Merge pull request #531 from DrChat/gl4_shader_cache

GL4 Shader Cache

src/xenia/gpu/gl4/gl4_command_processor.cc

@@ -56,7 +56,8 @@ GL4CommandProcessor::GL4CommandProcessor(GL4GraphicsSystem* graphics_system,
     : CommandProcessor(graphics_system, kernel_state),
       shader_translator_(GlslShaderTranslator::Dialect::kGL45),
       draw_batcher_(graphics_system_->register_file()),
-      scratch_buffer_(kScratchBufferCapacity, kScratchBufferAlignment) {}
+      scratch_buffer_(kScratchBufferCapacity, kScratchBufferAlignment),
+      shader_cache_(&shader_translator_) {}
 
 GL4CommandProcessor::~GL4CommandProcessor() = default;
 
@@ -324,8 +325,7 @@ void GL4CommandProcessor::ShutdownContext() {
   scratch_buffer_.Shutdown();
 
   all_pipelines_.clear();
-  all_shaders_.clear();
-  shader_cache_.clear();
+  shader_cache_.Reset();
 
   CommandProcessor::ShutdownContext();
 }
@@ -484,41 +484,8 @@ Shader* GL4CommandProcessor::LoadShader(ShaderType shader_type,
                                         uint32_t guest_address,
                                         const uint32_t* host_address,
                                         uint32_t dword_count) {
-  // Hash the input memory and lookup the shader.
-  GL4Shader* shader_ptr = nullptr;
-  uint64_t hash = XXH64(host_address, dword_count * sizeof(uint32_t), 0);
-  auto it = shader_cache_.find(hash);
-  if (it != shader_cache_.end()) {
-    // Found in the cache.
-    // TODO(benvanik): compare bytes? Likelyhood of collision is low.
-    shader_ptr = it->second;
-  } else {
-    // Not found in cache.
-    auto shader = std::make_unique<GL4Shader>(shader_type, hash, host_address,
-                                              dword_count);
-    shader_ptr = shader.get();
-    shader_cache_.insert({hash, shader_ptr});
-    all_shaders_.emplace_back(std::move(shader));
-
-    // Perform translation.
-    // If this fails the shader will be marked as invalid and ignored later.
-    if (shader_translator_.Translate(shader_ptr)) {
-      shader_ptr->Prepare();
-
-      // Dump shader files if desired.
-      if (!FLAGS_dump_shaders.empty()) {
-        shader_ptr->Dump(FLAGS_dump_shaders, "gl4");
-      }
-    } else {
-      XELOGE("Shader failed translation");
-    }
-
-    XELOGGPU("Set %s shader at %0.8X (%db):\n%s",
-             shader_type == ShaderType::kVertex ? "vertex" : "pixel",
-             guest_address, dword_count * 4,
-             shader_ptr->ucode_disassembly().c_str());
-  }
-  return shader_ptr;
+  return shader_cache_.LookupOrInsertShader(shader_type, host_address,
+                                            dword_count);
 }
 
 bool GL4CommandProcessor::IssueDraw(PrimitiveType prim_type,

src/xenia/gpu/gl4/gl4_command_processor.h

@@ -24,6 +24,7 @@
 #include "xenia/gpu/command_processor.h"
 #include "xenia/gpu/gl4/draw_batcher.h"
 #include "xenia/gpu/gl4/gl4_shader.h"
+#include "xenia/gpu/gl4/gl4_shader_cache.h"
 #include "xenia/gpu/gl4/texture_cache.h"
 #include "xenia/gpu/glsl_shader_translator.h"
 #include "xenia/gpu/register_file.h"
@@ -131,8 +132,7 @@ class GL4CommandProcessor : public CommandProcessor {
                                     GLuint depth_target);
 
   GlslShaderTranslator shader_translator_;
-  std::vector<std::unique_ptr<GL4Shader>> all_shaders_;
-  std::unordered_map<uint64_t, GL4Shader*> shader_cache_;
+  GL4ShaderCache shader_cache_;
   CachedFramebuffer* active_framebuffer_ = nullptr;
   GLuint last_framebuffer_texture_ = 0;
 

src/xenia/gpu/gl4/gl4_gpu_flags.cc

@@ -12,3 +12,6 @@
 DEFINE_bool(disable_framebuffer_readback, false,
             "Disable framebuffer readback.");
 DEFINE_bool(disable_textures, false, "Disable textures and use colors only.");
+DEFINE_string(shader_cache_dir, "",
+              "GL4 Shader cache directory (relative to Xenia). Specify an "
+              "empty string to disable the cache.");

src/xenia/gpu/gl4/gl4_gpu_flags.h

@@ -14,6 +14,7 @@
 
 DECLARE_bool(disable_framebuffer_readback);
 DECLARE_bool(disable_textures);
+DECLARE_string(shader_cache_dir);
 
 #define FINE_GRAINED_DRAW_SCOPES 0
 

src/xenia/gpu/gl4/gl4_shader.cc

@@ -51,6 +51,37 @@ bool GL4Shader::Prepare() {
   return success;
 }
 
+bool GL4Shader::LoadFromBinary(const uint8_t* blob, GLenum binary_format,
+                               size_t length) {
+  program_ = glCreateProgram();
+  glProgramBinary(program_, binary_format, blob, GLsizei(length));
+
+  GLint link_status = 0;
+  glGetProgramiv(program_, GL_LINK_STATUS, &link_status);
+  if (!link_status) {
+    // Failed to link. Not fatal - just clean up so we can get generated later.
+    XELOGD("GL4Shader::LoadFromBinary failed. Log:\n%s",
+           GetProgramInfoLog().c_str());
+    glDeleteProgram(program_);
+    program_ = 0;
+
+    return false;
+  }
+
+  // Build static vertex array descriptor.
+  if (!PrepareVertexArrayObject()) {
+    XELOGE("Unable to prepare vertex shader array object");
+    return false;
+  }
+
+  // Success!
+  host_binary_ = GetBinary();
+  host_disassembly_ = GetHostDisasmNV(host_binary_);
+
+  is_valid_ = true;
+  return true;
+}
+
 bool GL4Shader::PrepareVertexArrayObject() {
   glCreateVertexArrays(1, &vao_);
 
@@ -211,7 +242,7 @@ std::string GL4Shader::GetProgramInfoLog() {
   return log;
 }
 
-std::vector<uint8_t> GL4Shader::GetBinary() {
+std::vector<uint8_t> GL4Shader::GetBinary(GLenum* binary_format) {
   std::vector<uint8_t> binary;
 
   // Get program binary, if it's available.
@@ -219,9 +250,13 @@ std::vector<uint8_t> GL4Shader::GetBinary() {
   glGetProgramiv(program_, GL_PROGRAM_BINARY_LENGTH, &binary_length);
   if (binary_length) {
     binary.resize(binary_length);
-    GLenum binary_format;
-    glGetProgramBinary(program_, binary_length, &binary_length, &binary_format,
-                       binary.data());
+    GLenum binary_format_tmp = 0;
+    glGetProgramBinary(program_, binary_length, &binary_length,
+                       &binary_format_tmp, binary.data());
+
+    if (binary_format) {
+      *binary_format = binary_format_tmp;
+    }
   }
 
   return binary;

src/xenia/gpu/gl4/gl4_shader.h

@@ -29,7 +29,9 @@ class GL4Shader : public Shader {
   GLuint shader() const { return shader_; }
   GLuint vao() const { return vao_; }
 
-  bool Prepare() override;
+  bool Prepare();
+  bool LoadFromBinary(const uint8_t* blob, GLenum binary_format, size_t length);
+  std::vector<uint8_t> GetBinary(GLenum* binary_format = nullptr);
 
  protected:
   bool PrepareVertexArrayObject();
@@ -38,7 +40,6 @@ class GL4Shader : public Shader {
 
   std::string GetShaderInfoLog();
   std::string GetProgramInfoLog();
-  std::vector<uint8_t> GetBinary();
   static std::string GetHostDisasmNV(const std::vector<uint8_t>& binary);
 
   GLuint program_ = 0;

src/xenia/gpu/gl4/gl4_shader_cache.cc

@@ -0,0 +1,187 @@
+/**
+ ******************************************************************************
+ * Xenia : Xbox 360 Emulator Research Project                                 *
+ ******************************************************************************
+ * Copyright 2016 Ben Vanik. All rights reserved.                             *
+ * Released under the BSD license - see LICENSE in the root for more details. *
+ ******************************************************************************
+ */
+
+#include "xenia/gpu/gl4/gl4_shader_cache.h"
+
+#include <cinttypes>
+
+#include "xenia/base/filesystem.h"
+#include "xenia/base/logging.h"
+#include "xenia/base/mapped_memory.h"
+#include "xenia/gpu/gl4/gl4_gpu_flags.h"
+#include "xenia/gpu/gl4/gl4_shader.h"
+#include "xenia/gpu/glsl_shader_translator.h"
+#include "xenia/gpu/gpu_flags.h"
+
+#include "third_party/xxhash/xxhash.h"
+
+namespace xe {
+namespace gpu {
+namespace gl4 {
+
+GL4ShaderCache::GL4ShaderCache(GlslShaderTranslator* shader_translator)
+    : shader_translator_(shader_translator) {}
+
+GL4ShaderCache::~GL4ShaderCache() {}
+
+void GL4ShaderCache::Reset() {
+  shader_map_.clear();
+  all_shaders_.clear();
+}
+
+GL4Shader* GL4ShaderCache::LookupOrInsertShader(ShaderType shader_type,
+                                                const uint32_t* dwords,
+                                                uint32_t dword_count) {
+  // Hash the input memory and lookup the shader.
+  GL4Shader* shader_ptr = nullptr;
+  uint64_t hash = XXH64(dwords, dword_count * sizeof(uint32_t), 0);
+  auto it = shader_map_.find(hash);
+  if (it != shader_map_.end()) {
+    // Shader has been previously loaded.
+    // TODO(benvanik): compare bytes? Likelihood of collision is low.
+    shader_ptr = it->second;
+  } else {
+    // Check filesystem cache.
+    shader_ptr = FindCachedShader(shader_type, hash, dwords, dword_count);
+    if (shader_ptr) {
+      // Found!
+      XELOGGPU("Loaded %s shader from cache (hash: %.16" PRIX64 ")",
+               shader_type == ShaderType::kVertex ? "vertex" : "pixel", hash);
+      return shader_ptr;
+    }
+
+    // Not found in cache - load from scratch.
+    auto shader =
+        std::make_unique<GL4Shader>(shader_type, hash, dwords, dword_count);
+    shader_ptr = shader.get();
+    shader_map_.insert({hash, shader_ptr});
+    all_shaders_.emplace_back(std::move(shader));
+
+    // Perform translation.
+    // If this fails the shader will be marked as invalid and ignored later.
+    if (shader_translator_->Translate(shader_ptr)) {
+      shader_ptr->Prepare();
+      if (shader_ptr->is_valid()) {
+        CacheShader(shader_ptr);
+
+        XELOGGPU("Generated %s shader at 0x%.16" PRIX64 " (%db):\n%s",
+                 shader_type == ShaderType::kVertex ? "vertex" : "pixel",
+                 dwords, dword_count * 4,
+                 shader_ptr->ucode_disassembly().c_str());
+      }
+
+      // Dump shader files if desired.
+      if (!FLAGS_dump_shaders.empty()) {
+        shader_ptr->Dump(FLAGS_dump_shaders, "gl4");
+      }
+    } else {
+      XELOGE("Shader failed translation");
+    }
+  }
+
+  return shader_ptr;
+}
+
+void GL4ShaderCache::CacheShader(GL4Shader* shader) {
+  if (FLAGS_shader_cache_dir.empty()) {
+    // Cache disabled.
+    return;
+  }
+
+  GLenum binary_format = 0;
+  auto binary = shader->GetBinary(&binary_format);
+  if (binary.size() == 0) {
+    // No binary returned.
+    return;
+  }
+
+  auto cache_dir = xe::to_absolute_path(xe::to_wstring(FLAGS_shader_cache_dir));
+  xe::filesystem::CreateFolder(cache_dir);
+  auto filename =
+      cache_dir + xe::format_string(
+                      L"%.16" PRIX64 ".%s", shader->ucode_data_hash(),
+                      shader->type() == ShaderType::kPixel ? L"frag" : L"vert");
+  auto file = xe::filesystem::OpenFile(filename, "wb");
+  if (!file) {
+    // Not fatal, but not too good.
+    return;
+  }
+
+  std::vector<uint8_t> cached_shader_mem;
+  // Resize this vector to the final filesize (- 1 to account for dummy array
+  // in CachedShader)
+  cached_shader_mem.resize(sizeof(CachedShader) + binary.size() - 1);
+  auto cached_shader =
+      reinterpret_cast<CachedShader*>(cached_shader_mem.data());
+  cached_shader->magic = xe::byte_swap('XSHD');
+  cached_shader->version = 0;  // TODO
+  cached_shader->shader_type = uint8_t(shader->type());
+  cached_shader->binary_len = uint32_t(binary.size());
+  cached_shader->binary_format = binary_format;
+  std::memcpy(cached_shader->binary, binary.data(), binary.size());
+
+  fwrite(cached_shader_mem.data(), cached_shader_mem.size(), 1, file);
+  fclose(file);
+}
+
+GL4Shader* GL4ShaderCache::FindCachedShader(ShaderType shader_type,
+                                            uint64_t hash,
+                                            const uint32_t* dwords,
+                                            uint32_t dword_count) {
+  if (FLAGS_shader_cache_dir.empty()) {
+    // Cache disabled.
+    return nullptr;
+  }
+
+  auto cache_dir = xe::to_absolute_path(xe::to_wstring(FLAGS_shader_cache_dir));
+  auto filename =
+      cache_dir +
+      xe::format_string(L"%.16" PRIX64 ".%s", hash,
+                        shader_type == ShaderType::kPixel ? L"frag" : L"vert");
+  if (!xe::filesystem::PathExists(filename)) {
+    return nullptr;
+  }
+
+  // Shader is cached. Open it up.
+  auto map = xe::MappedMemory::Open(filename, MappedMemory::Mode::kRead);
+  if (!map) {
+    // Should not fail
+    assert_always();
+    return nullptr;
+  }
+
+  auto cached_shader = reinterpret_cast<CachedShader*>(map->data());
+  // TODO: Compare versions
+  if (cached_shader->magic != xe::byte_swap('XSHD')) {
+    return nullptr;
+  }
+
+  auto shader =
+      std::make_unique<GL4Shader>(shader_type, hash, dwords, dword_count);
+
+  // Gather the binding points.
+  // TODO: Make Shader do this on construction.
+  // TODO: Regenerate microcode disasm/etc on load.
+  shader_translator_->GatherAllBindingInformation(shader.get());
+  if (!shader->LoadFromBinary(cached_shader->binary,
+                              cached_shader->binary_format,
+                              cached_shader->binary_len)) {
+    // Failed to load from binary.
+    return nullptr;
+  }
+
+  auto shader_ptr = shader.get();
+  shader_map_.insert({hash, shader_ptr});
+  all_shaders_.emplace_back(std::move(shader));
+  return shader_ptr;
+}
+
+}  // namespace gl4
+}  // namespace gpu
+}  // namespace xe

src/xenia/gpu/gl4/gl4_shader_cache.h

@@ -0,0 +1,60 @@
+/**
+ ******************************************************************************
+ * Xenia : Xbox 360 Emulator Research Project                                 *
+ ******************************************************************************
+ * Copyright 2016 Ben Vanik. All rights reserved.                             *
+ * Released under the BSD license - see LICENSE in the root for more details. *
+ ******************************************************************************
+ */
+
+#ifndef XENIA_GPU_GL4_SHADER_CACHE_H_
+#define XENIA_GPU_GL4_SHADER_CACHE_H_
+
+#include <cstdint>
+#include <memory>
+#include <unordered_map>
+
+#include "xenia/gpu/xenos.h"
+
+namespace xe {
+namespace gpu {
+class GlslShaderTranslator;
+
+namespace gl4 {
+
+class GL4Shader;
+
+class GL4ShaderCache {
+ public:
+  GL4ShaderCache(GlslShaderTranslator* shader_translator);
+  ~GL4ShaderCache();
+
+  void Reset();
+  GL4Shader* LookupOrInsertShader(ShaderType shader_type,
+                                  const uint32_t* dwords, uint32_t dword_count);
+
+ private:
+  // Cached shader file format.
+  struct CachedShader {
+    uint32_t magic;
+    uint32_t version;        // Version of the shader translator used.
+    uint8_t shader_type;     // ShaderType enum
+    uint32_t binary_len;     // Code length
+    uint32_t binary_format;  // Binary format (from OpenGL)
+    uint8_t binary[1];       // Code
+  };
+
+  void CacheShader(GL4Shader* shader);
+  GL4Shader* FindCachedShader(ShaderType shader_type, uint64_t hash,
+                              const uint32_t* dwords, uint32_t dword_count);
+
+  GlslShaderTranslator* shader_translator_ = nullptr;
+  std::vector<std::unique_ptr<GL4Shader>> all_shaders_;
+  std::unordered_map<uint64_t, GL4Shader*> shader_map_;
+};
+
+}  // namespace gl4
+}  // namespace gpu
+}  // namespace xe
+
+#endif  // XENIA_GPU_GL4_SHADER_CACHE_H_

src/xenia/gpu/shader.cc

@@ -16,9 +16,11 @@
 #include "xenia/base/math.h"
 #include "xenia/base/memory.h"
 #include "xenia/base/string.h"
+#include "xenia/gpu/ucode.h"
 
 namespace xe {
 namespace gpu {
+using namespace ucode;
 
 Shader::Shader(ShaderType shader_type, uint64_t ucode_data_hash,
                const uint32_t* ucode_dwords, size_t ucode_dword_count)
@@ -38,7 +40,8 @@ std::string Shader::GetTranslatedBinaryString() const {
   return result;
 }
 
-void Shader::Dump(const std::string& base_path, const char* path_prefix) {
+std::pair<std::string, std::string> Shader::Dump(const std::string& base_path,
+                                                 const char* path_prefix) {
   // Ensure target path exists.
   auto target_path = xe::to_wstring(base_path);
   if (!target_path.empty()) {
@@ -79,6 +82,8 @@ void Shader::Dump(const std::string& base_path, const char* path_prefix) {
     fwrite(ucode_data_.data(), 4, ucode_data_.size(), f);
     fclose(f);
   }
+
+  return {std::string(txt_file_name), std::string(bin_file_name)};
 }
 
 }  //  namespace gpu

src/xenia/gpu/shader.h

@@ -502,6 +502,7 @@ class Shader {
 
   // Microcode dwords in host endianness.
   const std::vector<uint32_t>& ucode_data() const { return ucode_data_; }
+  uint64_t ucode_data_hash() const { return ucode_data_hash_; }
   const uint32_t* ucode_dwords() const { return ucode_data_.data(); }
   size_t ucode_dword_count() const { return ucode_data_.size(); }
 
@@ -547,13 +548,12 @@ class Shader {
   // May be empty if the host does not support saving binaries.
   const std::vector<uint8_t>& host_binary() const { return host_binary_; }
 
-  // Prepares the shader for use in the host graphics API.
-  virtual bool Prepare() { return is_valid_; }
-
   // Dumps the shader to a file in the given path based on ucode hash.
   // Both the ucode binary and disassembled and translated shader will be
   // written.
-  void Dump(const std::string& base_path, const char* path_prefix);
+  // Returns the filename of the shader and the binary.
+  std::pair<std::string, std::string> Dump(const std::string& base_path,
+                                           const char* path_prefix);
 
  protected:
   friend class ShaderTranslator;

src/xenia/gpu/shader_translator.cc

@@ -1,3 +1,4 @@
+#include "shader_translator.h"
 /**
  ******************************************************************************
  * Xenia : Xbox 360 Emulator Research Project                                 *
@@ -57,6 +58,41 @@ void ShaderTranslator::Reset() {
   }
 }
 
+bool ShaderTranslator::GatherAllBindingInformation(Shader* shader) {
+  // FIXME: This is kind of silly.
+  Reset();
+
+  shader_type_ = shader->type();
+  ucode_dwords_ = shader->ucode_dwords();
+  ucode_dword_count_ = shader->ucode_dword_count();
+
+  uint32_t max_cf_dword_index = static_cast<uint32_t>(ucode_dword_count_);
+  for (uint32_t i = 0; i < max_cf_dword_index; i += 3) {
+    ControlFlowInstruction cf_a;
+    ControlFlowInstruction cf_b;
+    UnpackControlFlowInstructions(ucode_dwords_ + i, &cf_a, &cf_b);
+    if (IsControlFlowOpcodeExec(cf_a.opcode())) {
+      max_cf_dword_index =
+          std::min(max_cf_dword_index, cf_a.exec.address() * 3);
+    }
+    if (IsControlFlowOpcodeExec(cf_b.opcode())) {
+      max_cf_dword_index =
+          std::min(max_cf_dword_index, cf_b.exec.address() * 3);
+    }
+
+    GatherBindingInformation(cf_a);
+    GatherBindingInformation(cf_b);
+  }
+
+  shader->vertex_bindings_ = std::move(vertex_bindings_);
+  shader->texture_bindings_ = std::move(texture_bindings_);
+  for (size_t i = 0; i < xe::countof(writes_color_targets_); ++i) {
+    shader->writes_color_targets_[i] = writes_color_targets_[i];
+  }
+
+  return true;
+}
+
 bool ShaderTranslator::Translate(Shader* shader) {
   Reset();
 
@@ -79,6 +115,7 @@ bool ShaderTranslator::Translate(Shader* shader) {
       max_cf_dword_index =
           std::min(max_cf_dword_index, cf_b.exec.address() * 3);
     }
+
     GatherBindingInformation(cf_a);
     GatherBindingInformation(cf_b);
   }

src/xenia/gpu/shader_translator.h

@@ -26,6 +26,9 @@ class ShaderTranslator {
  public:
   virtual ~ShaderTranslator();
 
+  // Gathers all vertex/texture bindings. Implicitly called in Translate.
+  // TODO: Move this functionality to Shader.
+  bool GatherAllBindingInformation(Shader* shader);
   bool Translate(Shader* shader);
 
  protected: