diff options
| author | jsmall-nvidia <jsmall@nvidia.com> | 2020-03-27 16:16:27 -0400 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2020-03-27 16:16:27 -0400 |
| commit | e267ce24e37b9b7f98921f75abc150c1463b1d6d (patch) | |
| tree | 331660a83ae0b72116b79b5cc8bf7a9c06555db5 /docs | |
| parent | 5b0b8436123aa2faa9b682ed45efe2bd7edbf01b (diff) | |
Adds WaveShuffle intrinsic (#1298)
* Support for WaveReadLaneAt with dynamic (but uniform across Wave) on Vk by enabling VK1.4.
Fixed wave-lane-at.slang test to test with laneId that is uniform across the Wave.
* Added WaveShuffle intrinsic.
Test for WaveShuffle intrinsic.
* Added some documentation on WaveShuffle
* Fix that version required for subgroupBroadcast to be non constexpr is actually 1.5
Diffstat (limited to 'docs')
| -rw-r--r-- | docs/target-compatibility.md | 11 |
1 files changed, 11 insertions, 0 deletions
diff --git a/docs/target-compatibility.md b/docs/target-compatibility.md index e8edeeaf5..6967f7454 100644 --- a/docs/target-compatibility.md +++ b/docs/target-compatibility.md @@ -20,6 +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 | Tesselation | Yes ^ | Yes ^ | No + | No | No | Graphics Pipeline | Yes | Yes | Yes | No | No | Ray Tracing DXR 1.0 | No | Yes ^ | Yes ^ | No | No @@ -56,6 +57,16 @@ tex.GetDimensions is the GetDimensions method on 'texture' objects. This is not SM6.5 Wave Intrinsics are supported, but requires a downstream DXC compiler that supports SM6.5. As it stands the DXC shipping with windows does not. +## WaveShuffle + +WaveShuffle is an intrinsic added to the Slang stdlibrary to expose the glsl `subgroupShuffle` intrinsics and allow loosened requirements on laneId. + +`HLSL` uses `WaveReadLaneAt` and this requires the `laneId` must be 'dynamically uniform' across the wave. WaveShuffle has the same functionality but relaxes this restriction. + +`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. + +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. + ## Tesselation Although tesselation stages should work on D3D11 and D3D12 they are not tested within our test framework, and may have problems. |
