v1.00 2021-04-20
Shader Model 6.6 introduces a new option that allows the shader author to specify a wave size that the shader is compatible with.
Wave aware HLSL code is becoming increasingly common, along with operations that operate at the level of the wave, instead of independently per thread. While HLSL is designed to abstract away the wave size being used on the hardware, there are currently some scenarios that require the shader author to write shader code dependent on a particular wave size. WaveMMA (wave_matrix) operations are one example. Other static dependencies on wave size may also be required for optimized shaders, such as threadgroup size and groupshared or local array sizes.
Without this feature, if the driver reports a range of supported wave sizes, the application still cannot guarantee that a shader will be run at the desired wave size. Without this guarantee, shaders may simply fail to produce the expected results at runtime.
With this option, D3D12 runtime validation will fail if shaders in a pipeline state object have a required wave size that is not in the range reported by the driver. This also enables additional compile-time validation for cases such as wave_matrix, where the thread group size must be a multiple of the wave size.
The allowed wave sizes that an HLSL shader may specify are
the powers of 2 between 4 and 128, inclusive.
In other words, the set: [4, 8, 16, 32, 64, 128].
A new attribute may be specified on compute shader entry points, to indicate that the function is only compatible with a specific wave size.
[WaveSize(<numLanes>)]
void main() ...
<numLanes> must be an immediate integer value of
an allowed wave size.
Compute shader entries using this attribute will keep track of this attribute when compiled into a library for use when linking to a final shader target.
Since HLSL is normally designed to be wave-size agnostic, use of this feature will result in a compiler warning. This warning can be turned off with a compiler option or a pragma. Warning details and option name TBD. See Issue #6.
The wave size will be captured in DXIL metadata, as a new extended property from the entry point metadata list, so that the driver can select an appropriate wave size for the shader.
In DxilMetadataHelper.h, assuming kDxilWaveSizeTag is assigned the next available value of 11:
class DxilMDHelper {
...
static const unsigned kDxilWaveSizeTag = 11;
; DXIL Example
!dx.entryPoints = !{!1}
!1 = !{..., !2}
!2 = !{i32 11, !3} ; kDxilWaveSizeTag = 11
!3 = !{i32 <numLanes>}
<numLanes> must be an immediate integer value of
an allowed wave size.
The D3D12 runtime will validate shaders with wave size specified against the device reported range of wave sizes to ensure the desired size is within the range of sizes supported by the device.
The device must run the shader at the wave size specified in the DXIL metadata, if any.
All power of 2 sizes in the range reported by the driver must be supported for selection by the shader, and wave ops must be supported at each size. Wave sizes supported by a device under some conditions, but not compatible with wave intrinsics and shader selected wave sizes for the compute shader stage, should not be included in the range of wave size support reported by the driver.
A device that reports support for Shader Model 6.6, and reports support for the WaveOps optional feature, must support this feature. See Issue 2
| Version | Date | Description |
|---|---|---|
| 1.00 | 20 Apr 2021 | Minor Edits for Publication |
| 0.3 | 2020-05-11 | Add warning, update justification |
| 0.2 | 2020-04-21 | Switch to using attribute only, exclude DXR |
| 0.1 | 2020-04-16 | First Draft |