10 files changed, 174 insertions, 64 deletions

diff --git a/src/video_core/engines/maxwell_3d.cpp b/src/video_core/engines/maxwell_3d.cpp
index a0dd7400d..50d8a94b1 100644
--- a/src/video_core/engines/maxwell_3d.cpp
+++ b/src/video_core/engines/maxwell_3d.cpp
@@ -182,8 +182,14 @@ u32 Maxwell3D::GetMaxCurrentVertices() {
 size_t Maxwell3D::EstimateIndexBufferSize() {
     GPUVAddr start_address = regs.index_buffer.StartAddress();
     GPUVAddr end_address = regs.index_buffer.EndAddress();
-    return std::min<size_t>(memory_manager.GetMemoryLayoutSize(start_address),
-                            static_cast<size_t>(end_address - start_address));
+    constexpr std::array<size_t, 4> max_sizes = {
+        std::numeric_limits<u8>::max(), std::numeric_limits<u16>::max(),
+        std::numeric_limits<u32>::max(), std::numeric_limits<u32>::max()};
+    const size_t byte_size = regs.index_buffer.FormatSizeInBytes();
+    return std::min<size_t>(
+        memory_manager.GetMemoryLayoutSize(start_address, byte_size * max_sizes[byte_size]) /
+            byte_size,
+        static_cast<size_t>(end_address - start_address));
 }
 
 u32 Maxwell3D::ProcessShadowRam(u32 method, u32 argument) {
@@ -572,4 +578,9 @@ u32 Maxwell3D::GetRegisterValue(u32 method) const {
     return regs.reg_array[method];
 }
 
+void Maxwell3D::setHLEReplacementName(u32 bank, u32 offset, HLEReplaceName name) {
+    const u64 key = (static_cast<u64>(bank) << 32) | offset;
+    replace_table.emplace(key, name);
+}
+
 } // namespace Tegra::Engines
diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h
index cfe1e4883..397e88f67 100644
--- a/src/video_core/engines/maxwell_3d.h
+++ b/src/video_core/engines/maxwell_3d.h
@@ -3020,6 +3020,23 @@ public:
     /// Store temporary hw register values, used by some calls to restore state after a operation
     Regs shadow_state;
 
+    // None Engine
+    enum class EngineHint : u32 {
+        None = 0x0,
+        OnHLEMacro = 0x1,
+    };
+
+    EngineHint engine_state{EngineHint::None};
+
+    enum class HLEReplaceName : u32 {
+        BaseVertex = 0x0,
+        BaseInstance = 0x1,
+    };
+
+    void setHLEReplacementName(u32 bank, u32 offset, HLEReplaceName name);
+
+    std::unordered_map<u64, HLEReplaceName> replace_table;
+
     static_assert(sizeof(Regs) == Regs::NUM_REGS * sizeof(u32), "Maxwell3D Regs has wrong size");
     static_assert(std::is_trivially_copyable_v<Regs>, "Maxwell3D Regs must be trivially copyable");
 
diff --git a/src/video_core/macro/macro_hle.cpp b/src/video_core/macro/macro_hle.cpp
index 93b6d42a4..638247e55 100644
--- a/src/video_core/macro/macro_hle.cpp
+++ b/src/video_core/macro/macro_hle.cpp
@@ -14,26 +14,29 @@
 #include "video_core/rasterizer_interface.h"
 
 namespace Tegra {
+
+using Maxwell = Engines::Maxwell3D;
+
 namespace {
 
-bool IsTopologySafe(Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology topology) {
+bool IsTopologySafe(Maxwell::Regs::PrimitiveTopology topology) {
     switch (topology) {
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::Points:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::Lines:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::LineLoop:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::LineStrip:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::Triangles:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::TriangleStrip:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::TriangleFan:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::LinesAdjacency:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::LineStripAdjacency:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::TrianglesAdjacency:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::TriangleStripAdjacency:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::Patches:
+    case Maxwell::Regs::PrimitiveTopology::Points:
+    case Maxwell::Regs::PrimitiveTopology::Lines:
+    case Maxwell::Regs::PrimitiveTopology::LineLoop:
+    case Maxwell::Regs::PrimitiveTopology::LineStrip:
+    case Maxwell::Regs::PrimitiveTopology::Triangles:
+    case Maxwell::Regs::PrimitiveTopology::TriangleStrip:
+    case Maxwell::Regs::PrimitiveTopology::TriangleFan:
+    case Maxwell::Regs::PrimitiveTopology::LinesAdjacency:
+    case Maxwell::Regs::PrimitiveTopology::LineStripAdjacency:
+    case Maxwell::Regs::PrimitiveTopology::TrianglesAdjacency:
+    case Maxwell::Regs::PrimitiveTopology::TriangleStripAdjacency:
+    case Maxwell::Regs::PrimitiveTopology::Patches:
         return true;
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::Quads:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::QuadStrip:
-    case Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology::Polygon:
+    case Maxwell::Regs::PrimitiveTopology::Quads:
+    case Maxwell::Regs::PrimitiveTopology::QuadStrip:
+    case Maxwell::Regs::PrimitiveTopology::Polygon:
     default:
         return false;
     }
@@ -82,8 +85,7 @@ public:
         : HLEMacroImpl(maxwell3d_), extended(extended_) {}
 
     void Execute(const std::vector<u32>& parameters, [[maybe_unused]] u32 method) override {
-        auto topology =
-            static_cast<Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology>(parameters[0]);
+        auto topology = static_cast<Maxwell::Regs::PrimitiveTopology>(parameters[0]);
         if (!IsTopologySafe(topology)) {
             Fallback(parameters);
             return;
@@ -99,18 +101,16 @@ public:
         params.stride = 0;
 
         if (extended) {
-            maxwell3d.CallMethod(0x8e3, 0x640, true);
-            maxwell3d.CallMethod(0x8e4, parameters[4], true);
+            maxwell3d.engine_state = Maxwell::EngineHint::OnHLEMacro;
+            maxwell3d.setHLEReplacementName(0, 0x640, Maxwell::HLEReplaceName::BaseInstance);
         }
 
         maxwell3d.draw_manager->DrawArrayIndirect(topology);
 
         if (extended) {
-            maxwell3d.CallMethod(0x8e3, 0x640, true);
-            maxwell3d.CallMethod(0x8e4, 0, true);
+            maxwell3d.engine_state = Maxwell::EngineHint::None;
+            maxwell3d.replace_table.clear();
         }
-        maxwell3d.regs.vertex_buffer.first = 0;
-        maxwell3d.regs.vertex_buffer.count = 0;
     }
 
 private:
@@ -134,13 +134,18 @@ private:
 
         const u32 base_instance = parameters[4];
         if (extended) {
-            maxwell3d.CallMethod(0x8e3, 0x640, true);
-            maxwell3d.CallMethod(0x8e4, base_instance, true);
+            maxwell3d.engine_state = Maxwell::EngineHint::OnHLEMacro;
+            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);
+
+        if (extended) {
+            maxwell3d.engine_state = Maxwell::EngineHint::None;
+            maxwell3d.replace_table.clear();
+        }
     }
 
     bool extended;
@@ -151,8 +156,7 @@ public:
     explicit HLE_DrawIndexedIndirect(Engines::Maxwell3D& maxwell3d_) : HLEMacroImpl(maxwell3d_) {}
 
     void Execute(const std::vector<u32>& parameters, [[maybe_unused]] u32 method) override {
-        auto topology =
-            static_cast<Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology>(parameters[0]);
+        auto topology = static_cast<Maxwell::Regs::PrimitiveTopology>(parameters[0]);
         if (!IsTopologySafe(topology)) {
             Fallback(parameters);
             return;
@@ -164,16 +168,12 @@ public:
             minimum_limit = std::max(parameters[3], minimum_limit);
         }
         const u32 estimate = static_cast<u32>(maxwell3d.EstimateIndexBufferSize());
-        const u32 base_size = std::max(minimum_limit, estimate);
-        const u32 element_base = parameters[4];
-        const u32 base_instance = parameters[5];
-        maxwell3d.regs.index_buffer.first = 0;
-        maxwell3d.regs.index_buffer.count = base_size; // Use a fixed size, just for mapping
+        const u32 base_size = std::max<u32>(minimum_limit, estimate);
         maxwell3d.regs.draw.topology.Assign(topology);
         maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
-        maxwell3d.CallMethod(0x8e3, 0x640, true);
-        maxwell3d.CallMethod(0x8e4, element_base, true);
-        maxwell3d.CallMethod(0x8e5, base_instance, true);
+        maxwell3d.engine_state = Maxwell::EngineHint::OnHLEMacro;
+        maxwell3d.setHLEReplacementName(0, 0x640, Maxwell::HLEReplaceName::BaseVertex);
+        maxwell3d.setHLEReplacementName(0, 0x644, Maxwell::HLEReplaceName::BaseInstance);
         auto& params = maxwell3d.draw_manager->GetIndirectParams();
         params.is_indexed = true;
         params.include_count = false;
@@ -184,9 +184,8 @@ public:
         params.stride = 0;
         maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
         maxwell3d.draw_manager->DrawIndexedIndirect(topology, 0, base_size);
-        maxwell3d.CallMethod(0x8e3, 0x640, true);
-        maxwell3d.CallMethod(0x8e4, 0x0, true);
-        maxwell3d.CallMethod(0x8e5, 0x0, true);
+        maxwell3d.engine_state = Maxwell::EngineHint::None;
+        maxwell3d.replace_table.clear();
     }
 
 private:
@@ -197,18 +196,17 @@ private:
         const u32 base_instance = parameters[5];
         maxwell3d.regs.vertex_id_base = element_base;
         maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
-        maxwell3d.CallMethod(0x8e3, 0x640, true);
-        maxwell3d.CallMethod(0x8e4, element_base, true);
-        maxwell3d.CallMethod(0x8e5, base_instance, true);
+        maxwell3d.engine_state = Maxwell::EngineHint::OnHLEMacro;
+        maxwell3d.setHLEReplacementName(0, 0x640, Maxwell::HLEReplaceName::BaseVertex);
+        maxwell3d.setHLEReplacementName(0, 0x644, Maxwell::HLEReplaceName::BaseInstance);
 
         maxwell3d.draw_manager->DrawIndex(
             static_cast<Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology>(parameters[0]),
             parameters[3], parameters[1], element_base, base_instance, instance_count);
 
         maxwell3d.regs.vertex_id_base = 0x0;
-        maxwell3d.CallMethod(0x8e3, 0x640, true);
-        maxwell3d.CallMethod(0x8e4, 0x0, true);
-        maxwell3d.CallMethod(0x8e5, 0x0, true);
+        maxwell3d.engine_state = Maxwell::EngineHint::None;
+        maxwell3d.replace_table.clear();
     }
 
     u32 minimum_limit{1 << 18};
@@ -238,8 +236,7 @@ public:
         : HLEMacroImpl(maxwell3d_) {}
 
     void Execute(const std::vector<u32>& parameters, [[maybe_unused]] u32 method) override {
-        const auto topology =
-            static_cast<Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology>(parameters[2]);
+        const auto topology = static_cast<Maxwell::Regs::PrimitiveTopology>(parameters[2]);
         if (!IsTopologySafe(topology)) {
             Fallback(parameters);
             return;
@@ -277,9 +274,6 @@ public:
         }
         const u32 estimate = static_cast<u32>(maxwell3d.EstimateIndexBufferSize());
         const u32 base_size = std::max(minimum_limit, estimate);
-
-        maxwell3d.regs.index_buffer.first = 0;
-        maxwell3d.regs.index_buffer.count = std::max(highest_limit, base_size);
         maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
         auto& params = maxwell3d.draw_manager->GetIndirectParams();
         params.is_indexed = true;
@@ -290,7 +284,12 @@ public:
         params.max_draw_counts = draw_count;
         params.stride = stride;
         maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
-        maxwell3d.draw_manager->DrawIndexedIndirect(topology, 0, highest_limit);
+        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.engine_state = Maxwell::EngineHint::None;
+        maxwell3d.replace_table.clear();
     }
 
 private:
@@ -299,9 +298,8 @@ private:
             // Clean everything.
             // Clean everything.
             maxwell3d.regs.vertex_id_base = 0x0;
-            maxwell3d.CallMethod(0x8e3, 0x640, true);
-            maxwell3d.CallMethod(0x8e4, 0x0, true);
-            maxwell3d.CallMethod(0x8e5, 0x0, true);
+            maxwell3d.engine_state = Maxwell::EngineHint::None;
+            maxwell3d.replace_table.clear();
         });
         maxwell3d.RefreshParameters();
         const u32 start_indirect = parameters[0];
@@ -310,8 +308,7 @@ private:
             // Nothing to do.
             return;
         }
-        const auto topology =
-            static_cast<Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology>(parameters[2]);
+        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];
@@ -326,9 +323,9 @@ private:
             const u32 base_vertex = parameters[base + 3];
             const u32 base_instance = parameters[base + 4];
             maxwell3d.regs.vertex_id_base = base_vertex;
-            maxwell3d.CallMethod(0x8e3, 0x640, true);
-            maxwell3d.CallMethod(0x8e4, base_vertex, true);
-            maxwell3d.CallMethod(0x8e5, base_instance, true);
+            maxwell3d.engine_state = Maxwell::EngineHint::OnHLEMacro;
+            maxwell3d.setHLEReplacementName(0, 0x640, Maxwell::HLEReplaceName::BaseVertex);
+            maxwell3d.setHLEReplacementName(0, 0x644, Maxwell::HLEReplaceName::BaseInstance);
             maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
             maxwell3d.draw_manager->DrawIndex(topology, parameters[base + 2], parameters[base],
                                               base_vertex, base_instance, parameters[base + 1]);
diff --git a/src/video_core/memory_manager.cpp b/src/video_core/memory_manager.cpp
index 8f6c51045..11e7d225e 100644
--- a/src/video_core/memory_manager.cpp
+++ b/src/video_core/memory_manager.cpp
@@ -577,7 +577,7 @@ size_t MemoryManager::MaxContinousRange(GPUVAddr gpu_addr, size_t size) const {
     return range_so_far;
 }
 
-size_t MemoryManager::GetMemoryLayoutSize(GPUVAddr gpu_addr) const {
+size_t MemoryManager::GetMemoryLayoutSize(GPUVAddr gpu_addr, size_t max_size) const {
     PTEKind base_kind = GetPageKind(gpu_addr);
     if (base_kind == PTEKind::INVALID) {
         return 0;
@@ -596,6 +596,10 @@ size_t MemoryManager::GetMemoryLayoutSize(GPUVAddr gpu_addr) const {
             return true;
         }
         range_so_far += copy_amount;
+        if (range_so_far >= max_size) {
+            result = true;
+            return true;
+        }
         return false;
     };
     auto big_check = [&](std::size_t page_index, std::size_t offset, std::size_t copy_amount) {
@@ -605,6 +609,10 @@ size_t MemoryManager::GetMemoryLayoutSize(GPUVAddr gpu_addr) const {
             return true;
         }
         range_so_far += copy_amount;
+        if (range_so_far >= max_size) {
+            result = true;
+            return true;
+        }
         return false;
     };
     auto check_short_pages = [&](std::size_t page_index, std::size_t offset,
diff --git a/src/video_core/memory_manager.h b/src/video_core/memory_manager.h
index 65f6e8134..ca22520d7 100644
--- a/src/video_core/memory_manager.h
+++ b/src/video_core/memory_manager.h
@@ -118,7 +118,8 @@ public:
 
     PTEKind GetPageKind(GPUVAddr gpu_addr) const;
 
-    size_t GetMemoryLayoutSize(GPUVAddr gpu_addr) const;
+    size_t GetMemoryLayoutSize(GPUVAddr gpu_addr,
+                               size_t max_size = std::numeric_limits<size_t>::max()) const;
 
 private:
     template <bool is_big_pages, typename FuncMapped, typename FuncReserved, typename FuncUnmapped>
diff --git a/src/video_core/renderer_vulkan/fixed_pipeline_state.cpp b/src/video_core/renderer_vulkan/fixed_pipeline_state.cpp
index e62b36822..df229f41b 100644
--- a/src/video_core/renderer_vulkan/fixed_pipeline_state.cpp
+++ b/src/video_core/renderer_vulkan/fixed_pipeline_state.cpp
@@ -97,6 +97,7 @@ void FixedPipelineState::Refresh(Tegra::Engines::Maxwell3D& maxwell3d,
     smooth_lines.Assign(regs.line_anti_alias_enable != 0 ? 1 : 0);
     alpha_to_coverage_enabled.Assign(regs.anti_alias_alpha_control.alpha_to_coverage != 0 ? 1 : 0);
     alpha_to_one_enabled.Assign(regs.anti_alias_alpha_control.alpha_to_one != 0 ? 1 : 0);
+    app_stage.Assign(maxwell3d.engine_state);
 
     for (size_t i = 0; i < regs.rt.size(); ++i) {
         color_formats[i] = static_cast<u8>(regs.rt[i].format);
diff --git a/src/video_core/renderer_vulkan/fixed_pipeline_state.h b/src/video_core/renderer_vulkan/fixed_pipeline_state.h
index ab79fb8f3..03bf64b57 100644
--- a/src/video_core/renderer_vulkan/fixed_pipeline_state.h
+++ b/src/video_core/renderer_vulkan/fixed_pipeline_state.h
@@ -197,6 +197,7 @@ struct FixedPipelineState {
         BitField<14, 1, u32> smooth_lines;
         BitField<15, 1, u32> alpha_to_coverage_enabled;
         BitField<16, 1, u32> alpha_to_one_enabled;
+        BitField<17, 3, Tegra::Engines::Maxwell3D::EngineHint> app_stage;
     };
     std::array<u8, Maxwell::NumRenderTargets> color_formats;
 
diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
index e7262420c..58b955821 100644
--- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
@@ -54,7 +54,7 @@ using VideoCommon::FileEnvironment;
 using VideoCommon::GenericEnvironment;
 using VideoCommon::GraphicsEnvironment;
 
-constexpr u32 CACHE_VERSION = 8;
+constexpr u32 CACHE_VERSION = 9;
 
 template <typename Container>
 auto MakeSpan(Container& container) {
diff --git a/src/video_core/shader_environment.cpp b/src/video_core/shader_environment.cpp
index 958810747..99d85bfb3 100644
--- a/src/video_core/shader_environment.cpp
+++ b/src/video_core/shader_environment.cpp
@@ -202,12 +202,15 @@ void GenericEnvironment::Serialize(std::ofstream& file) const {
     const u64 num_texture_types{static_cast<u64>(texture_types.size())};
     const u64 num_texture_pixel_formats{static_cast<u64>(texture_pixel_formats.size())};
     const u64 num_cbuf_values{static_cast<u64>(cbuf_values.size())};
+    const u64 num_cbuf_replacement_values{static_cast<u64>(cbuf_replacements.size())};
 
     file.write(reinterpret_cast<const char*>(&code_size), sizeof(code_size))
         .write(reinterpret_cast<const char*>(&num_texture_types), sizeof(num_texture_types))
         .write(reinterpret_cast<const char*>(&num_texture_pixel_formats),
                sizeof(num_texture_pixel_formats))
         .write(reinterpret_cast<const char*>(&num_cbuf_values), sizeof(num_cbuf_values))
+        .write(reinterpret_cast<const char*>(&num_cbuf_replacement_values),
+               sizeof(num_cbuf_replacement_values))
         .write(reinterpret_cast<const char*>(&local_memory_size), sizeof(local_memory_size))
         .write(reinterpret_cast<const char*>(&texture_bound), sizeof(texture_bound))
         .write(reinterpret_cast<const char*>(&start_address), sizeof(start_address))
@@ -229,6 +232,10 @@ void GenericEnvironment::Serialize(std::ofstream& file) const {
         file.write(reinterpret_cast<const char*>(&key), sizeof(key))
             .write(reinterpret_cast<const char*>(&type), sizeof(type));
     }
+    for (const auto& [key, type] : cbuf_replacements) {
+        file.write(reinterpret_cast<const char*>(&key), sizeof(key))
+            .write(reinterpret_cast<const char*>(&type), sizeof(type));
+    }
     if (stage == Shader::Stage::Compute) {
         file.write(reinterpret_cast<const char*>(&workgroup_size), sizeof(workgroup_size))
             .write(reinterpret_cast<const char*>(&shared_memory_size), sizeof(shared_memory_size));
@@ -318,6 +325,8 @@ GraphicsEnvironment::GraphicsEnvironment(Tegra::Engines::Maxwell3D& maxwell3d_,
     ASSERT(local_size <= std::numeric_limits<u32>::max());
     local_memory_size = static_cast<u32>(local_size) + sph.common3.shader_local_memory_crs_size;
     texture_bound = maxwell3d->regs.bindless_texture_const_buffer_slot;
+    has_hle_engine_state =
+        maxwell3d->engine_state == Tegra::Engines::Maxwell3D::EngineHint::OnHLEMacro;
 }
 
 u32 GraphicsEnvironment::ReadCbufValue(u32 cbuf_index, u32 cbuf_offset) {
@@ -331,6 +340,30 @@ u32 GraphicsEnvironment::ReadCbufValue(u32 cbuf_index, u32 cbuf_offset) {
     return value;
 }
 
+std::optional<Shader::ReplaceConstant> GraphicsEnvironment::GetReplaceConstBuffer(u32 bank,
+                                                                                  u32 offset) {
+    if (!has_hle_engine_state) {
+        return std::nullopt;
+    }
+    const u64 key = (static_cast<u64>(bank) << 32) | static_cast<u64>(offset);
+    auto it = maxwell3d->replace_table.find(key);
+    if (it == maxwell3d->replace_table.end()) {
+        return std::nullopt;
+    }
+    const auto converted_value = [](Tegra::Engines::Maxwell3D::HLEReplaceName name) {
+        switch (name) {
+        case Tegra::Engines::Maxwell3D::HLEReplaceName::BaseVertex:
+            return Shader::ReplaceConstant::BaseVertex;
+        case Tegra::Engines::Maxwell3D::HLEReplaceName::BaseInstance:
+            return Shader::ReplaceConstant::BaseInstance;
+        default:
+            UNREACHABLE();
+        }
+    }(it->second);
+    cbuf_replacements.emplace(key, converted_value);
+    return converted_value;
+}
+
 Shader::TextureType GraphicsEnvironment::ReadTextureType(u32 handle) {
     const auto& regs{maxwell3d->regs};
     const bool via_header_index{regs.sampler_binding == Maxwell::SamplerBinding::ViaHeaderBinding};
@@ -409,11 +442,14 @@ void FileEnvironment::Deserialize(std::ifstream& file) {
     u64 num_texture_types{};
     u64 num_texture_pixel_formats{};
     u64 num_cbuf_values{};
+    u64 num_cbuf_replacement_values{};
     file.read(reinterpret_cast<char*>(&code_size), sizeof(code_size))
         .read(reinterpret_cast<char*>(&num_texture_types), sizeof(num_texture_types))
         .read(reinterpret_cast<char*>(&num_texture_pixel_formats),
               sizeof(num_texture_pixel_formats))
         .read(reinterpret_cast<char*>(&num_cbuf_values), sizeof(num_cbuf_values))
+        .read(reinterpret_cast<char*>(&num_cbuf_replacement_values),
+              sizeof(num_cbuf_replacement_values))
         .read(reinterpret_cast<char*>(&local_memory_size), sizeof(local_memory_size))
         .read(reinterpret_cast<char*>(&texture_bound), sizeof(texture_bound))
         .read(reinterpret_cast<char*>(&start_address), sizeof(start_address))
@@ -444,6 +480,13 @@ void FileEnvironment::Deserialize(std::ifstream& file) {
             .read(reinterpret_cast<char*>(&value), sizeof(value));
         cbuf_values.emplace(key, value);
     }
+    for (size_t i = 0; i < num_cbuf_replacement_values; ++i) {
+        u64 key;
+        Shader::ReplaceConstant value;
+        file.read(reinterpret_cast<char*>(&key), sizeof(key))
+            .read(reinterpret_cast<char*>(&value), sizeof(value));
+        cbuf_replacements.emplace(key, value);
+    }
     if (stage == Shader::Stage::Compute) {
         file.read(reinterpret_cast<char*>(&workgroup_size), sizeof(workgroup_size))
             .read(reinterpret_cast<char*>(&shared_memory_size), sizeof(shared_memory_size));
@@ -512,6 +555,16 @@ std::array<u32, 3> FileEnvironment::WorkgroupSize() const {
     return workgroup_size;
 }
 
+std::optional<Shader::ReplaceConstant> FileEnvironment::GetReplaceConstBuffer(u32 bank,
+                                                                              u32 offset) {
+    const u64 key = (static_cast<u64>(bank) << 32) | static_cast<u64>(offset);
+    auto it = cbuf_replacements.find(key);
+    if (it == cbuf_replacements.end()) {
+        return std::nullopt;
+    }
+    return it->second;
+}
+
 void SerializePipeline(std::span<const char> key, std::span<const GenericEnvironment* const> envs,
                        const std::filesystem::path& filename, u32 cache_version) try {
     std::ofstream file(filename, std::ios::binary | std::ios::ate | std::ios::app);
diff --git a/src/video_core/shader_environment.h b/src/video_core/shader_environment.h
index 1342fab1e..d75987a52 100644
--- a/src/video_core/shader_environment.h
+++ b/src/video_core/shader_environment.h
@@ -60,6 +60,10 @@ public:
 
     void Serialize(std::ofstream& file) const;
 
+    bool HasHLEMacroState() const override {
+        return has_hle_engine_state;
+    }
+
 protected:
     std::optional<u64> TryFindSize();
 
@@ -73,6 +77,7 @@ protected:
     std::unordered_map<u32, Shader::TextureType> texture_types;
     std::unordered_map<u32, Shader::TexturePixelFormat> texture_pixel_formats;
     std::unordered_map<u64, u32> cbuf_values;
+    std::unordered_map<u64, Shader::ReplaceConstant> cbuf_replacements;
 
     u32 local_memory_size{};
     u32 texture_bound{};
@@ -89,6 +94,7 @@ protected:
     u32 viewport_transform_state = 1;
 
     bool has_unbound_instructions = false;
+    bool has_hle_engine_state = false;
 };
 
 class GraphicsEnvironment final : public GenericEnvironment {
@@ -109,6 +115,8 @@ public:
 
     u32 ReadViewportTransformState() override;
 
+    std::optional<Shader::ReplaceConstant> GetReplaceConstBuffer(u32 bank, u32 offset) override;
+
 private:
     Tegra::Engines::Maxwell3D* maxwell3d{};
     size_t stage_index{};
@@ -131,6 +139,11 @@ public:
 
     u32 ReadViewportTransformState() override;
 
+    std::optional<Shader::ReplaceConstant> GetReplaceConstBuffer(
+        [[maybe_unused]] u32 bank, [[maybe_unused]] u32 offset) override {
+        return std::nullopt;
+    }
+
 private:
     Tegra::Engines::KeplerCompute* kepler_compute{};
 };
@@ -166,6 +179,13 @@ public:
 
     [[nodiscard]] std::array<u32, 3> WorkgroupSize() const override;
 
+    [[nodiscard]] std::optional<Shader::ReplaceConstant> GetReplaceConstBuffer(u32 bank,
+                                                                               u32 offset) override;
+
+    [[nodiscard]] bool HasHLEMacroState() const override {
+        return cbuf_replacements.size() != 0;
+    }
+
     void Dump(u64 hash) override;
 
 private:
@@ -173,6 +193,7 @@ private:
     std::unordered_map<u32, Shader::TextureType> texture_types;
     std::unordered_map<u32, Shader::TexturePixelFormat> texture_pixel_formats;
     std::unordered_map<u64, u32> cbuf_values;
+    std::unordered_map<u64, Shader::ReplaceConstant> cbuf_replacements;
     std::array<u32, 3> workgroup_size{};
     u32 local_memory_size{};
     u32 shared_memory_size{};