Revision 8dd04c8730327d7dc2fd29b43b8566fabbc3aa2c authored by Yong He on 09 January 2024, 20:41:07 UTC, committed by GitHub on 09 January 2024, 20:41:07 UTC
Co-authored-by: Yong He <yhe@nvidia.com>
1 parent 69f3d79
cuda-target.md
Slang CUDA Target Support
=========================
Slang has preliminary support for producing CUDA source, and PTX binaries using [NVRTC](https://docs.nvidia.com/cuda/nvrtc/index.html).
NOTE! NVRTC is only available for 64-bit operating systems. On Windows Visual Studio make sure you are compiling for 'x64' and/or use 64 bit Slang binaries.
# Features
* Can compile Slang source into CUDA source code
* Supports compute style shaders
* Supports a 'bindless' CPU like model
* Can compile CUDA source to PTX through 'pass through' mechansism
# Limitations
These limitations apply to Slang transpiling to CUDA.
* Only supports the 'texture object' style binding (The texture object API is only supported on devices of compute capability 3.0 or higher. )
* Samplers are not separate objects in CUDA - they are combined into a single 'TextureObject'. So samplers are effectively ignored on CUDA targets.
* When using a TextureArray.Sample (layered texture in CUDA) - the index will be treated as an int, as this is all CUDA allows
* Care must be used in using `WaveGetLaneIndex` wave intrinsic - it will only give the right results for appropriate launches
* CUDA 'surfaces' are used for textures which are read/write (aka RWTexture).
The following are a work in progress or not implemented but are planned to be so in the future
* Some resource types remain unsupported, and not all methods on all types are supported
# How it works
For producing PTX binaries Slang uses [NVRTC](https://docs.nvidia.com/cuda/nvrtc/index.html). NVRTC dll/shared library has to be available to Slang (for example in the appropriate PATH for example) for it to be able to produce PTX.
The NVRTC compiler can be accessed directly via the pass through mechanism and is identifed by the enum value `SLANG_PASS_THROUGH_NVRTC`.
Much like other targets that use downstream compilers Slang can be used to compile CUDA source directly to PTX via the pass through mechansism. The Slang command line options will broadly be mapped down to the appropriate options for the NVRTC compilation. In the API the `SlangCompileTarget` for CUDA is `SLANG_CUDA_SOURCE` and for PTX is `SLANG_PTX`. These can also be specified on the Slang command line as `-target cuda` and `-target ptx`.
## Locating NVRTC
Finding NVRTC can require some nuance if a specific version is required. On the command line the `-nvrtc-path` option can be used to set the `path` to NVRTC. Also `spProcessCommandLineArguments`/`processCommandLineArguments` with `-nvrtc-path` or `setDownstreamCompilerPath` with `SLANG_PASS_THROUGH_NVRTC` can be used to set the location and/or name of NVRTC via the API.
Important points of note are
* The name of the shared library should *not* include any extension (such as `.dll`/`.so`/`.dynlib`) or prefix (such as `lib`).
* The path also *doesn't* have to be path, it can just be the shared library name. Doing so will mean it will be searched for by whatever the default mechanism is on the target.
* If a path and/or name is specified for NVRTC - this will be the *only* version searched for.
If a path/name is *not* specified for NVRTC, Slang will attempt to load a shared library called `nvrtc`. For non Windows targets this should be enough to find and load the latest version.
On Windows NVRTC dlls have a name the contains the version number, for example `nvrtc64_102_0.dll`. This will lead to the load of just `nvrtc` to fail. One approach to fix this is to place the NVRTC dll and associated files in the same directory as slang.dll, and rename the main dll to `nvrtc.dll`. Another approach is to specify directly on the command line the name including the version, as previously discussed. For example
`-nvrtc-path nvrtc64_102_0`
will load NVRTC 10.2 assuming that version of the dll can be found via the normal lookup mechanism.
On Windows if NVRTC is not loadable directly as 'nvrtc' Slang will attempt to search for the newest version of NVRTC on your system. The places searched are...
* The instance directory (where the slang.dll and/or program exe is)
* The CUDA_PATH enivonment variable (if set)
* Directories in PATH that look like a CUDA installation.
If a candidate is found via an earlier mechanism, subsequent searches are not performed. If multiple candidates are found, Slang tries the newest version first.
Binding
=======
Say we have some Slang source like the following:
```
struct Thing { int a; int b; }
Texture2D<float> tex;
SamplerState sampler;
RWStructuredBuffer<int> outputBuffer;
ConstantBuffer<Thing> thing3;
[numthreads(4, 1, 1)]
void computeMain(
uint3 dispatchThreadID : SV_DispatchThreadID,
uniform Thing thing,
uniform Thing thing2)
{
// ...
}
```
This will be turned into a CUDA entry point with
```
struct UniformEntryPointParams
{
Thing thing;
Thing thing2;
};
struct UniformState
{
CUtexObject tex; // This is the combination of a texture and a sampler(!)
SamplerState sampler; // This variable exists within the layout, but it's value is not used.
RWStructuredBuffer<int32_t> outputBuffer; // This is implemented as a template in the CUDA prelude. It's just a pointer, and a size
Thing* thing3; // Constant buffers map to pointers
};
// [numthreads(4, 1, 1)]
extern "C" __global__ void computeMain(UniformEntryPointParams* params, UniformState* uniformState)
```
With CUDA - the caller specifies how threading is broken up, so `[numthreads]` is available through reflection, and in a comment in output source code but does not produce varying code.
The UniformState and UniformEntryPointParams struct typically vary by shader. UniformState holds 'normal' bindings, whereas UniformEntryPointParams hold the uniform entry point parameters. Where specific bindings or parameters are located can be determined by reflection. The structures for the example above would be something like the following...
`StructuredBuffer<T>`,`RWStructuredBuffer<T>` become
```
T* data;
size_t count;
```
`ByteAddressBuffer`, `RWByteAddressBuffer` become
```
uint32_t* data;
size_t sizeInBytes;
```
## Texture
Read only textures will be bound as the opaque CUDA type CUtexObject. This type is the combination of both a texture AND a sampler. This is somewhat different from HLSL, where there can be separate `SamplerState` variables. This allows access of a single texture binding with different types of sampling.
If code relys on this behavior it will be necessary to bind multiple CtexObjects with different sampler settings, accessing the same texture data.
Slang has some preliminary support for TextureSampler type - a combined Texture and SamplerState. To write Slang code that can target CUDA and other platforms using this mechanism will expose the semantics appropriately within the source.
Load is only supported for Texture1D, and the mip map selection argument is ignored. This is because there is tex1Dfetch and no higher dimensional equivalents. CUDA also only allows such access if the backing array is linear memory - meaning the bound texture cannot have mip maps - thus making the mip map parameter superflous anyway. RWTexture does allow Load on other texture types.
## RWTexture
RWTexture types are converted into CUsurfObject type.
In regular CUDA it is not possible to do a format conversion on an access to a CUsurfObject. Slang does add support for hardware write conversions where they are available. To enable the feature it is necessary to attribute your RWTexture with `format`. For example
```
[format("rg16f")]
RWTexture2D<float2> rwt2D_2;
```
The format names used are the same as for [GLSL layout format types](https://www.khronos.org/opengl/wiki/Layout_Qualifier_(GLSL)). If no format is specified Slang will *assume* that the format is the same as the type specified.
Note that the format attribution is on variables/paramters/fields and not part of the type system. This means that if you have a scenario like...
```
[format(rg16f)]
RWTexture2d<float2> g_texture;
float2 getValue(RWTexture2D<float2> t)
{
return t[int2(0, 0];
}
void doThing()
{
float2 v = getValue(g_texture);
}
```
Even `getValue` will receive t *without* the format attribute, and so will access it, presumably erroneously. A work around for this specific scenario would be to attribute the parameter
```
float2 getValue([format("rg16f")] RWTexture2D<float2> t)
{
return t[int2(0, 0];
}
```
This will only work correctly if `getValue` is called with a `t` that has that format attribute. As it stands no checking is performed on this matching so no error or warning will be produced if there is a mismatch.
There is limited software support for doing a conversion on reading. Currently this only supports only 1D, 2D, 3D RWTexture, backed with half1, half2 or half4. For this path to work NVRTC must have the `cuda_fp16.h` and associated files available. Please check the section on `Half Support`.
If hardware read conversions are desired, this can be achieved by having a Texture<T> that uses the surface of a RWTexture<T>. Using the Texture<T> not only allows hardware conversion but also filtering.
It is also worth noting that CUsurfObjects in CUDA are NOT allowed to have mip maps.
By default surface access uses cudaBoundaryModeZero, this can be replaced using the macro SLANG_CUDA_BOUNDARY_MODE in the CUDA prelude. For HW format conversions the macro SLANG_PTX_BOUNDARY_MODE. These boundary settings are in effect global for the whole of the kernel.
`SLANG_CUDA_BOUNDARY_MODE` can be one of
* cudaBoundaryModeZero causes an execution trap on out-of-bounds addresses
* cudaBoundaryModeClamp stores data at the nearest surface location (sized appropriately)
* cudaBoundaryModeTrap drops stores to out-of-bounds addresses
`SLANG_PTX_BOUNDARY_MODE` can be one of `trap`, `clamp` or `zero`. In general it is recommended to have both set to the same type of value, for example `cudaBoundaryModeZero` and `zero`.
## Sampler
Samplers are in effect ignored in CUDA output. Currently we do output a variable `SamplerState`, but this value is never accessed within the kernel and so can be ignored. More discussion on this behavior is in `Texture` section.
## Unsized arrays
Unsized arrays can be used, which are indicated by an array with no size as in `[]`. For example
```
RWStructuredBuffer<int> arrayOfArrays[];
```
With normal 'sized' arrays, the elements are just stored contiguously within wherever they are defined. With an unsized array they map to `Array<T>` which is...
```
T* data;
size_t count;
```
Note that there is no method in the shader source to get the `count`, even though on the CUDA target it is stored and easily available. This is because of the behavior on GPU targets
* That the count has to be stored elsewhere (unlike with CUDA)
* On some GPU targets there is no bounds checking - accessing outside the bound values can cause *undefined behavior*
* The elements may be laid out *contiguously* on GPU
In practice this means if you want to access the `count` in shader code it will need to be passed by another mechanism - such as within a constant buffer. It is possible in the future support may be added to allow direct access of `count` work across targets transparently.
## Prelude
For CUDA the code to support the code generated by Slang is partly defined within the 'prelude'. The prelude is inserted text placed before the generated CUDA source code. For the Slang command line tools as well as the test infrastructure, the prelude functionality is achieved through a `#include` in the prelude text of the `prelude/slang-cuda-prelude.h` specified with an absolute path. Doing so means other files the `slang-cuda-prelude.h` might need can be specified relatively, and include paths for the backend compiler do not need to be modified.
The prelude needs to define
* 'Built in' types (vector, matrix, 'object'-like Texture, SamplerState etc)
* Scalar intrinsic function implementations
* Compiler based definations/tweaks
For a client application - as long as the requirements of the generated code are met, the prelude can be implemented by whatever mechanism is appropriate for the client. For example the implementation could be replaced with another implementation, or the prelude could contain all of the required text for compilation. Setting the prelude text can be achieved with the method on the global session...
```
/** Set the 'prelude' for generated code for a 'downstream compiler'.
@param passThrough The downstream compiler for generated code that will have the prelude applied to it.
@param preludeText The text added pre-pended verbatim before the generated source
That for pass-through usage, prelude is not pre-pended, preludes are for code generation only.
*/
void setDownstreamCompilerPrelude(SlangPassThrough passThrough, const char* preludeText);
```
The code that sets up the prelude for the test infrastucture and command line usage can be found in ```TestToolUtil::setSessionDefaultPrelude```. Essentially this determines what the absolute path is to `slang-cpp-prelude.h` is and then just makes the prelude `#include "the absolute path"`.
Half Support
============
Slang supports the half/float16 types on CUDA. To do so NVRTC must have access to the `cuda_fp16.h` and `cuda_fp16.hpp` files that are typically distributed as part of the CUDA SDK. When Slang detects the use of half in source, it will define `SLANG_CUDA_ENABLE_HALF` when `slang-cuda-prelude.h` is included. This will in turn try to include `cuda_fp16.h` and enable extra functionality within the prelude for half support.
Slang tries several mechanisms to locate `cuda_fp16.h` when NVRTC is initiated. The first mechanism is to look in the include paths that are passed to Slang. If `cuda_fp16.h` can be found in one of these paths, no more searching will be performed.
If this fails, the path where NVRTC is located will be searched. In that path "include" and "CUDA/include" paths will be searched. This is probably most suitable for Windows based targets, where NVRTC dll is placed along with other binaries. The "CUDA/include" path is used to try and make clear in this scenario what the contained files are for.
If this fails Slang will look for the CUDA_PATH environmental variable, as is typically set during a CUDA SDK installation.
If this fails - the prelude include of `cuda_fp16.h` will most likely fail on NVRTC invocation.
CUDA has the `__half` and `__half2` types defined in `cuda_fp16.h`. The `__half2` can produce results just as quickly as doing the same operation on `__half` - in essence for some operations `__half2` is [SIMD](https://en.wikipedia.org/wiki/SIMD) like. The half implementation in Slang tries to take advantage of this optimization.
Since Slang supports up to 4 wide vectors Slang has to build on CUDAs half support. The types _`_half3` and `__half4` are implemented in `slang-cuda-prelude.h` for this reason. It is worth noting that `__half3` is made up of a `__half2` and a `__half`. As `__half2` is 4 byte aligned, this means `__half3` is actually 8 bytes, rather than 6 bytes that might be expected.
One area where this optimization isn't fully used is in comparisons - as in effect Slang treats all the vector/matrix half comparisons as if they are scalar. This could be perhaps be improved on in the future. Doing so would require using features that are not directly available in the CUDA headers.
Wave Intrinsics
===============
There is broad support for [HLSL Wave intrinsics](https://docs.microsoft.com/en-us/windows/win32/direct3dhlsl/hlsl-shader-model-6-0-features-for-direct3d-12), including support for [SM 6.5 intrinsics](https://microsoft.github.io/DirectX-Specs/d3d/HLSL_ShaderModel6_5.html).
Most Wave intrinsics will work with vector, matrix or scalar types of typical built in types - uint, int, float, double, uint64_t, int64_t.
The support is provided via both the Slang stdlib as well as the Slang CUDA prelude found in 'prelude/slang-cuda-prelude.h'. Many Wave intrinsics are not directly applicable within CUDA which supplies a more low level mechanisms. The implementation of most Wave functions work most optimally if a 'Wave' where all lanes are used. If all lanes from index 0 to pow2(n) -1 are used (which is also true if all lanes are used) a binary reduction is typically applied. If this is not the case the implementation fallsback on a slow path which is linear in the number of active lanes, and so is typically significantly less performant.
For more a more concrete example take
```
int sum = WaveActiveSum(...);
```
When computing the sum, if all lanes (32 on CUDA), the computation will require 5 steps to complete (2^5 = 32). If say just one lane is not being used it will take 31 steps to complete (because it is now linear in amount of lanes). So just having one lane disabled required 6 times as many steps. If lanes with 0 - 15 are active, it will take 4 steps to complete (2^4 = 16).
In the future it may be possible to improve on the performance of the 'slow' path, however it will always remain the most efficient generally for all of 0 to pow2(n) - 1 lanes to be active.
It is also worth noting that lane communicating intrinsics performance will be impacted by the 'size' of the data communicated. The size here is at a minimum the amount of built in scalar types used in the processing. The CUDA language only allows direct communication with built in scalar types.
Thus
```
int3 v = ...;
int3 sum = WaveActiveSum(v);
```
Will require 3 times as many steps as the earlier scalar example just using a single int.
## WaveGetLaneIndex
'WaveGetLaneIndex' defaults to `(threadIdx.x & SLANG_CUDA_WARP_MASK)`. Depending on how the kernel is launched this could be incorrect. There other ways to get lane index, for example using inline assembly. This mechanism though is apparently slower than the simple method used here. There is support for using the asm mechnism in the CUDA prelude using the `SLANG_USE_ASM_LANE_ID` preprocessor define to enable the feature.
There is potential to calculate the lane id using the [numthreads] markup in Slang/HLSL, but that also requires some assumptions of how that maps to a lane index.
## Unsupported Intrinsics
* Intrinsics which only work in pixel shaders
+ QuadXXXX intrinsics
Limitations
===========
Some features are not available because they cannot be mapped with appropriate behavior to a target. Other features are unavailable because of resources to devote to more unusual features.
* Not all Wave intrinsics are supported
* There is not complete support for all methods on 'objects' like textures etc.
* Does not currently support combined 'TextureSampler'. A Texture behaves equivalently to a TextureSampler and Samplers are ignored.
* Half type is not currently supported
* GetDimensions is not available on any Texture type currently - as there doesn't appear to be a CUDA equivalent
Language aspects
================
# Arrays passed by Value
Slang follows the HLSL convention that arrays are passed by value. This is in contrast with CUDA where arrays follow C++ conventions and are passed by reference. To make generated CUDA follow this convention an array is turned into a 'FixedArray' struct type.
To get something more similar to CUDA/C++ operation the array can be marked in out or inout to make it passed by reference.
![swh spinner](/static/img/swh-spinner.gif)
Computing file changes ...