D3D: Uber shader support

2025-07-24 22:59:47 -06:00 · 2017-07-20 15:25:31 +10:00
parent cd502990fa
commit 4bf5625895
15 changed files with 720 additions and 193 deletions
--- a/Source/Core/VideoBackends/D3D/PixelShaderCache.cpp
+++ b/Source/Core/VideoBackends/D3D/PixelShaderCache.cpp
@ -8,12 +8,15 @@
 #include "Common/CommonTypes.h"
 #include "Common/FileUtil.h"
 #include "Common/LinearDiskCache.h"
+#include "Common/MsgHandler.h"
 #include "Common/StringUtil.h"

 #include "Core/ConfigManager.h"
+#include "Core/Host.h"

 #include "VideoBackends/D3D/D3DBase.h"
 #include "VideoBackends/D3D/D3DShader.h"
+#include "VideoBackends/D3D/D3DState.h"
 #include "VideoBackends/D3D/PixelShaderCache.h"

 #include "VideoCommon/Debugger.h"
@ -25,10 +28,15 @@
 namespace DX11
 {
 PixelShaderCache::PSCache PixelShaderCache::PixelShaders;
+PixelShaderCache::UberPSCache PixelShaderCache::UberPixelShaders;
 const PixelShaderCache::PSCacheEntry* PixelShaderCache::last_entry;
+const PixelShaderCache::PSCacheEntry* PixelShaderCache::last_uber_entry;
 PixelShaderUid PixelShaderCache::last_uid;
+UberShader::PixelShaderUid PixelShaderCache::last_uber_uid;

 LinearDiskCache<PixelShaderUid, u8> g_ps_disk_cache;
+LinearDiskCache<UberShader::PixelShaderUid, u8> g_uber_ps_disk_cache;
+extern std::unique_ptr<VideoCommon::AsyncShaderCompiler> g_async_compiler;

 ID3D11PixelShader* s_ColorMatrixProgram[2] = {nullptr};
 ID3D11PixelShader* s_ColorCopyProgram[2] = {nullptr};
@ -429,10 +437,8 @@ ID3D11PixelShader* PixelShaderCache::GetDepthResolveProgram()
  return s_DepthResolveProgram;
 }

-ID3D11Buffer*& PixelShaderCache::GetConstantBuffer()
+static void UpdateConstantBuffers()
 {
-  // TODO: divide the global variables of the generated shaders into about 5 constant buffers to
-  // speed this up
  if (PixelShaderManager::dirty)
  {
    D3D11_MAPPED_SUBRESOURCE map;
@ -443,14 +449,20 @@ ID3D11Buffer*& PixelShaderCache::GetConstantBuffer()

    ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(PixelShaderConstants));
  }
+}
+
+ID3D11Buffer* PixelShaderCache::GetConstantBuffer()
+{
+  UpdateConstantBuffers();
  return pscbuf;
 }

 // this class will load the precompiled shaders into our cache
-class PixelShaderCacheInserter : public LinearDiskCacheReader<PixelShaderUid, u8>
+template <typename UidType>
+class PixelShaderCacheInserter : public LinearDiskCacheReader<UidType, u8>
 {
 public:
-  void Read(const PixelShaderUid& key, const u8* value, u32 value_size)
+  void Read(const UidType& key, const u8* value, u32 value_size)
  {
    PixelShaderCache::InsertByteCode(key, value, value_size);
  }
@ -499,22 +511,34 @@ void PixelShaderCache::Init()

  if (g_ActiveConfig.bShaderCache)
    LoadShaderCache();
+
+  if (g_ActiveConfig.CanPrecompileUberShaders())
+    PrecompileUberShaders();
 }

 void PixelShaderCache::LoadShaderCache()
 {
-  PixelShaderCacheInserter inserter;
+  PixelShaderCacheInserter<PixelShaderUid> inserter;
  g_ps_disk_cache.OpenAndRead(GetDiskShaderCacheFileName(APIType::D3D, "PS", true, true), inserter);
+
+  PixelShaderCacheInserter<UberShader::PixelShaderUid> uber_inserter;
+  g_uber_ps_disk_cache.OpenAndRead(GetDiskShaderCacheFileName(APIType::D3D, "UberPS", false, true),
+                                   uber_inserter);
 }

 void PixelShaderCache::Reload()
 {
  g_ps_disk_cache.Sync();
  g_ps_disk_cache.Close();
+  g_uber_ps_disk_cache.Sync();
+  g_uber_ps_disk_cache.Close();
  Clear();

  if (g_ActiveConfig.bShaderCache)
    LoadShaderCache();
+
+  if (g_ActiveConfig.CanPrecompileUberShaders())
+    PrecompileUberShaders();
 }

 // ONLY to be used during shutdown.
@ -522,10 +546,15 @@ void PixelShaderCache::Clear()
 {
  for (auto& iter : PixelShaders)
    iter.second.Destroy();
+  for (auto& iter : UberPixelShaders)
+    iter.second.Destroy();
  PixelShaders.clear();
+  UberPixelShaders.clear();

  last_entry = nullptr;
+  last_uber_entry = nullptr;
  last_uid = {};
+  last_uber_uid = {};
 }

 // Used in Swap() when AA mode has changed
@ -558,82 +587,252 @@ void PixelShaderCache::Shutdown()
  Clear();
  g_ps_disk_cache.Sync();
  g_ps_disk_cache.Close();
+  g_uber_ps_disk_cache.Sync();
+  g_uber_ps_disk_cache.Close();
 }

 bool PixelShaderCache::SetShader()
 {
-  PixelShaderUid uid = GetPixelShaderUid();
-
-  // Check if the shader is already set
-  if (last_entry)
+  if (g_ActiveConfig.CanUseUberShaders() &&
+      (g_ActiveConfig.bDisableSpecializedShaders || g_ActiveConfig.bForcePixelUberShaders))
  {
-    if (uid == last_uid)
-    {
-      GFX_DEBUGGER_PAUSE_AT(NEXT_PIXEL_SHADER_CHANGE, true);
-      return (last_entry->shader != nullptr);
-    }
+    return SetUberShader();
  }

-  last_uid = uid;
+  PixelShaderUid uid = GetPixelShaderUid();
+  if (last_entry && uid == last_uid)
+  {
+    if (last_entry->pending)
+      return SetUberShader();
+
+    if (!last_entry->shader)
+      return false;
+
+    D3D::stateman->SetPixelShader(last_entry->shader);
+    return true;
+  }

  // Check if the shader is already in the cache
-  PSCache::iterator iter;
-  iter = PixelShaders.find(uid);
+  auto iter = PixelShaders.find(uid);
  if (iter != PixelShaders.end())
  {
    const PSCacheEntry& entry = iter->second;
+    if (entry.pending)
+      return SetUberShader();
+
+    last_uid = uid;
    last_entry = &entry;

    GFX_DEBUGGER_PAUSE_AT(NEXT_PIXEL_SHADER_CHANGE, true);
-    return (entry.shader != nullptr);
+    if (!last_entry->shader)
+      return false;
+
+    D3D::stateman->SetPixelShader(last_entry->shader);
+    return true;
+  }
+
+  // Background compiling?
+  if (g_ActiveConfig.CanBackgroundCompileShaders())
+  {
+    // Create a pending entry
+    PSCacheEntry entry;
+    entry.pending = true;
+    PixelShaders[uid] = entry;
+
+    // Queue normal shader compiling and use ubershader
+    g_async_compiler->QueueWorkItem(
+        g_async_compiler->CreateWorkItem<PixelShaderCompilerWorkItem>(uid));
+    return SetUberShader();
  }

  // Need to compile a new shader
+  D3DBlob* bytecode = nullptr;
  ShaderCode code =
      GeneratePixelShaderCode(APIType::D3D, ShaderHostConfig::GetCurrent(), uid.GetUidData());
-
-  D3DBlob* pbytecode;
-  if (!D3D::CompilePixelShader(code.GetBuffer(), &pbytecode))
+  D3D::CompilePixelShader(code.GetBuffer(), &bytecode);
+  if (!InsertByteCode(uid, bytecode->Data(), bytecode->Size()))
  {
-    GFX_DEBUGGER_PAUSE_AT(NEXT_ERROR, true);
+    SAFE_RELEASE(bytecode);
    return false;
  }

-  // Insert the bytecode into the caches
-  g_ps_disk_cache.Append(uid, pbytecode->Data(), pbytecode->Size());
-
-  bool success = InsertByteCode(uid, pbytecode->Data(), pbytecode->Size());
-  pbytecode->Release();
-
-  GFX_DEBUGGER_PAUSE_AT(NEXT_PIXEL_SHADER_CHANGE, true);
-  return success;
+  g_ps_disk_cache.Append(uid, bytecode->Data(), bytecode->Size());
+  return SetShader();
 }

-bool PixelShaderCache::InsertByteCode(const PixelShaderUid& uid, const void* bytecode,
-                                      unsigned int bytecodelen)
+bool PixelShaderCache::SetUberShader()
 {
-  ID3D11PixelShader* shader = D3D::CreatePixelShaderFromByteCode(bytecode, bytecodelen);
-  if (shader == nullptr)
-    return false;
+  UberShader::PixelShaderUid uid = UberShader::GetPixelShaderUid();

-  // TODO: Somehow make the debug name a bit more specific
-  D3D::SetDebugObjectName((ID3D11DeviceChild*)shader, "a pixel shader of PixelShaderCache");
-
-  // Make an entry in the table
-  PSCacheEntry newentry;
-  newentry.shader = shader;
-  PixelShaders[uid] = newentry;
-  last_entry = &PixelShaders[uid];
-
-  if (!shader)
+  if (last_uber_entry && last_uber_uid == uid)
  {
-    // INCSTAT(stats.numPixelShadersFailed);
+    if (!last_uber_entry->shader)
+      return false;
+
+    D3D::stateman->SetPixelShader(last_uber_entry->shader);
+    return true;
+  }
+
+  auto iter = UberPixelShaders.find(uid);
+  if (iter != UberPixelShaders.end())
+  {
+    const PSCacheEntry& entry = iter->second;
+    last_uber_uid = uid;
+    last_uber_entry = &entry;
+
+    GFX_DEBUGGER_PAUSE_AT(NEXT_PIXEL_SHADER_CHANGE, true);
+    if (!last_uber_entry->shader)
+      return false;
+
+    D3D::stateman->SetPixelShader(last_uber_entry->shader);
+    return true;
+  }
+
+  D3DBlob* bytecode = nullptr;
+  ShaderCode code =
+      UberShader::GenPixelShader(APIType::D3D, ShaderHostConfig::GetCurrent(), uid.GetUidData());
+  D3D::CompilePixelShader(code.GetBuffer(), &bytecode);
+  if (!InsertByteCode(uid, bytecode->Data(), bytecode->Size()))
+  {
+    SAFE_RELEASE(bytecode);
+    return false;
+  }
+
+  // Lookup map again.
+  g_uber_ps_disk_cache.Append(uid, bytecode->Data(), bytecode->Size());
+  bytecode->Release();
+  return SetUberShader();
+}
+
+bool PixelShaderCache::InsertByteCode(const PixelShaderUid& uid, const u8* data, size_t len)
+{
+  ID3D11PixelShader* shader = data ? D3D::CreatePixelShaderFromByteCode(data, len) : nullptr;
+  if (!InsertShader(uid, shader))
+  {
+    SAFE_RELEASE(shader);
    return false;
  }

-  INCSTAT(stats.numPixelShadersCreated);
-  SETSTAT(stats.numPixelShadersAlive, PixelShaders.size());
  return true;
 }

+bool PixelShaderCache::InsertByteCode(const UberShader::PixelShaderUid& uid, const u8* data,
+                                      size_t len)
+{
+  ID3D11PixelShader* shader = data ? D3D::CreatePixelShaderFromByteCode(data, len) : nullptr;
+  if (!InsertShader(uid, shader))
+  {
+    SAFE_RELEASE(shader);
+    return false;
+  }
+
+  return true;
+}
+
+bool PixelShaderCache::InsertShader(const PixelShaderUid& uid, ID3D11PixelShader* shader)
+{
+  auto iter = PixelShaders.find(uid);
+  if (iter != PixelShaders.end() && !iter->second.pending)
+    return false;
+
+  PSCacheEntry& newentry = PixelShaders[uid];
+  newentry.pending = false;
+  newentry.shader = shader;
+
+  INCSTAT(stats.numPixelShadersCreated);
+  SETSTAT(stats.numPixelShadersAlive, PixelShaders.size());
+  return (shader != nullptr);
+}
+
+bool PixelShaderCache::InsertShader(const UberShader::PixelShaderUid& uid,
+                                    ID3D11PixelShader* shader)
+{
+  auto iter = UberPixelShaders.find(uid);
+  if (iter != UberPixelShaders.end() && !iter->second.pending)
+    return false;
+
+  PSCacheEntry& newentry = UberPixelShaders[uid];
+  newentry.pending = false;
+  newentry.shader = shader;
+  return (shader != nullptr);
+}
+
+void PixelShaderCache::PrecompileUberShaders()
+{
+  UberShader::EnumeratePixelShaderUids([&](const UberShader::PixelShaderUid& uid) {
+    if (UberPixelShaders.find(uid) != UberPixelShaders.end())
+      return;
+
+    g_async_compiler->QueueWorkItem(
+        g_async_compiler->CreateWorkItem<UberPixelShaderCompilerWorkItem>(uid));
+  });
+
+  g_async_compiler->WaitUntilCompletion([](size_t completed, size_t total) {
+    Host_UpdateProgressDialog(GetStringT("Compiling shaders...").c_str(),
+                              static_cast<int>(completed), static_cast<int>(total));
+  });
+  g_async_compiler->RetrieveWorkItems();
+  Host_UpdateProgressDialog("", -1, -1);
+}
+
+PixelShaderCache::PixelShaderCompilerWorkItem::PixelShaderCompilerWorkItem(
+    const PixelShaderUid& uid)
+{
+  std::memcpy(&m_uid, &uid, sizeof(uid));
+}
+
+PixelShaderCache::PixelShaderCompilerWorkItem::~PixelShaderCompilerWorkItem()
+{
+  SAFE_RELEASE(m_bytecode);
+}
+
+bool PixelShaderCache::PixelShaderCompilerWorkItem::Compile()
+{
+  ShaderCode code =
+      GeneratePixelShaderCode(APIType::D3D, ShaderHostConfig::GetCurrent(), m_uid.GetUidData());
+
+  if (D3D::CompilePixelShader(code.GetBuffer(), &m_bytecode))
+    m_shader = D3D::CreatePixelShaderFromByteCode(m_bytecode);
+
+  return true;
+}
+
+void PixelShaderCache::PixelShaderCompilerWorkItem::Retrieve()
+{
+  if (InsertShader(m_uid, m_shader))
+    g_ps_disk_cache.Append(m_uid, m_bytecode->Data(), m_bytecode->Size());
+  else
+    SAFE_RELEASE(m_shader);
+}
+
+PixelShaderCache::UberPixelShaderCompilerWorkItem::UberPixelShaderCompilerWorkItem(
+    const UberShader::PixelShaderUid& uid)
+{
+  std::memcpy(&m_uid, &uid, sizeof(uid));
+}
+
+PixelShaderCache::UberPixelShaderCompilerWorkItem::~UberPixelShaderCompilerWorkItem()
+{
+  SAFE_RELEASE(m_bytecode);
+}
+
+bool PixelShaderCache::UberPixelShaderCompilerWorkItem::Compile()
+{
+  ShaderCode code =
+      UberShader::GenPixelShader(APIType::D3D, ShaderHostConfig::GetCurrent(), m_uid.GetUidData());
+
+  if (D3D::CompilePixelShader(code.GetBuffer(), &m_bytecode))
+    m_shader = D3D::CreatePixelShaderFromByteCode(m_bytecode);
+
+  return true;
+}
+
+void PixelShaderCache::UberPixelShaderCompilerWorkItem::Retrieve()
+{
+  if (InsertShader(m_uid, m_shader))
+    g_uber_ps_disk_cache.Append(m_uid, m_bytecode->Data(), m_bytecode->Size());
+  else
+    SAFE_RELEASE(m_shader);
+}
+
 }  // DX11