nintendo-back-up/yuzu
1
0
Fork 0
mirror of https://git.zaroz.cloud/nintendo-back-up/yuzu/yuzu.git synced 2025-05-12 00:45:25 +00:00
Code Issues Actions Packages Projects Releases 1 Wiki Activity

Optimize the vertex loader, nearly doubling its speed.

Browse Source
This commit is contained in:
Henrik Rydgard 2016-04-28 21:43:49 +02:00
parent 2403e86cbb
commit 251f29dd7f
2 changed files with 54 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

85
src/video_core/vertex_loader.cpp
View File

@ -46,13 +46,11 @@ void VertexLoader::Setup(const Pica::Regs& regs) {
u32 attribute_index = loader_config.GetComponent(component); u32 attribute_index = loader_config.GetComponent(component);
if (attribute_index < 12) { if (attribute_index < 12) {
int element_size = attribute_config.GetElementSizeInBytes(attribute_index); offset = Common::AlignUp(offset, attribute_config.GetElementSizeInBytes(attribute_index));
offset = Common::AlignUp(offset, element_size);
vertex_attribute_sources[attribute_index] = loader_config.data_offset + offset; vertex_attribute_sources[attribute_index] = loader_config.data_offset + offset;
vertex_attribute_strides[attribute_index] = static_cast<u32>(loader_config.byte_count); vertex_attribute_strides[attribute_index] = static_cast<u32>(loader_config.byte_count);
vertex_attribute_formats[attribute_index] = attribute_config.GetFormat(attribute_index); vertex_attribute_formats[attribute_index] = attribute_config.GetFormat(attribute_index);
vertex_attribute_elements[attribute_index] = attribute_config.GetNumElements(attribute_index); vertex_attribute_elements[attribute_index] = attribute_config.GetNumElements(attribute_index);
vertex_attribute_element_size[attribute_index] = element_size;
offset += attribute_config.GetStride(attribute_index); offset += attribute_config.GetStride(attribute_index);
} else if (attribute_index < 16) { } else if (attribute_index < 16) {
// Attribute ids 12, 13, 14 and 15 signify 4, 8, 12 and 16-byte paddings, respectively // Attribute ids 12, 13, 14 and 15 signify 4, 8, 12 and 16-byte paddings, respectively
@ -68,38 +66,63 @@ void VertexLoader::Setup(const Pica::Regs& regs) {
void VertexLoader::LoadVertex(u32 base_address, int index, int vertex, Shader::InputVertex& input, MemoryAccesses& memory_accesses) { void VertexLoader::LoadVertex(u32 base_address, int index, int vertex, Shader::InputVertex& input, MemoryAccesses& memory_accesses) {
for (int i = 0; i < num_total_attributes; ++i) { for (int i = 0; i < num_total_attributes; ++i) {
if (vertex_attribute_elements[i] != 0) { if (vertex_attribute_elements[i] != 0) {
// Load per-vertex data from the loader arrays
u32 source_addr = base_address + vertex_attribute_sources[i] + vertex_attribute_strides[i] * vertex;
if (g_debug_context && Pica::g_debug_context->recorder) {
memory_accesses.AddAccess(source_addr,
(vertex_attribute_formats[i] == Regs::VertexAttributeFormat::FLOAT) ? 4
: (vertex_attribute_formats[i] == Regs::VertexAttributeFormat::SHORT) ? 2 : 1);
}
switch (vertex_attribute_formats[i]) {
case Regs::VertexAttributeFormat::BYTE:
{
const s8* srcdata = reinterpret_cast<const s8*>(Memory::GetPhysicalPointer(source_addr));
for (unsigned int comp = 0; comp < vertex_attribute_elements[i]; ++comp) {
input.attr[i][comp] = float24::FromFloat32(srcdata[comp]);
}
break;
}
case Regs::VertexAttributeFormat::UBYTE:
{
const u8* srcdata = reinterpret_cast<const u8*>(Memory::GetPhysicalPointer(source_addr));
for (unsigned int comp = 0; comp < vertex_attribute_elements[i]; ++comp) {
input.attr[i][comp] = float24::FromFloat32(srcdata[comp]);
}
break;
}
case Regs::VertexAttributeFormat::SHORT:
{
const s16* srcdata = reinterpret_cast<const s16*>(Memory::GetPhysicalPointer(source_addr));
for (unsigned int comp = 0; comp < vertex_attribute_elements[i]; ++comp) {
input.attr[i][comp] = float24::FromFloat32(srcdata[comp]);
}
break;
}
case Regs::VertexAttributeFormat::FLOAT:
{
const float* srcdata = reinterpret_cast<const float*>(Memory::GetPhysicalPointer(source_addr));
for (unsigned int comp = 0; comp < vertex_attribute_elements[i]; ++comp) {
input.attr[i][comp] = float24::FromFloat32(srcdata[comp]);
}
break;
}
}
// Default attribute values set if array elements have < 4 components. This // Default attribute values set if array elements have < 4 components. This
// is *not* carried over from the default attribute settings even if they're // is *not* carried over from the default attribute settings even if they're
// enabled for this attribute. // enabled for this attribute.
static const float24 zero = float24::FromFloat32(0.0f); for (unsigned int comp = vertex_attribute_elements[i]; comp < 4; ++comp) {
static const float24 one = float24::FromFloat32(1.0f); input.attr[i][comp] = comp == 3 ? float24::FromFloat32(1.0f) : float24::FromFloat32(0.0f);
input.attr[i] = Math::Vec4<float24>(zero, zero, zero, one);
// Load per-vertex data from the loader arrays
for (unsigned int comp = 0; comp < vertex_attribute_elements[i]; ++comp) {
u32 source_addr = base_address + vertex_attribute_sources[i] + vertex_attribute_strides[i] * vertex + comp * vertex_attribute_element_size[i];
const u8* srcdata = Memory::GetPhysicalPointer(source_addr);
if (g_debug_context && Pica::g_debug_context->recorder) {
memory_accesses.AddAccess(source_addr,
(vertex_attribute_formats[i] == Regs::VertexAttributeFormat::FLOAT) ? 4
: (vertex_attribute_formats[i] == Regs::VertexAttributeFormat::SHORT) ? 2 : 1);
}
const float srcval =
(vertex_attribute_formats[i] == Regs::VertexAttributeFormat::BYTE) ? *reinterpret_cast<const s8*>(srcdata) :
(vertex_attribute_formats[i] == Regs::VertexAttributeFormat::UBYTE) ? *reinterpret_cast<const u8*>(srcdata) :
(vertex_attribute_formats[i] == Regs::VertexAttributeFormat::SHORT) ? *reinterpret_cast<const s16*>(srcdata) :
*reinterpret_cast<const float*>(srcdata);
input.attr[i][comp] = float24::FromFloat32(srcval);
LOG_TRACE(HW_GPU, "Loaded component %x of attribute %x for vertex %x (index %x) from 0x%08x + 0x%08x + 0x%04x: %f",
comp, i, vertex, index,
base_address,
vertex_attribute_sources[i],
vertex_attribute_strides[i] * vertex + comp * vertex_attribute_element_size[i],
input.attr[i][comp].ToFloat32());
} }
LOG_TRACE(HW_GPU, "Loaded %d components of attribute %x for vertex %x (index %x) from 0x%08x + 0x%08x + 0x%04x: %f %f %f %f",
vertex_attribute_elements[i], i, vertex, index,
base_address,
vertex_attribute_sources[i],
vertex_attribute_strides[i] * vertex,
input.attr[i][0].ToFloat32(), input.attr[i][1].ToFloat32(), input.attr[i][2].ToFloat32(), input.attr[i][3].ToFloat32());
} else if (vertex_attribute_is_default[i]) { } else if (vertex_attribute_is_default[i]) {
// Load the default attribute if we're configured to do so // Load the default attribute if we're configured to do so
input.attr[i] = g_state.vs.default_attributes[i]; input.attr[i] = g_state.vs.default_attributes[i];

1
src/video_core/vertex_loader.h
View File

@ -47,7 +47,6 @@ private:
u32 vertex_attribute_strides[16] = {}; u32 vertex_attribute_strides[16] = {};
Regs::VertexAttributeFormat vertex_attribute_formats[16] = {}; Regs::VertexAttributeFormat vertex_attribute_formats[16] = {};
u32 vertex_attribute_elements[16] = {}; u32 vertex_attribute_elements[16] = {};
u32 vertex_attribute_element_size[16] = {};
bool vertex_attribute_is_default[16]; bool vertex_attribute_is_default[16];
int num_total_attributes; int num_total_attributes;
}; };