Shader_IR: Address Feedback

src/video_core/CMakeLists.txt

@@ -29,6 +29,8 @@ add_library(video_core STATIC
     gpu_synch.h
     gpu_thread.cpp
     gpu_thread.h
+    guest_driver.cpp
+    guest_driver.h
     macro_interpreter.cpp
     macro_interpreter.h
     memory_manager.cpp

src/video_core/engines/const_buffer_engine_interface.h

@@ -9,6 +9,7 @@
 #include "common/common_types.h"
 #include "video_core/engines/shader_bytecode.h"
 #include "video_core/engines/shader_type.h"
+#include "video_core/guest_driver.h"
 #include "video_core/textures/texture.h"
 
 namespace Tegra::Engines {
@@ -106,6 +107,9 @@ public:
     virtual SamplerDescriptor AccessBindlessSampler(ShaderType stage, u64 const_buffer,
                                                     u64 offset) const = 0;
     virtual u32 GetBoundBuffer() const = 0;
+
+    virtual VideoCore::GuestDriverProfile& AccessGuestDriverProfile() = 0;
+    virtual const VideoCore::GuestDriverProfile& AccessGuestDriverProfile() const = 0;
 };
 
 } // namespace Tegra::Engines

src/video_core/engines/kepler_compute.cpp

@@ -94,6 +94,14 @@ SamplerDescriptor KeplerCompute::AccessBindlessSampler(ShaderType stage, u64 con
     return result;
 }
 
+VideoCore::GuestDriverProfile& KeplerCompute::AccessGuestDriverProfile() {
+    return rasterizer.AccessGuestDriverProfile();
+}
+
+const VideoCore::GuestDriverProfile& KeplerCompute::AccessGuestDriverProfile() const {
+    return rasterizer.AccessGuestDriverProfile();
+}
+
 void KeplerCompute::ProcessLaunch() {
     const GPUVAddr launch_desc_loc = regs.launch_desc_loc.Address();
     memory_manager.ReadBlockUnsafe(launch_desc_loc, &launch_description,

src/video_core/engines/kepler_compute.h

@@ -218,6 +218,10 @@ public:
         return regs.tex_cb_index;
     }
 
+    VideoCore::GuestDriverProfile& AccessGuestDriverProfile() override;
+
+    const VideoCore::GuestDriverProfile& AccessGuestDriverProfile() const override;
+
 private:
     Core::System& system;
     VideoCore::RasterizerInterface& rasterizer;

src/video_core/engines/maxwell_3d.cpp

@@ -783,4 +783,12 @@ SamplerDescriptor Maxwell3D::AccessBindlessSampler(ShaderType stage, u64 const_b
     return result;
 }
 
+VideoCore::GuestDriverProfile& Maxwell3D::AccessGuestDriverProfile() {
+    return rasterizer.AccessGuestDriverProfile();
+}
+
+const VideoCore::GuestDriverProfile& Maxwell3D::AccessGuestDriverProfile() const {
+    return rasterizer.AccessGuestDriverProfile();
+}
+
 } // namespace Tegra::Engines

src/video_core/engines/maxwell_3d.h

@@ -1301,6 +1301,10 @@ public:
         return regs.tex_cb_index;
     }
 
+    VideoCore::GuestDriverProfile& AccessGuestDriverProfile() override;
+
+    const VideoCore::GuestDriverProfile& AccessGuestDriverProfile() const override;
+
     /// Memory for macro code - it's undetermined how big this is, however 1MB is much larger than
     /// we've seen used.
     using MacroMemory = std::array<u32, 0x40000>;

src/video_core/guest_driver.cpp

@@ -0,0 +1,36 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <algorithm>
+#include <limits>
+
+#include "video_core/guest_driver.h"
+
+namespace VideoCore {
+
+void GuestDriverProfile::DeduceTextureHandlerSize(std::vector<u32>&& bound_offsets) {
+    if (texture_handler_size_deduced) {
+        return;
+    }
+    const std::size_t size = bound_offsets.size();
+    if (size < 2) {
+        return;
+    }
+    std::sort(bound_offsets.begin(), bound_offsets.end(), std::less{});
+    u32 min_val = std::numeric_limits<u32>::max();
+    for (std::size_t i = 1; i < size; ++i) {
+        if (bound_offsets[i] == bound_offsets[i - 1]) {
+            continue;
+        }
+        const u32 new_min = bound_offsets[i] - bound_offsets[i - 1];
+        min_val = std::min(min_val, new_min);
+    }
+    if (min_val > 2) {
+        return;
+    }
+    texture_handler_size_deduced = true;
+    texture_handler_size = min_texture_handler_size * min_val;
+}
+
+} // namespace VideoCore

src/video_core/guest_driver.h

@@ -0,0 +1,40 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <vector>
+
+#include "common/common_types.h"
+
+namespace VideoCore {
+
+/**
+ * The GuestDriverProfile class is used to learn about the GPU drivers behavior and collect
+ * information necessary for impossible to avoid HLE methods like shader tracks as they are
+ * Entscheidungsproblems.
+ */
+class GuestDriverProfile {
+public:
+    void DeduceTextureHandlerSize(std::vector<u32>&& bound_offsets);
+
+    u32 GetTextureHandlerSize() const {
+        return texture_handler_size;
+    }
+
+    bool TextureHandlerSizeKnown() const {
+        return texture_handler_size_deduced;
+    }
+
+private:
+    // Minimum size of texture handler any driver can use.
+    static constexpr u32 min_texture_handler_size = 4;
+    // This goes with Vulkan and OpenGL standards but Nvidia GPUs can easily
+    // use 4 bytes instead. Thus, certain drivers may squish the size.
+    static constexpr u32 default_texture_handler_size = 8;
+    u32 texture_handler_size = default_texture_handler_size;
+    bool texture_handler_size_deduced = false;
+};
+
+} // namespace VideoCore

src/video_core/rasterizer_interface.h

@@ -9,6 +9,7 @@
 #include "common/common_types.h"
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/gpu.h"
+#include "video_core/guest_driver.h"
 
 namespace Tegra {
 class MemoryManager;
@@ -78,5 +79,18 @@ public:
     /// Initialize disk cached resources for the game being emulated
     virtual void LoadDiskResources(const std::atomic_bool& stop_loading = false,
                                    const DiskResourceLoadCallback& callback = {}) {}
+
+    /// Grant access to the Guest Driver Profile for recording/obtaining info on the guest driver.
+    GuestDriverProfile& AccessGuestDriverProfile() {
+        return guest_driver_profile;
+    }
+
+    /// Grant access to the Guest Driver Profile for recording/obtaining info on the guest driver.
+    const GuestDriverProfile& AccessGuestDriverProfile() const {
+        return guest_driver_profile;
+    }
+
+private:
+    GuestDriverProfile guest_driver_profile{};
 };
 } // namespace VideoCore

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -217,6 +217,7 @@ std::unique_ptr<ConstBufferLocker> MakeLocker(Core::System& system, ShaderType s
 }
 
 void FillLocker(ConstBufferLocker& locker, const ShaderDiskCacheUsage& usage) {
+    locker.SetBoundBuffer(usage.bound_buffer);
     for (const auto& key : usage.keys) {
         const auto [buffer, offset] = key.first;
         locker.InsertKey(buffer, offset, key.second);
@@ -417,7 +418,8 @@ bool CachedShader::EnsureValidLockerVariant() {
 
 ShaderDiskCacheUsage CachedShader::GetUsage(const ProgramVariant& variant,
                                             const ConstBufferLocker& locker) const {
-    return ShaderDiskCacheUsage{unique_identifier, variant, locker.GetKeys(),
+    return ShaderDiskCacheUsage{unique_identifier,         variant,
+                                locker.GetBoundBuffer(),   locker.GetKeys(),
                                 locker.GetBoundSamplers(), locker.GetBindlessSamplers()};
 }
 

src/video_core/renderer_opengl/gl_shader_disk_cache.cpp

@@ -53,7 +53,7 @@ struct BindlessSamplerKey {
     Tegra::Engines::SamplerDescriptor sampler{};
 };
 
-constexpr u32 NativeVersion = 11;
+constexpr u32 NativeVersion = 12;
 
 // Making sure sizes doesn't change by accident
 static_assert(sizeof(ProgramVariant) == 20);
@@ -186,7 +186,8 @@ ShaderDiskCacheOpenGL::LoadTransferable() {
             u32 num_bound_samplers{};
             u32 num_bindless_samplers{};
             if (file.ReadArray(&usage.unique_identifier, 1) != 1 ||
-                file.ReadArray(&usage.variant, 1) != 1 || file.ReadArray(&num_keys, 1) != 1 ||
+                file.ReadArray(&usage.variant, 1) != 1 ||
+                file.ReadArray(&usage.bound_buffer, 1) != 1 || file.ReadArray(&num_keys, 1) != 1 ||
                 file.ReadArray(&num_bound_samplers, 1) != 1 ||
                 file.ReadArray(&num_bindless_samplers, 1) != 1) {
                 LOG_ERROR(Render_OpenGL, error_loading);
@@ -281,7 +282,9 @@ ShaderDiskCacheOpenGL::LoadPrecompiledFile(FileUtil::IOFile& file) {
         u32 num_bindless_samplers{};
         ShaderDiskCacheUsage usage;
         if (!LoadObjectFromPrecompiled(usage.unique_identifier) ||
-            !LoadObjectFromPrecompiled(usage.variant) || !LoadObjectFromPrecompiled(num_keys) ||
+            !LoadObjectFromPrecompiled(usage.variant) ||
+            !LoadObjectFromPrecompiled(usage.bound_buffer) ||
+            !LoadObjectFromPrecompiled(num_keys) ||
             !LoadObjectFromPrecompiled(num_bound_samplers) ||
             !LoadObjectFromPrecompiled(num_bindless_samplers)) {
             return {};
@@ -393,6 +396,7 @@ void ShaderDiskCacheOpenGL::SaveUsage(const ShaderDiskCacheUsage& usage) {
 
     if (file.WriteObject(TransferableEntryKind::Usage) != 1 ||
         file.WriteObject(usage.unique_identifier) != 1 || file.WriteObject(usage.variant) != 1 ||
+        file.WriteObject(usage.bound_buffer) != 1 ||
         file.WriteObject(static_cast<u32>(usage.keys.size())) != 1 ||
         file.WriteObject(static_cast<u32>(usage.bound_samplers.size())) != 1 ||
         file.WriteObject(static_cast<u32>(usage.bindless_samplers.size())) != 1) {
@@ -447,7 +451,7 @@ void ShaderDiskCacheOpenGL::SaveDump(const ShaderDiskCacheUsage& usage, GLuint p
     };
 
     if (!SaveObjectToPrecompiled(usage.unique_identifier) ||
-        !SaveObjectToPrecompiled(usage.variant) ||
+        !SaveObjectToPrecompiled(usage.variant) || !SaveObjectToPrecompiled(usage.bound_buffer) ||
         !SaveObjectToPrecompiled(static_cast<u32>(usage.keys.size())) ||
         !SaveObjectToPrecompiled(static_cast<u32>(usage.bound_samplers.size())) ||
         !SaveObjectToPrecompiled(static_cast<u32>(usage.bindless_samplers.size()))) {

src/video_core/renderer_opengl/gl_shader_disk_cache.h

@@ -79,6 +79,7 @@ static_assert(std::is_trivially_copyable_v<ProgramVariant>);
 struct ShaderDiskCacheUsage {
     u64 unique_identifier{};
     ProgramVariant variant;
+    u32 bound_buffer{};
     VideoCommon::Shader::KeyMap keys;
     VideoCommon::Shader::BoundSamplerMap bound_samplers;
     VideoCommon::Shader::BindlessSamplerMap bindless_samplers;

src/video_core/shader/const_buffer_locker.cpp

@@ -66,6 +66,18 @@ std::optional<Tegra::Engines::SamplerDescriptor> ConstBufferLocker::ObtainBindle
     return value;
 }
 
+std::optional<u32> ConstBufferLocker::ObtainBoundBuffer() {
+    if (bound_buffer_saved) {
+        return bound_buffer;
+    }
+    if (!engine) {
+        return std::nullopt;
+    }
+    bound_buffer_saved = true;
+    bound_buffer = engine->GetBoundBuffer();
+    return bound_buffer;
+}
+
 void ConstBufferLocker::InsertKey(u32 buffer, u32 offset, u32 value) {
     keys.insert_or_assign({buffer, offset}, value);
 }
@@ -78,6 +90,11 @@ void ConstBufferLocker::InsertBindlessSampler(u32 buffer, u32 offset, SamplerDes
     bindless_samplers.insert_or_assign({buffer, offset}, sampler);
 }
 
+void ConstBufferLocker::SetBoundBuffer(u32 buffer) {
+    bound_buffer_saved = true;
+    bound_buffer = buffer;
+}
+
 bool ConstBufferLocker::IsConsistent() const {
     if (!engine) {
         return false;

src/video_core/shader/const_buffer_locker.h

@@ -10,6 +10,7 @@
 #include "common/hash.h"
 #include "video_core/engines/const_buffer_engine_interface.h"
 #include "video_core/engines/shader_type.h"
+#include "video_core/guest_driver.h"
 
 namespace VideoCommon::Shader {
 
@@ -40,6 +41,8 @@ public:
 
     std::optional<Tegra::Engines::SamplerDescriptor> ObtainBindlessSampler(u32 buffer, u32 offset);
 
+    std::optional<u32> ObtainBoundBuffer();
+
     /// Inserts a key.
     void InsertKey(u32 buffer, u32 offset, u32 value);
 
@@ -49,6 +52,9 @@ public:
     /// Inserts a bindless sampler key.
     void InsertBindlessSampler(u32 buffer, u32 offset, Tegra::Engines::SamplerDescriptor sampler);
 
+    /// Set the bound buffer for this locker.
+    void SetBoundBuffer(u32 buffer);
+
     /// Checks keys and samplers against engine's current const buffers. Returns true if they are
     /// the same value, false otherwise;
     bool IsConsistent() const;
@@ -71,12 +77,25 @@ public:
         return bindless_samplers;
     }
 
+    u32 GetBoundBuffer() const {
+        return bound_buffer;
+    }
+
+    VideoCore::GuestDriverProfile* AccessGuestDriverProfile() const {
+        if (engine) {
+            return &engine->AccessGuestDriverProfile();
+        }
+        return nullptr;
+    }
+
 private:
     const Tegra::Engines::ShaderType stage;
     Tegra::Engines::ConstBufferEngineInterface* engine = nullptr;
     KeyMap keys;
     BoundSamplerMap bound_samplers;
     BindlessSamplerMap bindless_samplers;
+    bool bound_buffer_saved{};
+    u32 bound_buffer{};
 };
 
 } // namespace VideoCommon::Shader

src/video_core/shader/decode.cpp

@@ -33,6 +33,29 @@ constexpr bool IsSchedInstruction(u32 offset, u32 main_offset) {
     return (absolute_offset % SchedPeriod) == 0;
 }
 
+void DeduceTextureHandlerSize(VideoCore::GuestDriverProfile* gpu_driver,
+                              std::list<Sampler>& used_samplers) {
+    if (gpu_driver == nullptr) {
+        LOG_CRITICAL(HW_GPU, "GPU Driver profile has not been created yet");
+        return;
+    }
+    if (gpu_driver->TextureHandlerSizeKnown() || used_samplers.size() <= 1) {
+        return;
+    }
+    u32 count{};
+    std::vector<u32> bound_offsets;
+    for (const auto& sampler : used_samplers) {
+        if (sampler.IsBindless()) {
+            continue;
+        }
+        ++count;
+        bound_offsets.emplace_back(sampler.GetOffset());
+    }
+    if (count > 1) {
+        gpu_driver->DeduceTextureHandlerSize(std::move(bound_offsets));
+    }
+}
+
 } // Anonymous namespace
 
 class ASTDecoder {
@@ -315,4 +338,10 @@ u32 ShaderIR::DecodeInstr(NodeBlock& bb, u32 pc) {
     return pc + 1;
 }
 
+void ShaderIR::PostDecode() {
+    // Deduce texture handler size if needed
+    auto gpu_driver = locker.AccessGuestDriverProfile();
+    DeduceTextureHandlerSize(gpu_driver, used_samplers);
+}
+
 } // namespace VideoCommon::Shader

src/video_core/shader/shader_ir.cpp

@@ -27,6 +27,7 @@ ShaderIR::ShaderIR(const ProgramCode& program_code, u32 main_offset, CompilerSet
                    ConstBufferLocker& locker)
     : program_code{program_code}, main_offset{main_offset}, settings{settings}, locker{locker} {
     Decode();
+    PostDecode();
 }
 
 ShaderIR::~ShaderIR() = default;

src/video_core/shader/shader_ir.h

@@ -191,6 +191,7 @@ private:
     };
 
     void Decode();
+    void PostDecode();
 
     NodeBlock DecodeRange(u32 begin, u32 end);
     void DecodeRangeInner(NodeBlock& bb, u32 begin, u32 end);