Merge "Handle FP precision in kernel fusion"

3 files changed, 29 insertions, 7 deletions

diff --git a/include/bcc/Renderscript/RSCompilerDriver.h b/include/bcc/Renderscript/RSCompilerDriver.h
index 68f5fd4..5c0d6e8 100644
--- a/include/bcc/Renderscript/RSCompilerDriver.h
+++ b/include/bcc/Renderscript/RSCompilerDriver.h
@@ -121,7 +121,8 @@ public:
 
   bool buildScriptGroup(
       BCCContext& Context, const char* pOutputFilepath, const char* pRuntimePath,
-      bool dumpIR, const std::vector<Source*>& sources,
+      const char* pRuntimeRelaxedPath, bool dumpIR,
+      const std::vector<Source*>& sources,
       const std::list<std::list<std::pair<int, int>>>& toFuse,
       const std::list<std::string>& fused,
       const std::list<std::list<std::pair<int, int>>>& invokes,
diff --git a/lib/Renderscript/RSCompilerDriver.cpp b/lib/Renderscript/RSCompilerDriver.cpp
index 21beaa2..757d920 100644
--- a/lib/Renderscript/RSCompilerDriver.cpp
+++ b/lib/Renderscript/RSCompilerDriver.cpp
@@ -155,7 +155,8 @@ Compiler::ErrorCode RSCompilerDriver::compileScript(RSScript& pScript, const cha
   // Link RS script with Renderscript runtime.
   //===--------------------------------------------------------------------===//
   if (!RSScript::LinkRuntime(pScript, pRuntimePath)) {
-    ALOGE("Failed to link script '%s' with Renderscript runtime!", pScriptName);
+    ALOGE("Failed to link script '%s' with Renderscript runtime %s!",
+          pScriptName, pRuntimePath);
     return Compiler::kErrInvalidSource;
   }
 
@@ -331,7 +332,8 @@ bool RSCompilerDriver::build(BCCContext &pContext,
 
 bool RSCompilerDriver::buildScriptGroup(
     BCCContext& Context, const char* pOutputFilepath, const char* pRuntimePath,
-    bool dumpIR, const std::vector<Source*>& sources,
+    const char* pRuntimeRelaxedPath, bool dumpIR,
+    const std::vector<Source*>& sources,
     const std::list<std::list<std::pair<int, int>>>& toFuse,
     const std::list<std::string>& fused,
     const std::list<std::list<std::pair<int, int>>>& invokes,
@@ -414,7 +416,17 @@ bool RSCompilerDriver::buildScriptGroup(
   llvm::SmallString<80> output_path(pOutputFilepath);
   llvm::sys::path::replace_extension(output_path, ".o");
 
-  compileScript(script, pOutputFilepath, output_path.c_str(), pRuntimePath,
+  // Pick the right runtime lib
+  const char* coreLibPath = pRuntimePath;
+  if (strcmp(pRuntimeRelaxedPath, "")) {
+      bcinfo::MetadataExtractor me(&module);
+      me.extract();
+      if (me.getRSFloatPrecision() == bcinfo::RS_FP_Relaxed) {
+          coreLibPath = pRuntimeRelaxedPath;
+      }
+  }
+
+  compileScript(script, pOutputFilepath, output_path.c_str(), coreLibPath,
                 bitcode_sha1, compileCommandLineToEmbed, buildChecksum,
                 true, dumpIR);
 
@@ -443,8 +455,10 @@ bool RSCompilerDriver::buildForCompatLib(RSScript &pScript, const char *pOut,
   // offline (host) compilation.
   pScript.setEmbedInfo(true);
 
-  Compiler::ErrorCode status = compileScript(pScript, pOut, pOut, pRuntimePath, bitcode_sha1,
-                                             compileCommandLineToEmbed, pBuildChecksum, false, pDumpIR);
+  Compiler::ErrorCode status = compileScript(pScript, pOut, pOut, pRuntimePath,
+                                             bitcode_sha1,
+                                             compileCommandLineToEmbed,
+                                             pBuildChecksum, false, pDumpIR);
   if (status != Compiler::kSuccess) {
     return false;
   }
diff --git a/tools/bcc/Main.cpp b/tools/bcc/Main.cpp
index 6e2c8a4..a047e88 100644
--- a/tools/bcc/Main.cpp
+++ b/tools/bcc/Main.cpp
@@ -79,6 +79,12 @@ OptBCLibFilename("bclib", llvm::cl::desc("Specify the bclib filename"),
                  llvm::cl::value_desc("bclib"));
 
 llvm::cl::opt<std::string>
+OptBCLibRelaxedFilename("bclib_relaxed", llvm::cl::desc("Specify the bclib filename optimized for "
+                                                        "relaxed precision floating point maths"),
+                        llvm::cl::init(""),
+                        llvm::cl::value_desc("bclib_relaxed"));
+
+llvm::cl::opt<std::string>
 OptOutputPath("output_path", llvm::cl::desc("Specify the output path"),
               llvm::cl::value_desc("output path"),
               llvm::cl::init("."));
@@ -195,7 +201,8 @@ bool compileScriptGroup(BCCContext& Context, RSCompilerDriver& RSCD) {
   outputFilepath.append(OptOutputFilename);
 
   bool success = RSCD.buildScriptGroup(
-    Context, outputFilepath.c_str(), OptBCLibFilename.c_str(), true,
+    Context, outputFilepath.c_str(), OptBCLibFilename.c_str(),
+    OptBCLibRelaxedFilename.c_str(), true,
     sources, sourcesAndSlots, fusedKernelNames,
     invokeSourcesAndSlots, invokeBatchNames);