ComputePipeline Class Reference

Encapsulates a Vulkan compute pipeline and all necessary resources to use it. More...

#include <ComputePipeline.h>

Public Member Functions
	ComputePipeline (const std::string &shaderPath, size_t numSSBOs, size_t pushConstantSize, size_t numStorageImages=0, size_t numSampledImages=0)
	Construct a new compute pipeline. More...
void	Reinitialize (const std::string &shaderPath, size_t numSSBOs, size_t pushConstantSize, size_t numStorageImages=0, size_t numSampledImages=0)
	Wipes the state of this object and recreates it with a new shader binary and configuration. More...
template<class T >
void	BindBuffer (size_t i, AcceleratorBuffer< T > &buf, bool outputOnly=false)
	Binds an input or output SSBO to a descriptor slot. More...
void	BindStorageImage (size_t i, vk::Sampler sampler, vk::ImageView view, vk::ImageLayout layout)
	Binds a storage (output) image to a descriptor slot. More...
void	BindSampledImage (size_t i, vk::Sampler sampler, vk::ImageView view, vk::ImageLayout layout)
	Binds a sampled (input) image to a descriptor slot. More...
template<class T >
void	BindBufferNonblocking (size_t i, AcceleratorBuffer< T > &buf, vk::raii::CommandBuffer &cmdBuf, bool outputOnly=false)
	Binds an input or output SSBO to a descriptor slot. More...
void	Bind (vk::raii::CommandBuffer &cmdBuf)
	Binds the pipeline to a command buffer. More...
template<class T >
void	Dispatch (vk::raii::CommandBuffer &cmdBuf, T pushConstants, uint32_t x, uint32_t y=1, uint32_t z=1)
	Adds a vkCmdDispatch operation to a command buffer to execute the compute shader. More...
template<class T >
void	DispatchNoRebind (vk::raii::CommandBuffer &cmdBuf, T pushConstants, uint32_t x, uint32_t y=1, uint32_t z=1)
	Similar to Dispatch() but does not bind descriptor sets. More...

Static Public Member Functions
static void	AddComputeMemoryBarrier (vk::raii::CommandBuffer &cmdBuf)
	Helper function to insert a shader write-to-read memory barrier in a command buffer. More...

Protected Member Functions
void	DeferredInit ()
	Performs deferred initialization of the compute pipeline the first time the object is used. More...

Protected Attributes
std::string	m_shaderPath
	Filesystem path to the compiled SPIR-V shader binary.
size_t	m_numSSBOs
	Number of SSBO bindings in the shader.
size_t	m_numStorageImages
	Number of output image bindings in the shader.
size_t	m_numSampledImages
	Number of input image bindings in the shader.
size_t	m_pushConstantSize
	Size of the push constants, in bytes.
std::unique_ptr< vk::raii::ShaderModule >	m_shaderModule
	Handle to the shader module object.
std::unique_ptr< vk::raii::Pipeline >	m_computePipeline
	Handle to the Vulkan compute pipeline.
std::unique_ptr< vk::raii::PipelineLayout >	m_pipelineLayout
	Layout of the compute pipeline.
std::unique_ptr< vk::raii::DescriptorSetLayout >	m_descriptorSetLayout
	Layout of our descriptor set.
std::unique_ptr< vk::raii::DescriptorPool >	m_descriptorPool
	Pool for allocating m_descriptorSet from.
std::unique_ptr< vk::raii::DescriptorSet >	m_descriptorSet
	The actual descriptor set storing our inputs and outputs.
std::vector< vk::WriteDescriptorSet >	m_writeDescriptors
	Set of bindings to be written to m_descriptorSet.
std::vector< vk::DescriptorBufferInfo >	m_bufferInfo
	Details about our SSBOs.
std::vector< vk::DescriptorImageInfo >	m_storageImageInfo
	Details about our output images.
std::vector< vk::DescriptorImageInfo >	m_sampledImageInfo
	Details about our input images.

Detailed Description

Encapsulates a Vulkan compute pipeline and all necessary resources to use it.

Supported shaders must have all image bindings numerically after all SSBO bindings.

A ComputePipeline is typically owned by a filter instance.

Prefers KHR_push_descriptor (and some APIs are only available if it is present), but basic functionality is available without it.

Constructor & Destructor Documentation

ComputePipeline()

ComputePipeline::ComputePipeline	(	const std::string &	shaderPath,
		size_t	numSSBOs,
		size_t	pushConstantSize,
		size_t	numStorageImages = 0,
		size_t	numSampledImages = 0
	)

Construct a new compute pipeline.

Initialization is deferred until the first time the shader is bound. This allows creation of a ComputePipeline object to be relatively inexpensive (so many pipelines can be created which may not ever be used).

Parameters

shaderPath	Path to the compiled shader binary
numSSBOs	Number of SSBO bindings
pushConstantSize	Size in bytes of our push constants
numStorageImages	Number of output image bindings
numSampledImages	Number of input image bindings

Member Function Documentation

AddComputeMemoryBarrier()

static void ComputePipeline::AddComputeMemoryBarrier ( vk::raii::CommandBuffer &  cmdBuf )

inlinestatic

Helper function to insert a shader write-to-read memory barrier in a command buffer.

Parameters

cmdBuf

Command buffer to append the pipeline barrier to

Bind()

void ComputePipeline::Bind ( vk::raii::CommandBuffer &  cmdBuf )

inline

Binds the pipeline to a command buffer.

Parameters

cmdBuf

Command buffer to append the bind to

BindBuffer()

template<class T >

void ComputePipeline::BindBuffer ( size_t  i, AcceleratorBuffer< T > &  buf, bool  outputOnly = false  )

inline

Binds an input or output SSBO to a descriptor slot.

This method performs a blocking copy from the CPU to GPU views of the buffer if they are incoherent.

Parameters

i	Descriptor index
buf	The buffer to bind to the slot
outputOnly	Hint that the shader never reads from the buffer, so there is no need to ensure coherence between CPU and GPU cache views of the buffer before executing the shader.

BindBufferNonblocking()

template<class T >

void ComputePipeline::BindBufferNonblocking ( size_t  i, AcceleratorBuffer< T > &  buf, vk::raii::CommandBuffer &  cmdBuf, bool  outputOnly = false  )

inline

Binds an input or output SSBO to a descriptor slot.

This method performs a nonblocking copy from the CPU to GPU views of the buffer if they are incoherent.

Parameters

i	Descriptor index
buf	The buffer to bind to the slot
cmdBuf	Command buffer to append the copy operation, if needed, to
outputOnly	Hint that the shader never reads from the buffer, so there is no need to ensure coherence between CPU and GPU cache views of the buffer before executing the shader.

BindSampledImage()

void ComputePipeline::BindSampledImage ( size_t  i, vk::Sampler  sampler, vk::ImageView  view, vk::ImageLayout  layout  )

inline

Binds a sampled (input) image to a descriptor slot.

Parameters

i	Descriptor index
sampler	Vulkan sampler
view	Vulkan image view
layout	Vulkan image layout

BindStorageImage()

void ComputePipeline::BindStorageImage ( size_t  i, vk::Sampler  sampler, vk::ImageView  view, vk::ImageLayout  layout  )

inline

Binds a storage (output) image to a descriptor slot.

Parameters

i	Descriptor index
sampler	Vulkan sampler
view	Vulkan image view
layout	Vulkan image layout

DeferredInit()

void ComputePipeline::DeferredInit ( )

protected

Performs deferred initialization of the compute pipeline the first time the object is used.

This function actually loads the shader binary and creates descriptor sets etc.

Dispatch()

template<class T >

void ComputePipeline::Dispatch ( vk::raii::CommandBuffer &  cmdBuf, T  pushConstants, uint32_t  x, uint32_t  y = 1, uint32_t  z = 1  )

inline

Adds a vkCmdDispatch operation to a command buffer to execute the compute shader.

If KHR_push_descriptor is not available, performs an updateDescriptorSets. This means only one Dispatch() of a given ComputePipeline can be present in a command buffer at a time.

If KHR_push_descriptor is available, performs a pushDescriptorSetKHR. In this case, arbitrarily many Dispatch() calls on the same ComputePipeline may be submitted to the same command buffer in sequence.

Parameters

cmdBuf	Command buffer to append the dispatch operation to
pushConstants	Constants to pass to the shader
x	X size of the dispatch, in thread blocks
y	Y size of the dispatch, in thread blocks
z	Z size of the dispatch, in thread blocks

DispatchNoRebind()

template<class T >

void ComputePipeline::DispatchNoRebind ( vk::raii::CommandBuffer &  cmdBuf, T  pushConstants, uint32_t  x, uint32_t  y = 1, uint32_t  z = 1  )

inline

Similar to Dispatch() but does not bind descriptor sets.

This allows multiple consecutive invocations of the same shader (potentially with different dispatch sizes or push constant values) in the same command buffer, even without KHR_push_descriptor, as long as the same set of input and output descriptors are used by each invocation.

If KHR_push_descriptor is available, performs a vkPushDescriptorSetKHR. If not, descriptors are untouched.

Parameters

cmdBuf	Command buffer to append the dispatch operation to
pushConstants	Constants to pass to the shader
x	X size of the dispatch, in thread blocks
y	Y size of the dispatch, in thread blocks
z	Z size of the dispatch, in thread blocks

Reinitialize()

void ComputePipeline::Reinitialize	(	const std::string &	shaderPath,
		size_t	numSSBOs,
		size_t	pushConstantSize,
		size_t	numStorageImages = 0,
		size_t	numSampledImages = 0
	)

Wipes the state of this object and recreates it with a new shader binary and configuration.

Initialization is deferred until the first time the shader is bound. This allows creation of a ComputePipeline object to be relatively inexpensive (so many pipelines can be created which may not ever be used).

Parameters

shaderPath	Path to the compiled shader binary
numSSBOs	Number of SSBO bindings
pushConstantSize	Size in bytes of our push constants
numStorageImages	Number of output image bindings
numSampledImages	Number of input image bindings

---

The documentation for this class was generated from the following files:

• lib/scopehal/ComputePipeline.h
• lib/scopehal/ComputePipeline.cpp