fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[Vulkan] 32-bit index fetch without fullDrawIndexUint32

Browse Source
This commit is contained in:
Triang3l
2022-07-25 16:53:12 +03:00
parent 37579d3bf0
commit 77e85ecaa4
6 changed files with 170 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

78
src/xenia/gpu/vulkan/vulkan_command_processor.cc
View File

@@ -2383,9 +2383,10 @@ bool VulkanCommandProcessor::IssueDraw(xenos::PrimitiveType prim_type,
normalized_depth_control);
// Update system constants before uploading them.
UpdateSystemConstantValues(primitive_polygonal,
primitive_processing_result.host_index_endian,
viewport_info, used_texture_mask);
bool vertex_shader_index_load;
UpdateSystemConstantValues(primitive_polygonal, primitive_processing_result,
viewport_info, used_texture_mask,
vertex_shader_index_load);
// Update uniform buffers and descriptor sets after binding the pipeline with
// the new layout.
@@ -2451,7 +2452,8 @@ bool VulkanCommandProcessor::IssueDraw(xenos::PrimitiveType prim_type,
// Draw.
if (primitive_processing_result.index_buffer_type ==
PrimitiveProcessor::ProcessedIndexBufferType::kNone) {
PrimitiveProcessor::ProcessedIndexBufferType::kNone ||
vertex_shader_index_load) {
deferred_command_buffer_.CmdVkDraw(
primitive_processing_result.host_draw_vertex_count, 1, 0, 0);
} else {
@@ -3338,8 +3340,10 @@ void VulkanCommandProcessor::UpdateDynamicState(
}
void VulkanCommandProcessor::UpdateSystemConstantValues(
bool primitive_polygonal, xenos::Endian index_endian,
const draw_util::ViewportInfo& viewport_info, uint32_t used_texture_mask) {
bool primitive_polygonal,
const PrimitiveProcessor::ProcessingResult& primitive_processing_result,
const draw_util::ViewportInfo& viewport_info, uint32_t used_texture_mask,
bool& vertex_shader_index_load_out) {
#if XE_UI_VULKAN_FINE_GRAINED_DRAW_SCOPES
SCOPE_profile_cpu_f("gpu");
#endif // XE_UI_VULKAN_FINE_GRAINED_DRAW_SCOPES
@@ -3362,6 +3366,52 @@ void VulkanCommandProcessor::UpdateSystemConstantValues(
// Flags.
uint32_t flags = 0;
// Vertex index shader loading.
bool vertex_shader_index_load = false;
// Only for ProcessedIndexBufferType kGuest since kHostConverted indices may
// be not loaded into the GPU memory (only read on the CPU), though
// kHostConverted must never be used for point lists and rectangle lists
// without geometry shaders anyway. For regular 32-bit index fetching without
// fullDrawIndexUint32, kHostConverted indices are already byte-swapped and
// truncated to 24 bits, so indirect fetch is not needed.
if (primitive_processing_result.index_buffer_type ==
PrimitiveProcessor::ProcessedIndexBufferType::kGuest) {
switch (primitive_processing_result.host_vertex_shader_type) {
case Shader::HostVertexShaderType::kVertex: {
// For guest (usually big-endian) 32-bit indices when they're not
// supported by the device.
if (vgt_draw_initiator.index_size == xenos::IndexFormat::kInt32) {
const ui::vulkan::VulkanProvider& provider = GetVulkanProvider();
const VkPhysicalDeviceFeatures& device_features =
provider.device_features();
if (!device_features.fullDrawIndexUint32) {
vertex_shader_index_load = true;
flags |= SpirvShaderTranslator::kSysFlag_VertexIndexLoad;
}
}
} break;
// kMemexportCompute never comes out of the PrimitiveProcessor, as
// memexport compute shaders are executed alongside their vertex
// counterparts, since they may still result in drawing.
case Shader::HostVertexShaderType::kPointListAsTriangleStrip:
case Shader::HostVertexShaderType::kRectangleListAsTriangleStrip: {
// Always loading the guest index buffer indirectly if it's used, as
// host indexing contains a part needed specifically for the host for
// the construction of the primitive - host vertices don't map 1:1 to
// guest ones.
vertex_shader_index_load = true;
flags |=
SpirvShaderTranslator::kSysFlag_ComputeOrPrimitiveVertexIndexLoad;
if (vgt_draw_initiator.index_size == xenos::IndexFormat::kInt32) {
flags |= SpirvShaderTranslator ::
kSysFlag_ComputeOrPrimitiveVertexIndexLoad32Bit;
}
} break;
default:
break;
}
}
vertex_shader_index_load_out = vertex_shader_index_load;
// W0 division control.
// http://www.x.org/docs/AMD/old/evergreen_3D_registers_v2.pdf
// 8: VTX_XY_FMT = true: the incoming XY have already been multiplied by 1/W0.
@@ -3404,9 +3454,21 @@ void VulkanCommandProcessor::UpdateSystemConstantValues(
dirty |= system_constants_.flags != flags;
system_constants_.flags = flags;
// Index buffer address for loading in the shaders.
if (flags &
(SpirvShaderTranslator::kSysFlag_VertexIndexLoad |
SpirvShaderTranslator::kSysFlag_ComputeOrPrimitiveVertexIndexLoad)) {
dirty |= system_constants_.vertex_index_load_address !=
primitive_processing_result.guest_index_base;
system_constants_.vertex_index_load_address =
primitive_processing_result.guest_index_base;
}
// Index or tessellation edge factor buffer endianness.
dirty |= system_constants_.vertex_index_endian != index_endian;
system_constants_.vertex_index_endian = index_endian;
dirty |= system_constants_.vertex_index_endian !=
primitive_processing_result.host_index_endian;
system_constants_.vertex_index_endian =
primitive_processing_result.host_index_endian;
// Vertex index offset.
dirty |= system_constants_.vertex_base_index != vgt_indx_offset;