Vulkan acceleration

Collaboration diagram for Vulkan acceleration:

Files
file	ComputePipeline.cpp
	Implementation of ComputePipeline.
file	ComputePipeline.h
	Declaration of ComputePipeline.
file	NamedDebugRange.h
	Declaration of NamedDebugRange.
file	VulkanInit.cpp
	Vulkan initialization.

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

Functions
bool	VulkanInitInstance (bool skipGLFW, bool &vulkan11Available, bool &vulkan12Available, bool &hasPhysicalDeviceProperties2)
	Do context-level Vulkan initialization.
void	VulkanPrintPhysicalDeviceInfo (size_t i, vk::raii::PhysicalDevice &device, bool hasPhysicalDeviceProperties2)
	Log info about a Vulkan physical device.
void	VulkanCreateDevice (vk::raii::PhysicalDevice &device, bool vulkan11Available, bool vulkan12Available, bool hasPhysicalDeviceProperties2)
	Actually create the selected Vulkan device and set all of the global Vulkan feature flags appropriately.
bool	VulkanInit (bool skipGLFW)
	Initialize a Vulkan context for compute.
bool	IsDevicePreferred (const vk::PhysicalDeviceProperties &a, const vk::PhysicalDeviceProperties &b)
	Checks if a given Vulkan device is "better" than another.
void	VulkanCleanup ()
	Free all global Vulkan resources in the correct order.

Variables
vk::raii::Context	g_vkContext
	Global Vulkan context.
unique_ptr< vk::raii::Instance >	g_vkInstance
	Global Vulkan instance.
shared_ptr< vk::raii::Device >	g_vkComputeDevice
	The Vulkan device selected for compute operations (may or may not be same device as rendering)
unique_ptr< vk::raii::CommandPool >	g_vkTransferCommandPool
	Command pool for AcceleratorBuffer transfers.
unique_ptr< vk::raii::CommandBuffer >	g_vkTransferCommandBuffer
	Command buffer for AcceleratorBuffer transfers.
shared_ptr< QueueHandle >	g_vkTransferQueue
	Queue for AcceleratorBuffer transfers.
std::unique_ptr< QueueManager >	g_vkQueueManager
	Allocates QueueHandle objects.
mutex	g_vkTransferMutex
	Mutex for interlocking access to g_vkTransferCommandBuffer and g_vkTransferCommandPool.
uint32_t	g_vkPinnedMemoryType
	Vulkan memory type for CPU-based memory that is also GPU-readable.
uint32_t	g_vkLocalMemoryType
	Vulkan memory type for GPU-based memory (generally not CPU-readable, except on unified memory systems)
uint8_t	g_vkComputeDeviceUuid [16]
	UUID of g_vkComputeDevice.
uint32_t	g_vkComputeDeviceDriverVer
	Driver version of g_vkComputeDevice.
vk::raii::PhysicalDevice *	g_vkComputePhysicalDevice
	Physical device for g_vkComputeDevice.
uint32_t	g_vkPinnedMemoryHeap = 0
	Heap from which g_vkPinnedMemoryType is allocated.
uint32_t	g_vkLocalMemoryHeap = 0
	Heap from which g_vkLocalMemoryType is allocated.
bool	g_hasShaderFloat64 = false
	Indicates whether the float64 type is available for use in shaders and SSBOs.
bool	g_hasShaderInt64 = false
	Indicates whether the int64 type is available for use in shaders and SSBOs.
bool	g_hasShaderInt16 = false
	Indicates whether the int16 type is available for use in shaders and SSBOs.
bool	g_hasShaderInt8 = false
	Indicates whether the int8 type is available for use in shaders and SSBOs.
bool	g_hasShaderAtomicFloat = false
	Indicates whether atomic operations on the float type are available in shaders.
bool	g_hasShaderAtomicInt64 = false
	Indicates whether atomic operations on the int64 type are available in shaders.
bool	g_hasDebugUtils = false
	Indicates whether the VK_EXT_debug_utils extension is available.
bool	g_hasMemoryBudget = false
	Indicates whether the VK_EXT_memory_budget extension is available.
bool	g_hasPushDescriptor = false
	Indicates whether the VK_KHR_push_descriptor extension is available.
bool	g_vulkanDeviceHasUnifiedMemory = false
	Indicates whether the Vulkan device is unified memory.
size_t	g_maxComputeGroupCount [3] = {0}
	Maximum size of a Vulkan dispatch group for compute shaders, in each axis.
bool	g_vulkanDeviceIsIntelMesa = false
	Indicates that the Vulkan driver is Mesa on an Intel GPU.
bool	g_vulkanDeviceIsAnyMesa = false
	Indicates that the Vulkan driver is Mesa on any GPU.
bool	g_vulkanDeviceIsMoltenVK = false
	Indicates that the Vulkan driver is MoltenVK.
bool	g_vulkanDeviceIsApplePV = false
	Indicates that the Vulkan driver is an Apple paravirtualization device.

Detailed Description

Function Documentation

IsDevicePreferred()

bool IsDevicePreferred	(	const vk::PhysicalDeviceProperties &	a,
		const vk::PhysicalDeviceProperties &	b
	)

Checks if a given Vulkan device is "better" than another.

Parameters

a	One of two devices being considered
b	The second device being considered

Returns

True if we should use device B over A

VulkanInit()

bool VulkanInit

(

bool

skipGLFW

)

Initialize a Vulkan context for compute.

Parameters

skipGLFW

Do not initalize GLFW (workaround for what looks like gtk or video driver bug). This was a workaround used by glscopeclient and should probably be removed as it's not used now.

VulkanInitInstance()

bool VulkanInitInstance	(	bool	skipGLFW,
		bool &	vulkan11Available,
		bool &	vulkan12Available,
		bool &	hasPhysicalDeviceProperties2
	)

Do context-level Vulkan initialization.

Parameters

skipGLFW

Do not initialize GLFW.

This is needed for headless use of libscopehal in unit tests or ATE applications, which may be executed on a machine that does not have a display server running

VulkanPrintPhysicalDeviceInfo()

void VulkanPrintPhysicalDeviceInfo	(	size_t	i,
		vk::raii::PhysicalDevice &	device,
		bool	hasPhysicalDeviceProperties2
	)

Log info about a Vulkan physical device.

Does not actually set any "feature present" flags since we don't yet know if we're going to use this device

Variable Documentation

g_vkQueueManager

std::unique_ptr<QueueManager> g_vkQueueManager

Allocates QueueHandle objects.

This is a single global resource, all QueueHandles must be obtained through this object.

g_vkTransferCommandBuffer

unique_ptr<vk::raii::CommandBuffer> g_vkTransferCommandBuffer

Command buffer for AcceleratorBuffer transfers.

This is a single global resource interlocked by g_vkTransferMutex and is used for convenience and code simplicity when parallelism isn't that important.

g_vkTransferCommandPool

unique_ptr<vk::raii::CommandPool> g_vkTransferCommandPool

Command pool for AcceleratorBuffer transfers.

This is a single global resource interlocked by g_vkTransferMutex and is used for convenience and code simplicity when parallelism isn't that important.

g_vkTransferQueue

shared_ptr<QueueHandle> g_vkTransferQueue

Queue for AcceleratorBuffer transfers.

This is a single global resource interlocked by g_vkTransferMutex and is used for convenience and code simplicity when parallelism isn't that important.

g_vulkanDeviceHasUnifiedMemory

bool g_vulkanDeviceHasUnifiedMemory = false

Indicates whether the Vulkan device is unified memory.

Unified memory means that all memory allocations for compute shaders are from a single heap which is accessible to both the CPU and GPU with roughly equal performance, and there is no device-local memory with higher performance.

For example, AMD APUs, Intel integrated GPUs, and Apple Silicon platforms are unified memory.

Discrete PCIe GPUs typically are not.