GPU: Implement guest driver profile and deduce texture handler sizes.

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 +107,8 @@ public:
     virtual SamplerDescriptor AccessBindlessSampler(ShaderType stage, u64 const_buffer,
                                                     u64 offset) const = 0;
     virtual u32 GetBoundBuffer() const = 0;
+
+    virtual VideoCore::GuestDriverProfile& AccessGuestDriverProfile() = 0;
 };
 
 } // namespace Tegra::Engines

src/video_core/engines/kepler_compute.cpp

@@ -94,6 +94,10 @@ SamplerDescriptor KeplerCompute::AccessBindlessSampler(ShaderType stage, u64 con
     return result;
 }
 
+VideoCore::GuestDriverProfile& KeplerCompute::AccessGuestDriverProfile() {
+    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 +218,8 @@ public:
         return regs.tex_cb_index;
     }
 
+    VideoCore::GuestDriverProfile& AccessGuestDriverProfile() override;
+
 private:
     Core::System& system;
     VideoCore::RasterizerInterface& rasterizer;

src/video_core/engines/maxwell_3d.cpp

@@ -784,4 +784,8 @@ SamplerDescriptor Maxwell3D::AccessBindlessSampler(ShaderType stage, u64 const_b
     return result;
 }
 
+VideoCore::GuestDriverProfile& Maxwell3D::AccessGuestDriverProfile() {
+    return rasterizer.AccessGuestDriverProfile();
+}
+
 } // namespace Tegra::Engines

src/video_core/engines/maxwell_3d.h

@@ -1306,6 +1306,8 @@ public:
         return regs.tex_cb_index;
     }
 
+    VideoCore::GuestDriverProfile& AccessGuestDriverProfile() 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 +1,34 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "video_core/guest_driver.h"
+
+namespace VideoCore {
+
+void GuestDriverProfile::DeduceTextureHandlerSize(std::vector<u32>&& bound_offsets) {
+    if (texture_handler_size_deduced) {
+        return;
+    }
+    std::size_t size = bound_offsets.size();
+    if (size < 2) {
+        return;
+    }
+    std::sort(bound_offsets.begin(), bound_offsets.end(),
+              [](const u32& a, const u32& b) { return a < b; });
+    u32 min_val = 0xFFFFFFFF; // set to highest possible 32 bit integer;
+    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 = sizeof(u32) * min_val;
+}
+
+} // namespace VideoCore

src/video_core/guest_driver.h

@@ -0,0 +1,37 @@
+// Copyright 2019 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.
+ */
+class GuestDriverProfile {
+public:
+    u32 GetTextureHandlerSize() const {
+        return texture_handler_size;
+    }
+
+    bool TextureHandlerSizeKnown() const {
+        return texture_handler_size_deduced;
+    }
+
+    void DeduceTextureHandlerSize(std::vector<u32>&& bound_offsets);
+
+private:
+    // 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{};
+};
+
+} // 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 +79,12 @@ public:
     /// Initialize disk cached resources for the game being emulated
     virtual void LoadDiskResources(const std::atomic_bool& stop_loading = false,
                                    const DiskResourceLoadCallback& callback = {}) {}
+
+    GuestDriverProfile& AccessGuestDriverProfile() {
+        return guest_driver_profile;
+    }
+
+private:
+    GuestDriverProfile guest_driver_profile{};
 };
 } // namespace VideoCore

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 {
 
@@ -71,6 +72,13 @@ public:
         return bindless_samplers;
     }
 
+    VideoCore::GuestDriverProfile* AccessGuestDriverProfile() {
+        if (engine) {
+            return &(engine->AccessGuestDriverProfile());
+        }
+        return nullptr;
+    }
+
 private:
     const Tegra::Engines::ShaderType stage;
     Tegra::Engines::ConstBufferEngineInterface* engine = nullptr;

src/video_core/shader/decode.cpp

@@ -315,4 +315,25 @@ u32 ShaderIR::DecodeInstr(NodeBlock& bb, u32 pc) {
     return pc + 1;
 }
 
+void ShaderIR::PostDecode() {
+    // Deduce texture handler size if needed
+    auto* gpu_driver = locker.AccessGuestDriverProfile();
+    if (gpu_driver) {
+        if (!gpu_driver->TextureHandlerSizeKnown() && used_samplers.size() > 1) {
+            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));
+            }
+        }
+    }
+}
+
 } // 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);