luisagroup / luisacompute Goto Github PK

High-Performance Rendering Framework on Stream Architectures

License: BSD 3-Clause "New" or "Revised" License

CMake 0.52% C++ 78.91% C 11.81% Python 2.41% Objective-C++ 0.01% Cuda 0.44% Lua 0.37% Metal 0.34% Rust 5.18% Shell 0.01%

cpu gpu high-performance cross-platform cuda directx graphics ispc llvm metal

luisacompute's People

Contributors

Stargazers

Watchers

Forkers

kushinn automationkit riga27527 am1w1zz sondro 3aceshowhand gaoxinge huangxuezhen mymetalseed ajunlonglive fangwudi kaiyuetao jackzhousz shekn mike-gao zixuanvickylu nickyang sleepyeye faithzl sword-re ioding wenye01 lp249839965 xzy210 lanytcc phoohow gmh5225 zephyroal yangfengzzz acmlczh eva02prot mfkiwl tau233 hengle dxt00 implode-nz liameitimie airguanz saeruhikari hammerlwa asdlei99 examyes jeanschau wow55qq qx9575 leonkang130 summertree qx9557 needsmoar wrvsrx swfly phoenixdigitalfx shiinarinne ashheer codingeric

luisacompute's Issues

Command reordering gives wrong results in LuisaRender

Describe the bug
After reordering, submitting the should-be-independent commands to a single CUDA stream in the REVERSED order triggers illegal memory access on the device.

To Reproduce
As aforementioned; when rendering the Glass-of-Water scene.

Expected behavior
The correct rendering.

Screenshots
N/A

Desktop (please complete the following information):

OS: Manjaro 21.5
CUDA backend
RTX-2080Ti

Additional context
TODO: construct a minimal reproducible example...

LLVM backend: calling convention is not correct for x86

Bad.

buffer of arbitary type: copy from/to list

Converting to int is not safe in HLSL codegen

See this comment.

Occassional crushes when compiling Metal kernels, possibly related to concurrent cache updates

Maybe use file locks to avoid concurrent writes to the .cache folder.

inspect 在交互模式（REPL）得不到源码

>>> import inspect
>>> inspect.getsourcelines(f)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/Cellar/[email protected]/3.9.10/Frameworks/Python.framework/Versions/3.9/lib/python3.9/inspect.py", line 1006, in getsourcelines
    lines, lnum = findsource(object)
  File "/usr/local/Cellar/[email protected]/3.9.10/Frameworks/Python.framework/Versions/3.9/lib/python3.9/inspect.py", line 835, in findsource
    raise OSError('could not get source code')
OSError: could not get source code

ISPC后端 print导致了段错误

88e29e8 运行 build/bin/test_texture，其中

https://github.com/Mike-Leo-Smith/LuisaCompute/blob/88e29e8c76532c63f7ce4f6a8b434aa426a1ca71/src/backends/ispc/ispc_support/ispc_device_library.isph#L6338

会在输出一部分123之后段错误

Support block-wise synchronization in ISPC

i.e., support the synchronize_block() built-in

Atomic load/store with clear memory barrier semantics?

Currently, BUFFER_READ and BUFFER_WRITE are used for atomic load/store, which might not honor the correct memory barrier semantics.

Maybe we have to explicitly support built-in ATOMIC_LOAD and ATOMIC_STORE.

logging sometimes doesn't output anything

resolve host & luisa func with same name

e.g. luisa.struct in kernel should translate to luisa.builtin.struct

print formatted string in kernel/callable

ISPC does not support visiting struct members of an r-value

Bad.

Feature: array, struct & buffer

需求

arr = luisa.ArrayType(5, int)
sph = luisa.StructType(center=float3, radius=float)
geo = luisa.StructType(offset=float3, geometry=sph)

buf = luisa.Buffer(100000, dtype=geo)
buf.copy_from(...)
buf.copy_to(...)

@luisa.kernel
def f(a: arr, b: sph):
	a1 = arr()
	a2 = arr((1,2,3,4,5))
	a3 = arr([1,2,3,4,5])
	b1 = sph()
	b2 = sph(make_float3(1,2,3), 4)
	b3 = sph(center=make_float3(1,2,3), radius=4)
	# Note: 4 should be implicitly converted to float
	b0 = buf.read(dispatch_id().x)
	b0.geometry.radius += 1
	buf.write(dispatch_id().x, b0)

a = arr(...)
b = sph(...)
g = geo(...)
f(a,b, dispatch_size=...)

待定：Buffer上传下载对应的host数据类型

prompt when user forget to init

export operations on vector/matrix types (host side)

Path-tracing on Python is 10% slower than on C++

callable/kernel argument type deduction

Pybind11 not compiling on M1 Macs using LLVM Clang but is fine with AppleClang

Convenient interface for kernels

Probably, support kernel launching in the following style:

stream << kernel(device, args...).dispatch(dim);

Cannot build due to EASTL issue

Hi folks, I'm building with clang++15, ubuntu 20.04.

I encounter a compilation error that is related with EASTL

Err log:

In file included from /home/dev/LuisaCompute/src/ext/EASTL/include/EASTL/algorithm.h:243:
/home/dev/LuisaCompute/src/ext/EASTL/include/EASTL/internal/copy_help.h:135:90: error: no member named 'contiguous_iterator_tag' in namespace 'std'
                                                                  (eastl::is_pointer<InputIterator>::value  || eastl::is_same<IIC, EASTL_ITC_NS::contiguous_iterator_tag>::value) &&
                                                                                                                                   ~~~~~~~~~~~~~~^
/home/dev/LuisaCompute/src/ext/EASTL/include/EASTL/internal/copy_help.h:136:90: error: no member named 'contiguous_iterator_tag' in namespace 'std'
                                                                  (eastl::is_pointer<OutputIterator>::value || eastl::is_same<OIC, EASTL_ITC_NS::contiguous_iterator_tag>::value);
                                                                                                                                   ~~~~~~~~~~~~~~^
/home/dev/LuisaCompute/src/ext/EASTL/include/EASTL/internal/copy_help.h:138:51: error: non-type template argument is not a constant expression
                return eastl::move_and_copy_helper<IIC, isMove, canBeMemmoved>::move_or_copy(first, last, result); // Need to chose based on the input iterator tag and not the output iterator tag, because containers accept input ranges of iterator types different than self.
                                                                ^~~~~~~~~~~~~

This seems the same error with electronicarts/EASTL#482

EASTL commit hash b6f329dc9c09ffaf68a979c6591e66a12cf6c2b9

Do you have any suggestions about this?

vector is not subscriptable on host side

Implement SwapChain for CUDA and ISPC

Probably use OpenGL for display? Maybe we need to re-design the related interfaces before that...

spdlog in jupyter doesn't work as in terminal

spdlog output (output from c++ side) is not real-time & has no color in jupyter.
related: rapidsai/cuml#2604

chained comparison with side effects

consider a<f()<c where f has side effects. The expression will be interpreted as a<f() and f()<c, causing f to be potentially executed twice.

Unclear compile error information

when to_lctype is called on unsupported (non-data) types. e.g:

xx1 = dispatch_id.xy

error:

(kernel)f:5:10: Error: Exception: to_lctype(<class 'luisa.types.BuiltinFuncType'>): unrecognized type
    xx1 = dispatch_id.xy
         ~~~~~~~~~~~~~~~

need rand() in kernel/callable

LLVM backend: linking errors

I am having problems with LLVM backend.

I am using LLVM 15.0.5 on Linux and I get the following runtime error:

Failed to load dynamic module '/tmp/LuisaCompute/build/bin/libluisa-compute-backend-llvm.so', reason: /tmp/LuisaCompute/build/bin/libluisa-compute-backend-llvm.so: undefined symbol: _ZN4absl12lts_2022062318container_internal11kEmptyGroupE. [/tmp/LuisaCompute/src/core/platform.cpp:165]

Any ideas where the missing abseil symbol went?

DirectX backend --device choice error

When setting device via "--device 0", it will actually use device 1.

[ISPC] 可能错误的mipmap采样

当贴图尺寸不符合 2ⁿ×2ⁿ 时，调用 bindless_texture_sample2d_level 可能返回不准确的结果

"Too many sections" error when compiling with gcc

Use -Wa,-mbig-obj flag in GNU and /bigobj flag in MSVC

buffer index type check

type cast

Metal has trouble with non-zero-offset position and triangle buffers in RTX

Bad.

callable struct method?

def f(self, a:int):
    self.k += a
    return self.k

strt = luisa.StructType(k=int, f=f)

@luisa.kernel
f(s: strt):
    s.f(4)

r-value matrix subscripted causing crashes

>>> float2x2(1,2,3,4)
float2x2([1,2], [3,4])
>>> float2x2(1,2,3,4)[0]
float2(9.731018e-35,1e-45)

Refactor `Accel` and `Mesh` interfaces? Honor `allow_update`

Planned modifications:

uint64_t Device::Interface::create_accel(AccelBuildQuality quality, bool allow_update) noexcept;
uint64_t Device::Interface::create_mesh(
    uint64_t v_buffer, size_t v_offset, size_t v_stride, size_t v_count,
    uint64_t t_buffer, size_t t_offset, size_t t_count,
    AccelBuildQuality quality, bool allow_update) noexcept;

where

enum struct AccelBuildQuality {
    HIGH,    // high-quality build, fast trace, typically with compaction
    DEFAULT, // backend decided, balance between build time and trace performance
    LOW      // fast build, possibly lower trace performance, no compaction
};

tuple support in callable/kernel

Some meshes disappear after accel compaction

Describe the bug
Some meshes disappear after the top-level acceleration structure compaction.

To Reproduce
Steps to reproduce the behavior:
Enable accel compaction and render the Staircase scene.

Expected behavior
All meshes rendered.

Screenshots
None.

Desktop (please complete the following information):

OS: macOS

Additional context
Temporally disabling the compaction. May result in some increasing in VRAM usage. Will retry compaction some days later.

Cannot import luisa (Python) after build and setup

Describe the bug
After following the build guide for Python, I cannot import luisa or run test files.

in build_release/bin I see lcapi.cpython-310-darwin.so but no .so to import luisa. I'm assuming this is a PyBind11 thing that the lcapi .so will allow importing luisa but it doesn't see to work:

import lcapi
dir(lcapi)
['Accel', 'AccelBuildCommand', 'AccelBuildRequest', 'AccelModification', 'AccelUsageHint', 'AccessExpr', 'BinaryExpr', 'BinaryOp', 'BindlessArrayUpdateCommand', 'BufferCopyCommand', 'BufferDownloadCommand', 'BufferToTextureCopyCommand', 'BufferUploadCommand', 'CallExpr', 'CallOp', 'CastExpr', 'CastOp', 'Command', 'Context', 'Device', 'DeviceInterface', 'Expression', 'ForStmt', 'FsPath', 'Function', 'FunctionBuilder', 'IfStmt', 'LiteralExpr', 'LoopStmt', 'MemberExpr', 'MeshBuildCommand', 'PixelFormat', 'PixelStorage', 'RefExpr', 'Sampler', 'ScopeStmt', 'ShaderDispatchCommand', 'Stream', 'TextureCopyCommand', 'TextureDownloadCommand', 'TextureToBufferCopyCommand', 'TextureUploadCommand', 'Type', 'UnaryExpr', 'UnaryOp', 'doc', 'file', 'loader', 'name', 'package', 'spec', '_vectorstorage_bool2', '_vectorstorage_bool3', '_vectorstorage_bool4', '_vectorstorage_float2', '_vectorstorage_float3', '_vectorstorage_float4', '_vectorstorage_int2', '_vectorstorage_int3', '_vectorstorage_int4', '_vectorstorage_uint2', '_vectorstorage_uint3', '_vectorstorage_uint4', 'abs', 'acos', 'asin', 'atan', 'atan2', 'bool2', 'bool3', 'bool4', 'builder', 'ceil', 'clamp', 'cos', 'cross', 'degrees', 'determinant', 'distance', 'dot', 'exp', 'float2', 'float2x2', 'float3', 'float3x3', 'float4', 'float4x4', 'floor', 'int2', 'int3', 'int4', 'inverse', 'length', 'lerp', 'log', 'log10', 'log2', 'log_level_error', 'log_level_info', 'log_level_verbose', 'log_level_warning', 'make_bool2', 'make_bool3', 'make_bool4', 'make_float2', 'make_float2x2', 'make_float3', 'make_float3x3', 'make_float4', 'make_float4x4', 'make_int2', 'make_int3', 'make_int4', 'make_uint2', 'make_uint3', 'make_uint4', 'max', 'min', 'normalize', 'pixel_storage_channel_count', 'pixel_storage_size', 'pixel_storage_to_format_float', 'pixel_storage_to_format_int', 'pow', 'radians', 'rotation', 'round', 'scaling', 'select', 'sin', 'sqrt', 'tan', 'to_bytes', 'translation', 'transpose', 'uint2', 'uint3', 'uint4']
import luisa
Traceback (most recent call last):
File "", line 1, in
ModuleNotFoundError: No module named 'luisa'

To Reproduce
Steps to reproduce the behavior:

following https://github.com/LuisaGroup/LuisaCompute/blob/master/README_Python_en.md

cmake -S . -B build_release -D CMAKE_BUILD_TYPE=Release -D CMAKE_C_COMPILER=clang -D CMAKE_CXX_COMPILER=clang++ -D LUISA_COMPUTE_ENABLE_PYTHON=ON (note I removed -G Ninja in instructions and used default generator)
cmake --build build_release -j
source set_python_path.sh build_release
(note no test.py copied to build_release/bin directory as instructions imply)
echo $PYTHONPATH
/Users/bsavery/Code/LuisaCompute/build_release/bin:/Users/bsavery/Code/LuisaCompute/pyscenes:
run python3 src/py/tests/test_path_tracing.py

Traceback (most recent call last):
File "/Users/bsavery/Code/LuisaCompute/src/py/tests/test_path_tracing.py", line 6, in
import luisa
ModuleNotFoundError: No module named 'luisa'

Desktop (please complete the following information):

OS: Macos 13
Version Python 3.10

chained assignment / comparison / bool op

a = b = c
a or b or c
a < b < c

luisagroup / luisacompute Goto Github PK

luisacompute's People

Contributors

Stargazers

Watchers

Forkers

luisacompute's Issues

Recommend Projects

Recommend Topics

Recommend Org