| Age | Commit message (Collapse) | Author |
|---|
|
…function-parameter.slang #8315
Root Cause:
CUDA compilation crashed with `assert failure:
!seenFinalUnsizedArrayField` because unsized arrays like
`RWStructuredBuffer<uint> globalBuffer[]` were not the final field in
generated parameter structs, violating the layout constraint in
slang-ir-layout.cpp.
Fix:
Extended `collectGlobalUniformParameters` to automatically reorder
struct fields for CUDA targets - regular fields first, unsized arrays
last. Other targets preserve original order.
Impact:
- Enables CUDA support for nonuniform resource indexing as function
parameters
- Zero impact on existing GLSL/HLSL/SPIRV targets
- Automatic handling - no manual parameter reordering required
Files: slang-emit.cpp, slang-ir-collect-global-uniforms.cpp/.h, test
file
---------
Co-authored-by: slangbot <ellieh+slangbot@nvidia.com>
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
Co-authored-by: Ellie Hermaszewska <ellieh@nvidia.com>
|
|
## Problem
When generic functions with debug variables were specialized with
concrete types containing non-debuggable fields (e.g.,
`StructuredBuffer`), the IR cloning process would create invalid
`DebugVar` instructions without checking if the substituted types
remained debuggable.
## Solution
This fix adds a defensive check in the legalization pass that removes
the debugVar created for the non-debuggable types.
---------
Co-authored-by: slangbot <ellieh+slangbot@nvidia.com>
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
Fixes this regression:
```slang
struct MyType
{
// Regression Condition 1: there must be more than one member in the lookup scope.
float v;
int getSum() { return 0; }
}
void m(MyType t)
{
// Regression condition 2: the completion must be in an init expression.
// Regression condition 3: none of the candidate members can coerce to the expected type.
// Regression behavior: no completion candidates are shown, because
// SemanticsVisitor::resolveOverloadedLookup throws an error when there are 0 applicable candidates
// after type coercion filtering.
Texture2D x = t.; // completion request after . here
}
```
The root cause is that we shouldn't be applying candidate filtering on
the candidate list when in completion checking mode.
Closes #8417.
|
|
This updates SPIRV submodules.
Unfortunately we cannot use the latest because one of PRs in SPIRV-Tools
causes VVL errors on our slang-tests. The issue is tracked in
- https://github.com/KhronosGroup/SPIRV-Tools/issues/6316
- https://github.com/shader-slang/slang/issues/8414
This PR also updates/simplifies the document of updating SPIRV process.
|
|
|
|
- Enable ccache in slang bulid for CTS test
- Allow manual CI dispatch
|
|
This fixes an issue where non-raytracing kernels couldn't contain any
RaytracingAccelerationStructure resources even when not used.
|
|
macos 15.6 includes python 3.9.6 with Xcode, which doesn't understand
match/case. Changing it to to the less spiffy if/elif.
Co-authored-by: Yong He <yonghe@outlook.com>
Co-authored-by: Sam Estep <sam@samestep.com>
|
|
The test can be enabled
|
|
This PR modernizes the Windows-specific code by replacing ANSI Windows
API functions with their Unicode (wide character) counterparts. This
change ensures proper handling of Unicode file paths and strings on
Windows systems.
### File Operations (`source/core/slang-io.cpp`)
- `DeleteFileA` → `DeleteFileW`
- `GetTempPathA` → `GetTempPathW`
- `GetTempFileNameA` → `GetTempFileNameW`
- `RemoveDirectoryA` → `RemoveDirectoryW`
- `SHFileOperationA` → `SHFileOperationW`
- `GetModuleFileNameA` → `GetModuleFileNameW` with UTF-8 conversion
### Platform Operations (`source/core/slang-platform.cpp`)
- `GetModuleHandleExA` → `GetModuleHandleExW`
- `LoadLibraryExA` → `LoadLibraryExW`
- `LoadLibraryA` → `LoadLibraryW`
- `OutputDebugStringA` → `OutputDebugStringW`
### Runtime and Tools
- `MessageBoxA` → `MessageBoxW` in slang-rt
- `GetCurrentDirectoryA` → `GetCurrentDirectoryW` in slang-fiddle
- String literal conversion to wide strings in vk-pipeline-create
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Gangzheng Tong <gtong-nv@users.noreply.github.com>
Co-authored-by: slangbot <ellieh+slangbot@nvidia.com>
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
Fixes #8396 by not emitting the `ArrayStride` when it would've been
zero. The problem is caused by #7848, more details in the issue
description.
I checked that with equivalent GLSL code, glslangValidator does not emit
any `ArrayStride`. I assume that the addition of `storageClass ==
SpvStorageClassStorageBuffer` to line 1848 is not a mistake. If it is,
that addition could also be simply reverted to fix this issue, I tested
that option as well. With these changes, Slang tests work locally on my
PC again.
Related to this; it'd be nice to have GPUs from multiple vendors in the
CI to avoid this kind of thing happening again. Or even just llvmpipe;
that doesn't even require a GPU and would add at least one more driver
to test with.
|
|
1. Adds retry logic - Implements 3-attempt retry mechanism for
intermittent test failures
2. Reduces parallelism for Linux - Changes server counts back to 1 for
more stable execution, given the test is short in Linux for now
3. Adds detailed error logging - Enhanced diagnostic information for
test parsing failures
4. Add Python 3.10 setup for slangpy tests for github runners
5. Removes submodule checkout for slang and slasngpy tests
6. Adds check-ci job - Implements consolidated CI status for simplified
branch protection rule
---------
Co-authored-by: slangbot <ellieh+slangbot@nvidia.com>
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
This change relaxes a previous restriction on link-time types and
constants, so that we now allow them to be used to define shader
parameters.
Doing so will result in a parameter layout that is incomplete prior to
linking. The PR added a test to call the reflection API on a fully
linked program and ensure that we can report correct binding info.
|
|
- This PR removes "wasm" check from the test job, since there is no test
job for "wasm" yet.
- Also, move the check for slang-rhi and slangpy test to the job level
to skip the setup etc. if no run is needed.
- Update the macos compiler version in building.md to match the clang
compiler used in CI.
- Reduce the server count in Linux and Mac when running slang-test to
ease the machine load
- Run slangpy test with `-n auto --maxprocesses=4` to speed up
---------
Co-authored-by: slangbot <ellieh+slangbot@nvidia.com>
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
The header execinfo.h and the related backtrace functionality is not
available on all linux platforms. In particular it's missing on musl
linux and on Android before API version 33. This causes compilation
errors on those platforms.
With this change, we first check if backtrace functionality is available
by checking if we are using glibc or a compatible Android version.
Tested on manylinux_2_28 with glibc 2.28 and musllinux_1_2 with musl
1.2, has not been tested on Android.
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
WGSL requires Wave functions to be used only in uniform control flow.
The latest compiler, Tint, started to error out when Wave functions are
used in a dynamic control flow.
This commit disables some of tests using Wave functions in dynamic
branches. If possible, they are altered to call Wave functions in
uniform control flows.
|
|
The code int x4 = 0xFFFFFFFFFFFFFFFF previously did not produce a
warning due to the value being too large for the type. This patch now
checks for this and similar issues during parsing.
|
|
If a given library cannot be found using LoadLibraryExA then try again
using LoadLibraryA. Return an error only if both of these failed.
|
|
Previously we are missing `ccache --set-config=cache_dir="$ccache_dir"`
and the build was using the default cache_dir.
This PR fixed that and the cache are hit in reruns.
```
📊 ccache statistics (post-build):
Cacheable calls: 1198 / 1198 (100.0%)
Hits: 1195 / 1198 (99.75%)
Direct: 1193 / 1195 (99.83%)
Preprocessed: 2 / 1195 ( 0.17%)
Misses: 3 / 1198 ( 0.25%)
Local storage:
Cache size (GiB): 0.6 / 5.0 (11.11%)
Hits: 1195 / 1198 (99.75%)
Misses: 3 / 1198 ( 0.25%)
```
|
|
The CI has been re-organized with the following:
```
ci.yml (Main Orchestrator)
├── filter job
│ ├── Documentation Only? → Yes → Skip CI
│ └── Documentation Only? → No → Continue CI
│
├── Build Jobs
│ └── ci-slang-build.yml
│ ├── common-setup action
│ ├── Build & Package
│ └── Upload Artifacts (Build package and Tests package)
│
└── Test Jobs
└── ci-slang-test.yml
└── common-test-setup action
├── Download Tests Artifacts ← (from Build)
└── Run Tests
```
To achieve fine-grained build->test dependency, instead of using `matrix
strategy` in single build (or single test) job, the main ci.yml
statically defines the each config of the build and test job.
e.g. `build-windows-debug-cl-x86_64-gpu` and
`test-windows-debug-cl-x86_64-gpu` are a pair of the build-test job for
the windows/debug/ci config.
Closes: https://github.com/shader-slang/slang/issues/6728
---------
Co-authored-by: slangbot <ellieh+slangbot@nvidia.com>
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
Update intrinsics signature as per the nvapi header
|
|
Enable CUDA support for additional HLSL intrinsic tests by implementing
missing functionality and fixing compiler bugs affecting CUDA targets.
- Fix critical bug in InterlockedCompareStore64 where division used /4
instead of /8 for 64-bit types, causing incorrect memory addressing for
all signed int 64_t atomics
- Add signed int64_t atomic wrappers (atomicExch, atomicCAS) to CUDA
prelu de that properly cast to/from unsigned types as required by CUDA's
atomic API
- Enable tests: atomic-intrinsics-64bit.slang
- Implement CUDA support for QuadAny and QuadAll operations using warp
shu ffle primitives (__shfl_sync with quad-level lane masking)
- Add CUDA to quad_control capability definition in
slang-capabilities.capdef
- Add _slang_quadAny/_slang_quadAll helper functions to CUDA prelude
- Enable tests: quad-control-comp-functionality.slang,
subgroup-quad.slang
---------
Co-authored-by: szihs <675653+szihs@users.noreply.github.com>
|
|
Before this PR only the following was a valid line without any
white-space character nor additional `/` character,
```
//TEST:
```
This PR is to allow slang-test to handle the following variants of the
test command comments,
```
///TEST:
// TEST:
// TEST:
////// TEST:
```
This PR revealed a regression on two tests:
- tests/cpp-compiler/c-compile-shared-library.c (cpu)
- tests/cpp-compiler/cpp-compile-shared-library.cpp (cpu)
They are disabled as a part of this PR.
And there is a new github issue to track it later,
- https://github.com/shader-slang/slang/issues/8362
|
|
closes https://github.com/shader-slang/slang/issues/3313
|
|
|
|
Add 29 intrinsics to the list by new test
|
|
|
|
This avoids uploading the ccache to github and take the cache storage.
---------
Co-authored-by: slangbot <ellieh+slangbot@nvidia.com>
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
Related to https://github.com/shader-slang/slang/issues/6728
---------
Co-authored-by: slangbot <ellieh+slangbot@nvidia.com>
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
Changes default for render-test to sm_6_5.
Since sm_6_5 is the new default, remove the -use-dxil option, add
-use-dxcb option
Remove -use-dxil option from all test cases.
Add -use-dxcb to two tests that needed it.
Fixes #7611
|
|
This commit is to emit the debug-info for the entry point parameters.
Two things are implemented/fixed in this PR:
- We were not emitting the `DebugVar` and `DebugValue` at the IR
lowering level when the type of the entry point parameter is `ConstRef`.
This commit handles the `ConstRef` case in a same way that the other
types are handled so that `DebugVar` and `DebugValues` are properly
emitted at the IR lowering level.
- Two types for Geometry shaders were incorrectly treated as not valid
types for the DebugInfo. They are `InputPatch` and `OutputPatch`. This
commit handles them as valid types for DebugInfo.
|
|
The link-time specialization documentation contained an incorrect
example that used `[ForceUnroll]` with a link-time type method call,
which would cause a compilation error. The issue was that
`[ForceUnroll]` requires loop bounds to be known at compile time, but
`sampler.getSampleCount()` is a method call that returns a value at
runtime.
**Problem:**
The documentation example showed:
```csharp
Sampler sampler;
[ForceUnroll]
for (int i = 0; i < sampler.getSampleCount(); i++)
output[tid] += sampler.sample(i);
```
This would fail with error: `loop does not terminate within the limited
number of iterations, unrolling is aborted.`
**Solution:**
Removed the `[ForceUnroll]` attribute entirely, leaving a simple loop:
```csharp
Sampler sampler;
for (int i = 0; i < sampler.getSampleCount(); i++)
output[tid] += sampler.sample(i);
```
Since the loop bounds come from a runtime method call, there's no way
for the loop to be unrolled regardless of the directive used, so the
simplest solution is to remove the unroll attribute completely.
- [x] Remove ForceUnroll attribute from documentation example
- [x] Remove explanatory note about unroll vs ForceUnroll
- [x] Remove test cases for the removed functionality
- [x] Fix missing closing backticks in code block
Fixes #8161.
<!-- START COPILOT CODING AGENT TIPS -->
---
💡 You can make Copilot smarter by setting up custom instructions,
customizing its development environment and configuring Model Context
Protocol (MCP) servers. Learn more [Copilot coding agent
tips](https://gh.io/copilot-coding-agent-tips) in the docs.
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: bmillsNV <163073245+bmillsNV@users.noreply.github.com>
Co-authored-by: expipiplus1 <857308+expipiplus1@users.noreply.github.com>
|
|
|
|
Resolves #7628
Resolves: #8197
Primary Goals:
1. Add `Access` to pointer
2. AddressSpace::GroupShared support for pointers (SPIR-V)
3. Add `__getAddress()` to replace `&`
* `&` is not updated to `require(cpu)` since slangpy uses `&`. This
means we must: (1) merge PR; (2) replace `&` with `__getAddress()`; (3)
add `require(cpu)` to `&`
Changes:
* Added to `Ptr` the `Access` generic argument & logic (for
`Access::Read`).
* Moved the generic argument `AddressSpace` from `Ptr` to the end of the
type.
* Added pointer casting support between any `Ptr` as long as the
`AddressSpace` is the same
* Disallow globallycoherent T* and coherent T*
* Disallow const T*, T const*, and const T*
* Fixed .natvis display of `ConstantValue` `ValOperandNode`
* Support generic resolution of type-casted integers
* Added `VariablePointer` emitting for spirv + other minor logic needed
for groupshared pointers
Breaking Changes:
* Anyone using the `AddressSpace` of `Ptr` will now have to account for
the `Access` argument
* we disallow various syntax paired with `Ptr` and `T*`
---------
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
This PR enables slang-rhi and slangpy tests in the slang CI for MacOS.
* exclude the slang-rhi `sampler-array` test for mac; issue tracked in
https://github.com/shader-slang/slang/issues/8246
* update slang-rhi for fix of `nested-parameter-block-2`
* for slangpy test, install the required python package in the github
mac runner only each CI run.
Closes: https://github.com/shader-slang/slang/issues/7330
---------
Co-authored-by: slangbot <ellieh+slangbot@nvidia.com>
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
This cause the pthread not found issue on old glibc build.
This reverts commit 113327194d4cf750af6265a560615850a8e7e6fb.
|
|
When `SIMPLE` type test is used with `-g[1-3]` option, the filecheck
pattern will most likely to match to the string itself on the embedded
source code rather than match to the emitted spirv-asm code.
This commit avoids the problem by removing the embedded source code.
This commit also provides an option to keep the embedded source code,
`-preserve-embedded-source`.
The source code removal is happening in two steps:
1. iterate all output lines and find SPIRV-ASM in the following pattern:
`%N = OpExtInst %void %M DebugSource %fileId %sourceId`. And then, store
the "%sourceId" value to identify which SPIRV instructions are for the
embedded source code.
2. iterate all output lines again to find the `%sourceId = OpString
"...."` and replace the whole string with the following string, ``` %1 =
OpString "// slang-test removed the embedded source // Use
`-preserve-embedded-source` to keep it explicitly " ```
This change revealed problems in the existing tests:
- tests/bugs/spirv-debug-info.slang : The expected text was missing and
it had to be added. The file also had Carrage-Return character on all
lines and the pre-commit git hook removed them.
- tests/spirv/debug-info.slang : the expected keyword DebugValue had to
change to DebugDeclare, because that's what we get with ToT.
- tests/spirv/debug-value-dynamic-index.slang : This test is currently
failing, and it will pass once DebugLocalVariable instruction missing
for parameter of the entry point function #7693 is resolved.
---------
Co-authored-by: slangbot <ellieh+slangbot@nvidia.com>
|
|
Update docs/shader-execution-reordering.md with additional intrinsics
Add correct capability `LoadLocalRootTableConstant`
|
|
Emits the appropriate OpCapability for 8- and 16-bit type usage:
- UniformAndStorageBuffer8BitAccess: for 16-bit types in
SpvStorageClassUniform and SpvStorageClassStorageBuffer
- UniformAndStorageBuffer16BitAccess: for 16-bit types in
SpvStorageClassUniform and SpvStorageClassStorageBuffer
- StoragePushConstant8: for 8-bit types in SpvStorageClassPushConstant
- StoragePushConstant16: for 16-bit types in SpvStorageClassPushConstant
- StorageInputOutput16: for 16-bit types in SpvStorageClassInput and
SpvStorageClassOutput
Generated with Claude Code, with revisions.
Fixes #7879.
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: James Helferty (NVIDIA) <jhelferty-nv@users.noreply.github.com>
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
fixes https://github.com/shader-slang/slang/issues/8271
This PR does the following,
- Fail slang-test when there are VVL error messages.
- VVL error for `gfx-unit-test-tool/` were not captured properly by the
debug callback.
- Set an environment variable,
`VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation`, for CI and
VisualStudio project setup.
- Ignores VVL error about NullHandle is used for the acceleration
structure; a fix is at ToT of VVL and not available from release build
yet.
- Fix VVL error complaining about the varying inputs are not provided
for the tests, `gfx-unit-test-tool/linkTimeTypeLayout.internal` and
`gfx-unit-test-tool/linkTimeTypeLayoutNested.internal`.
---------
Co-authored-by: slangbot <ellieh+slangbot@nvidia.com>
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
(#8223)
The Metal backend was generating incorrect type names for 8-bit vector
types, causing compilation failures when targeting Metal. According to
the Metal specification, 8-bit vector types should be named `charN` and
`ucharN` (e.g., `char2`, `uchar3`) rather than `int8_tN` and `uint8_tN`.
## Problem
When compiling Slang code with 8-bit vector types for Metal, the
compiler would emit:
```metal
uint8_t2 _S8 = uint8_t2(uint8_t(0U), uint8_t(16U));
int8_t3 _S9 = int8_t3(int8_t(0), int8_t(16), int8_t(48));
```
But the Metal compiler expects:
```metal
uchar2 _S8 = uchar2(uint8_t(0U), uint8_t(16U));
char3 _S9 = char3(int8_t(0), int8_t(16), int8_t(48));
```
This caused errors like:
```
error: unknown type name 'uint8_t2'; did you mean 'uint8_t'?
```
## Solution
Modified `MetalSourceEmitter::emitSimpleTypeImpl()` to emit the correct
Metal-specific type names for 8-bit types:
- `kIROp_Int8Type` now emits `char` instead of `int8_t`
- `kIROp_UInt8Type` now emits `uchar` instead of `uint8_t`
This change only affects the Metal backend and ensures that vector types
like `int8_t2`, `uint8_t3`, etc. are correctly emitted as `char2`,
`uchar3`, etc.
## Testing
- Added a new test case `tests/metal/8bit-vector-types.slang` to verify
the fix
- Re-enabled the previously disabled Metal test in
`tests/hlsl-intrinsic/countbits8.slang`
- Updated `tests/metal/byte-address-buffer.slang` to expect the correct
type names
- Verified that existing Metal tests continue to pass
Fixes #8211.
<!-- START COPILOT CODING AGENT TIPS -->
---
💡 You can make Copilot smarter by setting up custom instructions,
customizing its development environment and configuring Model Context
Protocol (MCP) servers. Learn more [Copilot coding agent
tips](https://gh.io/copilot-coding-agent-tips) in the docs.
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: bmillsNV <163073245+bmillsNV@users.noreply.github.com>
|
|
This PR marks the `slang-rhi` a required dependecy for `platform` and
`gfx-unit-test`, and only build them when `SLANG_ENABLE_SLANG_RHI=ON`.
This should allow the slang still to be built without those tests
components when `SLANG_ENABLE_SLANG_RHI=OFF`.
---------
Co-authored-by: slangbot <ellieh+slangbot@nvidia.com>
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
Fixes: #8018
Changes:
* Do not emit true for `shouldEmitSPIRVDirectly` with a GLSL target
|
|
CMake 3.31 (released [last
November](https://www.kitware.com/cmake-3-31-0-available-for-download/))
deprecated compatibility with CMake versions older than 3.10, causing
warnings when running commands like `cmake --preset default` in the
Slang repo:
```
CMake Deprecation Warning at external/miniz/CMakeLists.txt:1 (cmake_minimum_required):
Compatibility with CMake < 3.10 will be removed from a future version of
CMake.
Update the VERSION argument <min> value. Or, use the <min>...<max> syntax
to tell CMake that the project requires at least <min> but has been updated
to work with policies introduced by <max> or earlier.
CMake Deprecation Warning at external/lz4/build/cmake/CMakeLists.txt:13 (cmake_minimum_required):
Compatibility with CMake < 3.10 will be removed from a future version of
CMake.
Update the VERSION argument <min> value. Or, use the <min>...<max> syntax
to tell CMake that the project requires at least <min> but has been updated
to work with policies introduced by <max> or earlier.
```
Those dependencies modified their `cmake_minimum_required` calls in
lz4/lz4#1601 and richgel999/miniz#344 respectively to fix those
warnings, so this PR bumps them both to include those changes.
Co-authored-by: Ellie Hermaszewska <ellieh@nvidia.com>
|
|
close #8068.
Currently the AutoDiff aggressively scan every IR inst in searching the
differentiable IR. This is not efficient and could have bug, details in
https://github.com/shader-slang/slang/issues/8068#issuecomment-3214856668.
This PR change the behavior. It will do a initial filter to only gather
the global differentiable IRs and IRFunc and IRGeneric as well. For
IRGeneric, we will pick it only when it's used in other generic function
(it's only useful when dealing with dynamic dispatch).
Then we will start searching reachable insts from this IR list by using
the same method as before.
|
|
This is a first pass at adding WASM builds to releases. ~~I haven't
tested the outputs yet but it appears to build successfully in my fork
at least.~~
Edit: Made a fake release in my fork for testing, seems to be working
well for my use case at least:
https://github.com/davidar/slang/releases/tag/v2025.999
Fixes #8207
|
|
This commit prints an additional information of GPU version while
running CI.yml.
This can help us to debug when things don't work as expected.
|
|
|
|
|
|
|
