseny Kapoulkine (arseny.kapoulkine@gmail.com)
* Report bugs and download new versions at https://github.com/zeux/meshoptimizer
*
* This library is distributed under the MIT License. See notice at the end of this file.
*/
#pragma once
#include <assert.h>
#include <stddef.h>
/* Version macro; major * 1000 + minor * 10 + patch */
#define MESHOPTIMIZER_VERSION 140
/* If no API is defined, assume default */
#ifndef MESHOPTIMIZER_API
#define MESHOPTIMIZER_API
#endif
/* Experimental APIs have unstable interface and might have implementation that's not fully tested or optimized */
#define MESHOPTIMIZER_EXPERIMENTAL MESHOPTIMIZER_API
/* C interface */
#ifdef __cplusplus
extern "C" {
#endif
/**
* Vertex attribute stream, similar to glVertexPointer
* Each element takes size bytes, with stride controlling the spacing between successive elements.
*/
struct meshopt_Stream
{
const void* data;
size_t size;
size_t stride;
};
/**
* Generates a vertex remap table from the vertex buffer and an optional index buffer and returns number of unique vertices
* As a result, all vertices that are binary equivalent map to the same (new) location, with no gaps in the resulting sequence.
* Resulting remap table maps old vertices to new vertices and can be used in meshopt_remapVertexBuffer/meshopt_remapIndexBuffer.
*
* destination must contain enough space for the resulting remap table (vertex_count elements)
* indices can be NULL if the input is unindexed
*/
MESHOPTIMIZER_API size_t meshopt_generateVertexRemap(unsigned int* destination, const unsigned int* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size);
/**
* Generates a vertex remap table from multiple vertex streams and an optional index buffer and returns number of unique vertices
* As a result, all vertices that are binary equivalent map to the same (new) location, with no gaps in the resulting sequence.
* Resulting remap table maps old vertices to new vertices and can be used in meshopt_remapVertexBuffer/meshopt_remapIndexBuffer.
* To remap vertex buffers, you will need to call meshopt_remapVertexBuffer for each vertex stream.
*
* destination must contain enough space for the resulting remap table (vertex_count elements)
* indices can be NULL if the input is unindexed
*/
MESHOPTIMIZER_API size_t meshopt_generateVertexRemapMulti(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, const struct meshopt_Stream* streams, size_t stream_count);
/**
* Generates vertex buffer from the source vertex buffer and remap table generated by meshopt_generateVertexRemap
*
* destination must contain enough space for the resulting vertex buffer (unique_vertex_count elements, returned by meshopt_generateVertexRemap)
* vertex_count should be the initial vertex count and not the value returned by meshopt_generateVertexRemap
*/
MESHOPTIMIZER_API void meshopt_remapVertexBuffer(void* destination, const void* vertices, size_t vertex_count, size_t vertex_size, const unsigned int* remap);
/**
* Generate index buffer from the source index buffer and remap table generated by meshopt_generateVertexRemap
*
* destination must contain enough space for the resulting index buffer (index_count elements)
* indices can be NULL if the input is unindexed
*/
MESHOPTIMIZER_API void meshopt_remapIndexBuffer(unsigned int* destination, const unsigned int* indices, size_t index_count, const unsigned int* remap);
/**
* Generate index buffer that can be used for more efficient rendering when only a subset of the vertex attributes is necessary
* All vertices that are binary equivalent (wrt first vertex_size bytes) map to the first vertex in the original vertex buffer.
* This makes it possible to use the index buffer for Z pre-pass or shadowmap rendering, while using the original index buffer for regular rendering.
*
* destination must contain enough space for the resulting index buffer (index_count elements)
*/
MESHOPTIMIZER_API void meshopt_generateShadowIndexBuffer(unsigned int* destination, const unsigned int* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size, size_t vertex_stride);
/**
* Generate index buffer that can be used for more efficient rendering when only a subset of the vertex attributes is necessary
* All vertices that are binary equivalent (wrt specified streams) map to the first vertex in the original vertex buffer.
* This makes it possible to use the index buffer for Z pre-pass or shadowmap rendering, while using the original index buffer for regular rendering.
*
* destination must contain enough space for the resulting index buffer (index_count elements)
*/
MESHOPTIMIZER_API void meshopt_generateShadowIndexBufferMulti(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, const struct meshopt_Stream* streams, size_t stream_count);
/**
* Vertex transform cache optimizer
* Reorders indices to reduce the number of GPU vertex shader invocations
* If index buffer contains multiple ranges for multiple draw calls, this functions needs to be called on each range individually.
*
* destination must contain enough space for the resulting index buffer (index_count elements)
*/
MESHOPTIMIZER_API void meshopt_optimizeVertexCache(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count);
/**
* Vertex transform cache optimizer for strip-like caches
* Produces inferior results to meshopt_optimizeVertexCache from the GPU vertex cache perspective
* However, the resulting index order is more optimal if the goal is to reduce the triangle strip length or improve compression efficiency
*
* destination must contain enough space for the resulting index buffer (index_count elements)
*/
MESHOPTIMIZER_API void meshopt_optimizeVertexCacheStrip(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count);
/**
* Vertex transform cache optimizer for FIFO caches
* Reorders indices to reduce the number of GPU vertex shader invocations
* Generally takes ~3x less time to optimize meshes but produces inferior results compared to meshopt_optimizeVertexCache
* If index buffer contains multiple ranges for multiple draw calls, this functions needs to be called on each range individually.
*
* destination must contain enough space for the resulting index buffer (index_count elements)
* cache_size should be less than the actual GPU cache size to avoid cache thrashing
*/
MESHOPTIMIZER_API void meshopt_optimizeVertexCacheFifo(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int cache_size);
/**
* Overdraw optimizer
* Reorders indices to reduce the number of GPU vertex shader invocations and the pixel overdraw
* If index buffer contains multiple ranges for multiple draw calls, this functions needs to be called on each range individually.
*
* destination must contain enough space for the resulting index buffer (index_count elements)
* indices must contain index data that is the result of meshopt_optimizeVertexCache (*not* the original mesh indices!)
* vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
* threshold indicates how much the overdraw optimizer can degrade vertex cache efficiency (1.05 = up to 5%) to reduce overdraw more efficiently
*/
MESHOPTIMIZER_API void meshopt_optimizeOverdraw(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, float threshold);
/**
* Vertex fetch cache optimizer
* Reorders vertices and changes indices to reduce the amount of GPU memory fetches during vertex processing
* Returns the number of unique vertices, which is the same as input vertex count unless some vertices are unused
* This functions works for a single vertex stream; for multiple vertex streams, use meshopt_optimizeVertexFetchRemap + meshopt_remapVertexBuffer for each stream.
*
* destination must contain enough space for the resulting vertex buffer (vertex_count elements)
* indices is used both as an input and as an output index buffer
*/
MESHOPTIMIZER_API size_t meshopt_optimizeVertexFetch(void* destination, unsigned int* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size);
/**
* Vertex fetch cache optimizer
* Generates vertex remap to reduce the amount of GPU memory fetches during vertex processing
* Returns the number of unique vertices, which is the same as input vertex count unless some vertices are unused
* The resulting remap table should be used to reorder vertex/index buffers using meshopt_remapVertexBuffer/meshopt_remapIndexBuffer
*
* destination must contain enough space for the resulting remap table (vertex_count elements)
*/
MESHOPTIMIZER_API size_t meshopt_optimizeVertexFetchRemap(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count);
/**
* Index buffer encoder
* Encodes index data into an array of bytes that is generally much smaller (<1.5 bytes/triangle) and compresses better (<1 bytes/triangle) compared to original.
* Input index buffer must represent a triangle list.
* Returns encoded data size on success, 0 on error; the only error condition is if buffer doesn't have enough space
* For maximum efficiency the index buffer being encoded has to be optimized for vertex cache and vertex fetch first.
*
* buffer must contain enough space for the encoded index buffer (use meshopt_encodeIndexBufferBound to compute worst case size)
*/
MESHOPTIMIZER_API size_t meshopt_encodeIndexBuffer(unsigned char* buffer, size_t buffer_size, const unsigned int* indices, size_t index_count);
MESHOPTIMIZER_API size_t meshopt_encodeIndexBufferBound(size_t index_count, size_t vertex_count);
/**
* Experimental: Set index encoder format version
* version must specify the data format version to encode; valid values are 0 (decodable by all library versions) and 1 (decodable by 0.14+)
*/
MESHOPTIMIZER_EXPERIMENTAL void meshopt_encodeIndexVersion(int version);
/**
* Index buffer decoder
* Decodes index data from an array of bytes generated by meshopt_encodeIndexBuffer
* Returns 0 if decoding was successful, and an error code otherwise
* The decoder is safe to use for untrusted input, but it may produce garbage data (e.g. out of range indices).
*
* destination must contain enough space for the resulting index buffer (index_count elements)
*/
MESHOPTIMIZER_API int meshopt_decodeIndexBuffer(void* destination, size_t index_count, size_t index_size, const unsigned char* buffer, size_t buffer_size);
/**
* Experimental: Index sequence encoder
* Encodes index sequence into an array of bytes that is generally smaller and compresses better compared to original.
* Input index sequence can represent arbitrary topology; for triangle lists meshopt_encodeIndexBuffer is likely to be better.
* Returns encoded data size on success, 0 on error; the only error condition is if buffer doesn't have enough space
*
* buffer must contain enough space for the encoded index sequence (use meshopt_encodeIndexSequenceBound to compute worst case size)
*/
MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_encodeIndexSequence(unsigned char* buffer, size_t buffer_size, const unsigned int* indices, size_t index_count);
MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_encodeIndexSequenceBound(size_t index_count, size_t vertex_count);
/**
* Index sequence decoder
* Decodes index data from an array of bytes generated by meshopt_encodeIndexSequence
* Returns 0 if decoding was successful, and an error code otherwise
* The decoder is safe to use for untrusted input, but it may produce garbage data (e.g. out of range indices).
*
* destination must contain enough space for the resulting index sequence (index_count elements)
*/
MESHOPTIMIZER_EXPERIMENTAL int meshopt_decodeIndexSequence(void* destination, size_t index_count, size_t index_size, const unsigned char* buffer, size_t buffer_size);
/**
* Vertex buffer encoder
* Encodes vertex data into an array of bytes that is generally smaller and compresses better compared to original.
* Returns encoded data size on success, 0 on error; the only error condition is if buffer doesn't have enough space
* This function works for a single vertex stream; for multiple vertex streams, call meshopt_encodeVertexBuffer for each stream.
*
* buffer must contain enough space for the encoded vertex buffer (use meshopt_encodeVertexBufferBound to compute worst case size)
*/
MESHOPTIMIZER_API size_t meshopt_encodeVertexBuffer(unsigned char* buffer, size_t buffer_size, const void* vertices, size_t vertex_count, size_t vertex_size);
MESHOPTIMIZER_API size_t meshopt_encodeVertexBufferBound(size_t vertex_count, size_t vertex_size);
/**
* Experimental: Set vertex encoder format version
* version must specify the data format version to encode; valid values are 0 (decodable by all library versions)
*/
MESHOPTIMIZER_EXPERIMENTAL void meshopt_encodeVertexVersion(int version);
/**
* Vertex buffer decoder
* Decodes vertex data from an array of bytes generated by meshopt_encodeVertexBuffer
* Returns 0 if decoding was successful, and an error code otherwise
* The decoder is safe to use for untrusted input, but it may produce garbage data.
*
* destination must contain enough space for the resulting vertex buffer (vertex_count * vertex_size bytes)
*/
MESHOPTIMIZER_API int meshopt_decodeVertexBuffer(void* destination, size_t vertex_count, size_t vertex_size, const unsigned char* buffer, size_t buffer_size);
/**
* Vertex buffer filters
* These functions can be used to filter output of meshopt_decodeVertexBuffer in-place.
* count must be aligned by 4 and stride is fixed for each function to facilitate SIMD implementation.
*
* meshopt_decodeFilterOct decodes octahedral encoding of a unit vector with K-bit (K <= 16) signed X/Y as an input; Z must store 1.0f.
* Each component is stored as an 8-bit or 16-bit normalized integer; stride must be equal to 4 or 8. W is preserved as is.
*
* meshopt_decodeFilterQuat decodes 3-component quaternion encoding with K-bit (4 <= K <= 16) component encoding and a 2-bit component index indicating which component to reconstruct.
* Each component is stored as an 16-bit integer; stride must be equal to 8.
*
* meshopt_decodeFilterExp decodes exponential encoding of floating-point data with 8-bit exponent and 24-bit integer mantissa as 2^E*M.
* Each 32-bit component is decoded in isolation; stride must be divisible by 4.
*/
MESHOPTIMIZER_EXPERIMENTAL void meshopt_decodeFilterOct(void* buffer, size_t vertex_count, size_t vertex_size);
MESHOPTIMIZER_EXPERIMENTAL void meshopt_decodeFilterQuat(void* buffer, size_t vertex_count, size_t vertex_size);
MESHOPTIMIZER_EXPERIMENTAL void meshopt_decodeFilterExp(void* buffer, size_t vertex_count, size_t vertex_size);
/**
* Experimental: Mesh simplifier
* Reduces the number of triangles in the mesh, attempting to preserve mesh appearance as much as possible
* The algorithm tries to preserve mesh topology and can stop short of the target goal based on topology constraints or target error.
* If not all attributes from the input mesh are required, it's recommended to reindex the mesh using meshopt_generateShadowIndexBuffer prior to simplification.
* Returns the number of indices after simplification, with destination containing new index data
* The resulting index buffer references vertices from the original vertex buffer.
* If the original vertex data isn't required, creating a compact vertex buffer using meshopt_optimizeVertexFetch is recommended.
*
* destination must contain enough space for the *source* index buffer (since optimization is iterative, this means index_count elements - *not* target_index_count!)
* vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
*/
MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_simplify(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error);
/**
* Experimental: Mesh simplifier (sloppy)
* Reduces the number of triangles in the mesh, sacrificing mesh apperance for simplification performance
* The algorithm doesn't preserve mesh topology but is always able to reach target triangle count.
* Returns the number of indices after simplification, with destination containing new index data
* The resulting index buffer references vertices from the original vertex buffer.
* If the original vertex data isn't required, creating a compact vertex buffer using meshopt_optimizeVertexFetch is recommended.
*
* destination must contain enough space for the target index buffer
* vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
*/
MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_simplifySloppy(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count);
/**
* Experimental: Point cloud simplifier
* Reduces the number of points in the cloud to reach the given target
* Returns the number of points after simplification, with destination containing new index data
* The resulting index buffer references vertices from the original vertex buffer.
* If the original vertex data isn't required, creating a compact vertex buffer using meshopt_optimizeVertexFetch is recommended.
*
* destination must contain enough space for the target index buffer
* vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
*/
MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_simplifyPoints(unsigned int* destination, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_vertex_count);
/**
* Mesh stripifier
* Converts a previously vertex cache optimized triangle list to triangle strip, stitching strips using restart index or degenerate triangles
* Returns the number of indices in the resulting strip, with destination containing new index data
* For maximum efficiency the index buffer being converted has to be optimized for vertex cache first.
* Using restart indices can result in ~10% smaller index buffers, but on some GPUs restart indices may result in decreased performance.
*
* destination must contain enough space for the target index buffer, worst case can be computed with meshopt_stripifyBound
* restart_index should be 0xffff or 0xffffffff depending on index size, or 0 to use degenerate triangles
*/
MESHOPTIMIZER_API size_t meshopt_stripify(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int restart_index);
MESHOPTIMIZER_API size_t meshopt_stripifyBound(size_t index_count);
/**
* Mesh unstripifier
* Converts a triangle strip to a triangle list
* Returns the number of indices in the resulting list, with destination containing new index data
*
* destination must contain enough space for the target index buffer, worst case can be computed with meshopt_unstripifyBound
*/
MESHOPTIMIZER_API size_t meshopt_unstripify(unsigned int* destination, const unsigned int* indices, size_t index_count, unsigned int restart_index);
MESHOPTIMIZER_API size_t meshopt_unstripifyBound(size_t index_count);
struct meshopt_VertexCacheStatistics
{
unsigned int vertices_transformed;
unsigned int warps_executed;
float acmr; /* transformed vertices / triangle count; best case 0.5, worst case 3.0, optimum depends on topology */
float atvr; /* transformed vertices / vertex count; best case 1.0, worst case 6.0, optimum is 1.0 (each vertex is transformed once) */
};
/**
* Vertex transform cache analyzer
* Returns cache hit statistics using a simplified FIFO model
* Results may not match actual GPU performance
*/
MESHOPTIMIZER_API struct meshopt_VertexCacheStatistics meshopt_analyzeVertexCache(const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int cache_size, unsigned int warp_size, unsigned int primgroup_size);
struct meshopt_OverdrawStatistics
{
unsigned int pixels_covered;
unsigned int pixels_shaded;
float overdraw; /* shaded pixels / covered pixels; best case 1.0 */
};
/**
* Overdraw analyzer
* Returns overdraw statistics using a software rasterizer
* Results may not match actual GPU performance
*
* vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
*/
MESHOPTIMIZER_API struct meshopt_OverdrawStatistics meshopt_analyzeOverdraw(const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
struct meshopt_VertexFetchStatistics
{
unsigned int bytes_fetched;
float overfetch; /* fetched bytes / vertex buffer size; best case 1.0 (each byte is fetched once) */
};
/**
* Vertex fetch cache analyzer
* Returns cache hit statistics using a simplified direct mapped model
* Results may not match actual GPU performance
*/
MESHOPTIMIZER_API struct meshopt_VertexFetchStatistics meshopt_analyzeVertexFetch(const unsigned int* indices, size_t index_count, size_t vertex_count, size_t vertex_size);
struct meshopt_Meshlet
{
unsigned int vertices[64];
unsigned char indices[126][3];
unsigned char triangle_count;
unsigned char vertex_count;
};
/**
* Experimental: Meshlet builder
* Splits the mesh into a set of meshlets where each meshlet has a micro index buffer indexing into meshlet vertices that refer to the original vertex buffer
* The resulting data can be used to render meshes using NVidia programmable mesh shading pipeline, or in other cluster-based renderers.
* For maximum efficiency the index buffer being converted has to be optimized for vertex cache first.
*
* destination must contain enough space for all meshlets, worst case size can be computed with meshopt_buildMeshletsBound
* max_vertices and max_triangles can't exceed limits statically declared in meshopt_Meshlet (max_vertices <= 64, max_triangles <= 126)
*/
MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_buildMeshlets(struct meshopt_Meshlet* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, size_t max_vertices, size_t max_triangles);
MESHOPTIMIZER_EXPERIMENTAL size_t meshopt_buildMeshletsBound(size_t index_count, size_t max_vertices, size_t max_triangles);
struct meshopt_Bounds
{
/* bounding sphere, useful for frustum and occlusion culling */
float center[3];
float radius;
/* normal cone, useful for backface culling */
float cone_apex[3];
float cone_axis[3];
float cone_cutoff; /* = cos(angle/2) */
/* normal cone axis and cutoff, stored in 8-bit SNORM format; decode using x/127.0 */
signed char cone_axis_s8[3];
signed char cone_cutoff_s8;
};
/**
* Experimental: Cluster bounds generator
* Creates bounding volumes that can be used for frustum, backface and occlusion culling.
*
* For backface culling with orthographic projection, use the following formula to reject backfacing clusters:
* dot(view, cone_axis) >= cone_cutoff
*
* For perspective projection, you can the formula that needs cone apex in addition to axis & cutoff:
* dot(normalize(cone_apex - camera_position), cone_axis) >= cone_cutoff
*
* Alternatively, you can use the formula that doesn't need cone apex and uses bounding sphere instead:
* dot(normalize(center - camera_position), cone_axis) >= cone_cutoff + radius / length(center - camera_position)
* or an equivalent formula that doesn't have a singularity at center = camera_position:
* dot(center - camera_position, cone_axis) >= cone_cutoff * length(center - camera_position) + radius
*
* The formula that uses the apex is slightly more accurate but needs the apex; if you are already using bounding sphere
* to do frustum/occlusion culling, the formula that doesn't use the apex may be preferable.
*
* vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
* index_count should be less than or equal to 256*3 (the function assumes clusters of limited size)
*/
MESHOPTIMIZER_EXPERIMENTAL struct meshopt_Bounds meshopt_computeClusterBounds(const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
MESHOPTIMIZER_EXPERIMENTAL struct meshopt_Bounds meshopt_computeMeshletBounds(const struct meshopt_Meshlet* meshlet, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
/**
* Experimental: Spatial sorter
* Generates a remap table that can be used to reorder points for spatial locality.
* Resulting remap table maps old vertices to new vertices and can be used in meshopt_remapVertexBuffer.
*
* destination must contain enough space for the resulting remap table (vertex_count elements)
*/
MESHOPTIMIZER_EXPERIMENTAL void meshopt_spatialSortRemap(unsigned int* destination, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
/**
* Experimental: Spatial sorter
* Reorders triangles for spatial locality, and generates a new index buffer. The resulting index buffer can be used with other functions like optimizeVertexCache.
*
* destination must contain enough space for the resulting index buffer (index_count elements)
* vertex_positions should have float3 position in the first 12 bytes of each vertex - similar to glVertexPointer
*/
MESHOPTIMIZER_EXPERIMENTAL void meshopt_spatialSortTriangles(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
/**
* Set allocation callbacks
* These callbacks will be used instead of the default operator new/operator delete for all temporary allocations in the library.
* Note that all algorithms only allocate memory for temporary use.
* allocate/deallocate are always called in a stack-like order - last pointer to be allocated is deallocated first.
*/
MESHOPTIMIZER_API void meshopt_setAllocator(void* (*allocate)(size_t), void (*deallocate)(void*));
#ifdef __cplusplus
} /* extern "C" */
#endif
/* Quantization into commonly supported data formats */
#ifdef __cplusplus
/**
* Quantize a float in [0..1] range into an N-bit fixed point unorm value
* Assumes reconstruction function (q / (2^N-1)), which is the case for fixed-function normalized fixed point conversion
* Maximum reconstruction error: 1/2^(N+1)
*/
inline int meshopt_quantizeUnorm(float v, int N);
/**
* Quantize a float in [-1..1] range into an N-bit fixed point snorm value
* Assumes reconstruction function (q / (2^(N-1)-1)), which is the case for fixed-function normalized fixed point conversion (except early OpenGL versions)
* Maximum reconstruction error: 1/2^N
*/
inline int meshopt_quantizeSnorm(float v, int N);
/**
* Quantize a float into half-precision floating point value
* Generates +-inf for overflow, preserves NaN, flushes denormals to zero, rounds to nearest
* Representable magnitude range: [6e-5; 65504]
* Maximum relative reconstruction error: 5e-4
*/
inline unsigned short meshopt_quantizeHalf(float v);
/**
* Quantize a float into a floating point value with a limited number of significant mantissa bits
* Generates +-inf for overflow, preserves NaN, flushes denormals to zero, rounds to nearest
* Assumes N is in a valid mantissa precision range, which is 1..23
*/
inline float meshopt_quantizeFloat(float v, int N);
#endif
/**
* C++ template interface
*
* These functions mirror the C interface the library provides, providing template-based overloads so that
* the caller can use an arbitrary type for the index data, both for input and output.
* When the supplied type is the same size as that of unsigned int, the wrappers are zero-cost; when it's not,
* the wrappers end up allocating memory and copying index data to convert from one type to another.
*/
#if defined(__cplusplus) && !defined(MESHOPTIMIZER_NO_WRAPPERS)
template <typename T>
inline size_t meshopt_generateVertexRemap(unsigned int* destination, const T* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size);
template <typename T>
inline size_t meshopt_generateVertexRemapMulti(unsigned int* destination, const T* indices, size_t index_count, size_t vertex_count, const meshopt_Stream* streams, size_t stream_count);
template <typename T>
inline void meshopt_remapIndexBuffer(T* destination, const T* indices, size_t index_count, const unsigned int* remap);
template <typename T>
inline void meshopt_generateShadowIndexBuffer(T* destination, const T* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size, size_t vertex_stride);
template <typename T>
inline void meshopt_generateShadowIndexBufferMulti(T* destination, const T* indices, size_t index_count, size_t vertex_count, const meshopt_Stream* streams, size_t stream_count);
template <typename T>
inline void meshopt_optimizeVertexCache(T* destination, const T* indices, size_t index_count, size_t vertex_count);
template <typename T>
inline void meshopt_optimizeVertexCacheStrip(T* destination, const T* indices, size_t index_count, size_t vertex_count);
template <typename T>
inline void meshopt_optimizeVertexCacheFifo(T* destination, const T* indices, size_t index_count, size_t vertex_count, unsigned int cache_size);
template <typename T>
inline void meshopt_optimizeOverdraw(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, float threshold);
template <typename T>
inline size_t meshopt_optimizeVertexFetchRemap(unsigned int* destination, const T* indices, size_t index_count, size_t vertex_count);
template <typename T>
inline size_t meshopt_optimizeVertexFetch(void* destination, T* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size);
template <typename T>
inline size_t meshopt_encodeIndexBuffer(unsigned char* buffer, size_t buffer_size, const T* indices, size_t index_count);
template <typename T>
inline int meshopt_decodeIndexBuffer(T* destination, size_t index_count, const unsigned char* buffer, size_t buffer_size);
template <typename T>
inline size_t meshopt_encodeIndexSequence(unsigned char* buffer, size_t buffer_size, const T* indices, size_t index_count);
template <typename T>
inline int meshopt_decodeIndexSequence(T* destination, size_t index_count, const unsigned char* buffer, size_t buffer_size);
template <typename T>
inline size_t meshopt_simplify(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error);
template <typename T>
inline size_t meshopt_simplifySloppy(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count);
template <typename T>
inline size_t meshopt_stripify(T* destination, const T* indices, size_t index_count, size_t vertex_count, T restart_index);
template <typename T>
inline size_t meshopt_unstripify(T* destination, const T* indices, size_t index_count, T restart_index);
template <typename T>
inline meshopt_VertexCacheStatistics meshopt_analyzeVertexCache(const T* indices, size_t index_count, size_t vertex_count, unsigned int cache_size, unsigned int warp_size, unsigned int buffer_size);
template <typename T>
inline meshopt_OverdrawStatistics meshopt_analyzeOverdraw(const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
template <typename T>
inline meshopt_VertexFetchStatistics meshopt_analyzeVertexFetch(const T* indices, size_t index_count, size_t vertex_count, size_t vertex_size);
template <typename T>
inline size_t meshopt_buildMeshlets(meshopt_Meshlet* destination, const T* indices, size_t index_count, size_t vertex_count, size_t max_vertices, size_t max_triangles);
template <typename T>
inline meshopt_Bounds meshopt_computeClusterBounds(const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
template <typename T>
inline void meshopt_spatialSortTriangles(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride);
#endif
/* Inline implementation */
#ifdef __cplusplus
inline int meshopt_quantizeUnorm(float v, int N)
{
const float scale = float((1 << N) - 1);
v = (v >= 0) ? v : 0;
v = (v <= 1) ? v : 1;
return int(v * scale + 0.5f);
}
inline int meshopt_quantizeSnorm(float v, int N)
{
const float scale = float((1 << (N - 1)) - 1);
float round = (v >= 0 ? 0.5f : -0.5f);
v = (v >= -1) ? v : -1;
v = (v <= +1) ? v : +1;
return int(v * scale + round);
}
inline unsigned short meshopt_quantizeHalf(float v)
{
union { float f; unsigned int ui; } u = {v};
unsigned int ui = u.ui;
int s = (ui >> 16) & 0x8000;
int em = ui & 0x7fffffff;
/* bias exponent and round to nearest; 112 is relative exponent bias (127-15) */
int h = (em - (112 << 23) + (1 << 12)) >> 13;
/* underflow: flush to zero; 113 encodes exponent -14 */
h = (em < (113 << 23)) ? 0 : h;
/* overflow: infinity; 143 encodes exponent 16 */
h = (em >= (143 << 23)) ? 0x7c00 : h;
/* NaN; note that we convert all types of NaN to qNaN */
h = (em > (255 << 23)) ? 0x7e00 : h;
return (unsigned short)(s | h);
}
inline float meshopt_quantizeFloat(float v, int N)
{
union { float f; unsigned int ui; } u = {v};
unsigned int ui = u.ui;
const int mask = (1 << (23 - N)) - 1;
const int round = (1 << (23 - N)) >> 1;
int e = ui & 0x7f800000;
unsigned int rui = (ui + round) & ~mask;
/* round all numbers except inf/nan; this is important to make sure nan doesn't overflow into -0 */
ui = e == 0x7f800000 ? ui : rui;
/* flush denormals to zero */
ui = e == 0 ? 0 : ui;
u.ui = ui;
return u.f;
}
#endif
/* Internal implementation helpers */
#ifdef __cplusplus
class meshopt_Allocator
{
public:
template <typename T>
struct StorageT
{
static void* (*allocate)(size_t);
static void (*deallocate)(void*);
};
typedef StorageT<void> Storage;
meshopt_Allocator()
: blocks()
, count(0)
{
}
~meshopt_Allocator()
{
for (size_t i = count; i > 0; --i)
Storage::deallocate(blocks[i - 1]);
}
template <typename T> T* allocate(size_t size)
{
assert(count < sizeof(blocks) / sizeof(blocks[0]));
T* result = static_cast<T*>(Storage::allocate(size > size_t(-1) / sizeof(T) ? size_t(-1) : size * sizeof(T)));
blocks[count++] = result;
return result;
}
private:
void* blocks[16];
size_t count;
};
// This makes sure that allocate/deallocate are lazily generated in translation units that need them and are deduplicated by the linker
template <typename T> void* (*meshopt_Allocator::StorageT<T>::allocate)(size_t) = operator new;
template <typename T> void (*meshopt_Allocator::StorageT<T>::deallocate)(void*) = operator delete;
#endif
/* Inline implementation for C++ templated wrappers */
#if defined(__cplusplus) && !defined(MESHOPTIMIZER_NO_WRAPPERS)
template <typename T, bool ZeroCopy = sizeof(T) == sizeof(unsigned int)>
struct meshopt_IndexAdapter;
template <typename T>
struct meshopt_IndexAdapter<T, false>
{
T* result;
unsigned int* data;
size_t count;
meshopt_IndexAdapter(T* result_, const T* input, size_t count_)
: result(result_)
, data(0)
, count(count_)
{
size_t size = count > size_t(-1) / sizeof(unsigned int) ? size_t(-1) : count * sizeof(unsigned int);
data = static_cast<unsigned int*>(meshopt_Allocator::Storage::allocate(size));
if (input)
{
for (size_t i = 0; i < count; ++i)
data[i] = input[i];
}
}
~meshopt_IndexAdapter()
{
if (result)
{
for (size_t i = 0; i < count; ++i)
result[i] = T(data[i]);
}
meshopt_Allocator::Storage::deallocate(data);
}
};
template <typename T>
struct meshopt_IndexAdapter<T, true>
{
unsigned int* data;
meshopt_IndexAdapter(T* result, const T* input, size_t)
: data(reinterpret_cast<unsigned int*>(result ? result : const_cast<T*>(input)))
{
}
};
template <typename T>
inline size_t meshopt_generateVertexRemap(unsigned int* destination, const T* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size)
{
meshopt_IndexAdapter<T> in(0, indices, indices ? index_count : 0);
return meshopt_generateVertexRemap(destination, indices ? in.data : 0, index_count, vertices, vertex_count, vertex_size);
}
template <typename T>
inline size_t meshopt_generateVertexRemapMulti(unsigned int* destination, const T* indices, size_t index_count, size_t vertex_count, const meshopt_Stream* streams, size_t stream_count)
{
meshopt_IndexAdapter<T> in(0, indices, indices ? index_count : 0);
return meshopt_generateVertexRemapMulti(destination, indices ? in.data : 0, index_count, vertex_count, streams, stream_count);
}
template <typename T>
inline void meshopt_remapIndexBuffer(T* destination, const T* indices, size_t index_count, const unsigned int* remap)
{
meshopt_IndexAdapter<T> in(0, indices, indices ? index_count : 0);
meshopt_IndexAdapter<T> out(destination, 0, index_count);
meshopt_remapIndexBuffer(out.data, indices ? in.data : 0, index_count, remap);
}
template <typename T>
inline void meshopt_generateShadowIndexBuffer(T* destination, const T* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size, size_t vertex_stride)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
meshopt_IndexAdapter<T> out(destination, 0, index_count);
meshopt_generateShadowIndexBuffer(out.data, in.data, index_count, vertices, vertex_count, vertex_size, vertex_stride);
}
template <typename T>
inline void meshopt_generateShadowIndexBufferMulti(T* destination, const T* indices, size_t index_count, size_t vertex_count, const meshopt_Stream* streams, size_t stream_count)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
meshopt_IndexAdapter<T> out(destination, 0, index_count);
meshopt_generateShadowIndexBufferMulti(out.data, in.data, index_count, vertex_count, streams, stream_count);
}
template <typename T>
inline void meshopt_optimizeVertexCache(T* destination, const T* indices, size_t index_count, size_t vertex_count)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
meshopt_IndexAdapter<T> out(destination, 0, index_count);
meshopt_optimizeVertexCache(out.data, in.data, index_count, vertex_count);
}
template <typename T>
inline void meshopt_optimizeVertexCacheStrip(T* destination, const T* indices, size_t index_count, size_t vertex_count)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
meshopt_IndexAdapter<T> out(destination, 0, index_count);
meshopt_optimizeVertexCacheStrip(out.data, in.data, index_count, vertex_count);
}
template <typename T>
inline void meshopt_optimizeVertexCacheFifo(T* destination, const T* indices, size_t index_count, size_t vertex_count, unsigned int cache_size)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
meshopt_IndexAdapter<T> out(destination, 0, index_count);
meshopt_optimizeVertexCacheFifo(out.data, in.data, index_count, vertex_count, cache_size);
}
template <typename T>
inline void meshopt_optimizeOverdraw(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, float threshold)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
meshopt_IndexAdapter<T> out(destination, 0, index_count);
meshopt_optimizeOverdraw(out.data, in.data, index_count, vertex_positions, vertex_count, vertex_positions_stride, threshold);
}
template <typename T>
inline size_t meshopt_optimizeVertexFetchRemap(unsigned int* destination, const T* indices, size_t index_count, size_t vertex_count)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
return meshopt_optimizeVertexFetchRemap(destination, in.data, index_count, vertex_count);
}
template <typename T>
inline size_t meshopt_optimizeVertexFetch(void* destination, T* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size)
{
meshopt_IndexAdapter<T> inout(indices, indices, index_count);
return meshopt_optimizeVertexFetch(destination, inout.data, index_count, vertices, vertex_count, vertex_size);
}
template <typename T>
inline size_t meshopt_encodeIndexBuffer(unsigned char* buffer, size_t buffer_size, const T* indices, size_t index_count)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
return meshopt_encodeIndexBuffer(buffer, buffer_size, in.data, index_count);
}
template <typename T>
inline int meshopt_decodeIndexBuffer(T* destination, size_t index_count, const unsigned char* buffer, size_t buffer_size)
{
char index_size_valid[sizeof(T) == 2 || sizeof(T) == 4 ? 1 : -1];
(void)index_size_valid;
return meshopt_decodeIndexBuffer(destination, index_count, sizeof(T), buffer, buffer_size);
}
template <typename T>
inline size_t meshopt_encodeIndexSequence(unsigned char* buffer, size_t buffer_size, const T* indices, size_t index_count)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
return meshopt_encodeIndexSequence(buffer, buffer_size, in.data, index_count);
}
template <typename T>
inline int meshopt_decodeIndexSequence(T* destination, size_t index_count, const unsigned char* buffer, size_t buffer_size)
{
char index_size_valid[sizeof(T) == 2 || sizeof(T) == 4 ? 1 : -1];
(void)index_size_valid;
return meshopt_decodeIndexSequence(destination, index_count, sizeof(T), buffer, buffer_size);
}
template <typename T>
inline size_t meshopt_simplify(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count, float target_error)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
meshopt_IndexAdapter<T> out(destination, 0, index_count);
return meshopt_simplify(out.data, in.data, index_count, vertex_positions, vertex_count, vertex_positions_stride, target_index_count, target_error);
}
template <typename T>
inline size_t meshopt_simplifySloppy(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, size_t target_index_count)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
meshopt_IndexAdapter<T> out(destination, 0, target_index_count);
return meshopt_simplifySloppy(out.data, in.data, index_count, vertex_positions, vertex_count, vertex_positions_stride, target_index_count);
}
template <typename T>
inline size_t meshopt_stripify(T* destination, const T* indices, size_t index_count, size_t vertex_count, T restart_index)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
meshopt_IndexAdapter<T> out(destination, 0, (index_count / 3) * 5);
return meshopt_stripify(out.data, in.data, index_count, vertex_count, unsigned(restart_index));
}
template <typename T>
inline size_t meshopt_unstripify(T* destination, const T* indices, size_t index_count, T restart_index)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
meshopt_IndexAdapter<T> out(destination, 0, (index_count - 2) * 3);
return meshopt_unstripify(out.data, in.data, index_count, unsigned(restart_index));
}
template <typename T>
inline meshopt_VertexCacheStatistics meshopt_analyzeVertexCache(const T* indices, size_t index_count, size_t vertex_count, unsigned int cache_size, unsigned int warp_size, unsigned int buffer_size)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
return meshopt_analyzeVertexCache(in.data, index_count, vertex_count, cache_size, warp_size, buffer_size);
}
template <typename T>
inline meshopt_OverdrawStatistics meshopt_analyzeOverdraw(const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
return meshopt_analyzeOverdraw(in.data, index_count, vertex_positions, vertex_count, vertex_positions_stride);
}
template <typename T>
inline meshopt_VertexFetchStatistics meshopt_analyzeVertexFetch(const T* indices, size_t index_count, size_t vertex_count, size_t vertex_size)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
return meshopt_analyzeVertexFetch(in.data, index_count, vertex_count, vertex_size);
}
template <typename T>
inline size_t meshopt_buildMeshlets(meshopt_Meshlet* destination, const T* indices, size_t index_count, size_t vertex_count, size_t max_vertices, size_t max_triangles)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
return meshopt_buildMeshlets(destination, in.data, index_count, vertex_count, max_vertices, max_triangles);
}
template <typename T>
inline meshopt_Bounds meshopt_computeClusterBounds(const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
return meshopt_computeClusterBounds(in.data, index_count, vertex_positions, vertex_count, vertex_positions_stride);
}
template <typename T>
inline void meshopt_spatialSortTriangles(T* destination, const T* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
{
meshopt_IndexAdapter<T> in(0, indices, index_count);
meshopt_IndexAdapter<T> out(destination, 0, index_count);
meshopt_spatialSortTriangles(out.data, in.data, index_count, vertex_positions, vertex_count, vertex_positions_stride);
}
#endif
/**
* Copyright (c) 2016-2020 Arseny Kapoulkine
*
* Permission is hereby granted, free of charge, to any person
* obtaining a copy of this software and associated documentation
* files (the "Software"), to deal in the Software without
* restriction, including without limitation the rights to use,
* copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following
* conditions:
*
* The above copyright notice and this permission notice shall be
* included in all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
* OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
* HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
* WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
* OTHER DEALINGS IN THE SOFTWARE.
*/

230
include/meshoptimizer/overdrawanalyzer.cpp Normal file
View File

@ -0,0 +1,230 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
#include <assert.h>
#include <float.h>
#include <string.h>
// This work is based on:
// Nicolas Capens. Advanced Rasterization. 2004
namespace meshopt
{
const int kViewport = 256;
struct OverdrawBuffer
{
float z[kViewport][kViewport][2];
unsigned int overdraw[kViewport][kViewport][2];
};
#ifndef min
#define min(a, b) ((a) < (b) ? (a) : (b))
#endif
#ifndef max
#define max(a, b) ((a) > (b) ? (a) : (b))
#endif
static float computeDepthGradients(float& dzdx, float& dzdy, float x1, float y1, float z1, float x2, float y2, float z2, float x3, float y3, float z3)
{
// z2 = z1 + dzdx * (x2 - x1) + dzdy * (y2 - y1)
// z3 = z1 + dzdx * (x3 - x1) + dzdy * (y3 - y1)
// (x2-x1 y2-y1)(dzdx) = (z2-z1)
// (x3-x1 y3-y1)(dzdy) (z3-z1)
// we'll solve it with Cramer's rule
float det = (x2 - x1) * (y3 - y1) - (y2 - y1) * (x3 - x1);
float invdet = (det == 0) ? 0 : 1 / det;
dzdx = (z2 - z1) * (y3 - y1) - (y2 - y1) * (z3 - z1) * invdet;
dzdy = (x2 - x1) * (z3 - z1) - (z2 - z1) * (x3 - x1) * invdet;
return det;
}
// half-space fixed point triangle rasterizer
static void rasterize(OverdrawBuffer* buffer, float v1x, float v1y, float v1z, float v2x, float v2y, float v2z, float v3x, float v3y, float v3z)
{
// compute depth gradients
float DZx, DZy;
float det = computeDepthGradients(DZx, DZy, v1x, v1y, v1z, v2x, v2y, v2z, v3x, v3y, v3z);
int sign = det > 0;
// flip backfacing triangles to simplify rasterization logic
if (sign)
{
// flipping v2 & v3 preserves depth gradients since they're based on v1
float t;
t = v2x, v2x = v3x, v3x = t;
t = v2y, v2y = v3y, v3y = t;
t = v2z, v2z = v3z, v3z = t;
// flip depth since we rasterize backfacing triangles to second buffer with reverse Z; only v1z is used below
v1z = kViewport - v1z;
DZx = -DZx;
DZy = -DZy;
}
// coordinates, 28.4 fixed point
int X1 = int(16.0f * v1x + 0.5f);
int X2 = int(16.0f * v2x + 0.5f);
int X3 = int(16.0f * v3x + 0.5f);
int Y1 = int(16.0f * v1y + 0.5f);
int Y2 = int(16.0f * v2y + 0.5f);
int Y3 = int(16.0f * v3y + 0.5f);
// bounding rectangle, clipped against viewport
// since we rasterize pixels with covered centers, min >0.5 should round up
// as for max, due to top-left filling convention we will never rasterize right/bottom edges
// so max >= 0.5 should round down
int minx = max((min(X1, min(X2, X3)) + 7) >> 4, 0);
int maxx = min((max(X1, max(X2, X3)) + 7) >> 4, kViewport);
int miny = max((min(Y1, min(Y2, Y3)) + 7) >> 4, 0);
int maxy = min((max(Y1, max(Y2, Y3)) + 7) >> 4, kViewport);
// deltas, 28.4 fixed point
int DX12 = X1 - X2;
int DX23 = X2 - X3;
int DX31 = X3 - X1;
int DY12 = Y1 - Y2;
int DY23 = Y2 - Y3;
int DY31 = Y3 - Y1;
// fill convention correction
int TL1 = DY12 < 0 || (DY12 == 0 && DX12 > 0);
int TL2 = DY23 < 0 || (DY23 == 0 && DX23 > 0);
int TL3 = DY31 < 0 || (DY31 == 0 && DX31 > 0);
// half edge equations, 24.8 fixed point
// note that we offset minx/miny by half pixel since we want to rasterize pixels with covered centers
int FX = (minx << 4) + 8;
int FY = (miny << 4) + 8;
int CY1 = DX12 * (FY - Y1) - DY12 * (FX - X1) + TL1 - 1;
int CY2 = DX23 * (FY - Y2) - DY23 * (FX - X2) + TL2 - 1;
int CY3 = DX31 * (FY - Y3) - DY31 * (FX - X3) + TL3 - 1;
float ZY = v1z + (DZx * float(FX - X1) + DZy * float(FY - Y1)) * (1 / 16.f);
for (int y = miny; y < maxy; y++)
{
int CX1 = CY1;
int CX2 = CY2;
int CX3 = CY3;
float ZX = ZY;
for (int x = minx; x < maxx; x++)
{
// check if all CXn are non-negative
if ((CX1 | CX2 | CX3) >= 0)
{
if (ZX >= buffer->z[y][x][sign])
{
buffer->z[y][x][sign] = ZX;
buffer->overdraw[y][x][sign]++;
}
}
// signed left shift is UB for negative numbers so use unsigned-signed casts
CX1 -= int(unsigned(DY12) << 4);
CX2 -= int(unsigned(DY23) << 4);
CX3 -= int(unsigned(DY31) << 4);
ZX += DZx;
}
// signed left shift is UB for negative numbers so use unsigned-signed casts
CY1 += int(unsigned(DX12) << 4);
CY2 += int(unsigned(DX23) << 4);
CY3 += int(unsigned(DX31) << 4);
ZY += DZy;
}
}
} // namespace meshopt
meshopt_OverdrawStatistics meshopt_analyzeOverdraw(const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
{
using namespace meshopt;
assert(index_count % 3 == 0);
assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
assert(vertex_positions_stride % sizeof(float) == 0);
meshopt_Allocator allocator;
size_t vertex_stride_float = vertex_positions_stride / sizeof(float);
meshopt_OverdrawStatistics result = {};
float minv[3] = {FLT_MAX, FLT_MAX, FLT_MAX};
float maxv[3] = {-FLT_MAX, -FLT_MAX, -FLT_MAX};
for (size_t i = 0; i < vertex_count; ++i)
{
const float* v = vertex_positions + i * vertex_stride_float;
for (int j = 0; j < 3; ++j)
{
minv[j] = min(minv[j], v[j]);
maxv[j] = max(maxv[j], v[j]);
}
}
float extent = max(maxv[0] - minv[0], max(maxv[1] - minv[1], maxv[2] - minv[2]));
float scale = kViewport / extent;
float* triangles = allocator.allocate<float>(index_count * 3);
for (size_t i = 0; i < index_count; ++i)
{
unsigned int index = indices[i];
assert(index < vertex_count);
const float* v = vertex_positions + index * vertex_stride_float;
triangles[i * 3 + 0] = (v[0] - minv[0]) * scale;
triangles[i * 3 + 1] = (v[1] - minv[1]) * scale;
triangles[i * 3 + 2] = (v[2] - minv[2]) * scale;
}
OverdrawBuffer* buffer = allocator.allocate<OverdrawBuffer>(1);
for (int axis = 0; axis < 3; ++axis)
{
memset(buffer, 0, sizeof(OverdrawBuffer));
for (size_t i = 0; i < index_count; i += 3)
{
const float* vn0 = &triangles[3 * (i + 0)];
const float* vn1 = &triangles[3 * (i + 1)];
const float* vn2 = &triangles[3 * (i + 2)];
switch (axis)
{
case 0:
rasterize(buffer, vn0[2], vn0[1], vn0[0], vn1[2], vn1[1], vn1[0], vn2[2], vn2[1], vn2[0]);
break;
case 1:
rasterize(buffer, vn0[0], vn0[2], vn0[1], vn1[0], vn1[2], vn1[1], vn2[0], vn2[2], vn2[1]);
break;
case 2:
rasterize(buffer, vn0[1], vn0[0], vn0[2], vn1[1], vn1[0], vn1[2], vn2[1], vn2[0], vn2[2]);
break;
}
}
for (int y = 0; y < kViewport; ++y)
for (int x = 0; x < kViewport; ++x)
for (int s = 0; s < 2; ++s)
{
unsigned int overdraw = buffer->overdraw[y][x][s];
result.pixels_covered += overdraw > 0;
result.pixels_shaded += overdraw;
}
}
result.overdraw = result.pixels_covered ? float(result.pixels_shaded) / float(result.pixels_covered) : 0.f;
return result;
}

333
include/meshoptimizer/overdrawoptimizer.cpp Normal file
View File

@ -0,0 +1,333 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
#include <assert.h>
#include <math.h>
#include <string.h>
// This work is based on:
// Pedro Sander, Diego Nehab and Joshua Barczak. Fast Triangle Reordering for Vertex Locality and Reduced Overdraw. 2007
namespace meshopt
{
static void calculateSortData(float* sort_data, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_positions_stride, const unsigned int* clusters, size_t cluster_count)
{
size_t vertex_stride_float = vertex_positions_stride / sizeof(float);
float mesh_centroid[3] = {};
for (size_t i = 0; i < index_count; ++i)
{
const float* p = vertex_positions + vertex_stride_float * indices[i];
mesh_centroid[0] += p[0];
mesh_centroid[1] += p[1];
mesh_centroid[2] += p[2];
}
mesh_centroid[0] /= index_count;
mesh_centroid[1] /= index_count;
mesh_centroid[2] /= index_count;
for (size_t cluster = 0; cluster < cluster_count; ++cluster)
{
size_t cluster_begin = clusters[cluster] * 3;
size_t cluster_end = (cluster + 1 < cluster_count) ? clusters[cluster + 1] * 3 : index_count;
assert(cluster_begin < cluster_end);
float cluster_area = 0;
float cluster_centroid[3] = {};
float cluster_normal[3] = {};
for (size_t i = cluster_begin; i < cluster_end; i += 3)
{
const float* p0 = vertex_positions + vertex_stride_float * indices[i + 0];
const float* p1 = vertex_positions + vertex_stride_float * indices[i + 1];
const float* p2 = vertex_positions + vertex_stride_float * indices[i + 2];
float p10[3] = {p1[0] - p0[0], p1[1] - p0[1], p1[2] - p0[2]};
float p20[3] = {p2[0] - p0[0], p2[1] - p0[1], p2[2] - p0[2]};
float normalx = p10[1] * p20[2] - p10[2] * p20[1];
float normaly = p10[2] * p20[0] - p10[0] * p20[2];
float normalz = p10[0] * p20[1] - p10[1] * p20[0];
float area = sqrtf(normalx * normalx + normaly * normaly + normalz * normalz);
cluster_centroid[0] += (p0[0] + p1[0] + p2[0]) * (area / 3);
cluster_centroid[1] += (p0[1] + p1[1] + p2[1]) * (area / 3);
cluster_centroid[2] += (p0[2] + p1[2] + p2[2]) * (area / 3);
cluster_normal[0] += normalx;
cluster_normal[1] += normaly;
cluster_normal[2] += normalz;
cluster_area += area;
}
float inv_cluster_area = cluster_area == 0 ? 0 : 1 / cluster_area;
cluster_centroid[0] *= inv_cluster_area;
cluster_centroid[1] *= inv_cluster_area;
cluster_centroid[2] *= inv_cluster_area;
float cluster_normal_length = sqrtf(cluster_normal[0] * cluster_normal[0] + cluster_normal[1] * cluster_normal[1] + cluster_normal[2] * cluster_normal[2]);
float inv_cluster_normal_length = cluster_normal_length == 0 ? 0 : 1 / cluster_normal_length;
cluster_normal[0] *= inv_cluster_normal_length;
cluster_normal[1] *= inv_cluster_normal_length;
cluster_normal[2] *= inv_cluster_normal_length;
float centroid_vector[3] = {cluster_centroid[0] - mesh_centroid[0], cluster_centroid[1] - mesh_centroid[1], cluster_centroid[2] - mesh_centroid[2]};
sort_data[cluster] = centroid_vector[0] * cluster_normal[0] + centroid_vector[1] * cluster_normal[1] + centroid_vector[2] * cluster_normal[2];
}
}
static void calculateSortOrderRadix(unsigned int* sort_order, const float* sort_data, unsigned short* sort_keys, size_t cluster_count)
{
// compute sort data bounds and renormalize, using fixed point snorm
float sort_data_max = 1e-3f;
for (size_t i = 0; i < cluster_count; ++i)
{
float dpa = fabsf(sort_data[i]);
sort_data_max = (sort_data_max < dpa) ? dpa : sort_data_max;
}
const int sort_bits = 11;
for (size_t i = 0; i < cluster_count; ++i)
{
// note that we flip distribution since high dot product should come first
float sort_key = 0.5f - 0.5f * (sort_data[i] / sort_data_max);
sort_keys[i] = meshopt_quantizeUnorm(sort_key, sort_bits) & ((1 << sort_bits) - 1);
}
// fill histogram for counting sort
unsigned int histogram[1 << sort_bits];
memset(histogram, 0, sizeof(histogram));
for (size_t i = 0; i < cluster_count; ++i)
{
histogram[sort_keys[i]]++;
}
// compute offsets based on histogram data
size_t histogram_sum = 0;
for (size_t i = 0; i < 1 << sort_bits; ++i)
{
size_t count = histogram[i];
histogram[i] = unsigned(histogram_sum);
histogram_sum += count;
}
assert(histogram_sum == cluster_count);
// compute sort order based on offsets
for (size_t i = 0; i < cluster_count; ++i)
{
sort_order[histogram[sort_keys[i]]++] = unsigned(i);
}
}
static unsigned int updateCache(unsigned int a, unsigned int b, unsigned int c, unsigned int cache_size, unsigned int* cache_timestamps, unsigned int& timestamp)
{
unsigned int cache_misses = 0;
// if vertex is not in cache, put it in cache
if (timestamp - cache_timestamps[a] > cache_size)
{
cache_timestamps[a] = timestamp++;
cache_misses++;
}
if (timestamp - cache_timestamps[b] > cache_size)
{
cache_timestamps[b] = timestamp++;
cache_misses++;
}
if (timestamp - cache_timestamps[c] > cache_size)
{
cache_timestamps[c] = timestamp++;
cache_misses++;
}
return cache_misses;
}
static size_t generateHardBoundaries(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int cache_size, unsigned int* cache_timestamps)
{
memset(cache_timestamps, 0, vertex_count * sizeof(unsigned int));
unsigned int timestamp = cache_size + 1;
size_t face_count = index_count / 3;
size_t result = 0;
for (size_t i = 0; i < face_count; ++i)
{
unsigned int m = updateCache(indices[i * 3 + 0], indices[i * 3 + 1], indices[i * 3 + 2], cache_size, &cache_timestamps[0], timestamp);
// when all three vertices are not in the cache it's usually relatively safe to assume that this is a new patch in the mesh
// that is disjoint from previous vertices; sometimes it might come back to reference existing vertices but that frequently
// suggests an inefficiency in the vertex cache optimization algorithm
// usually the first triangle has 3 misses unless it's degenerate - thus we make sure the first cluster always starts with 0
if (i == 0 || m == 3)
{
destination[result++] = unsigned(i);
}
}
assert(result <= index_count / 3);
return result;
}
static size_t generateSoftBoundaries(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, const unsigned int* clusters, size_t cluster_count, unsigned int cache_size, float threshold, unsigned int* cache_timestamps)
{
memset(cache_timestamps, 0, vertex_count * sizeof(unsigned int));
unsigned int timestamp = 0;
size_t result = 0;
for (size_t it = 0; it < cluster_count; ++it)
{
size_t start = clusters[it];
size_t end = (it + 1 < cluster_count) ? clusters[it + 1] : index_count / 3;
assert(start < end);
// reset cache
timestamp += cache_size + 1;
// measure cluster ACMR
unsigned int cluster_misses = 0;
for (size_t i = start; i < end; ++i)
{
unsigned int m = updateCache(indices[i * 3 + 0], indices[i * 3 + 1], indices[i * 3 + 2], cache_size, &cache_timestamps[0], timestamp);
cluster_misses += m;
}
float cluster_threshold = threshold * (float(cluster_misses) / float(end - start));
// first cluster always starts from the hard cluster boundary
destination[result++] = unsigned(start);
// reset cache
timestamp += cache_size + 1;
unsigned int running_misses = 0;
unsigned int running_faces = 0;
for (size_t i = start; i < end; ++i)
{
unsigned int m = updateCache(indices[i * 3 + 0], indices[i * 3 + 1], indices[i * 3 + 2], cache_size, &cache_timestamps[0], timestamp);
running_misses += m;
running_faces += 1;
if (float(running_misses) / float(running_faces) <= cluster_threshold)
{
// we have reached the target ACMR with the current triangle so we need to start a new cluster on the next one
// note that this may mean that we add 'end` to destination for the last triangle, which will imply that the last
// cluster is empty; however, the 'pop_back' after the loop will clean it up
destination[result++] = unsigned(i + 1);
// reset cache
timestamp += cache_size + 1;
running_misses = 0;
running_faces = 0;
}
}
// each time we reach the target ACMR we flush the cluster
// this means that the last cluster is by definition not very good - there are frequent cases where we are left with a few triangles
// in the last cluster, producing a very bad ACMR and significantly penalizing the overall results
// thus we remove the last cluster boundary, merging the last complete cluster with the last incomplete one
// there are sometimes cases when the last cluster is actually good enough - in which case the code above would have added 'end'
// to the cluster boundary array which we need to remove anyway - this code will do that automatically
if (destination[result - 1] != start)
{
result--;
}
}
assert(result >= cluster_count);
assert(result <= index_count / 3);
return result;
}
} // namespace meshopt
void meshopt_optimizeOverdraw(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride, float threshold)
{
using namespace meshopt;
assert(index_count % 3 == 0);
assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
assert(vertex_positions_stride % sizeof(float) == 0);
meshopt_Allocator allocator;
// guard for empty meshes
if (index_count == 0 || vertex_count == 0)
return;
// support in-place optimization
if (destination == indices)
{
unsigned int* indices_copy = allocator.allocate<unsigned int>(index_count);
memcpy(indices_copy, indices, index_count * sizeof(unsigned int));
indices = indices_copy;
}
unsigned int cache_size = 16;
unsigned int* cache_timestamps = allocator.allocate<unsigned int>(vertex_count);
// generate hard boundaries from full-triangle cache misses
unsigned int* hard_clusters = allocator.allocate<unsigned int>(index_count / 3);
size_t hard_cluster_count = generateHardBoundaries(hard_clusters, indices, index_count, vertex_count, cache_size, cache_timestamps);
// generate soft boundaries
unsigned int* soft_clusters = allocator.allocate<unsigned int>(index_count / 3 + 1);
size_t soft_cluster_count = generateSoftBoundaries(soft_clusters, indices, index_count, vertex_count, hard_clusters, hard_cluster_count, cache_size, threshold, cache_timestamps);
const unsigned int* clusters = soft_clusters;
size_t cluster_count = soft_cluster_count;
// fill sort data
float* sort_data = allocator.allocate<float>(cluster_count);
calculateSortData(sort_data, indices, index_count, vertex_positions, vertex_positions_stride, clusters, cluster_count);
// sort clusters using sort data
unsigned short* sort_keys = allocator.allocate<unsigned short>(cluster_count);
unsigned int* sort_order = allocator.allocate<unsigned int>(cluster_count);
calculateSortOrderRadix(sort_order, sort_data, sort_keys, cluster_count);
// fill output buffer
size_t offset = 0;
for (size_t it = 0; it < cluster_count; ++it)
{
unsigned int cluster = sort_order[it];
assert(cluster < cluster_count);
size_t cluster_begin = clusters[cluster] * 3;
size_t cluster_end = (cluster + 1 < cluster_count) ? clusters[cluster + 1] * 3 : index_count;
assert(cluster_begin < cluster_end);
memcpy(destination + offset, indices + cluster_begin, (cluster_end - cluster_begin) * sizeof(unsigned int));
offset += cluster_end - cluster_begin;
}
assert(offset == index_count);
}

1546
include/meshoptimizer/simplifier.cpp Normal file
View File

File diff suppressed because it is too large Load Diff

194
include/meshoptimizer/spatialorder.cpp Normal file
View File

@ -0,0 +1,194 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
#include <assert.h>
#include <float.h>
#include <string.h>
// This work is based on:
// Fabian Giesen. Decoding Morton codes. 2009
namespace meshopt
{
// "Insert" two 0 bits after each of the 10 low bits of x
inline unsigned int part1By2(unsigned int x)
{
x &= 0x000003ff; // x = ---- ---- ---- ---- ---- --98 7654 3210
x = (x ^ (x << 16)) & 0xff0000ff; // x = ---- --98 ---- ---- ---- ---- 7654 3210
x = (x ^ (x << 8)) & 0x0300f00f; // x = ---- --98 ---- ---- 7654 ---- ---- 3210
x = (x ^ (x << 4)) & 0x030c30c3; // x = ---- --98 ---- 76-- --54 ---- 32-- --10
x = (x ^ (x << 2)) & 0x09249249; // x = ---- 9--8 --7- -6-- 5--4 --3- -2-- 1--0
return x;
}
static void computeOrder(unsigned int* result, const float* vertex_positions_data, size_t vertex_count, size_t vertex_positions_stride)
{
size_t vertex_stride_float = vertex_positions_stride / sizeof(float);
float minv[3] = {FLT_MAX, FLT_MAX, FLT_MAX};
float maxv[3] = {-FLT_MAX, -FLT_MAX, -FLT_MAX};
for (size_t i = 0; i < vertex_count; ++i)
{
const float* v = vertex_positions_data + i * vertex_stride_float;
for (int j = 0; j < 3; ++j)
{
float vj = v[j];
minv[j] = minv[j] > vj ? vj : minv[j];
maxv[j] = maxv[j] < vj ? vj : maxv[j];
}
}
float extent = 0.f;
extent = (maxv[0] - minv[0]) < extent ? extent : (maxv[0] - minv[0]);
extent = (maxv[1] - minv[1]) < extent ? extent : (maxv[1] - minv[1]);
extent = (maxv[2] - minv[2]) < extent ? extent : (maxv[2] - minv[2]);
float scale = extent == 0 ? 0.f : 1.f / extent;
// generate Morton order based on the position inside a unit cube
for (size_t i = 0; i < vertex_count; ++i)
{
const float* v = vertex_positions_data + i * vertex_stride_float;
int x = int((v[0] - minv[0]) * scale * 1023.f + 0.5f);
int y = int((v[1] - minv[1]) * scale * 1023.f + 0.5f);
int z = int((v[2] - minv[2]) * scale * 1023.f + 0.5f);
result[i] = part1By2(x) | (part1By2(y) << 1) | (part1By2(z) << 2);
}
}
static void computeHistogram(unsigned int (&hist)[1024][3], const unsigned int* data, size_t count)
{
memset(hist, 0, sizeof(hist));
// compute 3 10-bit histograms in parallel
for (size_t i = 0; i < count; ++i)
{
unsigned int id = data[i];
hist[(id >> 0) & 1023][0]++;
hist[(id >> 10) & 1023][1]++;
hist[(id >> 20) & 1023][2]++;
}
unsigned int sumx = 0, sumy = 0, sumz = 0;
// replace histogram data with prefix histogram sums in-place
for (int i = 0; i < 1024; ++i)
{
unsigned int hx = hist[i][0], hy = hist[i][1], hz = hist[i][2];
hist[i][0] = sumx;
hist[i][1] = sumy;
hist[i][2] = sumz;
sumx += hx;
sumy += hy;
sumz += hz;
}
assert(sumx == count && sumy == count && sumz == count);
}
static void radixPass(unsigned int* destination, const unsigned int* source, const unsigned int* keys, size_t count, unsigned int (&hist)[1024][3], int pass)
{
int bitoff = pass * 10;
for (size_t i = 0; i < count; ++i)
{
unsigned int id = (keys[source[i]] >> bitoff) & 1023;
destination[hist[id][pass]++] = source[i];
}
}
} // namespace meshopt
void meshopt_spatialSortRemap(unsigned int* destination, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
{
using namespace meshopt;
assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
assert(vertex_positions_stride % sizeof(float) == 0);
meshopt_Allocator allocator;
unsigned int* keys = allocator.allocate<unsigned int>(vertex_count);
computeOrder(keys, vertex_positions, vertex_count, vertex_positions_stride);
unsigned int hist[1024][3];
computeHistogram(hist, keys, vertex_count);
unsigned int* scratch = allocator.allocate<unsigned int>(vertex_count);
for (size_t i = 0; i < vertex_count; ++i)
destination[i] = unsigned(i);
// 3-pass radix sort computes the resulting order into scratch
radixPass(scratch, destination, keys, vertex_count, hist, 0);
radixPass(destination, scratch, keys, vertex_count, hist, 1);
radixPass(scratch, destination, keys, vertex_count, hist, 2);
// since our remap table is mapping old=>new, we need to reverse it
for (size_t i = 0; i < vertex_count; ++i)
destination[scratch[i]] = unsigned(i);
}
void meshopt_spatialSortTriangles(unsigned int* destination, const unsigned int* indices, size_t index_count, const float* vertex_positions, size_t vertex_count, size_t vertex_positions_stride)
{
using namespace meshopt;
assert(index_count % 3 == 0);
assert(vertex_positions_stride > 0 && vertex_positions_stride <= 256);
assert(vertex_positions_stride % sizeof(float) == 0);
(void)vertex_count;
size_t face_count = index_count / 3;
size_t vertex_stride_float = vertex_positions_stride / sizeof(float);
meshopt_Allocator allocator;
float* centroids = allocator.allocate<float>(face_count * 3);
for (size_t i = 0; i < face_count; ++i)
{
unsigned int a = indices[i * 3 + 0], b = indices[i * 3 + 1], c = indices[i * 3 + 2];
assert(a < vertex_count && b < vertex_count && c < vertex_count);
const float* va = vertex_positions + a * vertex_stride_float;
const float* vb = vertex_positions + b * vertex_stride_float;
const float* vc = vertex_positions + c * vertex_stride_float;
centroids[i * 3 + 0] = (va[0] + vb[0] + vc[0]) / 3.f;
centroids[i * 3 + 1] = (va[1] + vb[1] + vc[1]) / 3.f;
centroids[i * 3 + 2] = (va[2] + vb[2] + vc[2]) / 3.f;
}
unsigned int* remap = allocator.allocate<unsigned int>(face_count);
meshopt_spatialSortRemap(remap, centroids, face_count, sizeof(float) * 3);
// support in-order remap
if (destination == indices)
{
unsigned int* indices_copy = allocator.allocate<unsigned int>(index_count);
memcpy(indices_copy, indices, index_count * sizeof(unsigned int));
indices = indices_copy;
}
for (size_t i = 0; i < face_count; ++i)
{
unsigned int a = indices[i * 3 + 0], b = indices[i * 3 + 1], c = indices[i * 3 + 2];
unsigned int r = remap[i];
destination[r * 3 + 0] = a;
destination[r * 3 + 1] = b;
destination[r * 3 + 2] = c;
}
}

295
include/meshoptimizer/stripifier.cpp Normal file
View File

@ -0,0 +1,295 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
#include <assert.h>
#include <limits.h>
#include <string.h>
// This work is based on:
// Francine Evans, Steven Skiena and Amitabh Varshney. Optimizing Triangle Strips for Fast Rendering. 1996
namespace meshopt
{
static unsigned int findStripFirst(const unsigned int buffer[][3], unsigned int buffer_size, const unsigned int* valence)
{
unsigned int index = 0;
unsigned int iv = ~0u;
for (size_t i = 0; i < buffer_size; ++i)
{
unsigned int va = valence[buffer[i][0]], vb = valence[buffer[i][1]], vc = valence[buffer[i][2]];
unsigned int v = (va < vb && va < vc) ? va : (vb < vc) ? vb : vc;
if (v < iv)
{
index = unsigned(i);
iv = v;
}
}
return index;
}
static int findStripNext(const unsigned int buffer[][3], unsigned int buffer_size, unsigned int e0, unsigned int e1)
{
for (size_t i = 0; i < buffer_size; ++i)
{
unsigned int a = buffer[i][0], b = buffer[i][1], c = buffer[i][2];
if (e0 == a && e1 == b)
return (int(i) << 2) | 2;
else if (e0 == b && e1 == c)
return (int(i) << 2) | 0;
else if (e0 == c && e1 == a)
return (int(i) << 2) | 1;
}
return -1;
}
} // namespace meshopt
size_t meshopt_stripify(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int restart_index)
{
assert(destination != indices);
assert(index_count % 3 == 0);
using namespace meshopt;
meshopt_Allocator allocator;
const size_t buffer_capacity = 8;
unsigned int buffer[buffer_capacity][3] = {};
unsigned int buffer_size = 0;
size_t index_offset = 0;
unsigned int strip[2] = {};
unsigned int parity = 0;
size_t strip_size = 0;
// compute vertex valence; this is used to prioritize starting triangle for strips
unsigned int* valence = allocator.allocate<unsigned int>(vertex_count);
memset(valence, 0, vertex_count * sizeof(unsigned int));
for (size_t i = 0; i < index_count; ++i)
{
unsigned int index = indices[i];
assert(index < vertex_count);
valence[index]++;
}
int next = -1;
while (buffer_size > 0 || index_offset < index_count)
{
assert(next < 0 || (size_t(next >> 2) < buffer_size && (next & 3) < 3));
// fill triangle buffer
while (buffer_size < buffer_capacity && index_offset < index_count)
{
buffer[buffer_size][0] = indices[index_offset + 0];
buffer[buffer_size][1] = indices[index_offset + 1];
buffer[buffer_size][2] = indices[index_offset + 2];
buffer_size++;
index_offset += 3;
}
assert(buffer_size > 0);
if (next >= 0)
{
unsigned int i = next >> 2;
unsigned int a = buffer[i][0], b = buffer[i][1], c = buffer[i][2];
unsigned int v = buffer[i][next & 3];
// ordered removal from the buffer
memmove(buffer[i], buffer[i + 1], (buffer_size - i - 1) * sizeof(buffer[0]));
buffer_size--;
// update vertex valences for strip start heuristic
valence[a]--;
valence[b]--;
valence[c]--;
// find next triangle (note that edge order flips on every iteration)
// in some cases we need to perform a swap to pick a different outgoing triangle edge
// for [a b c], the default strip edge is [b c], but we might want to use [a c]
int cont = findStripNext(buffer, buffer_size, parity ? strip[1] : v, parity ? v : strip[1]);
int swap = cont < 0 ? findStripNext(buffer, buffer_size, parity ? v : strip[0], parity ? strip[0] : v) : -1;
if (cont < 0 && swap >= 0)
{
// [a b c] => [a b a c]
destination[strip_size++] = strip[0];
destination[strip_size++] = v;
// next strip has same winding
// ? a b => b a v
strip[1] = v;
next = swap;
}
else
{
// emit the next vertex in the strip
destination[strip_size++] = v;
// next strip has flipped winding
strip[0] = strip[1];
strip[1] = v;
parity ^= 1;
next = cont;
}
}
else
{
// if we didn't find anything, we need to find the next new triangle
// we use a heuristic to maximize the strip length
unsigned int i = findStripFirst(buffer, buffer_size, &valence[0]);
unsigned int a = buffer[i][0], b = buffer[i][1], c = buffer[i][2];
// ordered removal from the buffer
memmove(buffer[i], buffer[i + 1], (buffer_size - i - 1) * sizeof(buffer[0]));
buffer_size--;
// update vertex valences for strip start heuristic
valence[a]--;
valence[b]--;
valence[c]--;
// we need to pre-rotate the triangle so that we will find a match in the existing buffer on the next iteration
int ea = findStripNext(buffer, buffer_size, c, b);
int eb = findStripNext(buffer, buffer_size, a, c);
int ec = findStripNext(buffer, buffer_size, b, a);
// in some cases we can have several matching edges; since we can pick any edge, we pick the one with the smallest
// triangle index in the buffer. this reduces the effect of stripification on ACMR and additionally - for unclear
// reasons - slightly improves the stripification efficiency
int mine = INT_MAX;
mine = (ea >= 0 && mine > ea) ? ea : mine;
mine = (eb >= 0 && mine > eb) ? eb : mine;
mine = (ec >= 0 && mine > ec) ? ec : mine;
if (ea == mine)
{
// keep abc
next = ea;
}
else if (eb == mine)
{
// abc -> bca
unsigned int t = a;
a = b, b = c, c = t;
next = eb;
}
else if (ec == mine)
{
// abc -> cab
unsigned int t = c;
c = b, b = a, a = t;
next = ec;
}
if (restart_index)
{
if (strip_size)
destination[strip_size++] = restart_index;
destination[strip_size++] = a;
destination[strip_size++] = b;
destination[strip_size++] = c;
// new strip always starts with the same edge winding
strip[0] = b;
strip[1] = c;
parity = 1;
}
else
{
if (strip_size)
{
// connect last strip using degenerate triangles
destination[strip_size++] = strip[1];
destination[strip_size++] = a;
}
// note that we may need to flip the emitted triangle based on parity
// we always end up with outgoing edge "cb" in the end
unsigned int e0 = parity ? c : b;
unsigned int e1 = parity ? b : c;
destination[strip_size++] = a;
destination[strip_size++] = e0;
destination[strip_size++] = e1;
strip[0] = e0;
strip[1] = e1;
parity ^= 1;
}
}
}
return strip_size;
}
size_t meshopt_stripifyBound(size_t index_count)
{
assert(index_count % 3 == 0);
// worst case without restarts is 2 degenerate indices and 3 indices per triangle
// worst case with restarts is 1 restart index and 3 indices per triangle
return (index_count / 3) * 5;
}
size_t meshopt_unstripify(unsigned int* destination, const unsigned int* indices, size_t index_count, unsigned int restart_index)
{
assert(destination != indices);
size_t offset = 0;
size_t start = 0;
for (size_t i = 0; i < index_count; ++i)
{
if (restart_index && indices[i] == restart_index)
{
start = i + 1;
}
else if (i - start >= 2)
{
unsigned int a = indices[i - 2], b = indices[i - 1], c = indices[i];
// flip winding for odd triangles
if ((i - start) & 1)
{
unsigned int t = a;
a = b, b = t;
}
// although we use restart indices, strip swaps still produce degenerate triangles, so skip them
if (a != b && a != c && b != c)
{
destination[offset + 0] = a;
destination[offset + 1] = b;
destination[offset + 2] = c;
offset += 3;
}
}
}
return offset;
}
size_t meshopt_unstripifyBound(size_t index_count)
{
assert(index_count == 0 || index_count >= 3);
return (index_count == 0) ? 0 : (index_count - 2) * 3;
}

73
include/meshoptimizer/vcacheanalyzer.cpp Normal file
View File

@ -0,0 +1,73 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
#include <assert.h>
#include <string.h>
meshopt_VertexCacheStatistics meshopt_analyzeVertexCache(const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int cache_size, unsigned int warp_size, unsigned int primgroup_size)
{
assert(index_count % 3 == 0);
assert(cache_size >= 3);
assert(warp_size == 0 || warp_size >= 3);
meshopt_Allocator allocator;
meshopt_VertexCacheStatistics result = {};
unsigned int warp_offset = 0;
unsigned int primgroup_offset = 0;
unsigned int* cache_timestamps = allocator.allocate<unsigned int>(vertex_count);
memset(cache_timestamps, 0, vertex_count * sizeof(unsigned int));
unsigned int timestamp = cache_size + 1;
for (size_t i = 0; i < index_count; i += 3)
{
unsigned int a = indices[i + 0], b = indices[i + 1], c = indices[i + 2];
assert(a < vertex_count && b < vertex_count && c < vertex_count);
bool ac = (timestamp - cache_timestamps[a]) > cache_size;
bool bc = (timestamp - cache_timestamps[b]) > cache_size;
bool cc = (timestamp - cache_timestamps[c]) > cache_size;
// flush cache if triangle doesn't fit into warp or into the primitive buffer
if ((primgroup_size && primgroup_offset == primgroup_size) || (warp_size && warp_offset + ac + bc + cc > warp_size))
{
result.warps_executed += warp_offset > 0;
warp_offset = 0;
primgroup_offset = 0;
// reset cache
timestamp += cache_size + 1;
}
// update cache and add vertices to warp
for (int j = 0; j < 3; ++j)
{
unsigned int index = indices[i + j];
if (timestamp - cache_timestamps[index] > cache_size)
{
cache_timestamps[index] = timestamp++;
result.vertices_transformed++;
warp_offset++;
}
}
primgroup_offset++;
}
size_t unique_vertex_count = 0;
for (size_t i = 0; i < vertex_count; ++i)
unique_vertex_count += cache_timestamps[i] > 0;
result.warps_executed += warp_offset > 0;
result.acmr = index_count == 0 ? 0 : float(result.vertices_transformed) / float(index_count / 3);
result.atvr = unique_vertex_count == 0 ? 0 : float(result.vertices_transformed) / float(unique_vertex_count);
return result;
}

473
include/meshoptimizer/vcacheoptimizer.cpp Normal file
View File

@ -0,0 +1,473 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
#include <assert.h>
#include <string.h>
// This work is based on:
// Tom Forsyth. Linear-Speed Vertex Cache Optimisation. 2006
// Pedro Sander, Diego Nehab and Joshua Barczak. Fast Triangle Reordering for Vertex Locality and Reduced Overdraw. 2007
namespace meshopt
{
const size_t kCacheSizeMax = 16;
const size_t kValenceMax = 8;
struct VertexScoreTable
{
float cache[1 + kCacheSizeMax];
float live[1 + kValenceMax];
};
// Tuned to minimize the ACMR of a GPU that has a cache profile similar to NVidia and AMD
static const VertexScoreTable kVertexScoreTable = {
{0.f, 0.779f, 0.791f, 0.789f, 0.981f, 0.843f, 0.726f, 0.847f, 0.882f, 0.867f, 0.799f, 0.642f, 0.613f, 0.600f, 0.568f, 0.372f, 0.234f},
{0.f, 0.995f, 0.713f, 0.450f, 0.404f, 0.059f, 0.005f, 0.147f, 0.006f},
};
// Tuned to minimize the encoded index buffer size
static const VertexScoreTable kVertexScoreTableStrip = {
{0.f, 1.000f, 1.000f, 1.000f, 0.453f, 0.561f, 0.490f, 0.459f, 0.179f, 0.526f, 0.000f, 0.227f, 0.184f, 0.490f, 0.112f, 0.050f, 0.131f},
{0.f, 0.956f, 0.786f, 0.577f, 0.558f, 0.618f, 0.549f, 0.499f, 0.489f},
};
struct TriangleAdjacency
{
unsigned int* counts;
unsigned int* offsets;
unsigned int* data;
};
static void buildTriangleAdjacency(TriangleAdjacency& adjacency, const unsigned int* indices, size_t index_count, size_t vertex_count, meshopt_Allocator& allocator)
{
size_t face_count = index_count / 3;
// allocate arrays
adjacency.counts = allocator.allocate<unsigned int>(vertex_count);
adjacency.offsets = allocator.allocate<unsigned int>(vertex_count);
adjacency.data = allocator.allocate<unsigned int>(index_count);
// fill triangle counts
memset(adjacency.counts, 0, vertex_count * sizeof(unsigned int));
for (size_t i = 0; i < index_count; ++i)
{
assert(indices[i] < vertex_count);
adjacency.counts[indices[i]]++;
}
// fill offset table
unsigned int offset = 0;
for (size_t i = 0; i < vertex_count; ++i)
{
adjacency.offsets[i] = offset;
offset += adjacency.counts[i];
}
assert(offset == index_count);
// fill triangle data
for (size_t i = 0; i < face_count; ++i)
{
unsigned int a = indices[i * 3 + 0], b = indices[i * 3 + 1], c = indices[i * 3 + 2];
adjacency.data[adjacency.offsets[a]++] = unsigned(i);
adjacency.data[adjacency.offsets[b]++] = unsigned(i);
adjacency.data[adjacency.offsets[c]++] = unsigned(i);
}
// fix offsets that have been disturbed by the previous pass
for (size_t i = 0; i < vertex_count; ++i)
{
assert(adjacency.offsets[i] >= adjacency.counts[i]);
adjacency.offsets[i] -= adjacency.counts[i];
}
}
static unsigned int getNextVertexDeadEnd(const unsigned int* dead_end, unsigned int& dead_end_top, unsigned int& input_cursor, const unsigned int* live_triangles, size_t vertex_count)
{
// check dead-end stack
while (dead_end_top)
{
unsigned int vertex = dead_end[--dead_end_top];
if (live_triangles[vertex] > 0)
return vertex;
}
// input order
while (input_cursor < vertex_count)
{
if (live_triangles[input_cursor] > 0)
return input_cursor;
++input_cursor;
}
return ~0u;
}
static unsigned int getNextVertexNeighbour(const unsigned int* next_candidates_begin, const unsigned int* next_candidates_end, const unsigned int* live_triangles, const unsigned int* cache_timestamps, unsigned int timestamp, unsigned int cache_size)
{
unsigned int best_candidate = ~0u;
int best_priority = -1;
for (const unsigned int* next_candidate = next_candidates_begin; next_candidate != next_candidates_end; ++next_candidate)
{
unsigned int vertex = *next_candidate;
// otherwise we don't need to process it
if (live_triangles[vertex] > 0)
{
int priority = 0;
// will it be in cache after fanning?
if (2 * live_triangles[vertex] + timestamp - cache_timestamps[vertex] <= cache_size)
{
priority = timestamp - cache_timestamps[vertex]; // position in cache
}
if (priority > best_priority)
{
best_candidate = vertex;
best_priority = priority;
}
}
}
return best_candidate;
}
static float vertexScore(const VertexScoreTable* table, int cache_position, unsigned int live_triangles)
{
assert(cache_position >= -1 && cache_position < int(kCacheSizeMax));
unsigned int live_triangles_clamped = live_triangles < kValenceMax ? live_triangles : kValenceMax;
return table->cache[1 + cache_position] + table->live[live_triangles_clamped];
}
static unsigned int getNextTriangleDeadEnd(unsigned int& input_cursor, const unsigned char* emitted_flags, size_t face_count)
{
// input order
while (input_cursor < face_count)
{
if (!emitted_flags[input_cursor])
return input_cursor;
++input_cursor;
}
return ~0u;
}
} // namespace meshopt
void meshopt_optimizeVertexCacheTable(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, const meshopt::VertexScoreTable* table)
{
using namespace meshopt;
assert(index_count % 3 == 0);
meshopt_Allocator allocator;
// guard for empty meshes
if (index_count == 0 || vertex_count == 0)
return;
// support in-place optimization
if (destination == indices)
{
unsigned int* indices_copy = allocator.allocate<unsigned int>(index_count);
memcpy(indices_copy, indices, index_count * sizeof(unsigned int));
indices = indices_copy;
}
unsigned int cache_size = 16;
assert(cache_size <= kCacheSizeMax);
size_t face_count = index_count / 3;
// build adjacency information
TriangleAdjacency adjacency = {};
buildTriangleAdjacency(adjacency, indices, index_count, vertex_count, allocator);
// live triangle counts
unsigned int* live_triangles = allocator.allocate<unsigned int>(vertex_count);
memcpy(live_triangles, adjacency.counts, vertex_count * sizeof(unsigned int));
// emitted flags
unsigned char* emitted_flags = allocator.allocate<unsigned char>(face_count);
memset(emitted_flags, 0, face_count);
// compute initial vertex scores
float* vertex_scores = allocator.allocate<float>(vertex_count);
for (size_t i = 0; i < vertex_count; ++i)
vertex_scores[i] = vertexScore(table, -1, live_triangles[i]);
// compute triangle scores
float* triangle_scores = allocator.allocate<float>(face_count);
for (size_t i = 0; i < face_count; ++i)
{
unsigned int a = indices[i * 3 + 0];
unsigned int b = indices[i * 3 + 1];
unsigned int c = indices[i * 3 + 2];
triangle_scores[i] = vertex_scores[a] + vertex_scores[b] + vertex_scores[c];
}
unsigned int cache_holder[2 * (kCacheSizeMax + 3)];
unsigned int* cache = cache_holder;
unsigned int* cache_new = cache_holder + kCacheSizeMax + 3;
size_t cache_count = 0;
unsigned int current_triangle = 0;
unsigned int input_cursor = 1;
unsigned int output_triangle = 0;
while (current_triangle != ~0u)
{
assert(output_triangle < face_count);
unsigned int a = indices[current_triangle * 3 + 0];
unsigned int b = indices[current_triangle * 3 + 1];
unsigned int c = indices[current_triangle * 3 + 2];
// output indices
destination[output_triangle * 3 + 0] = a;
destination[output_triangle * 3 + 1] = b;
destination[output_triangle * 3 + 2] = c;
output_triangle++;
// update emitted flags
emitted_flags[current_triangle] = true;
triangle_scores[current_triangle] = 0;
// new triangle
size_t cache_write = 0;
cache_new[cache_write++] = a;
cache_new[cache_write++] = b;
cache_new[cache_write++] = c;
// old triangles
for (size_t i = 0; i < cache_count; ++i)
{
unsigned int index = cache[i];
if (index != a && index != b && index != c)
{
cache_new[cache_write++] = index;
}
}
unsigned int* cache_temp = cache;
cache = cache_new, cache_new = cache_temp;
cache_count = cache_write > cache_size ? cache_size : cache_write;
// update live triangle counts
live_triangles[a]--;
live_triangles[b]--;
live_triangles[c]--;
// remove emitted triangle from adjacency data
// this makes sure that we spend less time traversing these lists on subsequent iterations
for (size_t k = 0; k < 3; ++k)
{
unsigned int index = indices[current_triangle * 3 + k];
unsigned int* neighbours = &adjacency.data[0] + adjacency.offsets[index];
size_t neighbours_size = adjacency.counts[index];
for (size_t i = 0; i < neighbours_size; ++i)
{
unsigned int tri = neighbours[i];
if (tri == current_triangle)
{
neighbours[i] = neighbours[neighbours_size - 1];
adjacency.counts[index]--;
break;
}
}
}
unsigned int best_triangle = ~0u;
float best_score = 0;
// update cache positions, vertex scores and triangle scores, and find next best triangle
for (size_t i = 0; i < cache_write; ++i)
{
unsigned int index = cache[i];
int cache_position = i >= cache_size ? -1 : int(i);
// update vertex score
float score = vertexScore(table, cache_position, live_triangles[index]);
float score_diff = score - vertex_scores[index];
vertex_scores[index] = score;
// update scores of vertex triangles
const unsigned int* neighbours_begin = &adjacency.data[0] + adjacency.offsets[index];
const unsigned int* neighbours_end = neighbours_begin + adjacency.counts[index];
for (const unsigned int* it = neighbours_begin; it != neighbours_end; ++it)
{
unsigned int tri = *it;
assert(!emitted_flags[tri]);
float tri_score = triangle_scores[tri] + score_diff;
assert(tri_score > 0);
if (best_score < tri_score)
{
best_triangle = tri;
best_score = tri_score;
}
triangle_scores[tri] = tri_score;
}
}
// step through input triangles in order if we hit a dead-end
current_triangle = best_triangle;
if (current_triangle == ~0u)
{
current_triangle = getNextTriangleDeadEnd(input_cursor, &emitted_flags[0], face_count);
}
}
assert(input_cursor == face_count);
assert(output_triangle == face_count);
}
void meshopt_optimizeVertexCache(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count)
{
meshopt_optimizeVertexCacheTable(destination, indices, index_count, vertex_count, &meshopt::kVertexScoreTable);
}
void meshopt_optimizeVertexCacheStrip(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count)
{
meshopt_optimizeVertexCacheTable(destination, indices, index_count, vertex_count, &meshopt::kVertexScoreTableStrip);
}
void meshopt_optimizeVertexCacheFifo(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count, unsigned int cache_size)
{
using namespace meshopt;
assert(index_count % 3 == 0);
assert(cache_size >= 3);
meshopt_Allocator allocator;
// guard for empty meshes
if (index_count == 0 || vertex_count == 0)
return;
// support in-place optimization
if (destination == indices)
{
unsigned int* indices_copy = allocator.allocate<unsigned int>(index_count);
memcpy(indices_copy, indices, index_count * sizeof(unsigned int));
indices = indices_copy;
}
size_t face_count = index_count / 3;
// build adjacency information
TriangleAdjacency adjacency = {};
buildTriangleAdjacency(adjacency, indices, index_count, vertex_count, allocator);
// live triangle counts
unsigned int* live_triangles = allocator.allocate<unsigned int>(vertex_count);
memcpy(live_triangles, adjacency.counts, vertex_count * sizeof(unsigned int));
// cache time stamps
unsigned int* cache_timestamps = allocator.allocate<unsigned int>(vertex_count);
memset(cache_timestamps, 0, vertex_count * sizeof(unsigned int));
// dead-end stack
unsigned int* dead_end = allocator.allocate<unsigned int>(index_count);
unsigned int dead_end_top = 0;
// emitted flags
unsigned char* emitted_flags = allocator.allocate<unsigned char>(face_count);
memset(emitted_flags, 0, face_count);
unsigned int current_vertex = 0;
unsigned int timestamp = cache_size + 1;
unsigned int input_cursor = 1; // vertex to restart from in case of dead-end
unsigned int output_triangle = 0;
while (current_vertex != ~0u)
{
const unsigned int* next_candidates_begin = &dead_end[0] + dead_end_top;
// emit all vertex neighbours
const unsigned int* neighbours_begin = &adjacency.data[0] + adjacency.offsets[current_vertex];
const unsigned int* neighbours_end = neighbours_begin + adjacency.counts[current_vertex];
for (const unsigned int* it = neighbours_begin; it != neighbours_end; ++it)
{
unsigned int triangle = *it;
if (!emitted_flags[triangle])
{
unsigned int a = indices[triangle * 3 + 0], b = indices[triangle * 3 + 1], c = indices[triangle * 3 + 2];
// output indices
destination[output_triangle * 3 + 0] = a;
destination[output_triangle * 3 + 1] = b;
destination[output_triangle * 3 + 2] = c;
output_triangle++;
// update dead-end stack
dead_end[dead_end_top + 0] = a;
dead_end[dead_end_top + 1] = b;
dead_end[dead_end_top + 2] = c;
dead_end_top += 3;
// update live triangle counts
live_triangles[a]--;
live_triangles[b]--;
live_triangles[c]--;
// update cache info
// if vertex is not in cache, put it in cache
if (timestamp - cache_timestamps[a] > cache_size)
cache_timestamps[a] = timestamp++;
if (timestamp - cache_timestamps[b] > cache_size)
cache_timestamps[b] = timestamp++;
if (timestamp - cache_timestamps[c] > cache_size)
cache_timestamps[c] = timestamp++;
// update emitted flags
emitted_flags[triangle] = true;
}
}
// next candidates are the ones we pushed to dead-end stack just now
const unsigned int* next_candidates_end = &dead_end[0] + dead_end_top;
// get next vertex
current_vertex = getNextVertexNeighbour(next_candidates_begin, next_candidates_end, &live_triangles[0], &cache_timestamps[0], timestamp, cache_size);
if (current_vertex == ~0u)
{
current_vertex = getNextVertexDeadEnd(&dead_end[0], dead_end_top, input_cursor, &live_triangles[0], vertex_count);
}
}
assert(output_triangle == face_count);
}

1283
include/meshoptimizer/vertexcodec.cpp Normal file
View File

File diff suppressed because it is too large Load Diff

821
include/meshoptimizer/vertexfilter.cpp Normal file
View File

@ -0,0 +1,821 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
#include <math.h>
// The block below auto-detects SIMD ISA that can be used on the target platform
#ifndef MESHOPTIMIZER_NO_SIMD
// The SIMD implementation requires SSE2, which can be enabled unconditionally through compiler settings
#if defined(__SSE2__)
#define SIMD_SSE
#endif
// MSVC supports compiling SSE2 code regardless of compile options; we assume all 32-bit CPUs support SSE2
#if !defined(SIMD_SSE) && defined(_MSC_VER) && !defined(__clang__) && (defined(_M_IX86) || defined(_M_X64))
#define SIMD_SSE
#endif
// GCC/clang define these when NEON support is available
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
#define SIMD_NEON
#endif
// On MSVC, we assume that ARM builds always target NEON-capable devices
#if !defined(SIMD_NEON) && defined(_MSC_VER) && (defined(_M_ARM) || defined(_M_ARM64))
#define SIMD_NEON
#endif
// When targeting Wasm SIMD we can't use runtime cpuid checks so we unconditionally enable SIMD
#if defined(__wasm_simd128__)
#define SIMD_WASM
#endif
#endif // !MESHOPTIMIZER_NO_SIMD
#ifdef SIMD_SSE
#include <emmintrin.h>
#include <stdint.h>
#endif
#ifdef _MSC_VER
#include <intrin.h>
#endif
#ifdef SIMD_NEON
#if defined(_MSC_VER) && defined(_M_ARM64)
#include <arm64_neon.h>
#else
#include <arm_neon.h>
#endif
#endif
#ifdef SIMD_WASM
#include <wasm_simd128.h>
#endif
#ifdef SIMD_WASM
#define wasmx_unpacklo_v16x8(a, b) wasm_v16x8_shuffle(a, b, 0, 8, 1, 9, 2, 10, 3, 11)
#define wasmx_unpackhi_v16x8(a, b) wasm_v16x8_shuffle(a, b, 4, 12, 5, 13, 6, 14, 7, 15)
#define wasmx_unziplo_v32x4(a, b) wasm_v32x4_shuffle(a, b, 0, 2, 4, 6)
#define wasmx_unziphi_v32x4(a, b) wasm_v32x4_shuffle(a, b, 1, 3, 5, 7)
#endif
namespace meshopt
{
#if !defined(SIMD_SSE) && !defined(SIMD_NEON) && !defined(SIMD_WASM)
template <typename T>
static void decodeFilterOct(T* data, size_t count)
{
const float max = float((1 << (sizeof(T) * 8 - 1)) - 1);
for (size_t i = 0; i < count; ++i)
{
// convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
float x = float(data[i * 4 + 0]);
float y = float(data[i * 4 + 1]);
float z = float(data[i * 4 + 2]) - fabsf(x) - fabsf(y);
// fixup octahedral coordinates for z<0
float t = (z >= 0.f) ? 0.f : z;
x += (x >= 0.f) ? t : -t;
y += (y >= 0.f) ? t : -t;
// compute normal length & scale
float l = sqrtf(x * x + y * y + z * z);
float s = max / l;
// rounded signed float->int
int xf = int(x * s + (x >= 0.f ? 0.5f : -0.5f));
int yf = int(y * s + (y >= 0.f ? 0.5f : -0.5f));
int zf = int(z * s + (z >= 0.f ? 0.5f : -0.5f));
data[i * 4 + 0] = T(xf);
data[i * 4 + 1] = T(yf);
data[i * 4 + 2] = T(zf);
}
}
static void decodeFilterQuat(short* data, size_t count)
{
const float scale = 1.f / sqrtf(2.f);
for (size_t i = 0; i < count; ++i)
{
// recover scale from the high byte of the component
int sf = data[i * 4 + 3] | 3;
float ss = scale / float(sf);
// convert x/y/z to [-1..1] (scaled...)
float x = float(data[i * 4 + 0]) * ss;
float y = float(data[i * 4 + 1]) * ss;
float z = float(data[i * 4 + 2]) * ss;
// reconstruct w as a square root; we clamp to 0.f to avoid NaN due to precision errors
float ww = 1.f - x * x - y * y - z * z;
float w = sqrtf(ww >= 0.f ? ww : 0.f);
// rounded signed float->int
int xf = int(x * 32767.f + (x >= 0.f ? 0.5f : -0.5f));
int yf = int(y * 32767.f + (y >= 0.f ? 0.5f : -0.5f));
int zf = int(z * 32767.f + (z >= 0.f ? 0.5f : -0.5f));
int wf = int(w * 32767.f + 0.5f);
int qc = data[i * 4 + 3] & 3;
// output order is dictated by input index
data[i * 4 + ((qc + 1) & 3)] = short(xf);
data[i * 4 + ((qc + 2) & 3)] = short(yf);
data[i * 4 + ((qc + 3) & 3)] = short(zf);
data[i * 4 + ((qc + 0) & 3)] = short(wf);
}
}
static void decodeFilterExp(unsigned int* data, size_t count)
{
for (size_t i = 0; i < count; ++i)
{
unsigned int v = data[i];
// decode mantissa and exponent
int m = int(v << 8) >> 8;
int e = int(v) >> 24;
union {
float f;
unsigned int ui;
} u;
// optimized version of ldexp(float(m), e)
u.ui = unsigned(e + 127) << 23;
u.f = u.f * float(m);
data[i] = u.ui;
}
}
#endif
#if defined(SIMD_SSE) || defined(SIMD_NEON) || defined(SIMD_WASM)
inline uint64_t rotateleft64(uint64_t v, int x)
{
#if defined(_MSC_VER) && !defined(__clang__)
return _rotl64(v, x);
#elif defined(__clang__) && __clang_major__ >= 8
return __builtin_rotateleft64(v, x);
#else
return (v << (x & 63)) | (v >> ((64 - x) & 63));
#endif
}
#endif
#ifdef SIMD_SSE
static void decodeFilterOctSimd(signed char* data, size_t count)
{
const __m128 sign = _mm_set1_ps(-0.f);
for (size_t i = 0; i < count; i += 4)
{
__m128i n4 = _mm_loadu_si128(reinterpret_cast<__m128i*>(&data[i * 4]));
// sign-extends each of x,y in [x y ? ?] with arithmetic shifts
__m128i xf = _mm_srai_epi32(_mm_slli_epi32(n4, 24), 24);
__m128i yf = _mm_srai_epi32(_mm_slli_epi32(n4, 16), 24);
// unpack z; note that z is unsigned so we technically don't need to sign extend it
__m128i zf = _mm_srai_epi32(_mm_slli_epi32(n4, 8), 24);
// convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
__m128 x = _mm_cvtepi32_ps(xf);
__m128 y = _mm_cvtepi32_ps(yf);
__m128 z = _mm_sub_ps(_mm_cvtepi32_ps(zf), _mm_add_ps(_mm_andnot_ps(sign, x), _mm_andnot_ps(sign, y)));
// fixup octahedral coordinates for z<0
__m128 t = _mm_min_ps(z, _mm_setzero_ps());
x = _mm_add_ps(x, _mm_xor_ps(t, _mm_and_ps(x, sign)));
y = _mm_add_ps(y, _mm_xor_ps(t, _mm_and_ps(y, sign)));
// compute normal length & scale
__m128 ll = _mm_add_ps(_mm_mul_ps(x, x), _mm_add_ps(_mm_mul_ps(y, y), _mm_mul_ps(z, z)));
__m128 s = _mm_mul_ps(_mm_set1_ps(127.f), _mm_rsqrt_ps(ll));
// rounded signed float->int
__m128i xr = _mm_cvtps_epi32(_mm_mul_ps(x, s));
__m128i yr = _mm_cvtps_epi32(_mm_mul_ps(y, s));
__m128i zr = _mm_cvtps_epi32(_mm_mul_ps(z, s));
// combine xr/yr/zr into final value
__m128i res = _mm_and_si128(n4, _mm_set1_epi32(0xff000000));
res = _mm_or_si128(res, _mm_and_si128(xr, _mm_set1_epi32(0xff)));
res = _mm_or_si128(res, _mm_slli_epi32(_mm_and_si128(yr, _mm_set1_epi32(0xff)), 8));
res = _mm_or_si128(res, _mm_slli_epi32(_mm_and_si128(zr, _mm_set1_epi32(0xff)), 16));
_mm_storeu_si128(reinterpret_cast<__m128i*>(&data[i * 4]), res);
}
}
static void decodeFilterOctSimd(short* data, size_t count)
{
const __m128 sign = _mm_set1_ps(-0.f);
for (size_t i = 0; i < count; i += 4)
{
__m128 n4_0 = _mm_loadu_ps(reinterpret_cast<float*>(&data[(i + 0) * 4]));
__m128 n4_1 = _mm_loadu_ps(reinterpret_cast<float*>(&data[(i + 2) * 4]));
// gather both x/y 16-bit pairs in each 32-bit lane
__m128i n4 = _mm_castps_si128(_mm_shuffle_ps(n4_0, n4_1, _MM_SHUFFLE(2, 0, 2, 0)));
// sign-extends each of x,y in [x y] with arithmetic shifts
__m128i xf = _mm_srai_epi32(_mm_slli_epi32(n4, 16), 16);
__m128i yf = _mm_srai_epi32(n4, 16);
// unpack z; note that z is unsigned so we don't need to sign extend it
__m128i z4 = _mm_castps_si128(_mm_shuffle_ps(n4_0, n4_1, _MM_SHUFFLE(3, 1, 3, 1)));
__m128i zf = _mm_and_si128(z4, _mm_set1_epi32(0x7fff));
// convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
__m128 x = _mm_cvtepi32_ps(xf);
__m128 y = _mm_cvtepi32_ps(yf);
__m128 z = _mm_sub_ps(_mm_cvtepi32_ps(zf), _mm_add_ps(_mm_andnot_ps(sign, x), _mm_andnot_ps(sign, y)));
// fixup octahedral coordinates for z<0
__m128 t = _mm_min_ps(z, _mm_setzero_ps());
x = _mm_add_ps(x, _mm_xor_ps(t, _mm_and_ps(x, sign)));
y = _mm_add_ps(y, _mm_xor_ps(t, _mm_and_ps(y, sign)));
// compute normal length & scale
__m128 ll = _mm_add_ps(_mm_mul_ps(x, x), _mm_add_ps(_mm_mul_ps(y, y), _mm_mul_ps(z, z)));
__m128 s = _mm_div_ps(_mm_set1_ps(32767.f), _mm_sqrt_ps(ll));
// rounded signed float->int
__m128i xr = _mm_cvtps_epi32(_mm_mul_ps(x, s));
__m128i yr = _mm_cvtps_epi32(_mm_mul_ps(y, s));
__m128i zr = _mm_cvtps_epi32(_mm_mul_ps(z, s));
// mix x/z and y/0 to make 16-bit unpack easier
__m128i xzr = _mm_or_si128(_mm_and_si128(xr, _mm_set1_epi32(0xffff)), _mm_slli_epi32(zr, 16));
__m128i y0r = _mm_and_si128(yr, _mm_set1_epi32(0xffff));
// pack x/y/z using 16-bit unpacks; note that this has 0 where we should have .w
__m128i res_0 = _mm_unpacklo_epi16(xzr, y0r);
__m128i res_1 = _mm_unpackhi_epi16(xzr, y0r);
// patch in .w
res_0 = _mm_or_si128(res_0, _mm_and_si128(_mm_castps_si128(n4_0), _mm_set1_epi64x(0xffff000000000000)));
res_1 = _mm_or_si128(res_1, _mm_and_si128(_mm_castps_si128(n4_1), _mm_set1_epi64x(0xffff000000000000)));
_mm_storeu_si128(reinterpret_cast<__m128i*>(&data[(i + 0) * 4]), res_0);
_mm_storeu_si128(reinterpret_cast<__m128i*>(&data[(i + 2) * 4]), res_1);
}
}
static void decodeFilterQuatSimd(short* data, size_t count)
{
const float scale = 1.f / sqrtf(2.f);
for (size_t i = 0; i < count; i += 4)
{
__m128 q4_0 = _mm_loadu_ps(reinterpret_cast<float*>(&data[(i + 0) * 4]));
__m128 q4_1 = _mm_loadu_ps(reinterpret_cast<float*>(&data[(i + 2) * 4]));
// gather both x/y 16-bit pairs in each 32-bit lane
__m128i q4_xy = _mm_castps_si128(_mm_shuffle_ps(q4_0, q4_1, _MM_SHUFFLE(2, 0, 2, 0)));
__m128i q4_zc = _mm_castps_si128(_mm_shuffle_ps(q4_0, q4_1, _MM_SHUFFLE(3, 1, 3, 1)));
// sign-extends each of x,y in [x y] with arithmetic shifts
__m128i xf = _mm_srai_epi32(_mm_slli_epi32(q4_xy, 16), 16);
__m128i yf = _mm_srai_epi32(q4_xy, 16);
__m128i zf = _mm_srai_epi32(_mm_slli_epi32(q4_zc, 16), 16);
__m128i cf = _mm_srai_epi32(q4_zc, 16);
// get a floating-point scaler using zc with bottom 2 bits set to 1 (which represents 1.f)
__m128i sf = _mm_or_si128(cf, _mm_set1_epi32(3));
__m128 ss = _mm_div_ps(_mm_set1_ps(scale), _mm_cvtepi32_ps(sf));
// convert x/y/z to [-1..1] (scaled...)
__m128 x = _mm_mul_ps(_mm_cvtepi32_ps(xf), ss);
__m128 y = _mm_mul_ps(_mm_cvtepi32_ps(yf), ss);
__m128 z = _mm_mul_ps(_mm_cvtepi32_ps(zf), ss);
// reconstruct w as a square root; we clamp to 0.f to avoid NaN due to precision errors
__m128 ww = _mm_sub_ps(_mm_set1_ps(1.f), _mm_add_ps(_mm_mul_ps(x, x), _mm_add_ps(_mm_mul_ps(y, y), _mm_mul_ps(z, z))));
__m128 w = _mm_sqrt_ps(_mm_max_ps(ww, _mm_setzero_ps()));
__m128 s = _mm_set1_ps(32767.f);
// rounded signed float->int
__m128i xr = _mm_cvtps_epi32(_mm_mul_ps(x, s));
__m128i yr = _mm_cvtps_epi32(_mm_mul_ps(y, s));
__m128i zr = _mm_cvtps_epi32(_mm_mul_ps(z, s));
__m128i wr = _mm_cvtps_epi32(_mm_mul_ps(w, s));
// mix x/z and w/y to make 16-bit unpack easier
__m128i xzr = _mm_or_si128(_mm_and_si128(xr, _mm_set1_epi32(0xffff)), _mm_slli_epi32(zr, 16));
__m128i wyr = _mm_or_si128(_mm_and_si128(wr, _mm_set1_epi32(0xffff)), _mm_slli_epi32(yr, 16));
// pack x/y/z/w using 16-bit unpacks; we pack wxyz by default (for qc=0)
__m128i res_0 = _mm_unpacklo_epi16(wyr, xzr);
__m128i res_1 = _mm_unpackhi_epi16(wyr, xzr);
// store results to stack so that we can rotate using scalar instructions
uint64_t res[4];
_mm_storeu_si128(reinterpret_cast<__m128i*>(&res[0]), res_0);
_mm_storeu_si128(reinterpret_cast<__m128i*>(&res[2]), res_1);
// rotate and store
uint64_t* out = reinterpret_cast<uint64_t*>(&data[i * 4]);
out[0] = rotateleft64(res[0], data[(i + 0) * 4 + 3] << 4);
out[1] = rotateleft64(res[1], data[(i + 1) * 4 + 3] << 4);
out[2] = rotateleft64(res[2], data[(i + 2) * 4 + 3] << 4);
out[3] = rotateleft64(res[3], data[(i + 3) * 4 + 3] << 4);
}
}
static void decodeFilterExpSimd(unsigned int* data, size_t count)
{
for (size_t i = 0; i < count; i += 4)
{
__m128i v = _mm_loadu_si128(reinterpret_cast<__m128i*>(&data[i]));
// decode exponent into 2^x directly
__m128i ef = _mm_srai_epi32(v, 24);
__m128i es = _mm_slli_epi32(_mm_add_epi32(ef, _mm_set1_epi32(127)), 23);
// decode 24-bit mantissa into floating-point value
__m128i mf = _mm_srai_epi32(_mm_slli_epi32(v, 8), 8);
__m128 m = _mm_cvtepi32_ps(mf);
__m128 r = _mm_mul_ps(_mm_castsi128_ps(es), m);
_mm_storeu_ps(reinterpret_cast<float*>(&data[i]), r);
}
}
#endif
#if defined(SIMD_NEON) && !defined(__aarch64__) && !defined(_M_ARM64)
inline float32x4_t vsqrtq_f32(float32x4_t x)
{
float32x4_t r = vrsqrteq_f32(x);
r = vmulq_f32(r, vrsqrtsq_f32(vmulq_f32(r, x), r)); // refine rsqrt estimate
return vmulq_f32(r, x);
}
inline float32x4_t vdivq_f32(float32x4_t x, float32x4_t y)
{
float32x4_t r = vrecpeq_f32(y);
r = vmulq_f32(r, vrecpsq_f32(y, r)); // refine rcp estimate
return vmulq_f32(x, r);
}
#endif
#ifdef SIMD_NEON
static void decodeFilterOctSimd(signed char* data, size_t count)
{
const int32x4_t sign = vdupq_n_s32(0x80000000);
for (size_t i = 0; i < count; i += 4)
{
int32x4_t n4 = vld1q_s32(reinterpret_cast<int32_t*>(&data[i * 4]));
// sign-extends each of x,y in [x y ? ?] with arithmetic shifts
int32x4_t xf = vshrq_n_s32(vshlq_n_s32(n4, 24), 24);
int32x4_t yf = vshrq_n_s32(vshlq_n_s32(n4, 16), 24);
// unpack z; note that z is unsigned so we technically don't need to sign extend it
int32x4_t zf = vshrq_n_s32(vshlq_n_s32(n4, 8), 24);
// convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
float32x4_t x = vcvtq_f32_s32(xf);
float32x4_t y = vcvtq_f32_s32(yf);
float32x4_t z = vsubq_f32(vcvtq_f32_s32(zf), vaddq_f32(vabsq_f32(x), vabsq_f32(y)));
// fixup octahedral coordinates for z<0
float32x4_t t = vminq_f32(z, vdupq_n_f32(0.f));
x = vaddq_f32(x, vreinterpretq_f32_s32(veorq_s32(vreinterpretq_s32_f32(t), vandq_s32(vreinterpretq_s32_f32(x), sign))));
y = vaddq_f32(y, vreinterpretq_f32_s32(veorq_s32(vreinterpretq_s32_f32(t), vandq_s32(vreinterpretq_s32_f32(y), sign))));
// compute normal length & scale
float32x4_t ll = vaddq_f32(vmulq_f32(x, x), vaddq_f32(vmulq_f32(y, y), vmulq_f32(z, z)));
float32x4_t rl = vrsqrteq_f32(ll);
float32x4_t s = vmulq_f32(vdupq_n_f32(127.f), rl);
// fast rounded signed float->int: addition triggers renormalization after which mantissa stores the integer value
// note: the result is offset by 0x4B40_0000, but we only need the low 16 bits so we can omit the subtraction
const float32x4_t fsnap = vdupq_n_f32(3 << 22);
int32x4_t xr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(x, s), fsnap));
int32x4_t yr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(y, s), fsnap));
int32x4_t zr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(z, s), fsnap));
// combine xr/yr/zr into final value
int32x4_t res = vandq_s32(n4, vdupq_n_s32(0xff000000));
res = vorrq_s32(res, vandq_s32(xr, vdupq_n_s32(0xff)));
res = vorrq_s32(res, vshlq_n_s32(vandq_s32(yr, vdupq_n_s32(0xff)), 8));
res = vorrq_s32(res, vshlq_n_s32(vandq_s32(zr, vdupq_n_s32(0xff)), 16));
vst1q_s32(reinterpret_cast<int32_t*>(&data[i * 4]), res);
}
}
static void decodeFilterOctSimd(short* data, size_t count)
{
const int32x4_t sign = vdupq_n_s32(0x80000000);
for (size_t i = 0; i < count; i += 4)
{
int32x4_t n4_0 = vld1q_s32(reinterpret_cast<int32_t*>(&data[(i + 0) * 4]));
int32x4_t n4_1 = vld1q_s32(reinterpret_cast<int32_t*>(&data[(i + 2) * 4]));
// gather both x/y 16-bit pairs in each 32-bit lane
int32x4_t n4 = vuzpq_s32(n4_0, n4_1).val[0];
// sign-extends each of x,y in [x y] with arithmetic shifts
int32x4_t xf = vshrq_n_s32(vshlq_n_s32(n4, 16), 16);
int32x4_t yf = vshrq_n_s32(n4, 16);
// unpack z; note that z is unsigned so we don't need to sign extend it
int32x4_t z4 = vuzpq_s32(n4_0, n4_1).val[1];
int32x4_t zf = vandq_s32(z4, vdupq_n_s32(0x7fff));
// convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
float32x4_t x = vcvtq_f32_s32(xf);
float32x4_t y = vcvtq_f32_s32(yf);
float32x4_t z = vsubq_f32(vcvtq_f32_s32(zf), vaddq_f32(vabsq_f32(x), vabsq_f32(y)));
// fixup octahedral coordinates for z<0
float32x4_t t = vminq_f32(z, vdupq_n_f32(0.f));
x = vaddq_f32(x, vreinterpretq_f32_s32(veorq_s32(vreinterpretq_s32_f32(t), vandq_s32(vreinterpretq_s32_f32(x), sign))));
y = vaddq_f32(y, vreinterpretq_f32_s32(veorq_s32(vreinterpretq_s32_f32(t), vandq_s32(vreinterpretq_s32_f32(y), sign))));
// compute normal length & scale
float32x4_t ll = vaddq_f32(vmulq_f32(x, x), vaddq_f32(vmulq_f32(y, y), vmulq_f32(z, z)));
float32x4_t rl = vrsqrteq_f32(ll);
rl = vmulq_f32(rl, vrsqrtsq_f32(vmulq_f32(rl, ll), rl)); // refine rsqrt estimate
float32x4_t s = vmulq_f32(vdupq_n_f32(32767.f), rl);
// fast rounded signed float->int: addition triggers renormalization after which mantissa stores the integer value
// note: the result is offset by 0x4B40_0000, but we only need the low 16 bits so we can omit the subtraction
const float32x4_t fsnap = vdupq_n_f32(3 << 22);
int32x4_t xr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(x, s), fsnap));
int32x4_t yr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(y, s), fsnap));
int32x4_t zr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(z, s), fsnap));
// mix x/z and y/0 to make 16-bit unpack easier
int32x4_t xzr = vorrq_s32(vandq_s32(xr, vdupq_n_s32(0xffff)), vshlq_n_s32(zr, 16));
int32x4_t y0r = vandq_s32(yr, vdupq_n_s32(0xffff));
// pack x/y/z using 16-bit unpacks; note that this has 0 where we should have .w
int32x4_t res_0 = vreinterpretq_s32_s16(vzipq_s16(vreinterpretq_s16_s32(xzr), vreinterpretq_s16_s32(y0r)).val[0]);
int32x4_t res_1 = vreinterpretq_s32_s16(vzipq_s16(vreinterpretq_s16_s32(xzr), vreinterpretq_s16_s32(y0r)).val[1]);
// patch in .w
res_0 = vbslq_s32(vreinterpretq_u32_u64(vdupq_n_u64(0xffff000000000000)), n4_0, res_0);
res_1 = vbslq_s32(vreinterpretq_u32_u64(vdupq_n_u64(0xffff000000000000)), n4_1, res_1);
vst1q_s32(reinterpret_cast<int32_t*>(&data[(i + 0) * 4]), res_0);
vst1q_s32(reinterpret_cast<int32_t*>(&data[(i + 2) * 4]), res_1);
}
}
static void decodeFilterQuatSimd(short* data, size_t count)
{
const float scale = 1.f / sqrtf(2.f);
for (size_t i = 0; i < count; i += 4)
{
int32x4_t q4_0 = vld1q_s32(reinterpret_cast<int32_t*>(&data[(i + 0) * 4]));
int32x4_t q4_1 = vld1q_s32(reinterpret_cast<int32_t*>(&data[(i + 2) * 4]));
// gather both x/y 16-bit pairs in each 32-bit lane
int32x4_t q4_xy = vuzpq_s32(q4_0, q4_1).val[0];
int32x4_t q4_zc = vuzpq_s32(q4_0, q4_1).val[1];
// sign-extends each of x,y in [x y] with arithmetic shifts
int32x4_t xf = vshrq_n_s32(vshlq_n_s32(q4_xy, 16), 16);
int32x4_t yf = vshrq_n_s32(q4_xy, 16);
int32x4_t zf = vshrq_n_s32(vshlq_n_s32(q4_zc, 16), 16);
int32x4_t cf = vshrq_n_s32(q4_zc, 16);
// get a floating-point scaler using zc with bottom 2 bits set to 1 (which represents 1.f)
int32x4_t sf = vorrq_s32(cf, vdupq_n_s32(3));
float32x4_t ss = vdivq_f32(vdupq_n_f32(scale), vcvtq_f32_s32(sf));
// convert x/y/z to [-1..1] (scaled...)
float32x4_t x = vmulq_f32(vcvtq_f32_s32(xf), ss);
float32x4_t y = vmulq_f32(vcvtq_f32_s32(yf), ss);
float32x4_t z = vmulq_f32(vcvtq_f32_s32(zf), ss);
// reconstruct w as a square root; we clamp to 0.f to avoid NaN due to precision errors
float32x4_t ww = vsubq_f32(vdupq_n_f32(1.f), vaddq_f32(vmulq_f32(x, x), vaddq_f32(vmulq_f32(y, y), vmulq_f32(z, z))));
float32x4_t w = vsqrtq_f32(vmaxq_f32(ww, vdupq_n_f32(0.f)));
float32x4_t s = vdupq_n_f32(32767.f);
// fast rounded signed float->int: addition triggers renormalization after which mantissa stores the integer value
// note: the result is offset by 0x4B40_0000, but we only need the low 16 bits so we can omit the subtraction
const float32x4_t fsnap = vdupq_n_f32(3 << 22);
int32x4_t xr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(x, s), fsnap));
int32x4_t yr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(y, s), fsnap));
int32x4_t zr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(z, s), fsnap));
int32x4_t wr = vreinterpretq_s32_f32(vaddq_f32(vmulq_f32(w, s), fsnap));
// mix x/z and w/y to make 16-bit unpack easier
int32x4_t xzr = vorrq_s32(vandq_s32(xr, vdupq_n_s32(0xffff)), vshlq_n_s32(zr, 16));
int32x4_t wyr = vorrq_s32(vandq_s32(wr, vdupq_n_s32(0xffff)), vshlq_n_s32(yr, 16));
// pack x/y/z/w using 16-bit unpacks; we pack wxyz by default (for qc=0)
int32x4_t res_0 = vreinterpretq_s32_s16(vzipq_s16(vreinterpretq_s16_s32(wyr), vreinterpretq_s16_s32(xzr)).val[0]);
int32x4_t res_1 = vreinterpretq_s32_s16(vzipq_s16(vreinterpretq_s16_s32(wyr), vreinterpretq_s16_s32(xzr)).val[1]);
// rotate and store
uint64_t* out = (uint64_t*)&data[i * 4];
out[0] = rotateleft64(vgetq_lane_u64(vreinterpretq_u64_s32(res_0), 0), vgetq_lane_s32(cf, 0) << 4);
out[1] = rotateleft64(vgetq_lane_u64(vreinterpretq_u64_s32(res_0), 1), vgetq_lane_s32(cf, 1) << 4);
out[2] = rotateleft64(vgetq_lane_u64(vreinterpretq_u64_s32(res_1), 0), vgetq_lane_s32(cf, 2) << 4);
out[3] = rotateleft64(vgetq_lane_u64(vreinterpretq_u64_s32(res_1), 1), vgetq_lane_s32(cf, 3) << 4);
}
}
static void decodeFilterExpSimd(unsigned int* data, size_t count)
{
for (size_t i = 0; i < count; i += 4)
{
int32x4_t v = vld1q_s32(reinterpret_cast<int32_t*>(&data[i]));
// decode exponent into 2^x directly
int32x4_t ef = vshrq_n_s32(v, 24);
int32x4_t es = vshlq_n_s32(vaddq_s32(ef, vdupq_n_s32(127)), 23);
// decode 24-bit mantissa into floating-point value
int32x4_t mf = vshrq_n_s32(vshlq_n_s32(v, 8), 8);
float32x4_t m = vcvtq_f32_s32(mf);
float32x4_t r = vmulq_f32(vreinterpretq_f32_s32(es), m);
vst1q_f32(reinterpret_cast<float*>(&data[i]), r);
}
}
#endif
#ifdef SIMD_WASM
static void decodeFilterOctSimd(signed char* data, size_t count)
{
const v128_t sign = wasm_f32x4_splat(-0.f);
for (size_t i = 0; i < count; i += 4)
{
v128_t n4 = wasm_v128_load(&data[i * 4]);
// sign-extends each of x,y in [x y ? ?] with arithmetic shifts
v128_t xf = wasm_i32x4_shr(wasm_i32x4_shl(n4, 24), 24);
v128_t yf = wasm_i32x4_shr(wasm_i32x4_shl(n4, 16), 24);
// unpack z; note that z is unsigned so we technically don't need to sign extend it
v128_t zf = wasm_i32x4_shr(wasm_i32x4_shl(n4, 8), 24);
// convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
v128_t x = wasm_f32x4_convert_i32x4(xf);
v128_t y = wasm_f32x4_convert_i32x4(yf);
v128_t z = wasm_f32x4_sub(wasm_f32x4_convert_i32x4(zf), wasm_f32x4_add(wasm_f32x4_abs(x), wasm_f32x4_abs(y)));
// fixup octahedral coordinates for z<0
// note: i32x4_min with 0 is equvalent to f32x4_min
v128_t t = wasm_i32x4_min(z, wasm_i32x4_splat(0));
x = wasm_f32x4_add(x, wasm_v128_xor(t, wasm_v128_and(x, sign)));
y = wasm_f32x4_add(y, wasm_v128_xor(t, wasm_v128_and(y, sign)));
// compute normal length & scale
v128_t ll = wasm_f32x4_add(wasm_f32x4_mul(x, x), wasm_f32x4_add(wasm_f32x4_mul(y, y), wasm_f32x4_mul(z, z)));
v128_t s = wasm_f32x4_div(wasm_f32x4_splat(127.f), wasm_f32x4_sqrt(ll));
// fast rounded signed float->int: addition triggers renormalization after which mantissa stores the integer value
// note: the result is offset by 0x4B40_0000, but we only need the low 8 bits so we can omit the subtraction
const v128_t fsnap = wasm_f32x4_splat(3 << 22);
v128_t xr = wasm_f32x4_add(wasm_f32x4_mul(x, s), fsnap);
v128_t yr = wasm_f32x4_add(wasm_f32x4_mul(y, s), fsnap);
v128_t zr = wasm_f32x4_add(wasm_f32x4_mul(z, s), fsnap);
// combine xr/yr/zr into final value
v128_t res = wasm_v128_and(n4, wasm_i32x4_splat(0xff000000));
res = wasm_v128_or(res, wasm_v128_and(xr, wasm_i32x4_splat(0xff)));
res = wasm_v128_or(res, wasm_i32x4_shl(wasm_v128_and(yr, wasm_i32x4_splat(0xff)), 8));
res = wasm_v128_or(res, wasm_i32x4_shl(wasm_v128_and(zr, wasm_i32x4_splat(0xff)), 16));
wasm_v128_store(&data[i * 4], res);
}
}
static void decodeFilterOctSimd(short* data, size_t count)
{
const v128_t sign = wasm_f32x4_splat(-0.f);
volatile v128_t zmask = wasm_i32x4_splat(0x7fff); // TODO: volatile works around LLVM shuffle "optimizations"
for (size_t i = 0; i < count; i += 4)
{
v128_t n4_0 = wasm_v128_load(&data[(i + 0) * 4]);
v128_t n4_1 = wasm_v128_load(&data[(i + 2) * 4]);
// gather both x/y 16-bit pairs in each 32-bit lane
v128_t n4 = wasmx_unziplo_v32x4(n4_0, n4_1);
// sign-extends each of x,y in [x y] with arithmetic shifts
v128_t xf = wasm_i32x4_shr(wasm_i32x4_shl(n4, 16), 16);
v128_t yf = wasm_i32x4_shr(n4, 16);
// unpack z; note that z is unsigned so we don't need to sign extend it
v128_t z4 = wasmx_unziphi_v32x4(n4_0, n4_1);
v128_t zf = wasm_v128_and(z4, zmask);
// convert x and y to floats and reconstruct z; this assumes zf encodes 1.f at the same bit count
v128_t x = wasm_f32x4_convert_i32x4(xf);
v128_t y = wasm_f32x4_convert_i32x4(yf);
v128_t z = wasm_f32x4_sub(wasm_f32x4_convert_i32x4(zf), wasm_f32x4_add(wasm_f32x4_abs(x), wasm_f32x4_abs(y)));
// fixup octahedral coordinates for z<0
// note: i32x4_min with 0 is equvalent to f32x4_min
v128_t t = wasm_i32x4_min(z, wasm_i32x4_splat(0));
x = wasm_f32x4_add(x, wasm_v128_xor(t, wasm_v128_and(x, sign)));
y = wasm_f32x4_add(y, wasm_v128_xor(t, wasm_v128_and(y, sign)));
// compute normal length & scale
v128_t ll = wasm_f32x4_add(wasm_f32x4_mul(x, x), wasm_f32x4_add(wasm_f32x4_mul(y, y), wasm_f32x4_mul(z, z)));
v128_t s = wasm_f32x4_div(wasm_f32x4_splat(32767.f), wasm_f32x4_sqrt(ll));
// fast rounded signed float->int: addition triggers renormalization after which mantissa stores the integer value
// note: the result is offset by 0x4B40_0000, but we only need the low 16 bits so we can omit the subtraction
const v128_t fsnap = wasm_f32x4_splat(3 << 22);
v128_t xr = wasm_f32x4_add(wasm_f32x4_mul(x, s), fsnap);
v128_t yr = wasm_f32x4_add(wasm_f32x4_mul(y, s), fsnap);
v128_t zr = wasm_f32x4_add(wasm_f32x4_mul(z, s), fsnap);
// mix x/z and y/0 to make 16-bit unpack easier
v128_t xzr = wasm_v128_or(wasm_v128_and(xr, wasm_i32x4_splat(0xffff)), wasm_i32x4_shl(zr, 16));
v128_t y0r = wasm_v128_and(yr, wasm_i32x4_splat(0xffff));
// pack x/y/z using 16-bit unpacks; note that this has 0 where we should have .w
v128_t res_0 = wasmx_unpacklo_v16x8(xzr, y0r);
v128_t res_1 = wasmx_unpackhi_v16x8(xzr, y0r);
// patch in .w
res_0 = wasm_v128_or(res_0, wasm_v128_and(n4_0, wasm_i64x2_splat(0xffff000000000000)));
res_1 = wasm_v128_or(res_1, wasm_v128_and(n4_1, wasm_i64x2_splat(0xffff000000000000)));
wasm_v128_store(&data[(i + 0) * 4], res_0);
wasm_v128_store(&data[(i + 2) * 4], res_1);
}
}
static void decodeFilterQuatSimd(short* data, size_t count)
{
const float scale = 1.f / sqrtf(2.f);
for (size_t i = 0; i < count; i += 4)
{
v128_t q4_0 = wasm_v128_load(&data[(i + 0) * 4]);
v128_t q4_1 = wasm_v128_load(&data[(i + 2) * 4]);
// gather both x/y 16-bit pairs in each 32-bit lane
v128_t q4_xy = wasmx_unziplo_v32x4(q4_0, q4_1);
v128_t q4_zc = wasmx_unziphi_v32x4(q4_0, q4_1);
// sign-extends each of x,y in [x y] with arithmetic shifts
v128_t xf = wasm_i32x4_shr(wasm_i32x4_shl(q4_xy, 16), 16);
v128_t yf = wasm_i32x4_shr(q4_xy, 16);
v128_t zf = wasm_i32x4_shr(wasm_i32x4_shl(q4_zc, 16), 16);
v128_t cf = wasm_i32x4_shr(q4_zc, 16);
// get a floating-point scaler using zc with bottom 2 bits set to 1 (which represents 1.f)
v128_t sf = wasm_v128_or(cf, wasm_i32x4_splat(3));
v128_t ss = wasm_f32x4_div(wasm_f32x4_splat(scale), wasm_f32x4_convert_i32x4(sf));
// convert x/y/z to [-1..1] (scaled...)
v128_t x = wasm_f32x4_mul(wasm_f32x4_convert_i32x4(xf), ss);
v128_t y = wasm_f32x4_mul(wasm_f32x4_convert_i32x4(yf), ss);
v128_t z = wasm_f32x4_mul(wasm_f32x4_convert_i32x4(zf), ss);
// reconstruct w as a square root; we clamp to 0.f to avoid NaN due to precision errors
// note: i32x4_max with 0 is equivalent to f32x4_max
v128_t ww = wasm_f32x4_sub(wasm_f32x4_splat(1.f), wasm_f32x4_add(wasm_f32x4_mul(x, x), wasm_f32x4_add(wasm_f32x4_mul(y, y), wasm_f32x4_mul(z, z))));
v128_t w = wasm_f32x4_sqrt(wasm_i32x4_max(ww, wasm_i32x4_splat(0)));
v128_t s = wasm_f32x4_splat(32767.f);
// fast rounded signed float->int: addition triggers renormalization after which mantissa stores the integer value
// note: the result is offset by 0x4B40_0000, but we only need the low 16 bits so we can omit the subtraction
const v128_t fsnap = wasm_f32x4_splat(3 << 22);
v128_t xr = wasm_f32x4_add(wasm_f32x4_mul(x, s), fsnap);
v128_t yr = wasm_f32x4_add(wasm_f32x4_mul(y, s), fsnap);
v128_t zr = wasm_f32x4_add(wasm_f32x4_mul(z, s), fsnap);
v128_t wr = wasm_f32x4_add(wasm_f32x4_mul(w, s), fsnap);
// mix x/z and w/y to make 16-bit unpack easier
v128_t xzr = wasm_v128_or(wasm_v128_and(xr, wasm_i32x4_splat(0xffff)), wasm_i32x4_shl(zr, 16));
v128_t wyr = wasm_v128_or(wasm_v128_and(wr, wasm_i32x4_splat(0xffff)), wasm_i32x4_shl(yr, 16));
// pack x/y/z/w using 16-bit unpacks; we pack wxyz by default (for qc=0)
v128_t res_0 = wasmx_unpacklo_v16x8(wyr, xzr);
v128_t res_1 = wasmx_unpackhi_v16x8(wyr, xzr);
// compute component index shifted left by 4 (and moved into i32x4 slot)
v128_t cm = wasm_i32x4_shl(cf, 4);
// rotate and store
uint64_t* out = reinterpret_cast<uint64_t*>(&data[i * 4]);
out[0] = rotateleft64(wasm_i64x2_extract_lane(res_0, 0), wasm_i32x4_extract_lane(cm, 0));
out[1] = rotateleft64(wasm_i64x2_extract_lane(res_0, 1), wasm_i32x4_extract_lane(cm, 1));
out[2] = rotateleft64(wasm_i64x2_extract_lane(res_1, 0), wasm_i32x4_extract_lane(cm, 2));
out[3] = rotateleft64(wasm_i64x2_extract_lane(res_1, 1), wasm_i32x4_extract_lane(cm, 3));
}
}
static void decodeFilterExpSimd(unsigned int* data, size_t count)
{
for (size_t i = 0; i < count; i += 4)
{
v128_t v = wasm_v128_load(&data[i]);
// decode exponent into 2^x directly
v128_t ef = wasm_i32x4_shr(v, 24);
v128_t es = wasm_i32x4_shl(wasm_i32x4_add(ef, wasm_i32x4_splat(127)), 23);
// decode 24-bit mantissa into floating-point value
v128_t mf = wasm_i32x4_shr(wasm_i32x4_shl(v, 8), 8);
v128_t m = wasm_f32x4_convert_i32x4(mf);
v128_t r = wasm_f32x4_mul(es, m);
wasm_v128_store(&data[i], r);
}
}
#endif
} // namespace meshopt
void meshopt_decodeFilterOct(void* buffer, size_t vertex_count, size_t vertex_size)
{
using namespace meshopt;
assert(vertex_count % 4 == 0);
assert(vertex_size == 4 || vertex_size == 8);
#if defined(SIMD_SSE) || defined(SIMD_NEON) || defined(SIMD_WASM)
if (vertex_size == 4)
decodeFilterOctSimd(static_cast<signed char*>(buffer), vertex_count);
else
decodeFilterOctSimd(static_cast<short*>(buffer), vertex_count);
#else
if (vertex_size == 4)
decodeFilterOct(static_cast<signed char*>(buffer), vertex_count);
else
decodeFilterOct(static_cast<short*>(buffer), vertex_count);
#endif
}
void meshopt_decodeFilterQuat(void* buffer, size_t vertex_count, size_t vertex_size)
{
using namespace meshopt;
assert(vertex_count % 4 == 0);
assert(vertex_size == 8);
(void)vertex_size;
#if defined(SIMD_SSE) || defined(SIMD_NEON) || defined(SIMD_WASM)
decodeFilterQuatSimd(static_cast<short*>(buffer), vertex_count);
#else
decodeFilterQuat(static_cast<short*>(buffer), vertex_count);
#endif
}
void meshopt_decodeFilterExp(void* buffer, size_t vertex_count, size_t vertex_size)
{
using namespace meshopt;
assert(vertex_count % 4 == 0);
assert(vertex_size % 4 == 0);
#if defined(SIMD_SSE) || defined(SIMD_NEON) || defined(SIMD_WASM)
decodeFilterExpSimd(static_cast<unsigned int*>(buffer), vertex_count * (vertex_size / 4));
#else
decodeFilterExp(static_cast<unsigned int*>(buffer), vertex_count * (vertex_size / 4));
#endif
}
#undef SIMD_SSE
#undef SIMD_NEON
#undef SIMD_WASM

58
include/meshoptimizer/vfetchanalyzer.cpp Normal file
View File

@ -0,0 +1,58 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
#include <assert.h>
#include <string.h>
meshopt_VertexFetchStatistics meshopt_analyzeVertexFetch(const unsigned int* indices, size_t index_count, size_t vertex_count, size_t vertex_size)
{
assert(index_count % 3 == 0);
assert(vertex_size > 0 && vertex_size <= 256);
meshopt_Allocator allocator;
meshopt_VertexFetchStatistics result = {};
unsigned char* vertex_visited = allocator.allocate<unsigned char>(vertex_count);
memset(vertex_visited, 0, vertex_count);
const size_t kCacheLine = 64;
const size_t kCacheSize = 128 * 1024;
// simple direct mapped cache; on typical mesh data this is close to 4-way cache, and this model is a gross approximation anyway
size_t cache[kCacheSize / kCacheLine] = {};
for (size_t i = 0; i < index_count; ++i)
{
unsigned int index = indices[i];
assert(index < vertex_count);
vertex_visited[index] = 1;
size_t start_address = index * vertex_size;
size_t end_address = start_address + vertex_size;
size_t start_tag = start_address / kCacheLine;
size_t end_tag = (end_address + kCacheLine - 1) / kCacheLine;
assert(start_tag < end_tag);
for (size_t tag = start_tag; tag < end_tag; ++tag)
{
size_t line = tag % (sizeof(cache) / sizeof(cache[0]));
// we store +1 since cache is filled with 0 by default
result.bytes_fetched += (cache[line] != tag + 1) * kCacheLine;
cache[line] = tag + 1;
}
}
size_t unique_vertex_count = 0;
for (size_t i = 0; i < vertex_count; ++i)
unique_vertex_count += vertex_visited[i];
result.overfetch = unique_vertex_count == 0 ? 0 : float(result.bytes_fetched) / float(unique_vertex_count * vertex_size);
return result;
}

74
include/meshoptimizer/vfetchoptimizer.cpp Normal file
View File

@ -0,0 +1,74 @@
// This file is part of meshoptimizer library; see meshoptimizer.h for version/license details
#include "meshoptimizer.h"
#include <assert.h>
#include <string.h>
size_t meshopt_optimizeVertexFetchRemap(unsigned int* destination, const unsigned int* indices, size_t index_count, size_t vertex_count)
{
assert(index_count % 3 == 0);
memset(destination, -1, vertex_count * sizeof(unsigned int));
unsigned int next_vertex = 0;
for (size_t i = 0; i < index_count; ++i)
{
unsigned int index = indices[i];
assert(index < vertex_count);
if (destination[index] == ~0u)
{
destination[index] = next_vertex++;
}
}
assert(next_vertex <= vertex_count);
return next_vertex;
}
size_t meshopt_optimizeVertexFetch(void* destination, unsigned int* indices, size_t index_count, const void* vertices, size_t vertex_count, size_t vertex_size)
{
assert(index_count % 3 == 0);
assert(vertex_size > 0 && vertex_size <= 256);
meshopt_Allocator allocator;
// support in-place optimization
if (destination == vertices)
{
unsigned char* vertices_copy = allocator.allocate<unsigned char>(vertex_count * vertex_size);
memcpy(vertices_copy, vertices, vertex_count * vertex_size);
vertices = vertices_copy;
}
// build vertex remap table
unsigned int* vertex_remap = allocator.allocate<unsigned int>(vertex_count);
memset(vertex_remap, -1, vertex_count * sizeof(unsigned int));
unsigned int next_vertex = 0;
for (size_t i = 0; i < index_count; ++i)
{
unsigned int index = indices[i];
assert(index < vertex_count);
unsigned int& remap = vertex_remap[index];
if (remap == ~0u) // vertex was not added to destination VB
{
// add vertex
memcpy(static_cast<unsigned char*>(destination) + next_vertex * vertex_size, static_cast<const unsigned char*>(vertices) + index * vertex_size, vertex_size);
remap = next_vertex++;
}
// modify indices in place
indices[i] = remap;
}
assert(next_vertex <= vertex_count);
return next_vertex;
}

11
src/contouring/FlatDualMC.cpp
View File

@ -7,6 +7,7 @@
#include <imgui.h> // NOLINT
#include <toml.h>
#include "dualmc.h"
#include <meshoptimizer.h>
namespace contouring {
FlatDualMC::FlatDualMC(const std::string &str) : AbstractFlat(str) {
@ -83,11 +84,12 @@ namespace contouring {
}
void FlatDualMC::update(const voxel_pos& pos, const world::area_map& areas) {
ZoneScopedN("Ct");
std::pair<area_<chunk_pos>, buffer::ShortIndexed::Data> out;
TracyPlot("CtLoad", static_cast<int64_t>(loadQueue.size()));
//MAYBE: clear out of range loadQueue.trim(keepDistance * keepDistance)
TracyPlot("CtLoaded", static_cast<int64_t>(loadedQueue.size()));
while(loadedQueue.pop(out)) {
for(auto handle = loadedQueue.extractor(); handle.first(out);) {
const auto buffer = new buffer::ShortIndexed(GL_TRIANGLES, out.second);
auto &bfs = buffers[out.first.first].second; //NOTE: buffer.first uninitialized (will be set in clear())
if (const auto it = bfs.find(out.first.second); it != bfs.end()) {
@ -142,7 +144,7 @@ namespace contouring {
out.normals.emplace_back(0);
}
out.indices.reserve(dmc_tris.size());
out.indices.reserve(dmc_tris.size() * 3);
for (const auto& t: dmc_tris) {
glm::vec3 edge1 = out.vertices[t.i1] - out.vertices[t.i0];
glm::vec3 edge2 = out.vertices[t.i2] - out.vertices[t.i0];
@ -167,5 +169,10 @@ namespace contouring {
n = glm::normalize(n);
}
}
meshopt_optimizeVertexCache(out.indices.data(), out.indices.data(), out.indices.size(), out.vertices.size()); //NOTE: pretty minimal gain
// reorder indices for overdraw, balancing overdraw and vertex cache efficiency
const float kThreshold = 1.01f; // allow up to 1% worse ACMR to get more reordering opportunities for overdraw
meshopt_optimizeOverdraw(out.indices.data(), out.indices.data(), out.indices.size(), &out.vertices.front()[0], out.vertices.size(), sizeof(glm::vec3), kThreshold);
}
}

16
src/data/safe_priority_queue.hpp
View File

@ -23,6 +23,14 @@ namespace data {
std::condition_variable_any cv;
public:
std::pair<std::function<void(const K&, const V&, const W&)>, std::unique_lock<LockableBase(std::mutex)>> inserter() {
return std::make_pair([&](const K& key, const V& val, const W& weight) {
heap.emplace_back(key, weight);
std::push_heap(heap.begin(), heap.end(), cmpByWeight);
map.insert_or_assign(key, val);
}, std::unique_lock<LockableBase(std::mutex)>(mutex));
}
void push(const K& key, const V& val, const W& weight) {
std::unique_lock<LockableBase(std::mutex)> lock(mutex);
heap.emplace_back(key, weight);
@ -88,6 +96,14 @@ namespace data {
std::condition_variable_any cv;
public:
std::pair<std::function<void(const K&, const W&)>, std::unique_lock<LockableBase(std::mutex)>> inserter() {
return std::make_pair([&](const K& key, const W& weight) {
heap.emplace_back(key, weight);
std::push_heap(heap.begin(), heap.end(), cmpByWeight);
set.insert(key);
}, std::unique_lock<LockableBase(std::mutex)>(mutex));
}
void push(const K& key, const W& weight) {
std::unique_lock<LockableBase(std::mutex)> lock(mutex);
heap.emplace_back(key, weight);

11
src/data/safe_queue.hpp
View File

@ -15,6 +15,17 @@ namespace data {
std::condition_variable_any cv;
public:
std::pair<std::function<bool(T&)>, std::unique_lock<LockableBase(std::mutex)>> extractor() {
return std::make_pair([&](T& out) {
if (queue.empty())
return false;
out = queue.front();
queue.pop();
return true;
}, std::unique_lock<LockableBase(std::mutex)>(mutex));
}
void push(const T& in) {
std::unique_lock<LockableBase(std::mutex)> lock(mutex);
queue.push(in);

2
src/main.cpp
View File

@ -169,7 +169,7 @@ int main(int /*unused*/, char */*unused*/[]){
renderer->SkyEnable = options.renderer.skybox;
}
{ // Rendering
ZoneScopedNS("Render", 5);
ZoneScopedN("Render");
TracyGpuZone("Render");
auto pass = renderer->getPass();
pass.start();

9
src/world/Universe.cpp
View File

@ -212,7 +212,7 @@ void Universe::update(const voxel_pos& pos, float deltaTime) {
auto it_c = chunks.begin();
while(it_c != chunks.end()) {
if (glm::length2(diff - it_c->first) > glm::pow2(keepDistance)) {
saveQueue.emplace(*it, *it_c);
saveQueue.emplace(*it, *it_c); //MAYBE: take look
lazyArea = false;
it_c = chunks.erase(it_c);
}else {
@ -231,6 +231,7 @@ void Universe::update(const voxel_pos& pos, float deltaTime) {
}
if (chunkChangeArea) { // Enqueue missing chunks
ZoneScopedN("Missing");
auto handle = loadQueue.inserter();
//TODO: need dist so no easy sphere fill
for (int x = -loadDistance; x <= loadDistance; x++) {
for (int y = -loadDistance; y <= loadDistance; y++) {
@ -239,11 +240,13 @@ void Universe::update(const voxel_pos& pos, float deltaTime) {
if (dist2 <= loadDistance * loadDistance) {
const auto p = diff + chunk_pos(x, y, z);
if (chunks.inRange(p) && chunks.find(p) == chunks.end()) {
loadQueue.push(std::make_pair(it->first, p), it->second, -dist2);
handle.first(std::make_pair(it->first, p), it->second, -dist2);
lazyArea = false;
}
}
}}}
if(!lazyArea)
loadQueue.notify_all();
}
allLazy &= lazyArea;
if (lazyArea) { // Clear un-used regions
@ -298,7 +301,7 @@ void Universe::update(const voxel_pos& pos, float deltaTime) {
{ // Store loaded chunks
ZoneScopedN("Load");
robin_hood::pair<area_<chunk_pos>, std::shared_ptr<Chunk>> loaded;
while (loadedQueue.pop(loaded)) {
for (auto handle = loadedQueue.extractor(); handle.first(loaded);) {
if (const auto it = areas.find(loaded.first.first); it != areas.end()) {
auto &chunks = it->second->setChunks();
chunks.emplace(loaded.first.second, loaded.second);

2
src/zstd_sampler.cpp
View File

@ -36,7 +36,7 @@ int main(int /*unused*/, char * /*unused*/[])
std::chrono::nanoseconds gen_time(0);
while(samples.size() < SIZE * SAMPLES) {
const auto start = std::chrono::high_resolution_clock::now();
world::Chunk chunk(chunk_pos(std::rand() % RANGE, std::rand() % RANGE, std::rand() % RANGE), generator);
world::Chunk chunk(chunk_pos(-(std::rand() % RANGE), -(std::rand() % RANGE), -(std::rand() % RANGE)), generator);
gen_time += (std::chrono::high_resolution_clock::now() - start);
std::ostringstream oss;
chunk.write(oss);