Merge pull request #1987 from ReinUsesLisp/explicit-shader-ldg

gl_shader_cache: Use explicit bindings

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -297,11 +297,7 @@ void RasterizerOpenGL::SetupShaders(GLenum primitive_mode) {
     MICROPROFILE_SCOPE(OpenGL_Shader);
     auto& gpu = Core::System::GetInstance().GPU().Maxwell3D();
 
-    // Next available bindpoints to use when uploading the const buffers and textures to the GLSL
+    BaseBindings base_bindings;
-    // shaders. The constbuffer bindpoint starts after the shader stage configuration bind points.
-    u32 current_constbuffer_bindpoint = Tegra::Engines::Maxwell3D::Regs::MaxShaderStage;
-    u32 current_gmem_bindpoint = 0;
-    u32 current_texture_bindpoint = 0;
     std::array<bool, Maxwell::NumClipDistances> clip_distances{};
 
     for (std::size_t index = 0; index < Maxwell::MaxShaderProgram; ++index) {
@@ -325,47 +321,35 @@ void RasterizerOpenGL::SetupShaders(GLenum primitive_mode) {
         const GLintptr offset = buffer_cache.UploadHostMemory(
             &ubo, sizeof(ubo), static_cast<std::size_t>(uniform_buffer_alignment));
 
-        // Bind the buffer
+        // Bind the emulation info buffer
-        glBindBufferRange(GL_UNIFORM_BUFFER, static_cast<GLuint>(stage), buffer_cache.GetHandle(),
+        glBindBufferRange(GL_UNIFORM_BUFFER, base_bindings.cbuf, buffer_cache.GetHandle(), offset,
-                          offset, static_cast<GLsizeiptr>(sizeof(ubo)));
+                          static_cast<GLsizeiptr>(sizeof(ubo)));
 
         Shader shader{shader_cache.GetStageProgram(program)};
+        const auto [program_handle, next_bindings] =
+            shader->GetProgramHandle(primitive_mode, base_bindings);
 
         switch (program) {
         case Maxwell::ShaderProgram::VertexA:
-        case Maxwell::ShaderProgram::VertexB: {
+        case Maxwell::ShaderProgram::VertexB:
-            shader_program_manager->UseProgrammableVertexShader(
+            shader_program_manager->UseProgrammableVertexShader(program_handle);
-                shader->GetProgramHandle(primitive_mode));
             break;
-        }
+        case Maxwell::ShaderProgram::Geometry:
-        case Maxwell::ShaderProgram::Geometry: {
+            shader_program_manager->UseProgrammableGeometryShader(program_handle);
-            shader_program_manager->UseProgrammableGeometryShader(
-                shader->GetProgramHandle(primitive_mode));
             break;
-        }
+        case Maxwell::ShaderProgram::Fragment:
-        case Maxwell::ShaderProgram::Fragment: {
+            shader_program_manager->UseProgrammableFragmentShader(program_handle);
-            shader_program_manager->UseProgrammableFragmentShader(
-                shader->GetProgramHandle(primitive_mode));
             break;
-        }
         default:
             LOG_CRITICAL(HW_GPU, "Unimplemented shader index={}, enable={}, offset=0x{:08X}", index,
                          shader_config.enable.Value(), shader_config.offset);
             UNREACHABLE();
         }
 
-        // Configure the const buffers for this shader stage.
+        const auto stage_enum = static_cast<Maxwell::ShaderStage>(stage);
-        current_constbuffer_bindpoint =
+        SetupConstBuffers(stage_enum, shader, program_handle, base_bindings);
-            SetupConstBuffers(static_cast<Maxwell::ShaderStage>(stage), shader, primitive_mode,
+        SetupGlobalRegions(stage_enum, shader, program_handle, base_bindings);
-                              current_constbuffer_bindpoint);
+        SetupTextures(stage_enum, shader, program_handle, base_bindings);
-
-        // Configure global memory regions for this shader stage.
-        current_gmem_bindpoint = SetupGlobalRegions(static_cast<Maxwell::ShaderStage>(stage),
-                                                    shader, primitive_mode, current_gmem_bindpoint);
-
-        // Configure the textures for this shader stage.
-        current_texture_bindpoint = SetupTextures(static_cast<Maxwell::ShaderStage>(stage), shader,
-                                                  primitive_mode, current_texture_bindpoint);
 
         // Workaround for Intel drivers.
         // When a clip distance is enabled but not set in the shader it crops parts of the screen
@@ -380,6 +364,8 @@ void RasterizerOpenGL::SetupShaders(GLenum primitive_mode) {
             // VertexB was combined with VertexA, so we skip the VertexB iteration
             index++;
         }
+
+        base_bindings = next_bindings;
     }
 
     SyncClipEnabled(clip_distances);
@@ -929,8 +915,9 @@ void RasterizerOpenGL::SamplerInfo::SyncWithConfig(const Tegra::Texture::TSCEntr
     }
 }
 
-u32 RasterizerOpenGL::SetupConstBuffers(Maxwell::ShaderStage stage, Shader& shader,
+void RasterizerOpenGL::SetupConstBuffers(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage,
-                                        GLenum primitive_mode, u32 current_bindpoint) {
+                                         const Shader& shader, GLuint program_handle,
+                                         BaseBindings base_bindings) {
     MICROPROFILE_SCOPE(OpenGL_UBO);
     const auto& gpu = Core::System::GetInstance().GPU();
     const auto& maxwell3d = gpu.Maxwell3D();
@@ -978,92 +965,73 @@ u32 RasterizerOpenGL::SetupConstBuffers(Maxwell::ShaderStage stage, Shader& shad
         size = Common::AlignUp(size, sizeof(GLvec4));
         ASSERT_MSG(size <= MaxConstbufferSize, "Constbuffer too big");
 
-        GLintptr const_buffer_offset = buffer_cache.UploadMemory(
+        const GLintptr const_buffer_offset = buffer_cache.UploadMemory(
             buffer.address, size, static_cast<std::size_t>(uniform_buffer_alignment));
 
-        // Now configure the bindpoint of the buffer inside the shader
-        glUniformBlockBinding(shader->GetProgramHandle(primitive_mode),
-                              shader->GetProgramResourceIndex(used_buffer),
-                              current_bindpoint + bindpoint);
-
         // Prepare values for multibind
         bind_buffers[bindpoint] = buffer_cache.GetHandle();
         bind_offsets[bindpoint] = const_buffer_offset;
         bind_sizes[bindpoint] = size;
     }
 
-    glBindBuffersRange(GL_UNIFORM_BUFFER, current_bindpoint, static_cast<GLsizei>(entries.size()),
+    // The first binding is reserved for emulation values
+    const GLuint ubo_base_binding = base_bindings.cbuf + 1;
+    glBindBuffersRange(GL_UNIFORM_BUFFER, ubo_base_binding, static_cast<GLsizei>(entries.size()),
                        bind_buffers.data(), bind_offsets.data(), bind_sizes.data());
-
-    return current_bindpoint + static_cast<u32>(entries.size());
 }
 
-u32 RasterizerOpenGL::SetupGlobalRegions(Maxwell::ShaderStage stage, Shader& shader,
+void RasterizerOpenGL::SetupGlobalRegions(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage,
-                                         GLenum primitive_mode, u32 current_bindpoint) {
+                                          const Shader& shader, GLenum primitive_mode,
-    for (const auto& global_region : shader->GetShaderEntries().global_memory_entries) {
+                                          BaseBindings base_bindings) {
-        const auto& region =
+    // TODO(Rodrigo): Use ARB_multi_bind here
-            global_cache.GetGlobalRegion(global_region, static_cast<Maxwell::ShaderStage>(stage));
+    const auto& entries = shader->GetShaderEntries().global_memory_entries;
-        const GLuint block_index{shader->GetProgramResourceIndex(global_region)};
+
-        ASSERT(block_index != GL_INVALID_INDEX);
+    for (u32 bindpoint = 0; bindpoint < static_cast<u32>(entries.size()); ++bindpoint) {
+        const auto& entry = entries[bindpoint];
+        const u32 current_bindpoint = base_bindings.gmem + bindpoint;
+        const auto& region = global_cache.GetGlobalRegion(entry, stage);
 
         glBindBufferBase(GL_SHADER_STORAGE_BUFFER, current_bindpoint, region->GetBufferHandle());
-        glShaderStorageBlockBinding(shader->GetProgramHandle(primitive_mode), block_index,
-                                    current_bindpoint);
-        ++current_bindpoint;
     }
-
-    return current_bindpoint;
 }
 
-u32 RasterizerOpenGL::SetupTextures(Maxwell::ShaderStage stage, Shader& shader,
+void RasterizerOpenGL::SetupTextures(Maxwell::ShaderStage stage, const Shader& shader,
-                                    GLenum primitive_mode, u32 current_unit) {
+                                     GLuint program_handle, BaseBindings base_bindings) {
     MICROPROFILE_SCOPE(OpenGL_Texture);
     const auto& gpu = Core::System::GetInstance().GPU();
     const auto& maxwell3d = gpu.Maxwell3D();
     const auto& entries = shader->GetShaderEntries().samplers;
 
-    ASSERT_MSG(current_unit + entries.size() <= std::size(state.texture_units),
+    ASSERT_MSG(base_bindings.sampler + entries.size() <= std::size(state.texture_units),
                "Exceeded the number of active textures.");
 
     for (u32 bindpoint = 0; bindpoint < entries.size(); ++bindpoint) {
         const auto& entry = entries[bindpoint];
-        const u32 current_bindpoint = current_unit + bindpoint;
+        const u32 current_bindpoint = base_bindings.sampler + bindpoint;
-
+        auto& unit = state.texture_units[current_bindpoint];
-        // Bind the uniform to the sampler.
-
-        glProgramUniform1i(shader->GetProgramHandle(primitive_mode),
-                           shader->GetUniformLocation(entry), current_bindpoint);
 
         const auto texture = maxwell3d.GetStageTexture(entry.GetStage(), entry.GetOffset());
-
         if (!texture.enabled) {
-            state.texture_units[current_bindpoint].texture = 0;
+            unit.texture = 0;
             continue;
         }
 
         texture_samplers[current_bindpoint].SyncWithConfig(texture.tsc);
+
         Surface surface = res_cache.GetTextureSurface(texture, entry);
         if (surface != nullptr) {
-            const GLuint handle =
+            unit.texture =
                 entry.IsArray() ? surface->TextureLayer().handle : surface->Texture().handle;
-            const GLenum target = entry.IsArray() ? surface->TargetLayer() : surface->Target();
+            unit.target = entry.IsArray() ? surface->TargetLayer() : surface->Target();
-            state.texture_units[current_bindpoint].texture = handle;
+            unit.swizzle.r = MaxwellToGL::SwizzleSource(texture.tic.x_source);
-            state.texture_units[current_bindpoint].target = target;
+            unit.swizzle.g = MaxwellToGL::SwizzleSource(texture.tic.y_source);
-            state.texture_units[current_bindpoint].swizzle.r =
+            unit.swizzle.b = MaxwellToGL::SwizzleSource(texture.tic.z_source);
-                MaxwellToGL::SwizzleSource(texture.tic.x_source);
+            unit.swizzle.a = MaxwellToGL::SwizzleSource(texture.tic.w_source);
-            state.texture_units[current_bindpoint].swizzle.g =
-                MaxwellToGL::SwizzleSource(texture.tic.y_source);
-            state.texture_units[current_bindpoint].swizzle.b =
-                MaxwellToGL::SwizzleSource(texture.tic.z_source);
-            state.texture_units[current_bindpoint].swizzle.a =
-                MaxwellToGL::SwizzleSource(texture.tic.w_source);
         } else {
             // Can occur when texture addr is null or its memory is unmapped/invalid
-            state.texture_units[current_bindpoint].texture = 0;
+            unit.texture = 0;
         }
     }
-
-    return current_unit + static_cast<u32>(entries.size());
 }
 
 void RasterizerOpenGL::SyncViewport(OpenGLState& current_state) {

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -127,35 +127,18 @@ private:
                                bool using_depth_fb = true, bool preserve_contents = true,
                                std::optional<std::size_t> single_color_target = {});
 
-    /**
+    /// Configures the current constbuffers to use for the draw command.
-     * Configures the current constbuffers to use for the draw command.
+    void SetupConstBuffers(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage, const Shader& shader,
-     * @param stage The shader stage to configure buffers for.
+                           GLuint program_handle, BaseBindings base_bindings);
-     * @param shader The shader object that contains the specified stage.
-     * @param current_bindpoint The offset at which to start counting new buffer bindpoints.
-     * @returns The next available bindpoint for use in the next shader stage.
-     */
-    u32 SetupConstBuffers(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage, Shader& shader,
-                          GLenum primitive_mode, u32 current_bindpoint);
 
-    /**
+    /// Configures the current global memory entries to use for the draw command.
-     * Configures the current global memory regions to use for the draw command.
+    void SetupGlobalRegions(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage,
-     * @param stage The shader stage to configure buffers for.
+                            const Shader& shader, GLenum primitive_mode,
-     * @param shader The shader object that contains the specified stage.
+                            BaseBindings base_bindings);
-     * @param current_bindpoint The offset at which to start counting new buffer bindpoints.
-     * @returns The next available bindpoint for use in the next shader stage.
-     */
-    u32 SetupGlobalRegions(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage, Shader& shader,
-                           GLenum primitive_mode, u32 current_bindpoint);
 
-    /**
+    /// Configures the current textures to use for the draw command.
-     * Configures the current textures to use for the draw command.
+    void SetupTextures(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage, const Shader& shader,
-     * @param stage The shader stage to configure textures for.
+                       GLuint program_handle, BaseBindings base_bindings);
-     * @param shader The shader object that contains the specified stage.
-     * @param current_unit The offset at which to start counting unused texture units.
-     * @returns The next available bindpoint for use in the next shader stage.
-     */
-    u32 SetupTextures(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage, Shader& shader,
-                      GLenum primitive_mode, u32 current_unit);
 
     /// Syncs the viewport and depth range to match the guest state
     void SyncViewport(OpenGLState& current_state);

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -34,36 +34,25 @@ static ProgramCode GetShaderCode(VAddr addr) {
     return program_code;
 }
 
-/// Helper function to set shader uniform block bindings for a single shader stage
+/// Gets the shader type from a Maxwell program type
-static void SetShaderUniformBlockBinding(GLuint shader, const char* name,
+constexpr GLenum GetShaderType(Maxwell::ShaderProgram program_type) {
-                                         Maxwell::ShaderStage binding, std::size_t expected_size) {
+    switch (program_type) {
-    const GLuint ub_index = glGetUniformBlockIndex(shader, name);
+    case Maxwell::ShaderProgram::VertexA:
-    if (ub_index == GL_INVALID_INDEX) {
+    case Maxwell::ShaderProgram::VertexB:
-        return;
+        return GL_VERTEX_SHADER;
+    case Maxwell::ShaderProgram::Geometry:
+        return GL_GEOMETRY_SHADER;
+    case Maxwell::ShaderProgram::Fragment:
+        return GL_FRAGMENT_SHADER;
+    default:
+        return GL_NONE;
     }
-
-    GLint ub_size = 0;
-    glGetActiveUniformBlockiv(shader, ub_index, GL_UNIFORM_BLOCK_DATA_SIZE, &ub_size);
-    ASSERT_MSG(static_cast<std::size_t>(ub_size) == expected_size,
-               "Uniform block size did not match! Got {}, expected {}", ub_size, expected_size);
-    glUniformBlockBinding(shader, ub_index, static_cast<GLuint>(binding));
-}
-
-/// Sets shader uniform block bindings for an entire shader program
-static void SetShaderUniformBlockBindings(GLuint shader) {
-    SetShaderUniformBlockBinding(shader, "vs_config", Maxwell::ShaderStage::Vertex,
-                                 sizeof(GLShader::MaxwellUniformData));
-    SetShaderUniformBlockBinding(shader, "gs_config", Maxwell::ShaderStage::Geometry,
-                                 sizeof(GLShader::MaxwellUniformData));
-    SetShaderUniformBlockBinding(shader, "fs_config", Maxwell::ShaderStage::Fragment,
-                                 sizeof(GLShader::MaxwellUniformData));
 }
 
 CachedShader::CachedShader(VAddr addr, Maxwell::ShaderProgram program_type)
     : addr{addr}, program_type{program_type}, setup{GetShaderCode(addr)} {
 
     GLShader::ProgramResult program_result;
-    GLenum gl_type{};
 
     switch (program_type) {
     case Maxwell::ShaderProgram::VertexA:
@@ -74,17 +63,14 @@ CachedShader::CachedShader(VAddr addr, Maxwell::ShaderProgram program_type)
     case Maxwell::ShaderProgram::VertexB:
         CalculateProperties();
         program_result = GLShader::GenerateVertexShader(setup);
-        gl_type = GL_VERTEX_SHADER;
         break;
     case Maxwell::ShaderProgram::Geometry:
         CalculateProperties();
         program_result = GLShader::GenerateGeometryShader(setup);
-        gl_type = GL_GEOMETRY_SHADER;
         break;
     case Maxwell::ShaderProgram::Fragment:
         CalculateProperties();
         program_result = GLShader::GenerateFragmentShader(setup);
-        gl_type = GL_FRAGMENT_SHADER;
         break;
     default:
         LOG_CRITICAL(HW_GPU, "Unimplemented program_type={}", static_cast<u32>(program_type));
@@ -92,71 +78,105 @@ CachedShader::CachedShader(VAddr addr, Maxwell::ShaderProgram program_type)
         return;
     }
 
+    code = program_result.first;
     entries = program_result.second;
     shader_length = entries.shader_length;
+}
 
-    if (program_type != Maxwell::ShaderProgram::Geometry) {
+std::tuple<GLuint, BaseBindings> CachedShader::GetProgramHandle(GLenum primitive_mode,
-        OGLShader shader;
+                                                                BaseBindings base_bindings) {
-        shader.Create(program_result.first.c_str(), gl_type);
+    GLuint handle{};
-        program.Create(true, shader.handle);
+    if (program_type == Maxwell::ShaderProgram::Geometry) {
-        SetShaderUniformBlockBindings(program.handle);
+        handle = GetGeometryShader(primitive_mode, base_bindings);
-        LabelGLObject(GL_PROGRAM, program.handle, addr);
     } else {
-        // Store shader's code to lazily build it on draw
+        const auto [entry, is_cache_miss] = programs.try_emplace(base_bindings);
-        geometry_programs.code = program_result.first;
+        auto& program = entry->second;
+        if (is_cache_miss) {
+            std::string source = AllocateBindings(base_bindings);
+            source += code;
+
+            OGLShader shader;
+            shader.Create(source.c_str(), GetShaderType(program_type));
+            program.Create(true, shader.handle);
+            LabelGLObject(GL_PROGRAM, program.handle, addr);
+        }
+
+        handle = program.handle;
+    }
+
+    // Add const buffer and samplers offset reserved by this shader. One UBO binding is reserved for
+    // emulation values
+    base_bindings.cbuf += static_cast<u32>(entries.const_buffers.size()) + 1;
+    base_bindings.gmem += static_cast<u32>(entries.global_memory_entries.size());
+    base_bindings.sampler += static_cast<u32>(entries.samplers.size());
+
+    return {handle, base_bindings};
+}
+
+std::string CachedShader::AllocateBindings(BaseBindings base_bindings) {
+    std::string code = "#version 430 core\n";
+    code += fmt::format("#define EMULATION_UBO_BINDING {}\n", base_bindings.cbuf++);
+
+    for (const auto& cbuf : entries.const_buffers) {
+        code += fmt::format("#define CBUF_BINDING_{} {}\n", cbuf.GetIndex(), base_bindings.cbuf++);
+    }
+
+    for (const auto& gmem : entries.global_memory_entries) {
+        code += fmt::format("#define GMEM_BINDING_{}_{} {}\n", gmem.GetCbufIndex(),
+                            gmem.GetCbufOffset(), base_bindings.gmem++);
+    }
+
+    for (const auto& sampler : entries.samplers) {
+        code += fmt::format("#define SAMPLER_BINDING_{} {}\n", sampler.GetIndex(),
+                            base_bindings.sampler++);
+    }
+
+    return code;
+}
+
+GLuint CachedShader::GetGeometryShader(GLenum primitive_mode, BaseBindings base_bindings) {
+    const auto [entry, is_cache_miss] = geometry_programs.try_emplace(base_bindings);
+    auto& programs = entry->second;
+
+    switch (primitive_mode) {
+    case GL_POINTS:
+        return LazyGeometryProgram(programs.points, base_bindings, "points", 1, "ShaderPoints");
+    case GL_LINES:
+    case GL_LINE_STRIP:
+        return LazyGeometryProgram(programs.lines, base_bindings, "lines", 2, "ShaderLines");
+    case GL_LINES_ADJACENCY:
+    case GL_LINE_STRIP_ADJACENCY:
+        return LazyGeometryProgram(programs.lines_adjacency, base_bindings, "lines_adjacency", 4,
+                                   "ShaderLinesAdjacency");
+    case GL_TRIANGLES:
+    case GL_TRIANGLE_STRIP:
+    case GL_TRIANGLE_FAN:
+        return LazyGeometryProgram(programs.triangles, base_bindings, "triangles", 3,
+                                   "ShaderTriangles");
+    case GL_TRIANGLES_ADJACENCY:
+    case GL_TRIANGLE_STRIP_ADJACENCY:
+        return LazyGeometryProgram(programs.triangles_adjacency, base_bindings,
+                                   "triangles_adjacency", 6, "ShaderTrianglesAdjacency");
+    default:
+        UNREACHABLE_MSG("Unknown primitive mode.");
+        return LazyGeometryProgram(programs.points, base_bindings, "points", 1, "ShaderPoints");
     }
 }
 
-GLuint CachedShader::GetProgramResourceIndex(const GLShader::ConstBufferEntry& buffer) {
+GLuint CachedShader::LazyGeometryProgram(OGLProgram& target_program, BaseBindings base_bindings,
-    const auto search{cbuf_resource_cache.find(buffer.GetHash())};
-    if (search == cbuf_resource_cache.end()) {
-        const GLuint index{
-            glGetProgramResourceIndex(program.handle, GL_UNIFORM_BLOCK, buffer.GetName().c_str())};
-        cbuf_resource_cache[buffer.GetHash()] = index;
-        return index;
-    }
-
-    return search->second;
-}
-
-GLuint CachedShader::GetProgramResourceIndex(const GLShader::GlobalMemoryEntry& global_mem) {
-    const auto search{gmem_resource_cache.find(global_mem.GetHash())};
-    if (search == gmem_resource_cache.end()) {
-        const GLuint index{glGetProgramResourceIndex(program.handle, GL_SHADER_STORAGE_BLOCK,
-                                                     global_mem.GetName().c_str())};
-        gmem_resource_cache[global_mem.GetHash()] = index;
-        return index;
-    }
-
-    return search->second;
-}
-
-GLint CachedShader::GetUniformLocation(const GLShader::SamplerEntry& sampler) {
-    const auto search{uniform_cache.find(sampler.GetHash())};
-    if (search == uniform_cache.end()) {
-        const GLint index{glGetUniformLocation(program.handle, sampler.GetName().c_str())};
-        uniform_cache[sampler.GetHash()] = index;
-        return index;
-    }
-
-    return search->second;
-}
-
-GLuint CachedShader::LazyGeometryProgram(OGLProgram& target_program,
                                          const std::string& glsl_topology, u32 max_vertices,
                                          const std::string& debug_name) {
     if (target_program.handle != 0) {
         return target_program.handle;
     }
-    std::string source = "#version 430 core\n";
+    std::string source = AllocateBindings(base_bindings);
     source += "layout (" + glsl_topology + ") in;\n";
     source += "#define MAX_VERTEX_INPUT " + std::to_string(max_vertices) + '\n';
-    source += geometry_programs.code;
+    source += code;
 
     OGLShader shader;
     shader.Create(source.c_str(), GL_GEOMETRY_SHADER);
     target_program.Create(true, shader.handle);
-    SetShaderUniformBlockBindings(target_program.handle);
     LabelGLObject(GL_PROGRAM, target_program.handle, addr, debug_name);
     return target_program.handle;
 };

src/video_core/renderer_opengl/gl_shader_cache.h

@@ -7,6 +7,9 @@
 #include <array>
 #include <map>
 #include <memory>
+#include <tuple>
+
+#include <glad/glad.h>
 
 #include "common/assert.h"
 #include "common/common_types.h"
@@ -23,6 +26,16 @@ class RasterizerOpenGL;
 using Shader = std::shared_ptr<CachedShader>;
 using Maxwell = Tegra::Engines::Maxwell3D::Regs;
 
+struct BaseBindings {
+    u32 cbuf{};
+    u32 gmem{};
+    u32 sampler{};
+
+    bool operator<(const BaseBindings& rhs) const {
+        return std::tie(cbuf, gmem, sampler) < std::tie(rhs.cbuf, rhs.gmem, rhs.sampler);
+    }
+};
+
 class CachedShader final : public RasterizerCacheObject {
 public:
     CachedShader(VAddr addr, Maxwell::ShaderProgram program_type);
@@ -44,71 +57,42 @@ public:
     }
 
     /// Gets the GL program handle for the shader
-    GLuint GetProgramHandle(GLenum primitive_mode) {
+    std::tuple<GLuint, BaseBindings> GetProgramHandle(GLenum primitive_mode,
-        if (program_type != Maxwell::ShaderProgram::Geometry) {
+                                                      BaseBindings base_bindings);
-            return program.handle;
-        }
-        switch (primitive_mode) {
-        case GL_POINTS:
-            return LazyGeometryProgram(geometry_programs.points, "points", 1, "ShaderPoints");
-        case GL_LINES:
-        case GL_LINE_STRIP:
-            return LazyGeometryProgram(geometry_programs.lines, "lines", 2, "ShaderLines");
-        case GL_LINES_ADJACENCY:
-        case GL_LINE_STRIP_ADJACENCY:
-            return LazyGeometryProgram(geometry_programs.lines_adjacency, "lines_adjacency", 4,
-                                       "ShaderLinesAdjacency");
-        case GL_TRIANGLES:
-        case GL_TRIANGLE_STRIP:
-        case GL_TRIANGLE_FAN:
-            return LazyGeometryProgram(geometry_programs.triangles, "triangles", 3,
-                                       "ShaderTriangles");
-        case GL_TRIANGLES_ADJACENCY:
-        case GL_TRIANGLE_STRIP_ADJACENCY:
-            return LazyGeometryProgram(geometry_programs.triangles_adjacency, "triangles_adjacency",
-                                       6, "ShaderTrianglesAdjacency");
-        default:
-            UNREACHABLE_MSG("Unknown primitive mode.");
-            return LazyGeometryProgram(geometry_programs.points, "points", 1, "ShaderPoints");
-        }
-    }
-
-    /// Gets the GL program resource location for the specified resource, caching as needed
-    GLuint GetProgramResourceIndex(const GLShader::ConstBufferEntry& buffer);
-
-    /// Gets the GL program resource location for the specified resource, caching as needed
-    GLuint GetProgramResourceIndex(const GLShader::GlobalMemoryEntry& global_mem);
-
-    /// Gets the GL uniform location for the specified resource, caching as needed
-    GLint GetUniformLocation(const GLShader::SamplerEntry& sampler);
 
 private:
-    /// Generates a geometry shader or returns one that already exists.
-    GLuint LazyGeometryProgram(OGLProgram& target_program, const std::string& glsl_topology,
-                               u32 max_vertices, const std::string& debug_name);
-
-    void CalculateProperties();
-
-    VAddr addr;
-    std::size_t shader_length;
-    Maxwell::ShaderProgram program_type;
-    GLShader::ShaderSetup setup;
-    GLShader::ShaderEntries entries;
-
-    // Non-geometry program.
-    OGLProgram program;
-
     // Geometry programs. These are needed because GLSL needs an input topology but it's not
     // declared by the hardware. Workaround this issue by generating a different shader per input
     // topology class.
-    struct {
+    struct GeometryPrograms {
-        std::string code;
         OGLProgram points;
         OGLProgram lines;
         OGLProgram lines_adjacency;
         OGLProgram triangles;
         OGLProgram triangles_adjacency;
-    } geometry_programs;
+    };
+
+    std::string AllocateBindings(BaseBindings base_bindings);
+
+    GLuint GetGeometryShader(GLenum primitive_mode, BaseBindings base_bindings);
+
+    /// Generates a geometry shader or returns one that already exists.
+    GLuint LazyGeometryProgram(OGLProgram& target_program, BaseBindings base_bindings,
+                               const std::string& glsl_topology, u32 max_vertices,
+                               const std::string& debug_name);
+
+    void CalculateProperties();
+
+    VAddr addr{};
+    std::size_t shader_length{};
+    Maxwell::ShaderProgram program_type{};
+    GLShader::ShaderSetup setup;
+    GLShader::ShaderEntries entries;
+
+    std::string code;
+
+    std::map<BaseBindings, OGLProgram> programs;
+    std::map<BaseBindings, GeometryPrograms> geometry_programs;
 
     std::map<u32, GLuint> cbuf_resource_cache;
     std::map<u32, GLuint> gmem_resource_cache;

src/video_core/renderer_opengl/gl_shader_decompiler.cpp

@@ -374,7 +374,8 @@ private:
     void DeclareConstantBuffers() {
         for (const auto& entry : ir.GetConstantBuffers()) {
             const auto [index, size] = entry;
-            code.AddLine("layout (std140) uniform " + GetConstBufferBlock(index) + " {");
+            code.AddLine("layout (std140, binding = CBUF_BINDING_" + std::to_string(index) +
+                         ") uniform " + GetConstBufferBlock(index) + " {");
             code.AddLine("    vec4 " + GetConstBuffer(index) + "[MAX_CONSTBUFFER_ELEMENTS];");
             code.AddLine("};");
             code.AddNewLine();
@@ -383,7 +384,10 @@ private:
 
     void DeclareGlobalMemory() {
         for (const auto& entry : ir.GetGlobalMemoryBases()) {
-            code.AddLine("layout (std430) buffer " + GetGlobalMemoryBlock(entry) + " {");
+            const std::string binding =
+                fmt::format("GMEM_BINDING_{}_{}", entry.cbuf_index, entry.cbuf_offset);
+            code.AddLine("layout (std430, binding = " + binding + ") buffer " +
+                         GetGlobalMemoryBlock(entry) + " {");
             code.AddLine("    float " + GetGlobalMemory(entry) + "[MAX_GLOBALMEMORY_ELEMENTS];");
             code.AddLine("};");
             code.AddNewLine();
@@ -413,7 +417,8 @@ private:
             if (sampler.IsShadow())
                 sampler_type += "Shadow";
 
-            code.AddLine("uniform " + sampler_type + ' ' + GetSampler(sampler) + ';');
+            code.AddLine("layout (binding = SAMPLER_BINDING_" + std::to_string(sampler.GetIndex()) +
+                         ") uniform " + sampler_type + ' ' + GetSampler(sampler) + ';');
         }
         if (!samplers.empty())
             code.AddNewLine();

src/video_core/renderer_opengl/gl_shader_decompiler.h

@@ -38,10 +38,6 @@ public:
         return index;
     }
 
-    u32 GetHash() const {
-        return (static_cast<u32>(stage) << 16) | index;
-    }
-
 private:
     std::string name;
     Maxwell::ShaderStage stage{};
@@ -62,10 +58,6 @@ public:
         return stage;
     }
 
-    u32 GetHash() const {
-        return (static_cast<u32>(stage) << 16) | static_cast<u32>(GetIndex());
-    }
-
 private:
     std::string name;
     Maxwell::ShaderStage stage{};
@@ -93,10 +85,6 @@ public:
         return stage;
     }
 
-    u32 GetHash() const {
-        return (static_cast<u32>(stage) << 24) | (cbuf_index << 16) | cbuf_offset;
-    }
-
 private:
     u32 cbuf_index{};
     u32 cbuf_offset{};

src/video_core/renderer_opengl/gl_shader_gen.cpp

@@ -20,15 +20,14 @@ static constexpr u32 PROGRAM_OFFSET{10};
 ProgramResult GenerateVertexShader(const ShaderSetup& setup) {
     const std::string id = fmt::format("{:016x}", setup.program.unique_identifier);
 
-    std::string out = "#version 430 core\n";
+    std::string out = "#extension GL_ARB_separate_shader_objects : enable\n\n";
-    out += "#extension GL_ARB_separate_shader_objects : enable\n\n";
     out += "// Shader Unique Id: VS" + id + "\n\n";
     out += GetCommonDeclarations();
 
     out += R"(
 layout (location = 0) out vec4 position;
 
-layout(std140) uniform vs_config {
+layout (std140, binding = EMULATION_UBO_BINDING) uniform vs_config {
     vec4 viewport_flip;
     uvec4 config_pack; // instance_id, flip_stage, y_direction, padding
     uvec4 alpha_test;
@@ -78,7 +77,6 @@ void main() {
 }
 
 ProgramResult GenerateGeometryShader(const ShaderSetup& setup) {
-    // Version is intentionally skipped in shader generation, it's added by the lazy compilation.
     const std::string id = fmt::format("{:016x}", setup.program.unique_identifier);
 
     std::string out = "#extension GL_ARB_separate_shader_objects : enable\n\n";
@@ -89,7 +87,7 @@ ProgramResult GenerateGeometryShader(const ShaderSetup& setup) {
 layout (location = 0) in vec4 gs_position[];
 layout (location = 0) out vec4 position;
 
-layout (std140) uniform gs_config {
+layout (std140, binding = EMULATION_UBO_BINDING) uniform gs_config {
     vec4 viewport_flip;
     uvec4 config_pack; // instance_id, flip_stage, y_direction, padding
     uvec4 alpha_test;
@@ -112,8 +110,7 @@ void main() {
 ProgramResult GenerateFragmentShader(const ShaderSetup& setup) {
     const std::string id = fmt::format("{:016x}", setup.program.unique_identifier);
 
-    std::string out = "#version 430 core\n";
+    std::string out = "#extension GL_ARB_separate_shader_objects : enable\n\n";
-    out += "#extension GL_ARB_separate_shader_objects : enable\n\n";
     out += "// Shader Unique Id: FS" + id + "\n\n";
     out += GetCommonDeclarations();
 
@@ -129,7 +126,7 @@ layout (location = 7) out vec4 FragColor7;
 
 layout (location = 0) in vec4 position;
 
-layout (std140) uniform fs_config {
+layout (std140, binding = EMULATION_UBO_BINDING) uniform fs_config {
     vec4 viewport_flip;
     uvec4 config_pack; // instance_id, flip_stage, y_direction, padding
     uvec4 alpha_test;