diff options
Diffstat (limited to 'docs/target-compatibility.md')
| -rw-r--r-- | docs/target-compatibility.md | 20 |
1 files changed, 15 insertions, 5 deletions
diff --git a/docs/target-compatibility.md b/docs/target-compatibility.md index 6967f7454..ee5341733 100644 --- a/docs/target-compatibility.md +++ b/docs/target-compatibility.md @@ -20,7 +20,7 @@ Items with ^ means there is some discussion about support later in the document | SM6.0 Wave Intrinsics | No | Yes | Partial | Yes | No | SM6.0 Quad Intrinsics | No | Yes | No + | No | No | SM6.5 Wave Intrinsics | No | Yes ^ | No + | Yes | No -| WaveShuffle | No | Limited ^ | Yes + | Yes | No +| WaveShuffle | No | Limited ^ | Yes | Yes | No | Tesselation | Yes ^ | Yes ^ | No + | No | No | Graphics Pipeline | Yes | Yes | Yes | No | No | Ray Tracing DXR 1.0 | No | Yes ^ | Yes ^ | No | No @@ -59,13 +59,23 @@ SM6.5 Wave Intrinsics are supported, but requires a downstream DXC compiler that ## WaveShuffle -WaveShuffle is an intrinsic added to the Slang stdlibrary to expose the glsl `subgroupShuffle` intrinsics and allow loosened requirements on laneId. +`WaveShuffle` and `WaveBroadcastLaneAt` are Slang specific intrinsic additions to expand the options available around `WaveReadLaneAt`. -`HLSL` uses `WaveReadLaneAt` and this requires the `laneId` must be 'dynamically uniform' across the wave. WaveShuffle has the same functionality but relaxes this restriction. +To be clear this means they will not compile directly on 'standard' HLSL compilers such as `dxc`, but Slang HLSL *output* (which will not contain these intrinsics) can (and typically is) compiled via dxc. -`WaveReadLaneAt` most obviously maps to `subgroupBroadcast` in GLSL. This has the extra restriction the index must be compile time consts. With SPIR-V 1.5 it is allowed to be 'dynamically uniform', but doesn't work on current glslang. +The difference between them can be summarized as follows -NOTE! That using WaveShuffle to target `HLSL` will produce `WaveReadLaneAt` - that means strictly speaking the restriction *still applies*, and the correct behavior will only be seen on hardware that allows the loosed requirements of laneId, on hardware that does not result of `WaveShuffle` is the same as `WaveReadLaneId` which is undefined. +* WaveBroadcastLaneAt - laneId must be a compile time constant +* WaveReadLaneAt - laneId can be dynamic but *MUST* be the same value across the Wave ie 'dynamically uniform' across the Wave +* WaveShuffle - laneId can be truly dynamic (NOTE! That it is not strictly truly available currently on all targets, specifically HLSL) + +Other than the different restrictions on laneId they act identically to WaveReadLaneAt. + +`WaveBroadcastLaneAt` and `WaveReadLaneAt` will work on all targets that support wave intrinsics, with the only current restriction being that on GLSL targets, only scalars and vectors are supported. + +`WaveShuffle` will always work on CUDA/Vulkan. + +On HLSL based targets currently `WaveShuffle` will be converted into `WaveReadLaneAt`. Strictly speaking this means it *requires* the `laneId` to be `dynamically uniform` across the Wave. In practice some hardware supports the loosened usage, and others does not. In the future this may be fixed in Slang and/or HLSL to work across all hardware. For now if you use `WaveShuffle` on HLSL based targets it will be necessary to confirm that `WaveReadLaneAt` has the loosened behavior for all the hardware intended. If target hardware does not support the loosened restrictions it's behavior is undefined. ## Tesselation |