Merge pull request #1730 from hrydgard/vertex-loader
* Remove late accesses to attribute_config * Refactor: Extract VertexLoader from command_processor.cpp. Preparation for a similar concept to Dolphin or PPSSPP. These can be JIT-ed and cached. * Move "&" to their proper place, add missing includes and make some properly relative. * Don't keep base_address in the loader, it doesn't belong there (with it, the loader can't be cached). * Optimize the vertex loader, nearly doubling its speed. * Debugger fix * Move and rename the MemoryAccesses class to MemoryAccessTracker.
This commit is contained in:
bunnei
commit
90243c56fb
|
|
@ -16,6 +16,7 @@ set(SRCS
|
|||
shader/shader_interpreter.cpp
|
||||
swrasterizer.cpp
|
||||
utils.cpp
|
||||
vertex_loader.cpp
|
||||
video_core.cpp
|
||||
)
|
||||
|
||||
|
|
@ -43,6 +44,7 @@ set(HEADERS
|
|||
shader/shader_interpreter.h
|
||||
swrasterizer.h
|
||||
utils.h
|
||||
vertex_loader.h
|
||||
video_core.h
|
||||
)
|
||||
|
||||
|
|
|
|||
|
|
@ -21,6 +21,7 @@
|
|||
#include "video_core/video_core.h"
|
||||
#include "video_core/debug_utils/debug_utils.h"
|
||||
#include "video_core/shader/shader_interpreter.h"
|
||||
#include "video_core/vertex_loader.h"
|
||||
|
||||
namespace Pica {
|
||||
|
||||
|
|
@ -188,54 +189,14 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
|
|||
#if PICA_LOG_TEV
|
||||
DebugUtils::DumpTevStageConfig(regs.GetTevStages());
|
||||
#endif
|
||||
|
||||
if (g_debug_context)
|
||||
g_debug_context->OnEvent(DebugContext::Event::IncomingPrimitiveBatch, nullptr);
|
||||
|
||||
const auto& attribute_config = regs.vertex_attributes;
|
||||
const u32 base_address = attribute_config.GetPhysicalBaseAddress();
|
||||
|
||||
// Information about internal vertex attributes
|
||||
u32 vertex_attribute_sources[16];
|
||||
boost::fill(vertex_attribute_sources, 0xdeadbeef);
|
||||
u32 vertex_attribute_strides[16] = {};
|
||||
Regs::VertexAttributeFormat vertex_attribute_formats[16] = {};
|
||||
|
||||
u32 vertex_attribute_elements[16] = {};
|
||||
u32 vertex_attribute_element_size[16] = {};
|
||||
|
||||
// Setup attribute data from loaders
|
||||
for (int loader = 0; loader < 12; ++loader) {
|
||||
const auto& loader_config = attribute_config.attribute_loaders[loader];
|
||||
|
||||
u32 offset = 0;
|
||||
|
||||
// TODO: What happens if a loader overwrites a previous one's data?
|
||||
for (unsigned component = 0; component < loader_config.component_count; ++component) {
|
||||
if (component >= 12) {
|
||||
LOG_ERROR(HW_GPU, "Overflow in the vertex attribute loader %u trying to load component %u", loader, component);
|
||||
continue;
|
||||
}
|
||||
|
||||
u32 attribute_index = loader_config.GetComponent(component);
|
||||
if (attribute_index < 12) {
|
||||
int element_size = attribute_config.GetElementSizeInBytes(attribute_index);
|
||||
offset = Common::AlignUp(offset, element_size);
|
||||
vertex_attribute_sources[attribute_index] = base_address + loader_config.data_offset + offset;
|
||||
vertex_attribute_strides[attribute_index] = static_cast<u32>(loader_config.byte_count);
|
||||
vertex_attribute_formats[attribute_index] = attribute_config.GetFormat(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);
|
||||
} else if (attribute_index < 16) {
|
||||
// Attribute ids 12, 13, 14 and 15 signify 4, 8, 12 and 16-byte paddings, respectively
|
||||
offset = Common::AlignUp(offset, 4);
|
||||
offset += (attribute_index - 11) * 4;
|
||||
} else {
|
||||
UNREACHABLE(); // This is truly unreachable due to the number of bits for each component
|
||||
}
|
||||
}
|
||||
}
|
||||
// Processes information about internal vertex attributes to figure out how a vertex is loaded.
|
||||
// Later, these can be compiled and cached.
|
||||
VertexLoader loader;
|
||||
const u32 base_address = regs.vertex_attributes.GetPhysicalBaseAddress();
|
||||
loader.Setup(regs);
|
||||
|
||||
// Load vertices
|
||||
bool is_indexed = (id == PICA_REG_INDEX(trigger_draw_indexed));
|
||||
|
|
@ -259,32 +220,7 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
|
|||
}
|
||||
}
|
||||
|
||||
class {
|
||||
/// Combine overlapping and close ranges
|
||||
void SimplifyRanges() {
|
||||
for (auto it = ranges.begin(); it != ranges.end(); ++it) {
|
||||
// NOTE: We add 32 to the range end address to make sure "close" ranges are combined, too
|
||||
auto it2 = std::next(it);
|
||||
while (it2 != ranges.end() && it->first + it->second + 32 >= it2->first) {
|
||||
it->second = std::max(it->second, it2->first + it2->second - it->first);
|
||||
it2 = ranges.erase(it2);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
public:
|
||||
/// Record a particular memory access in the list
|
||||
void AddAccess(u32 paddr, u32 size) {
|
||||
// Create new range or extend existing one
|
||||
ranges[paddr] = std::max(ranges[paddr], size);
|
||||
|
||||
// Simplify ranges...
|
||||
SimplifyRanges();
|
||||
}
|
||||
|
||||
/// Map of accessed ranges (mapping start address to range size)
|
||||
std::map<u32, u32> ranges;
|
||||
} memory_accesses;
|
||||
DebugUtils::MemoryAccessTracker memory_accesses;
|
||||
|
||||
// Simple circular-replacement vertex cache
|
||||
// The size has been tuned for optimal balance between hit-rate and the cost of lookup
|
||||
|
|
@ -328,60 +264,13 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
|
|||
if (!vertex_cache_hit) {
|
||||
// Initialize data for the current vertex
|
||||
Shader::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) ? *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,
|
||||
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.
|
||||
}
|
||||
}
|
||||
loader.LoadVertex(base_address, index, vertex, input, memory_accesses);
|
||||
|
||||
if (g_debug_context)
|
||||
g_debug_context->OnEvent(DebugContext::Event::VertexLoaded, (void*)&input);
|
||||
|
||||
// Send to vertex shader
|
||||
output = Shader::Run(shader_unit, input, attribute_config.GetNumTotalAttributes());
|
||||
output = Shader::Run(shader_unit, input, loader.GetNumTotalAttributes());
|
||||
|
||||
if (is_indexed) {
|
||||
vertex_cache[vertex_cache_pos] = output;
|
||||
|
|
|
|||
|
|
@ -216,6 +216,36 @@ void DumpTexture(const Pica::Regs::TextureConfig& texture_config, u8* data);
|
|||
|
||||
void DumpTevStageConfig(const std::array<Pica::Regs::TevStageConfig,6>& stages);
|
||||
|
||||
/**
|
||||
* Used in the vertex loader to merge access records. TODO: Investigate if actually useful.
|
||||
*/
|
||||
class MemoryAccessTracker {
|
||||
/// Combine overlapping and close ranges
|
||||
void SimplifyRanges() {
|
||||
for (auto it = ranges.begin(); it != ranges.end(); ++it) {
|
||||
// NOTE: We add 32 to the range end address to make sure "close" ranges are combined, too
|
||||
auto it2 = std::next(it);
|
||||
while (it2 != ranges.end() && it->first + it->second + 32 >= it2->first) {
|
||||
it->second = std::max(it->second, it2->first + it2->second - it->first);
|
||||
it2 = ranges.erase(it2);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
public:
|
||||
/// Record a particular memory access in the list
|
||||
void AddAccess(u32 paddr, u32 size) {
|
||||
// Create new range or extend existing one
|
||||
ranges[paddr] = std::max(ranges[paddr], size);
|
||||
|
||||
// Simplify ranges...
|
||||
SimplifyRanges();
|
||||
}
|
||||
|
||||
/// Map of accessed ranges (mapping start address to range size)
|
||||
std::map<u32, u32> ranges;
|
||||
};
|
||||
|
||||
} // namespace
|
||||
|
||||
} // namespace
|
||||
|
|
|
|||
|
|
@ -25,7 +25,7 @@ namespace Pica {
|
|||
namespace Shader {
|
||||
|
||||
struct InputVertex {
|
||||
Math::Vec4<float24> attr[16];
|
||||
alignas(16) Math::Vec4<float24> attr[16];
|
||||
};
|
||||
|
||||
struct OutputVertex {
|
||||
|
|
|
|||
|
|
@ -0,0 +1,140 @@
|
|||
#include <cmath>
|
||||
#include <string>
|
||||
|
||||
#include "boost/range/algorithm/fill.hpp"
|
||||
|
||||
#include "common/assert.h"
|
||||
#include "common/alignment.h"
|
||||
#include "common/bit_field.h"
|
||||
#include "common/common_funcs.h"
|
||||
#include "common/common_types.h"
|
||||
#include "common/logging/log.h"
|
||||
|
||||
#include "core/memory.h"
|
||||
|
||||
#include "video_core/debug_utils/debug_utils.h"
|
||||
#include "video_core/pica.h"
|
||||
#include "video_core/pica_state.h"
|
||||
#include "video_core/pica_types.h"
|
||||
#include "video_core/vertex_loader.h"
|
||||
|
||||
namespace Pica {
|
||||
|
||||
void VertexLoader::Setup(const Pica::Regs& regs) {
|
||||
const auto& attribute_config = regs.vertex_attributes;
|
||||
num_total_attributes = attribute_config.GetNumTotalAttributes();
|
||||
|
||||
boost::fill(vertex_attribute_sources, 0xdeadbeef);
|
||||
|
||||
for (int i = 0; i < 16; i++) {
|
||||
vertex_attribute_is_default[i] = attribute_config.IsDefaultAttribute(i);
|
||||
}
|
||||
|
||||
// Setup attribute data from loaders
|
||||
for (int loader = 0; loader < 12; ++loader) {
|
||||
const auto& loader_config = attribute_config.attribute_loaders[loader];
|
||||
|
||||
u32 offset = 0;
|
||||
|
||||
// TODO: What happens if a loader overwrites a previous one's data?
|
||||
for (unsigned component = 0; component < loader_config.component_count; ++component) {
|
||||
if (component >= 12) {
|
||||
LOG_ERROR(HW_GPU, "Overflow in the vertex attribute loader %u trying to load component %u", loader, component);
|
||||
continue;
|
||||
}
|
||||
|
||||
u32 attribute_index = loader_config.GetComponent(component);
|
||||
if (attribute_index < 12) {
|
||||
offset = Common::AlignUp(offset, attribute_config.GetElementSizeInBytes(attribute_index));
|
||||
vertex_attribute_sources[attribute_index] = loader_config.data_offset + offset;
|
||||
vertex_attribute_strides[attribute_index] = static_cast<u32>(loader_config.byte_count);
|
||||
vertex_attribute_formats[attribute_index] = attribute_config.GetFormat(attribute_index);
|
||||
vertex_attribute_elements[attribute_index] = attribute_config.GetNumElements(attribute_index);
|
||||
offset += attribute_config.GetStride(attribute_index);
|
||||
} else if (attribute_index < 16) {
|
||||
// Attribute ids 12, 13, 14 and 15 signify 4, 8, 12 and 16-byte paddings, respectively
|
||||
offset = Common::AlignUp(offset, 4);
|
||||
offset += (attribute_index - 11) * 4;
|
||||
} else {
|
||||
UNREACHABLE(); // This is truly unreachable due to the number of bits for each component
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void VertexLoader::LoadVertex(u32 base_address, int index, int vertex, Shader::InputVertex& input, DebugUtils::MemoryAccessTracker& memory_accesses) {
|
||||
for (int i = 0; i < num_total_attributes; ++i) {
|
||||
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_elements[i] * (
|
||||
(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
|
||||
// is *not* carried over from the default attribute settings even if they're
|
||||
// enabled for this attribute.
|
||||
for (unsigned int comp = vertex_attribute_elements[i]; comp < 4; ++comp) {
|
||||
input.attr[i][comp] = comp == 3 ? float24::FromFloat32(1.0f) : float24::FromFloat32(0.0f);
|
||||
}
|
||||
|
||||
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]) {
|
||||
// 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.
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
} // namespace Pica
|
||||
|
|
@ -0,0 +1,28 @@
|
|||
#pragma once
|
||||
|
||||
#include <iterator>
|
||||
#include <algorithm>
|
||||
|
||||
#include "video_core/pica.h"
|
||||
#include "video_core/shader/shader.h"
|
||||
#include "video_core/debug_utils/debug_utils.h"
|
||||
|
||||
namespace Pica {
|
||||
|
||||
class VertexLoader {
|
||||
public:
|
||||
void Setup(const Pica::Regs& regs);
|
||||
void LoadVertex(u32 base_address, int index, int vertex, Shader::InputVertex& input, DebugUtils::MemoryAccessTracker& memory_accesses);
|
||||
|
||||
int GetNumTotalAttributes() const { return num_total_attributes; }
|
||||
|
||||
private:
|
||||
u32 vertex_attribute_sources[16];
|
||||
u32 vertex_attribute_strides[16] = {};
|
||||
Regs::VertexAttributeFormat vertex_attribute_formats[16] = {};
|
||||
u32 vertex_attribute_elements[16] = {};
|
||||
bool vertex_attribute_is_default[16];
|
||||
int num_total_attributes;
|
||||
};
|
||||
|
||||
} // namespace Pica
|
||||
Reference in New Issue