Videocore: Implement simple vertex caching

This gives a ~2/3 reduction in the amount of vertices that need to be
processed through the vertex loaders and the vertex shader, yielding a
good speedup.
This commit is contained in:
Yuri Kunde Schlesner 2015-07-25 03:19:17 -03:00
parent 4d086a4db4
commit a96502edd3

View file

@ -206,88 +206,115 @@ static inline void WritePicaReg(u32 id, u32 value, u32 mask) {
std::map<u32, u32> ranges; std::map<u32, u32> ranges;
} memory_accesses; } memory_accesses;
// Simple circular-replacement vertex cache
// The size has been tuned for optimal balance between hit-rate and the cost of lookup
const size_t VERTEX_CACHE_SIZE = 32;
std::array<u16, VERTEX_CACHE_SIZE> vertex_cache_ids;
std::array<VertexShader::OutputVertex, VERTEX_CACHE_SIZE> vertex_cache;
unsigned int vertex_cache_pos = 0;
vertex_cache_ids.fill(-1);
for (unsigned int index = 0; index < regs.num_vertices; ++index) for (unsigned int index = 0; index < regs.num_vertices; ++index)
{ {
unsigned int vertex = is_indexed ? (index_u16 ? index_address_16[index] : index_address_8[index]) : index; unsigned int vertex = is_indexed ? (index_u16 ? index_address_16[index] : index_address_8[index]) : index;
// -1 is a common special value used for primitive restart. Since it's unknown if
// the PICA supports it, and it would mess up the caching, guard against it here.
ASSERT(vertex != -1);
bool vertex_cache_hit = false;
VertexShader::OutputVertex output;
if (is_indexed) { if (is_indexed) {
// TODO: Implement some sort of vertex cache!
if (g_debug_context && Pica::g_debug_context->recorder) { if (g_debug_context && Pica::g_debug_context->recorder) {
int size = index_u16 ? 2 : 1; int size = index_u16 ? 2 : 1;
memory_accesses.AddAccess(base_address + index_info.offset + size * index, size); memory_accesses.AddAccess(base_address + index_info.offset + size * index, size);
} }
}
// Initialize data for the current vertex for (unsigned int i = 0; i < VERTEX_CACHE_SIZE; ++i) {
VertexShader::InputVertex input; if (vertex == vertex_cache_ids[i]) {
output = vertex_cache[i];
for (int i = 0; i < attribute_config.GetNumTotalAttributes(); ++i) { vertex_cache_hit = true;
if (vertex_attribute_elements[i] != 0) { break;
// Default attribute values set if array elements have < 4 components. This
// is *not* carried over from the default attribute settings even if they're
// enabled for this attribute.
static const float24 zero = float24::FromFloat32(0.0f);
static const float24 one = float24::FromFloat32(1.0f);
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 = 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) ? *(s8*)srcdata :
(vertex_attribute_formats[i] == Regs::VertexAttributeFormat::UBYTE) ? *(u8*)srcdata :
(vertex_attribute_formats[i] == Regs::VertexAttributeFormat::SHORT) ? *(s16*)srcdata :
*(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%08lx + 0x%04lx: %f",
comp, i, vertex, index,
attribute_config.GetPhysicalBaseAddress(),
vertex_attribute_sources[i] - base_address,
vertex_attribute_strides[i] * vertex + comp * vertex_attribute_element_size[i],
input.attr[i][comp].ToFloat32());
} }
} else if (attribute_config.IsDefaultAttribute(i)) {
// Load the default attribute if we're configured to do so
input.attr[i] = g_state.vs.default_attributes[i];
LOG_TRACE(HW_GPU, "Loaded default attribute %x for vertex %x (index %x): (%f, %f, %f, %f)",
i, vertex, index,
input.attr[i][0].ToFloat32(), input.attr[i][1].ToFloat32(),
input.attr[i][2].ToFloat32(), input.attr[i][3].ToFloat32());
} else {
// TODO(yuriks): In this case, no data gets loaded and the vertex remains
// with the last value it had. This isn't currently maintained
// as global state, however, and so won't work in Cita yet.
} }
} }
if (g_debug_context) if (!vertex_cache_hit) {
g_debug_context->OnEvent(DebugContext::Event::VertexLoaded, (void*)&input); // Initialize data for the current vertex
VertexShader::InputVertex input;
for (int i = 0; i < attribute_config.GetNumTotalAttributes(); ++i) {
if (vertex_attribute_elements[i] != 0) {
// Default attribute values set if array elements have < 4 components. This
// is *not* carried over from the default attribute settings even if they're
// enabled for this attribute.
static const float24 zero = float24::FromFloat32(0.0f);
static const float24 one = float24::FromFloat32(1.0f);
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 = 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) ? *(s8*)srcdata :
(vertex_attribute_formats[i] == Regs::VertexAttributeFormat::UBYTE) ? *(u8*)srcdata :
(vertex_attribute_formats[i] == Regs::VertexAttributeFormat::SHORT) ? *(s16*)srcdata :
*(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%08lx + 0x%04lx: %f",
comp, i, vertex, index,
attribute_config.GetPhysicalBaseAddress(),
vertex_attribute_sources[i] - base_address,
vertex_attribute_strides[i] * vertex + comp * vertex_attribute_element_size[i],
input.attr[i][comp].ToFloat32());
}
} else if (attribute_config.IsDefaultAttribute(i)) {
// Load the default attribute if we're configured to do so
input.attr[i] = g_state.vs.default_attributes[i];
LOG_TRACE(HW_GPU, "Loaded default attribute %x for vertex %x (index %x): (%f, %f, %f, %f)",
i, vertex, index,
input.attr[i][0].ToFloat32(), input.attr[i][1].ToFloat32(),
input.attr[i][2].ToFloat32(), input.attr[i][3].ToFloat32());
} else {
// TODO(yuriks): In this case, no data gets loaded and the vertex
// remains with the last value it had. This isn't currently maintained
// as global state, however, and so won't work in Citra yet.
}
}
if (g_debug_context)
g_debug_context->OnEvent(DebugContext::Event::VertexLoaded, (void*)&input);
#if PICA_DUMP_GEOMETRY #if PICA_DUMP_GEOMETRY
// NOTE: When dumping geometry, we simply assume that the first input attribute // NOTE: When dumping geometry, we simply assume that the first input attribute
// corresponds to the position for now. // corresponds to the position for now.
DebugUtils::GeometryDumper::Vertex dumped_vertex = { DebugUtils::GeometryDumper::Vertex dumped_vertex = {
input.attr[0][0].ToFloat32(), input.attr[0][1].ToFloat32(), input.attr[0][2].ToFloat32() input.attr[0][0].ToFloat32(), input.attr[0][1].ToFloat32(), input.attr[0][2].ToFloat32()
}; };
using namespace std::placeholders; using namespace std::placeholders;
dumping_primitive_assembler.SubmitVertex(dumped_vertex, dumping_primitive_assembler.SubmitVertex(dumped_vertex,
std::bind(&DebugUtils::GeometryDumper::AddTriangle, std::bind(&DebugUtils::GeometryDumper::AddTriangle,
&geometry_dumper, _1, _2, _3)); &geometry_dumper, _1, _2, _3));
#endif #endif
// Send to vertex shader // Send to vertex shader
VertexShader::OutputVertex output = VertexShader::RunShader(input, attribute_config.GetNumTotalAttributes(), g_state.regs.vs, g_state.vs); output = VertexShader::RunShader(input, attribute_config.GetNumTotalAttributes(), g_state.regs.vs, g_state.vs);
if (is_indexed) { if (is_indexed) {
// TODO: Add processed vertex to vertex cache! vertex_cache[vertex_cache_pos] = output;
vertex_cache_ids[vertex_cache_pos] = vertex;
vertex_cache_pos = (vertex_cache_pos + 1) % VERTEX_CACHE_SIZE;
}
} }
if (Settings::values.use_hw_renderer) { if (Settings::values.use_hw_renderer) {