Merge pull request #5589 from stenzek/vulkan-validation-fixes

Vulkan: Don't transition image layouts inside render passes
This commit is contained in:
shuffle2 2017-06-11 16:59:02 -07:00 committed by GitHub
commit 2b86cf02a8
5 changed files with 81 additions and 46 deletions

View File

@ -329,6 +329,10 @@ void Renderer::BeginFrame()
// Activate a new command list, and restore state ready for the next draw
g_command_buffer_mgr->ActivateCommandBuffer();
// Restore the EFB color texture to color attachment ready for rendering the next frame.
FramebufferManager::GetInstance()->GetEFBColorTexture()->TransitionToLayout(
g_command_buffer_mgr->GetCurrentCommandBuffer(), VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL);
// Ensure that the state tracker rebinds everything, and allocates a new set
// of descriptors out of the next pool.
StateTracker::GetInstance()->InvalidateDescriptorSets();
@ -516,8 +520,7 @@ void Renderer::SwapImpl(u32 xfb_addr, u32 fb_width, u32 fb_stride, u32 fb_height
// If MSAA is enabled, and we're not using XFB, we need to resolve the EFB framebuffer before
// rendering the final image to the screen, or dumping the frame. This is because we can't resolve
// an image within a render pass, which will have already started by the time it is used.
if (g_ActiveConfig.iMultisamples > 1 && !g_ActiveConfig.bUseXFB)
ResolveEFBForSwap(scaled_efb_rect);
TransitionBuffersForSwap(scaled_efb_rect, xfb_sources, xfb_count);
// Render the frame dump image if enabled.
if (IsFrameDumping())
@ -588,16 +591,44 @@ void Renderer::SwapImpl(u32 xfb_addr, u32 fb_width, u32 fb_stride, u32 fb_height
TextureCache::GetInstance()->Cleanup(frameCount);
}
void Renderer::ResolveEFBForSwap(const TargetRectangle& scaled_rect)
void Renderer::TransitionBuffersForSwap(const TargetRectangle& scaled_rect,
const XFBSourceBase* const* xfb_sources, u32 xfb_count)
{
// While the source rect can be out-of-range when drawing, the resolve rectangle must be within
// the bounds of the texture.
VkRect2D region = {
{scaled_rect.left, scaled_rect.top},
{static_cast<u32>(scaled_rect.GetWidth()), static_cast<u32>(scaled_rect.GetHeight())}};
region = Util::ClampRect2D(region, FramebufferManager::GetInstance()->GetEFBWidth(),
FramebufferManager::GetInstance()->GetEFBHeight());
FramebufferManager::GetInstance()->ResolveEFBColorTexture(region);
VkCommandBuffer command_buffer = g_command_buffer_mgr->GetCurrentCommandBuffer();
if (!g_ActiveConfig.bUseXFB)
{
// Drawing EFB direct.
if (g_ActiveConfig.iMultisamples > 1)
{
// While the source rect can be out-of-range when drawing, the resolve rectangle must be
// within the bounds of the texture.
VkRect2D region = {
{scaled_rect.left, scaled_rect.top},
{static_cast<u32>(scaled_rect.GetWidth()), static_cast<u32>(scaled_rect.GetHeight())}};
region = Util::ClampRect2D(region, FramebufferManager::GetInstance()->GetEFBWidth(),
FramebufferManager::GetInstance()->GetEFBHeight());
Vulkan::Texture2D* rtex = FramebufferManager::GetInstance()->ResolveEFBColorTexture(region);
rtex->TransitionToLayout(command_buffer, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
}
else
{
FramebufferManager::GetInstance()->GetEFBColorTexture()->TransitionToLayout(
command_buffer, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
}
return;
}
// Drawing XFB sources, so transition all of them.
// Don't need the EFB, so leave it as-is.
for (u32 i = 0; i < xfb_count; i++)
{
const XFBSource* xfb_source = static_cast<const XFBSource*>(xfb_sources[i]);
xfb_source->GetTexture()->GetTexture()->TransitionToLayout(
command_buffer, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
}
}
void Renderer::DrawFrame(VkRenderPass render_pass, const TargetRectangle& target_rect,
@ -622,18 +653,9 @@ void Renderer::DrawEFB(VkRenderPass render_pass, const TargetRectangle& target_r
g_ActiveConfig.iMultisamples > 1 ?
FramebufferManager::GetInstance()->GetResolvedEFBColorTexture() :
FramebufferManager::GetInstance()->GetEFBColorTexture();
efb_color_texture->TransitionToLayout(g_command_buffer_mgr->GetCurrentCommandBuffer(),
VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
// Copy EFB -> backbuffer
BlitScreen(render_pass, target_rect, scaled_efb_rect, efb_color_texture);
// Restore the EFB color texture to color attachment ready for rendering the next frame.
if (efb_color_texture == FramebufferManager::GetInstance()->GetEFBColorTexture())
{
FramebufferManager::GetInstance()->GetEFBColorTexture()->TransitionToLayout(
g_command_buffer_mgr->GetCurrentCommandBuffer(), VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL);
}
}
void Renderer::DrawVirtualXFB(VkRenderPass render_pass, const TargetRectangle& target_rect,
@ -643,9 +665,6 @@ void Renderer::DrawVirtualXFB(VkRenderPass render_pass, const TargetRectangle& t
for (u32 i = 0; i < xfb_count; ++i)
{
const XFBSource* xfb_source = static_cast<const XFBSource*>(xfb_sources[i]);
xfb_source->GetTexture()->GetTexture()->TransitionToLayout(
g_command_buffer_mgr->GetCurrentCommandBuffer(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
TargetRectangle source_rect = xfb_source->sourceRc;
TargetRectangle draw_rect;
@ -679,9 +698,6 @@ void Renderer::DrawRealXFB(VkRenderPass render_pass, const TargetRectangle& targ
for (u32 i = 0; i < xfb_count; ++i)
{
const XFBSource* xfb_source = static_cast<const XFBSource*>(xfb_sources[i]);
xfb_source->GetTexture()->GetTexture()->TransitionToLayout(
g_command_buffer_mgr->GetCurrentCommandBuffer(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
TargetRectangle source_rect = xfb_source->sourceRc;
TargetRectangle draw_rect = target_rect;
source_rect.right -= fb_stride - fb_width;

View File

@ -87,7 +87,10 @@ private:
bool CompileShaders();
void DestroyShaders();
void ResolveEFBForSwap(const TargetRectangle& scaled_rect);
// Transitions EFB/XFB buffers to SHADER_READ_ONLY, ready for presenting/dumping.
// If MSAA is enabled, and XFB is disabled, also resolves the EFB buffer.
void TransitionBuffersForSwap(const TargetRectangle& scaled_rect,
const XFBSourceBase* const* xfb_sources, u32 xfb_count);
// Draw either the EFB, or specified XFB sources to the currently-bound framebuffer.
void DrawFrame(VkRenderPass render_pass, const TargetRectangle& target_rect,

View File

@ -152,9 +152,19 @@ void TextureCache::DecodeTextureOnGPU(TCacheEntryBase* entry, u32 dst_level, con
u32 aligned_width, u32 aligned_height, u32 row_stride,
const u8* palette, TlutFormat palette_format)
{
m_texture_converter->DecodeTexture(static_cast<TCacheEntry*>(entry), dst_level, data, data_size,
format, width, height, aligned_width, aligned_height,
row_stride, palette, palette_format);
// Group compute shader dispatches together in the init command buffer. That way we don't have to
// pay a penalty for switching from graphics->compute, or end/restart our render pass.
VkCommandBuffer command_buffer = g_command_buffer_mgr->GetCurrentInitCommandBuffer();
m_texture_converter->DecodeTexture(command_buffer, static_cast<TCacheEntry*>(entry), dst_level,
data, data_size, format, width, height, aligned_width,
aligned_height, row_stride, palette, palette_format);
// Last mip level? Ensure the texture is ready for use.
if (dst_level == (entry->config.levels - 1))
{
static_cast<TCacheEntry*>(entry)->GetTexture()->TransitionToLayout(
command_buffer, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
}
}
void TextureCache::CopyTextureRectangle(TCacheEntry* dst_texture,
@ -207,7 +217,8 @@ void TextureCache::ScaleTextureRectangle(TCacheEntry* dst_texture,
_assert_msg_(VIDEO, dst_texture->config.rendertarget,
"Destination texture for partial copy is not a rendertarget");
// Render pass expects dst_texture to be in SHADER_READ_ONLY state.
// Render pass expects dst_texture to be in COLOR_ATTACHMENT_OPTIMAL state.
// src_texture should already be in SHADER_READ_ONLY state, but transition in case (XFB).
src_texture->TransitionToLayout(g_command_buffer_mgr->GetCurrentCommandBuffer(),
VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
dst_texture->GetTexture()->TransitionToLayout(g_command_buffer_mgr->GetCurrentCommandBuffer(),
@ -363,8 +374,9 @@ void TextureCache::TCacheEntry::Load(u32 level, u32 width, u32 height, u32 row_l
// should insert an explicit pipeline barrier just in case (done by TransitionToLayout).
//
// We transition to TRANSFER_DST, ready for the image copy, and leave the texture in this state.
// This is so that the remaining mip levels can be uploaded without barriers, and then when the
// texture is used, it can be transitioned to SHADER_READ_ONLY (see TCacheEntry::Bind).
// When the last mip level is uploaded, we transition to SHADER_READ_ONLY, ready for use. This is
// because we can't transition in a render pass, and we don't necessarily know when this texture
// is going to be used.
m_texture->TransitionToLayout(g_command_buffer_mgr->GetCurrentInitCommandBuffer(),
VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL);
@ -430,6 +442,13 @@ void TextureCache::TCacheEntry::Load(u32 level, u32 width, u32 height, u32 row_l
vkCmdCopyBufferToImage(g_command_buffer_mgr->GetCurrentInitCommandBuffer(), upload_buffer,
m_texture->GetImage(), VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 1,
&image_copy);
// Last mip level? We shouldn't be doing any further uploads now, so transition for rendering.
if (level == (config.levels - 1))
{
m_texture->TransitionToLayout(g_command_buffer_mgr->GetCurrentInitCommandBuffer(),
VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
}
}
void TextureCache::TCacheEntry::FromRenderTarget(bool is_depth_copy, const EFBRectangle& src_rect,
@ -497,7 +516,7 @@ void TextureCache::TCacheEntry::FromRenderTarget(bool is_depth_copy, const EFBRe
// Transition the EFB back to its original layout.
src_texture->TransitionToLayout(command_buffer, original_layout);
// Render pass transitions texture to SHADER_READ_ONLY.
// Ensure texture is in SHADER_READ_ONLY layout, ready for usage.
m_texture->TransitionToLayout(command_buffer, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
}
@ -518,8 +537,9 @@ void TextureCache::TCacheEntry::CopyRectangleFromTexture(const TCacheEntryBase*
void TextureCache::TCacheEntry::Bind(unsigned int stage)
{
m_texture->TransitionToLayout(g_command_buffer_mgr->GetCurrentCommandBuffer(),
VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
// Texture should always be in SHADER_READ_ONLY layout prior to use.
// This is so we don't need to transition during render passes.
_assert_(m_texture->GetLayout() == VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
StateTracker::GetInstance()->SetTexture(stage, m_texture->GetView());
}

View File

@ -416,7 +416,8 @@ bool TextureConverter::SupportsTextureDecoding(TextureFormat format, TlutFormat
return true;
}
void TextureConverter::DecodeTexture(TextureCache::TCacheEntry* entry, u32 dst_level,
void TextureConverter::DecodeTexture(VkCommandBuffer command_buffer,
TextureCache::TCacheEntry* entry, u32 dst_level,
const u8* data, size_t data_size, TextureFormat format,
u32 width, u32 height, u32 aligned_width, u32 aligned_height,
u32 row_stride, const u8* palette, TlutFormat palette_format)
@ -498,11 +499,6 @@ void TextureConverter::DecodeTexture(TextureCache::TCacheEntry* entry, u32 dst_l
break;
}
// Place compute shader dispatches together in the init command buffer.
// That way we don't have to pay a penalty for switching from graphics->compute,
// or end/restart our render pass.
VkCommandBuffer command_buffer = g_command_buffer_mgr->GetCurrentInitCommandBuffer();
// Dispatch compute to temporary texture.
ComputeShaderDispatcher dispatcher(command_buffer,
g_object_cache->GetPipelineLayout(PIPELINE_LAYOUT_COMPUTE),

View File

@ -49,10 +49,10 @@ public:
u32 src_width, u32 src_stride, u32 src_height);
bool SupportsTextureDecoding(TextureFormat format, TlutFormat palette_format);
void DecodeTexture(TextureCache::TCacheEntry* entry, u32 dst_level, const u8* data,
size_t data_size, TextureFormat format, u32 width, u32 height,
u32 aligned_width, u32 aligned_height, u32 row_stride, const u8* palette,
TlutFormat palette_format);
void DecodeTexture(VkCommandBuffer command_buffer, TextureCache::TCacheEntry* entry,
u32 dst_level, const u8* data, size_t data_size, TextureFormat format,
u32 width, u32 height, u32 aligned_width, u32 aligned_height, u32 row_stride,
const u8* palette, TlutFormat palette_format);
private:
static const u32 ENCODING_TEXTURE_WIDTH = EFB_WIDTH * 4;