hlsl-specs

Expanded Comparison Sampling

Introduction

Sampler operations in HLSL perform either a comparison sampler or non-comparison sampler operation. Comparison sample operations will perform the compare operation specified in the sampler between the sampled value and a provided value and produce a 0 or 1 value indicating the failure or success of the comparison respectively. In linear filtered samples, multiple values will be blended to produce the complete sample operation result. Sampling operations sample from a texture according to the settings in the provided sampler object. The most useful forms of sampling involve the calculation of a MIP level to determine which level(s) of detail(LOD) of the sampled texture should be used and how they should be weighted.

Motivation

Different non-comparison sampler operations calculate the MIP level in different ways:

Note that throughout this document, the template parameter CoordsCt represents the number of indexable coordinate dimensions and DimsCt represents the number of native dimension axes in the given texture resource. This means that CoordsCt includes the native dimensions as well as an additional dimension for the slices in array textures and DimsCt excludes the additional array texture slice dimension.

Comparison sampler operations have fewer options for how the LOD is calculated:

Additionally, the level of detail value can be calculated and retrieved directly for non-comparison samplers:

There are no CalculateLevelOfDetail variants for comparison samplers.

Without the full complement of comparison sampler operations, techniques that require LOD calculation beyond the implicit calculations are missing the full support for different approaches to LOD calculation. In particular, this can affect algorithms involving shadow maps and any case where standard samples have comparison sample counterparts that need to match. In addition, this represents a discrepancy with comparable platforms where this deficiency makes porting, emulating, or developing for multiple platforms difficult or impossible.

Proposed solution

To add full comparison sampling support, HLSL will introduce an optional feature that adds these new comparison sampling builtin methods to all texture objects that currently support the non-comparison methods:

Only the simplest overloads of the new and existing builtin methods are described in detail here for the sake of brevity, Though not included above, all relevant overloads of SampleBias and SampleGrad will be reflected in their new SampleCmpBias and SampleCmpGrad counterparts. The parameters not explicitly described here will function exactly as they do when used by the non-comparison builtin methods:

To add LOD calculation for comparison samplers, HLSL will introduce and require these new comparison sampler method overloads to all texture objects that currently support the non-comparison method overloads:

Detailed Design

HLSL additions

These new HLSL texture methods will be added:

template<typename ElementType, int CoordsCt, int DimsCt>
vector<ElementType, DimsCt> TextureBase::SampleCmpGrad(
    SamplerComparisonState S,
    vector<float, CoordsCt> Location,
    float CompareValue,
    vector<float, DimsCt> DDX,
    vector<float, DimsCt> DDY,
    vector<int, DimsCt> Offset = 0,
    float Clamp = 0.0f);

template<typename ElementType, int CoordsCt, int DimsCt>
vector<ElementType, DimsCt> TextureCube::SampleCmpGrad(
    SamplerComparisonState S,
    vector<float, CoordsCt> Location,
    float CompareValue,
    vector<float, DimsCt> DDX,
    vector<float, DimsCt> DDY,
    float Clamp = 0.0f);

Where TextureBase represents any of Texture1D[Array] or Texture2D[Array] texture object types. The TextureCube method does not have an Offset parameter in keeping with other TextureCube methods. TextureCubeArray does not support the SampleCmpGrad method. Note that these both have an additional overload that includes all the default parameters explicitly specified and a final output status parameter. SampleCmpGrad is available in all shader stages.

template<typename ElementType, int CoordsCt, int DimsCt>
vector<ElementType, DimsCt> TextureBase::SampleCmpBias(
      SamplerComparisonState S,
      vector<float, CoordsCt> Location,
      float CompareValue,
      float Bias,
      vector<int, DimsCt> Offset = 0,
      float Clamp = 0.0f,
);
template<typename ElementType, int CoordsCt, int DimsCt>
vector<ElementType, DimsCt> TextureCube[Array]::SampleCmpBias(
      SamplerComparisonState S,
      vector<float, CoordsCt> Location,
      float CompareValue,
      float Bias,
      float Clamp = 0.0f
);

Where TextureBase represents any of Texture1D[Array] or Texture2D[Array] texture object types. The TextureCube* methods do not have an Offset parameter in keeping with other TextureCube* methods. Both methods have an additional overload that includes all the default parameters explicitly specified and a final output status parameter. SampleCmpBias is only available in the pixel and compute shader stages where implicit derivatives can be calculated.

The semantics of the comparisons are identical to the existing SampleCmp and SampleCmpLevel[Zero] methods. The comparison is performed on the R channel of the texture after applying SRV swizzling. The number of components in the Location and Offset value depends on the type of the texture being sampled from.

The only difference between these new functions and the existing comparison sample ops is how the LOD is computed. The computation of the LOD is identical to the equivalent non-comparison sampling operation. E.g. SampleCmpGrad behaves like SampleGrad, except that the comparison is performed on the result, and likewise for SampleCmpBias relative to SampleBias. More specifically, SampleCmpGrad uses explicit gradients to compute the LOD rather than implicit ones from other pixels in the current pixel shader quad, and SampleCmpBias computes an implicit LOD and then adds a shader-computed bias value to the result.

Two new HLSL overloads for calculating the level of detail will be added:

template<int CoordsCt>
float TextureBase::CalculateLevelOfDetail(
      SamplerComparisonState S,
      vector<float, CoordsCt> Location
);
template<int CoordsCt>
float TextureBase::CalculateLevelOfDetailUnclamped(
      SamplerComparisonState S,
      vector<float, CoordsCt> Location
);

Where TextureBase represents any texture object type except Texture2DMS*. CalculateLevelOfDetail is only available in the pixel and compute shader stages where implicit derivatives can be calculated.

The semantics of the level of detail calculations are identical to the existing CalculateLevelOfDetail overloads. The only difference with the new overloads is that they calculate the level of detail using comparison sampler objects.

DXIL Additions

Two new DXIL intrinsics are added:

declare %dx.types.ResRet.f32 @dx.op.sampleCmpGrad.f32(
    i32,                  ; opcode
    %dx.types.Handle,     ; texture handle
    %dx.types.Handle,     ; sampler handle
    float,                ; coordinate c0
    float,                ; coordinate c1
    float,                ; coordinate c2
    float,                ; coordinate c3
    i32,                  ; offset o0
    i32,                  ; offset o1
    i32,                  ; offset o2
    float,                ; compare value
    float,                ; ddx0
    float,                ; ddx1
    float,                ; ddx2
    float,                ; ddy0
    float,                ; ddy1
    float,                ; ddy2
    float)                ; clamp

Valid resource types for the texture handle parameter and the corresponding active coordinates and offsets are the same as for the sampleGrad operation except that TextureCubeArray resources are invalid. Valid active ddx and ddy are the same as offsets. The sampler handle parameter must be of type SamplerComparisonState.

declare %dx.types.ResRet.f32 @dx.op.sampleCmpBias.f32(
    i32,                  ; opcode
    %dx.types.Handle,     ; texture handle
    %dx.types.Handle,     ; sampler handle
    float,                ; coordinate c0
    float,                ; coordinate c1
    float,                ; coordinate c2
    float,                ; coordinate c3
    i32,                  ; offset o0
    i32,                  ; offset o1
    i32,                  ; offset o2
    float,                ; compare value
    float,                ; bias: in [-16.f,15.99f]
    float)                ; clamp

Valid resource types for the texture handle parameter and the corresponding active coordinates and offsets are the same as for the sampleBias operation except that TextureCubeArray resources are invalid for sampleCmpGrad. The sampler handle parameter must be of type SamplerComparisonState.

The existing dx.op.calculateLOD intrinsic is reused for the new CalculateLevelOfDetail overloads taking SamplerComparisonState parameters.

SPIR-V Additions

To implement the new SampleCmpGrad and SampleCmpBias texture methods, the existing OpImage*SampleDrefImplicitLod operations can be used with the optional image operands set to the value for Grad(0x4) or Bias(0x1) set to the values provided by the user.

To implement the new CalculateLevelOfDetail[Unclamped] overloads, the existing OpImageQueryLod operation may be used with comparison state images.

Diagnostic Changes

Removed errors

New Errors

Validation Changes

Runtime Additions

Runtime information

A new feature info flag is added to SFI0:

Tag Bit Value Description
SampleCmpGradientOrBias 31 0x80000000 SampleCmpBias or SampleCmpGrad are used

It indicates to the runtime that the shader makes use of SampleCmpBias or SampleCmpGrad. This indicates to the runtime that that the shader requires the presence of the corresponding capability bit (described below) indicating support.

Device Capability

Applications can query the availability of these features by passing D3D12_FEATURE_D3D12_OPTIONS21 as the Feature parameter and retrieving the pFeatureSupportData parameter as a struct of type D3D12_FEATURE_DATA_D3D12_OPTIONS21. The relevant parts of these structs are defined below.

typedef enum D3D12_FEATURE {
    ...
    D3D12_FEATURE_D3D12_OPTIONS21
} D3D12_FEATURE;

typedef struct D3D12_FEATURE_DATA_D3D12_OPTIONS21 {
    ...
    BOOL SampleCmpGradientAndBiasSupported;
} D3D12_FEATURE_DATA_D3D12_OPTIONS21;

ExpandedComparisonSamplingSupported is a boolean that specifies whether the expanded comparison sample methods indicated here are supported by the given hardware and runtime.

Testing

Correct Behavior Testing

Verify that SampleCmpGrad and the new CalculateLevelOfDetail* overloads generate appropriate code including the new DXIL and SPIRV intrinsics for all shader stages. Verify that SampleCmpBias produces appropriate code including the new DXIL and SPIRV intrinsics for pixel, compute, and node shader stages. Where the DerivativesInMeshAndAmplificationShadersSupported capability bit is supported, verify that SampleCmpBias produces appropriate code including the new DXIL and SPIRV intrinsics for mesh and amplification shader stages. For all of the above, verify that no validation errors are produced for the generated code. For all of the above, wherever SampleCmpGrad and SampleCmpBias are used, check that the appropriate shader flag is set. The above should all be doable with LIT filecheck tests.

Diagnostics Testing

Use SampleCmpBias in code reachable from the entry point of each unsupported shader stage (vertex, geometry, tesselation stages, ray tracing stages, and mesh stages if lacking hardware support for mesh derivatives) Using SampleCmpBias, SampleCmpGrad, and the new CalculateLevelOfDetail* overloads in an unsupported shader model on a supported stage and verify that a special error indicating the required version is produced.

Validation Testing

Test that the Bias range limits tested by SampleBiasFail are respected by dx.op.sampleCmpBias. (see Instr.ImmBiasForSampleB) Test that DXIL validation of allowed resource types must be performed for dx.op.sampleCmpBias and dx.op.sampleCmpGrad:

Execution Testing

Since there is no defined equation to determine which MIP level an implementation should use, the best approach is to compare the results of different operations involving MIP levels to ensure that they are consistent.

To test CalculateLevelOfDetail* methods for each supported shader stage, first fill each MIP level of a texture with a single sentinel value matching the MIP level. Then use CalculateLevelOfDetail to calculate the expected MIP level. Finally, use SampleCmp with the comparison sampler with comparison mode equal and the comparison value set to the calculated MIP level to determine if the calculated value matches the sentinal values in the texture. Return the result of the comparison sample. It should be 1 if the calculated level was correct. This will ensure that CalculateLevelOfDetail correctly returns the MIP value that sample would have used. Ensure as well that CalculateLevelOfDetail clamps the MIP result and CalculateLevelOfDetailUnclamped does not.

Given the nature of SampleCmp operations, the tests will have to be performed with comparison values and success determined based on whether it matches what is expected. At any point, the tests described below have two possible values. This allows quickly testing for one or the other to determine how to advance the test. If it’s the current value, the next iteration starts. If it’s the next expected value, the expected value is updated before the next iteration starts. If any other value is found, the test should fail.

To test the SampleCmpBias method, use a texture with sentinel values in each MIP level, start with no bias and determine the default MIP level, then increase the bias ensuring that the MIP level the sample uses switches to the expected next level at a reasonable bias: somewhere between just over 0.0 and 1.0. Repeat for a few levels.

To test the SampleCmpGrad method, use a texture with sentinel values in each MIP level, start with gradients of 0 to determine the base MIP level. Increase the gradients gradually and ensure that MIP level the sample uses switches to the expected next level at reasonable gradient values. Repeat for a few levels.

This site is open source. Improve this page.