summaryrefslogtreecommitdiffstats
path: root/src/video_core/renderer_vulkan
diff options
context:
space:
mode:
Diffstat (limited to 'src/video_core/renderer_vulkan')
-rw-r--r--src/video_core/renderer_vulkan/renderer_vulkan.cpp3
-rw-r--r--src/video_core/renderer_vulkan/vk_blit_screen.cpp6
-rw-r--r--src/video_core/renderer_vulkan/vk_descriptor_pool.cpp1
-rw-r--r--src/video_core/renderer_vulkan/vk_pipeline_cache.cpp2
-rw-r--r--src/video_core/renderer_vulkan/vk_rasterizer.cpp25
-rw-r--r--src/video_core/renderer_vulkan/vk_scheduler.cpp40
-rw-r--r--src/video_core/renderer_vulkan/vk_scheduler.h6
-rw-r--r--src/video_core/renderer_vulkan/vk_state_tracker.h4
-rw-r--r--src/video_core/renderer_vulkan/vk_swapchain.cpp30
-rw-r--r--src/video_core/renderer_vulkan/vk_swapchain.h10
-rw-r--r--src/video_core/renderer_vulkan/vk_texture_cache.cpp5
-rw-r--r--src/video_core/renderer_vulkan/vk_texture_cache.h14
12 files changed, 92 insertions, 54 deletions
diff --git a/src/video_core/renderer_vulkan/renderer_vulkan.cpp b/src/video_core/renderer_vulkan/renderer_vulkan.cpp
index 7c9b0d6db..9ff0a28cd 100644
--- a/src/video_core/renderer_vulkan/renderer_vulkan.cpp
+++ b/src/video_core/renderer_vulkan/renderer_vulkan.cpp
@@ -164,7 +164,8 @@ void RendererVulkan::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
blit_screen.Recreate();
}
const VkSemaphore render_semaphore = blit_screen.DrawToSwapchain(*framebuffer, use_accelerated);
- scheduler.Flush(render_semaphore);
+ const VkSemaphore present_semaphore = swapchain.CurrentPresentSemaphore();
+ scheduler.Flush(render_semaphore, present_semaphore);
scheduler.WaitWorker();
swapchain.Present(render_semaphore);
diff --git a/src/video_core/renderer_vulkan/vk_blit_screen.cpp b/src/video_core/renderer_vulkan/vk_blit_screen.cpp
index 5c43b8acf..888bc7392 100644
--- a/src/video_core/renderer_vulkan/vk_blit_screen.cpp
+++ b/src/video_core/renderer_vulkan/vk_blit_screen.cpp
@@ -159,11 +159,13 @@ VkSemaphore VKBlitScreen::Draw(const Tegra::FramebufferConfig& framebuffer,
const VAddr framebuffer_addr = framebuffer.address + framebuffer.offset;
const u8* const host_ptr = cpu_memory.GetPointer(framebuffer_addr);
- const size_t size_bytes = GetSizeInBytes(framebuffer);
// TODO(Rodrigo): Read this from HLE
constexpr u32 block_height_log2 = 4;
const u32 bytes_per_pixel = GetBytesPerPixel(framebuffer);
+ const u64 size_bytes{Tegra::Texture::CalculateSize(true, bytes_per_pixel,
+ framebuffer.stride, framebuffer.height,
+ 1, block_height_log2, 0)};
Tegra::Texture::UnswizzleTexture(
mapped_span.subspan(image_offset, size_bytes), std::span(host_ptr, size_bytes),
bytes_per_pixel, framebuffer.width, framebuffer.height, 1, block_height_log2, 0);
@@ -356,7 +358,7 @@ void VKBlitScreen::CreateDescriptorPool() {
void VKBlitScreen::CreateRenderPass() {
const VkAttachmentDescription color_attachment{
.flags = 0,
- .format = swapchain.GetImageFormat(),
+ .format = swapchain.GetImageViewFormat(),
.samples = VK_SAMPLE_COUNT_1_BIT,
.loadOp = VK_ATTACHMENT_LOAD_OP_CLEAR,
.storeOp = VK_ATTACHMENT_STORE_OP_STORE,
diff --git a/src/video_core/renderer_vulkan/vk_descriptor_pool.cpp b/src/video_core/renderer_vulkan/vk_descriptor_pool.cpp
index 60e45f1b9..d87da2a34 100644
--- a/src/video_core/renderer_vulkan/vk_descriptor_pool.cpp
+++ b/src/video_core/renderer_vulkan/vk_descriptor_pool.cpp
@@ -2,6 +2,7 @@
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
+#include <algorithm>
#include <mutex>
#include <span>
#include <vector>
diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
index f316c4f92..31bfbcb06 100644
--- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
@@ -325,6 +325,8 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, Tegra::Engines::Maxw
host_info = Shader::HostTranslateInfo{
.support_float16 = device.IsFloat16Supported(),
.support_int64 = device.IsShaderInt64Supported(),
+ .needs_demote_reorder = driver_id == VK_DRIVER_ID_AMD_PROPRIETARY_KHR ||
+ driver_id == VK_DRIVER_ID_AMD_OPEN_SOURCE_KHR,
};
}
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 3ac18ea54..3bcd6d6cc 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -228,9 +228,7 @@ void RasterizerVulkan::Clear() {
};
const u32 color_attachment = regs.clear_buffers.RT;
- const auto attachment_aspect_mask = framebuffer->ImageRanges()[color_attachment].aspectMask;
- const bool is_color_rt = (attachment_aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT) != 0;
- if (use_color && is_color_rt) {
+ if (use_color && framebuffer->HasAspectColorBit(color_attachment)) {
VkClearValue clear_value;
std::memcpy(clear_value.color.float32, regs.clear_color, sizeof(regs.clear_color));
@@ -248,12 +246,15 @@ void RasterizerVulkan::Clear() {
return;
}
VkImageAspectFlags aspect_flags = 0;
- if (use_depth) {
+ if (use_depth && framebuffer->HasAspectDepthBit()) {
aspect_flags |= VK_IMAGE_ASPECT_DEPTH_BIT;
}
- if (use_stencil) {
+ if (use_stencil && framebuffer->HasAspectStencilBit()) {
aspect_flags |= VK_IMAGE_ASPECT_STENCIL_BIT;
}
+ if (aspect_flags == 0) {
+ return;
+ }
scheduler.Record([clear_depth = regs.clear_depth, clear_stencil = regs.clear_stencil,
clear_rect, aspect_flags](vk::CommandBuffer cmdbuf) {
VkClearAttachment attachment;
@@ -764,12 +765,7 @@ void RasterizerVulkan::UpdateStencilOp(Tegra::Engines::Maxwell3D::Regs& regs) {
const Maxwell::StencilOp zpass = regs.stencil_front_op_zpass;
const Maxwell::ComparisonOp compare = regs.stencil_front_func_func;
if (regs.stencil_two_side_enable) {
- scheduler.Record([fail, zfail, zpass, compare](vk::CommandBuffer cmdbuf) {
- cmdbuf.SetStencilOpEXT(VK_STENCIL_FACE_FRONT_AND_BACK, MaxwellToVK::StencilOp(fail),
- MaxwellToVK::StencilOp(zpass), MaxwellToVK::StencilOp(zfail),
- MaxwellToVK::ComparisonOp(compare));
- });
- } else {
+ // Separate stencil op per face
const Maxwell::StencilOp back_fail = regs.stencil_back_op_fail;
const Maxwell::StencilOp back_zfail = regs.stencil_back_op_zfail;
const Maxwell::StencilOp back_zpass = regs.stencil_back_op_zpass;
@@ -784,6 +780,13 @@ void RasterizerVulkan::UpdateStencilOp(Tegra::Engines::Maxwell3D::Regs& regs) {
MaxwellToVK::StencilOp(back_zfail),
MaxwellToVK::ComparisonOp(back_compare));
});
+ } else {
+ // Front face defines the stencil op of both faces
+ scheduler.Record([fail, zfail, zpass, compare](vk::CommandBuffer cmdbuf) {
+ cmdbuf.SetStencilOpEXT(VK_STENCIL_FACE_FRONT_AND_BACK, MaxwellToVK::StencilOp(fail),
+ MaxwellToVK::StencilOp(zpass), MaxwellToVK::StencilOp(zfail),
+ MaxwellToVK::ComparisonOp(compare));
+ });
}
}
diff --git a/src/video_core/renderer_vulkan/vk_scheduler.cpp b/src/video_core/renderer_vulkan/vk_scheduler.cpp
index 4840962de..1d438787a 100644
--- a/src/video_core/renderer_vulkan/vk_scheduler.cpp
+++ b/src/video_core/renderer_vulkan/vk_scheduler.cpp
@@ -55,14 +55,14 @@ VKScheduler::~VKScheduler() {
worker_thread.join();
}
-void VKScheduler::Flush(VkSemaphore semaphore) {
- SubmitExecution(semaphore);
+void VKScheduler::Flush(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
+ SubmitExecution(signal_semaphore, wait_semaphore);
AllocateNewContext();
}
-void VKScheduler::Finish(VkSemaphore semaphore) {
+void VKScheduler::Finish(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
const u64 presubmit_tick = CurrentTick();
- SubmitExecution(semaphore);
+ SubmitExecution(signal_semaphore, wait_semaphore);
WaitWorker();
Wait(presubmit_tick);
AllocateNewContext();
@@ -171,37 +171,41 @@ void VKScheduler::AllocateWorkerCommandBuffer() {
});
}
-void VKScheduler::SubmitExecution(VkSemaphore semaphore) {
+void VKScheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
EndPendingOperations();
InvalidateState();
const u64 signal_value = master_semaphore->NextTick();
- Record([semaphore, signal_value, this](vk::CommandBuffer cmdbuf) {
+ Record([signal_semaphore, wait_semaphore, signal_value, this](vk::CommandBuffer cmdbuf) {
cmdbuf.End();
-
- const u32 num_signal_semaphores = semaphore ? 2U : 1U;
-
- const u64 wait_value = signal_value - 1;
- const VkPipelineStageFlags wait_stage_mask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
-
const VkSemaphore timeline_semaphore = master_semaphore->Handle();
+
+ const u32 num_signal_semaphores = signal_semaphore ? 2U : 1U;
const std::array signal_values{signal_value, u64(0)};
- const std::array signal_semaphores{timeline_semaphore, semaphore};
+ const std::array signal_semaphores{timeline_semaphore, signal_semaphore};
+
+ const u32 num_wait_semaphores = wait_semaphore ? 2U : 1U;
+ const std::array wait_values{signal_value - 1, u64(1)};
+ const std::array wait_semaphores{timeline_semaphore, wait_semaphore};
+ static constexpr std::array<VkPipelineStageFlags, 2> wait_stage_masks{
+ VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+ VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+ };
const VkTimelineSemaphoreSubmitInfoKHR timeline_si{
.sType = VK_STRUCTURE_TYPE_TIMELINE_SEMAPHORE_SUBMIT_INFO_KHR,
.pNext = nullptr,
- .waitSemaphoreValueCount = 1,
- .pWaitSemaphoreValues = &wait_value,
+ .waitSemaphoreValueCount = num_wait_semaphores,
+ .pWaitSemaphoreValues = wait_values.data(),
.signalSemaphoreValueCount = num_signal_semaphores,
.pSignalSemaphoreValues = signal_values.data(),
};
const VkSubmitInfo submit_info{
.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
.pNext = &timeline_si,
- .waitSemaphoreCount = 1,
- .pWaitSemaphores = &timeline_semaphore,
- .pWaitDstStageMask = &wait_stage_mask,
+ .waitSemaphoreCount = num_wait_semaphores,
+ .pWaitSemaphores = wait_semaphores.data(),
+ .pWaitDstStageMask = wait_stage_masks.data(),
.commandBufferCount = 1,
.pCommandBuffers = cmdbuf.address(),
.signalSemaphoreCount = num_signal_semaphores,
diff --git a/src/video_core/renderer_vulkan/vk_scheduler.h b/src/video_core/renderer_vulkan/vk_scheduler.h
index cf39a2363..759ed5a48 100644
--- a/src/video_core/renderer_vulkan/vk_scheduler.h
+++ b/src/video_core/renderer_vulkan/vk_scheduler.h
@@ -34,10 +34,10 @@ public:
~VKScheduler();
/// Sends the current execution context to the GPU.
- void Flush(VkSemaphore semaphore = nullptr);
+ void Flush(VkSemaphore signal_semaphore = nullptr, VkSemaphore wait_semaphore = nullptr);
/// Sends the current execution context to the GPU and waits for it to complete.
- void Finish(VkSemaphore semaphore = nullptr);
+ void Finish(VkSemaphore signal_semaphore = nullptr, VkSemaphore wait_semaphore = nullptr);
/// Waits for the worker thread to finish executing everything. After this function returns it's
/// safe to touch worker resources.
@@ -191,7 +191,7 @@ private:
void AllocateWorkerCommandBuffer();
- void SubmitExecution(VkSemaphore semaphore);
+ void SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore);
void AllocateNewContext();
diff --git a/src/video_core/renderer_vulkan/vk_state_tracker.h b/src/video_core/renderer_vulkan/vk_state_tracker.h
index 5f78f6950..d90935f52 100644
--- a/src/video_core/renderer_vulkan/vk_state_tracker.h
+++ b/src/video_core/renderer_vulkan/vk_state_tracker.h
@@ -110,10 +110,6 @@ public:
return Exchange(Dirty::DepthTestEnable, false);
}
- bool TouchDepthBoundsEnable() {
- return Exchange(Dirty::DepthBoundsEnable, false);
- }
-
bool TouchDepthWriteEnable() {
return Exchange(Dirty::DepthWriteEnable, false);
}
diff --git a/src/video_core/renderer_vulkan/vk_swapchain.cpp b/src/video_core/renderer_vulkan/vk_swapchain.cpp
index d990eefba..aadf03cb0 100644
--- a/src/video_core/renderer_vulkan/vk_swapchain.cpp
+++ b/src/video_core/renderer_vulkan/vk_swapchain.cpp
@@ -20,16 +20,15 @@ namespace Vulkan {
namespace {
-VkSurfaceFormatKHR ChooseSwapSurfaceFormat(vk::Span<VkSurfaceFormatKHR> formats, bool srgb) {
+VkSurfaceFormatKHR ChooseSwapSurfaceFormat(vk::Span<VkSurfaceFormatKHR> formats) {
if (formats.size() == 1 && formats[0].format == VK_FORMAT_UNDEFINED) {
VkSurfaceFormatKHR format;
format.format = VK_FORMAT_B8G8R8A8_UNORM;
format.colorSpace = VK_COLOR_SPACE_SRGB_NONLINEAR_KHR;
return format;
}
- const auto& found = std::find_if(formats.begin(), formats.end(), [srgb](const auto& format) {
- const auto request_format = srgb ? VK_FORMAT_B8G8R8A8_SRGB : VK_FORMAT_B8G8R8A8_UNORM;
- return format.format == request_format &&
+ const auto& found = std::find_if(formats.begin(), formats.end(), [](const auto& format) {
+ return format.format == VK_FORMAT_B8G8R8A8_UNORM &&
format.colorSpace == VK_COLOR_SPACE_SRGB_NONLINEAR_KHR;
});
return found != formats.end() ? *found : formats[0];
@@ -107,14 +106,12 @@ void VKSwapchain::AcquireNextImage() {
}
void VKSwapchain::Present(VkSemaphore render_semaphore) {
- const VkSemaphore present_semaphore{*present_semaphores[frame_index]};
- const std::array<VkSemaphore, 2> semaphores{present_semaphore, render_semaphore};
const auto present_queue{device.GetPresentQueue()};
const VkPresentInfoKHR present_info{
.sType = VK_STRUCTURE_TYPE_PRESENT_INFO_KHR,
.pNext = nullptr,
- .waitSemaphoreCount = render_semaphore ? 2U : 1U,
- .pWaitSemaphores = semaphores.data(),
+ .waitSemaphoreCount = render_semaphore ? 1U : 0U,
+ .pWaitSemaphores = &render_semaphore,
.swapchainCount = 1,
.pSwapchains = swapchain.address(),
.pImageIndices = &image_index,
@@ -145,7 +142,7 @@ void VKSwapchain::CreateSwapchain(const VkSurfaceCapabilitiesKHR& capabilities,
const auto formats{physical_device.GetSurfaceFormatsKHR(surface)};
const auto present_modes{physical_device.GetSurfacePresentModesKHR(surface)};
- const VkSurfaceFormatKHR surface_format{ChooseSwapSurfaceFormat(formats, srgb)};
+ const VkSurfaceFormatKHR surface_format{ChooseSwapSurfaceFormat(formats)};
const VkPresentModeKHR present_mode{ChooseSwapPresentMode(present_modes)};
u32 requested_image_count{capabilities.minImageCount + 1};
@@ -180,6 +177,17 @@ void VKSwapchain::CreateSwapchain(const VkSurfaceCapabilitiesKHR& capabilities,
swapchain_ci.queueFamilyIndexCount = static_cast<u32>(queue_indices.size());
swapchain_ci.pQueueFamilyIndices = queue_indices.data();
}
+ static constexpr std::array view_formats{VK_FORMAT_B8G8R8A8_UNORM, VK_FORMAT_B8G8R8A8_SRGB};
+ VkImageFormatListCreateInfo format_list{
+ .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_LIST_CREATE_INFO_KHR,
+ .pNext = nullptr,
+ .viewFormatCount = static_cast<u32>(view_formats.size()),
+ .pViewFormats = view_formats.data(),
+ };
+ if (device.IsKhrSwapchainMutableFormatEnabled()) {
+ format_list.pNext = std::exchange(swapchain_ci.pNext, &format_list);
+ swapchain_ci.flags |= VK_SWAPCHAIN_CREATE_MUTABLE_FORMAT_BIT_KHR;
+ }
// Request the size again to reduce the possibility of a TOCTOU race condition.
const auto updated_capabilities = physical_device.GetSurfaceCapabilitiesKHR(surface);
swapchain_ci.imageExtent = ChooseSwapExtent(updated_capabilities, width, height);
@@ -191,7 +199,7 @@ void VKSwapchain::CreateSwapchain(const VkSurfaceCapabilitiesKHR& capabilities,
images = swapchain.GetImages();
image_count = static_cast<u32>(images.size());
- image_format = surface_format.format;
+ image_view_format = srgb ? VK_FORMAT_B8G8R8A8_SRGB : VK_FORMAT_B8G8R8A8_UNORM;
}
void VKSwapchain::CreateSemaphores() {
@@ -207,7 +215,7 @@ void VKSwapchain::CreateImageViews() {
.flags = 0,
.image = {},
.viewType = VK_IMAGE_VIEW_TYPE_2D,
- .format = image_format,
+ .format = image_view_format,
.components =
{
.r = VK_COMPONENT_SWIZZLE_IDENTITY,
diff --git a/src/video_core/renderer_vulkan/vk_swapchain.h b/src/video_core/renderer_vulkan/vk_swapchain.h
index 35c2cdc14..5bce41e21 100644
--- a/src/video_core/renderer_vulkan/vk_swapchain.h
+++ b/src/video_core/renderer_vulkan/vk_swapchain.h
@@ -68,8 +68,12 @@ public:
return *image_views[index];
}
- VkFormat GetImageFormat() const {
- return image_format;
+ VkFormat GetImageViewFormat() const {
+ return image_view_format;
+ }
+
+ VkSemaphore CurrentPresentSemaphore() const {
+ return *present_semaphores[frame_index];
}
private:
@@ -96,7 +100,7 @@ private:
u32 image_index{};
u32 frame_index{};
- VkFormat image_format{};
+ VkFormat image_view_format{};
VkExtent2D extent{};
bool current_srgb{};
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.cpp b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
index 8f4df7122..ff979a7ac 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
@@ -1186,9 +1186,12 @@ Framebuffer::Framebuffer(TextureCacheRuntime& runtime, std::span<ImageView*, NUM
renderpass_key.depth_format = depth_buffer->format;
num_layers = std::max(num_layers, depth_buffer->range.extent.layers);
images[num_images] = depth_buffer->ImageHandle();
- image_ranges[num_images] = MakeSubresourceRange(depth_buffer);
+ const VkImageSubresourceRange subresource_range = MakeSubresourceRange(depth_buffer);
+ image_ranges[num_images] = subresource_range;
samples = depth_buffer->Samples();
++num_images;
+ has_depth = (subresource_range.aspectMask & VK_IMAGE_ASPECT_DEPTH_BIT) != 0;
+ has_stencil = (subresource_range.aspectMask & VK_IMAGE_ASPECT_STENCIL_BIT) != 0;
} else {
renderpass_key.depth_format = PixelFormat::Invalid;
}
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.h b/src/video_core/renderer_vulkan/vk_texture_cache.h
index 5fe6b7ba3..6d5a68bfe 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.h
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.h
@@ -232,6 +232,18 @@ public:
return image_ranges;
}
+ [[nodiscard]] bool HasAspectColorBit(size_t index) const noexcept {
+ return (image_ranges.at(index).aspectMask & VK_IMAGE_ASPECT_COLOR_BIT) != 0;
+ }
+
+ [[nodiscard]] bool HasAspectDepthBit() const noexcept {
+ return has_depth;
+ }
+
+ [[nodiscard]] bool HasAspectStencilBit() const noexcept {
+ return has_stencil;
+ }
+
private:
vk::Framebuffer framebuffer;
VkRenderPass renderpass{};
@@ -241,6 +253,8 @@ private:
u32 num_images = 0;
std::array<VkImage, 9> images{};
std::array<VkImageSubresourceRange, 9> image_ranges{};
+ bool has_depth{};
+ bool has_stencil{};
};
struct TextureCacheParams {
generated by cgit v1.2.3 (git 2.25.1) at 2024-07-02 02:50:24 +0200