RPCS3/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2024-11-26 12:42:41 +01:00
Code Issues Releases Wiki Activity

d3d12: Use a shared root signature and add more slots

Browse Source 
This allows for finer grained rebinding later.
This commit is contained in:
Vincent Lejeune 2016-02-21 17:55:34 +01:00
parent 4b54d7ece1
commit 9ef24509cb
7 changed files with 96 additions and 76 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
rpcs3/Emu/RSX/D3D12/D3D12Buffer.cpp
View File

@ -252,7 +252,7 @@ void D3D12GSRender::upload_and_bind_vertex_shader_constants(size_t descriptor_in
.Offset((INT)descriptor_index, m_descriptor_stride_srv_cbv_uav));
}
void D3D12GSRender::upload_and_bind_fragment_shader_constants(size_t descriptor_index)
D3D12_CONSTANT_BUFFER_VIEW_DESC D3D12GSRender::upload_fragment_shader_constants()
{
// Get constant from fragment program
size_t buffer_size = m_pso_cache.get_fragment_constants_buffer_size(m_fragment_program);
@ -266,13 +266,10 @@ void D3D12GSRender::upload_and_bind_fragment_shader_constants(size_t descriptor_
m_pso_cache.fill_fragment_constans_buffer({ mapped_buffer, gsl::narrow<int>(buffer_size) }, m_fragment_program);
m_buffer_data.unmap(CD3DX12_RANGE(heap_offset, heap_offset + buffer_size));
D3D12_CONSTANT_BUFFER_VIEW_DESC constant_buffer_view_desc = {
return {
m_buffer_data.get_heap()->GetGPUVirtualAddress() + heap_offset,
(UINT)buffer_size
};
m_device->CreateConstantBufferView(&constant_buffer_view_desc,
CD3DX12_CPU_DESCRIPTOR_HANDLE(get_current_resource_storage().descriptors_heap->GetCPUDescriptorHandleForHeapStart())
.Offset((INT)descriptor_index, m_descriptor_stride_srv_cbv_uav));
}

8
rpcs3/Emu/RSX/D3D12/D3D12FragmentProgramDecompiler.cpp
View File

@ -107,7 +107,7 @@ void D3D12FragmentDecompiler::insertConstants(std::stringstream & OS)
for (const ParamItem &PI : PT.items)
{
size_t textureIndex = atoi(PI.name.data() + 3);
OS << "Texture1D " << PI.name << " : register(t" << textureIndex + 16 << ");" << std::endl;
OS << "Texture1D " << PI.name << " : register(t" << textureIndex << ");" << std::endl;
OS << "sampler " << PI.name << "sampler : register(s" << textureIndex << ");" << std::endl;
}
}
@ -116,7 +116,7 @@ void D3D12FragmentDecompiler::insertConstants(std::stringstream & OS)
for (const ParamItem &PI : PT.items)
{
size_t textureIndex = atoi(PI.name.data() + 3);
OS << "Texture2D " << PI.name << " : register(t" << textureIndex + 16 << ");" << std::endl;
OS << "Texture2D " << PI.name << " : register(t" << textureIndex << ");" << std::endl;
OS << "sampler " << PI.name << "sampler : register(s" << textureIndex << ");" << std::endl;
}
}
@ -125,7 +125,7 @@ void D3D12FragmentDecompiler::insertConstants(std::stringstream & OS)
for (const ParamItem &PI : PT.items)
{
size_t textureIndex = atoi(PI.name.data() + 3);
OS << "Texture3D " << PI.name << " : register(t" << textureIndex + 16 << ");" << std::endl;
OS << "Texture3D " << PI.name << " : register(t" << textureIndex << ");" << std::endl;
OS << "sampler " << PI.name << "sampler : register(s" << textureIndex << ");" << std::endl;
}
}
@ -134,7 +134,7 @@ void D3D12FragmentDecompiler::insertConstants(std::stringstream & OS)
for (const ParamItem &PI : PT.items)
{
size_t textureIndex = atoi(PI.name.data() + 3);
OS << "TextureCube " << PI.name << " : register(t" << textureIndex + 16 << ");" << std::endl;
OS << "TextureCube " << PI.name << " : register(t" << textureIndex << ");" << std::endl;
OS << "sampler " << PI.name << "sampler : register(s" << textureIndex << ");" << std::endl;
}
}

137
rpcs3/Emu/RSX/D3D12/D3D12GSRender.cpp
View File

@ -18,6 +18,13 @@ PFN_D3D12_SERIALIZE_ROOT_SIGNATURE wrapD3D12SerializeRootSignature;
PFN_D3D11ON12_CREATE_DEVICE wrapD3D11On12CreateDevice;
pD3DCompile wrapD3DCompile;
#define VERTEX_BUFFERS_SLOT 0
#define FRAGMENT_CONSTANT_BUFFERS_SLOT 1
#define VERTEX_CONSTANT_BUFFERS_SLOT 2
#define TEXTURES_SLOT 3
#define SAMPLERS_SLOT 4
#define SCALE_OFFSET_SLOT 5
namespace
{
HMODULE D3D12Module;
@ -84,6 +91,44 @@ D3D12DLLManagement::~D3D12DLLManagement()
unloadD3D12FunctionPointers();
}
namespace
{
ComPtr<ID3DBlob> get_shared_root_signature_blob()
{
CD3DX12_ROOT_PARAMETER RP[6];
// vertex buffer are bound each draw calls
CD3DX12_DESCRIPTOR_RANGE vertex_buffer_descriptors(D3D12_DESCRIPTOR_RANGE_TYPE_SRV, 16, 0);
RP[VERTEX_BUFFERS_SLOT].InitAsDescriptorTable(1, &vertex_buffer_descriptors, D3D12_SHADER_VISIBILITY_VERTEX);
// fragment constants are bound each draw calls
RP[FRAGMENT_CONSTANT_BUFFERS_SLOT].InitAsConstantBufferView(2, 0, D3D12_SHADER_VISIBILITY_PIXEL);
// vertex constants are bound often
CD3DX12_DESCRIPTOR_RANGE vertex_constant_buffer_descriptors(D3D12_DESCRIPTOR_RANGE_TYPE_CBV, 1, 1);
RP[VERTEX_CONSTANT_BUFFERS_SLOT].InitAsDescriptorTable(1, &vertex_constant_buffer_descriptors, D3D12_SHADER_VISIBILITY_VERTEX);
// textures are bound often
CD3DX12_DESCRIPTOR_RANGE texture_descriptors(D3D12_DESCRIPTOR_RANGE_TYPE_SRV, 16, 0);
RP[TEXTURES_SLOT].InitAsDescriptorTable(1, &texture_descriptors, D3D12_SHADER_VISIBILITY_PIXEL);
// samplers are bound often
CD3DX12_DESCRIPTOR_RANGE sampler_descriptors(D3D12_DESCRIPTOR_RANGE_TYPE_SAMPLER, 16, 0);
RP[SAMPLERS_SLOT].InitAsDescriptorTable(1, &sampler_descriptors, D3D12_SHADER_VISIBILITY_PIXEL);
// scale offset matrix are bound once in a while
CD3DX12_DESCRIPTOR_RANGE scale_offset_descriptors(D3D12_DESCRIPTOR_RANGE_TYPE_CBV, 1, 0);
RP[SCALE_OFFSET_SLOT].InitAsDescriptorTable(1, &scale_offset_descriptors, D3D12_SHADER_VISIBILITY_ALL);
Microsoft::WRL::ComPtr<ID3DBlob> root_signature_blob;
Microsoft::WRL::ComPtr<ID3DBlob> error_blob;
CHECK_HRESULT(wrapD3D12SerializeRootSignature(
&CD3DX12_ROOT_SIGNATURE_DESC(6, RP, 0, 0),
D3D_ROOT_SIGNATURE_VERSION_1, root_signature_blob.GetAddressOf(), error_blob.GetAddressOf()));
return root_signature_blob;
}
}
D3D12GSRender::D3D12GSRender()
: GSRender(frame_type::DX12), m_d3d12_lib(), m_current_pso({})
{
@ -134,45 +179,12 @@ D3D12GSRender::D3D12GSRender()
m_device->CreateDescriptorHeap(&render_target_descriptor_heap_desc, IID_PPV_ARGS(&m_backbuffer_descriptor_heap[1]));
m_device->CreateRenderTargetView(m_backbuffer[1].Get(), &renter_target_view_desc, m_backbuffer_descriptor_heap[1]->GetCPUDescriptorHandleForHeapStart());
// Common root signatures
for (int vertex_buffer_count = 0; vertex_buffer_count < 17; vertex_buffer_count++) // Some app (naruto ultimate ninja storm 2) uses a shader without inputs...
{
for (unsigned texture_count = 0; texture_count < 17; texture_count++)
{
CD3DX12_DESCRIPTOR_RANGE descriptorRange[] =
{
// Vertex buffer
CD3DX12_DESCRIPTOR_RANGE(D3D12_DESCRIPTOR_RANGE_TYPE_SRV, vertex_buffer_count, 0),
// Scale Offset data
CD3DX12_DESCRIPTOR_RANGE(D3D12_DESCRIPTOR_RANGE_TYPE_CBV, 1, 0),
// Constants
CD3DX12_DESCRIPTOR_RANGE(D3D12_DESCRIPTOR_RANGE_TYPE_CBV, 2, 1),
// Textures
CD3DX12_DESCRIPTOR_RANGE(D3D12_DESCRIPTOR_RANGE_TYPE_SRV, texture_count, 16),
// Samplers
CD3DX12_DESCRIPTOR_RANGE(D3D12_DESCRIPTOR_RANGE_TYPE_SAMPLER, texture_count, 0),
};
CD3DX12_ROOT_PARAMETER RP[2];
UINT cbv_srv_uav_descriptor_size = 4;
if (texture_count == 0)
cbv_srv_uav_descriptor_size -= 1;
if (vertex_buffer_count == 0)
cbv_srv_uav_descriptor_size -= 1;
RP[0].InitAsDescriptorTable(cbv_srv_uav_descriptor_size, (vertex_buffer_count > 0) ? &descriptorRange[0] : &descriptorRange[1]);
RP[1].InitAsDescriptorTable(1, &descriptorRange[4]);
Microsoft::WRL::ComPtr<ID3DBlob> rootSignatureBlob;
Microsoft::WRL::ComPtr<ID3DBlob> errorBlob;
CHECK_HRESULT(wrapD3D12SerializeRootSignature(
&CD3DX12_ROOT_SIGNATURE_DESC((texture_count > 0) ? 2 : 1, RP, 0, 0),
D3D_ROOT_SIGNATURE_VERSION_1, &rootSignatureBlob, &errorBlob));
ComPtr<ID3DBlob> root_signature_blob = get_shared_root_signature_blob();
m_device->CreateRootSignature(0,
rootSignatureBlob->GetBufferPointer(),
rootSignatureBlob->GetBufferSize(),
IID_PPV_ARGS(m_root_signatures[texture_count][vertex_buffer_count].GetAddressOf()));
}
}
root_signature_blob->GetBufferPointer(),
root_signature_blob->GetBufferSize(),
IID_PPV_ARGS(m_shared_root_signature.GetAddressOf()));
m_per_frame_storage[0].init(m_device.Get());
m_per_frame_storage[0].reset();
@ -269,7 +281,7 @@ void D3D12GSRender::end()
std::vector<D3D12_SHADER_RESOURCE_VIEW_DESC> vertex_buffer_views;
std::tie(indexed_draw, vertex_count, vertex_buffer_views) = upload_and_set_vertex_index_data(get_current_resource_storage().command_list.Get());
size_t vertex_buffer_count = vertex_buffer_views.size();
UINT vertex_buffer_count = static_cast<UINT>(vertex_buffer_views.size());
std::chrono::time_point<std::chrono::system_clock> vertex_index_duration_end = std::chrono::system_clock::now();
m_timers.vertex_index_duration += std::chrono::duration_cast<std::chrono::microseconds>(vertex_index_duration_end - vertex_index_duration_start).count();
@ -279,7 +291,7 @@ void D3D12GSRender::end()
std::chrono::time_point<std::chrono::system_clock> program_load_end = std::chrono::system_clock::now();
m_timers.program_load_duration += std::chrono::duration_cast<std::chrono::microseconds>(program_load_end - program_load_start).count();
get_current_resource_storage().command_list->SetGraphicsRootSignature(m_root_signatures[std::get<2>(m_current_pso)][vertex_buffer_count].Get());
get_current_resource_storage().command_list->SetGraphicsRootSignature(m_shared_root_signature.Get());
get_current_resource_storage().command_list->OMSetStencilRef(rsx::method_registers[NV4097_SET_STENCIL_FUNC_REF]);
std::chrono::time_point<std::chrono::system_clock> constants_duration_start = std::chrono::system_clock::now();
@ -291,11 +303,27 @@ void D3D12GSRender::end()
CD3DX12_CPU_DESCRIPTOR_HANDLE(get_current_resource_storage().descriptors_heap->GetCPUDescriptorHandleForHeapStart())
.Offset((INT)currentDescriptorIndex + offset++, m_descriptor_stride_srv_cbv_uav));
}
// Bind vertex buffer
get_current_resource_storage().command_list->SetGraphicsRootDescriptorTable(VERTEX_BUFFERS_SLOT,
CD3DX12_GPU_DESCRIPTOR_HANDLE(get_current_resource_storage().descriptors_heap->GetGPUDescriptorHandleForHeapStart())
.Offset((INT)currentDescriptorIndex, m_descriptor_stride_srv_cbv_uav)
);
// Constants
const D3D12_CONSTANT_BUFFER_VIEW_DESC &fragment_constant_view = upload_fragment_shader_constants();
get_current_resource_storage().command_list->SetGraphicsRootConstantBufferView(FRAGMENT_CONSTANT_BUFFERS_SLOT, fragment_constant_view.BufferLocation);
upload_and_bind_scale_offset_matrix(currentDescriptorIndex + vertex_buffer_count);
get_current_resource_storage().command_list->SetGraphicsRootDescriptorTable(SCALE_OFFSET_SLOT,
CD3DX12_GPU_DESCRIPTOR_HANDLE(get_current_resource_storage().descriptors_heap->GetGPUDescriptorHandleForHeapStart())
.Offset((INT)currentDescriptorIndex + vertex_buffer_count, m_descriptor_stride_srv_cbv_uav)
);
upload_and_bind_vertex_shader_constants(currentDescriptorIndex + 1 + vertex_buffer_count);
upload_and_bind_fragment_shader_constants(currentDescriptorIndex + 2 + vertex_buffer_count);
get_current_resource_storage().command_list->SetGraphicsRootDescriptorTable(VERTEX_CONSTANT_BUFFERS_SLOT,
CD3DX12_GPU_DESCRIPTOR_HANDLE(get_current_resource_storage().descriptors_heap->GetGPUDescriptorHandleForHeapStart())
.Offset((INT)currentDescriptorIndex + 1 + vertex_buffer_count, m_descriptor_stride_srv_cbv_uav)
);
std::chrono::time_point<std::chrono::system_clock> constants_duration_end = std::chrono::system_clock::now();
m_timers.constants_duration += std::chrono::duration_cast<std::chrono::microseconds>(constants_duration_end - constants_duration_start).count();
@ -303,44 +331,39 @@ void D3D12GSRender::end()
get_current_resource_storage().command_list->SetPipelineState(std::get<0>(m_current_pso).Get());
std::chrono::time_point<std::chrono::system_clock> texture_duration_start = std::chrono::system_clock::now();
get_current_resource_storage().descriptors_heap_index += 2 + vertex_buffer_count;
size_t texture_count = std::get<2>(m_current_pso);
if (texture_count > 0)
{
upload_and_bind_textures(get_current_resource_storage().command_list.Get(), texture_count);
upload_textures(get_current_resource_storage().command_list.Get(), texture_count);
for (unsigned i = 0; i < texture_count; i++)
// Bind texture and samplers
for (u32 i = 0; i < texture_count; i++)
{
ID3D12Resource *tex_resource;
D3D12_SHADER_RESOURCE_VIEW_DESC srv;
std::tie(tex_resource, srv) = m_current_shader_resources[i];
m_device->CreateShaderResourceView(tex_resource, &srv,
CD3DX12_CPU_DESCRIPTOR_HANDLE(get_current_resource_storage().descriptors_heap->GetCPUDescriptorHandleForHeapStart())
.Offset((INT)currentDescriptorIndex + 3 + (INT)vertex_buffer_count + (INT)i, m_descriptor_stride_srv_cbv_uav)
.Offset((UINT)get_current_resource_storage().descriptors_heap_index + i, m_descriptor_stride_srv_cbv_uav)
);
m_device->CreateSampler(&m_current_samplers[i],
CD3DX12_CPU_DESCRIPTOR_HANDLE(get_current_resource_storage().sampler_descriptor_heap[get_current_resource_storage().sampler_descriptors_heap_index]->GetCPUDescriptorHandleForHeapStart())
.Offset((UINT)get_current_resource_storage().current_sampler_index + (UINT)i, m_descriptor_stride_samplers));
.Offset((UINT)get_current_resource_storage().current_sampler_index + i, m_descriptor_stride_samplers));
}
get_current_resource_storage().command_list->SetGraphicsRootDescriptorTable(0,
get_current_resource_storage().command_list->SetGraphicsRootDescriptorTable(TEXTURES_SLOT,
CD3DX12_GPU_DESCRIPTOR_HANDLE(get_current_resource_storage().descriptors_heap->GetGPUDescriptorHandleForHeapStart())
.Offset((INT)currentDescriptorIndex, m_descriptor_stride_srv_cbv_uav)
.Offset((INT)get_current_resource_storage().descriptors_heap_index, m_descriptor_stride_srv_cbv_uav)
);
get_current_resource_storage().command_list->SetGraphicsRootDescriptorTable(1,
get_current_resource_storage().command_list->SetGraphicsRootDescriptorTable(SAMPLERS_SLOT,
CD3DX12_GPU_DESCRIPTOR_HANDLE(get_current_resource_storage().sampler_descriptor_heap[get_current_resource_storage().sampler_descriptors_heap_index]->GetGPUDescriptorHandleForHeapStart())
.Offset((INT)get_current_resource_storage().current_sampler_index, m_descriptor_stride_samplers)
);
get_current_resource_storage().current_sampler_index += std::get<2>(m_current_pso);
get_current_resource_storage().descriptors_heap_index += std::get<2>(m_current_pso) + 3 + vertex_buffer_count;
}
else
{
get_current_resource_storage().command_list->SetGraphicsRootDescriptorTable(0,
CD3DX12_GPU_DESCRIPTOR_HANDLE(get_current_resource_storage().descriptors_heap->GetGPUDescriptorHandleForHeapStart())
.Offset((INT)currentDescriptorIndex, m_descriptor_stride_srv_cbv_uav)
);
get_current_resource_storage().descriptors_heap_index += 3 + vertex_buffer_count;
get_current_resource_storage().current_sampler_index += texture_count;
get_current_resource_storage().descriptors_heap_index += texture_count;
}
std::chrono::time_point<std::chrono::system_clock> texture_duration_end = std::chrono::system_clock::now();

8
rpcs3/Emu/RSX/D3D12/D3D12GSRender.h
View File

@ -55,8 +55,8 @@ private:
ComPtr<struct IDXGISwapChain3> m_swap_chain;
ComPtr<ID3D12Resource> m_backbuffer[2];
ComPtr<ID3D12DescriptorHeap> m_backbuffer_descriptor_heap[2];
// m_rootSignatures[N] is RS with N texture/sample
ComPtr<ID3D12RootSignature> m_root_signatures[17][17]; // indexed by [texture count][vertex count]
ComPtr<ID3D12RootSignature> m_shared_root_signature;
// TODO: Use a tree structure to parse more efficiently
data_cache m_texture_cache;
@ -161,14 +161,14 @@ private:
void upload_and_bind_scale_offset_matrix(size_t descriptor_index);
void upload_and_bind_vertex_shader_constants(size_t descriptor_index);
void upload_and_bind_fragment_shader_constants(size_t descriptorIndex);
D3D12_CONSTANT_BUFFER_VIEW_DESC upload_fragment_shader_constants();
/**
* Fetch all textures recorded in the state in the render target cache and in the texture cache.
* If a texture is not cached, populate cmdlist with uploads command.
* Create necessary resource view/sampler descriptors in the per frame storage struct.
* If the count of enabled texture is below texture_count, fills with dummy texture and sampler.
*/
void upload_and_bind_textures(ID3D12GraphicsCommandList *command_list, size_t texture_count);
void upload_textures(ID3D12GraphicsCommandList *command_list, size_t texture_count);
/**
* Creates render target if necessary.

2
rpcs3/Emu/RSX/D3D12/D3D12PipelineState.cpp
View File

@ -236,7 +236,7 @@ void D3D12GSRender::load_program()
}
}
m_current_pso = m_pso_cache.getGraphicPipelineState(m_vertex_program, m_fragment_program, prop, m_device.Get(), m_root_signatures);
m_current_pso = m_pso_cache.getGraphicPipelineState(m_vertex_program, m_fragment_program, prop, m_device.Get(), m_shared_root_signature.Get());
return;
}

4
rpcs3/Emu/RSX/D3D12/D3D12PipelineState.h
View File

@ -148,7 +148,7 @@ struct D3D12Traits
static
pipeline_storage_type build_pipeline(
const vertex_program_type &vertexProgramData, const fragment_program_type &fragmentProgramData, const pipeline_properties &pipelineProperties,
ID3D12Device *device, gsl::span<ComPtr<ID3D12RootSignature>, 17, 17> root_signatures)
ID3D12Device *device, ID3D12RootSignature* root_signatures)
{
std::tuple<ID3D12PipelineState *, std::vector<size_t>, size_t> result = {};
D3D12_GRAPHICS_PIPELINE_STATE_DESC graphicPipelineStateDesc = {};
@ -163,7 +163,7 @@ struct D3D12Traits
graphicPipelineStateDesc.PS.BytecodeLength = fragmentProgramData.bytecode->GetBufferSize();
graphicPipelineStateDesc.PS.pShaderBytecode = fragmentProgramData.bytecode->GetBufferPointer();
graphicPipelineStateDesc.pRootSignature = root_signatures[fragmentProgramData.m_textureCount][vertexProgramData.vertex_shader_input_count].Get();
graphicPipelineStateDesc.pRootSignature = root_signatures;
graphicPipelineStateDesc.BlendState = pipelineProperties.Blend;
graphicPipelineStateDesc.DepthStencilState = pipelineProperties.DepthStencil;

4
rpcs3/Emu/RSX/D3D12/D3D12Texture.cpp
View File

@ -173,9 +173,9 @@ D3D12_SHADER_RESOURCE_VIEW_DESC get_srv_descriptor_with_dimensions(const rsx::te
}
}
void D3D12GSRender::upload_and_bind_textures(ID3D12GraphicsCommandList *command_list, size_t texture_count)
void D3D12GSRender::upload_textures(ID3D12GraphicsCommandList *command_list, size_t texture_count)
{
for (u32 i = 0; i < 16; ++i)
for (u32 i = 0; i < texture_count; ++i)
{
if (!m_textures_dirty[i])
continue;