mirror of
https://github.com/dolphin-emu/dolphin.git
synced 2025-07-23 14:19:46 -06:00
D3D12: Implement non-blocking EFB access when EFB has not been modified
This commit is contained in:
Stenzek
@ -51,8 +51,6 @@ static bool s_last_xfb_mode = false;
|
||||
|
||||
static Television s_television;
|
||||
|
||||
static ID3D12Resource* s_access_efb_constant_buffer = nullptr;
|
||||
|
||||
enum CLEAR_BLEND_DESC
|
||||
{
|
||||
CLEAR_BLEND_DESC_ALL_CHANNELS_ENABLED = 0,
|
||||
@ -110,25 +108,6 @@ static void SetupDeviceObjects()
|
||||
|
||||
g_framebuffer_manager = std::make_unique<FramebufferManager>();
|
||||
|
||||
float colmat[20] = { 0.0f };
|
||||
colmat[0] = colmat[5] = colmat[10] = 1.0f;
|
||||
|
||||
CheckHR(
|
||||
D3D::device12->CreateCommittedResource(
|
||||
&CD3DX12_HEAP_PROPERTIES(D3D12_HEAP_TYPE_UPLOAD),
|
||||
D3D12_HEAP_FLAG_NONE,
|
||||
&CD3DX12_RESOURCE_DESC::Buffer(sizeof(colmat)),
|
||||
D3D12_RESOURCE_STATE_GENERIC_READ,
|
||||
nullptr,
|
||||
IID_PPV_ARGS(&s_access_efb_constant_buffer)
|
||||
)
|
||||
);
|
||||
|
||||
// Copy inital data to access_efb_cbuf12.
|
||||
void* access_efb_constant_buffer_data = nullptr;
|
||||
CheckHR(s_access_efb_constant_buffer->Map(0, nullptr, &access_efb_constant_buffer_data));
|
||||
memcpy(access_efb_constant_buffer_data, colmat, sizeof(colmat));
|
||||
|
||||
D3D12_DEPTH_STENCIL_DESC depth_desc;
|
||||
depth_desc.DepthEnable = FALSE;
|
||||
depth_desc.DepthWriteMask = D3D12_DEPTH_WRITE_MASK_ZERO;
|
||||
@ -197,9 +176,6 @@ static void TeardownDeviceObjects()
|
||||
s_screenshot_texture = nullptr;
|
||||
}
|
||||
|
||||
D3D::command_list_mgr->DestroyResourceAfterCurrentCommandListExecuted(s_access_efb_constant_buffer);
|
||||
s_access_efb_constant_buffer = nullptr;
|
||||
|
||||
s_television.Shutdown();
|
||||
|
||||
gx_state_cache.Clear();
|
||||
@ -394,192 +370,60 @@ void Renderer::SetColorMask()
|
||||
// - GX_PokeZMode (TODO)
|
||||
u32 Renderer::AccessEFB(EFBAccessType type, u32 x, u32 y, u32 poke_data)
|
||||
{
|
||||
// EXISTINGD3D11TODO: This function currently is broken if anti-aliasing is enabled
|
||||
|
||||
// Convert EFB dimensions to the ones of our render target
|
||||
EFBRectangle efb_pixel_rc;
|
||||
efb_pixel_rc.left = x;
|
||||
efb_pixel_rc.top = y;
|
||||
efb_pixel_rc.right = x + 1;
|
||||
efb_pixel_rc.bottom = y + 1;
|
||||
TargetRectangle target_pixel_rc = Renderer::ConvertEFBRectangle(efb_pixel_rc);
|
||||
|
||||
// Take the mean of the resulting dimensions; TODO: Don't use the center pixel, compute the average color instead
|
||||
D3D12_RECT rect_to_lock;
|
||||
if (type == PEEK_COLOR || type == PEEK_Z)
|
||||
if (type == PEEK_COLOR)
|
||||
{
|
||||
rect_to_lock.left = (target_pixel_rc.left + target_pixel_rc.right) / 2;
|
||||
rect_to_lock.top = (target_pixel_rc.top + target_pixel_rc.bottom) / 2;
|
||||
rect_to_lock.right = rect_to_lock.left + 1;
|
||||
rect_to_lock.bottom = rect_to_lock.top + 1;
|
||||
}
|
||||
else
|
||||
{
|
||||
rect_to_lock.left = target_pixel_rc.left;
|
||||
rect_to_lock.right = target_pixel_rc.right;
|
||||
rect_to_lock.top = target_pixel_rc.top;
|
||||
rect_to_lock.bottom = target_pixel_rc.bottom;
|
||||
}
|
||||
u32 color = FramebufferManager::ReadEFBColorAccessCopy(x, y);
|
||||
|
||||
if (type == PEEK_Z)
|
||||
{
|
||||
D3D::command_list_mgr->CPUAccessNotify();
|
||||
|
||||
// depth buffers can only be completely CopySubresourceRegion'ed, so we're using DrawShadedTexQuad instead
|
||||
// D3D12TODO: Is above statement true on D3D12?
|
||||
D3D12_VIEWPORT vp12 = { 0.f, 0.f, 1.f, 1.f, D3D12_MIN_DEPTH, D3D12_MAX_DEPTH };
|
||||
D3D::current_command_list->RSSetViewports(1, &vp12);
|
||||
|
||||
D3D::current_command_list->SetGraphicsRootConstantBufferView(DESCRIPTOR_TABLE_PS_CBVONE, s_access_efb_constant_buffer->GetGPUVirtualAddress());
|
||||
D3D::command_list_mgr->SetCommandListDirtyState(COMMAND_LIST_STATE_PS_CBV, true);
|
||||
|
||||
FramebufferManager::GetEFBDepthReadTexture()->TransitionToResourceState(D3D::current_command_list, D3D12_RESOURCE_STATE_RENDER_TARGET);
|
||||
D3D::current_command_list->OMSetRenderTargets(1, &FramebufferManager::GetEFBDepthReadTexture()->GetRTV12(), FALSE, nullptr);
|
||||
|
||||
D3D::SetPointCopySampler();
|
||||
|
||||
D3D::DrawShadedTexQuad(
|
||||
FramebufferManager::GetEFBDepthTexture(),
|
||||
&rect_to_lock,
|
||||
Renderer::GetTargetWidth(),
|
||||
Renderer::GetTargetHeight(),
|
||||
StaticShaderCache::GetColorCopyPixelShader(true),
|
||||
StaticShaderCache::GetSimpleVertexShader(),
|
||||
StaticShaderCache::GetSimpleVertexShaderInputLayout(),
|
||||
D3D12_SHADER_BYTECODE(),
|
||||
1.0f,
|
||||
0,
|
||||
DXGI_FORMAT_R32_FLOAT,
|
||||
false,
|
||||
FramebufferManager::GetEFBDepthReadTexture()->GetMultisampled()
|
||||
);
|
||||
|
||||
// copy to system memory
|
||||
D3D12_BOX src_box = CD3DX12_BOX(0, 0, 0, 1, 1, 1);
|
||||
ID3D12Resource* readback_buffer = FramebufferManager::GetEFBDepthStagingBuffer();
|
||||
|
||||
D3D12_TEXTURE_COPY_LOCATION dst_location = {};
|
||||
dst_location.pResource = readback_buffer;
|
||||
dst_location.Type = D3D12_TEXTURE_COPY_TYPE_PLACED_FOOTPRINT;
|
||||
dst_location.PlacedFootprint.Offset = 0;
|
||||
dst_location.PlacedFootprint.Footprint.Format = DXGI_FORMAT_R32_FLOAT;
|
||||
dst_location.PlacedFootprint.Footprint.Width = 1;
|
||||
dst_location.PlacedFootprint.Footprint.Height = 1;
|
||||
dst_location.PlacedFootprint.Footprint.Depth = 1;
|
||||
dst_location.PlacedFootprint.Footprint.RowPitch = D3D::AlignValue(dst_location.PlacedFootprint.Footprint.Width * 4, D3D12_TEXTURE_DATA_PITCH_ALIGNMENT);
|
||||
|
||||
D3D12_TEXTURE_COPY_LOCATION src_location = {};
|
||||
src_location.pResource = FramebufferManager::GetEFBDepthReadTexture()->GetTex12();
|
||||
src_location.SubresourceIndex = 0;
|
||||
src_location.Type = D3D12_TEXTURE_COPY_TYPE_SUBRESOURCE_INDEX;
|
||||
|
||||
FramebufferManager::GetEFBDepthReadTexture()->TransitionToResourceState(D3D::current_command_list, D3D12_RESOURCE_STATE_COPY_SOURCE);
|
||||
D3D::current_command_list->CopyTextureRegion(&dst_location, 0, 0, 0, &src_location, &src_box);
|
||||
|
||||
// Need to wait for the CPU to complete the copy (and all prior operations) before we can read it on the CPU.
|
||||
D3D::command_list_mgr->ExecuteQueuedWork(true);
|
||||
|
||||
FramebufferManager::GetEFBColorTexture()->TransitionToResourceState(D3D::current_command_list, D3D12_RESOURCE_STATE_RENDER_TARGET);
|
||||
FramebufferManager::GetEFBDepthTexture()->TransitionToResourceState(D3D::current_command_list, D3D12_RESOURCE_STATE_DEPTH_WRITE );
|
||||
D3D::current_command_list->OMSetRenderTargets(1, &FramebufferManager::GetEFBColorTexture()->GetRTV12(), FALSE, &FramebufferManager::GetEFBDepthTexture()->GetDSV12());
|
||||
|
||||
// Restores proper viewport/scissor settings.
|
||||
g_renderer->RestoreAPIState();
|
||||
|
||||
// read the data from system memory
|
||||
void* readback_buffer_data = nullptr;
|
||||
CheckHR(readback_buffer->Map(0, nullptr, &readback_buffer_data));
|
||||
|
||||
// depth buffer is inverted in the d3d backend
|
||||
float val = 1.0f - reinterpret_cast<float*>(readback_buffer_data)[0];
|
||||
u32 ret = 0;
|
||||
|
||||
if (bpmem.zcontrol.pixel_format == PEControl::RGB565_Z16)
|
||||
{
|
||||
// if Z is in 16 bit format you must return a 16 bit integer
|
||||
ret = MathUtil::Clamp<u32>(static_cast<u32>(val * 65536.0f), 0, 0xFFFF);
|
||||
}
|
||||
else
|
||||
{
|
||||
ret = MathUtil::Clamp<u32>(static_cast<u32>(val * 16777216.0f), 0, 0xFFFFFF);
|
||||
}
|
||||
|
||||
// EXISTINGD3D11TODO: in RE0 this value is often off by one in Video_DX9 (where this code is derived from), which causes lighting to disappear
|
||||
return ret;
|
||||
}
|
||||
else if (type == PEEK_COLOR)
|
||||
{
|
||||
D3D::command_list_mgr->CPUAccessNotify();
|
||||
|
||||
ID3D12Resource* readback_buffer = FramebufferManager::GetEFBColorStagingBuffer();
|
||||
|
||||
D3D12_BOX src_box = CD3DX12_BOX(rect_to_lock.left, rect_to_lock.top, 0, rect_to_lock.right, rect_to_lock.bottom, 1);
|
||||
|
||||
D3D12_TEXTURE_COPY_LOCATION dst_location = {};
|
||||
dst_location.pResource = readback_buffer;
|
||||
dst_location.Type = D3D12_TEXTURE_COPY_TYPE_PLACED_FOOTPRINT;
|
||||
dst_location.PlacedFootprint.Offset = 0;
|
||||
dst_location.PlacedFootprint.Footprint.Format = DXGI_FORMAT_R8G8B8A8_UNORM;
|
||||
dst_location.PlacedFootprint.Footprint.Width = 1;
|
||||
dst_location.PlacedFootprint.Footprint.Height = 1;
|
||||
dst_location.PlacedFootprint.Footprint.Depth = 1;
|
||||
dst_location.PlacedFootprint.Footprint.RowPitch = D3D::AlignValue(dst_location.PlacedFootprint.Footprint.Width * 4, D3D12_TEXTURE_DATA_PITCH_ALIGNMENT);
|
||||
|
||||
D3D12_TEXTURE_COPY_LOCATION src_location = {};
|
||||
src_location.pResource = FramebufferManager::GetResolvedEFBColorTexture()->GetTex12();
|
||||
src_location.SubresourceIndex = 0;
|
||||
src_location.Type = D3D12_TEXTURE_COPY_TYPE_SUBRESOURCE_INDEX;
|
||||
|
||||
FramebufferManager::GetResolvedEFBColorTexture()->TransitionToResourceState(D3D::current_command_list, D3D12_RESOURCE_STATE_COPY_SOURCE);
|
||||
D3D::current_command_list->CopyTextureRegion(&dst_location, 0, 0, 0, &src_location, &src_box);
|
||||
|
||||
// Need to wait for the CPU to complete the copy (and all prior operations) before we can read it on the CPU.
|
||||
D3D::command_list_mgr->ExecuteQueuedWork(true);
|
||||
|
||||
FramebufferManager::GetEFBColorTexture()->TransitionToResourceState(D3D::current_command_list, D3D12_RESOURCE_STATE_RENDER_TARGET);
|
||||
FramebufferManager::GetEFBDepthTexture()->TransitionToResourceState(D3D::current_command_list, D3D12_RESOURCE_STATE_DEPTH_WRITE);
|
||||
D3D::current_command_list->OMSetRenderTargets(1, &FramebufferManager::GetEFBColorTexture()->GetRTV12(), FALSE, &FramebufferManager::GetEFBDepthTexture()->GetDSV12());
|
||||
|
||||
// Restores proper viewport/scissor settings.
|
||||
g_renderer->RestoreAPIState();
|
||||
|
||||
// read the data from system memory
|
||||
void* readback_buffer_data = nullptr;
|
||||
CheckHR(readback_buffer->Map(0, nullptr, &readback_buffer_data));
|
||||
|
||||
u32 ret = reinterpret_cast<u32*>(readback_buffer_data)[0];
|
||||
// a little-endian value is expected to be returned
|
||||
color = ((color & 0xFF00FF00) | ((color >> 16) & 0xFF) | ((color << 16) & 0xFF0000));
|
||||
|
||||
// check what to do with the alpha channel (GX_PokeAlphaRead)
|
||||
PixelEngine::UPEAlphaReadReg alpha_read_mode = PixelEngine::GetAlphaReadMode();
|
||||
|
||||
if (bpmem.zcontrol.pixel_format == PEControl::RGBA6_Z24)
|
||||
{
|
||||
ret = RGBA8ToRGBA6ToRGBA8(ret);
|
||||
color = RGBA8ToRGBA6ToRGBA8(color);
|
||||
}
|
||||
else if (bpmem.zcontrol.pixel_format == PEControl::RGB565_Z16)
|
||||
{
|
||||
ret = RGBA8ToRGB565ToRGBA8(ret);
|
||||
color = RGBA8ToRGB565ToRGBA8(color);
|
||||
}
|
||||
if (bpmem.zcontrol.pixel_format != PEControl::RGBA6_Z24)
|
||||
{
|
||||
ret |= 0xFF000000;
|
||||
color |= 0xFF000000;
|
||||
}
|
||||
|
||||
if (alpha_read_mode.ReadMode == 2)
|
||||
{
|
||||
return ret; // GX_READ_NONE
|
||||
return color; // GX_READ_NONE
|
||||
}
|
||||
else if (alpha_read_mode.ReadMode == 1)
|
||||
{
|
||||
return (ret | 0xFF000000); // GX_READ_FF
|
||||
return (color | 0xFF000000); // GX_READ_FF
|
||||
}
|
||||
else /*if(alpha_read_mode.ReadMode == 0)*/
|
||||
{
|
||||
return (ret & 0x00FFFFFF); // GX_READ_00
|
||||
return (color & 0x00FFFFFF); // GX_READ_00
|
||||
}
|
||||
}
|
||||
else // if (type == PEEK_Z)
|
||||
{
|
||||
// depth buffer is inverted in the d3d backend
|
||||
float depth = 1.0f - FramebufferManager::ReadEFBDepthAccessCopy(x, y);
|
||||
u32 ret = 0;
|
||||
|
||||
return 0;
|
||||
if (bpmem.zcontrol.pixel_format == PEControl::RGB565_Z16)
|
||||
{
|
||||
// if Z is in 16 bit format you must return a 16 bit integer
|
||||
ret = MathUtil::Clamp<u32>(static_cast<u32>(depth * 65536.0f), 0, 0xFFFF);
|
||||
}
|
||||
else
|
||||
{
|
||||
ret = MathUtil::Clamp<u32>(static_cast<u32>(depth * 16777216.0f), 0, 0xFFFFFF);
|
||||
}
|
||||
|
||||
return ret;
|
||||
}
|
||||
}
|
||||
|
||||
void Renderer::PokeEFB(EFBAccessType type, const EfbPokeData* points, size_t num_points)
|
||||
@ -706,6 +550,8 @@ void Renderer::ClearScreen(const EFBRectangle& rc, bool color_enable, bool alpha
|
||||
|
||||
// Restores proper viewport/scissor settings.
|
||||
g_renderer->RestoreAPIState();
|
||||
|
||||
FramebufferManager::InvalidateEFBAccessCopies();
|
||||
}
|
||||
|
||||
void Renderer::ReinterpretPixelData(unsigned int convtype)
|
||||
@ -906,6 +752,9 @@ void Renderer::SwapImpl(u32 xfb_addr, u32 fb_width, u32 fb_stride, u32 fb_height
|
||||
return;
|
||||
}
|
||||
|
||||
// Invalidate EFB access copies. Not strictly necessary, but this avoids having the buffers mapped when calling Present().
|
||||
FramebufferManager::InvalidateEFBAccessCopies();
|
||||
|
||||
// Prepare to copy the XFBs to our backbuffer
|
||||
UpdateDrawRectangle(s_backbuffer_width, s_backbuffer_height);
|
||||
TargetRectangle target_rc = GetTargetRectangle();
|
||||
@ -1272,6 +1121,9 @@ void Renderer::ApplyState(bool use_dst_alpha)
|
||||
|
||||
D3D::command_list_mgr->SetCommandListDirtyState(COMMAND_LIST_STATE_PSO, false);
|
||||
}
|
||||
|
||||
// Always called prior to drawing, so we can invalidate the CPU EFB copies here.
|
||||
FramebufferManager::InvalidateEFBAccessCopies();
|
||||
}
|
||||
|
||||
void Renderer::RestoreState()
|
||||
|
||||
Reference in New Issue
Block a user