renderer_opengl: Add assembly program code paths

Add code required to use OpenGL assembly programs based on
NV_gpu_program5. Decompilation for ARB programs is intended to be added
in a follow up commit. This does **not** include ARB decompilation and
it's not in an usable state.

The intention behind assembly programs is to reduce shader stutter
significantly on drivers supporting NV_gpu_program5 (and other required
extensions). Currently only Nvidia's proprietary driver supports these
extensions.

Add a UI option hidden for now to avoid people enabling this option
accidentally.

This code path has some limitations that OpenGL compatibility doesn't
have:
- NV_shader_storage_buffer_object is limited to 16 entries for a single
OpenGL context state (I don't know if this is an intended limitation, an
specification issue or I am missing something). Currently causes issues
on The Legend of Zelda: Link's Awakening.
- NV_parameter_buffer_object can't bind buffers using an offset
different to zero. The used workaround is to copy to a temporary buffer
(this doesn't happen often so it's not an issue).

On the other hand, it has the following advantages:
- Shaders build a lot faster.
- We have control over how floating point rounding is done over
individual instructions (SPIR-V on Vulkan can't do this).
- Operations on shared memory can be unsigned and signed.
- Transform feedbacks are dynamic state (not yet implemented).
- Parameter buffers (uniform buffers) are per stage, matching NVN and
hardware's behavior.
- The API to bind and create assembly programs makes sense, unlike
ARB_separate_shader_objects.

src/video_core/renderer_opengl/gl_device.cpp

@@ -13,6 +13,7 @@
 
 #include "common/logging/log.h"
 #include "common/scope_exit.h"
+#include "core/settings.h"
 #include "video_core/renderer_opengl/gl_device.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 
@@ -183,10 +184,16 @@ Device::Device() : base_bindings{BuildBaseBindings()} {
     has_precise_bug = TestPreciseBug();
     has_broken_compute = is_intel_proprietary;
     has_fast_buffer_sub_data = is_nvidia;
+    use_assembly_shaders = Settings::values.use_assembly_shaders && GLAD_GL_NV_gpu_program5 &&
+                           GLAD_GL_NV_compute_program5;
 
     LOG_INFO(Render_OpenGL, "Renderer_VariableAOFFI: {}", has_variable_aoffi);
     LOG_INFO(Render_OpenGL, "Renderer_ComponentIndexingBug: {}", has_component_indexing_bug);
     LOG_INFO(Render_OpenGL, "Renderer_PreciseBug: {}", has_precise_bug);
+
+    if (Settings::values.use_assembly_shaders && !use_assembly_shaders) {
+        LOG_ERROR(Render_OpenGL, "Assembly shaders enabled but not supported");
+    }
 }
 
 Device::Device(std::nullptr_t) {

src/video_core/renderer_opengl/gl_device.h

@@ -88,6 +88,10 @@ public:
         return has_fast_buffer_sub_data;
     }
 
+    bool UseAssemblyShaders() const {
+        return use_assembly_shaders;
+    }
+
 private:
     static bool TestVariableAoffi();
     static bool TestPreciseBug();
@@ -107,6 +111,7 @@ private:
     bool has_precise_bug{};
     bool has_broken_compute{};
     bool has_fast_buffer_sub_data{};
+    bool use_assembly_shaders{};
 };
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -94,17 +94,30 @@ void oglEnable(GLenum cap, bool state) {
 } // Anonymous namespace
 
 RasterizerOpenGL::RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWindow& emu_window,
-                                   ScreenInfo& info, GLShader::ProgramManager& program_manager,
+                                   const Device& device, ScreenInfo& info,
-                                   StateTracker& state_tracker)
+                                   ProgramManager& program_manager, StateTracker& state_tracker)
-    : RasterizerAccelerated{system.Memory()}, texture_cache{system, *this, device, state_tracker},
+    : RasterizerAccelerated{system.Memory()}, device{device}, texture_cache{system, *this, device,
+                                                                            state_tracker},
       shader_cache{*this, system, emu_window, device}, query_cache{system, *this},
       buffer_cache{*this, system, device, STREAM_BUFFER_SIZE},
       fence_manager{system, *this, texture_cache, buffer_cache, query_cache}, system{system},
       screen_info{info}, program_manager{program_manager}, state_tracker{state_tracker} {
     CheckExtensions();
+
+    if (device.UseAssemblyShaders()) {
+        glCreateBuffers(static_cast<GLsizei>(staging_cbufs.size()), staging_cbufs.data());
+        for (const GLuint cbuf : staging_cbufs) {
+            glNamedBufferStorage(cbuf, static_cast<GLsizeiptr>(Maxwell::MaxConstBufferSize),
+                                 nullptr, 0);
+        }
+    }
 }
 
-RasterizerOpenGL::~RasterizerOpenGL() {}
+RasterizerOpenGL::~RasterizerOpenGL() {
+    if (device.UseAssemblyShaders()) {
+        glDeleteBuffers(static_cast<GLsizei>(staging_cbufs.size()), staging_cbufs.data());
+    }
+}
 
 void RasterizerOpenGL::CheckExtensions() {
     if (!GLAD_GL_ARB_texture_filter_anisotropic && !GLAD_GL_EXT_texture_filter_anisotropic) {
@@ -230,6 +243,7 @@ GLintptr RasterizerOpenGL::SetupIndexBuffer() {
 void RasterizerOpenGL::SetupShaders(GLenum primitive_mode) {
     MICROPROFILE_SCOPE(OpenGL_Shader);
     auto& gpu = system.GPU().Maxwell3D();
+    std::size_t num_ssbos = 0;
     u32 clip_distances = 0;
 
     for (std::size_t index = 0; index < Maxwell::MaxShaderProgram; ++index) {
@@ -261,6 +275,14 @@ void RasterizerOpenGL::SetupShaders(GLenum primitive_mode) {
 
         Shader shader{shader_cache.GetStageProgram(program)};
 
+        if (device.UseAssemblyShaders()) {
+            // Check for ARB limitation. We only have 16 SSBOs per context state. To workaround this
+            // all stages share the same bindings.
+            const std::size_t num_stage_ssbos = shader->GetEntries().global_memory_entries.size();
+            ASSERT_MSG(num_stage_ssbos == 0 || num_ssbos == 0, "SSBOs on more than one stage");
+            num_ssbos += num_stage_ssbos;
+        }
+
         // Stage indices are 0 - 5
         const std::size_t stage = index == 0 ? 0 : index - 1;
         SetupDrawConstBuffers(stage, shader);
@@ -526,6 +548,7 @@ void RasterizerOpenGL::Draw(bool is_indexed, bool is_instanced) {
     SyncFramebufferSRGB();
 
     buffer_cache.Acquire();
+    current_cbuf = 0;
 
     std::size_t buffer_size = CalculateVertexArraysSize();
 
@@ -535,9 +558,9 @@ void RasterizerOpenGL::Draw(bool is_indexed, bool is_instanced) {
     }
 
     // Uniform space for the 5 shader stages
-    buffer_size = Common::AlignUp<std::size_t>(buffer_size, 4) +
+    buffer_size =
-                  (sizeof(GLShader::MaxwellUniformData) + device.GetUniformBufferAlignment()) *
+        Common::AlignUp<std::size_t>(buffer_size, 4) +
-                      Maxwell::MaxShaderStage;
+        (sizeof(MaxwellUniformData) + device.GetUniformBufferAlignment()) * Maxwell::MaxShaderStage;
 
     // Add space for at least 18 constant buffers
     buffer_size += Maxwell::MaxConstBuffers *
@@ -558,12 +581,14 @@ void RasterizerOpenGL::Draw(bool is_indexed, bool is_instanced) {
     }
 
     // Setup emulation uniform buffer.
-    GLShader::MaxwellUniformData ubo;
+    if (!device.UseAssemblyShaders()) {
+        MaxwellUniformData ubo;
         ubo.SetFromRegs(gpu);
         const auto [buffer, offset] =
             buffer_cache.UploadHostMemory(&ubo, sizeof(ubo), device.GetUniformBufferAlignment());
         glBindBufferRange(GL_UNIFORM_BUFFER, EmulationUniformBlockBinding, buffer, offset,
                           static_cast<GLsizeiptr>(sizeof(ubo)));
+    }
 
     // Setup shaders and their used resources.
     texture_cache.GuardSamplers(true);
@@ -635,11 +660,11 @@ void RasterizerOpenGL::DispatchCompute(GPUVAddr code_addr) {
     }
 
     buffer_cache.Acquire();
+    current_cbuf = 0;
 
     auto kernel = shader_cache.GetComputeKernel(code_addr);
     SetupComputeTextures(kernel);
     SetupComputeImages(kernel);
-    program_manager.BindComputeShader(kernel->GetHandle());
 
     const std::size_t buffer_size =
         Tegra::Engines::KeplerCompute::NumConstBuffers *
@@ -652,6 +677,7 @@ void RasterizerOpenGL::DispatchCompute(GPUVAddr code_addr) {
     buffer_cache.Unmap();
 
     const auto& launch_desc = system.GPU().KeplerCompute().launch_description;
+    program_manager.BindCompute(kernel->GetHandle());
     glDispatchCompute(launch_desc.grid_dim_x, launch_desc.grid_dim_y, launch_desc.grid_dim_z);
     ++num_queued_commands;
 }
@@ -812,14 +838,20 @@ bool RasterizerOpenGL::AccelerateDisplay(const Tegra::FramebufferConfig& config,
 }
 
 void RasterizerOpenGL::SetupDrawConstBuffers(std::size_t stage_index, const Shader& shader) {
+    static constexpr std::array PARAMETER_LUT = {
+        GL_VERTEX_PROGRAM_PARAMETER_BUFFER_NV, GL_TESS_CONTROL_PROGRAM_PARAMETER_BUFFER_NV,
+        GL_TESS_EVALUATION_PROGRAM_PARAMETER_BUFFER_NV, GL_GEOMETRY_PROGRAM_PARAMETER_BUFFER_NV,
+        GL_FRAGMENT_PROGRAM_PARAMETER_BUFFER_NV};
+
     MICROPROFILE_SCOPE(OpenGL_UBO);
     const auto& stages = system.GPU().Maxwell3D().state.shader_stages;
     const auto& shader_stage = stages[stage_index];
 
-    u32 binding = device.GetBaseBindings(stage_index).uniform_buffer;
+    u32 binding =
+        device.UseAssemblyShaders() ? 0 : device.GetBaseBindings(stage_index).uniform_buffer;
     for (const auto& entry : shader->GetEntries().const_buffers) {
         const auto& buffer = shader_stage.const_buffers[entry.GetIndex()];
-        SetupConstBuffer(binding++, buffer, entry);
+        SetupConstBuffer(PARAMETER_LUT[stage_index], binding++, buffer, entry);
     }
 }
 
@@ -835,16 +867,21 @@ void RasterizerOpenGL::SetupComputeConstBuffers(const Shader& kernel) {
         buffer.address = config.Address();
         buffer.size = config.size;
         buffer.enabled = mask[entry.GetIndex()];
-        SetupConstBuffer(binding++, buffer, entry);
+        SetupConstBuffer(GL_COMPUTE_PROGRAM_PARAMETER_BUFFER_NV, binding++, buffer, entry);
     }
 }
 
-void RasterizerOpenGL::SetupConstBuffer(u32 binding, const Tegra::Engines::ConstBufferInfo& buffer,
+void RasterizerOpenGL::SetupConstBuffer(GLenum stage, u32 binding,
+                                        const Tegra::Engines::ConstBufferInfo& buffer,
                                         const ConstBufferEntry& entry) {
     if (!buffer.enabled) {
         // Set values to zero to unbind buffers
-        glBindBufferRange(GL_UNIFORM_BUFFER, binding, buffer_cache.GetEmptyBuffer(sizeof(float)), 0,
+        if (device.UseAssemblyShaders()) {
-                          sizeof(float));
+            glBindBufferRangeNV(stage, entry.GetIndex(), 0, 0, 0);
+        } else {
+            glBindBufferRange(GL_UNIFORM_BUFFER, binding,
+                              buffer_cache.GetEmptyBuffer(sizeof(float)), 0, sizeof(float));
+        }
         return;
     }
 
@@ -853,9 +890,19 @@ void RasterizerOpenGL::SetupConstBuffer(u32 binding, const Tegra::Engines::Const
     const std::size_t size = Common::AlignUp(GetConstBufferSize(buffer, entry), sizeof(GLvec4));
 
     const auto alignment = device.GetUniformBufferAlignment();
-    const auto [cbuf, offset] = buffer_cache.UploadMemory(buffer.address, size, alignment, false,
+    auto [cbuf, offset] = buffer_cache.UploadMemory(buffer.address, size, alignment, false,
                                                     device.HasFastBufferSubData());
+    if (!device.UseAssemblyShaders()) {
         glBindBufferRange(GL_UNIFORM_BUFFER, binding, cbuf, offset, size);
+        return;
+    }
+    if (offset != 0) {
+        const GLuint staging_cbuf = staging_cbufs[current_cbuf++];
+        glCopyNamedBufferSubData(cbuf, staging_cbuf, offset, 0, size);
+        cbuf = staging_cbuf;
+        offset = 0;
+    }
+    glBindBufferRangeNV(stage, binding, cbuf, offset, size);
 }
 
 void RasterizerOpenGL::SetupDrawGlobalMemory(std::size_t stage_index, const Shader& shader) {
@@ -863,7 +910,8 @@ void RasterizerOpenGL::SetupDrawGlobalMemory(std::size_t stage_index, const Shad
     auto& memory_manager{gpu.MemoryManager()};
     const auto cbufs{gpu.Maxwell3D().state.shader_stages[stage_index]};
 
-    u32 binding = device.GetBaseBindings(stage_index).shader_storage_buffer;
+    u32 binding =
+        device.UseAssemblyShaders() ? 0 : device.GetBaseBindings(stage_index).shader_storage_buffer;
     for (const auto& entry : shader->GetEntries().global_memory_entries) {
         const GPUVAddr addr{cbufs.const_buffers[entry.cbuf_index].address + entry.cbuf_offset};
         const GPUVAddr gpu_addr{memory_manager.Read<u64>(addr)};

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -56,8 +56,8 @@ struct DrawParameters;
 class RasterizerOpenGL : public VideoCore::RasterizerAccelerated {
 public:
     explicit RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWindow& emu_window,
-                              ScreenInfo& info, GLShader::ProgramManager& program_manager,
+                              const Device& device, ScreenInfo& info,
-                              StateTracker& state_tracker);
+                              ProgramManager& program_manager, StateTracker& state_tracker);
     ~RasterizerOpenGL() override;
 
     void Draw(bool is_indexed, bool is_instanced) override;
@@ -106,7 +106,7 @@ private:
     void SetupComputeConstBuffers(const Shader& kernel);
 
     /// Configures a constant buffer.
-    void SetupConstBuffer(u32 binding, const Tegra::Engines::ConstBufferInfo& buffer,
+    void SetupConstBuffer(GLenum stage, u32 binding, const Tegra::Engines::ConstBufferInfo& buffer,
                           const ConstBufferEntry& entry);
 
     /// Configures the current global memory entries to use for the draw command.
@@ -224,7 +224,7 @@ private:
 
     void SetupShaders(GLenum primitive_mode);
 
-    const Device device;
+    const Device& device;
 
     TextureCacheOpenGL texture_cache;
     ShaderCacheOpenGL shader_cache;
@@ -236,7 +236,7 @@ private:
 
     Core::System& system;
     ScreenInfo& screen_info;
-    GLShader::ProgramManager& program_manager;
+    ProgramManager& program_manager;
     StateTracker& state_tracker;
 
     static constexpr std::size_t STREAM_BUFFER_SIZE = 128 * 1024 * 1024;
@@ -248,6 +248,12 @@ private:
     std::bitset<Tegra::Engines::Maxwell3D::Regs::NumTransformFeedbackBuffers>
         enabled_transform_feedback_buffers;
 
+    static constexpr std::size_t NUM_CONSTANT_BUFFERS =
+        Tegra::Engines::Maxwell3D::Regs::MaxConstBuffers *
+        Tegra::Engines::Maxwell3D::Regs::MaxShaderProgram;
+    std::array<GLuint, NUM_CONSTANT_BUFFERS> staging_cbufs{};
+    std::size_t current_cbuf = 0;
+
     /// Number of commands queued to the OpenGL driver. Reseted on flush.
     std::size_t num_queued_commands = 0;
 

src/video_core/renderer_opengl/gl_resource_manager.cpp

@@ -125,6 +125,15 @@ void OGLProgram::Release() {
     handle = 0;
 }
 
+void OGLAssemblyProgram::Release() {
+    if (handle == 0) {
+        return;
+    }
+    MICROPROFILE_SCOPE(OpenGL_ResourceDeletion);
+    glDeleteProgramsARB(1, &handle);
+    handle = 0;
+}
+
 void OGLPipeline::Create() {
     if (handle != 0)
         return;

src/video_core/renderer_opengl/gl_resource_manager.h

@@ -167,6 +167,22 @@ public:
     GLuint handle = 0;
 };
 
+class OGLAssemblyProgram : private NonCopyable {
+public:
+    OGLAssemblyProgram() = default;
+
+    OGLAssemblyProgram(OGLAssemblyProgram&& o) noexcept : handle(std::exchange(o.handle, 0)) {}
+
+    ~OGLAssemblyProgram() {
+        Release();
+    }
+
+    /// Deletes the internal OpenGL resource
+    void Release();
+
+    GLuint handle = 0;
+};
+
 class OGLPipeline : private NonCopyable {
 public:
     OGLPipeline() = default;

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -97,6 +97,24 @@ constexpr ShaderType GetShaderType(Maxwell::ShaderProgram program_type) {
     return {};
 }
 
+constexpr GLenum AssemblyEnum(ShaderType shader_type) {
+    switch (shader_type) {
+    case ShaderType::Vertex:
+        return GL_VERTEX_PROGRAM_NV;
+    case ShaderType::TesselationControl:
+        return GL_TESS_CONTROL_PROGRAM_NV;
+    case ShaderType::TesselationEval:
+        return GL_TESS_EVALUATION_PROGRAM_NV;
+    case ShaderType::Geometry:
+        return GL_GEOMETRY_PROGRAM_NV;
+    case ShaderType::Fragment:
+        return GL_FRAGMENT_PROGRAM_NV;
+    case ShaderType::Compute:
+        return GL_COMPUTE_PROGRAM_NV;
+    }
+    return {};
+}
+
 std::string MakeShaderID(u64 unique_identifier, ShaderType shader_type) {
     return fmt::format("{}{:016X}", GetShaderTypeName(shader_type), unique_identifier);
 }
@@ -120,18 +138,43 @@ std::shared_ptr<Registry> MakeRegistry(const ShaderDiskCacheEntry& entry) {
     return registry;
 }
 
-std::shared_ptr<OGLProgram> BuildShader(const Device& device, ShaderType shader_type,
+ProgramSharedPtr BuildShader(const Device& device, ShaderType shader_type, u64 unique_identifier,
-                                        u64 unique_identifier, const ShaderIR& ir,
+                             const ShaderIR& ir, const Registry& registry,
-                                        const Registry& registry, bool hint_retrievable = false) {
+                             bool hint_retrievable = false) {
     const std::string shader_id = MakeShaderID(unique_identifier, shader_type);
     LOG_INFO(Render_OpenGL, "{}", shader_id);
 
+    auto program = std::make_shared<ProgramHandle>();
+
+    if (device.UseAssemblyShaders()) {
+        const std::string arb = "Not implemented";
+
+        GLuint& arb_prog = program->assembly_program.handle;
+
+// Commented out functions signal OpenGL errors but are compatible with apitrace.
+// Use them only to capture and replay on apitrace.
+#if 0
+        glGenProgramsNV(1, &arb_prog);
+        glLoadProgramNV(AssemblyEnum(shader_type), arb_prog, static_cast<GLsizei>(arb.size()),
+                        reinterpret_cast<const GLubyte*>(arb.data()));
+#else
+        glGenProgramsARB(1, &arb_prog);
+        glNamedProgramStringEXT(arb_prog, AssemblyEnum(shader_type), GL_PROGRAM_FORMAT_ASCII_ARB,
+                                static_cast<GLsizei>(arb.size()), arb.data());
+#endif
+        const auto err = reinterpret_cast<const char*>(glGetString(GL_PROGRAM_ERROR_STRING_NV));
+        if (err && *err) {
+            LOG_CRITICAL(Render_OpenGL, "{}", err);
+            LOG_INFO(Render_OpenGL, "\n{}", arb);
+        }
+    } else {
         const std::string glsl = DecompileShader(device, ir, registry, shader_type, shader_id);
         OGLShader shader;
         shader.Create(glsl.c_str(), GetGLShaderType(shader_type));
 
-    auto program = std::make_shared<OGLProgram>();
+        program->source_program.Create(true, hint_retrievable, shader.handle);
-    program->Create(true, hint_retrievable, shader.handle);
+    }
+
     return program;
 }
 
@@ -153,15 +196,22 @@ std::unordered_set<GLenum> GetSupportedFormats() {
 
 CachedShader::CachedShader(VAddr cpu_addr, std::size_t size_in_bytes,
                            std::shared_ptr<VideoCommon::Shader::Registry> registry,
-                           ShaderEntries entries, std::shared_ptr<OGLProgram> program)
+                           ShaderEntries entries, ProgramSharedPtr program_)
     : RasterizerCacheObject{cpu_addr}, registry{std::move(registry)}, entries{std::move(entries)},
-      size_in_bytes{size_in_bytes}, program{std::move(program)} {}
+      size_in_bytes{size_in_bytes}, program{std::move(program_)} {
+    // Assign either the assembly program or source program. We can't have both.
+    handle = program->assembly_program.handle;
+    if (handle == 0) {
+        handle = program->source_program.handle;
+    }
+    ASSERT(handle != 0);
+}
 
 CachedShader::~CachedShader() = default;
 
 GLuint CachedShader::GetHandle() const {
     DEBUG_ASSERT(registry->IsConsistent());
-    return program->handle;
+    return handle;
 }
 
 Shader CachedShader::CreateStageFromMemory(const ShaderParameters& params,
@@ -239,7 +289,11 @@ void ShaderCacheOpenGL::LoadDiskCache(const std::atomic_bool& stop_loading,
         return;
     }
 
-    const std::vector gl_cache = disk_cache.LoadPrecompiled();
+    std::vector<ShaderDiskCachePrecompiled> gl_cache;
+    if (!device.UseAssemblyShaders()) {
+        // Only load precompiled cache when we are not using assembly shaders
+        gl_cache = disk_cache.LoadPrecompiled();
+    }
     const auto supported_formats = GetSupportedFormats();
 
     // Track if precompiled cache was altered during loading to know if we have to
@@ -278,7 +332,7 @@ void ShaderCacheOpenGL::LoadDiskCache(const std::atomic_bool& stop_loading,
             auto registry = MakeRegistry(entry);
             const ShaderIR ir(entry.code, main_offset, COMPILER_SETTINGS, *registry);
 
-            std::shared_ptr<OGLProgram> program;
+            ProgramSharedPtr program;
             if (precompiled_entry) {
                 // If the shader is precompiled, attempt to load it with
                 program = GeneratePrecompiledProgram(entry, *precompiled_entry, supported_formats);
@@ -332,6 +386,11 @@ void ShaderCacheOpenGL::LoadDiskCache(const std::atomic_bool& stop_loading,
         return;
     }
 
+    if (device.UseAssemblyShaders()) {
+        // Don't store precompiled binaries for assembly shaders.
+        return;
+    }
+
     // TODO(Rodrigo): Do state tracking for transferable shaders and do a dummy draw
     // before precompiling them
 
@@ -339,7 +398,7 @@ void ShaderCacheOpenGL::LoadDiskCache(const std::atomic_bool& stop_loading,
         const u64 id = (*transferable)[i].unique_identifier;
         const auto it = find_precompiled(id);
         if (it == gl_cache.end()) {
-            const GLuint program = runtime_cache.at(id).program->handle;
+            const GLuint program = runtime_cache.at(id).program->source_program.handle;
             disk_cache.SavePrecompiled(id, program);
             precompiled_cache_altered = true;
         }
@@ -350,7 +409,7 @@ void ShaderCacheOpenGL::LoadDiskCache(const std::atomic_bool& stop_loading,
     }
 }
 
-std::shared_ptr<OGLProgram> ShaderCacheOpenGL::GeneratePrecompiledProgram(
+ProgramSharedPtr ShaderCacheOpenGL::GeneratePrecompiledProgram(
     const ShaderDiskCacheEntry& entry, const ShaderDiskCachePrecompiled& precompiled_entry,
     const std::unordered_set<GLenum>& supported_formats) {
     if (supported_formats.find(precompiled_entry.binary_format) == supported_formats.end()) {
@@ -358,15 +417,15 @@ std::shared_ptr<OGLProgram> ShaderCacheOpenGL::GeneratePrecompiledProgram(
         return {};
     }
 
-    auto program = std::make_shared<OGLProgram>();
+    auto program = std::make_shared<ProgramHandle>();
-    program->handle = glCreateProgram();
+    GLuint& handle = program->source_program.handle;
-    glProgramParameteri(program->handle, GL_PROGRAM_SEPARABLE, GL_TRUE);
+    handle = glCreateProgram();
-    glProgramBinary(program->handle, precompiled_entry.binary_format,
+    glProgramParameteri(handle, GL_PROGRAM_SEPARABLE, GL_TRUE);
-                    precompiled_entry.binary.data(),
+    glProgramBinary(handle, precompiled_entry.binary_format, precompiled_entry.binary.data(),
                     static_cast<GLsizei>(precompiled_entry.binary.size()));
 
     GLint link_status;
-    glGetProgramiv(program->handle, GL_LINK_STATUS, &link_status);
+    glGetProgramiv(handle, GL_LINK_STATUS, &link_status);
     if (link_status == GL_FALSE) {
         LOG_INFO(Render_OpenGL, "Precompiled cache rejected by the driver, removing");
         return {};

src/video_core/renderer_opengl/gl_shader_cache.h

@@ -43,8 +43,14 @@ struct UnspecializedShader;
 using Shader = std::shared_ptr<CachedShader>;
 using Maxwell = Tegra::Engines::Maxwell3D::Regs;
 
+struct ProgramHandle {
+    OGLProgram source_program;
+    OGLAssemblyProgram assembly_program;
+};
+using ProgramSharedPtr = std::shared_ptr<ProgramHandle>;
+
 struct PrecompiledShader {
-    std::shared_ptr<OGLProgram> program;
+    ProgramSharedPtr program;
     std::shared_ptr<VideoCommon::Shader::Registry> registry;
     ShaderEntries entries;
 };
@@ -87,12 +93,13 @@ public:
 private:
     explicit CachedShader(VAddr cpu_addr, std::size_t size_in_bytes,
                           std::shared_ptr<VideoCommon::Shader::Registry> registry,
-                          ShaderEntries entries, std::shared_ptr<OGLProgram> program);
+                          ShaderEntries entries, ProgramSharedPtr program);
 
     std::shared_ptr<VideoCommon::Shader::Registry> registry;
     ShaderEntries entries;
     std::size_t size_in_bytes = 0;
-    std::shared_ptr<OGLProgram> program;
+    ProgramSharedPtr program;
+    GLuint handle = 0;
 };
 
 class ShaderCacheOpenGL final : public RasterizerCache<Shader> {
@@ -115,7 +122,7 @@ protected:
     void FlushObjectInner(const Shader& object) override {}
 
 private:
-    std::shared_ptr<OGLProgram> GeneratePrecompiledProgram(
+    ProgramSharedPtr GeneratePrecompiledProgram(
         const ShaderDiskCacheEntry& entry, const ShaderDiskCachePrecompiled& precompiled_entry,
         const std::unordered_set<GLenum>& supported_formats);
 

src/video_core/renderer_opengl/gl_shader_manager.cpp

@@ -6,47 +6,107 @@
 
 #include "common/common_types.h"
 #include "video_core/engines/maxwell_3d.h"
+#include "video_core/renderer_opengl/gl_device.h"
 #include "video_core/renderer_opengl/gl_shader_manager.h"
 
-namespace OpenGL::GLShader {
+namespace OpenGL {
 
-ProgramManager::ProgramManager() = default;
+ProgramManager::ProgramManager(const Device& device) {
+    use_assembly_programs = device.UseAssemblyShaders();
+    if (use_assembly_programs) {
+        glEnable(GL_COMPUTE_PROGRAM_NV);
+    } else {
+        graphics_pipeline.Create();
+        glBindProgramPipeline(graphics_pipeline.handle);
+    }
+}
 
 ProgramManager::~ProgramManager() = default;
 
-void ProgramManager::Create() {
+void ProgramManager::BindCompute(GLuint program) {
-    graphics_pipeline.Create();
+    if (use_assembly_programs) {
-    glBindProgramPipeline(graphics_pipeline.handle);
+        glBindProgramARB(GL_COMPUTE_PROGRAM_NV, program);
+    } else {
+        is_graphics_bound = false;
+        glUseProgram(program);
+    }
 }
 
 void ProgramManager::BindGraphicsPipeline() {
+    if (use_assembly_programs) {
+        UpdateAssemblyPrograms();
+    } else {
+        UpdateSourcePrograms();
+    }
+}
+
+void ProgramManager::BindHostPipeline(GLuint pipeline) {
+    if (use_assembly_programs) {
+        if (geometry_enabled) {
+            geometry_enabled = false;
+            old_state.geometry = 0;
+            glDisable(GL_GEOMETRY_PROGRAM_NV);
+        }
+    }
+    glBindProgramPipeline(pipeline);
+}
+
+void ProgramManager::RestoreGuestPipeline() {
+    if (use_assembly_programs) {
+        glBindProgramPipeline(0);
+    } else {
+        glBindProgramPipeline(graphics_pipeline.handle);
+    }
+}
+
+void ProgramManager::UpdateAssemblyPrograms() {
+    const auto update_state = [](GLenum stage, bool& enabled, GLuint current, GLuint old) {
+        if (current == old) {
+            return;
+        }
+        if (current == 0) {
+            if (enabled) {
+                enabled = false;
+                glDisable(stage);
+            }
+            return;
+        }
+        if (!enabled) {
+            enabled = true;
+            glEnable(stage);
+        }
+        glBindProgramARB(stage, current);
+    };
+
+    update_state(GL_VERTEX_PROGRAM_NV, vertex_enabled, current_state.vertex, old_state.vertex);
+    update_state(GL_GEOMETRY_PROGRAM_NV, geometry_enabled, current_state.geometry,
+                 old_state.geometry);
+    update_state(GL_FRAGMENT_PROGRAM_NV, fragment_enabled, current_state.fragment,
+                 old_state.fragment);
+
+    old_state = current_state;
+}
+
+void ProgramManager::UpdateSourcePrograms() {
     if (!is_graphics_bound) {
         is_graphics_bound = true;
         glUseProgram(0);
     }
 
-    // Avoid updating the pipeline when values have no changed
+    const GLuint handle = graphics_pipeline.handle;
-    if (old_state == current_state) {
+    const auto update_state = [handle](GLenum stage, GLuint current, GLuint old) {
+        if (current == old) {
             return;
         }
-
+        glUseProgramStages(handle, stage, current);
-    // Workaround for AMD bug
+    };
-    static constexpr GLenum all_used_stages{GL_VERTEX_SHADER_BIT | GL_GEOMETRY_SHADER_BIT |
+    update_state(GL_VERTEX_SHADER_BIT, current_state.vertex, old_state.vertex);
-                                            GL_FRAGMENT_SHADER_BIT};
+    update_state(GL_GEOMETRY_SHADER_BIT, current_state.geometry, old_state.geometry);
-    const GLuint handle = graphics_pipeline.handle;
+    update_state(GL_FRAGMENT_SHADER_BIT, current_state.fragment, old_state.fragment);
-    glUseProgramStages(handle, all_used_stages, 0);
-    glUseProgramStages(handle, GL_VERTEX_SHADER_BIT, current_state.vertex_shader);
-    glUseProgramStages(handle, GL_GEOMETRY_SHADER_BIT, current_state.geometry_shader);
-    glUseProgramStages(handle, GL_FRAGMENT_SHADER_BIT, current_state.fragment_shader);
 
     old_state = current_state;
 }
 
-void ProgramManager::BindComputeShader(GLuint program) {
-    is_graphics_bound = false;
-    glUseProgram(program);
-}
-
 void MaxwellUniformData::SetFromRegs(const Tegra::Engines::Maxwell3D& maxwell) {
     const auto& regs = maxwell.regs;
 
@@ -54,4 +114,4 @@ void MaxwellUniformData::SetFromRegs(const Tegra::Engines::Maxwell3D& maxwell) {
     y_direction = regs.screen_y_control.y_negate == 0 ? 1.0f : -1.0f;
 }
 
-} // namespace OpenGL::GLShader
+} // namespace OpenGL

src/video_core/renderer_opengl/gl_shader_manager.h

@@ -11,7 +11,9 @@
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 #include "video_core/renderer_opengl/maxwell_to_gl.h"
 
-namespace OpenGL::GLShader {
+namespace OpenGL {
+
+class Device;
 
 /// Uniform structure for the Uniform Buffer Object, all vectors must be 16-byte aligned
 /// @note Always keep a vec4 at the end. The GL spec is not clear whether the alignment at
@@ -28,50 +30,58 @@ static_assert(sizeof(MaxwellUniformData) < 16384,
 
 class ProgramManager {
 public:
-    explicit ProgramManager();
+    explicit ProgramManager(const Device& device);
     ~ProgramManager();
 
-    void Create();
+    /// Binds a compute program
+    void BindCompute(GLuint program);
 
-    /// Updates the graphics pipeline and binds it.
+    /// Updates bound programs.
     void BindGraphicsPipeline();
 
-    /// Binds a compute shader.
+    /// Binds an OpenGL pipeline object unsynchronized with the guest state.
-    void BindComputeShader(GLuint program);
+    void BindHostPipeline(GLuint pipeline);
+
+    /// Rewinds BindHostPipeline state changes.
+    void RestoreGuestPipeline();
 
     void UseVertexShader(GLuint program) {
-        current_state.vertex_shader = program;
+        current_state.vertex = program;
     }
 
     void UseGeometryShader(GLuint program) {
-        current_state.geometry_shader = program;
+        current_state.geometry = program;
     }
 
     void UseFragmentShader(GLuint program) {
-        current_state.fragment_shader = program;
+        current_state.fragment = program;
     }
 
 private:
     struct PipelineState {
-        bool operator==(const PipelineState& rhs) const noexcept {
+        GLuint vertex = 0;
-            return vertex_shader == rhs.vertex_shader && fragment_shader == rhs.fragment_shader &&
+        GLuint geometry = 0;
-                   geometry_shader == rhs.geometry_shader;
+        GLuint fragment = 0;
-        }
-
-        bool operator!=(const PipelineState& rhs) const noexcept {
-            return !operator==(rhs);
-        }
-
-        GLuint vertex_shader = 0;
-        GLuint fragment_shader = 0;
-        GLuint geometry_shader = 0;
     };
 
+    /// Update NV_gpu_program5 programs.
+    void UpdateAssemblyPrograms();
+
+    /// Update GLSL programs.
+    void UpdateSourcePrograms();
+
     OGLPipeline graphics_pipeline;
-    OGLPipeline compute_pipeline;
+
     PipelineState current_state;
     PipelineState old_state;
+
+    bool use_assembly_programs = false;
+
     bool is_graphics_bound = true;
+
+    bool vertex_enabled = false;
+    bool geometry_enabled = false;
+    bool fragment_enabled = false;
 };
 
-} // namespace OpenGL::GLShader
+} // namespace OpenGL

src/video_core/renderer_opengl/renderer_opengl.cpp

@@ -316,7 +316,7 @@ public:
 RendererOpenGL::RendererOpenGL(Core::Frontend::EmuWindow& emu_window, Core::System& system,
                                Core::Frontend::GraphicsContext& context)
     : RendererBase{emu_window}, emu_window{emu_window}, system{system}, context{context},
-      has_debug_tool{HasDebugTool()} {}
+      program_manager{device}, has_debug_tool{HasDebugTool()} {}
 
 RendererOpenGL::~RendererOpenGL() = default;
 
@@ -468,8 +468,9 @@ void RendererOpenGL::InitOpenGLObjects() {
     vertex_program.Create(true, false, vertex_shader.handle);
     fragment_program.Create(true, false, fragment_shader.handle);
 
-    // Create program pipeline
+    pipeline.Create();
-    program_manager.Create();
+    glUseProgramStages(pipeline.handle, GL_VERTEX_SHADER_BIT, vertex_program.handle);
+    glUseProgramStages(pipeline.handle, GL_FRAGMENT_SHADER_BIT, fragment_program.handle);
 
     // Generate VBO handle for drawing
     vertex_buffer.Create();
@@ -508,7 +509,7 @@ void RendererOpenGL::CreateRasterizer() {
     if (rasterizer) {
         return;
     }
-    rasterizer = std::make_unique<RasterizerOpenGL>(system, emu_window, screen_info,
+    rasterizer = std::make_unique<RasterizerOpenGL>(system, emu_window, device, screen_info,
                                                     program_manager, state_tracker);
 }
 
@@ -620,10 +621,7 @@ void RendererOpenGL::DrawScreen(const Layout::FramebufferLayout& layout) {
     state_tracker.NotifyClipControl();
     state_tracker.NotifyAlphaTest();
 
-    program_manager.UseVertexShader(vertex_program.handle);
+    program_manager.BindHostPipeline(pipeline.handle);
-    program_manager.UseGeometryShader(0);
-    program_manager.UseFragmentShader(fragment_program.handle);
-    program_manager.BindGraphicsPipeline();
 
     glEnable(GL_CULL_FACE);
     if (screen_info.display_srgb) {
@@ -665,6 +663,8 @@ void RendererOpenGL::DrawScreen(const Layout::FramebufferLayout& layout) {
 
     glClear(GL_COLOR_BUFFER_BIT);
     glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+
+    program_manager.RestoreGuestPipeline();
 }
 
 bool RendererOpenGL::TryPresent(int timeout_ms) {

src/video_core/renderer_opengl/renderer_opengl.h

@@ -9,6 +9,7 @@
 #include "common/common_types.h"
 #include "common/math_util.h"
 #include "video_core/renderer_base.h"
+#include "video_core/renderer_opengl/gl_device.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 #include "video_core/renderer_opengl/gl_shader_manager.h"
 #include "video_core/renderer_opengl/gl_state_tracker.h"
@@ -95,6 +96,7 @@ private:
     Core::Frontend::EmuWindow& emu_window;
     Core::System& system;
     Core::Frontend::GraphicsContext& context;
+    const Device device;
 
     StateTracker state_tracker{system};
 
@@ -102,13 +104,14 @@ private:
     OGLBuffer vertex_buffer;
     OGLProgram vertex_program;
     OGLProgram fragment_program;
+    OGLPipeline pipeline;
     OGLFramebuffer screenshot_framebuffer;
 
     /// Display information for Switch screen
     ScreenInfo screen_info;
 
     /// Global dummy shader pipeline
-    GLShader::ProgramManager program_manager;
+    ProgramManager program_manager;
 
     /// OpenGL framebuffer data
     std::vector<u8> gl_framebuffer_data;