Adding barriers

This commit is contained in:
Dynamitos
2024-07-12 22:05:52 +02:00
parent 8387ad6ffe
commit b756f17a09
5 changed files with 97 additions and 53 deletions
+22 -2
View File
@@ -39,6 +39,7 @@ PFN_vkGetAccelerationStructureBuildSizesKHR getAccelerationStructureBuildSize;
PFN_vkCreateRayTracingPipelinesKHR createRayTracingPipelines;
PFN_vkGetRayTracingShaderGroupHandlesKHR getRayTracingShaderGroupHandles;
PFN_vkCmdTraceRaysKHR cmdTraceRays;
PFN_vkGetAccelerationStructureDeviceAddressKHR getAccelerationStructureDeviceAddress;
void vkCmdDrawMeshTasksEXT(VkCommandBuffer command, uint32 groupX, uint32 groupY, uint32 groupZ) {
cmdDrawMeshTasks(command, groupX, groupY, groupZ);
@@ -92,6 +93,10 @@ void vkCmdTraceRaysKHR(VkCommandBuffer commandBuffer, const VkStridedDeviceAddre
width, height, depth);
}
VkDeviceAddress vkGetAccelerationStructureDeviceAddressKHR(VkDevice device, const VkAccelerationStructureDeviceAddressInfoKHR* pInfo) {
return getAccelerationStructureDeviceAddress(device, pInfo);
}
Graphics::Graphics() : instance(VK_NULL_HANDLE), handle(VK_NULL_HANDLE), physicalDevice(VK_NULL_HANDLE), callback(VK_NULL_HANDLE) {}
Graphics::~Graphics() {
@@ -399,6 +404,13 @@ void Graphics::buildBottomLevelAccelerationStructures(Array<Gfx::PBottomLevelAS>
OBufferAllocation transformBuffer =
new BufferAllocation(this, "TransformBuffer", transformBufferInfo, transformAllocInfo, Gfx::QueueType::GRAPHICS);
transformBuffer->updateContents(0, sizeof(VkTransformMatrixKHR) * matrices.size(), matrices.data());
transformBuffer->pipelineBarrier(VK_ACCESS_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_TRANSFER_READ_BIT,
VK_PIPELINE_STAGE_ACCELERATION_STRUCTURE_BUILD_BIT_KHR | VK_PIPELINE_STAGE_TRANSFER_BIT);
positionBuffer->pipelineBarrier(VK_ACCESS_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_ACCESS_SHADER_READ_BIT,
VK_PIPELINE_STAGE_ACCELERATION_STRUCTURE_BUILD_BIT_KHR);
indexBuffer->pipelineBarrier(VK_ACCESS_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, VK_ACCESS_SHADER_READ_BIT,
VK_PIPELINE_STAGE_ACCELERATION_STRUCTURE_BUILD_BIT_KHR);
Array<VkAccelerationStructureGeometryKHR> geometries(data.size());
Array<VkAccelerationStructureBuildGeometryInfoKHR> buildGeometries(data.size());
@@ -656,7 +668,7 @@ void Graphics::setupDebugCallback() {
.sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_MESSENGER_CREATE_INFO_EXT,
.pNext = nullptr,
.flags = 0,
.messageSeverity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT | VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT | VK_DEBUG_UTILS_MESSAGE_SEVERITY_INFO_BIT_EXT | VK_DEBUG_UTILS_MESSAGE_SEVERITY_VERBOSE_BIT_EXT,
.messageSeverity = VK_DEBUG_UTILS_MESSAGE_SEVERITY_ERROR_BIT_EXT | VK_DEBUG_UTILS_MESSAGE_SEVERITY_WARNING_BIT_EXT,
.messageType = VK_DEBUG_UTILS_MESSAGE_TYPE_GENERAL_BIT_EXT | VK_DEBUG_UTILS_MESSAGE_TYPE_VALIDATION_BIT_EXT |
VK_DEBUG_UTILS_MESSAGE_TYPE_PERFORMANCE_BIT_EXT,
.pfnUserCallback = &debugCallback,
@@ -725,6 +737,7 @@ void Graphics::pickPhysicalDevice() {
.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES,
.pNext = &meshShaderFeatures,
.storageBuffer8BitAccess = true,
.uniformAndStorageBuffer8BitAccess = true,
.shaderUniformBufferArrayNonUniformIndexing = true,
.shaderSampledImageArrayNonUniformIndexing = true,
.shaderStorageBufferArrayNonUniformIndexing = true,
@@ -735,9 +748,14 @@ void Graphics::pickPhysicalDevice() {
.runtimeDescriptorArray = true,
.bufferDeviceAddress = true,
};
features11 = {
.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES,
.pNext = &features12,
.uniformAndStorageBuffer16BitAccess = true,
};
features = {
.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2,
.pNext = &features12,
.pNext = &features11,
.features =
{
.fillModeNonSolid = true,
@@ -883,4 +901,6 @@ void Graphics::createDevice(GraphicsInitializer initializer) {
getRayTracingShaderGroupHandles =
(PFN_vkGetRayTracingShaderGroupHandlesKHR)vkGetDeviceProcAddr(handle, "vkGetRayTracingShaderGroupHandlesKHR");
cmdTraceRays = (PFN_vkCmdTraceRaysKHR)vkGetDeviceProcAddr(handle, "vkCmdTraceRaysKHR");
getAccelerationStructureDeviceAddress =
(PFN_vkGetAccelerationStructureDeviceAddressKHR)vkGetDeviceProcAddr(handle, "vkGetAccelerationStructureDeviceAddressKHR");
}
+1
View File
@@ -113,6 +113,7 @@ class Graphics : public Gfx::Graphics {
VkQueueFamilyProperties graphicsProps;
VkPhysicalDeviceProperties2 props;
VkPhysicalDeviceFeatures2 features;
VkPhysicalDeviceVulkan11Features features11;
VkPhysicalDeviceVulkan12Features features12;
VkPhysicalDeviceMeshShaderFeaturesEXT meshShaderFeatures;
VkPhysicalDeviceAccelerationStructureFeaturesKHR accelerationFeatures;
+42 -43
View File
@@ -610,47 +610,6 @@ PRayTracingPipeline PipelineCache::createPipeline(Gfx::RayTracingPipelineCreateI
std::memcpy(rayGenSbt.data(), sbt.data(), handleSize);
std::memcpy(rayGenSbt.data() + handleSize, createInfo.rayGenGroup.parameters.data(), createInfo.rayGenGroup.parameters.size());
uint64 sbtOffset = handleSizeAligned;
uint32 maxParamSize = 0;
for (auto& hitgroup : createInfo.hitGroups) {
maxParamSize = std::max<uint32>(maxParamSize, hitgroup.parameters.size());
}
uint64 hitStride = align(handleSize + maxParamSize, handleAlignment);
Array<uint8> hitSbt(hitStride * createInfo.hitGroups.size());
for (uint64 i = 0; i < createInfo.hitGroups.size(); ++i) {
std::memcpy(hitSbt.data() + i * hitStride, sbt.data() + sbtOffset, handleSize);
std::memcpy(hitSbt.data() + i * hitStride + handleSize, createInfo.hitGroups[i].parameters.data(),
createInfo.hitGroups[i].parameters.size());
sbtOffset += handleSizeAligned;
}
maxParamSize = 0;
for (auto& missGroup : createInfo.missGroups) {
maxParamSize = std::max<uint32>(maxParamSize, missGroup.parameters.size());
}
uint64 missStride = align(handleSize + maxParamSize, handleAlignment);
Array<uint8> missSbt(missStride * createInfo.missGroups.size());
for (uint64 i = 0; i < createInfo.missGroups.size(); ++i) {
std::memcpy(missSbt.data() + i * missStride, sbt.data() + sbtOffset, handleSize);
std::memcpy(missSbt.data() + i * missStride + handleSize, createInfo.missGroups[i].parameters.data(),
createInfo.missGroups[i].parameters.size());
sbtOffset += handleSizeAligned;
}
maxParamSize = 0;
for (auto& callableGroup : createInfo.callableGroups) {
maxParamSize = std::max<uint32>(maxParamSize, callableGroup.parameters.size());
}
uint64 callableStride = align(handleSize + maxParamSize, handleAlignment);
Array<uint8> callableSbt(callableStride * createInfo.callableGroups.size());
for (uint64 i = 0; i < createInfo.callableGroups.size(); ++i) {
std::memcpy(callableSbt.data() + i * callableStride, sbt.data() + sbtOffset, handleSize);
std::memcpy(callableSbt.data() + i * callableStride + handleSize, createInfo.callableGroups[i].parameters.data(),
createInfo.callableGroups[i].parameters.size());
sbtOffset += handleSizeAligned;
}
OBufferAllocation rayGenBuffer =
new BufferAllocation(graphics, "RayGenSBT",
VkBufferCreateInfo{
@@ -669,6 +628,21 @@ PRayTracingPipeline PipelineCache::createPipeline(Gfx::RayTracingPipelineCreateI
rayGenBuffer->pipelineBarrier(Gfx::SE_ACCESS_TRANSFER_WRITE_BIT, Gfx::SE_PIPELINE_STAGE_TRANSFER_BIT, Gfx::SE_ACCESS_SHADER_READ_BIT,
Gfx::SE_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR);
uint64 sbtOffset = handleSizeAligned;
uint32 maxParamSize = 0;
for (auto& hitgroup : createInfo.hitGroups) {
maxParamSize = std::max<uint32>(maxParamSize, hitgroup.parameters.size());
}
uint64 hitStride = align(handleSize + maxParamSize, handleAlignment);
Array<uint8> hitSbt(hitStride * createInfo.hitGroups.size());
for (uint64 i = 0; i < createInfo.hitGroups.size(); ++i) {
std::memcpy(hitSbt.data() + i * hitStride, sbt.data() + sbtOffset, handleSize);
std::memcpy(hitSbt.data() + i * hitStride + handleSize, createInfo.hitGroups[i].parameters.data(),
createInfo.hitGroups[i].parameters.size());
sbtOffset += handleSizeAligned;
}
OBufferAllocation hitBuffer =
new BufferAllocation(graphics, "HitSBT",
VkBufferCreateInfo{
@@ -685,7 +659,20 @@ PRayTracingPipeline PipelineCache::createPipeline(Gfx::RayTracingPipelineCreateI
Gfx::QueueType::GRAPHICS);
hitBuffer->updateContents(0, hitSbt.size(), hitSbt.data());
hitBuffer->pipelineBarrier(Gfx::SE_ACCESS_TRANSFER_WRITE_BIT, Gfx::SE_PIPELINE_STAGE_TRANSFER_BIT, Gfx::SE_ACCESS_SHADER_READ_BIT,
Gfx::SE_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR);
Gfx::SE_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR);
maxParamSize = 0;
for (auto& missGroup : createInfo.missGroups) {
maxParamSize = std::max<uint32>(maxParamSize, missGroup.parameters.size());
}
uint64 missStride = align(handleSize + maxParamSize, handleAlignment);
Array<uint8> missSbt(missStride * createInfo.missGroups.size());
for (uint64 i = 0; i < createInfo.missGroups.size(); ++i) {
std::memcpy(missSbt.data() + i * missStride, sbt.data() + sbtOffset, handleSize);
std::memcpy(missSbt.data() + i * missStride + handleSize, createInfo.missGroups[i].parameters.data(),
createInfo.missGroups[i].parameters.size());
sbtOffset += handleSizeAligned;
}
OBufferAllocation missBuffer =
new BufferAllocation(graphics, "MissSBT",
@@ -703,8 +690,20 @@ PRayTracingPipeline PipelineCache::createPipeline(Gfx::RayTracingPipelineCreateI
Gfx::QueueType::GRAPHICS);
missBuffer->updateContents(0, missSbt.size(), missSbt.data());
missBuffer->pipelineBarrier(Gfx::SE_ACCESS_TRANSFER_WRITE_BIT, Gfx::SE_PIPELINE_STAGE_TRANSFER_BIT, Gfx::SE_ACCESS_SHADER_READ_BIT,
Gfx::SE_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR);
Gfx::SE_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR);
maxParamSize = 0;
for (auto& callableGroup : createInfo.callableGroups) {
maxParamSize = std::max<uint32>(maxParamSize, callableGroup.parameters.size());
}
uint64 callableStride = align(handleSize + maxParamSize, handleAlignment);
Array<uint8> callableSbt(callableStride * createInfo.callableGroups.size());
for (uint64 i = 0; i < createInfo.callableGroups.size(); ++i) {
std::memcpy(callableSbt.data() + i * callableStride, sbt.data() + sbtOffset, handleSize);
std::memcpy(callableSbt.data() + i * callableStride + handleSize, createInfo.callableGroups[i].parameters.data(),
createInfo.callableGroups[i].parameters.size());
sbtOffset += handleSizeAligned;
}
OBufferAllocation callableBuffer =
new BufferAllocation(graphics, "CallableSBT",
+24 -7
View File
@@ -83,13 +83,16 @@ TopLevelAS::TopLevelAS(PGraphics graphics, const Gfx::TopLevelASCreateInfo& crea
.sType = VK_STRUCTURE_TYPE_ACCELERATION_STRUCTURE_GEOMETRY_KHR,
.pNext = nullptr,
.geometryType = VK_GEOMETRY_TYPE_INSTANCES_KHR,
.geometry = {.instances =
{
.sType = VK_STRUCTURE_TYPE_ACCELERATION_STRUCTURE_GEOMETRY_INSTANCES_DATA_KHR,
.pNext = nullptr,
.arrayOfPointers = VK_FALSE,
.data = instanceDeviceAddress,
}},
.geometry =
{
.instances =
{
.sType = VK_STRUCTURE_TYPE_ACCELERATION_STRUCTURE_GEOMETRY_INSTANCES_DATA_KHR,
.pNext = nullptr,
.arrayOfPointers = VK_FALSE,
.data = instanceDeviceAddress,
},
},
.flags = VK_GEOMETRY_OPAQUE_BIT_KHR,
};
VkAccelerationStructureBuildGeometryInfoKHR structureBuildGeometry = {
@@ -171,7 +174,21 @@ TopLevelAS::TopLevelAS(PGraphics graphics, const Gfx::TopLevelASCreateInfo& crea
VkAccelerationStructureBuildRangeInfoKHR* buildRangeInfos[] = {&buildRange};
auto cmd = graphics->getGraphicsCommands()->getCommands();
vkCmdBuildAccelerationStructuresKHR(cmd->getHandle(), 1, &buildGeometry, buildRangeInfos);
VkBufferMemoryBarrier barrier = {
.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
.pNext = nullptr,
.srcAccessMask = VK_ACCESS_ACCELERATION_STRUCTURE_WRITE_BIT_KHR,
.dstAccessMask = VK_ACCESS_ACCELERATION_STRUCTURE_READ_BIT_KHR,
.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.buffer = buffer->buffer,
.offset = 0,
.size = buffer->size,
};
vkCmdPipelineBarrier(cmd->getHandle(), VK_PIPELINE_STAGE_ACCELERATION_STRUCTURE_BUILD_BIT_KHR, VK_PIPELINE_STAGE_RAY_TRACING_SHADER_BIT_KHR, 0, 0,
nullptr, 1, &barrier, 0, nullptr);
scratchBuffer->bind();
cmd->bindResource(PBufferAllocation(scratchBuffer));
graphics->getDestructionManager()->queueResourceForDestruction(std::move(scratchBuffer));
+8 -1
View File
@@ -12,7 +12,14 @@ class BottomLevelAS : public Gfx::BottomLevelAS {
public:
BottomLevelAS(PGraphics graphics, const Gfx::BottomLevelASCreateInfo& createInfo);
~BottomLevelAS();
uint64 getDeviceAddress() const { return buffer->deviceAddress; }
uint64 getDeviceAddress() const {
VkAccelerationStructureDeviceAddressInfoKHR accelerationStructureDeviceInfo{
.sType = VK_STRUCTURE_TYPE_ACCELERATION_STRUCTURE_DEVICE_ADDRESS_INFO_KHR,
.pNext = nullptr,
.accelerationStructure = handle,
};
return vkGetAccelerationStructureDeviceAddressKHR(graphics->getDevice(), &accelerationStructureDeviceInfo);
}
constexpr VkTransformMatrixKHR getTransform() const { return matrix; }
constexpr uint64 getIndexOffset() const { return indexOffset; }
constexpr uint64 getVertexOffset() const { return vertexOffset; }