stuff/suyu
stuff/suyu
1
0
Fork 0
mirror of https://git.suyu.dev/suyu/suyu.git synced 2024-11-25 12:56:27 -05:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

MacroHLE: Final cleanup and fixes.

Browse source
This commit is contained in:
Fernando Sahmkow 2022-12-24 19:19:41 -05:00
parent 581a7d785b
commit d09aa0182f
14 changed files with 94 additions and 128 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
src/common/range_map.h
View file

@ -13,8 +13,8 @@ namespace Common {
template <typename KeyTBase, typename ValueT>
class RangeMap {
private:
using KeyT = std::conditional_t<std::is_signed_v<KeyTBase>, typename KeyTBase,
std::make_signed_t<KeyTBase>>;
using KeyT =
std::conditional_t<std::is_signed_v<KeyTBase>, KeyTBase, std::make_signed_t<KeyTBase>>;
public:
explicit RangeMap(ValueT null_value_) : null_value{null_value_} {
@ -56,8 +56,8 @@ public:
private:
using MapType = std::map<KeyT, ValueT>;
using IteratorType = MapType::iterator;
using ConstIteratorType = MapType::const_iterator;
using IteratorType = typename MapType::iterator;
using ConstIteratorType = typename MapType::const_iterator;
size_t ContinousSizeInternal(KeyT address) const {
const auto it = GetFirstElemnentBeforeOrOn(address);

4
src/shader_recompiler/environment.h
View file

@ -36,8 +36,8 @@ public:
[[nodiscard]] virtual bool HasHLEMacroState() const = 0;
[[nodiscard]] virtual std::optional<ReplaceConstant> GetReplaceConstBuffer(
u32 bank, u32 offset) = 0;
[[nodiscard]] virtual std::optional<ReplaceConstant> GetReplaceConstBuffer(u32 bank,
u32 offset) = 0;
virtual void Dump(u64 hash) = 0;

3
src/video_core/buffer_cache/buffer_cache.h
View file

@ -200,7 +200,8 @@ public:
/// Return true when a CPU region is modified from the CPU
[[nodiscard]] bool IsRegionCpuModified(VAddr addr, size_t size);
void SetDrawIndirect(const Tegra::Engines::DrawManager::IndirectParams* current_draw_indirect_) {
void SetDrawIndirect(
const Tegra::Engines::DrawManager::IndirectParams* current_draw_indirect_) {
current_draw_indirect = current_draw_indirect_;
}

3
src/video_core/engines/draw_manager.cpp
View file

@ -97,7 +97,8 @@ void DrawManager::DrawArrayIndirect(PrimitiveTopology topology) {
ProcessDrawIndirect(true);
}
void DrawManager::DrawIndexedIndirect(PrimitiveTopology topology, u32 index_first, u32 index_count) {
void DrawManager::DrawIndexedIndirect(PrimitiveTopology topology, u32 index_first,
u32 index_count) {
const auto& regs{maxwell3d->regs};
draw_state.topology = topology;
draw_state.index_buffer = regs.index_buffer;

98
src/video_core/macro/macro_hle.cpp
View file

@ -47,21 +47,7 @@ public:
explicit HLEMacroImpl(Engines::Maxwell3D& maxwell3d_) : maxwell3d{maxwell3d_} {}
protected:
void advanceCheck() {
current_value = (current_value + 1) % fibonacci_post;
check_limit = current_value == 0;
if (check_limit) {
const u32 new_fibonacci = fibonacci_pre + fibonacci_post;
fibonacci_pre = fibonacci_post;
fibonacci_post = new_fibonacci;
}
}
Engines::Maxwell3D& maxwell3d;
u32 fibonacci_pre{89};
u32 fibonacci_post{144};
u32 current_value{fibonacci_post - 1};
bool check_limit{};
};
class HLE_771BB18C62444DA0 final : public HLEMacroImpl {
@ -124,12 +110,13 @@ private:
maxwell3d.RefreshParameters();
const u32 instance_count = (maxwell3d.GetRegisterValue(0xD1B) & parameters[2]);
auto topology = static_cast<Maxwell::Regs::PrimitiveTopology>(parameters[0]);
const u32 vertex_first = parameters[3];
const u32 vertex_count = parameters[1];
if (maxwell3d.AnyParametersDirty() &&
maxwell3d.GetMaxCurrentVertices() < vertex_first + vertex_count) {
if (!IsTopologySafe(topology) &&
static_cast<size_t>(maxwell3d.GetMaxCurrentVertices()) <
static_cast<size_t>(vertex_first) + static_cast<size_t>(vertex_count)) {
ASSERT_MSG(false, "Faulty draw!");
return;
}
@ -141,9 +128,8 @@ private:
maxwell3d.setHLEReplacementName(0, 0x640, Maxwell::HLEReplaceName::BaseInstance);
}
maxwell3d.draw_manager->DrawArray(
static_cast<Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology>(parameters[0]),
vertex_first, vertex_count, base_instance, instance_count);
maxwell3d.draw_manager->DrawArray(topology, vertex_first, vertex_count, base_instance,
instance_count);
if (extended) {
maxwell3d.regs.global_base_instance_index = 0;
@ -166,13 +152,7 @@ public:
return;
}
advanceCheck();
if (check_limit) {
maxwell3d.RefreshParameters();
minimum_limit = std::max(parameters[3], minimum_limit);
}
const u32 estimate = static_cast<u32>(maxwell3d.EstimateIndexBufferSize());
const u32 base_size = std::max<u32>(minimum_limit, estimate);
const u32 element_base = parameters[4];
const u32 base_instance = parameters[5];
maxwell3d.regs.vertex_id_base = element_base;
@ -191,7 +171,7 @@ public:
params.max_draw_counts = 1;
params.stride = 0;
maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
maxwell3d.draw_manager->DrawIndexedIndirect(topology, 0, base_size);
maxwell3d.draw_manager->DrawIndexedIndirect(topology, 0, estimate);
maxwell3d.engine_state = Maxwell::EngineHint::None;
maxwell3d.replace_table.clear();
maxwell3d.regs.vertex_id_base = 0x0;
@ -223,8 +203,6 @@ private:
maxwell3d.engine_state = Maxwell::EngineHint::None;
maxwell3d.replace_table.clear();
}
u32 minimum_limit{1 << 18};
};
class HLE_MultiLayerClear final : public HLEMacroImpl {
@ -257,10 +235,6 @@ public:
return;
}
advanceCheck();
if (check_limit) {
maxwell3d.RefreshParameters();
}
const u32 start_indirect = parameters[0];
const u32 end_indirect = parameters[1];
if (start_indirect >= end_indirect) {
@ -274,20 +248,7 @@ public:
const u32 indirect_words = 5 + padding;
const u32 stride = indirect_words * sizeof(u32);
const std::size_t draw_count = end_indirect - start_indirect;
u32 lowest_first = std::numeric_limits<u32>::max();
u32 highest_limit = std::numeric_limits<u32>::min();
for (std::size_t index = 0; index < draw_count; index++) {
const std::size_t base = index * indirect_words + 5;
const u32 count = parameters[base];
const u32 first_index = parameters[base + 2];
lowest_first = std::min(lowest_first, first_index);
highest_limit = std::max(highest_limit, first_index + count);
}
if (check_limit) {
minimum_limit = std::max(highest_limit, minimum_limit);
}
const u32 estimate = static_cast<u32>(maxwell3d.EstimateIndexBufferSize());
const u32 base_size = std::max(minimum_limit, estimate);
maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
auto& params = maxwell3d.draw_manager->GetIndirectParams();
params.is_indexed = true;
@ -301,7 +262,7 @@ public:
maxwell3d.engine_state = Maxwell::EngineHint::OnHLEMacro;
maxwell3d.setHLEReplacementName(0, 0x640, Maxwell::HLEReplaceName::BaseVertex);
maxwell3d.setHLEReplacementName(0, 0x644, Maxwell::HLEReplaceName::BaseInstance);
maxwell3d.draw_manager->DrawIndexedIndirect(topology, 0, base_size);
maxwell3d.draw_manager->DrawIndexedIndirect(topology, 0, estimate);
maxwell3d.engine_state = Maxwell::EngineHint::None;
maxwell3d.replace_table.clear();
}
@ -323,7 +284,6 @@ private:
return;
}
const auto topology = static_cast<Maxwell::Regs::PrimitiveTopology>(parameters[2]);
maxwell3d.regs.draw.topology.Assign(topology);
const u32 padding = parameters[3];
const std::size_t max_draws = parameters[4];
@ -345,8 +305,6 @@ private:
base_vertex, base_instance, parameters[base + 1]);
}
}
u32 minimum_limit{1 << 12};
};
class HLE_C713C83D8F63CCF3 final : public HLEMacroImpl {
@ -431,53 +389,53 @@ public:
HLEMacro::HLEMacro(Engines::Maxwell3D& maxwell3d_) : maxwell3d{maxwell3d_} {
builders.emplace(0x771BB18C62444DA0ULL,
std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>(
[](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_771BB18C62444DA0>(maxwell3d);
[](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_771BB18C62444DA0>(maxwell3d__);
}));
builders.emplace(0x0D61FC9FAAC9FCADULL,
std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>(
[](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_DrawArraysIndirect>(maxwell3d);
[](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_DrawArraysIndirect>(maxwell3d__);
}));
builders.emplace(0x8A4D173EB99A8603ULL,
std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>(
[](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_DrawArraysIndirect>(maxwell3d, true);
[](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_DrawArraysIndirect>(maxwell3d__, true);
}));
builders.emplace(0x0217920100488FF7ULL,
std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>(
[](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_DrawIndexedIndirect>(maxwell3d);
[](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_DrawIndexedIndirect>(maxwell3d__);
}));
builders.emplace(0x3F5E74B9C9A50164ULL,
std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>(
[](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_MultiDrawIndexedIndirectCount>(maxwell3d);
[](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_MultiDrawIndexedIndirectCount>(maxwell3d__);
}));
builders.emplace(0xEAD26C3E2109B06BULL,
std::function<std::unique_ptr<CachedMacro>(Engines::Maxwell3D&)>(
[](Engines::Maxwell3D& maxwell3d) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_MultiLayerClear>(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);
[](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);
[](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);
[](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);
[](Engines::Maxwell3D& maxwell3d__) -> std::unique_ptr<CachedMacro> {
return std::make_unique<HLE_SetRasterBoundingBox>(maxwell3d__);
}));
}

3
src/video_core/renderer_opengl/gl_rasterizer.h
View file

@ -93,8 +93,7 @@ public:
void SignalReference() override;
void ReleaseFences() override;
void FlushAndInvalidateRegion(
VAddr addr, u64 size,
VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
VAddr addr, u64 size, VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
void WaitForIdle() override;
void FragmentBarrier() override;
void TiledCacheBarrier() override;

3
src/video_core/renderer_vulkan/fixed_pipeline_state.h
View file

@ -164,7 +164,8 @@ struct FixedPipelineState {
};
void Refresh(const Maxwell& regs);
void Refresh2(const Maxwell& regs, Maxwell::PrimitiveTopology topology, bool base_feautures_supported);
void Refresh2(const Maxwell& regs, Maxwell::PrimitiveTopology topology,
bool base_feautures_supported);
void Refresh3(const Maxwell& regs);
Maxwell::ComparisonOp DepthTestFunc() const noexcept {

5
src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp
View file

@ -227,7 +227,8 @@ struct DefaultSpec {
ConfigureFuncPtr ConfigureFunc(const std::array<vk::ShaderModule, NUM_STAGES>& modules,
const std::array<Shader::Info, NUM_STAGES>& infos) {
return FindSpec<SimpleVertexSpec, SimpleVertexFragmentSpec, SimpleStorageSpec, SimpleImageSpec, DefaultSpec>(modules, infos);
return FindSpec<SimpleVertexSpec, SimpleVertexFragmentSpec, SimpleStorageSpec, SimpleImageSpec,
DefaultSpec>(modules, infos);
}
} // Anonymous namespace
@ -505,11 +506,9 @@ void GraphicsPipeline::ConfigureDraw(const RescalingPushConstant& rescaling,
if (bind_pipeline) {
cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_GRAPHICS, *pipeline);
}
if (is_rescaling) {
cmdbuf.PushConstants(*pipeline_layout, VK_SHADER_STAGE_ALL_GRAPHICS,
RESCALING_LAYOUT_WORDS_OFFSET, sizeof(rescaling_data),
rescaling_data.data());
}
if (update_rescaling) {
const f32 config_down_factor{Settings::values.resolution_info.down_factor};
const f32 scale_down_factor{is_rescaling ? config_down_factor : 1.0f};

3
src/video_core/renderer_vulkan/vk_rasterizer.h
View file

@ -89,8 +89,7 @@ public:
void SignalReference() override;
void ReleaseFences() override;
void FlushAndInvalidateRegion(
VAddr addr, u64 size,
VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
VAddr addr, u64 size, VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
void WaitForIdle() override;
void FragmentBarrier() override;
void TiledCacheBarrier() override;

5
src/video_core/renderer_vulkan/vk_staging_buffer_pool.h
View file

@ -1,6 +1,5 @@
// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
// SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
// SPDX-License-Identifier: GPL-3.0-or-later
#pragma once

24
src/video_core/vulkan_common/vulkan_device.cpp
View file

@ -587,11 +587,16 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
dynamic_state_3 = {
.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_3_FEATURES_EXT,
.pNext = nullptr,
.extendedDynamicState3DepthClampEnable = ext_extended_dynamic_state_3_enables ? VK_TRUE : VK_FALSE,
.extendedDynamicState3LogicOpEnable = ext_extended_dynamic_state_3_enables ? VK_TRUE : VK_FALSE,
.extendedDynamicState3ColorBlendEnable = ext_extended_dynamic_state_3_blend ? VK_TRUE : VK_FALSE,
.extendedDynamicState3ColorBlendEquation = ext_extended_dynamic_state_3_blend ? VK_TRUE : VK_FALSE,
.extendedDynamicState3ColorWriteMask = ext_extended_dynamic_state_3_blend ? VK_TRUE : VK_FALSE,
.extendedDynamicState3DepthClampEnable =
ext_extended_dynamic_state_3_enables ? VK_TRUE : VK_FALSE,
.extendedDynamicState3LogicOpEnable =
ext_extended_dynamic_state_3_enables ? VK_TRUE : VK_FALSE,
.extendedDynamicState3ColorBlendEnable =
ext_extended_dynamic_state_3_blend ? VK_TRUE : VK_FALSE,
.extendedDynamicState3ColorBlendEquation =
ext_extended_dynamic_state_3_blend ? VK_TRUE : VK_FALSE,
.extendedDynamicState3ColorWriteMask =
ext_extended_dynamic_state_3_blend ? VK_TRUE : VK_FALSE,
};
SetNext(next, dynamic_state_3);
} else {
@ -1342,14 +1347,17 @@ std::vector<const char*> Device::LoadExtensions(bool requires_surface) {
features.pNext = &extended_dynamic_state_3;
physical.GetFeatures2(features);
ext_extended_dynamic_state_3_blend = extended_dynamic_state_3.extendedDynamicState3ColorBlendEnable &&
ext_extended_dynamic_state_3_blend =
extended_dynamic_state_3.extendedDynamicState3ColorBlendEnable &&
extended_dynamic_state_3.extendedDynamicState3ColorBlendEquation &&
extended_dynamic_state_3.extendedDynamicState3ColorWriteMask;
ext_extended_dynamic_state_3_enables = extended_dynamic_state_3.extendedDynamicState3DepthClampEnable &&
ext_extended_dynamic_state_3_enables =
extended_dynamic_state_3.extendedDynamicState3DepthClampEnable &&
extended_dynamic_state_3.extendedDynamicState3LogicOpEnable;
ext_extended_dynamic_state_3 = ext_extended_dynamic_state_3_blend || ext_extended_dynamic_state_3_enables;
ext_extended_dynamic_state_3 =
ext_extended_dynamic_state_3_blend || ext_extended_dynamic_state_3_enables;
if (ext_extended_dynamic_state_3) {
extensions.push_back(VK_EXT_EXTENDED_DYNAMIC_STATE_3_EXTENSION_NAME);
}

3
src/video_core/vulkan_common/vulkan_wrapper.h
View file

@ -1269,7 +1269,8 @@ public:
dld->vkCmdSetColorBlendEnableEXT(handle, first, enables.size(), enables.data());
}
void SetColorBlendEquationEXT(u32 first, Span<VkColorBlendEquationEXT> equations) const noexcept {
void SetColorBlendEquationEXT(u32 first,
Span<VkColorBlendEquationEXT> equations) const noexcept {
dld->vkCmdSetColorBlendEquationEXT(handle, first, equations.size(), equations.data());
}