void D3D12GSRender::end()
{
std::chrono::time_point<steady_clock> start_duration = steady_clock::now();
std::chrono::time_point<steady_clock> rtt_duration_start = steady_clock::now();
prepare_render_targets(get_current_resource_storage().command_list.Get());
std::chrono::time_point<steady_clock> rtt_duration_end = steady_clock::now();
m_timers.prepare_rtt_duration += std::chrono::duration_cast<std::chrono::microseconds>(rtt_duration_end - rtt_duration_start).count();
std::chrono::time_point<steady_clock> vertex_index_duration_start = steady_clock::now();
size_t currentDescriptorIndex = get_current_resource_storage().descriptors_heap_index;
size_t vertex_count;
bool indexed_draw;
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());
UINT vertex_buffer_count = static_cast<UINT>(vertex_buffer_views.size());
std::chrono::time_point<steady_clock> vertex_index_duration_end = steady_clock::now();
m_timers.vertex_index_duration += std::chrono::duration_cast<std::chrono::microseconds>(vertex_index_duration_end - vertex_index_duration_start).count();
std::chrono::time_point<steady_clock> program_load_start = steady_clock::now();
load_program();
std::chrono::time_point<steady_clock> program_load_end = steady_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_shared_root_signature.Get());
get_current_resource_storage().command_list->OMSetStencilRef(rsx::method_registers.stencil_func_ref());
std::chrono::time_point<steady_clock> constants_duration_start = steady_clock::now();
INT offset = 0;
for (const auto view : vertex_buffer_views)
{
m_device->CreateShaderResourceView(m_vertex_buffer_data.Get(), &view,
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)
);
if (m_transform_constants_dirty && !g_cfg.video.debug_output)
{
m_current_transform_constants_buffer_descriptor_id = (u32)currentDescriptorIndex + 1 + vertex_buffer_count;
upload_and_bind_vertex_shader_constants(currentDescriptorIndex + 1 + vertex_buffer_count);
m_transform_constants_dirty = false;
get_current_resource_storage().command_list->SetGraphicsRootDescriptorTable(VERTEX_CONSTANT_BUFFERS_SLOT,
CD3DX12_GPU_DESCRIPTOR_HANDLE(get_current_resource_storage().descriptors_heap->GetGPUDescriptorHandleForHeapStart())
.Offset(m_current_transform_constants_buffer_descriptor_id, m_descriptor_stride_srv_cbv_uav)
);
}
std::chrono::time_point<steady_clock> constants_duration_end = steady_clock::now();
m_timers.constants_duration += std::chrono::duration_cast<std::chrono::microseconds>(constants_duration_end - constants_duration_start).count();
get_current_resource_storage().command_list->SetPipelineState(std::get<0>(m_current_pso).Get());
std::chrono::time_point<steady_clock> texture_duration_start = steady_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)
{
if (get_current_resource_storage().current_sampler_index + 16 > 2048)
{
get_current_resource_storage().sampler_descriptors_heap_index = 1;
get_current_resource_storage().current_sampler_index = 0;
ID3D12DescriptorHeap *descriptors[] =
{
get_current_resource_storage().descriptors_heap.Get(),
get_current_resource_storage().sampler_descriptor_heap[get_current_resource_storage().sampler_descriptors_heap_index].Get(),
};
get_current_resource_storage().command_list->SetDescriptorHeaps(2, descriptors);
}
upload_textures(get_current_resource_storage().command_list.Get(), texture_count);
m_device->CopyDescriptorsSimple(16,
CD3DX12_CPU_DESCRIPTOR_HANDLE(get_current_resource_storage().descriptors_heap->GetCPUDescriptorHandleForHeapStart())
.Offset((UINT)get_current_resource_storage().descriptors_heap_index, m_descriptor_stride_srv_cbv_uav),
CD3DX12_CPU_DESCRIPTOR_HANDLE(m_current_texture_descriptors->GetCPUDescriptorHandleForHeapStart()),
D3D12_DESCRIPTOR_HEAP_TYPE_CBV_SRV_UAV
);
m_device->CopyDescriptorsSimple(16,
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, m_descriptor_stride_samplers),
CD3DX12_CPU_DESCRIPTOR_HANDLE(m_current_sampler_descriptors->GetCPUDescriptorHandleForHeapStart()),
D3D12_DESCRIPTOR_HEAP_TYPE_SAMPLER
);
get_current_resource_storage().command_list->SetGraphicsRootDescriptorTable(TEXTURES_SLOT,
CD3DX12_GPU_DESCRIPTOR_HANDLE(get_current_resource_storage().descriptors_heap->GetGPUDescriptorHandleForHeapStart())
.Offset((INT)get_current_resource_storage().descriptors_heap_index, m_descriptor_stride_srv_cbv_uav)
);
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 += texture_count;
get_current_resource_storage().descriptors_heap_index += texture_count;
}
std::chrono::time_point<steady_clock> texture_duration_end = steady_clock::now();
m_timers.texture_duration += std::chrono::duration_cast<std::chrono::microseconds>(texture_duration_end - texture_duration_start).count();
set_rtt_and_ds(get_current_resource_storage().command_list.Get());
int clip_w = rsx::method_registers.surface_clip_width();
int clip_h = rsx::method_registers.surface_clip_height();
D3D12_VIEWPORT viewport =
{
0.f,
0.f,
(float)clip_w,
(float)clip_h,
rsx::method_registers.clip_min(),
rsx::method_registers.clip_max(),
};
get_current_resource_storage().command_list->RSSetViewports(1, &viewport);
get_current_resource_storage().command_list->RSSetScissorRects(1, &get_scissor(rsx::method_registers.scissor_origin_x(), rsx::method_registers.scissor_origin_y(),
rsx::method_registers.scissor_width(), rsx::method_registers.scissor_height()));
get_current_resource_storage().command_list->IASetPrimitiveTopology(get_primitive_topology(rsx::method_registers.current_draw_clause.primitive));
if (indexed_draw)
get_current_resource_storage().command_list->DrawIndexedInstanced((UINT)vertex_count, 1, 0, 0, 0);
else
get_current_resource_storage().command_list->DrawInstanced((UINT)vertex_count, 1, 0, 0);
std::chrono::time_point<steady_clock> end_duration = steady_clock::now();
m_timers.draw_calls_duration += std::chrono::duration_cast<std::chrono::microseconds>(end_duration - start_duration).count();
m_timers.draw_calls_count++;
if (g_cfg.video.debug_output)
{
CHECK_HRESULT(get_current_resource_storage().command_list->Close());
m_command_queue->ExecuteCommandLists(1, (ID3D12CommandList**)get_current_resource_storage().command_list.GetAddressOf());
get_current_resource_storage().set_new_command_list();
}
thread::end();
}
void D3D12GSRender::end()
{
std::chrono::time_point<std::chrono::system_clock> start_duration = std::chrono::system_clock::now();
std::chrono::time_point<std::chrono::system_clock> rtt_duration_start = std::chrono::system_clock::now();
prepare_render_targets(get_current_resource_storage().command_list.Get());
std::chrono::time_point<std::chrono::system_clock> rtt_duration_end = std::chrono::system_clock::now();
m_timers.prepare_rtt_duration += std::chrono::duration_cast<std::chrono::microseconds>(rtt_duration_end - rtt_duration_start).count();
std::chrono::time_point<std::chrono::system_clock> vertex_index_duration_start = std::chrono::system_clock::now();
size_t currentDescriptorIndex = get_current_resource_storage().descriptors_heap_index;
size_t vertex_count;
bool indexed_draw;
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();
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();
std::chrono::time_point<std::chrono::system_clock> program_load_start = std::chrono::system_clock::now();
load_program();
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->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();
INT offset = 0;
for (const auto view : vertex_buffer_views)
{
m_device->CreateShaderResourceView(m_vertex_buffer_data.Get(), &view,
CD3DX12_CPU_DESCRIPTOR_HANDLE(get_current_resource_storage().descriptors_heap->GetCPUDescriptorHandleForHeapStart())
.Offset((INT)currentDescriptorIndex + offset++, m_descriptor_stride_srv_cbv_uav));
}
// Constants
upload_and_bind_scale_offset_matrix(currentDescriptorIndex + vertex_buffer_count);
upload_and_bind_vertex_shader_constants(currentDescriptorIndex + 1 + vertex_buffer_count);
upload_and_bind_fragment_shader_constants(currentDescriptorIndex + 2 + vertex_buffer_count);
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();
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();
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);
for (unsigned 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)
);
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));
}
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().command_list->SetGraphicsRootDescriptorTable(1,
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;
}
std::chrono::time_point<std::chrono::system_clock> texture_duration_end = std::chrono::system_clock::now();
m_timers.texture_duration += std::chrono::duration_cast<std::chrono::microseconds>(texture_duration_end - texture_duration_start).count();
set_rtt_and_ds(get_current_resource_storage().command_list.Get());
int clip_w = rsx::method_registers[NV4097_SET_SURFACE_CLIP_HORIZONTAL] >> 16;
int clip_h = rsx::method_registers[NV4097_SET_SURFACE_CLIP_VERTICAL] >> 16;
D3D12_VIEWPORT viewport =
{
0.f,
0.f,
(float)clip_w,
(float)clip_h,
(f32&)rsx::method_registers[NV4097_SET_CLIP_MIN],
(f32&)rsx::method_registers[NV4097_SET_CLIP_MAX]
};
get_current_resource_storage().command_list->RSSetViewports(1, &viewport);
get_current_resource_storage().command_list->RSSetScissorRects(1, &get_scissor(rsx::method_registers[NV4097_SET_SCISSOR_HORIZONTAL], rsx::method_registers[NV4097_SET_SCISSOR_VERTICAL]));
get_current_resource_storage().command_list->IASetPrimitiveTopology(get_primitive_topology(draw_mode));
if (indexed_draw)
get_current_resource_storage().command_list->DrawIndexedInstanced((UINT)vertex_count, 1, 0, 0, 0);
else
get_current_resource_storage().command_list->DrawInstanced((UINT)vertex_count, 1, 0, 0);
std::chrono::time_point<std::chrono::system_clock> end_duration = std::chrono::system_clock::now();
m_timers.draw_calls_duration += std::chrono::duration_cast<std::chrono::microseconds>(end_duration - start_duration).count();
m_timers.draw_calls_count++;
if (rpcs3::config.rsx.d3d12.debug_output.value())
{
CHECK_HRESULT(get_current_resource_storage().command_list->Close());
m_command_queue->ExecuteCommandLists(1, (ID3D12CommandList**)get_current_resource_storage().command_list.GetAddressOf());
get_current_resource_storage().set_new_command_list();
}
thread::end();
}
void D3D12GSRender::end()
{
std::chrono::time_point<std::chrono::system_clock> start_duration = std::chrono::system_clock::now();
std::chrono::time_point<std::chrono::system_clock> rtt_duration_start = std::chrono::system_clock::now();
prepare_render_targets(getCurrentResourceStorage().command_list.Get());
std::chrono::time_point<std::chrono::system_clock> rtt_duration_end = std::chrono::system_clock::now();
m_timers.m_rttDuration += std::chrono::duration_cast<std::chrono::microseconds>(rtt_duration_end - rtt_duration_start).count();
std::chrono::time_point<std::chrono::system_clock> vertex_index_duration_start = std::chrono::system_clock::now();
if (!vertex_index_array.empty() || vertex_draw_count)
upload_and_set_vertex_index_data(getCurrentResourceStorage().command_list.Get());
std::chrono::time_point<std::chrono::system_clock> vertex_index_duration_end = std::chrono::system_clock::now();
m_timers.m_vertexIndexDuration += std::chrono::duration_cast<std::chrono::microseconds>(vertex_index_duration_end - vertex_index_duration_start).count();
std::chrono::time_point<std::chrono::system_clock> program_load_start = std::chrono::system_clock::now();
if (!load_program())
{
LOG_ERROR(RSX, "LoadProgram failed.");
Emu.Pause();
return;
}
std::chrono::time_point<std::chrono::system_clock> program_load_end = std::chrono::system_clock::now();
m_timers.m_programLoadDuration += std::chrono::duration_cast<std::chrono::microseconds>(program_load_end - program_load_start).count();
getCurrentResourceStorage().command_list->SetGraphicsRootSignature(m_rootSignatures[std::get<2>(*m_PSO)].Get());
getCurrentResourceStorage().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();
size_t currentDescriptorIndex = getCurrentResourceStorage().descriptors_heap_index;
// Constants
upload_and_bind_scale_offset_matrix(currentDescriptorIndex);
upload_and_bind_vertex_shader_constants(currentDescriptorIndex + 1);
upload_and_bind_fragment_shader_constants(currentDescriptorIndex + 2);
std::chrono::time_point<std::chrono::system_clock> constants_duration_end = std::chrono::system_clock::now();
m_timers.m_constantsDuration += std::chrono::duration_cast<std::chrono::microseconds>(constants_duration_end - constants_duration_start).count();
getCurrentResourceStorage().command_list->SetPipelineState(std::get<0>(*m_PSO));
std::chrono::time_point<std::chrono::system_clock> texture_duration_start = std::chrono::system_clock::now();
if (std::get<2>(*m_PSO) > 0)
{
upload_and_bind_textures(getCurrentResourceStorage().command_list.Get(), currentDescriptorIndex + 3, std::get<2>(*m_PSO) > 0);
ID3D12DescriptorHeap *descriptors[] =
{
getCurrentResourceStorage().descriptors_heap.Get(),
getCurrentResourceStorage().sampler_descriptor_heap[getCurrentResourceStorage().sampler_descriptors_heap_index].Get(),
};
getCurrentResourceStorage().command_list->SetDescriptorHeaps(2, descriptors);
getCurrentResourceStorage().command_list->SetGraphicsRootDescriptorTable(0,
CD3DX12_GPU_DESCRIPTOR_HANDLE(getCurrentResourceStorage().descriptors_heap->GetGPUDescriptorHandleForHeapStart())
.Offset((INT)currentDescriptorIndex, g_descriptorStrideSRVCBVUAV)
);
getCurrentResourceStorage().command_list->SetGraphicsRootDescriptorTable(1,
CD3DX12_GPU_DESCRIPTOR_HANDLE(getCurrentResourceStorage().sampler_descriptor_heap[getCurrentResourceStorage().sampler_descriptors_heap_index]->GetGPUDescriptorHandleForHeapStart())
.Offset((INT)getCurrentResourceStorage().current_sampler_index, g_descriptorStrideSamplers)
);
getCurrentResourceStorage().current_sampler_index += std::get<2>(*m_PSO);
getCurrentResourceStorage().descriptors_heap_index += std::get<2>(*m_PSO) + 3;
}
else
{
getCurrentResourceStorage().command_list->SetDescriptorHeaps(1, getCurrentResourceStorage().descriptors_heap.GetAddressOf());
getCurrentResourceStorage().command_list->SetGraphicsRootDescriptorTable(0,
CD3DX12_GPU_DESCRIPTOR_HANDLE(getCurrentResourceStorage().descriptors_heap->GetGPUDescriptorHandleForHeapStart())
.Offset((INT)currentDescriptorIndex, g_descriptorStrideSRVCBVUAV)
);
getCurrentResourceStorage().descriptors_heap_index += 3;
}
std::chrono::time_point<std::chrono::system_clock> texture_duration_end = std::chrono::system_clock::now();
m_timers.m_textureDuration += std::chrono::duration_cast<std::chrono::microseconds>(texture_duration_end - texture_duration_start).count();
set_rtt_and_ds(getCurrentResourceStorage().command_list.Get());
int clip_w = rsx::method_registers[NV4097_SET_SURFACE_CLIP_HORIZONTAL] >> 16;
int clip_h = rsx::method_registers[NV4097_SET_SURFACE_CLIP_VERTICAL] >> 16;
D3D12_VIEWPORT viewport =
{
0.f,
0.f,
(float)clip_w,
(float)clip_h,
-1.f,
1.f
};
getCurrentResourceStorage().command_list->RSSetViewports(1, &viewport);
D3D12_RECT box =
{
0,
0,
(LONG)clip_w,
(LONG)clip_h,
};
getCurrentResourceStorage().command_list->RSSetScissorRects(1, &box);
getCurrentResourceStorage().command_list->IASetPrimitiveTopology(get_primitive_topology(draw_mode));
if (m_renderingInfo.m_indexed)
getCurrentResourceStorage().command_list->DrawIndexedInstanced((UINT)m_renderingInfo.m_count, 1, 0, 0, 0);
else
getCurrentResourceStorage().command_list->DrawInstanced((UINT)m_renderingInfo.m_count, 1, 0, 0);
vertex_index_array.clear();
std::chrono::time_point<std::chrono::system_clock> end_duration = std::chrono::system_clock::now();
m_timers.m_drawCallDuration += std::chrono::duration_cast<std::chrono::microseconds>(end_duration - start_duration).count();
m_timers.m_drawCallCount++;
if (rpcs3::config.rsx.d3d12.debug_output.value())
{
ThrowIfFailed(getCurrentResourceStorage().command_list->Close());
m_commandQueueGraphic->ExecuteCommandLists(1, (ID3D12CommandList**)getCurrentResourceStorage().command_list.GetAddressOf());
getCurrentResourceStorage().set_new_command_list();
}
m_first_count_pairs.clear();
m_renderingInfo.m_indexed = false;
thread::end();
}