Revert Buffer cache changes and setup additional macros.

This commit is contained in:
Fernando Sahmkow 2022-11-18 00:21:13 +01:00
parent 18637766ef
commit ce448ce770
7 changed files with 179 additions and 128 deletions

View file

@ -170,11 +170,6 @@ public:
void BindComputeTextureBuffer(size_t tbo_index, GPUVAddr gpu_addr, u32 size, PixelFormat format,
bool is_written, bool is_image);
[[nodiscard]] std::pair<Buffer*, u32> ObtainBuffer(GPUVAddr gpu_addr, u32 size,
bool synchronize = true,
bool mark_as_written = false,
bool discard_downloads = false);
void FlushCachedWrites();
/// Return true when there are uncommitted buffers to be downloaded
@ -354,8 +349,6 @@ private:
bool SynchronizeBufferImpl(Buffer& buffer, VAddr cpu_addr, u32 size);
bool SynchronizeBufferNoModified(Buffer& buffer, VAddr cpu_addr, u32 size);
void UploadMemory(Buffer& buffer, u64 total_size_bytes, u64 largest_copy,
std::span<BufferCopy> copies);
@ -442,7 +435,6 @@ private:
std::vector<BufferId> cached_write_buffer_ids;
IntervalSet discarded_ranges;
IntervalSet uncommitted_ranges;
IntervalSet common_ranges;
std::deque<IntervalSet> committed_ranges;
@ -600,17 +592,13 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
}};
boost::container::small_vector<IntervalType, 4> tmp_intervals;
const bool is_high_accuracy =
Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::High;
auto mirror = [&](VAddr base_address, VAddr base_address_end) {
const u64 size = base_address_end - base_address;
const VAddr diff = base_address - *cpu_src_address;
const VAddr new_base_address = *cpu_dest_address + diff;
const IntervalType add_interval{new_base_address, new_base_address + size};
tmp_intervals.push_back(add_interval);
if (is_high_accuracy) {
uncommitted_ranges.add(add_interval);
}
tmp_intervals.push_back(add_interval);
};
ForEachWrittenRange(*cpu_src_address, amount, mirror);
// This subtraction in this order is important for overlapping copies.
@ -821,32 +809,6 @@ void BufferCache<P>::BindComputeTextureBuffer(size_t tbo_index, GPUVAddr gpu_add
compute_texture_buffers[tbo_index] = GetTextureBufferBinding(gpu_addr, size, format);
}
template <class P>
std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainBuffer(GPUVAddr gpu_addr, u32 size,
bool synchronize,
bool mark_as_written,
bool discard_downloads) {
const std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr);
if (!cpu_addr) {
return {&slot_buffers[NULL_BUFFER_ID], 0};
}
const BufferId buffer_id = FindBuffer(*cpu_addr, size);
Buffer& buffer = slot_buffers[buffer_id];
if (synchronize) {
// SynchronizeBuffer(buffer, *cpu_addr, size);
SynchronizeBufferNoModified(buffer, *cpu_addr, size);
}
if (mark_as_written) {
MarkWrittenBuffer(buffer_id, *cpu_addr, size);
}
if (discard_downloads) {
IntervalType interval{*cpu_addr, size};
ClearDownload(interval);
discarded_ranges.subtract(interval);
}
return {&buffer, buffer.Offset(*cpu_addr)};
}
template <class P>
void BufferCache<P>::FlushCachedWrites() {
for (const BufferId buffer_id : cached_write_buffer_ids) {
@ -862,6 +824,10 @@ bool BufferCache<P>::HasUncommittedFlushes() const noexcept {
template <class P>
void BufferCache<P>::AccumulateFlushes() {
if (Settings::values.gpu_accuracy.GetValue() != Settings::GPUAccuracy::High) {
uncommitted_ranges.clear();
return;
}
if (uncommitted_ranges.empty()) {
return;
}
@ -877,14 +843,12 @@ template <class P>
void BufferCache<P>::CommitAsyncFlushesHigh() {
AccumulateFlushes();
for (const auto& interval : discarded_ranges) {
common_ranges.subtract(interval);
}
if (committed_ranges.empty()) {
return;
}
MICROPROFILE_SCOPE(GPU_DownloadMemory);
const bool is_accuracy_normal =
Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::Normal;
auto it = committed_ranges.begin();
while (it != committed_ranges.end()) {
@ -909,6 +873,9 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
ForEachBufferInRange(cpu_addr, size, [&](BufferId buffer_id, Buffer& buffer) {
buffer.ForEachDownloadRangeAndClear(
cpu_addr, size, [&](u64 range_offset, u64 range_size) {
if (is_accuracy_normal) {
return;
}
const VAddr buffer_addr = buffer.CpuAddr();
const auto add_download = [&](VAddr start, VAddr end) {
const u64 new_offset = start - buffer_addr;
@ -973,7 +940,12 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
template <class P>
void BufferCache<P>::CommitAsyncFlushes() {
if (Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::High) {
CommitAsyncFlushesHigh();
} else {
uncommitted_ranges.clear();
committed_ranges.clear();
}
}
template <class P>
@ -1353,7 +1325,7 @@ void BufferCache<P>::UpdateIndexBuffer() {
const auto& draw_state = maxwell3d->draw_manager->GetDrawState();
const auto& index_array = draw_state.index_buffer;
auto& flags = maxwell3d->dirty.flags;
if (!flags[Dirty::IndexBuffer] && last_index_count == index_array.count) {
if (!flags[Dirty::IndexBuffer]) {
return;
}
flags[Dirty::IndexBuffer] = false;
@ -1574,11 +1546,7 @@ void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 s
if (!is_async) {
return;
}
const bool is_high_accuracy =
Settings::values.gpu_accuracy.GetValue() == Settings::GPUAccuracy::High;
if (is_high_accuracy) {
uncommitted_ranges.add(base_interval);
}
}
template <class P>
@ -1771,51 +1739,6 @@ bool BufferCache<P>::SynchronizeBufferImpl(Buffer& buffer, VAddr cpu_addr, u32 s
return false;
}
template <class P>
bool BufferCache<P>::SynchronizeBufferNoModified(Buffer& buffer, VAddr cpu_addr, u32 size) {
boost::container::small_vector<BufferCopy, 4> copies;
u64 total_size_bytes = 0;
u64 largest_copy = 0;
IntervalSet found_sets{};
auto make_copies = [&] {
for (auto& interval : found_sets) {
const std::size_t sub_size = interval.upper() - interval.lower();
const VAddr cpu_addr = interval.lower();
copies.push_back(BufferCopy{
.src_offset = total_size_bytes,
.dst_offset = cpu_addr - buffer.CpuAddr(),
.size = sub_size,
});
total_size_bytes += sub_size;
largest_copy = std::max(largest_copy, sub_size);
}
const std::span<BufferCopy> copies_span(copies.data(), copies.size());
UploadMemory(buffer, total_size_bytes, largest_copy, copies_span);
};
buffer.ForEachUploadRange(cpu_addr, size, [&](u64 range_offset, u64 range_size) {
const VAddr base_adr = buffer.CpuAddr() + range_offset;
const VAddr end_adr = base_adr + range_size;
const IntervalType add_interval{base_adr, end_adr};
found_sets.add(add_interval);
});
if (found_sets.empty()) {
return true;
}
const IntervalType search_interval{cpu_addr, cpu_addr + size};
auto it = common_ranges.lower_bound(search_interval);
auto it_end = common_ranges.upper_bound(search_interval);
if (it == common_ranges.end()) {
make_copies();
return false;
}
while (it != it_end) {
found_sets.subtract(*it);
it++;
}
make_copies();
return false;
}
template <class P>
void BufferCache<P>::UploadMemory(Buffer& buffer, u64 total_size_bytes, u64 largest_copy,
std::span<BufferCopy> copies) {

View file

@ -77,12 +77,21 @@ bool DmaPusher::Step() {
command_headers.resize_destructive(command_list_header.size);
constexpr u32 MacroRegistersStart = 0xE00;
if (dma_state.method < MacroRegistersStart) {
if (Settings::IsGPULevelHigh()) {
memory_manager.ReadBlock(dma_state.dma_get, command_headers.data(),
command_list_header.size * sizeof(u32));
} else {
memory_manager.ReadBlockUnsafe(dma_state.dma_get, command_headers.data(),
command_list_header.size * sizeof(u32));
}
} else {
const size_t copy_size = command_list_header.size * sizeof(u32);
if (subchannels[dma_state.subchannel]) {
subchannels[dma_state.subchannel]->current_dirty =
memory_manager.IsMemoryDirty(dma_state.dma_get, copy_size);
}
memory_manager.ReadBlockUnsafe(dma_state.dma_get, command_headers.data(), copy_size);
}
ProcessCommands(command_headers);
}

View file

@ -18,6 +18,7 @@ public:
virtual void CallMultiMethod(u32 method, const u32* base_start, u32 amount,
u32 methods_pending) = 0;
bool current_dirty{};
GPUVAddr current_dma_segment;
};

View file

@ -4,6 +4,7 @@
#include <cstring>
#include <optional>
#include "common/assert.h"
#include "common/settings.h"
#include "core/core.h"
#include "core/core_timing.h"
#include "video_core/dirty_flags.h"
@ -14,6 +15,7 @@
#include "video_core/rasterizer_interface.h"
#include "video_core/textures/texture.h"
namespace Tegra::Engines {
using VideoCore::QueryType;
@ -134,6 +136,8 @@ void Maxwell3D::ProcessMacro(u32 method, const u32* base_start, u32 amount, bool
macro_addresses.push_back(current_dma_segment + i * sizeof(u32));
}
macro_segments.emplace_back(current_dma_segment, amount);
current_macro_dirty |= current_dirty;
current_dirty = false;
// Call the macro when there are no more parameters in the command buffer
if (is_last_call) {
@ -141,10 +145,14 @@ void Maxwell3D::ProcessMacro(u32 method, const u32* base_start, u32 amount, bool
macro_params.clear();
macro_addresses.clear();
macro_segments.clear();
current_macro_dirty = false;
}
}
void Maxwell3D::RefreshParameters() {
void Maxwell3D::RefreshParametersImpl() {
if (!Settings::IsGPULevelHigh()) {
return;
}
size_t current_index = 0;
for (auto& segment : macro_segments) {
if (segment.first == 0) {
@ -157,21 +165,6 @@ void Maxwell3D::RefreshParameters() {
}
}
bool Maxwell3D::AnyParametersDirty() {
size_t current_index = 0;
for (auto& segment : macro_segments) {
if (segment.first == 0) {
current_index += segment.second;
continue;
}
if (memory_manager.IsMemoryDirty(segment.first, sizeof(u32) * segment.second)) {
return true;
}
current_index += segment.second;
}
return false;
}
u32 Maxwell3D::GetMaxCurrentVertices() {
u32 num_vertices = 0;
for (size_t index = 0; index < Regs::NumVertexArrays; ++index) {
@ -332,7 +325,6 @@ void Maxwell3D::CallMethod(u32 method, u32 method_argument, bool is_last_call) {
const u32 argument = ProcessShadowRam(method, method_argument);
ProcessDirtyRegisters(method, argument);
ProcessMethodCall(method, argument, method_argument, is_last_call);
}

View file

@ -272,6 +272,7 @@ public:
};
union {
u32 raw;
BitField<0, 1, Mode> mode;
BitField<4, 8, u32> pad;
};
@ -1217,10 +1218,12 @@ public:
struct Window {
union {
u32 raw_1;
BitField<0, 16, u32> x_min;
BitField<16, 16, u32> x_max;
};
union {
u32 raw_2;
BitField<0, 16, u32> y_min;
BitField<16, 16, u32> y_max;
};
@ -3090,9 +3093,16 @@ public:
return macro_addresses[index];
}
void RefreshParameters();
void RefreshParameters() {
if (!current_macro_dirty) {
return;
}
RefreshParametersImpl();
}
bool AnyParametersDirty();
bool AnyParametersDirty() {
return current_macro_dirty;
}
u32 GetMaxCurrentVertices();
@ -3101,6 +3111,9 @@ public:
/// Handles a write to the CLEAR_BUFFERS register.
void ProcessClearBuffers(u32 layer_count);
/// Handles a write to the CB_BIND register.
void ProcessCBBind(size_t stage_index);
private:
void InitializeRegisterDefaults();
@ -3154,12 +3167,11 @@ private:
void ProcessCBData(u32 value);
void ProcessCBMultiData(const u32* start_base, u32 amount);
/// Handles a write to the CB_BIND register.
void ProcessCBBind(size_t stage_index);
/// Returns a query's value or an empty object if the value will be deferred through a cache.
std::optional<u64> GetQueryResult();
void RefreshParametersImpl();
Core::System& system;
MemoryManager& memory_manager;
@ -3187,6 +3199,7 @@ private:
bool draw_indexed{};
std::vector<std::pair<GPUVAddr, size_t>> macro_segments;
std::vector<GPUVAddr> macro_addresses;
bool current_macro_dirty{};
};
#define ASSERT_REG_POSITION(field_name, position) \

View file

@ -12,6 +12,7 @@
#include "common/assert.h"
#include "common/fs/fs.h"
#include "common/fs/path_util.h"
#include "common/microprofile.h"
#include "common/settings.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/macro/macro.h"
@ -22,6 +23,8 @@
#include "video_core/macro/macro_jit_x64.h"
#endif
MICROPROFILE_DEFINE(MacroHLE, "GPU", "Execute macro hle", MP_RGB(128, 192, 192));
namespace Tegra {
static void Dump(u64 hash, std::span<const u32> code) {
@ -60,6 +63,7 @@ void MacroEngine::Execute(u32 method, const std::vector<u32>& parameters) {
if (compiled_macro != macro_cache.end()) {
const auto& cache_info = compiled_macro->second;
if (cache_info.has_hle_program) {
MICROPROFILE_SCOPE(MacroHLE);
cache_info.hle_program->Execute(parameters, method);
} else {
maxwell3d.RefreshParameters();
@ -106,6 +110,7 @@ void MacroEngine::Execute(u32 method, const std::vector<u32>& parameters) {
if (auto hle_program = hle_macros->GetHLEProgram(cache_info.hash)) {
cache_info.has_hle_program = true;
cache_info.hle_program = std::move(hle_program);
MICROPROFILE_SCOPE(MacroHLE);
cache_info.hle_program->Execute(parameters, method);
} else {
maxwell3d.RefreshParameters();

View file

@ -86,7 +86,7 @@ public:
void Execute(const std::vector<u32>& parameters, [[maybe_unused]] u32 method) override {
auto topology = static_cast<Maxwell::Regs::PrimitiveTopology>(parameters[0]);
if (!IsTopologySafe(topology)) {
if (!maxwell3d.AnyParametersDirty() || !IsTopologySafe(topology)) {
Fallback(parameters);
return;
}
@ -117,8 +117,8 @@ private:
void Fallback(const std::vector<u32>& parameters) {
SCOPE_EXIT({
if (extended) {
maxwell3d.CallMethod(0x8e3, 0x640, true);
maxwell3d.CallMethod(0x8e4, 0, true);
maxwell3d.engine_state = Maxwell::EngineHint::None;
maxwell3d.replace_table.clear();
}
});
maxwell3d.RefreshParameters();
@ -127,7 +127,8 @@ private:
const u32 vertex_first = parameters[3];
const u32 vertex_count = parameters[1];
if (maxwell3d.GetMaxCurrentVertices() < vertex_first + vertex_count) {
if (maxwell3d.AnyParametersDirty() &&
maxwell3d.GetMaxCurrentVertices() < vertex_first + vertex_count) {
ASSERT_MSG(false, "Faulty draw!");
return;
}
@ -157,7 +158,7 @@ public:
void Execute(const std::vector<u32>& parameters, [[maybe_unused]] u32 method) override {
auto topology = static_cast<Maxwell::Regs::PrimitiveTopology>(parameters[0]);
if (!IsTopologySafe(topology)) {
if (!maxwell3d.AnyParametersDirty() || !IsTopologySafe(topology)) {
Fallback(parameters);
return;
}
@ -169,7 +170,11 @@ public:
}
const u32 estimate = static_cast<u32>(maxwell3d.EstimateIndexBufferSize());
const u32 base_size = std::max<u32>(minimum_limit, estimate);
maxwell3d.regs.draw.topology.Assign(topology);
const u32 element_base = parameters[4];
const u32 base_instance = parameters[5];
maxwell3d.regs.vertex_id_base = element_base;
maxwell3d.regs.global_base_vertex_index = element_base;
maxwell3d.regs.global_base_instance_index = base_instance;
maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
maxwell3d.engine_state = Maxwell::EngineHint::OnHLEMacro;
maxwell3d.setHLEReplacementName(0, 0x640, Maxwell::HLEReplaceName::BaseVertex);
@ -186,6 +191,9 @@ public:
maxwell3d.draw_manager->DrawIndexedIndirect(topology, 0, base_size);
maxwell3d.engine_state = Maxwell::EngineHint::None;
maxwell3d.replace_table.clear();
maxwell3d.regs.vertex_id_base = 0x0;
maxwell3d.regs.global_base_vertex_index = 0x0;
maxwell3d.regs.global_base_instance_index = 0x0;
}
private:
@ -195,6 +203,8 @@ private:
const u32 element_base = parameters[4];
const u32 base_instance = parameters[5];
maxwell3d.regs.vertex_id_base = element_base;
maxwell3d.regs.global_base_vertex_index = element_base;
maxwell3d.regs.global_base_instance_index = base_instance;
maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
maxwell3d.engine_state = Maxwell::EngineHint::OnHLEMacro;
maxwell3d.setHLEReplacementName(0, 0x640, Maxwell::HLEReplaceName::BaseVertex);
@ -205,6 +215,8 @@ private:
parameters[3], parameters[1], element_base, base_instance, instance_count);
maxwell3d.regs.vertex_id_base = 0x0;
maxwell3d.regs.global_base_vertex_index = 0x0;
maxwell3d.regs.global_base_instance_index = 0x0;
maxwell3d.engine_state = Maxwell::EngineHint::None;
maxwell3d.replace_table.clear();
}
@ -253,7 +265,6 @@ public:
return;
}
maxwell3d.regs.draw.topology.Assign(topology);
const u32 padding = parameters[3]; // padding is in words
// size of each indirect segment
@ -335,6 +346,83 @@ private:
u32 minimum_limit{1 << 12};
};
class HLE_C713C83D8F63CCF3 final : public HLEMacroImpl {
public:
explicit HLE_C713C83D8F63CCF3(Engines::Maxwell3D& maxwell3d_) : HLEMacroImpl(maxwell3d_) {}
void Execute(const std::vector<u32>& parameters, [[maybe_unused]] u32 method) override {
maxwell3d.RefreshParameters();
const u32 offset = (parameters[0] & 0x3FFFFFFF) << 2;
const u32 address = maxwell3d.regs.shadow_scratch[24];
auto& const_buffer = maxwell3d.regs.const_buffer;
const_buffer.size = 0x7000;
const_buffer.address_high = (address >> 24) & 0xFF;
const_buffer.address_low = address << 8;
const_buffer.offset = offset;
}
};
class HLE_D7333D26E0A93EDE final : public HLEMacroImpl {
public:
explicit HLE_D7333D26E0A93EDE(Engines::Maxwell3D& maxwell3d_) : HLEMacroImpl(maxwell3d_) {}
void Execute(const std::vector<u32>& parameters, [[maybe_unused]] u32 method) override {
maxwell3d.RefreshParameters();
const size_t index = parameters[0];
const u32 address = maxwell3d.regs.shadow_scratch[42 + index];
const u32 size = maxwell3d.regs.shadow_scratch[47 + index];
auto& const_buffer = maxwell3d.regs.const_buffer;
const_buffer.size = size;
const_buffer.address_high = (address >> 24) & 0xFF;
const_buffer.address_low = address << 8;
}
};
class HLE_BindShader final : public HLEMacroImpl {
public:
explicit HLE_BindShader(Engines::Maxwell3D& maxwell3d_) : HLEMacroImpl(maxwell3d_) {}
void Execute(const std::vector<u32>& parameters, [[maybe_unused]] u32 method) override {
maxwell3d.RefreshParameters();
auto& regs = maxwell3d.regs;
const u32 index = parameters[0];
if ((parameters[1] - regs.shadow_scratch[28 + index]) == 0) {
return;
}
regs.pipelines[index & 0xF].offset = parameters[2];
maxwell3d.dirty.flags[VideoCommon::Dirty::Shaders] = true;
regs.shadow_scratch[28 + index] = parameters[1];
regs.shadow_scratch[34 + index] = parameters[2];
const u32 address = parameters[4];
auto& const_buffer = regs.const_buffer;
const_buffer.size = 0x10000;
const_buffer.address_high = (address >> 24) & 0xFF;
const_buffer.address_low = address << 8;
const size_t bind_group_id = parameters[3] & 0x7F;
auto& bind_group = regs.bind_groups[bind_group_id];
bind_group.raw_config = 0x11;
maxwell3d.ProcessCBBind(bind_group_id);
}
};
class HLE_SetRasterBoundingBox final : public HLEMacroImpl {
public:
explicit HLE_SetRasterBoundingBox(Engines::Maxwell3D& maxwell3d_) : HLEMacroImpl(maxwell3d_) {}
void Execute(const std::vector<u32>& parameters, [[maybe_unused]] u32 method) override {
maxwell3d.RefreshParameters();
const u32 raster_mode = parameters[0];
auto& regs = maxwell3d.regs;
const u32 raster_enabled = maxwell3d.regs.conservative_raster_enable;
const u32 scratch_data = maxwell3d.regs.shadow_scratch[52];
regs.raster_bounding_box.raw = raster_mode & 0xFFFFF00F;
regs.raster_bounding_box.pad.Assign(scratch_data & raster_enabled);
}
};
} // Anonymous namespace
HLEMacro::HLEMacro(Engines::Maxwell3D& maxwell3d_) : maxwell3d{maxwell3d_} {
@ -368,6 +456,26 @@ HLEMacro::HLEMacro(Engines::Maxwell3D& maxwell3d_) : maxwell3d{maxwell3d_} {
[](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_MultiLayerClear>(maxwell3d);
}));
builders.emplace(0xC713C83D8F63CCF3ULL,
std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>(
[](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_C713C83D8F63CCF3>(maxwell3d);
}));
builders.emplace(0xD7333D26E0A93EDEULL,
std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>(
[](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_D7333D26E0A93EDE>(maxwell3d);
}));
builders.emplace(0xEB29B2A09AA06D38ULL,
std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>(
[](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_BindShader>(maxwell3d);
}));
builders.emplace(0xDB1341DBEB4C8AF7ULL,
std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>(
[](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_SetRasterBoundingBox>(maxwell3d);
}));
}
HLEMacro::~HLEMacro() = default;