From 1a58f45d76fe7756dd365e099d1536da769c1eab Mon Sep 17 00:00:00 2001
From: Fernando Sahmkow <fsahmkow27@gmail.com>
Date: Mon, 23 Sep 2019 14:02:02 -0400
Subject: [PATCH] VideoCore: Unify const buffer accessing along engines and
provide ConstBufferLocker class to shaders.
---
CMakeModules/GenerateSCMRev.cmake | 6 +-
src/common/CMakeLists.txt | 6 +-
src/common/hash.h | 11 +++
src/video_core/CMakeLists.txt | 7 +-
.../engines/const_buffer_engine_interface.h | 26 +++++++
src/video_core/engines/kepler_compute.cpp | 3 +-
src/video_core/engines/kepler_compute.h | 5 +-
src/video_core/engines/maxwell_3d.cpp | 3 +-
src/video_core/engines/maxwell_3d.h | 5 +-
.../renderer_opengl/gl_rasterizer.cpp | 7 +-
src/video_core/shader/const_buffer_locker.cpp | 72 +++++++++++++++++++
src/video_core/shader/const_buffer_locker.h | 50 +++++++++++++
src/video_core/shader/shader_ir.h | 1 +
13 files changed, 187 insertions(+), 15 deletions(-)
create mode 100644 src/video_core/engines/const_buffer_engine_interface.h
create mode 100644 src/video_core/shader/const_buffer_locker.cpp
create mode 100644 src/video_core/shader/const_buffer_locker.h
diff --git a/CMakeModules/GenerateSCMRev.cmake b/CMakeModules/GenerateSCMRev.cmake
index 09eabe2c73..21e03ae98f 100644
--- a/CMakeModules/GenerateSCMRev.cmake
+++ b/CMakeModules/GenerateSCMRev.cmake
@@ -85,10 +85,12 @@ set(HASH_FILES
"${VIDEO_CORE}/shader/decode/xmad.cpp"
"${VIDEO_CORE}/shader/ast.cpp"
"${VIDEO_CORE}/shader/ast.h"
- "${VIDEO_CORE}/shader/control_flow.cpp"
- "${VIDEO_CORE}/shader/control_flow.h"
"${VIDEO_CORE}/shader/compiler_settings.cpp"
"${VIDEO_CORE}/shader/compiler_settings.h"
+ "${VIDEO_CORE}/shader/const_buffer_locker.cpp"
+ "${VIDEO_CORE}/shader/const_buffer_locker.h"
+ "${VIDEO_CORE}/shader/control_flow.cpp"
+ "${VIDEO_CORE}/shader/control_flow.h"
"${VIDEO_CORE}/shader/decode.cpp"
"${VIDEO_CORE}/shader/expr.cpp"
"${VIDEO_CORE}/shader/expr.h"
diff --git a/src/common/CMakeLists.txt b/src/common/CMakeLists.txt
index 5b51fcafa9..9c6f1c07c2 100644
--- a/src/common/CMakeLists.txt
+++ b/src/common/CMakeLists.txt
@@ -74,10 +74,12 @@ add_custom_command(OUTPUT scm_rev.cpp
"${VIDEO_CORE}/shader/decode/xmad.cpp"
"${VIDEO_CORE}/shader/ast.cpp"
"${VIDEO_CORE}/shader/ast.h"
- "${VIDEO_CORE}/shader/control_flow.cpp"
- "${VIDEO_CORE}/shader/control_flow.h"
"${VIDEO_CORE}/shader/compiler_settings.cpp"
"${VIDEO_CORE}/shader/compiler_settings.h"
+ "${VIDEO_CORE}/shader/const_buffer_locker.cpp"
+ "${VIDEO_CORE}/shader/const_buffer_locker.h"
+ "${VIDEO_CORE}/shader/control_flow.cpp"
+ "${VIDEO_CORE}/shader/control_flow.h"
"${VIDEO_CORE}/shader/decode.cpp"
"${VIDEO_CORE}/shader/expr.cpp"
"${VIDEO_CORE}/shader/expr.h"
diff --git a/src/common/hash.h b/src/common/hash.h
index 40194d1ee4..c939709bc9 100644
--- a/src/common/hash.h
+++ b/src/common/hash.h
@@ -6,6 +6,8 @@
#include <cstddef>
#include <cstring>
+#include <utility>
+#include <boost/functional/hash.hpp>
#include "common/cityhash.h"
#include "common/common_types.h"
@@ -68,4 +70,13 @@ struct HashableStruct {
}
};
+struct PairHash {
+ template <class T1, class T2>
+ std::size_t operator()(const std::pair<T1, T2>& pair) const {
+ std::size_t seed = std::hash<T1>()(pair.first);
+ boost::hash_combine(seed, std::hash<T2>()(pair.second));
+ return seed;
+ }
+};
+
} // namespace Common
diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt
index eaa694ff8d..cb6eda1b86 100644
--- a/src/video_core/CMakeLists.txt
+++ b/src/video_core/CMakeLists.txt
@@ -6,6 +6,7 @@ add_library(video_core STATIC
dma_pusher.h
debug_utils/debug_utils.cpp
debug_utils/debug_utils.h
+ engines/const_buffer_engine_interface.h
engines/const_buffer_info.h
engines/engine_upload.cpp
engines/engine_upload.h
@@ -107,10 +108,12 @@ add_library(video_core STATIC
shader/decode/other.cpp
shader/ast.cpp
shader/ast.h
- shader/control_flow.cpp
- shader/control_flow.h
shader/compiler_settings.cpp
shader/compiler_settings.h
+ shader/const_buffer_locker.cpp
+ shader/const_buffer_locker.h
+ shader/control_flow.cpp
+ shader/control_flow.h
shader/decode.cpp
shader/expr.cpp
shader/expr.h
diff --git a/src/video_core/engines/const_buffer_engine_interface.h b/src/video_core/engines/const_buffer_engine_interface.h
new file mode 100644
index 0000000000..cc41a9cacb
--- /dev/null
+++ b/src/video_core/engines/const_buffer_engine_interface.h
@@ -0,0 +1,26 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "common/common_types.h"
+
+namespace Tegra::Engines {
+
+enum class ShaderType : u32 {
+ Vertex = 0,
+ TesselationControl = 1,
+ TesselationEval = 2,
+ Geometry = 3,
+ Fragment = 4,
+ Compute = 5,
+};
+
+class ConstBufferEngineInterface {
+public:
+ virtual ~ConstBufferEngineInterface() {}
+ virtual u32 AccessConstBuffer32(ShaderType stage, u64 const_buffer, u64 offset) const = 0;
+};
+
+}
diff --git a/src/video_core/engines/kepler_compute.cpp b/src/video_core/engines/kepler_compute.cpp
index 63d4491352..ba97c28949 100644
--- a/src/video_core/engines/kepler_compute.cpp
+++ b/src/video_core/engines/kepler_compute.cpp
@@ -70,7 +70,8 @@ Texture::FullTextureInfo KeplerCompute::GetTextureInfo(const Texture::TextureHan
GetTSCEntry(tex_handle.tsc_id)};
}
-u32 KeplerCompute::AccessConstBuffer32(u64 const_buffer, u64 offset) const {
+u32 KeplerCompute::AccessConstBuffer32(ShaderType stage, u64 const_buffer, u64 offset) const {
+ ASSERT(stage == ShaderType::Compute);
const auto& buffer = launch_description.const_buffer_config[const_buffer];
u32 result;
std::memcpy(&result, memory_manager.GetPointer(buffer.Address() + offset), sizeof(u32));
diff --git a/src/video_core/engines/kepler_compute.h b/src/video_core/engines/kepler_compute.h
index 90cf650d2f..d7e0dfcd6f 100644
--- a/src/video_core/engines/kepler_compute.h
+++ b/src/video_core/engines/kepler_compute.h
@@ -11,6 +11,7 @@
#include "common/common_funcs.h"
#include "common/common_types.h"
#include "video_core/engines/engine_upload.h"
+#include "video_core/engines/const_buffer_engine_interface.h"
#include "video_core/gpu.h"
#include "video_core/textures/texture.h"
@@ -37,7 +38,7 @@ namespace Tegra::Engines {
#define KEPLER_COMPUTE_REG_INDEX(field_name) \
(offsetof(Tegra::Engines::KeplerCompute::Regs, field_name) / sizeof(u32))
-class KeplerCompute final {
+class KeplerCompute final : public ConstBufferEngineInterface {
public:
explicit KeplerCompute(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
MemoryManager& memory_manager);
@@ -201,7 +202,7 @@ public:
Texture::FullTextureInfo GetTextureInfo(const Texture::TextureHandle tex_handle,
std::size_t offset) const;
- u32 AccessConstBuffer32(u64 const_buffer, u64 offset) const;
+ u32 AccessConstBuffer32(ShaderType stage, u64 const_buffer, u64 offset) const override;
private:
Core::System& system;
diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index 59976943a2..92e38b0713 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -847,7 +847,8 @@ void Maxwell3D::ProcessClearBuffers() {
rasterizer.Clear();
}
-u32 Maxwell3D::AccessConstBuffer32(Regs::ShaderStage stage, u64 const_buffer, u64 offset) const {
+u32 Maxwell3D::AccessConstBuffer32(ShaderType stage, u64 const_buffer, u64 offset) const {
+ ASSERT(stage != ShaderType::Compute);
const auto& shader_stage = state.shader_stages[static_cast<std::size_t>(stage)];
const auto& buffer = shader_stage.const_buffers[const_buffer];
u32 result;
diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index e3f1047d5c..04d02d2086 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -16,6 +16,7 @@
#include "common/common_types.h"
#include "common/math_util.h"
#include "video_core/engines/const_buffer_info.h"
+#include "video_core/engines/const_buffer_engine_interface.h"
#include "video_core/engines/engine_upload.h"
#include "video_core/gpu.h"
#include "video_core/macro_interpreter.h"
@@ -44,7 +45,7 @@ namespace Tegra::Engines {
#define MAXWELL3D_REG_INDEX(field_name) \
(offsetof(Tegra::Engines::Maxwell3D::Regs, field_name) / sizeof(u32))
-class Maxwell3D final {
+class Maxwell3D final : public ConstBufferEngineInterface {
public:
explicit Maxwell3D(Core::System& system, VideoCore::RasterizerInterface& rasterizer,
MemoryManager& memory_manager);
@@ -1257,7 +1258,7 @@ public:
/// Returns the texture information for a specific texture in a specific shader stage.
Texture::FullTextureInfo GetStageTexture(Regs::ShaderStage stage, std::size_t offset) const;
- u32 AccessConstBuffer32(Regs::ShaderStage stage, u64 const_buffer, u64 offset) const;
+ u32 AccessConstBuffer32(ShaderType stage, u64 const_buffer, u64 offset) const override;
/// Memory for macro code - it's undetermined how big this is, however 1MB is much larger than
/// we've seen used.
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index cbcf81414b..10114909b1 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -975,7 +975,8 @@ TextureBufferUsage RasterizerOpenGL::SetupDrawTextures(Maxwell::ShaderStage stag
}
const auto cbuf = entry.GetBindlessCBuf();
Tegra::Texture::TextureHandle tex_handle;
- tex_handle.raw = maxwell3d.AccessConstBuffer32(stage, cbuf.first, cbuf.second);
+ Tegra::Engines::ShaderType shader_type = static_cast<Tegra::Engines::ShaderType>(stage);
+ tex_handle.raw = maxwell3d.AccessConstBuffer32(shader_type, cbuf.first, cbuf.second);
return maxwell3d.GetTextureInfo(tex_handle, entry.GetOffset());
}();
@@ -1005,7 +1006,7 @@ TextureBufferUsage RasterizerOpenGL::SetupComputeTextures(const Shader& kernel)
}
const auto cbuf = entry.GetBindlessCBuf();
Tegra::Texture::TextureHandle tex_handle;
- tex_handle.raw = compute.AccessConstBuffer32(cbuf.first, cbuf.second);
+ tex_handle.raw = compute.AccessConstBuffer32(Tegra::Engines::ShaderType::Compute, cbuf.first, cbuf.second);
return compute.GetTextureInfo(tex_handle, entry.GetOffset());
}();
@@ -1050,7 +1051,7 @@ void RasterizerOpenGL::SetupComputeImages(const Shader& shader) {
}
const auto cbuf = entry.GetBindlessCBuf();
Tegra::Texture::TextureHandle tex_handle;
- tex_handle.raw = compute.AccessConstBuffer32(cbuf.first, cbuf.second);
+ tex_handle.raw = compute.AccessConstBuffer32(Tegra::Engines::ShaderType::Compute, cbuf.first, cbuf.second);
return compute.GetTextureInfo(tex_handle, entry.GetOffset()).tic;
}();
SetupImage(bindpoint, tic, entry);
diff --git a/src/video_core/shader/const_buffer_locker.cpp b/src/video_core/shader/const_buffer_locker.cpp
new file mode 100644
index 0000000000..6a9e0ed5e3
--- /dev/null
+++ b/src/video_core/shader/const_buffer_locker.cpp
@@ -0,0 +1,72 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "common/assert.h"
+#include "common/common_types.h"
+#include "video_core/engines/maxwell_3d.h"
+#include "video_core/shader/const_buffer_locker.h"
+
+namespace VideoCommon::Shader {
+
+ConstBufferLocker::ConstBufferLocker(Tegra::Engines::ShaderType shader_stage)
+ : engine{nullptr}, shader_stage{shader_stage} {}
+
+ConstBufferLocker::ConstBufferLocker(Tegra::Engines::ShaderType shader_stage,
+ Tegra::Engines::ConstBufferEngineInterface* engine)
+ : engine{engine}, shader_stage{shader_stage} {}
+
+bool ConstBufferLocker::IsEngineSet() const {
+ return engine != nullptr;
+}
+
+void ConstBufferLocker::SetEngine(Tegra::Engines::ConstBufferEngineInterface* engine_) {
+ engine = engine_;
+}
+
+std::optional<u32> ConstBufferLocker::ObtainKey(u32 buffer, u32 offset) {
+ const std::pair<u32, u32> key = {buffer, offset};
+ const auto iter = keys.find(key);
+ if (iter != keys.end()) {
+ return {iter->second};
+ }
+ if (!IsEngineSet()) {
+ return {};
+ }
+ const u32 value = engine->AccessConstBuffer32(shader_stage, buffer, offset);
+ keys.emplace(key, value);
+ return {value};
+}
+
+void ConstBufferLocker::InsertKey(u32 buffer, u32 offset, u32 value) {
+ const std::pair<u32, u32> key = {buffer, offset};
+ keys[key] = value;
+}
+
+u32 ConstBufferLocker::NumKeys() const {
+ return keys.size();
+}
+
+const std::unordered_map<std::pair<u32, u32>, u32, Common::PairHash>&
+ConstBufferLocker::AccessKeys() const {
+ return keys;
+}
+
+bool ConstBufferLocker::AreKeysConsistant() const {
+ if (!IsEngineSet()) {
+ return false;
+ }
+ for (const auto& key_val : keys) {
+ const std::pair<u32, u32> key = key_val.first;
+ const u32 value = key_val.second;
+ const u32 other_value = engine->AccessConstBuffer32(shader_stage, key.first, key.second);
+ if (other_value != value) {
+ return false;
+ }
+ }
+ return true;
+}
+
+} // namespace VideoCommon::Shader
diff --git a/src/video_core/shader/const_buffer_locker.h b/src/video_core/shader/const_buffer_locker.h
new file mode 100644
index 0000000000..39e62584da
--- /dev/null
+++ b/src/video_core/shader/const_buffer_locker.h
@@ -0,0 +1,50 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <unordered_map>
+#include "common/common_types.h"
+#include "common/hash.h"
+#include "video_core/engines/const_buffer_engine_interface.h"
+
+namespace VideoCommon::Shader {
+
+class ConstBufferLocker {
+public:
+ explicit ConstBufferLocker(Tegra::Engines::ShaderType shader_stage);
+
+ explicit ConstBufferLocker(Tegra::Engines::ShaderType shader_stage,
+ Tegra::Engines::ConstBufferEngineInterface* engine);
+
+ // Checks if an engine is setup, it may be possible that during disk shader
+ // cache run, the engines have not been created yet.
+ bool IsEngineSet() const;
+
+ // Use this to set/change the engine used for this shader.
+ void SetEngine(Tegra::Engines::ConstBufferEngineInterface* engine);
+
+ // Retrieves a key from the locker, if it's registered, it will give the
+ // registered value, if not it will obtain it from maxwell3d and register it.
+ std::optional<u32> ObtainKey(u32 buffer, u32 offset);
+
+ // Manually inserts a key.
+ void InsertKey(u32 buffer, u32 offset, u32 value);
+
+ // Retrieves the number of keys registered.
+ u32 NumKeys() const;
+
+ // Gives an accessor to the key's database.
+ const std::unordered_map<std::pair<u32, u32>, u32, Common::PairHash>& AccessKeys() const;
+
+ // Checks keys against maxwell3d's current const buffers. Returns true if they
+ // are the same value, false otherwise;
+ bool AreKeysConsistant() const;
+
+private:
+ Tegra::Engines::ConstBufferEngineInterface* engine;
+ Tegra::Engines::ShaderType shader_stage;
+ std::unordered_map<std::pair<u32, u32>, u32, Common::PairHash> keys{};
+};
+} // namespace VideoCommon::Shader
diff --git a/src/video_core/shader/shader_ir.h b/src/video_core/shader/shader_ir.h
index 91cd0a5349..68818643c1 100644
--- a/src/video_core/shader/shader_ir.h
+++ b/src/video_core/shader/shader_ir.h
@@ -17,6 +17,7 @@
#include "video_core/engines/shader_header.h"
#include "video_core/shader/ast.h"
#include "video_core/shader/compiler_settings.h"
+#include "video_core/shader/const_buffer_locker.h"
#include "video_core/shader/node.h"
namespace VideoCommon::Shader {