Describes the wave operation intrinsics added to HLSL Shader Model 6.0.

• Shader model 6.0
• Terminology
• Shading language intrinsics
  • Wave Query
  • Wave Vote
  • Wave Broadcast
  • Wave Reduction
  • Wave Scan and Prefix
  • Quad-wide Shuffle operations
• Hardware capability
• Related topics

Shader Model 6.0

For earlier shader models, HLSL programming exposes only a single thread of execution. New wave-level operations are provided, starting with model 6.0, to explicitly take advantage of the parallelism of current GPUs - many threads can be executing in lockstep on the same core simultaneously. For example, the model 6.0 intrinsics enable the elimination of barrier constructs when the scope of synchronization is within the width of the SIMD processor, or some other set of threads that are known to be atomic relative to each other.

Potential use cases include: stream compaction, reductions, block transpose, bitonic sort or Fast Fourier Transforms (FFT), binning, stream de-duplication, and similar scenarios.

Most of the intrinsics appear in pixel shaders and compute shaders, though there are some exceptions (noted for each function). The functions have been added to the requirements for DirectX Feature Level 12.0, under API level 12.

The<type> parameter and return value for these functions implies the type of the expression, the supported types are those from the following list that arealso present in the target shader model for your app:

• half, half2, half3, half4
• float, float2, float3, float4
• double, double2, double3, double4
• int, int2, int3, int4
• uint, uint2, uint3, uint4
• short, short2, short3, short4
• ushort, ushort2, ushort3, ushort4
• uint64_t, uint64_t2, uint64_t3, uint64_t4

Some operations (such as the bitwise operators) only support the integer types.

Terminology

Shading language intrinsics

All the operations of this shader model have been added in a range of intrinsic functions.

Wave Query

The intrinsics for querying a single wave.

Wave Vote

This set of intrinsics compare values across threads currently active from the current wave.

Wave Broadcast

These intrinsics enable all active lanes in the current wave to receive the value from the specified lane, effectively broadcasting it. The return value from an invalid lane is undefined.

Wave Reduction

These intrinsics compute the specified operation across all active lanes in the wave and broadcast the final result to all active lanes. Therefore, the final output is guaranteed uniform across the wave.

Wave Scan and Prefix

These intrinsics apply the operation to each lane and leave each partial result of the computation in the corresponding lane.

Quad-wide Shuffle operations

These intrinsics perform swap operations on the values across a wave known to contain pixel shader quads as defined here. The indices of the pixels in the quad are defined in scan-line or reading order - where the coordinates within a quad are:

+---------> X

| [0] [1]

| [2] [3]

v

Y

These routines work in either compute shaders or pixel shaders. In compute shaders they operate in quads defined as evenly divided groups of 4 within an SIMD wave. In pixel shaders they should be used on waves captured by WaveQuadLanes, otherwise results are undefined.

Hardware capability

In order to check that the wave operation features are available on any specific hardware, callID3D12Device::CheckFeatureSupport, noting the description and use of theD3D12_FEATURE_DATA_D3D12_OPTIONS1 structure.

Related topics

• Programming Guide for HLSL
• Shader Model 6 intrinsics