Various bug fixes

blender/hecl/frme.py

@@ -183,7 +183,7 @@ def recursive_cook(buffer, obj, version, path_hasher, parent_name):
 
     elif obj.retro_widget_type == 'RETRO_ENRG':
         path_hash = path_hasher.hashpath32(obj.retro_energybar_texture_path)
-        buffer += struct.pack('>ffff', path_hash)
+        buffer += struct.pack('>I', path_hash)
 
     elif obj.retro_widget_type == 'RETRO_METR':
         buffer += struct.pack('>bbII',

include/hecl/ClientProcess.hpp

@@ -4,6 +4,7 @@
 #include "hecl.hpp"
 #include "Database.hpp"
 #include "hecl/Blender/BlenderConnection.hpp"
+#include "boo/ThreadLocalPtr.hpp"
 #include <list>
 #include <thread>
 #include <mutex>
@@ -80,6 +81,7 @@ class ClientProcess
         void proc();
     };
     std::vector<Worker> m_workers;
+    static ThreadLocalPtr<ClientProcess::Worker> ThreadWorker;
 
 public:
     ClientProcess(int verbosityLevel=1);
@@ -95,6 +97,15 @@ class ClientProcess
     void swapCompletedQueue(std::list<std::shared_ptr<Transaction>>& queue);
     void waitUntilComplete();
     void shutdown();
+    bool isBusy() const { return m_pendingQueue.size() || m_inProgress; }
+
+    static int GetThreadWorkerIdx()
+    {
+        Worker* w = ThreadWorker.get();
+        if (w)
+            return w->m_idx;
+        return -1;
+    }
 };
 
 }

include/hecl/Runtime.hpp

@@ -145,6 +145,7 @@ class ShaderCacheExtensions
         bool noDepthWrite = false;
         bool noColorWrite = false;
         bool noAlphaWrite = false;
+        bool noReflection = false;
     };
     std::vector<ExtensionSlot> m_extensionSlots;
 
@@ -164,7 +165,8 @@ class ShaderCacheExtensions
                                    bool frontfaceCull = false,
                                    bool noDepthWrite = false,
                                    bool noColorWrite = false,
-                                   bool noAlphaWrite = false)
+                                   bool noAlphaWrite = false,
+                                   bool noReflection = false)
     {
         m_extensionSlots.emplace_back();
         ExtensionSlot& slot = m_extensionSlots.back();
@@ -181,6 +183,7 @@ class ShaderCacheExtensions
         slot.noDepthWrite = noDepthWrite;
         slot.noColorWrite = noColorWrite;
         slot.noAlphaWrite = noAlphaWrite;
+        slot.noReflection = noReflection;
         return m_extensionSlots.size() - 1;
     }
 };
@@ -247,7 +250,7 @@ class ShaderCacheManager
     };
     std::vector<IndexEntry> m_entries;
     std::unordered_map<Hash, size_t> m_entryLookup;
-    std::unordered_map<Hash, std::weak_ptr<ShaderPipelines>> m_pipelineLookup;
+    std::unordered_map<Hash, std::shared_ptr<ShaderPipelines>> m_pipelineLookup;
 
     uint64_t m_timeHash = 0;
     void bootstrapIndex();
@@ -265,6 +268,7 @@ class ShaderCacheManager
                        boo::IGraphicsDataFactory* gfxFactory)
     : ShaderCacheManager(storeMgr, gfxFactory, ShaderCacheExtensions()) {}
     void reload();
+    void clearCachedPipelines() { m_pipelineLookup.clear(); }
 
     /* Some platforms (like Metal) require information about the render target
      * for encoding the pipeline state. This must be called before building shaders */

lib/Backend/GLSL.cpp

@@ -216,8 +216,8 @@ std::string GLSL::makeVert(const char* glslVer, unsigned col, unsigned uv, unsig
     }
 
     if (reflectionType != ReflectionType::None)
-        retval += "    vtf.reflectTcgs[0] = normalize((indMtx * vec4(v.posIn, 1.0)).xz) * vec2(0.5, 0.5) + vec2(0.5, 0.5);\n"
-                  "    vtf.reflectTcgs[1] = (reflectMtx * vec4(v.posIn, 1.0)).xy;\n"
+        retval += "    vtf.reflectTcgs[0] = normalize((indMtx * vec4(posIn, 1.0)).xz) * vec2(0.5, 0.5) + vec2(0.5, 0.5);\n"
+                  "    vtf.reflectTcgs[1] = (reflectMtx * vec4(posIn, 1.0)).xy;\n"
                   "    vtf.reflectAlpha = reflectAlpha;\n";
 
     return retval + "}\n";

lib/Backend/Metal.cpp

@@ -396,10 +396,8 @@ std::string Metal::makeFrag(size_t blockCount, const char** blockNames, bool alp
     {
         if (lighting.m_entry)
         {
-            if (!strncmp(lighting.m_entry, "EXT", 3))
-                retval += "    float4 lighting = " + lightingEntry + "(" + blockCall + ", vtf.mvPos, vtf.mvNorm, vtf, " + (extTexCall.size() ? (extTexCall + ", ") : "") + ");\n";
-            else
-                retval += "    float4 lighting = " + lightingEntry + "(" + blockCall + ", vtf.mvPos, vtf.mvNorm, vtf);\n";
+                retval += "    float4 lighting = " + lightingEntry + "(" + blockCall + ", vtf.mvPos, vtf.mvNorm, vtf" +
+                    (!strncmp(lighting.m_entry, "EXT", 3) ? (extTexCall.size() ? (", " + extTexCall) : "") : "") + ");\n";
         }
         else
             retval += "    float4 lighting = float4(1.0,1.0,1.0,1.0);\n";
@@ -413,11 +411,19 @@ std::string Metal::makeFrag(size_t blockCount, const char** blockNames, bool alp
     std::string reflectionExpr = GenerateReflectionExpr(reflectionType);
 
     if (m_alphaExpr.size())
-        retval += "    out.color = " + postEntry + "(" + (postEntry.size() ? ("vtf, " + (blockCall.size() ? (blockCall + ", ") : "") + (extTexCall.size() ? (extTexCall + ", ") : "")) : "") +
+    {
+        retval += "    out.color = " + postEntry + "(" +
+                  (postEntry.size() ? ("vtf, " + (blockCall.size() ? (blockCall + ", ") : "") +
+                      (!strncmp(post.m_entry, "EXT", 3) ? (extTexCall.size() ? (extTexCall + ", ") : "") : "")) : "") +
                   "float4(" + m_colorExpr + " + " + reflectionExpr + ", " + m_alphaExpr + ")) * mulColor;\n";
+    }
     else
-        retval += "    out.color = " + postEntry + "(" + (postEntry.size() ? ("vtf, " + (blockCall.size() ? (blockCall + ", ") : "") + (extTexCall.size() ? (extTexCall + ", ") : "")) : "") +
+    {
+        retval += "    out.color = " + postEntry + "(" +
+                  (postEntry.size() ? ("vtf, " + (blockCall.size() ? (blockCall + ", ") : "") +
+                      (!strncmp(post.m_entry, "EXT", 3) ? (extTexCall.size() ? (extTexCall + ", ") : "") : "")) : "") +
                   "float4(" + m_colorExpr + " + " + reflectionExpr + ", 1.0)) * mulColor;\n";
+    }
 
     return retval + (alphaTest ? GenerateAlphaTest() : "") +
            "    //out.depth = 1.0 - float(int((1.0 - vtf.mvpPos.z) * 16777216.0)) / 16777216.0;\n"
@@ -526,10 +532,11 @@ struct MetalBackendFactory : IShaderBackendFactory
         {
             sources.emplace_back(m_backend.makeVert(tag.getColorCount(), tag.getUvCount(), tag.getWeightCount(),
                                                     tag.getSkinSlotCount(), tag.getTexMtxCount(), slot.texCount, slot.texs,
-                                                    tag.getReflectionType()),
+                                                    slot.noReflection ? Backend::ReflectionType::None : tag.getReflectionType()),
                                  m_backend.makeFrag(slot.blockCount, slot.blockNames,
                                                     tag.getDepthWrite() && m_backend.m_blendDst == hecl::Backend::BlendFactor::InvSrcAlpha,
-                                                    tag.getReflectionType(), slot.lighting, slot.post, slot.texCount, slot.texs));
+                                                    slot.noReflection ? Backend::ReflectionType::None : tag.getReflectionType(),
+                                                    slot.lighting, slot.post, slot.texCount, slot.texs));
             cachedSz += sources.back().first.size() + 1;
             cachedSz += sources.back().second.size() + 1;
 

lib/Backend/ProgrammableCommon.cpp

@@ -191,7 +191,7 @@ std::string ProgrammableCommon::RecursiveTraceAlpha(const IR& ir, Diagnostics& d
         {
             const IR::Instruction& idxInst = inst.getChildInst(ir, 0);
             unsigned idx = unsigned(idxInst.getImmVec().vec[0]);
-            return EmitColorRegUseAlpha(idx);
+            return toSwizzle ? EmitColorRegUseRaw(idx) : EmitColorRegUseAlpha(idx);
         }
         else if (!name.compare("Lighting"))
         {

lib/Blender/BlenderConnection.cpp

@@ -518,7 +518,8 @@ BlenderConnection::BlenderConnection(int verbosityLevel)
             InstallAddon(blenderAddonPath.c_str());
             ++installAttempt;
             if (installAttempt >= 2)
-                BlenderLog.report(logvisor::Fatal, _S("unable to install blender addon using '%s'"), blenderAddonPath.c_str());
+                BlenderLog.report(logvisor::Fatal, _S("unable to install blender addon using '%s'"),
+                                  blenderAddonPath.c_str());
             continue;
         }
         else if (!strcmp(lineBuf, "ADDONINSTALLED"))
@@ -600,7 +601,8 @@ bool BlenderConnection::createBlend(const ProjectPath& path, BlendType type)
                           "BlenderConnection::createBlend() musn't be called with stream active");
         return false;
     }
-    _writeStr(("CREATE \"" + path.getAbsolutePathUTF8() + "\" " + BlendTypeStrs[int(type)] + " \"" + m_startupBlend + "\"").c_str());
+    _writeStr(("CREATE \"" + path.getAbsolutePathUTF8() + "\" " + BlendTypeStrs[int(type)] +
+        " \"" + m_startupBlend + "\"").c_str());
     char lineBuf[256];
     _readStr(lineBuf, sizeof(lineBuf));
     if (!strcmp(lineBuf, "FINISHED"))

lib/ClientProcess.cpp

@@ -10,10 +10,14 @@
 #include <sys/wait.h>
 #endif
 
+#define HECL_MULTIPROCESSOR 1
+
 namespace hecl
 {
 static logvisor::Module Log("hecl::ClientProcess");
 
+ThreadLocalPtr<ClientProcess::Worker> ClientProcess::ThreadWorker;
+
 static int GetCPUCount()
 {
 #if _WIN32
@@ -68,6 +72,8 @@ ClientProcess::Worker::Worker(ClientProcess& proc, int idx)
 
 void ClientProcess::Worker::proc()
 {
+    ClientProcess::ThreadWorker.reset(this);
+
     char thrName[64];
     snprintf(thrName, 64, "HECL Client Worker %d", m_idx);
     logvisor::RegisterThreadName(thrName);
@@ -172,7 +178,7 @@ void ClientProcess::swapCompletedQueue(std::list<std::shared_ptr<Transaction>>&
 void ClientProcess::waitUntilComplete()
 {
     std::unique_lock<std::mutex> lk(m_mutex);
-    while (m_pendingQueue.size() || m_inProgress)
+    while (isBusy())
         m_waitCv.wait(lk);
 }
 

lib/Runtime/ShaderCacheManager.cpp

@@ -384,8 +384,7 @@ ShaderCacheManager::buildShader(const ShaderTag& tag, const std::string& source,
 {
     auto search = m_pipelineLookup.find(tag);
     if (search != m_pipelineLookup.cend())
-        if (auto ret = search->second.lock())
-            return ret;
+        return search->second;
 
     std::shared_ptr<ShaderPipelines> ret = std::make_shared<ShaderPipelines>();
     ShaderCachedData foundData = lookupData(tag);
@@ -432,8 +431,7 @@ ShaderCacheManager::buildShader(const ShaderTag& tag, const hecl::Frontend::IR&
 {
     auto search = m_pipelineLookup.find(tag);
     if (search != m_pipelineLookup.cend())
-        if (auto ret = search->second.lock())
-            return ret;
+        return search->second;
 
     std::shared_ptr<ShaderPipelines> ret = std::make_shared<ShaderPipelines>();
     ShaderCachedData foundData = lookupData(tag);
@@ -494,8 +492,7 @@ ShaderCacheManager::buildExtendedShader(const ShaderTag& tag, const std::string&
 {
     auto search = m_pipelineLookup.find(tag);
     if (search != m_pipelineLookup.cend())
-        if (auto ret = search->second.lock())
-            return ret;
+        return search->second;
 
     std::shared_ptr<ShaderPipelines> ret = std::make_shared<ShaderPipelines>();
     ShaderCachedData foundData = lookupData(tag);
@@ -545,8 +542,7 @@ ShaderCacheManager::buildExtendedShader(const ShaderTag& tag, const hecl::Fronte
 {
     auto search = m_pipelineLookup.find(tag);
     if (search != m_pipelineLookup.cend())
-        if (auto ret = search->second.lock())
-            return ret;
+        return search->second;
 
     std::shared_ptr<ShaderPipelines> ret = std::make_shared<ShaderPipelines>();
     ShaderCachedData foundData = lookupData(tag);