Files
test2/source/blender/gpu/vulkan/vk_device.hh
Jeroen Bakker 47fc709987 Fix #146313: Vulkan: Memory leak HDR backbuffer shader
Oversight in !146170 where shader pointers where overwritten each
frame. Should also be a performance regression.

Pull Request: https://projects.blender.org/blender/blender/pulls/146319
2025-09-15 21:01:25 +02:00

489 lines
15 KiB
C++

/* SPDX-FileCopyrightText: 2023 Blender Authors
*
* SPDX-License-Identifier: GPL-2.0-or-later */
/** \file
* \ingroup gpu
*/
#pragma once
#include <atomic>
#include "BLI_task.h"
#include "BLI_threads.h"
#include "BLI_utility_mixins.hh"
#include "BLI_vector.hh"
#include "render_graph/vk_render_graph.hh"
#include "render_graph/vk_resource_state_tracker.hh"
#include "vk_buffer.hh"
#include "vk_common.hh"
#include "vk_debug.hh"
#include "vk_descriptor_pools.hh"
#include "vk_descriptor_set_layouts.hh"
#include "vk_memory_pool.hh"
#include "vk_pipeline_pool.hh"
#include "vk_resource_pool.hh"
#include "vk_samplers.hh"
namespace blender::gpu {
class VKBackend;
struct VKExtensions {
/** Does the device support VkPhysicalDeviceVulkan12Features::shaderOutputViewportIndex. */
bool shader_output_viewport_index = false;
/** Does the device support VkPhysicalDeviceVulkan12Features::shaderOutputLayer. */
bool shader_output_layer = false;
/**
* Does the device support
* VkPhysicalDeviceFragmentShaderBarycentricFeaturesKHR::fragmentShaderBarycentric.
*/
bool fragment_shader_barycentric = false;
/**
* Does the device support VK_KHR_dynamic_rendering_local_read enabled.
*/
bool dynamic_rendering_local_read = false;
/**
* Does the device support VK_EXT_dynamic_rendering_unused_attachments.
*/
bool dynamic_rendering_unused_attachments = false;
/**
* Does the device support VK_EXT_external_memory_win32/VK_EXT_external_memory_fd
*/
bool external_memory = false;
/** VK_KHR_maintenance4 */
bool maintenance4 = false;
/**
* Does the device support VK_EXT_descriptor_buffer.
*/
bool descriptor_buffer = false;
/**
* Does the device support logic ops.
*/
bool logic_ops = false;
/**
* Does the device support VK_EXT_memory_priority
*/
bool memory_priority = false;
/**
* Does the device support VK_EXT_pageable_device_local_memory
*/
bool pageable_device_local_memory = false;
/** Log enabled features and extensions. */
void log() const;
};
/* TODO: Split into VKWorkarounds and VKExtensions to remove the negating when an extension isn't
* supported. */
struct VKWorkarounds {
/**
* Some devices don't support pixel formats that are aligned to 24 and 48 bits.
* In this case we need to use a different texture format.
*
* If set to true we should work around this issue by using a different texture format.
*/
bool not_aligned_pixel_formats = false;
struct {
/**
* Is the workaround enabled for devices that don't support using VK_FORMAT_R8G8B8_* as vertex
* buffer.
*/
bool r8g8b8 = false;
} vertex_formats;
};
/**
* Shared resources between contexts that run in the same thread.
*/
class VKThreadData : public NonCopyable, NonMovable {
public:
/** Thread ID this instance belongs to. */
pthread_t thread_id;
VKDescriptorPools descriptor_pools;
VKDescriptorSetTracker descriptor_set;
/**
* The current rendering depth.
*
* GPU_rendering_begin can be called multiple times forming a hierarchy. The same resource pool
* should be used for the whole hierarchy. rendering_depth is increased for every
* GPU_rendering_begin and decreased when GPU_rendering_end is called. Resources pools are cycled
* when the rendering_depth set to 0.
*/
int32_t rendering_depth = 0;
VKThreadData(VKDevice &device, pthread_t thread_id);
};
class VKDevice : public NonCopyable {
private:
/** Copies of the handles owned by the GHOST context. */
VkInstance vk_instance_ = VK_NULL_HANDLE;
VkPhysicalDevice vk_physical_device_ = VK_NULL_HANDLE;
VkDevice vk_device_ = VK_NULL_HANDLE;
uint32_t vk_queue_family_ = 0;
VkQueue vk_queue_ = VK_NULL_HANDLE;
std::mutex *queue_mutex_ = nullptr;
bool is_initialized_ = false;
/**
* Task pool for render graph submission.
*
* Multiple threads in Blender can build a render graph. Building the command buffer for a render
* graph is faster when doing it in serial. Submission pool ensures that only one task is
* building at a time (background_serial).
*/
TaskPool *submission_pool_ = nullptr;
/**
* All created render graphs.
*/
Vector<render_graph::VKRenderGraph *> render_graphs_;
ThreadQueue *submitted_render_graphs_ = nullptr;
ThreadQueue *unused_render_graphs_ = nullptr;
VkSemaphore vk_timeline_semaphore_ = VK_NULL_HANDLE;
/**
* Last used timeline value.
*
* Must be externally synced by orphaned_data.mutex_get()
*/
TimelineValue timeline_value_ = 0;
VKSamplers samplers_;
VKDescriptorSetLayouts descriptor_set_layouts_;
/**
* Available Contexts for this device.
*
* Device keeps track of each contexts. When buffers/images are freed they need to be removed
* from all contexts state managers.
*
* The contexts inside this list aren't owned by the VKDevice. Caller of `GPU_context_create`
* holds the ownership.
*/
Vector<std::reference_wrapper<VKContext>> contexts_;
/** Allocator used for texture and buffers and other resources. */
VmaAllocator mem_allocator_ = VK_NULL_HANDLE;
/** Limits of the device linked to this context. */
VkPhysicalDeviceProperties vk_physical_device_properties_ = {};
VkPhysicalDeviceDriverProperties vk_physical_device_driver_properties_ = {};
VkPhysicalDeviceIDProperties vk_physical_device_id_properties_ = {};
VkPhysicalDeviceMemoryProperties vk_physical_device_memory_properties_ = {};
VkPhysicalDeviceMaintenance4Properties vk_physical_device_maintenance4_properties_ = {
VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_4_PROPERTIES};
VkPhysicalDeviceDescriptorBufferPropertiesEXT vk_physical_device_descriptor_buffer_properties_ =
{};
/** Features support. */
VkPhysicalDeviceFeatures vk_physical_device_features_ = {};
VkPhysicalDeviceVulkan11Features vk_physical_device_vulkan_11_features_ = {};
VkPhysicalDeviceVulkan12Features vk_physical_device_vulkan_12_features_ = {};
Array<VkExtensionProperties> device_extensions_;
/** Functions of vk_ext_debugutils for this device/instance. */
debug::VKDebuggingTools debugging_tools_;
/* Workarounds */
VKWorkarounds workarounds_;
VKExtensions extensions_;
std::string glsl_vert_patch_;
std::string glsl_geom_patch_;
std::string glsl_frag_patch_;
std::string glsl_comp_patch_;
Vector<VKThreadData *> thread_data_;
Shader *vk_backbuffer_blit_sh_ = nullptr;
public:
render_graph::VKResourceStateTracker resources;
VKDiscardPool orphaned_data;
/** Discard pool for resources that could still be used during rendering. */
VKDiscardPool orphaned_data_render;
VKPipelinePool pipelines;
/** Buffer to bind to unbound resource locations. */
VKBuffer dummy_buffer;
/**
* This struct contains the functions pointer to extension provided functions.
*/
struct {
/* Extension: VK_KHR_dynamic_rendering */
PFN_vkCmdBeginRendering vkCmdBeginRendering = nullptr;
PFN_vkCmdEndRendering vkCmdEndRendering = nullptr;
/* Extension: VK_EXT_debug_utils */
PFN_vkCmdBeginDebugUtilsLabelEXT vkCmdBeginDebugUtilsLabel = nullptr;
PFN_vkCmdEndDebugUtilsLabelEXT vkCmdEndDebugUtilsLabel = nullptr;
PFN_vkSetDebugUtilsObjectNameEXT vkSetDebugUtilsObjectName = nullptr;
PFN_vkCreateDebugUtilsMessengerEXT vkCreateDebugUtilsMessenger = nullptr;
PFN_vkDestroyDebugUtilsMessengerEXT vkDestroyDebugUtilsMessenger = nullptr;
/* Extension: VK_KHR_external_memory_fd */
PFN_vkGetMemoryFdKHR vkGetMemoryFd = nullptr;
#ifdef _WIN32
/* Extension: VK_KHR_external_memory_win32 */
PFN_vkGetMemoryWin32HandleKHR vkGetMemoryWin32Handle = nullptr;
#endif
/* Extension: VK_EXT_descriptor_buffer */
PFN_vkGetDescriptorSetLayoutSizeEXT vkGetDescriptorSetLayoutSize = nullptr;
PFN_vkGetDescriptorSetLayoutBindingOffsetEXT vkGetDescriptorSetLayoutBindingOffset = nullptr;
PFN_vkGetDescriptorEXT vkGetDescriptor = nullptr;
PFN_vkCmdBindDescriptorBuffersEXT vkCmdBindDescriptorBuffers = nullptr;
PFN_vkCmdSetDescriptorBufferOffsetsEXT vkCmdSetDescriptorBufferOffsets = nullptr;
} functions;
VKMemoryPools vma_pools;
const char *extension_name_get(int index) const
{
return device_extensions_[index].extensionName;
}
VkPhysicalDevice physical_device_get() const
{
return vk_physical_device_;
}
const VkPhysicalDeviceProperties &physical_device_properties_get() const
{
return vk_physical_device_properties_;
}
inline const VkPhysicalDeviceMaintenance4Properties &
physical_device_maintenance4_properties_get() const
{
return vk_physical_device_maintenance4_properties_;
}
const VkPhysicalDeviceIDProperties &physical_device_id_properties_get() const
{
return vk_physical_device_id_properties_;
}
inline const VkPhysicalDeviceDescriptorBufferPropertiesEXT &
physical_device_descriptor_buffer_properties_get() const
{
return vk_physical_device_descriptor_buffer_properties_;
}
const VkPhysicalDeviceFeatures &physical_device_features_get() const
{
return vk_physical_device_features_;
}
const VkPhysicalDeviceVulkan11Features &physical_device_vulkan_11_features_get() const
{
return vk_physical_device_vulkan_11_features_;
}
const VkPhysicalDeviceVulkan12Features &physical_device_vulkan_12_features_get() const
{
return vk_physical_device_vulkan_12_features_;
}
VkInstance instance_get() const
{
return vk_instance_;
};
VkDevice vk_handle() const
{
return vk_device_;
}
uint32_t queue_family_get() const
{
return vk_queue_family_;
}
inline VmaAllocator mem_allocator_get() const
{
return mem_allocator_;
}
VKDescriptorSetLayouts &descriptor_set_layouts_get()
{
return descriptor_set_layouts_;
}
debug::VKDebuggingTools &debugging_tools_get()
{
return debugging_tools_;
}
const debug::VKDebuggingTools &debugging_tools_get() const
{
return debugging_tools_;
}
const VKSamplers &samplers() const
{
return samplers_;
}
void init(void *ghost_context);
void reinit();
void deinit();
bool is_initialized() const
{
return is_initialized_;
}
GPUDeviceType device_type() const;
GPUDriverType driver_type() const;
std::string vendor_name() const;
std::string driver_version() const;
/**
* Check if a specific extension is supported by the device.
*
* This should be called from vk_backend to set the correct capabilities and workarounds needed
* for this device.
*/
bool supports_extension(const char *extension_name) const;
const VKWorkarounds &workarounds_get() const
{
return workarounds_;
}
inline const VKExtensions &extensions_get() const
{
return extensions_;
}
std::string glsl_vertex_patch_get() const;
std::string glsl_geometry_patch_get() const;
std::string glsl_fragment_patch_get() const;
std::string glsl_compute_patch_get() const;
shader::GeneratedSource extensions_define(StringRefNull stage_define) const;
/* -------------------------------------------------------------------- */
/** \name Render graph
* \{ */
static void submission_runner(TaskPool *__restrict pool, void *task_data);
render_graph::VKRenderGraph *render_graph_new();
TimelineValue render_graph_submit(render_graph::VKRenderGraph *render_graph,
VKDiscardPool &context_discard_pool,
bool submit_to_device,
bool wait_for_completion,
VkPipelineStageFlags wait_dst_stage_mask,
VkSemaphore wait_semaphore,
VkSemaphore signal_semaphore,
VkFence signal_fence);
void wait_for_timeline(TimelineValue timeline);
void wait_queue_idle();
/**
* Retrieve the last finished submission timeline.
*/
TimelineValue submission_finished_timeline_get() const
{
BLI_assert(vk_timeline_semaphore_ != VK_NULL_HANDLE);
TimelineValue current_timeline;
VkResult result = vkGetSemaphoreCounterValue(
vk_device_, vk_timeline_semaphore_, &current_timeline);
UNUSED_VARS(result);
BLI_assert_msg(
result == VK_SUCCESS && current_timeline != UINT64_MAX,
"Potential driver crash has happened. Several drivers will report UINT64_MAX when "
"requesting a counter value of an timeline semaphore right after/during a driver reset. "
"If this happen we should investigate what makes the driver crash. In the past this has "
"been detected on QUALCOMM and NVIDIA drivers. The result code of the call is "
"VK_SUCCESS.");
return current_timeline;
}
/** \} */
/* -------------------------------------------------------------------- */
/** \name Resource management
* \{ */
/**
* Get or create current thread data.
*/
VKThreadData &current_thread_data();
#if 0
/**
* Get the discard pool for the current thread.
*
* When the active thread has a context a discard pool associated to the thread is returned.
* When there is no context the orphan discard pool is returned.
*
* A thread with a context can have multiple discard pools. One for each swap-chain image.
* A thread without a context is most likely a discarded resource triggered during dependency
* graph update. A dependency graph update from the viewport during playback or editing;
* or a dependency graph update when rendering.
* These can happen from a different thread which will don't have a context at all.
* \param thread_safe: Caller thread already owns the resources mutex and is safe to run this
* function without trying to reacquire resources mutex making a deadlock.
*/
VKDiscardPool &discard_pool_for_current_thread(bool thread_safe = false);
#endif
void context_register(VKContext &context);
void context_unregister(VKContext &context);
Span<std::reference_wrapper<VKContext>> contexts_get() const;
void memory_statistics_get(int *r_total_mem_kb, int *r_free_mem_kb) const;
static void debug_print(std::ostream &os, const VKDiscardPool &discard_pool);
void debug_print();
/** \} */
Shader *vk_backbuffer_blit_sh_get()
{
if (vk_backbuffer_blit_sh_ == nullptr) {
/* See display_as_extended_srgb in libocio_display_processor.cc for details on this choice. */
#if defined(_WIN32) || defined(__APPLE__)
vk_backbuffer_blit_sh_ = GPU_shader_create_from_info_name("vk_backbuffer_blit");
#else
vk_backbuffer_blit_sh_ = GPU_shader_create_from_info_name("vk_backbuffer_blit_gamma22");
#endif
}
return vk_backbuffer_blit_sh_;
}
private:
void init_physical_device_properties();
void init_physical_device_memory_properties();
void init_physical_device_features();
void init_physical_device_extensions();
void init_debug_callbacks();
void init_memory_allocator();
void init_submission_pool();
void deinit_submission_pool();
/**
* Initialize the functions struct with extension specific function pointer.
*/
void init_functions();
/**
* Initialize a dummy buffer that can be bound for missing attributes.
*/
void init_dummy_buffer();
/* During initialization the backend requires access to update the workarounds. */
friend VKBackend;
};
} // namespace blender::gpu