Implement experimental Vulkan backend

This commit is contained in:
Stenzek
2016-10-01 02:40:01 +10:00
parent fdd954e7e7
commit 77a128ab87
59 changed files with 14533 additions and 1 deletions
+1
View File
@@ -233,6 +233,7 @@ set(LIBS
sfml-network
sfml-system
videonull
videovulkan
videoogl
videosoftware
z
@@ -207,6 +207,9 @@
<ProjectReference Include="..\VideoBackends\D3D12\D3D12.vcxproj">
<Project>{570215b7-e32f-4438-95ae-c8d955f9fca3}</Project>
</ProjectReference>
<ProjectReference Include="..\VideoBackends\Vulkan\Vulkan.vcxproj">
<Project>{29f29a19-f141-45ad-9679-5a2923b49da3}</Project>
</ProjectReference>
</ItemGroup>
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
+3
View File
@@ -238,6 +238,9 @@
<ProjectReference Include="$(CoreDir)VideoBackends\Null\Null.vcxproj">
<Project>{53A5391B-737E-49A8-BC8F-312ADA00736F}</Project>
</ProjectReference>
<ProjectReference Include="$(CoreDir)VideoBackends\Vulkan\Vulkan.vcxproj">
<Project>{29F29A19-F141-45AD-9679-5A2923B49DA3}</Project>
</ProjectReference>
<ProjectReference Include="$(CoreDir)VideoCommon\VideoCommon.vcxproj">
<Project>{3de9ee35-3e91-4f27-a014-2866ad8c3fe3}</Project>
</ProjectReference>
+1
View File
@@ -1,4 +1,5 @@
add_subdirectory(OGL)
add_subdirectory(Null)
add_subdirectory(Software)
add_subdirectory(Vulkan)
# TODO: Add other backends here!
@@ -0,0 +1,249 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#include <vector>
#include "Common/Assert.h"
#include "VideoBackends/Vulkan/BoundingBox.h"
#include "VideoBackends/Vulkan/CommandBufferManager.h"
#include "VideoBackends/Vulkan/ObjectCache.h"
#include "VideoBackends/Vulkan/StagingBuffer.h"
#include "VideoBackends/Vulkan/StateTracker.h"
#include "VideoBackends/Vulkan/Util.h"
#include "VideoBackends/Vulkan/VulkanContext.h"
namespace Vulkan
{
BoundingBox::BoundingBox()
{
}
BoundingBox::~BoundingBox()
{
if (m_gpu_buffer != VK_NULL_HANDLE)
{
vkDestroyBuffer(g_vulkan_context->GetDevice(), m_gpu_buffer, nullptr);
vkFreeMemory(g_vulkan_context->GetDevice(), m_gpu_memory, nullptr);
}
}
bool BoundingBox::Initialize()
{
if (!g_vulkan_context->SupportsBoundingBox())
{
WARN_LOG(VIDEO, "Vulkan: Bounding box is unsupported by your device.");
return true;
}
if (!CreateGPUBuffer())
return false;
if (!CreateReadbackBuffer())
return false;
return true;
}
void BoundingBox::Flush(StateTracker* state_tracker)
{
if (m_gpu_buffer == VK_NULL_HANDLE)
return;
// Combine updates together, chances are the game would have written all 4.
bool updated_buffer = false;
for (size_t start = 0; start < 4; start++)
{
if (!m_values_dirty[start])
continue;
size_t count = 0;
std::array<s32, 4> write_values;
for (; (start + count) < 4; count++)
{
if (!m_values_dirty[start + count])
break;
m_readback_buffer->Read((start + count) * sizeof(s32), &write_values[count], sizeof(s32),
false);
m_values_dirty[start + count] = false;
}
// We can't issue vkCmdUpdateBuffer within a render pass.
// However, the writes must be serialized, so we can't put it in the init buffer.
if (!updated_buffer)
{
state_tracker->EndRenderPass();
// Ensure GPU buffer is in a state where it can be transferred to.
Util::BufferMemoryBarrier(
g_command_buffer_mgr->GetCurrentCommandBuffer(), m_gpu_buffer,
VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT, VK_ACCESS_TRANSFER_WRITE_BIT, 0,
BUFFER_SIZE, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT);
updated_buffer = true;
}
vkCmdUpdateBuffer(g_command_buffer_mgr->GetCurrentCommandBuffer(), m_gpu_buffer,
start * sizeof(s32), count * sizeof(s32),
reinterpret_cast<const u32*>(write_values.data()));
}
// Restore fragment shader access to the buffer.
if (updated_buffer)
{
Util::BufferMemoryBarrier(
g_command_buffer_mgr->GetCurrentCommandBuffer(), m_gpu_buffer, VK_ACCESS_TRANSFER_WRITE_BIT,
VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT, 0, BUFFER_SIZE,
VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
}
// We're now up-to-date.
m_valid = true;
}
void BoundingBox::Invalidate(StateTracker* state_tracker)
{
if (m_gpu_buffer == VK_NULL_HANDLE)
return;
m_valid = false;
}
s32 BoundingBox::Get(StateTracker* state_tracker, size_t index)
{
_assert_(index < NUM_VALUES);
if (!m_valid)
Readback(state_tracker);
s32 value;
m_readback_buffer->Read(index * sizeof(s32), &value, sizeof(value), false);
return value;
}
void BoundingBox::Set(StateTracker* state_tracker, size_t index, s32 value)
{
_assert_(index < NUM_VALUES);
// If we're currently valid, update the stored value in both our cache and the GPU buffer.
if (m_valid)
{
// Skip when it hasn't changed.
s32 current_value;
m_readback_buffer->Read(index * sizeof(s32), &current_value, sizeof(current_value), false);
if (current_value == value)
return;
}
// Flag as dirty, and update values.
m_readback_buffer->Write(index * sizeof(s32), &value, sizeof(value), true);
m_values_dirty[index] = true;
}
bool BoundingBox::CreateGPUBuffer()
{
VkBufferUsageFlags buffer_usage = VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
VK_BUFFER_USAGE_TRANSFER_SRC_BIT |
VK_BUFFER_USAGE_TRANSFER_DST_BIT;
VkBufferCreateInfo info = {
VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO, // VkStructureType sType
nullptr, // const void* pNext
0, // VkBufferCreateFlags flags
BUFFER_SIZE, // VkDeviceSize size
buffer_usage, // VkBufferUsageFlags usage
VK_SHARING_MODE_EXCLUSIVE, // VkSharingMode sharingMode
0, // uint32_t queueFamilyIndexCount
nullptr // const uint32_t* pQueueFamilyIndices
};
VkBuffer buffer;
VkResult res = vkCreateBuffer(g_vulkan_context->GetDevice(), &info, nullptr, &buffer);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkCreateBuffer failed: ");
return false;
}
VkMemoryRequirements memory_requirements;
vkGetBufferMemoryRequirements(g_vulkan_context->GetDevice(), buffer, &memory_requirements);
uint32_t memory_type_index = g_vulkan_context->GetMemoryType(memory_requirements.memoryTypeBits,
VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
VkMemoryAllocateInfo memory_allocate_info = {
VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO, // VkStructureType sType
nullptr, // const void* pNext
memory_requirements.size, // VkDeviceSize allocationSize
memory_type_index // uint32_t memoryTypeIndex
};
VkDeviceMemory memory;
res = vkAllocateMemory(g_vulkan_context->GetDevice(), &memory_allocate_info, nullptr, &memory);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkAllocateMemory failed: ");
vkDestroyBuffer(g_vulkan_context->GetDevice(), buffer, nullptr);
return false;
}
res = vkBindBufferMemory(g_vulkan_context->GetDevice(), buffer, memory, 0);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkBindBufferMemory failed: ");
vkDestroyBuffer(g_vulkan_context->GetDevice(), buffer, nullptr);
vkFreeMemory(g_vulkan_context->GetDevice(), memory, nullptr);
return false;
}
m_gpu_buffer = buffer;
m_gpu_memory = memory;
return true;
}
bool BoundingBox::CreateReadbackBuffer()
{
m_readback_buffer = StagingBuffer::Create(STAGING_BUFFER_TYPE_READBACK, BUFFER_SIZE,
VK_BUFFER_USAGE_TRANSFER_DST_BIT);
if (!m_readback_buffer || !m_readback_buffer->Map())
return false;
return true;
}
void BoundingBox::Readback(StateTracker* state_tracker)
{
// Can't be done within a render pass.
state_tracker->EndRenderPass();
// Ensure all writes are completed to the GPU buffer prior to the transfer.
Util::BufferMemoryBarrier(
g_command_buffer_mgr->GetCurrentCommandBuffer(), m_gpu_buffer,
VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT, VK_ACCESS_TRANSFER_READ_BIT, 0,
BUFFER_SIZE, VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT);
m_readback_buffer->PrepareForGPUWrite(g_command_buffer_mgr->GetCurrentCommandBuffer(),
VK_ACCESS_TRANSFER_WRITE_BIT,
VK_PIPELINE_STAGE_TRANSFER_BIT);
// Copy from GPU -> readback buffer.
VkBufferCopy region = {0, 0, BUFFER_SIZE};
vkCmdCopyBuffer(g_command_buffer_mgr->GetCurrentCommandBuffer(), m_gpu_buffer,
m_readback_buffer->GetBuffer(), 1, &region);
// Restore GPU buffer access.
Util::BufferMemoryBarrier(g_command_buffer_mgr->GetCurrentCommandBuffer(), m_gpu_buffer,
VK_ACCESS_TRANSFER_READ_BIT,
VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT, 0, BUFFER_SIZE,
VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT);
m_readback_buffer->FlushGPUCache(g_command_buffer_mgr->GetCurrentCommandBuffer(),
VK_ACCESS_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT);
// Wait until these commands complete.
Util::ExecuteCurrentCommandsAndRestoreState(state_tracker, false, true);
// Cache is now valid.
m_readback_buffer->InvalidateCPUCache();
m_valid = true;
}
} // namespace Vulkan
@@ -0,0 +1,52 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#pragma once
#include <memory>
#include <string>
#include "Common/CommonTypes.h"
#include "VideoBackends/Vulkan/VulkanLoader.h"
namespace Vulkan
{
class StagingBuffer;
class StateTracker;
class BoundingBox
{
public:
BoundingBox();
~BoundingBox();
bool Initialize();
VkBuffer GetGPUBuffer() const { return m_gpu_buffer; }
VkDeviceSize GetGPUBufferOffset() const { return 0; }
VkDeviceSize GetGPUBufferSize() const { return BUFFER_SIZE; }
s32 Get(StateTracker* state_tracker, size_t index);
void Set(StateTracker* state_tracker, size_t index, s32 value);
void Invalidate(StateTracker* state_tracker);
void Flush(StateTracker* state_tracker);
private:
bool CreateGPUBuffer();
bool CreateReadbackBuffer();
void Readback(StateTracker* state_tracker);
VkBuffer m_gpu_buffer = VK_NULL_HANDLE;
VkDeviceMemory m_gpu_memory = nullptr;
static const size_t NUM_VALUES = 4;
static const size_t BUFFER_SIZE = sizeof(u32) * NUM_VALUES;
std::unique_ptr<StagingBuffer> m_readback_buffer;
std::array<bool, NUM_VALUES> m_values_dirty = {};
bool m_valid = true;
};
} // namespace Vulkan
@@ -0,0 +1,42 @@
set(SRCS
BoundingBox.cpp
CommandBufferManager.cpp
FramebufferManager.cpp
ObjectCache.cpp
PaletteTextureConverter.cpp
PerfQuery.cpp
RasterFont.cpp
Renderer.cpp
ShaderCompiler.cpp
StateTracker.cpp
StagingBuffer.cpp
StagingTexture2D.cpp
StreamBuffer.cpp
SwapChain.cpp
Texture2D.cpp
TextureCache.cpp
TextureEncoder.cpp
Util.cpp
VertexFormat.cpp
VertexManager.cpp
VulkanContext.cpp
VulkanLoader.cpp
main.cpp
)
set(LIBS
videocommon
common
)
# Only include the Vulkan headers when building the Vulkan backend
include_directories(${CMAKE_SOURCE_DIR}/Externals/Vulkan/Include)
# Silence warnings on glslang by flagging it as a system include
include_directories(SYSTEM ${CMAKE_SOURCE_DIR}/Externals/glslang/glslang/Public)
include_directories(SYSTEM ${CMAKE_SOURCE_DIR}/Externals/glslang/SPIRV)
# Link against glslang, the other necessary libraries are referenced by the executable.
add_dolphin_library(videovulkan "${SRCS}" "${LIBS}")
target_link_libraries(videovulkan glslang)
@@ -0,0 +1,457 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#include <algorithm>
#include <cstdint>
#include "Common/Assert.h"
#include "Common/CommonFuncs.h"
#include "Common/MsgHandler.h"
#include "VideoBackends/Vulkan/CommandBufferManager.h"
#include "VideoBackends/Vulkan/VulkanContext.h"
namespace Vulkan
{
CommandBufferManager::CommandBufferManager(bool use_threaded_submission)
: m_submit_semaphore(1, 1), m_use_threaded_submission(use_threaded_submission)
{
}
CommandBufferManager::~CommandBufferManager()
{
// If the worker thread is enabled, wait for it to exit.
if (m_use_threaded_submission)
{
// Wait for all command buffers to be consumed by the worker thread.
m_submit_semaphore.Wait();
m_submit_loop->Stop();
m_submit_thread.join();
}
vkDeviceWaitIdle(g_vulkan_context->GetDevice());
DestroyCommandBuffers();
DestroyCommandPool();
}
bool CommandBufferManager::Initialize()
{
if (!CreateCommandPool())
return false;
if (!CreateCommandBuffers())
return false;
if (m_use_threaded_submission && !CreateSubmitThread())
return false;
return true;
}
bool CommandBufferManager::CreateCommandPool()
{
VkCommandPoolCreateInfo info = {VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO, nullptr,
VK_COMMAND_POOL_CREATE_TRANSIENT_BIT |
VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT,
g_vulkan_context->GetGraphicsQueueFamilyIndex()};
VkResult res =
vkCreateCommandPool(g_vulkan_context->GetDevice(), &info, nullptr, &m_command_pool);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkCreateCommandPool failed: ");
return false;
}
return true;
}
void CommandBufferManager::DestroyCommandPool()
{
if (m_command_pool)
{
vkDestroyCommandPool(g_vulkan_context->GetDevice(), m_command_pool, nullptr);
m_command_pool = nullptr;
}
}
bool CommandBufferManager::CreateCommandBuffers()
{
VkDevice device = g_vulkan_context->GetDevice();
for (FrameResources& resources : m_frame_resources)
{
resources.needs_fence_wait = false;
VkCommandBufferAllocateInfo allocate_info = {
VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO, nullptr, m_command_pool,
VK_COMMAND_BUFFER_LEVEL_PRIMARY, static_cast<uint32_t>(resources.command_buffers.size())};
VkResult res =
vkAllocateCommandBuffers(device, &allocate_info, resources.command_buffers.data());
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkAllocateCommandBuffers failed: ");
return false;
}
VkFenceCreateInfo fence_info = {VK_STRUCTURE_TYPE_FENCE_CREATE_INFO, nullptr,
VK_FENCE_CREATE_SIGNALED_BIT};
res = vkCreateFence(device, &fence_info, nullptr, &resources.fence);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkCreateFence failed: ");
return false;
}
// TODO: A better way to choose the number of descriptors.
VkDescriptorPoolSize pool_sizes[] = {{VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC, 500000},
{VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 500000},
{VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, 16},
{VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER, 1024}};
VkDescriptorPoolCreateInfo pool_create_info = {VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
nullptr,
0,
100000, // tweak this
static_cast<u32>(ArraySize(pool_sizes)),
pool_sizes};
res = vkCreateDescriptorPool(device, &pool_create_info, nullptr, &resources.descriptor_pool);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkCreateDescriptorPool failed: ");
return false;
}
}
// Activate the first command buffer. ActivateCommandBuffer moves forward, so start with the last
m_current_frame = m_frame_resources.size() - 1;
ActivateCommandBuffer();
return true;
}
void CommandBufferManager::DestroyCommandBuffers()
{
VkDevice device = g_vulkan_context->GetDevice();
for (FrameResources& resources : m_frame_resources)
{
for (const auto& it : resources.cleanup_resources)
it.destroy_callback(device, it.object);
resources.cleanup_resources.clear();
if (resources.fence != VK_NULL_HANDLE)
{
vkDestroyFence(device, resources.fence, nullptr);
resources.fence = VK_NULL_HANDLE;
}
if (resources.descriptor_pool != VK_NULL_HANDLE)
{
vkDestroyDescriptorPool(device, resources.descriptor_pool, nullptr);
resources.descriptor_pool = VK_NULL_HANDLE;
}
if (resources.command_buffers[0] != VK_NULL_HANDLE)
{
vkFreeCommandBuffers(device, m_command_pool,
static_cast<u32>(resources.command_buffers.size()),
resources.command_buffers.data());
resources.command_buffers.fill(VK_NULL_HANDLE);
}
}
}
VkDescriptorSet CommandBufferManager::AllocateDescriptorSet(VkDescriptorSetLayout set_layout)
{
VkDescriptorSetAllocateInfo allocate_info = {
VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO, nullptr,
m_frame_resources[m_current_frame].descriptor_pool, 1, &set_layout};
VkDescriptorSet descriptor_set;
VkResult res =
vkAllocateDescriptorSets(g_vulkan_context->GetDevice(), &allocate_info, &descriptor_set);
if (res != VK_SUCCESS)
{
// Failing to allocate a descriptor set is not a fatal error, we can
// recover by moving to the next command buffer.
return VK_NULL_HANDLE;
}
return descriptor_set;
}
bool CommandBufferManager::CreateSubmitThread()
{
m_submit_loop = std::make_unique<Common::BlockingLoop>();
m_submit_thread = std::thread([this]() {
m_submit_loop->Run([this]() {
PendingCommandBufferSubmit submit;
{
std::lock_guard<std::mutex> guard(m_pending_submit_lock);
if (m_pending_submits.empty())
{
m_submit_loop->AllowSleep();
return;
}
submit = m_pending_submits.front();
m_pending_submits.pop_front();
}
SubmitCommandBuffer(submit.index, submit.wait_semaphore, submit.signal_semaphore,
submit.present_swap_chain, submit.present_image_index);
});
});
return true;
}
void CommandBufferManager::PrepareToSubmitCommandBuffer()
{
// Grab the semaphore before submitting command buffer either on-thread or off-thread.
// This prevents a race from occurring where a second command buffer is executed
// before the worker thread has woken and executed the first one yet.
m_submit_semaphore.Wait();
}
void CommandBufferManager::WaitForWorkerThreadIdle()
{
// Drain the semaphore, then allow another request in the future.
m_submit_semaphore.Wait();
m_submit_semaphore.Post();
}
void CommandBufferManager::WaitForGPUIdle()
{
WaitForWorkerThreadIdle();
vkDeviceWaitIdle(g_vulkan_context->GetDevice());
}
void CommandBufferManager::WaitForFence(VkFence fence)
{
// Find the command buffer that this fence corresponds to.
size_t command_buffer_index = 0;
for (; command_buffer_index < m_frame_resources.size(); command_buffer_index++)
{
if (m_frame_resources[command_buffer_index].fence == fence)
break;
}
_assert_(command_buffer_index < m_frame_resources.size());
// Has this command buffer already been waited for?
if (!m_frame_resources[command_buffer_index].needs_fence_wait)
return;
// Wait for this command buffer to be completed.
VkResult res =
vkWaitForFences(g_vulkan_context->GetDevice(), 1,
&m_frame_resources[command_buffer_index].fence, VK_TRUE, UINT64_MAX);
if (res != VK_SUCCESS)
LOG_VULKAN_ERROR(res, "vkWaitForFences failed: ");
// Immediately fire callbacks and cleanups, since the commands has been completed.
m_frame_resources[command_buffer_index].needs_fence_wait = false;
OnCommandBufferExecuted(command_buffer_index);
}
void CommandBufferManager::SubmitCommandBuffer(bool submit_on_worker_thread,
VkSemaphore wait_semaphore,
VkSemaphore signal_semaphore,
VkSwapchainKHR present_swap_chain,
uint32_t present_image_index)
{
FrameResources& resources = m_frame_resources[m_current_frame];
// Fire fence tracking callbacks. This can't happen on the worker thread.
// We invoke these before submitting so that any last-minute commands can be added.
for (const auto& iter : m_fence_point_callbacks)
iter.second.first(resources.command_buffers[1], resources.fence);
// End the current command buffer.
for (VkCommandBuffer command_buffer : resources.command_buffers)
{
VkResult res = vkEndCommandBuffer(command_buffer);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkEndCommandBuffer failed: ");
PanicAlert("Failed to end command buffer");
}
}
// This command buffer now has commands, so can't be re-used without waiting.
resources.needs_fence_wait = true;
// Submitting off-thread?
if (m_use_threaded_submission && submit_on_worker_thread)
{
// Push to the pending submit queue.
{
std::lock_guard<std::mutex> guard(m_pending_submit_lock);
m_pending_submits.push_back({m_current_frame, wait_semaphore, signal_semaphore,
present_swap_chain, present_image_index});
}
// Wake up the worker thread for a single iteration.
m_submit_loop->Wakeup();
}
else
{
// Pass through to normal submission path.
SubmitCommandBuffer(m_current_frame, wait_semaphore, signal_semaphore, present_swap_chain,
present_image_index);
}
}
void CommandBufferManager::SubmitCommandBuffer(size_t index, VkSemaphore wait_semaphore,
VkSemaphore signal_semaphore,
VkSwapchainKHR present_swap_chain,
uint32_t present_image_index)
{
FrameResources& resources = m_frame_resources[index];
// This may be executed on the worker thread, so don't modify any state of the manager class.
uint32_t wait_bits = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
VkSubmitInfo submit_info = {VK_STRUCTURE_TYPE_SUBMIT_INFO,
nullptr,
0,
nullptr,
&wait_bits,
static_cast<u32>(resources.command_buffers.size()),
resources.command_buffers.data(),
0,
nullptr};
if (wait_semaphore != VK_NULL_HANDLE)
{
submit_info.pWaitSemaphores = &wait_semaphore;
submit_info.waitSemaphoreCount = 1;
}
if (signal_semaphore != VK_NULL_HANDLE)
{
submit_info.signalSemaphoreCount = 1;
submit_info.pSignalSemaphores = &signal_semaphore;
}
VkResult res =
vkQueueSubmit(g_vulkan_context->GetGraphicsQueue(), 1, &submit_info, resources.fence);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkQueueSubmit failed: ");
PanicAlert("Failed to submit command buffer.");
}
// Do we have a swap chain to present?
if (present_swap_chain != VK_NULL_HANDLE)
{
// Should have a signal semaphore.
_assert_(signal_semaphore != VK_NULL_HANDLE);
VkPresentInfoKHR present_info = {VK_STRUCTURE_TYPE_PRESENT_INFO_KHR,
nullptr,
1,
&signal_semaphore,
1,
&present_swap_chain,
&present_image_index,
nullptr};
res = vkQueuePresentKHR(g_vulkan_context->GetGraphicsQueue(), &present_info);
if (res != VK_SUCCESS && res != VK_ERROR_OUT_OF_DATE_KHR && res != VK_SUBOPTIMAL_KHR)
LOG_VULKAN_ERROR(res, "vkQueuePresentKHR failed: ");
}
// Command buffer has been queued, so permit the next one.
m_submit_semaphore.Post();
}
void CommandBufferManager::OnCommandBufferExecuted(size_t index)
{
FrameResources& resources = m_frame_resources[index];
// Fire fence tracking callbacks.
for (const auto& iter : m_fence_point_callbacks)
iter.second.second(resources.fence);
// Clean up all objects pending destruction on this command buffer
for (const auto& it : resources.cleanup_resources)
it.destroy_callback(g_vulkan_context->GetDevice(), it.object);
resources.cleanup_resources.clear();
}
void CommandBufferManager::ActivateCommandBuffer()
{
// Move to the next command buffer.
m_current_frame = (m_current_frame + 1) % NUM_COMMAND_BUFFERS;
FrameResources& resources = m_frame_resources[m_current_frame];
// Wait for the GPU to finish with all resources for this command buffer.
if (resources.needs_fence_wait)
{
VkResult res =
vkWaitForFences(g_vulkan_context->GetDevice(), 1, &resources.fence, true, UINT64_MAX);
if (res != VK_SUCCESS)
LOG_VULKAN_ERROR(res, "vkWaitForFences failed: ");
OnCommandBufferExecuted(m_current_frame);
}
// Reset fence to unsignaled before starting.
VkResult res = vkResetFences(g_vulkan_context->GetDevice(), 1, &resources.fence);
if (res != VK_SUCCESS)
LOG_VULKAN_ERROR(res, "vkResetFences failed: ");
// Reset command buffer to beginning since we can re-use the memory now
VkCommandBufferBeginInfo begin_info = {VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO, nullptr,
VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT, nullptr};
for (VkCommandBuffer command_buffer : resources.command_buffers)
{
res = vkResetCommandBuffer(command_buffer, 0);
if (res != VK_SUCCESS)
LOG_VULKAN_ERROR(res, "vkResetCommandBuffer failed: ");
res = vkBeginCommandBuffer(command_buffer, &begin_info);
if (res != VK_SUCCESS)
LOG_VULKAN_ERROR(res, "vkBeginCommandBuffer failed: ");
}
// Also can do the same for the descriptor pools
res = vkResetDescriptorPool(g_vulkan_context->GetDevice(), resources.descriptor_pool, 0);
if (res != VK_SUCCESS)
LOG_VULKAN_ERROR(res, "vkResetDescriptorPool failed: ");
}
void CommandBufferManager::ExecuteCommandBuffer(bool submit_off_thread, bool wait_for_completion)
{
VkFence pending_fence = GetCurrentCommandBufferFence();
// If we're waiting for completion, don't bother waking the worker thread.
PrepareToSubmitCommandBuffer();
SubmitCommandBuffer((submit_off_thread && wait_for_completion));
ActivateCommandBuffer();
if (wait_for_completion)
WaitForFence(pending_fence);
}
void CommandBufferManager::AddFencePointCallback(
const void* key, const CommandBufferQueuedCallback& queued_callback,
const CommandBufferExecutedCallback& executed_callback)
{
// Shouldn't be adding twice.
_assert_(m_fence_point_callbacks.find(key) == m_fence_point_callbacks.end());
m_fence_point_callbacks.emplace(key, std::make_pair(queued_callback, executed_callback));
}
void CommandBufferManager::RemoveFencePointCallback(const void* key)
{
auto iter = m_fence_point_callbacks.find(key);
_assert_(iter != m_fence_point_callbacks.end());
m_fence_point_callbacks.erase(iter);
}
std::unique_ptr<CommandBufferManager> g_command_buffer_mgr;
}
@@ -0,0 +1,154 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#pragma once
#include <array>
#include <deque>
#include <functional>
#include <map>
#include <memory>
#include <mutex>
#include <thread>
#include <utility>
#include <vector>
#include "Common/BlockingLoop.h"
#include "Common/Semaphore.h"
#include "VideoCommon/VideoCommon.h"
#include "VideoBackends/Vulkan/Constants.h"
#include "VideoBackends/Vulkan/Util.h"
namespace Vulkan
{
class CommandBufferManager
{
public:
explicit CommandBufferManager(bool use_threaded_submission);
~CommandBufferManager();
bool Initialize();
VkCommandPool GetCommandPool() const { return m_command_pool; }
// These command buffers are allocated per-frame. They are valid until the command buffer
// is submitted, after that you should call these functions again.
VkCommandBuffer GetCurrentInitCommandBuffer() const
{
return m_frame_resources[m_current_frame].command_buffers[0];
}
VkCommandBuffer GetCurrentCommandBuffer() const
{
return m_frame_resources[m_current_frame].command_buffers[1];
}
VkDescriptorPool GetCurrentDescriptorPool() const
{
return m_frame_resources[m_current_frame].descriptor_pool;
}
// Allocates a descriptors set from the pool reserved for the current frame.
VkDescriptorSet AllocateDescriptorSet(VkDescriptorSetLayout set_layout);
// Gets the fence that will be signaled when the currently executing command buffer is
// queued and executed. Do not wait for this fence before the buffer is executed.
VkFence GetCurrentCommandBufferFence() const { return m_frame_resources[m_current_frame].fence; }
// Ensure the worker thread has submitted the previous frame's command buffer.
void PrepareToSubmitCommandBuffer();
// Ensure that the worker thread has submitted any previous command buffers and is idle.
void WaitForWorkerThreadIdle();
// Ensure that the worker thread has both submitted all commands, and the GPU has caught up.
// Use with caution, huge performance penalty.
void WaitForGPUIdle();
// Wait for a fence to be completed.
// Also invokes callbacks for completion.
void WaitForFence(VkFence fence);
void SubmitCommandBuffer(bool submit_on_worker_thread,
VkSemaphore wait_semaphore = VK_NULL_HANDLE,
VkSemaphore signal_semaphore = VK_NULL_HANDLE,
VkSwapchainKHR present_swap_chain = VK_NULL_HANDLE,
uint32_t present_image_index = 0xFFFFFFFF);
void ActivateCommandBuffer();
void ExecuteCommandBuffer(bool submit_off_thread, bool wait_for_completion);
// Schedule a vulkan resource for destruction later on. This will occur when the command buffer
// is next re-used, and the GPU has finished working with the specified resource.
template <typename T>
void DeferResourceDestruction(T object)
{
DeferredResourceDestruction wrapper = DeferredResourceDestruction::Wrapper<T>(object);
m_frame_resources[m_current_frame].cleanup_resources.push_back(wrapper);
}
// Instruct the manager to fire the specified callback when a fence is flagged to be signaled.
// This happens when command buffers are executed, and can be tested if signaled, which means
// that all commands up to the point when the callback was fired have completed.
using CommandBufferQueuedCallback = std::function<void(VkCommandBuffer, VkFence)>;
using CommandBufferExecutedCallback = std::function<void(VkFence)>;
void AddFencePointCallback(const void* key, const CommandBufferQueuedCallback& queued_callback,
const CommandBufferExecutedCallback& executed_callback);
void RemoveFencePointCallback(const void* key);
private:
bool CreateCommandPool();
void DestroyCommandPool();
bool CreateCommandBuffers();
void DestroyCommandBuffers();
bool CreateSubmitThread();
void SubmitCommandBuffer(size_t index, VkSemaphore wait_semaphore, VkSemaphore signal_semaphore,
VkSwapchainKHR present_swap_chain, uint32_t present_image_index);
void OnCommandBufferExecuted(size_t index);
VkCommandPool m_command_pool = VK_NULL_HANDLE;
struct FrameResources
{
// [0] - Init (upload) command buffer, [1] - draw command buffer
std::array<VkCommandBuffer, 2> command_buffers;
VkDescriptorPool descriptor_pool;
VkFence fence;
bool needs_fence_wait;
std::vector<DeferredResourceDestruction> cleanup_resources;
};
std::array<FrameResources, NUM_COMMAND_BUFFERS> m_frame_resources = {};
size_t m_current_frame;
// callbacks when a fence point is set
std::map<const void*, std::pair<CommandBufferQueuedCallback, CommandBufferExecutedCallback>>
m_fence_point_callbacks;
// Threaded command buffer execution
// Semaphore determines when a command buffer can be queued
Common::Semaphore m_submit_semaphore;
std::thread m_submit_thread;
std::unique_ptr<Common::BlockingLoop> m_submit_loop;
struct PendingCommandBufferSubmit
{
size_t index;
VkSemaphore wait_semaphore;
VkSemaphore signal_semaphore;
VkSwapchainKHR present_swap_chain;
uint32_t present_image_index;
};
std::deque<PendingCommandBufferSubmit> m_pending_submits;
std::mutex m_pending_submit_lock;
bool m_use_threaded_submission = false;
};
extern std::unique_ptr<CommandBufferManager> g_command_buffer_mgr;
} // namespace Vulkan
@@ -0,0 +1,136 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#pragma once
#include "Common/BitField.h"
#include "VideoBackends/Vulkan/VulkanLoader.h"
namespace Vulkan
{
// Number of command buffers. Having two allows one buffer to be
// executed whilst another is being built.
constexpr size_t NUM_COMMAND_BUFFERS = 2;
// Staging buffer usage - optimize for uploads or readbacks
enum STAGING_BUFFER_TYPE
{
STAGING_BUFFER_TYPE_UPLOAD,
STAGING_BUFFER_TYPE_READBACK
};
// Descriptor sets
enum DESCRIPTOR_SET
{
DESCRIPTOR_SET_UNIFORM_BUFFERS,
DESCRIPTOR_SET_PIXEL_SHADER_SAMPLERS,
DESCRIPTOR_SET_SHADER_STORAGE_BUFFERS,
NUM_DESCRIPTOR_SETS
};
// Uniform buffer bindings within the first descriptor set
enum UNIFORM_BUFFER_DESCRIPTOR_SET_BINDING
{
UBO_DESCRIPTOR_SET_BINDING_PS,
UBO_DESCRIPTOR_SET_BINDING_VS,
UBO_DESCRIPTOR_SET_BINDING_GS,
NUM_UBO_DESCRIPTOR_SET_BINDINGS
};
// Maximum number of attributes per vertex (we don't have any more than this?)
constexpr size_t MAX_VERTEX_ATTRIBUTES = 16;
// Number of pixel shader texture slots
constexpr size_t NUM_PIXEL_SHADER_SAMPLERS = 8;
// Total number of binding points in the pipeline layout
constexpr size_t TOTAL_PIPELINE_BINDING_POINTS =
NUM_UBO_DESCRIPTOR_SET_BINDINGS + NUM_PIXEL_SHADER_SAMPLERS + 1;
// Format of EFB textures
constexpr VkFormat EFB_COLOR_TEXTURE_FORMAT = VK_FORMAT_R8G8B8A8_UNORM;
constexpr VkFormat EFB_DEPTH_TEXTURE_FORMAT = VK_FORMAT_D32_SFLOAT;
constexpr VkFormat EFB_DEPTH_AS_COLOR_TEXTURE_FORMAT = VK_FORMAT_R32_SFLOAT;
// Format of texturecache textures
constexpr VkFormat TEXTURECACHE_TEXTURE_FORMAT = VK_FORMAT_R8G8B8A8_UNORM;
// Textures that don't fit into this buffer will be uploaded with a separate buffer (see below).
constexpr size_t INITIAL_TEXTURE_UPLOAD_BUFFER_SIZE = 16 * 1024 * 1024;
constexpr size_t MAXIMUM_TEXTURE_UPLOAD_BUFFER_SIZE = 64 * 1024 * 1024;
// Textures greater than 1024*1024 will be put in staging textures that are released after
// execution instead. A 2048x2048 texture is 16MB, and we'd only fit four of these in our
// streaming buffer and be blocking frequently. Games are unlikely to have textures this
// large anyway, so it's only really an issue for HD texture packs, and memory is not
// a limiting factor in these scenarios anyway.
constexpr size_t STAGING_TEXTURE_UPLOAD_THRESHOLD = 1024 * 1024 * 4;
// Streaming uniform buffer size
constexpr size_t INITIAL_UNIFORM_STREAM_BUFFER_SIZE = 16 * 1024 * 1024;
constexpr size_t MAXIMUM_UNIFORM_STREAM_BUFFER_SIZE = 32 * 1024 * 1024;
// Push constant buffer size for utility shaders
constexpr u32 PUSH_CONSTANT_BUFFER_SIZE = 128;
// Rasterization state info
union RasterizationState {
BitField<0, 2, VkCullModeFlags> cull_mode;
BitField<2, 7, VkSampleCountFlagBits> samples;
BitField<9, 1, VkBool32> per_sample_shading;
BitField<10, 1, VkBool32> depth_clamp;
u32 bits;
};
// Depth state info
union DepthStencilState {
BitField<0, 1, VkBool32> test_enable;
BitField<1, 1, VkBool32> write_enable;
BitField<2, 3, VkCompareOp> compare_op;
u32 bits;
};
// Blend state info
union BlendState {
struct
{
union {
BitField<0, 1, VkBool32> blend_enable;
BitField<1, 3, VkBlendOp> blend_op;
BitField<4, 5, VkBlendFactor> src_blend;
BitField<9, 5, VkBlendFactor> dst_blend;
BitField<14, 3, VkBlendOp> alpha_blend_op;
BitField<17, 5, VkBlendFactor> src_alpha_blend;
BitField<22, 5, VkBlendFactor> dst_alpha_blend;
BitField<27, 4, VkColorComponentFlags> write_mask;
u32 low_bits;
};
union {
BitField<0, 1, VkBool32> logic_op_enable;
BitField<1, 4, VkLogicOp> logic_op;
u32 high_bits;
};
};
u64 bits;
};
// Sampler info
union SamplerState {
BitField<0, 1, VkFilter> min_filter;
BitField<1, 1, VkFilter> mag_filter;
BitField<2, 1, VkSamplerMipmapMode> mipmap_mode;
BitField<3, 2, VkSamplerAddressMode> wrap_u;
BitField<5, 2, VkSamplerAddressMode> wrap_v;
BitField<7, 8, u32> min_lod;
BitField<15, 8, u32> max_lod;
BitField<23, 6, s32> lod_bias; // tm0.lod_bias (8 bits) / 32 gives us 0-7.
BitField<29, 3, u32> anisotropy; // max_anisotropy = 1 << anisotropy, max of 16, so range 0-4.
u32 bits;
};
} // namespace Vulkan
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,170 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#include <memory>
#include "VideoCommon/FramebufferManagerBase.h"
#include "VideoBackends/Vulkan/Constants.h"
namespace Vulkan
{
class StagingTexture2D;
class StateTracker;
class StreamBuffer;
class Texture2D;
class VertexFormat;
class XFBSource : public XFBSourceBase
{
void DecodeToTexture(u32 xfb_addr, u32 fb_width, u32 fb_height) override {}
void CopyEFB(float gamma) override {}
};
class FramebufferManager : public FramebufferManagerBase
{
public:
FramebufferManager();
~FramebufferManager();
bool Initialize();
VkRenderPass GetEFBRenderPass() const { return m_efb_render_pass; }
u32 GetEFBWidth() const { return m_efb_width; }
u32 GetEFBHeight() const { return m_efb_height; }
u32 GetEFBLayers() const { return m_efb_layers; }
VkSampleCountFlagBits GetEFBSamples() const { return m_efb_samples; }
Texture2D* GetEFBColorTexture() const { return m_efb_color_texture.get(); }
Texture2D* GetEFBDepthTexture() const { return m_efb_depth_texture.get(); }
VkFramebuffer GetEFBFramebuffer() const { return m_efb_framebuffer; }
void GetTargetSize(unsigned int* width, unsigned int* height) override;
std::unique_ptr<XFBSourceBase> CreateXFBSource(unsigned int target_width,
unsigned int target_height,
unsigned int layers) override
{
return std::make_unique<XFBSource>();
}
void CopyToRealXFB(u32 xfb_addr, u32 fb_stride, u32 fb_height, const EFBRectangle& source_rc,
float gamma = 1.0f) override
{
}
void ResizeEFBTextures();
// Recompile shaders, use when MSAA mode changes.
void RecreateRenderPass();
void RecompileShaders();
// Reinterpret pixel format of EFB color texture.
// Assumes no render pass is currently in progress.
// Swaps EFB framebuffers, so re-bind afterwards.
void ReinterpretPixelData(int convtype);
// Resolve color/depth textures to a non-msaa texture, and return it.
Texture2D* ResolveEFBColorTexture(StateTracker* state_tracker, const VkRect2D& region);
Texture2D* ResolveEFBDepthTexture(StateTracker* state_tracker, const VkRect2D& region);
// Reads a framebuffer value back from the GPU. This may block if the cache is not current.
u32 PeekEFBColor(StateTracker* state_tracker, u32 x, u32 y);
float PeekEFBDepth(StateTracker* state_tracker, u32 x, u32 y);
void InvalidatePeekCache();
// Writes a value to the framebuffer. This will never block, and writes will be batched.
void PokeEFBColor(StateTracker* state_tracker, u32 x, u32 y, u32 color);
void PokeEFBDepth(StateTracker* state_tracker, u32 x, u32 y, float depth);
void FlushEFBPokes(StateTracker* state_tracker);
private:
struct EFBPokeVertex
{
float position[4];
u32 color;
};
bool CreateEFBRenderPass();
void DestroyEFBRenderPass();
bool CreateEFBFramebuffer();
void DestroyEFBFramebuffer();
bool CompileConversionShaders();
void DestroyConversionShaders();
bool CreateReadbackRenderPasses();
void DestroyReadbackRenderPasses();
bool CompileReadbackShaders();
void DestroyReadbackShaders();
bool CreateReadbackTextures();
void DestroyReadbackTextures();
bool CreateReadbackFramebuffer();
void DestroyReadbackFramebuffer();
void CreatePokeVertexFormat();
bool CreatePokeVertexBuffer();
void DestroyPokeVertexBuffer();
bool CompilePokeShaders();
void DestroyPokeShaders();
bool PopulateColorReadbackTexture(StateTracker* state_tracker);
bool PopulateDepthReadbackTexture(StateTracker* state_tracker);
void CreatePokeVertices(std::vector<EFBPokeVertex>* destination_list, u32 x, u32 y, float z,
u32 color);
void DrawPokeVertices(StateTracker* state_tracker, const EFBPokeVertex* vertices,
size_t vertex_count, bool write_color, bool write_depth);
VkRenderPass m_efb_render_pass = VK_NULL_HANDLE;
VkRenderPass m_depth_resolve_render_pass = VK_NULL_HANDLE;
u32 m_efb_width = 0;
u32 m_efb_height = 0;
u32 m_efb_layers = 1;
VkSampleCountFlagBits m_efb_samples = VK_SAMPLE_COUNT_1_BIT;
std::unique_ptr<Texture2D> m_efb_color_texture;
std::unique_ptr<Texture2D> m_efb_convert_color_texture;
std::unique_ptr<Texture2D> m_efb_depth_texture;
std::unique_ptr<Texture2D> m_efb_resolve_color_texture;
std::unique_ptr<Texture2D> m_efb_resolve_depth_texture;
VkFramebuffer m_efb_framebuffer = VK_NULL_HANDLE;
VkFramebuffer m_efb_convert_framebuffer = VK_NULL_HANDLE;
VkFramebuffer m_depth_resolve_framebuffer = VK_NULL_HANDLE;
// Format conversion shaders
VkShaderModule m_ps_rgb8_to_rgba6 = VK_NULL_HANDLE;
VkShaderModule m_ps_rgba6_to_rgb8 = VK_NULL_HANDLE;
VkShaderModule m_ps_depth_resolve = VK_NULL_HANDLE;
// EFB readback texture
std::unique_ptr<Texture2D> m_color_copy_texture;
std::unique_ptr<Texture2D> m_depth_copy_texture;
VkFramebuffer m_color_copy_framebuffer = VK_NULL_HANDLE;
VkFramebuffer m_depth_copy_framebuffer = VK_NULL_HANDLE;
// CPU-side EFB readback texture
std::unique_ptr<StagingTexture2D> m_color_readback_texture;
std::unique_ptr<StagingTexture2D> m_depth_readback_texture;
bool m_color_readback_texture_valid = false;
bool m_depth_readback_texture_valid = false;
// EFB poke drawing setup
std::unique_ptr<VertexFormat> m_poke_vertex_format;
std::unique_ptr<StreamBuffer> m_poke_vertex_stream_buffer;
std::vector<EFBPokeVertex> m_color_poke_vertices;
std::vector<EFBPokeVertex> m_depth_poke_vertices;
VkPrimitiveTopology m_poke_primitive_topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST;
VkRenderPass m_copy_color_render_pass = VK_NULL_HANDLE;
VkRenderPass m_copy_depth_render_pass = VK_NULL_HANDLE;
VkShaderModule m_copy_color_shader = VK_NULL_HANDLE;
VkShaderModule m_copy_depth_shader = VK_NULL_HANDLE;
VkShaderModule m_poke_vertex_shader = VK_NULL_HANDLE;
VkShaderModule m_poke_geometry_shader = VK_NULL_HANDLE;
VkShaderModule m_poke_fragment_shader = VK_NULL_HANDLE;
};
} // namespace Vulkan
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,189 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#pragma once
#include <array>
#include <map>
#include <memory>
#include <string>
#include <unordered_map>
#include "Common/LinearDiskCache.h"
#include "VideoBackends/Vulkan/Constants.h"
#include "VideoCommon/GeometryShaderGen.h"
#include "VideoCommon/PixelShaderGen.h"
#include "VideoCommon/VertexShaderGen.h"
namespace Vulkan
{
class CommandBufferManager;
class VertexFormat;
class StreamBuffer;
struct PipelineInfo
{
// These are packed in descending order of size, to avoid any padding so that the structure
// can be copied/compared as a single block of memory. 64-bit pointer size is assumed.
const VertexFormat* vertex_format;
VkPipelineLayout pipeline_layout;
VkShaderModule vs;
VkShaderModule gs;
VkShaderModule ps;
VkRenderPass render_pass;
BlendState blend_state;
RasterizationState rasterization_state;
DepthStencilState depth_stencil_state;
VkPrimitiveTopology primitive_topology;
};
struct PipelineInfoHash
{
std::size_t operator()(const PipelineInfo& key) const;
};
bool operator==(const PipelineInfo& lhs, const PipelineInfo& rhs);
bool operator!=(const PipelineInfo& lhs, const PipelineInfo& rhs);
bool operator<(const PipelineInfo& lhs, const PipelineInfo& rhs);
bool operator>(const PipelineInfo& lhs, const PipelineInfo& rhs);
bool operator==(const SamplerState& lhs, const SamplerState& rhs);
bool operator!=(const SamplerState& lhs, const SamplerState& rhs);
bool operator>(const SamplerState& lhs, const SamplerState& rhs);
bool operator<(const SamplerState& lhs, const SamplerState& rhs);
class ObjectCache
{
public:
ObjectCache();
~ObjectCache();
// We have four shared pipeline layouts:
// - Standard
// - Per-stage UBO (VS/GS/PS, VS constants accessible from PS)
// - 8 combined image samplers (accessible from PS)
// - BBox Enabled
// - Same as standard, plus a single SSBO accessible from PS
// - Push Constant
// - Same as standard, plus 128 bytes of push constants, accessible from all stages.
//
// All three pipeline layouts use the same descriptor set layouts, but the final descriptor set
// (SSBO) is only required when using the BBox Enabled pipeline layout.
//
VkDescriptorSetLayout GetDescriptorSetLayout(DESCRIPTOR_SET set) const
{
return m_descriptor_set_layouts[set];
}
VkPipelineLayout GetStandardPipelineLayout() const { return m_standard_pipeline_layout; }
VkPipelineLayout GetBBoxPipelineLayout() const { return m_bbox_pipeline_layout; }
VkPipelineLayout GetPushConstantPipelineLayout() const { return m_push_constant_pipeline_layout; }
// Shared utility shader resources
VertexFormat* GetUtilityShaderVertexFormat() const
{
return m_utility_shader_vertex_format.get();
}
StreamBuffer* GetUtilityShaderVertexBuffer() const
{
return m_utility_shader_vertex_buffer.get();
}
StreamBuffer* GetUtilityShaderUniformBuffer() const
{
return m_utility_shader_uniform_buffer.get();
}
// Get utility shader header based on current config.
std::string GetUtilityShaderHeader() const;
// Accesses ShaderGen shader caches
VkShaderModule GetVertexShaderForUid(const VertexShaderUid& uid);
VkShaderModule GetGeometryShaderForUid(const GeometryShaderUid& uid);
VkShaderModule GetPixelShaderForUid(const PixelShaderUid& uid, DSTALPHA_MODE dstalpha_mode);
// Static samplers
VkSampler GetPointSampler() const { return m_point_sampler; }
VkSampler GetLinearSampler() const { return m_linear_sampler; }
VkSampler GetSampler(const SamplerState& info);
// Perform at startup, create descriptor layouts, compiles all static shaders.
bool Initialize();
// Find a pipeline by the specified description, if not found, attempts to create it
VkPipeline GetPipeline(const PipelineInfo& info);
// Wipes out the pipeline cache, use when MSAA modes change, for example
// Also destroys the data that would be stored in the disk cache.
void ClearPipelineCache();
// Saves the pipeline cache to disk. Call when shutting down.
void SavePipelineCache();
// Clear sampler cache, use when anisotropy mode changes
// WARNING: Ensure none of the objects from here are in use when calling
void ClearSamplerCache();
// Recompile shared shaders, call when stereo mode changes.
void RecompileSharedShaders();
// Shared shader accessors
VkShaderModule GetScreenQuadVertexShader() const { return m_screen_quad_vertex_shader; }
VkShaderModule GetPassthroughVertexShader() const { return m_passthrough_vertex_shader; }
VkShaderModule GetScreenQuadGeometryShader() const { return m_screen_quad_geometry_shader; }
VkShaderModule GetPassthroughGeometryShader() const { return m_passthrough_geometry_shader; }
private:
bool CreatePipelineCache(bool load_from_disk);
void DestroyPipelineCache();
void LoadShaderCaches();
void DestroyShaderCaches();
bool CreateDescriptorSetLayouts();
void DestroyDescriptorSetLayouts();
bool CreatePipelineLayouts();
void DestroyPipelineLayouts();
bool CreateUtilityShaderVertexFormat();
bool CreateStaticSamplers();
bool CompileSharedShaders();
void DestroySharedShaders();
void DestroySamplers();
std::string GetDiskCacheFileName(const char* type);
std::array<VkDescriptorSetLayout, NUM_DESCRIPTOR_SETS> m_descriptor_set_layouts = {};
VkPipelineLayout m_standard_pipeline_layout = VK_NULL_HANDLE;
VkPipelineLayout m_bbox_pipeline_layout = VK_NULL_HANDLE;
VkPipelineLayout m_push_constant_pipeline_layout = VK_NULL_HANDLE;
std::unique_ptr<VertexFormat> m_utility_shader_vertex_format;
std::unique_ptr<StreamBuffer> m_utility_shader_vertex_buffer;
std::unique_ptr<StreamBuffer> m_utility_shader_uniform_buffer;
template <typename Uid>
struct ShaderCache
{
std::map<Uid, VkShaderModule> shader_map;
LinearDiskCache<Uid, u32> disk_cache;
};
ShaderCache<VertexShaderUid> m_vs_cache;
ShaderCache<GeometryShaderUid> m_gs_cache;
ShaderCache<PixelShaderUid> m_ps_cache;
std::unordered_map<PipelineInfo, VkPipeline, PipelineInfoHash> m_pipeline_objects;
VkPipelineCache m_pipeline_cache = VK_NULL_HANDLE;
std::string m_pipeline_cache_filename;
VkSampler m_point_sampler = VK_NULL_HANDLE;
VkSampler m_linear_sampler = VK_NULL_HANDLE;
std::map<SamplerState, VkSampler> m_sampler_cache;
// Utility/shared shaders
VkShaderModule m_screen_quad_vertex_shader = VK_NULL_HANDLE;
VkShaderModule m_passthrough_vertex_shader = VK_NULL_HANDLE;
VkShaderModule m_screen_quad_geometry_shader = VK_NULL_HANDLE;
VkShaderModule m_passthrough_geometry_shader = VK_NULL_HANDLE;
};
extern std::unique_ptr<ObjectCache> g_object_cache;
} // namespace Vulkan
@@ -0,0 +1,311 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#include <algorithm>
#include <cstring>
#include "Common/Assert.h"
#include "Common/CommonFuncs.h"
#include "VideoBackends/Vulkan/CommandBufferManager.h"
#include "VideoBackends/Vulkan/FramebufferManager.h"
#include "VideoBackends/Vulkan/ObjectCache.h"
#include "VideoBackends/Vulkan/PaletteTextureConverter.h"
#include "VideoBackends/Vulkan/Renderer.h"
#include "VideoBackends/Vulkan/StateTracker.h"
#include "VideoBackends/Vulkan/StreamBuffer.h"
#include "VideoBackends/Vulkan/Texture2D.h"
#include "VideoBackends/Vulkan/Util.h"
#include "VideoBackends/Vulkan/VulkanContext.h"
namespace Vulkan
{
PaletteTextureConverter::PaletteTextureConverter()
{
}
PaletteTextureConverter::~PaletteTextureConverter()
{
for (const auto& it : m_shaders)
{
if (it != VK_NULL_HANDLE)
vkDestroyShaderModule(g_vulkan_context->GetDevice(), it, nullptr);
}
if (m_palette_buffer_view != VK_NULL_HANDLE)
vkDestroyBufferView(g_vulkan_context->GetDevice(), m_palette_buffer_view, nullptr);
if (m_palette_set_layout != VK_NULL_HANDLE)
vkDestroyDescriptorSetLayout(g_vulkan_context->GetDevice(), m_palette_set_layout, nullptr);
}
bool PaletteTextureConverter::Initialize()
{
if (!CreateBuffers())
return false;
if (!CompileShaders())
return false;
if (!CreateDescriptorLayout())
return false;
return true;
}
void PaletteTextureConverter::ConvertTexture(StateTracker* state_tracker, VkRenderPass render_pass,
VkFramebuffer dst_framebuffer, Texture2D* src_texture,
u32 width, u32 height, void* palette,
TlutFormat format)
{
struct PSUniformBlock
{
float multiplier;
int texel_buffer_offset;
int pad[2];
};
_assert_(static_cast<size_t>(format) < NUM_PALETTE_CONVERSION_SHADERS);
size_t palette_size = ((format & 0xF) == GX_TF_I4) ? 32 : 512;
VkDescriptorSet texel_buffer_descriptor_set;
// Allocate memory for the palette, and descriptor sets for the buffer.
// If any of these fail, execute a command buffer, and try again.
if (!m_palette_stream_buffer->ReserveMemory(palette_size,
g_vulkan_context->GetTexelBufferAlignment()) ||
(texel_buffer_descriptor_set =
g_command_buffer_mgr->AllocateDescriptorSet(m_palette_set_layout)) == VK_NULL_HANDLE)
{
WARN_LOG(VIDEO, "Executing command list while waiting for space in palette buffer");
Util::ExecuteCurrentCommandsAndRestoreState(state_tracker, false);
if (!m_palette_stream_buffer->ReserveMemory(palette_size,
g_vulkan_context->GetTexelBufferAlignment()) ||
(texel_buffer_descriptor_set =
g_command_buffer_mgr->AllocateDescriptorSet(m_palette_set_layout)) == VK_NULL_HANDLE)
{
PanicAlert("Failed to allocate space for texture conversion");
return;
}
}
// Fill descriptor set #2 (texel buffer)
u32 palette_offset = static_cast<u32>(m_palette_stream_buffer->GetCurrentOffset());
VkWriteDescriptorSet texel_set_write = {VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
nullptr,
texel_buffer_descriptor_set,
0,
0,
1,
VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER,
nullptr,
nullptr,
&m_palette_buffer_view};
vkUpdateDescriptorSets(g_vulkan_context->GetDevice(), 1, &texel_set_write, 0, nullptr);
Util::BufferMemoryBarrier(g_command_buffer_mgr->GetCurrentInitCommandBuffer(),
m_palette_stream_buffer->GetBuffer(), VK_ACCESS_HOST_WRITE_BIT,
VK_ACCESS_SHADER_READ_BIT, palette_offset, palette_size,
VK_PIPELINE_STAGE_HOST_BIT, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT);
// Set up draw
UtilityShaderDraw draw(g_command_buffer_mgr->GetCurrentInitCommandBuffer(), m_pipeline_layout,
render_pass, g_object_cache->GetScreenQuadVertexShader(), VK_NULL_HANDLE,
m_shaders[format]);
VkRect2D region = {{0, 0}, {width, height}};
draw.BeginRenderPass(dst_framebuffer, region);
// Copy in palette
memcpy(m_palette_stream_buffer->GetCurrentHostPointer(), palette, palette_size);
m_palette_stream_buffer->CommitMemory(palette_size);
// PS Uniforms/Samplers
PSUniformBlock uniforms = {};
uniforms.multiplier = ((format & 0xF)) == GX_TF_I4 ? 15.0f : 255.0f;
uniforms.texel_buffer_offset = static_cast<int>(palette_offset / sizeof(u16));
draw.SetPushConstants(&uniforms, sizeof(uniforms));
draw.SetPSSampler(0, src_texture->GetView(), g_object_cache->GetPointSampler());
// We have to bind the texel buffer descriptor set separately.
vkCmdBindDescriptorSets(g_command_buffer_mgr->GetCurrentInitCommandBuffer(),
VK_PIPELINE_BIND_POINT_GRAPHICS, m_pipeline_layout, 0, 1,
&texel_buffer_descriptor_set, 0, nullptr);
// Draw
draw.SetViewportAndScissor(0, 0, width, height);
draw.DrawWithoutVertexBuffer(VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP, 4);
draw.EndRenderPass();
}
bool PaletteTextureConverter::CreateBuffers()
{
// TODO: Check against maximum size
static const size_t BUFFER_SIZE = 1024 * 1024;
m_palette_stream_buffer =
StreamBuffer::Create(VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT, BUFFER_SIZE, BUFFER_SIZE);
if (!m_palette_stream_buffer)
return false;
// Create a view of the whole buffer, we'll offset our texel load into it
VkBufferViewCreateInfo view_info = {
VK_STRUCTURE_TYPE_BUFFER_VIEW_CREATE_INFO, // VkStructureType sType
nullptr, // const void* pNext
0, // VkBufferViewCreateFlags flags
m_palette_stream_buffer->GetBuffer(), // VkBuffer buffer
VK_FORMAT_R16_UINT, // VkFormat format
0, // VkDeviceSize offset
BUFFER_SIZE // VkDeviceSize range
};
VkResult res = vkCreateBufferView(g_vulkan_context->GetDevice(), &view_info, nullptr,
&m_palette_buffer_view);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkCreateBufferView failed: ");
return false;
}
return true;
}
bool PaletteTextureConverter::CompileShaders()
{
static const char PALETTE_CONVERSION_FRAGMENT_SHADER_SOURCE[] = R"(
layout(std140, push_constant) uniform PCBlock
{
float multiplier;
int texture_buffer_offset;
} PC;
layout(set = 1, binding = 0) uniform sampler2DArray samp0;
layout(set = 0, binding = 0) uniform usamplerBuffer samp1;
layout(location = 0) in vec3 f_uv0;
layout(location = 0) out vec4 ocol0;
int Convert3To8(int v)
{
// Swizzle bits: 00000123 -> 12312312
return (v << 5) | (v << 2) | (v >> 1);
}
int Convert4To8(int v)
{
// Swizzle bits: 00001234 -> 12341234
return (v << 4) | v;
}
int Convert5To8(int v)
{
// Swizzle bits: 00012345 -> 12345123
return (v << 3) | (v >> 2);
}
int Convert6To8(int v)
{
// Swizzle bits: 00123456 -> 12345612
return (v << 2) | (v >> 4);
}
float4 DecodePixel_RGB5A3(int val)
{
int r,g,b,a;
if ((val&0x8000) > 0)
{
r=Convert5To8((val>>10) & 0x1f);
g=Convert5To8((val>>5 ) & 0x1f);
b=Convert5To8((val ) & 0x1f);
a=0xFF;
}
else
{
a=Convert3To8((val>>12) & 0x7);
r=Convert4To8((val>>8 ) & 0xf);
g=Convert4To8((val>>4 ) & 0xf);
b=Convert4To8((val ) & 0xf);
}
return float4(r, g, b, a) / 255.0;
}
float4 DecodePixel_RGB565(int val)
{
int r, g, b, a;
r = Convert5To8((val >> 11) & 0x1f);
g = Convert6To8((val >> 5) & 0x3f);
b = Convert5To8((val) & 0x1f);
a = 0xFF;
return float4(r, g, b, a) / 255.0;
}
float4 DecodePixel_IA8(int val)
{
int i = val & 0xFF;
int a = val >> 8;
return float4(i, i, i, a) / 255.0;
}
void main()
{
int src = int(round(texture(samp0, f_uv0).r * PC.multiplier));
src = int(texelFetch(samp1, src + PC.texture_buffer_offset).r);
src = ((src << 8) & 0xFF00) | (src >> 8);
ocol0 = DECODE(src);
}
)";
std::string palette_ia8_program = StringFromFormat("%s\n%s", "#define DECODE DecodePixel_IA8",
PALETTE_CONVERSION_FRAGMENT_SHADER_SOURCE);
std::string palette_rgb565_program = StringFromFormat(
"%s\n%s", "#define DECODE DecodePixel_RGB565", PALETTE_CONVERSION_FRAGMENT_SHADER_SOURCE);
std::string palette_rgb5a3_program = StringFromFormat(
"%s\n%s", "#define DECODE DecodePixel_RGB5A3", PALETTE_CONVERSION_FRAGMENT_SHADER_SOURCE);
m_shaders[GX_TL_IA8] = Util::CompileAndCreateFragmentShader(palette_ia8_program);
m_shaders[GX_TL_RGB565] = Util::CompileAndCreateFragmentShader(palette_rgb565_program);
m_shaders[GX_TL_RGB5A3] = Util::CompileAndCreateFragmentShader(palette_rgb5a3_program);
return (m_shaders[GX_TL_IA8] != VK_NULL_HANDLE && m_shaders[GX_TL_RGB565] != VK_NULL_HANDLE &&
m_shaders[GX_TL_RGB5A3] != VK_NULL_HANDLE);
}
bool PaletteTextureConverter::CreateDescriptorLayout()
{
static const VkDescriptorSetLayoutBinding set_bindings[] = {
{0, VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER, 1, VK_SHADER_STAGE_FRAGMENT_BIT},
};
static const VkDescriptorSetLayoutCreateInfo set_info = {
VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO, nullptr, 0,
static_cast<u32>(ArraySize(set_bindings)), set_bindings};
VkResult res = vkCreateDescriptorSetLayout(g_vulkan_context->GetDevice(), &set_info, nullptr,
&m_palette_set_layout);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkCreateDescriptorSetLayout failed: ");
return false;
}
VkDescriptorSetLayout sets[] = {m_palette_set_layout, g_object_cache->GetDescriptorSetLayout(
DESCRIPTOR_SET_PIXEL_SHADER_SAMPLERS)};
VkPushConstantRange push_constant_range = {
VK_SHADER_STAGE_VERTEX_BIT | VK_SHADER_STAGE_FRAGMENT_BIT, 0, PUSH_CONSTANT_BUFFER_SIZE};
VkPipelineLayoutCreateInfo pipeline_layout_info = {VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
nullptr,
0,
static_cast<u32>(ArraySize(sets)),
sets,
1,
&push_constant_range};
res = vkCreatePipelineLayout(g_vulkan_context->GetDevice(), &pipeline_layout_info, nullptr,
&m_pipeline_layout);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkCreatePipelineLayout failed: ");
return false;
}
return true;
}
} // namespace Vulkan
@@ -0,0 +1,50 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#pragma once
#include <memory>
#include "VideoBackends/Vulkan/StreamBuffer.h"
#include "VideoCommon/TextureDecoder.h"
namespace Vulkan
{
class StateTracker;
class Texture2D;
// Since this converter uses a uniform texel buffer, we can't use the general pipeline generators.
class PaletteTextureConverter
{
public:
PaletteTextureConverter();
~PaletteTextureConverter();
bool Initialize();
void ConvertTexture(StateTracker* state_tracker, VkRenderPass render_pass,
VkFramebuffer dst_framebuffer, Texture2D* src_texture, u32 width, u32 height,
void* palette, TlutFormat format);
private:
static const size_t NUM_PALETTE_CONVERSION_SHADERS = 3;
bool CreateBuffers();
bool CompileShaders();
bool CreateDescriptorLayout();
VkDescriptorSetLayout m_palette_set_layout = VK_NULL_HANDLE;
VkPipelineLayout m_pipeline_layout = VK_NULL_HANDLE;
std::array<VkShaderModule, NUM_PALETTE_CONVERSION_SHADERS> m_shaders = {};
std::unique_ptr<StreamBuffer> m_palette_stream_buffer;
VkBufferView m_palette_buffer_view = VK_NULL_HANDLE;
std::unique_ptr<StreamBuffer> m_uniform_buffer;
};
} // namespace Vulkan
@@ -0,0 +1,370 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#include <algorithm>
#include <cstring>
#include <functional>
#include "Common/Assert.h"
#include "VideoBackends/Vulkan/CommandBufferManager.h"
#include "VideoBackends/Vulkan/PerfQuery.h"
#include "VideoBackends/Vulkan/Renderer.h"
#include "VideoBackends/Vulkan/StagingBuffer.h"
#include "VideoBackends/Vulkan/StateTracker.h"
#include "VideoBackends/Vulkan/Util.h"
#include "VideoBackends/Vulkan/VulkanContext.h"
namespace Vulkan
{
PerfQuery::PerfQuery()
{
}
PerfQuery::~PerfQuery()
{
g_command_buffer_mgr->RemoveFencePointCallback(this);
if (m_query_pool != VK_NULL_HANDLE)
vkDestroyQueryPool(g_vulkan_context->GetDevice(), m_query_pool, nullptr);
}
bool PerfQuery::Initialize(StateTracker* state_tracker)
{
m_state_tracker = state_tracker;
if (!CreateQueryPool())
{
PanicAlert("Failed to create query pool");
return false;
}
if (!CreateReadbackBuffer())
{
PanicAlert("Failed to create readback buffer");
return false;
}
g_command_buffer_mgr->AddFencePointCallback(
this, std::bind(&PerfQuery::OnCommandBufferQueued, this, std::placeholders::_1,
std::placeholders::_2),
std::bind(&PerfQuery::OnCommandBufferExecuted, this, std::placeholders::_1));
return true;
}
void PerfQuery::EnableQuery(PerfQueryGroup type)
{
// Have we used half of the query buffer already?
if (m_query_count > m_query_buffer.size() / 2)
NonBlockingPartialFlush();
// Block if there are no free slots.
if (m_query_count == PERF_QUERY_BUFFER_SIZE)
{
// ERROR_LOG(VIDEO, "Flushed query buffer early!");
BlockingPartialFlush();
}
if (type == PQG_ZCOMP_ZCOMPLOC || type == PQG_ZCOMP)
{
u32 index = (m_query_read_pos + m_query_count) % PERF_QUERY_BUFFER_SIZE;
ActiveQuery& entry = m_query_buffer[index];
_assert_(!entry.active && !entry.available);
entry.active = true;
m_query_count++;
DEBUG_LOG(VIDEO, "start query %u", index);
// Use precise queries if supported, otherwise boolean (which will be incorrect).
VkQueryControlFlags flags = 0;
if (g_vulkan_context->SupportsPreciseOcclusionQueries())
flags = VK_QUERY_CONTROL_PRECISE_BIT;
// Ensure the query starts within a render pass.
// TODO: Is this needed?
m_state_tracker->BeginRenderPass();
vkCmdBeginQuery(g_command_buffer_mgr->GetCurrentCommandBuffer(), m_query_pool, index, flags);
// Prevent background command buffer submission while the query is active.
m_state_tracker->SetBackgroundCommandBufferExecution(false);
}
}
void PerfQuery::DisableQuery(PerfQueryGroup type)
{
if (type == PQG_ZCOMP_ZCOMPLOC || type == PQG_ZCOMP)
{
// DisableQuery should be called for each EnableQuery, so subtract one to get the previous one.
u32 index = (m_query_read_pos + m_query_count - 1) % PERF_QUERY_BUFFER_SIZE;
vkCmdEndQuery(g_command_buffer_mgr->GetCurrentCommandBuffer(), m_query_pool, index);
m_state_tracker->SetBackgroundCommandBufferExecution(true);
DEBUG_LOG(VIDEO, "end query %u", index);
}
}
void PerfQuery::ResetQuery()
{
m_query_count = 0;
m_query_read_pos = 0;
std::fill_n(m_results, ArraySize(m_results), 0);
// Reset entire query pool, ensuring all queries are ready to write to.
m_state_tracker->EndRenderPass();
vkCmdResetQueryPool(g_command_buffer_mgr->GetCurrentCommandBuffer(), m_query_pool, 0,
PERF_QUERY_BUFFER_SIZE);
for (auto& entry : m_query_buffer)
{
entry.pending_fence = VK_NULL_HANDLE;
entry.available = false;
entry.active = false;
}
}
u32 PerfQuery::GetQueryResult(PerfQueryType type)
{
u32 result = 0;
if (type == PQ_ZCOMP_INPUT_ZCOMPLOC || type == PQ_ZCOMP_OUTPUT_ZCOMPLOC)
{
result = m_results[PQG_ZCOMP_ZCOMPLOC];
}
else if (type == PQ_ZCOMP_INPUT || type == PQ_ZCOMP_OUTPUT)
{
result = m_results[PQG_ZCOMP];
}
else if (type == PQ_BLEND_INPUT)
{
result = m_results[PQG_ZCOMP] + m_results[PQG_ZCOMP_ZCOMPLOC];
}
else if (type == PQ_EFB_COPY_CLOCKS)
{
result = m_results[PQG_EFB_COPY_CLOCKS];
}
return result / 4;
}
void PerfQuery::FlushResults()
{
while (!IsFlushed())
BlockingPartialFlush();
}
bool PerfQuery::IsFlushed() const
{
return m_query_count == 0;
}
bool PerfQuery::CreateQueryPool()
{
VkQueryPoolCreateInfo info = {
VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO, // VkStructureType sType
nullptr, // const void* pNext
0, // VkQueryPoolCreateFlags flags
VK_QUERY_TYPE_OCCLUSION, // VkQueryType queryType
PERF_QUERY_BUFFER_SIZE, // uint32_t queryCount
0 // VkQueryPipelineStatisticFlags pipelineStatistics;
};
VkResult res = vkCreateQueryPool(g_vulkan_context->GetDevice(), &info, nullptr, &m_query_pool);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkCreateQueryPool failed: ");
return false;
}
return true;
}
bool PerfQuery::CreateReadbackBuffer()
{
m_readback_buffer = StagingBuffer::Create(STAGING_BUFFER_TYPE_READBACK,
PERF_QUERY_BUFFER_SIZE * sizeof(PerfQueryDataType),
VK_BUFFER_USAGE_TRANSFER_DST_BIT);
// Leave the buffer persistently mapped, we invalidate it when we need to read.
if (!m_readback_buffer || !m_readback_buffer->Map())
return false;
return true;
}
void PerfQuery::QueueCopyQueryResults(VkCommandBuffer command_buffer, VkFence fence,
u32 start_index, u32 query_count)
{
DEBUG_LOG(VIDEO, "queue copy of queries %u-%u", start_index, start_index + query_count - 1);
// Transition buffer for GPU write
// TODO: Is this needed?
m_readback_buffer->PrepareForGPUWrite(command_buffer, VK_ACCESS_TRANSFER_WRITE_BIT,
VK_PIPELINE_STAGE_TRANSFER_BIT);
// Copy from queries -> buffer
vkCmdCopyQueryPoolResults(command_buffer, m_query_pool, start_index, query_count,
m_readback_buffer->GetBuffer(), start_index * sizeof(PerfQueryDataType),
sizeof(PerfQueryDataType), VK_QUERY_RESULT_WAIT_BIT);
// Prepare for host readback
m_readback_buffer->FlushGPUCache(command_buffer, VK_ACCESS_TRANSFER_WRITE_BIT,
VK_PIPELINE_STAGE_TRANSFER_BIT);
// Reset queries so they're ready to use again
vkCmdResetQueryPool(command_buffer, m_query_pool, start_index, query_count);
// Flag all queries as available, but with a fence that has to be completed first
for (u32 i = 0; i < query_count; i++)
{
u32 index = start_index + i;
ActiveQuery& entry = m_query_buffer[index];
entry.pending_fence = fence;
entry.available = true;
entry.active = false;
}
}
void PerfQuery::OnCommandBufferQueued(VkCommandBuffer command_buffer, VkFence fence)
{
// Flag all pending queries that aren't available as available after execution.
u32 copy_start_index = 0;
u32 copy_count = 0;
for (u32 i = 0; i < m_query_count; i++)
{
u32 index = (m_query_read_pos + i) % PERF_QUERY_BUFFER_SIZE;
ActiveQuery& entry = m_query_buffer[index];
// Skip already-copied queries (will happen if a flush hasn't occurred and
// a command buffer hasn't finished executing).
if (entry.available)
{
// These should be grouped together, and at the start.
_assert_(copy_count == 0);
continue;
}
// If this wrapped around, we need to flush the entries before the end of the buffer.
_assert_(entry.active);
if (index < copy_start_index)
{
QueueCopyQueryResults(command_buffer, fence, copy_start_index, copy_count);
copy_start_index = index;
copy_count = 0;
}
else if (copy_count == 0)
{
copy_start_index = index;
}
copy_count++;
}
if (copy_count > 0)
QueueCopyQueryResults(command_buffer, fence, copy_start_index, copy_count);
}
void PerfQuery::OnCommandBufferExecuted(VkFence fence)
{
// Need to save these since ProcessResults will modify them.
u32 query_read_pos = m_query_read_pos;
u32 query_count = m_query_count;
// Flush as many queries as are bound to this fence.
u32 flush_start_index = 0;
u32 flush_count = 0;
for (u32 i = 0; i < query_count; i++)
{
u32 index = (query_read_pos + i) % PERF_QUERY_BUFFER_SIZE;
if (m_query_buffer[index].pending_fence != fence)
{
// These should be grouped together, at the end.
break;
}
// If this wrapped around, we need to flush the entries before the end of the buffer.
if (index < flush_start_index)
{
ProcessResults(flush_start_index, flush_count);
flush_start_index = index;
flush_count = 0;
}
else if (flush_count == 0)
{
flush_start_index = index;
}
flush_count++;
}
if (flush_count > 0)
ProcessResults(flush_start_index, flush_count);
}
void PerfQuery::ProcessResults(u32 start_index, u32 query_count)
{
// Invalidate CPU caches before reading back.
m_readback_buffer->InvalidateCPUCache(start_index * sizeof(PerfQueryDataType),
query_count * sizeof(PerfQueryDataType));
// Should be at maximum query_count queries pending.
_assert_(query_count <= m_query_count);
DEBUG_LOG(VIDEO, "process queries %u-%u", start_index, start_index + query_count - 1);
// Remove pending queries.
for (u32 i = 0; i < query_count; i++)
{
u32 index = (m_query_read_pos + i) % PERF_QUERY_BUFFER_SIZE;
ActiveQuery& entry = m_query_buffer[index];
// Should have a fence associated with it (waiting for a result).
_assert_(entry.pending_fence != VK_NULL_HANDLE);
entry.pending_fence = VK_NULL_HANDLE;
entry.available = false;
entry.active = false;
// Grab result from readback buffer, it will already have been invalidated.
u32 result;
m_readback_buffer->Read(index * sizeof(PerfQueryDataType), &result, sizeof(result), false);
DEBUG_LOG(VIDEO, " query result %u", result);
// NOTE: Reported pixel metrics should be referenced to native resolution
m_results[entry.query_type] +=
static_cast<u32>(static_cast<u64>(result) * EFB_WIDTH / g_renderer->GetTargetWidth() *
EFB_HEIGHT / g_renderer->GetTargetHeight());
}
m_query_read_pos = (m_query_read_pos + query_count) % PERF_QUERY_BUFFER_SIZE;
m_query_count -= query_count;
}
void PerfQuery::NonBlockingPartialFlush()
{
if (IsFlushed())
return;
// Submit a command buffer in the background if the front query is not bound to one.
// Ideally this will complete before the buffer fills.
if (m_query_buffer[m_query_read_pos].pending_fence == VK_NULL_HANDLE)
Util::ExecuteCurrentCommandsAndRestoreState(m_state_tracker, true, false);
}
void PerfQuery::BlockingPartialFlush()
{
if (IsFlushed())
return;
// If the first pending query is needing command buffer execution, do that.
ActiveQuery& entry = m_query_buffer[m_query_read_pos];
if (entry.pending_fence == VK_NULL_HANDLE)
{
// This will callback OnCommandBufferQueued which will set the fence on the entry.
// We wait for completion, which will also call OnCommandBufferExecuted, and clear the fence.
Util::ExecuteCurrentCommandsAndRestoreState(m_state_tracker, false, true);
}
else
{
// The command buffer has been submitted, but is awaiting completion.
// Wait for the fence to complete, which will call OnCommandBufferExecuted.
g_command_buffer_mgr->WaitForFence(entry.pending_fence);
}
}
}
@@ -0,0 +1,70 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#pragma once
#include <array>
#include <memory>
#include "VideoBackends/Vulkan/Constants.h"
#include "VideoCommon/PerfQueryBase.h"
namespace Vulkan
{
class StagingBuffer;
class StateTracker;
class PerfQuery : public PerfQueryBase
{
public:
PerfQuery();
~PerfQuery();
bool Initialize(StateTracker* state_tracker);
void EnableQuery(PerfQueryGroup type) override;
void DisableQuery(PerfQueryGroup type) override;
void ResetQuery() override;
u32 GetQueryResult(PerfQueryType type) override;
void FlushResults() override;
bool IsFlushed() const override;
private:
struct ActiveQuery
{
PerfQueryType query_type;
VkFence pending_fence;
bool available;
bool active;
};
bool CreateQueryPool();
bool CreateReadbackBuffer();
void QueueCopyQueryResults(VkCommandBuffer command_buffer, VkFence fence, u32 start_index,
u32 query_count);
void ProcessResults(u32 start_index, u32 query_count);
void OnCommandBufferQueued(VkCommandBuffer command_buffer, VkFence fence);
void OnCommandBufferExecuted(VkFence fence);
void NonBlockingPartialFlush();
void BlockingPartialFlush();
StateTracker* m_state_tracker = nullptr;
// when testing in SMS: 64 was too small, 128 was ok
// TODO: This should be size_t, but the base class uses u32s
using PerfQueryDataType = u32;
static const u32 PERF_QUERY_BUFFER_SIZE = 512;
std::array<ActiveQuery, PERF_QUERY_BUFFER_SIZE> m_query_buffer = {};
u32 m_query_read_pos = 0;
// TODO: Investigate using pipeline statistics to implement other query types
VkQueryPool m_query_pool = VK_NULL_HANDLE;
// Buffer containing query results. Each query is a u32.
std::unique_ptr<StagingBuffer> m_readback_buffer;
};
} // namespace Vulkan
@@ -0,0 +1,408 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#include <vector>
#include "VideoBackends/Vulkan/CommandBufferManager.h"
#include "VideoBackends/Vulkan/RasterFont.h"
#include "VideoBackends/Vulkan/Texture2D.h"
#include "VideoBackends/Vulkan/Util.h"
#include "VideoBackends/Vulkan/VulkanContext.h"
// Based on OGL RasterFont
// TODO: We should move this to common.
namespace Vulkan
{
constexpr int CHAR_WIDTH = 8;
constexpr int CHAR_HEIGHT = 13;
constexpr int CHAR_OFFSET = 32;
constexpr int CHAR_COUNT = 95;
static const u8 rasters[CHAR_COUNT][CHAR_HEIGHT] = {
{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x18, 0x18, 0x00, 0x00, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18},
{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x36, 0x36, 0x36, 0x36},
{0x00, 0x00, 0x00, 0x66, 0x66, 0xff, 0x66, 0x66, 0xff, 0x66, 0x66, 0x00, 0x00},
{0x00, 0x00, 0x18, 0x7e, 0xff, 0x1b, 0x1f, 0x7e, 0xf8, 0xd8, 0xff, 0x7e, 0x18},
{0x00, 0x00, 0x0e, 0x1b, 0xdb, 0x6e, 0x30, 0x18, 0x0c, 0x76, 0xdb, 0xd8, 0x70},
{0x00, 0x00, 0x7f, 0xc6, 0xcf, 0xd8, 0x70, 0x70, 0xd8, 0xcc, 0xcc, 0x6c, 0x38},
{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x18, 0x1c, 0x0c, 0x0e},
{0x00, 0x00, 0x0c, 0x18, 0x30, 0x30, 0x30, 0x30, 0x30, 0x30, 0x30, 0x18, 0x0c},
{0x00, 0x00, 0x30, 0x18, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x18, 0x30},
{0x00, 0x00, 0x00, 0x00, 0x99, 0x5a, 0x3c, 0xff, 0x3c, 0x5a, 0x99, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x18, 0x18, 0x18, 0xff, 0xff, 0x18, 0x18, 0x18, 0x00, 0x00},
{0x00, 0x00, 0x30, 0x18, 0x1c, 0x1c, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x38, 0x38, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x60, 0x60, 0x30, 0x30, 0x18, 0x18, 0x0c, 0x0c, 0x06, 0x06, 0x03, 0x03},
{0x00, 0x00, 0x3c, 0x66, 0xc3, 0xe3, 0xf3, 0xdb, 0xcf, 0xc7, 0xc3, 0x66, 0x3c},
{0x00, 0x00, 0x7e, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x78, 0x38, 0x18},
{0x00, 0x00, 0xff, 0xc0, 0xc0, 0x60, 0x30, 0x18, 0x0c, 0x06, 0x03, 0xe7, 0x7e},
{0x00, 0x00, 0x7e, 0xe7, 0x03, 0x03, 0x07, 0x7e, 0x07, 0x03, 0x03, 0xe7, 0x7e},
{0x00, 0x00, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0xff, 0xcc, 0x6c, 0x3c, 0x1c, 0x0c},
{0x00, 0x00, 0x7e, 0xe7, 0x03, 0x03, 0x07, 0xfe, 0xc0, 0xc0, 0xc0, 0xc0, 0xff},
{0x00, 0x00, 0x7e, 0xe7, 0xc3, 0xc3, 0xc7, 0xfe, 0xc0, 0xc0, 0xc0, 0xe7, 0x7e},
{0x00, 0x00, 0x30, 0x30, 0x30, 0x30, 0x18, 0x0c, 0x06, 0x03, 0x03, 0x03, 0xff},
{0x00, 0x00, 0x7e, 0xe7, 0xc3, 0xc3, 0xe7, 0x7e, 0xe7, 0xc3, 0xc3, 0xe7, 0x7e},
{0x00, 0x00, 0x7e, 0xe7, 0x03, 0x03, 0x03, 0x7f, 0xe7, 0xc3, 0xc3, 0xe7, 0x7e},
{0x00, 0x00, 0x00, 0x38, 0x38, 0x00, 0x00, 0x38, 0x38, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x30, 0x18, 0x1c, 0x1c, 0x00, 0x00, 0x1c, 0x1c, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x06, 0x0c, 0x18, 0x30, 0x60, 0xc0, 0x60, 0x30, 0x18, 0x0c, 0x06},
{0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0x00, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x60, 0x30, 0x18, 0x0c, 0x06, 0x03, 0x06, 0x0c, 0x18, 0x30, 0x60},
{0x00, 0x00, 0x18, 0x00, 0x00, 0x18, 0x18, 0x0c, 0x06, 0x03, 0xc3, 0xc3, 0x7e},
{0x00, 0x00, 0x3f, 0x60, 0xcf, 0xdb, 0xd3, 0xdd, 0xc3, 0x7e, 0x00, 0x00, 0x00},
{0x00, 0x00, 0xc3, 0xc3, 0xc3, 0xc3, 0xff, 0xc3, 0xc3, 0xc3, 0x66, 0x3c, 0x18},
{0x00, 0x00, 0xfe, 0xc7, 0xc3, 0xc3, 0xc7, 0xfe, 0xc7, 0xc3, 0xc3, 0xc7, 0xfe},
{0x00, 0x00, 0x7e, 0xe7, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0, 0xe7, 0x7e},
{0x00, 0x00, 0xfc, 0xce, 0xc7, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xc7, 0xce, 0xfc},
{0x00, 0x00, 0xff, 0xc0, 0xc0, 0xc0, 0xc0, 0xfc, 0xc0, 0xc0, 0xc0, 0xc0, 0xff},
{0x00, 0x00, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0, 0xfc, 0xc0, 0xc0, 0xc0, 0xff},
{0x00, 0x00, 0x7e, 0xe7, 0xc3, 0xc3, 0xcf, 0xc0, 0xc0, 0xc0, 0xc0, 0xe7, 0x7e},
{0x00, 0x00, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xff, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3},
{0x00, 0x00, 0x7e, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x7e},
{0x00, 0x00, 0x7c, 0xee, 0xc6, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06},
{0x00, 0x00, 0xc3, 0xc6, 0xcc, 0xd8, 0xf0, 0xe0, 0xf0, 0xd8, 0xcc, 0xc6, 0xc3},
{0x00, 0x00, 0xff, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0},
{0x00, 0x00, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xdb, 0xff, 0xff, 0xe7, 0xc3},
{0x00, 0x00, 0xc7, 0xc7, 0xcf, 0xcf, 0xdf, 0xdb, 0xfb, 0xf3, 0xf3, 0xe3, 0xe3},
{0x00, 0x00, 0x7e, 0xe7, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xe7, 0x7e},
{0x00, 0x00, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0, 0xfe, 0xc7, 0xc3, 0xc3, 0xc7, 0xfe},
{0x00, 0x00, 0x3f, 0x6e, 0xdf, 0xdb, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0x66, 0x3c},
{0x00, 0x00, 0xc3, 0xc6, 0xcc, 0xd8, 0xf0, 0xfe, 0xc7, 0xc3, 0xc3, 0xc7, 0xfe},
{0x00, 0x00, 0x7e, 0xe7, 0x03, 0x03, 0x07, 0x7e, 0xe0, 0xc0, 0xc0, 0xe7, 0x7e},
{0x00, 0x00, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0xff},
{0x00, 0x00, 0x7e, 0xe7, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3},
{0x00, 0x00, 0x18, 0x3c, 0x3c, 0x66, 0x66, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3},
{0x00, 0x00, 0xc3, 0xe7, 0xff, 0xff, 0xdb, 0xdb, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3},
{0x00, 0x00, 0xc3, 0x66, 0x66, 0x3c, 0x3c, 0x18, 0x3c, 0x3c, 0x66, 0x66, 0xc3},
{0x00, 0x00, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x3c, 0x3c, 0x66, 0x66, 0xc3},
{0x00, 0x00, 0xff, 0xc0, 0xc0, 0x60, 0x30, 0x7e, 0x0c, 0x06, 0x03, 0x03, 0xff},
{0x00, 0x00, 0x3c, 0x30, 0x30, 0x30, 0x30, 0x30, 0x30, 0x30, 0x30, 0x30, 0x3c},
{0x00, 0x03, 0x03, 0x06, 0x06, 0x0c, 0x0c, 0x18, 0x18, 0x30, 0x30, 0x60, 0x60},
{0x00, 0x00, 0x3c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x3c},
{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xc3, 0x66, 0x3c, 0x18},
{0xff, 0xff, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x18, 0x38, 0x30, 0x70},
{0x00, 0x00, 0x7f, 0xc3, 0xc3, 0x7f, 0x03, 0xc3, 0x7e, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0xfe, 0xc3, 0xc3, 0xc3, 0xc3, 0xfe, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0},
{0x00, 0x00, 0x7e, 0xc3, 0xc0, 0xc0, 0xc0, 0xc3, 0x7e, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x7f, 0xc3, 0xc3, 0xc3, 0xc3, 0x7f, 0x03, 0x03, 0x03, 0x03, 0x03},
{0x00, 0x00, 0x7f, 0xc0, 0xc0, 0xfe, 0xc3, 0xc3, 0x7e, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x30, 0x30, 0x30, 0x30, 0x30, 0xfc, 0x30, 0x30, 0x30, 0x33, 0x1e},
{0x7e, 0xc3, 0x03, 0x03, 0x7f, 0xc3, 0xc3, 0xc3, 0x7e, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xc3, 0xfe, 0xc0, 0xc0, 0xc0, 0xc0},
{0x00, 0x00, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x00, 0x00, 0x18, 0x00},
{0x38, 0x6c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x00, 0x00, 0x0c, 0x00},
{0x00, 0x00, 0xc6, 0xcc, 0xf8, 0xf0, 0xd8, 0xcc, 0xc6, 0xc0, 0xc0, 0xc0, 0xc0},
{0x00, 0x00, 0x7e, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x78},
{0x00, 0x00, 0xdb, 0xdb, 0xdb, 0xdb, 0xdb, 0xdb, 0xfe, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0xc6, 0xc6, 0xc6, 0xc6, 0xc6, 0xc6, 0xfc, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x7c, 0xc6, 0xc6, 0xc6, 0xc6, 0xc6, 0x7c, 0x00, 0x00, 0x00, 0x00},
{0xc0, 0xc0, 0xc0, 0xfe, 0xc3, 0xc3, 0xc3, 0xc3, 0xfe, 0x00, 0x00, 0x00, 0x00},
{0x03, 0x03, 0x03, 0x7f, 0xc3, 0xc3, 0xc3, 0xc3, 0x7f, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0xc0, 0xc0, 0xc0, 0xc0, 0xc0, 0xe0, 0xfe, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0xfe, 0x03, 0x03, 0x7e, 0xc0, 0xc0, 0x7f, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x1c, 0x36, 0x30, 0x30, 0x30, 0x30, 0xfc, 0x30, 0x30, 0x30, 0x00},
{0x00, 0x00, 0x7e, 0xc6, 0xc6, 0xc6, 0xc6, 0xc6, 0xc6, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x18, 0x3c, 0x3c, 0x66, 0x66, 0xc3, 0xc3, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0xc3, 0xe7, 0xff, 0xdb, 0xc3, 0xc3, 0xc3, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0xc3, 0x66, 0x3c, 0x18, 0x3c, 0x66, 0xc3, 0x00, 0x00, 0x00, 0x00},
{0xc0, 0x60, 0x60, 0x30, 0x18, 0x3c, 0x66, 0x66, 0xc3, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0xff, 0x60, 0x30, 0x18, 0x0c, 0x06, 0xff, 0x00, 0x00, 0x00, 0x00},
{0x00, 0x00, 0x0f, 0x18, 0x18, 0x18, 0x38, 0xf0, 0x38, 0x18, 0x18, 0x18, 0x0f},
{0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18, 0x18},
{0x00, 0x00, 0xf0, 0x18, 0x18, 0x18, 0x1c, 0x0f, 0x1c, 0x18, 0x18, 0x18, 0xf0},
{0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x06, 0x8f, 0xf1, 0x60, 0x00, 0x00, 0x00}};
static const char VERTEX_SHADER_SOURCE[] = R"(
layout(std140, push_constant) uniform PCBlock {
vec2 char_size;
vec2 offset;
vec4 color;
} PC;
layout(location = 0) in vec4 ipos;
layout(location = 5) in vec4 icol0;
layout(location = 8) in vec3 itex0;
layout(location = 0) out vec2 uv0;
void main()
{
gl_Position = vec4(ipos.xy + PC.offset, 0.0f, 1.0f);
gl_Position.y = -gl_Position.y;
uv0 = itex0.xy * PC.char_size;
}
)";
static const char FRAGMENT_SHADER_SOURCE[] = R"(
layout(std140, push_constant) uniform PCBlock {
vec2 char_size;
vec2 offset;
vec4 color;
} PC;
layout(set = 1, binding = 0) uniform sampler2D samp0;
layout(location = 0) in vec2 uv0;
layout(location = 0) out vec4 ocol0;
void main()
{
ocol0 = texture(samp0, uv0) * PC.color;
}
)";
RasterFont::RasterFont()
{
}
RasterFont::~RasterFont()
{
if (m_vertex_shader != VK_NULL_HANDLE)
g_command_buffer_mgr->DeferResourceDestruction(m_vertex_shader);
if (m_fragment_shader != VK_NULL_HANDLE)
g_command_buffer_mgr->DeferResourceDestruction(m_fragment_shader);
}
bool RasterFont::Initialize()
{
// Create shaders and texture
if (!CreateShaders() || !CreateTexture())
return false;
return true;
}
bool RasterFont::CreateTexture()
{
// generate the texture
std::vector<u32> texture_data(CHAR_WIDTH * CHAR_COUNT * CHAR_HEIGHT);
for (int y = 0; y < CHAR_HEIGHT; y++)
{
for (int c = 0; c < CHAR_COUNT; c++)
{
for (int x = 0; x < CHAR_WIDTH; x++)
{
bool pixel = (0 != (rasters[c][y] & (1 << (CHAR_WIDTH - x - 1))));
texture_data[CHAR_WIDTH * CHAR_COUNT * y + CHAR_WIDTH * c + x] = pixel ? -1 : 0;
}
}
}
// create the actual texture object
m_texture =
Texture2D::Create(CHAR_WIDTH * CHAR_COUNT, CHAR_HEIGHT, 1, 1, VK_FORMAT_R8G8B8A8_UNORM,
VK_SAMPLE_COUNT_1_BIT, VK_IMAGE_VIEW_TYPE_2D, VK_IMAGE_TILING_OPTIMAL,
VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT);
if (!m_texture)
return false;
// create temporary buffer for uploading texture
VkBufferCreateInfo buffer_info = {VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
nullptr,
0,
static_cast<VkDeviceSize>(texture_data.size() * sizeof(u32)),
VK_BUFFER_USAGE_TRANSFER_SRC_BIT,
VK_SHARING_MODE_EXCLUSIVE,
0,
nullptr};
VkBuffer temp_buffer;
VkResult res = vkCreateBuffer(g_vulkan_context->GetDevice(), &buffer_info, nullptr, &temp_buffer);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkCreateBuffer failed: ");
return false;
}
VkMemoryRequirements memory_requirements;
vkGetBufferMemoryRequirements(g_vulkan_context->GetDevice(), temp_buffer, &memory_requirements);
uint32_t memory_type_index = g_vulkan_context->GetMemoryType(memory_requirements.memoryTypeBits,
VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT);
VkMemoryAllocateInfo memory_allocate_info = {VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO, nullptr,
memory_requirements.size, memory_type_index};
VkDeviceMemory temp_buffer_memory;
res = vkAllocateMemory(g_vulkan_context->GetDevice(), &memory_allocate_info, nullptr,
&temp_buffer_memory);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkAllocateMemory failed: ");
vkDestroyBuffer(g_vulkan_context->GetDevice(), temp_buffer, nullptr);
return false;
}
// Bind buffer to memory
res = vkBindBufferMemory(g_vulkan_context->GetDevice(), temp_buffer, temp_buffer_memory, 0);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkBindBufferMemory failed: ");
vkDestroyBuffer(g_vulkan_context->GetDevice(), temp_buffer, nullptr);
vkFreeMemory(g_vulkan_context->GetDevice(), temp_buffer_memory, nullptr);
return false;
}
// Copy into buffer
void* mapped_ptr;
res = vkMapMemory(g_vulkan_context->GetDevice(), temp_buffer_memory, 0, buffer_info.size, 0,
&mapped_ptr);
if (res != VK_SUCCESS)
{
LOG_VULKAN_ERROR(res, "vkMapMemory failed: ");
vkDestroyBuffer(g_vulkan_context->GetDevice(), temp_buffer, nullptr);
vkFreeMemory(g_vulkan_context->GetDevice(), temp_buffer_memory, nullptr);
return false;
}
// Copy texture data into staging buffer
memcpy(mapped_ptr, texture_data.data(), texture_data.size() * sizeof(u32));
vkUnmapMemory(g_vulkan_context->GetDevice(), temp_buffer_memory);
// Copy from staging buffer to the final texture
VkBufferImageCopy region = {0, CHAR_WIDTH * CHAR_COUNT,
CHAR_HEIGHT, {VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1},
{0, 0, 0}, {CHAR_WIDTH * CHAR_COUNT, CHAR_HEIGHT, 1}};
m_texture->TransitionToLayout(g_command_buffer_mgr->GetCurrentInitCommandBuffer(),
VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
vkCmdCopyBufferToImage(g_command_buffer_mgr->GetCurrentInitCommandBuffer(), temp_buffer,
m_texture->GetImage(), VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &region);
// Free temp buffers after command buffer executes
m_texture->TransitionToLayout(g_command_buffer_mgr->GetCurrentInitCommandBuffer(),
VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
g_command_buffer_mgr->DeferResourceDestruction(temp_buffer);
g_command_buffer_mgr->DeferResourceDestruction(temp_buffer_memory);
return true;
}
bool RasterFont::CreateShaders()
{
m_vertex_shader = Util::CompileAndCreateVertexShader(VERTEX_SHADER_SOURCE);
m_fragment_shader = Util::CompileAndCreateFragmentShader(FRAGMENT_SHADER_SOURCE);
return m_vertex_shader != VK_NULL_HANDLE && m_fragment_shader != VK_NULL_HANDLE;
}
void RasterFont::PrintMultiLineText(VkRenderPass render_pass, const std::string& text,
float start_x, float start_y, u32 bbWidth, u32 bbHeight,
u32 color)
{
// skip empty strings
if (text.empty())
return;
UtilityShaderDraw draw(g_command_buffer_mgr->GetCurrentCommandBuffer(),
g_object_cache->GetPushConstantPipelineLayout(), render_pass,
m_vertex_shader, VK_NULL_HANDLE, m_fragment_shader);
UtilityShaderVertex* vertices =
draw.ReserveVertices(VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST, text.length() * 6);
size_t num_vertices = 0;
if (!vertices)
return;
float delta_x = float(2 * CHAR_WIDTH) / float(bbWidth);
float delta_y = float(2 * CHAR_HEIGHT) / float(bbHeight);
float border_x = 2.0f / float(bbWidth);
float border_y = 4.0f / float(bbHeight);
float x = float(start_x);
float y = float(start_y);
for (const char& c : text)
{
if (c == '\n')
{
x = float(start_x);
y -= delta_y + border_y;
continue;
}
// do not print spaces, they can be skipped easily
if (c == ' ')
{
x += delta_x + border_x;
continue;
}
if (c < CHAR_OFFSET || c >= CHAR_COUNT + CHAR_OFFSET)
continue;
vertices[num_vertices].SetPosition(x, y);
vertices[num_vertices].SetTextureCoordinates(static_cast<float>(c - CHAR_OFFSET), 0.0f);
num_vertices++;
vertices[num_vertices].SetPosition(x + delta_x, y);
vertices[num_vertices].SetTextureCoordinates(static_cast<float>(c - CHAR_OFFSET + 1), 0.0f);
num_vertices++;
vertices[num_vertices].SetPosition(x + delta_x, y + delta_y);
vertices[num_vertices].SetTextureCoordinates(static_cast<float>(c - CHAR_OFFSET + 1), 1.0f);
num_vertices++;
vertices[num_vertices].SetPosition(x, y);
vertices[num_vertices].SetTextureCoordinates(static_cast<float>(c - CHAR_OFFSET), 0.0f);
num_vertices++;
vertices[num_vertices].SetPosition(x + delta_x, y + delta_y);
vertices[num_vertices].SetTextureCoordinates(static_cast<float>(c - CHAR_OFFSET + 1), 1.0f);
num_vertices++;
vertices[num_vertices].SetPosition(x, y + delta_y);
vertices[num_vertices].SetTextureCoordinates(static_cast<float>(c - CHAR_OFFSET), 1.0f);
num_vertices++;
x += delta_x + border_x;
}
// skip all whitespace strings
if (num_vertices == 0)
return;
draw.CommitVertices(num_vertices);
struct PCBlock
{
float char_size[2];
float offset[2];
float color[4];
} pc_block = {};
pc_block.char_size[0] = 1.0f / static_cast<float>(CHAR_COUNT);
pc_block.char_size[1] = 1.0f;
// shadows
pc_block.offset[0] = 2.0f / bbWidth;
pc_block.offset[1] = -2.0f / bbHeight;
pc_block.color[3] = (color >> 24) / 255.0f;
draw.SetPushConstants(&pc_block, sizeof(pc_block));
draw.SetPSSampler(0, m_texture->GetView(), g_object_cache->GetLinearSampler());
// Setup alpha blending
BlendState blend_state = Util::GetNoBlendingBlendState();
blend_state.blend_enable = VK_TRUE;
blend_state.src_blend = VK_BLEND_FACTOR_SRC_ALPHA;
blend_state.blend_op = VK_BLEND_OP_ADD;
blend_state.dst_blend = VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA;
draw.SetBlendState(blend_state);
draw.Draw();
// non-shadowed part
pc_block.offset[0] = 0.0f;
pc_block.offset[1] = 0.0f;
pc_block.color[0] = ((color >> 16) & 0xFF) / 255.0f;
pc_block.color[1] = ((color >> 8) & 0xFF) / 255.0f;
pc_block.color[2] = (color & 0xFF) / 255.0f;
draw.SetPushConstants(&pc_block, sizeof(pc_block));
draw.Draw();
}
} // namespace Vulkan
@@ -0,0 +1,39 @@
// Copyright 2016 Dolphin Emulator Project
// Licensed under GPLv2+
// Refer to the license.txt file included.
#pragma once
#include <memory>
#include <string>
#include "Common/CommonTypes.h"
#include "VideoBackends/Vulkan/Constants.h"
namespace Vulkan
{
class Texture2D;
class RasterFont
{
public:
RasterFont();
~RasterFont();
bool Initialize();
void PrintMultiLineText(VkRenderPass render_pass, const std::string& text, float start_x,
float start_y, u32 bbWidth, u32 bbHeight, u32 color);
private:
bool CreateTexture();
bool CreateShaders();
std::unique_ptr<Texture2D> m_texture;
VkShaderModule m_vertex_shader = VK_NULL_HANDLE;
VkShaderModule m_fragment_shader = VK_NULL_HANDLE;
};
} // namespace Vulkan

Some files were not shown because too many files have changed in this diff Show More