VK_HUAWEI_cluster_culling_shader

1. Problem Statement

When drawing a scene with a massive amount of geometry, it is necessary to remove the invisible geometry to decrease redundant drawing, a common approach used to remove invisible geometry is called cluster culling, a cluster is a subset of a mesh that has as many shared vertices as possible, alternatively, a cluster can also be an entire mesh, cluster will be pre-computed and stored with the geometry to avoid computation at runtime.

The GPU has numerous of thread and parallel computing capabilities, so it is more suitable for culling many clusters than the CPU. Many developers use compute shader for GPU culling tasks. Because compute shader can not generate output which directly connected to the existing rendering pipeline. It is necessary to separate culling and rendering into two passes. First, the culling pass processes the whole scene and updates the MDI command, and then uses the MDI method to draw during the rendering pass.

2. Solution space

Provide a new extension to connect the output of the compute shader to the existing rendering pipeline, in addition, when drawing a visible cluster, an appropriately shading rate can also be configured. developers who originally used compute shader for culling can easily migrate to this new extension and have better performance.

3. Proposal.

3.1. Cluster culling shader

This extension allowing application to use a new programmable shader type — Cluster Culling Shader — to execute geometry culling on GPU. This mechanism does not require pipeline barrier between compute shader and other rendering pipeline.

This new shader type have execution environments similar to that of compute shaders, where a collection of shader invocations form a workgroup and cooperate to perform cluster based culling and level-of-detail selection, a shader invocation can emit a group of built-in output variables treated as a drawing command which can drives subsequent rendering pipeline to draw geometries of cluster with a specific shading rate, e.g. the distance between a cluster and the view point can be used to determine the shading rate of the cluster. These capabilities enables the cluster culling shader to reduce the rendering loading more effectively.

It should be noted that the usage of per-cluster shading rate has the following restrictions: 1. CCS and Vertex Shader cannot output shading rate at the same time. 2. The per-cluster shading rate output by CCS will be regarded as per-primitive shading rate in combiner operation. 3. If CCS does not output per-cluster shading rate, the rules of combiner operation remain unchanged.

3.2. API changes

3.2.1. shader stage and synchronization

Extending VkShaderStageFlagBits:

VK_SHADER_STAGE_CLUSTER_CULLING_BIT_HUAWEI specifies the cluster shader stage.

Extending VkPipelineStageFlagBits2:

VK_PIPELINE_STAGE2_CLUSTER_CULLING_SHADER_BIT_HUAWEI specifies the cluster pipeline stage for synchronization.

3.2.2. New structure

Extending VkPhysicalDeviceFeatures2, VkDeviceCreateInfo:

VkPhysicalDeviceClusterCullingShaderFeaturesHUAWEI

Extending VkPhysicalDeviceProperties2:

VkPhysicalDeviceClusterCullingShaderPropertiesHUAWEI

3.2.3. drawcall

dispatching command are recording into a command buffer and when executed by a queue, it will produce work which executes according to the bound Cluster Culling Shader pipeline.

To Record a Cluster Culling Shader command:

void vkCmdDrawClusterHUAWEI(
    VkCommandBuffer     commandBuffer,
    uint32_t            groupCountX,
    uint32_t            groupCountY,
    uint32_t            groupCountZ );
  • commandBuffer is the command buffer into which the command will be recorded.

  • groupCountX is the number of local workgroups to dispatch in the X dimension.

  • groupCountY is the number of local workgroups to dispatch in the Y dimension

  • groupCountZ is the number of local workgroups to dispatch in the Z dimension When the command is executed, a global workgroup consisting of groupCountX * groupCountY * groupCountZ local workgroup is assembled.

To record an indirect Cluster Culling Shader command:

void vkCmdDrawClusterIndirectHUAWEI(
    VkCommandBuffer     commandBuffer,
    vkBuffer            buffer,
    vkDeviceSize        offset );
  • commandBuffer is the command buffer into which the command will be recorded.

  • buffer is the buffer containing dispatch parameters.

  • offset is the byte offset into buffer where parameters begin.

vkCmdDrawClusterIndirectHUAWEI behaves similarly to vkCmdDrawClusterHUAWEI except that the parameters are read by the device from a buffer during execution.

3.2.4. feature

VkPhysicalDeviceClusterCullingShaderFeaturesHUAWEI - Structure describing cluster culling shading features that can be supported by an implementation.

VkPhysicalDeviceClusterCullingShaderFeaturesHUAWEI structure is defined as:

Typedef struct VkPhysicalDeviceClusterCullingShaderFeaturesHUAWEI {
    VkStructureType             sType;
    void*                       pNext;
    VkBool32                    clustercullingShader;
    VkBool32                    multiviewClusterCullingShader;
    VkBool32                    clusterShadingRate;
}VkPhysicalDeviceClusterCullingShaderFeaturesHUAWEI;
  • sType is the type of this structure.

  • pNext is NULL or a pointer to a structure extending this structure.

  • clustercullingShader indicates whether the cluster culling stage is supported.

  • multiviewClusterCullingShader indicates whether multiview can be used with cluster culling shader.

  • clusterShadingRate specifies whether the per-cluster shading rate is supported.

If the VkPhysicalDeviceClusterCullingShaderFeaturesHUAWEI structure is included in the pNext chain of the VkPhysicalDeviceFeature2 structure passed to vkPhysicalDeviceFeature2, it is filled in to indicate whether each corresponding feature is supported. VkPhysicalDeviceClusterCullingShaderFeaturesHUAWEI can also be used in the pNext chain of VkDeviceCreateInfo to selectively enable these features.

VkPhysicalDeviceClusterCullingShaderVrsFeaturesHUAWEI - Structure describing whether cluster culling shading supported per-cluster shading rate.

VkPhysicalDeviceClusterCullingShaderVrsFeaturesHUAWEI structure is defined as:

Typedef struct VkPhysicalDeviceClusterCullingShaderVrsFeaturesHUAWEI {
    VkStructureType		sType;
    void*				pNext;
    VkBool32            clusterShadingRate;
}VkPhysicalDeviceClusterCullingShaderVrsFeaturesHUAWEI;
  • sType is the type of this structure.

  • pNext is NULL or a pointer to a structure extending this structure.

  • clusterShadingRate specifies whether the per-cluster shading rate is supported.

To query whether Cluster Culling Shader support per-cluster shading rate, include a VkPhysicalDeviceClusterCullingShaderVrsFeaturesHUAWEI structure in the pNext chain of the VkPhysicalDeviceClusterCullingShaderFeaturesHUAWEI structure passed to vkPhysicalDeviceFeature2.

3.2.5. property

VkPhysicalDeviceClusterCullingShaderPropertiesHUAWEI - Structure describing cluster culling shading properties.

Typedef struct VkPhysicalDeviceClusterCullingShaderPropertiesHUAWEI {
    VkStructureType             sType;
    void*                       pNext;
    uint32_t                    maxWorkGroupCount[3];
    uint32_t                    maxWorkGroupSize[3];
    uint32_t                    maxOutputClusterCount;
}VkPhysicalDeviceClusterCullingShaderFeaturesHUAWEI;
  • sType is the type of this structure.

  • pNext is NULL or a pointer to a structure extending this structure.

  • maxWorkgroupCount is the maximum number of local workgroups that can be launched by a single command. These three value represent the maximum local workgroup count in the X, Y and Z dimensions, respectively. In the current implementation, the values of Y and Z are both implicitly set as one. groupCountX of DrawCluster* command must be less than or equal to maxWorkGroupCount[0].

  • maxWorkGroupSize is the maximum size of a local workgroup. These three value represent the maximum local workgroup size in the X, Y and Z dimensions, respectively. The x, y and z sizes, as specified by the LocalSize or LocalSizeId execution mode or by the object decorated by the WorkgroupSize decoration in shader modules, must be less than or equal to the corresponding limit.

  • maxOutputClusterCount is the maximum number of output clusters that a single workgroup may emit.

If the VkPhysicalDeviceClusterCullingShaderPropertiesHUAWEI structure is included in the pNext chain of the VkPhysicalDeviceProperties2 structure passed to vkGetPhysicalDeviceProperties2, it is filled in with each corresponding implementation-dependent property.

3.3. SPIR-V changes

3.3.1. new capability

ClusterCullingShadingHUAWEI

3.3.2. execution model

ClusterCullingHUAWEI

3.3.3. built-in

Cluster Culling Shaders have the following built-in output variables, these variables form an aforementioned drawing command.

  • IndexCountHUAWEI is the number of vertices to draw.

  • VertexCountHUAWEI is the number of vertices to draw.

  • InstanceCountHUAWEI is the number of instances to draw.

  • FirstIndexHUAWEI is the base index within the index buffer.

  • FirstVertexHUAWEI is the index of the first vertex to draw.

  • VertexOffsetHUAWEI is the value added to the vertex index before indexing into the vertex buffer.

  • FirstInstanceHUAWEI┬áis the instance ID of the first instance to draw.

  • ClusterIdHUAWEI is the index of cluster being rendered by this drawing command. Cluster Culling Shader passes this id to vertex shader for cluster related information fetching. When cluster culling shader enable, gl_DrawID will be replaced by gl_ClusterIDHUAWEI in Vertex Shader.

  • ClusterShadingRateHUAWEI is the shading rate of cluster being rendering by this drawing command. if VkPhysicalDeviceClusterCullingShaderFeaturesHUAWEI::clusterShadingRate is enabled, ClusterShadingRateHUAWEI is settable from Cluster Culling Shader which support coarse shading.

3.3.4. new function.

  • OpDispatchClusterHUAWEI

Any invocation in Cluster Culling Shader can execute this instruction more than once, after execution, it will emite the Cluster Culling Shader built-in output variables which describe in 3.3.3 to the subsequent rendering pipeline. While a workgroup is done, GPU creates warps for VS according to these output variables, all invocations in VertexShader are responsible for shading the vertices.

3.4. GLSL changes

New write-only output blocks are defined for built-in output variables:

Type 1 (non-indexed mode):
out gl_PerClusterHUAWEI
{
    uint gl_VertexCountHUAWEI;
    uint gl_InstanceCountHUAWEI;
    uint gl_FirstVertexHUAWEI;
    uint gl_FirstInstanceHUAWEI;
    uint gl_ClusterIdHUAWEI;
    uint gl_ClusterShadingRateHUAWEI;
}
Type 2 (indexed mode):
 out gl_PerClusterHUAWEI
{
    uint gl_IndexCountHUAWEI;
    uint gl_InstanceCountHUAWEI;
    uint gl_FirstIndexHUAWEI ;
    int  gl_VertexOffsetHUAWEI;
    uint gl_FirstInstanceHUAWEI;
    uint gl_ClusterIdHUAWEI;
    uint gl_ClusterShadingRateHUAWEI;
}

A new function is added:

void dispatchClusterHUAWEI(void);