claude_code/procedural-3d-engine
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Add proper queue family ownership transfer barriers between compute and graphics (for separate queue families)

Browse source 
Refs #544
This commit is contained in:
Sascha Willems 2020-02-22 12:20:30 +01:00
parent fff003315b
commit acdec3e127
1 changed files with 192 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

204
examples/computenbody/computenbody.cpp
View file

@ -49,6 +49,7 @@ public:
// Resources for the graphics part of the example
struct {
uint32_t queueFamilyIndex; // Used to check if compute and graphics queue families differ and require additional barriers
vks::Buffer uniformBuffer; // Contains scene matrices
VkDescriptorSetLayout descriptorSetLayout; // Particle system rendering shader binding layout
VkDescriptorSet descriptorSet; // Particle system rendering shader bindings
@ -64,6 +65,7 @@ public:
// Resources for the compute part of the example
struct {
uint32_t queueFamilyIndex; // Used to check if compute and graphics queue families differ and require additional barriers
vks::Buffer storageBuffer; // (Shader) storage buffer object containing the particles
vks::Buffer uniformBuffer; // Uniform buffer object containing particle system parameters
VkQueue queue; // Separate queue for compute commands (queue family may differ from the one used for graphics)
@ -162,8 +164,33 @@ public:
VK_CHECK_RESULT(vkBeginCommandBuffer(drawCmdBuffers[i], &cmdBufInfo));
// Draw the particle system using the update vertex buffer
// Acquire barrier
if (graphics.queueFamilyIndex != compute.queueFamilyIndex)
{
VkBufferMemoryBarrier buffer_barrier =
{
VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
nullptr,
0,
VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT,
compute.queueFamilyIndex,
graphics.queueFamilyIndex,
compute.storageBuffer.buffer,
0,
compute.storageBuffer.size
};
vkCmdPipelineBarrier(
drawCmdBuffers[i],
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
VK_PIPELINE_STAGE_VERTEX_INPUT_BIT,
0,
0, nullptr,
1, &buffer_barrier,
0, nullptr);
}
// Draw the particle system using the update vertex buffer
vkCmdBeginRenderPass(drawCmdBuffers[i], &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);
VkViewport viewport = vks::initializers::viewport((float)width, (float)height, 0.0f, 1.0f);
@ -173,7 +200,7 @@ public:
vkCmdSetScissor(drawCmdBuffers[i], 0, 1, &scissor);
vkCmdBindPipeline(drawCmdBuffers[i], VK_PIPELINE_BIND_POINT_GRAPHICS, graphics.pipeline);
vkCmdBindDescriptorSets(drawCmdBuffers[i], VK_PIPELINE_BIND_POINT_GRAPHICS, graphics.pipelineLayout, 0, 1, &graphics.descriptorSet, 0, NULL);
vkCmdBindDescriptorSets(drawCmdBuffers[i], VK_PIPELINE_BIND_POINT_GRAPHICS, graphics.pipelineLayout, 0, 1, &graphics.descriptorSet, 0, nullptr);
VkDeviceSize offsets[1] = { 0 };
vkCmdBindVertexBuffers(drawCmdBuffers[i], VERTEX_BUFFER_BIND_ID, 1, &compute.storageBuffer.buffer, offsets);
@ -183,6 +210,32 @@ public:
vkCmdEndRenderPass(drawCmdBuffers[i]);
// Release barrier
if (graphics.queueFamilyIndex != compute.queueFamilyIndex)
{
VkBufferMemoryBarrier buffer_barrier =
{
VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
nullptr,
VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT,
0,
graphics.queueFamilyIndex,
compute.queueFamilyIndex,
compute.storageBuffer.buffer,
0,
compute.storageBuffer.size
};
vkCmdPipelineBarrier(
drawCmdBuffers[i],
VK_PIPELINE_STAGE_VERTEX_INPUT_BIT,
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
0,
0, nullptr,
1, &buffer_barrier,
0, nullptr);
}
VK_CHECK_RESULT(vkEndCommandBuffer(drawCmdBuffers[i]));
}
@ -194,6 +247,32 @@ public:
VK_CHECK_RESULT(vkBeginCommandBuffer(compute.commandBuffer, &cmdBufInfo));
// Acquire barrier
if (graphics.queueFamilyIndex != compute.queueFamilyIndex)
{
VkBufferMemoryBarrier buffer_barrier =
{
VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
nullptr,
0,
VK_ACCESS_SHADER_WRITE_BIT,
graphics.queueFamilyIndex,
compute.queueFamilyIndex,
compute.storageBuffer.buffer,
0,
compute.storageBuffer.size
};
vkCmdPipelineBarrier(
compute.commandBuffer,
VK_PIPELINE_STAGE_VERTEX_INPUT_BIT,
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
0,
0, nullptr,
1, &buffer_barrier,
0, nullptr);
}
// First pass: Calculate particle movement
// -------------------------------------------------------------------------------------------------------
vkCmdBindPipeline(compute.commandBuffer, VK_PIPELINE_BIND_POINT_COMPUTE, compute.pipelineCalculate);
@ -224,6 +303,32 @@ public:
vkCmdBindPipeline(compute.commandBuffer, VK_PIPELINE_BIND_POINT_COMPUTE, compute.pipelineIntegrate);
vkCmdDispatch(compute.commandBuffer, numParticles / 256, 1, 1);
// Release barrier
if (graphics.queueFamilyIndex != compute.queueFamilyIndex)
{
VkBufferMemoryBarrier buffer_barrier =
{
VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
nullptr,
VK_ACCESS_SHADER_WRITE_BIT,
0,
compute.queueFamilyIndex,
graphics.queueFamilyIndex,
compute.storageBuffer.buffer,
0,
compute.storageBuffer.size
};
vkCmdPipelineBarrier(
compute.commandBuffer,
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
VK_PIPELINE_STAGE_VERTEX_INPUT_BIT,
0,
0, nullptr,
1, &buffer_barrier,
0, nullptr);
}
vkEndCommandBuffer(compute.commandBuffer);
}
@ -310,11 +415,36 @@ public:
&compute.storageBuffer,
storageBufferSize);
// Copy to staging buffer
// Copy from staging buffer to storage buffer
VkCommandBuffer copyCmd = VulkanExampleBase::createCommandBuffer(VK_COMMAND_BUFFER_LEVEL_PRIMARY, true);
VkBufferCopy copyRegion = {};
copyRegion.size = storageBufferSize;
vkCmdCopyBuffer(copyCmd, stagingBuffer.buffer, compute.storageBuffer.buffer, 1, &copyRegion);
// Execute a transfer barrier to the compute queue, if necessary
if (graphics.queueFamilyIndex != compute.queueFamilyIndex)
{
VkBufferMemoryBarrier buffer_barrier =
{
VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
nullptr,
VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT,
0,
graphics.queueFamilyIndex,
compute.queueFamilyIndex,
compute.storageBuffer.buffer,
0,
compute.storageBuffer.size
};
vkCmdPipelineBarrier(
copyCmd,
VK_PIPELINE_STAGE_VERTEX_INPUT_BIT,
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
0,
0, nullptr,
1, &buffer_barrier,
0, nullptr);
}
VulkanExampleBase::flushCommandBuffer(copyCmd, queue, true);
stagingBuffer.destroy();
@ -411,7 +541,7 @@ public:
vks::initializers::writeDescriptorSet(graphics.descriptorSet, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, 1, &textures.gradient.descriptor),
vks::initializers::writeDescriptorSet(graphics.descriptorSet, VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER, 2, &graphics.uniformBuffer.descriptor),
};
vkUpdateDescriptorSets(device, static_cast<uint32_t>(writeDescriptorSets.size()), writeDescriptorSets.data(), 0, NULL);
vkUpdateDescriptorSets(device, static_cast<uint32_t>(writeDescriptorSets.size()), writeDescriptorSets.data(), 0, nullptr);
}
void preparePipelines()
@ -520,12 +650,7 @@ public:
// The VulkanDevice::createLogicalDevice functions finds a compute capable queue and prefers queue families that only support compute
// Depending on the implementation this may result in different queue family indices for graphics and computes,
// requiring proper synchronization (see the memory barriers in buildComputeCommandBuffer)
VkDeviceQueueCreateInfo queueCreateInfo = {};
queueCreateInfo.sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO;
queueCreateInfo.pNext = NULL;
queueCreateInfo.queueFamilyIndex = vulkanDevice->queueFamilyIndices.compute;
queueCreateInfo.queueCount = 1;
vkGetDeviceQueue(device, vulkanDevice->queueFamilyIndices.compute, 0, &compute.queue);
vkGetDeviceQueue(device, compute.queueFamilyIndex, 0, &compute.queue);
// Create compute pipeline
// Compute pipelines are created separate from graphics pipelines even if they use the same queue (family index)
@ -581,7 +706,7 @@ public:
&compute.uniformBuffer.descriptor)
};
vkUpdateDescriptorSets(device, static_cast<uint32_t>(computeWriteDescriptorSets.size()), computeWriteDescriptorSets.data(), 0, NULL);
vkUpdateDescriptorSets(device, static_cast<uint32_t>(computeWriteDescriptorSets.size()), computeWriteDescriptorSets.data(), 0, nullptr);
// Create pipelines
VkComputePipelineCreateInfo computePipelineCreateInfo = vks::initializers::computePipelineCreateInfo(compute.pipelineLayout, 0);
@ -622,7 +747,7 @@ public:
// Separate command pool as queue family for compute may be different than graphics
VkCommandPoolCreateInfo cmdPoolInfo = {};
cmdPoolInfo.sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO;
cmdPoolInfo.queueFamilyIndex = vulkanDevice->queueFamilyIndices.compute;
cmdPoolInfo.queueFamilyIndex = compute.queueFamilyIndex;
cmdPoolInfo.flags = VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT;
VK_CHECK_RESULT(vkCreateCommandPool(device, &cmdPoolInfo, nullptr, &compute.commandPool));
@ -648,6 +773,57 @@ public:
// Build a single command buffer containing the compute dispatch commands
buildComputeCommandBuffer();
// If graphics and compute queue family indices differ, acquire and immediately release the storage buffer, so that the initial acquire from the graphics command buffers are matched up properly
if (graphics.queueFamilyIndex != compute.queueFamilyIndex)
{
// Create a transient command buffer for setting up the initial buffer transfer state
VkCommandBuffer transferCmd = vulkanDevice->createCommandBuffer(VK_COMMAND_BUFFER_LEVEL_PRIMARY, compute.commandPool, true);
VkBufferMemoryBarrier acquire_buffer_barrier =
{
VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
nullptr,
0,
VK_ACCESS_SHADER_WRITE_BIT,
graphics.queueFamilyIndex,
compute.queueFamilyIndex,
compute.storageBuffer.buffer,
0,
compute.storageBuffer.size
};
vkCmdPipelineBarrier(
transferCmd,
VK_PIPELINE_STAGE_VERTEX_INPUT_BIT,
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
0,
0, nullptr,
1, &acquire_buffer_barrier,
0, nullptr);
VkBufferMemoryBarrier release_buffer_barrier =
{
VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
nullptr,
VK_ACCESS_SHADER_WRITE_BIT,
0,
compute.queueFamilyIndex,
graphics.queueFamilyIndex,
compute.storageBuffer.buffer,
0,
compute.storageBuffer.size
};
vkCmdPipelineBarrier(
transferCmd,
VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
VK_PIPELINE_STAGE_VERTEX_INPUT_BIT,
0,
0, nullptr,
1, &release_buffer_barrier,
0, nullptr);
vulkanDevice->flushCommandBuffer(transferCmd, compute.queue, compute.commandPool);
}
}
// Prepare and initialize uniform buffer containing shader uniforms
@ -729,6 +905,10 @@ public:
void prepare()
{
VulkanExampleBase::prepare();
// We will be using the queue family indices to check if graphics and compute queue families differ
// If that's the case, we need additional barriers for acquiring and releasing resources
graphics.queueFamilyIndex = vulkanDevice->queueFamilyIndices.graphics;
compute.queueFamilyIndex = vulkanDevice->queueFamilyIndices.compute;
loadAssets();
setupDescriptorPool();
prepareGraphics();