Skip to content

Commit

Permalink
[gpu.vulkan] passes that write to the system image will no longer do …
Browse files Browse the repository at this point in the history 
…the system image transition + blit->swapchain straight after their gpu work. this is because if multiple graphics passes are in a graph then they will both try to do it which is wrong. instead at the end of the graph it should do it if any writes to the system image were detected. this is now the behaviour
  • Loading branch information
@harrand
harrand committed Nov 16, 2024
1 parent 73034d4 commit 0ae688a
Showing 1 changed file with 97 additions and 67 deletions.
164 changes: 97 additions & 67 deletions src/tz/gpu/rhi_vulkan.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -206,11 +206,12 @@ namespace tz::gpu
tz::error_code impl_cmd_resource_write(VkCommandBuffer cmds, resource_handle resource, std::span<const std::byte> newdata, std::size_t offset = 0);
void impl_write_all_resources(pass_handle pass);
void impl_write_single_resource(resource_handle resh);
tz::error_code impl_record_gpu_work(pass_handle pass, std::size_t i, std::uint32_t swapchain_image_id, std::span<const pass_handle> deps);
tz::error_code impl_record_gpu_work(pass_handle pass, std::size_t i, std::span<const pass_handle> deps);
void impl_pass_go(pass_handle pass);
void impl_destroy_system_images();
void impl_check_for_resize();
bool impl_graph_will_present(graph_handle graphh);
bool impl_graph_writes_to_system_image(graph_handle graphh);
void impl_execute_subgraph(graph_handle graphh, std::size_t image_index);

/////////////////// tz::gpu api ///////////////////
Expand Down Expand Up @@ -1721,7 +1722,8 @@ namespace tz::gpu
return;
}
const auto& frame = frames[current_frame];
bool will_present = impl_graph_will_present(graphh);
const bool will_present = impl_graph_will_present(graphh);
const bool writes_to_system_image = impl_graph_writes_to_system_image(graphh);

std::uint32_t image_index = -1u;
if(will_present)
Expand Down Expand Up @@ -1783,6 +1785,64 @@ namespace tz::gpu
// do stuff
impl_execute_subgraph(graphh, image_index);

if(writes_to_system_image)
{
tz_assert(image_index != -1u, "ruh roh");
VkImageBlit blit
{
.srcSubresource = VkImageSubresourceLayers
{
.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
.mipLevel = 0,
.baseArrayLayer = 0,
.layerCount = 1
},
.srcOffsets =
{
VkOffset3D{0, 0, 0},
VkOffset3D{static_cast<std::int32_t>(swapchain_width), static_cast<std::int32_t>(swapchain_height), 1},
},
.dstSubresource = VkImageSubresourceLayers
{
.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
.mipLevel = 0,
.baseArrayLayer = 0,
.layerCount = 1
},
.dstOffsets =
{
VkOffset3D{0, 0, 0},
VkOffset3D{static_cast<std::int32_t>(swapchain_width), static_cast<std::int32_t>(swapchain_height), 1},
},
};
// we're about to blit the system image into the swapchain image.
// however we've just rendered into the system image, so its layout must be color_attachment.
// to do the transfer we must first transition the system image to transfer_src.
// we can assume the swapchain image is already in transfer_dst.
VkImageMemoryBarrier system_image_transfer
{
.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
.pNext = nullptr,
.srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
.oldLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
.newLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.image = system_image,
.subresourceRange = VkImageSubresourceRange
{
.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
.baseMipLevel = 0,
.levelCount = 1,
.baseArrayLayer = 0,
.layerCount = 1,
}
};
vkCmdPipelineBarrier(frame.cmds, VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 0, nullptr, 1, &system_image_transfer);
vkCmdBlitImage(frame.cmds, system_image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, swapchain_images[image_index], VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &blit, VK_FILTER_NEAREST);
}

if(will_present)
{
// transition swapchain image to PRESENT
Expand Down Expand Up @@ -2609,9 +2669,9 @@ namespace tz::gpu
}

tz::error_code impl_record_compute_work(const pass_data& pass, const frame_data_t& frame, std::uint32_t id);
tz::error_code impl_record_graphics_work(const pass_data& pass, const frame_data_t& frame, std::uint32_t id, std::uint32_t swapchain_image_id);
tz::error_code impl_record_graphics_work(const pass_data& pass, const frame_data_t& frame, std::uint32_t id);

tz::error_code impl_record_gpu_work(pass_handle passh, std::size_t i, std::uint32_t swapchain_image_id, std::span<const pass_handle> dependencies)
tz::error_code impl_record_gpu_work(pass_handle passh, std::size_t i, std::span<const pass_handle> dependencies)
{
const auto& pass = passes[passh.peek()];
tz::error_code ret = tz::error_code::success;
Expand Down Expand Up @@ -2641,7 +2701,7 @@ namespace tz::gpu
}
else
{
ret = impl_record_graphics_work(pass, frame, i, swapchain_image_id);
ret = impl_record_graphics_work(pass, frame, i);
}
#if TOPAZ_DEBUG
vkCmdEndDebugUtilsLabelEXT(frame.cmds);
Expand All @@ -2659,13 +2719,12 @@ namespace tz::gpu
return tz::error_code::success;
}

tz::error_code impl_record_graphics_work(const pass_data& pass, const frame_data_t& frame, std::uint32_t id, std::uint32_t swapchain_image_id)
tz::error_code impl_record_graphics_work(const pass_data& pass, const frame_data_t& frame, std::uint32_t id)
{
std::vector<VkRenderingAttachmentInfo> colour_attachments;
std::vector<VkImageMemoryBarrier> colour_transitions;
colour_attachments.reserve(pass.info.graphics.colour_targets.size());
colour_transitions.reserve(pass.info.graphics.colour_targets.size());
bool render_into_system_image = false;
VkClearColorValue clear_colour{.float32 = {}};
// BGRA
clear_colour.float32[0] = pass.info.graphics.clear_colour[0];
Expand All @@ -2681,7 +2740,6 @@ namespace tz::gpu
{
rt = system_image;
rtv = system_image_view;
render_into_system_image = true;
}
else
{
Expand Down Expand Up @@ -2898,63 +2956,6 @@ namespace tz::gpu
}
}
vkCmdEndRendering(frame.cmds);
if(render_into_system_image)
{
tz_assert(swapchain_image_id != -1u, "graphics pass targets system image but swapchain image was not acquired.");
VkImageBlit blit
{
.srcSubresource = VkImageSubresourceLayers
{
.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
.mipLevel = 0,
.baseArrayLayer = 0,
.layerCount = 1
},
.srcOffsets =
{
VkOffset3D{0, 0, 0},
VkOffset3D{static_cast<std::int32_t>(swapchain_width), static_cast<std::int32_t>(swapchain_height), 1},
},
.dstSubresource = VkImageSubresourceLayers
{
.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
.mipLevel = 0,
.baseArrayLayer = 0,
.layerCount = 1
},
.dstOffsets =
{
VkOffset3D{0, 0, 0},
VkOffset3D{static_cast<std::int32_t>(swapchain_width), static_cast<std::int32_t>(swapchain_height), 1},
},
};
// we're about to blit the system image into the swapchain image.
// however we've just rendered into the system image, so its layout must be color_attachment.
// to do the transfer we must first transition the system image to transfer_src.
// we can assume the swapchain image is already in transfer_dst.
VkImageMemoryBarrier system_image_transfer
{
.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
.pNext = nullptr,
.srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
.dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
.oldLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
.newLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.image = system_image,
.subresourceRange = VkImageSubresourceRange
{
.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
.baseMipLevel = 0,
.levelCount = 1,
.baseArrayLayer = 0,
.layerCount = 1,
}
};
vkCmdPipelineBarrier(frame.cmds, VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0, 0, nullptr, 0, nullptr, 1, &system_image_transfer);
vkCmdBlitImage(frame.cmds, system_image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, swapchain_images[swapchain_image_id], VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1, &blit, VK_FILTER_NEAREST);
}
return tz::error_code::success;
}

Expand Down Expand Up @@ -3283,6 +3284,36 @@ namespace tz::gpu
return false;
}

bool impl_graph_writes_to_system_image(graph_handle graphh)
{
const auto& graph = graphs[graphh.peek()];
for(const graph_data::entry& entry : graph.timeline)
{
if(entry.is_graph)
{
if(impl_graph_writes_to_system_image(entry.han))
{
return true;
}
}
else
{
if(entry.han == tz::gpu::present_pass)
{
continue;
}
const auto& pass = passes[static_cast<std::size_t>(entry.han)];
for(resource_handle colour_target : pass.colour_targets)
{
if(colour_target == tz::gpu::window_resource)
{
return true;
}
}
}
}
return false;
}

void impl_execute_subgraph(graph_handle graphh, std::size_t image_index)
{
Expand Down Expand Up @@ -3322,14 +3353,13 @@ namespace tz::gpu
tz_assert(!dep_entry.is_graph, "a subgraph cannot be a dependency - only a pass can be a dependency");
dep_passes.push_back(dep_entry.han);
}
tz_must(impl_record_gpu_work(entry.han, current_frame, image_index, dep_passes));
tz_must(impl_record_gpu_work(entry.han, current_frame, dep_passes));
}
}

#if TOPAZ_DEBUG
vkCmdEndDebugUtilsLabelEXT(frames[current_frame].cmds);
#endif

}
}
#endif

0 comments on commit 0ae688a

Please sign in to comment.