Automatic detection of SPMD, ESIMD and SPMD+ESIMD SPIR-V modules

PawelJurek · igcbot · commit 316afc59b94e · 2021-10-27T16:41:16.000+02:00
Before this change there were 3 paths of SPIR-V compilation:
1. If -vc-codegen option was passed, further translation was passed to VC codegen
2. If -spmd_esimd_compile was passed, vISA Linker Driver was invoked to split
the module and dispatch the separate SPIR-V modules to SPMD and ESIMD (VC) backends.
3. If no option was provided, SPMD backend was invoked.

This change introduces automatic ESIMD module detection - user can now pass
ESIMD module without providing -vc-codegen option.
Current flow is:
1. If no option is provided, detection algorithm is invoked and the SPIR-V
module is directed to detected backend.
2. If -vc-codegen option is provided, VC backend is chosen without invoking the
detection code.
3. There is no -spmd_esimd_compile option, so SPMD+ESIMD path is chosen after
detection.

Currently module is marked as ESIMD only, if there is an entry point marked with
VectorComputeFunctionINTEL decoration. This is forbidden in SPMD module and
SPMD+ESIMD module.
diff --git a/IGC/AdaptorOCL/dllInterfaceCompute.cpp b/IGC/AdaptorOCL/dllInterfaceCompute.cpp
@@ -919,53 +919,12 @@ static void WriteSpecConstantsDump(const STB_TranslateInputArgs *pInputArgs,
                    hash, "_specconst.txt");
 }
 
-bool TranslateBuild(
-    const STB_TranslateInputArgs* pInputArgs,
-    STB_TranslateOutputArgs* pOutputArgs,
-    TB_DATA_FORMAT inputDataFormatTemp,
-    const IGC::CPlatform& IGCPlatform,
-    float profilingTimerResolution)
-{
-    ShaderHash inputShHash = ShaderHashOCL(reinterpret_cast<const UINT *>(pInputArgs->pInput),
-                                           pInputArgs->InputSize / 4);
-
-    // on wrong spec constants, vc::translateBuild may fail
-    // so lets dump those early
-    if (pInputArgs->SpecConstantsSize > 0 &&
-        IGC_IS_FLAG_ENABLED(ShaderDumpEnable))
-      WriteSpecConstantsDump(pInputArgs, inputShHash.getAsmHash());
-
-#if defined(IGC_VC_ENABLED)
-    if (pInputArgs->pOptions) {
-        // Currently, VC compiler effectively uses global variables to store
-        // some configuration information. This may lead to problems
-        // during multi-threaded compilations. The mutex below serializes
-        // the whole compilation process.
-        // This is a temporary measure till a proper re-design is done.
-        const std::lock_guard<std::mutex> lock(llvm_mutex);
-
-        std::error_code Status =
-            vc::translateBuild(pInputArgs, pOutputArgs, inputDataFormatTemp,
-                               IGCPlatform, profilingTimerResolution);
-        if (!Status)
-            return true;
-        // If vc codegen option was not specified, then vc was not called.
-        if (static_cast<vc::errc>(Status.value()) != vc::errc::not_vc_codegen)
-            return false;
-    }
-#endif // defined(IGC_VC_ENABLED)
-    if (pInputArgs->pOptions &&
-        strstr(pInputArgs->pOptions,
-               VLD::VLD_compilation_enable_option)) {
-        std::string errorMessage;
-        bool ret = VLD::TranslateBuildSPMDAndESIMD(
-            pInputArgs, pOutputArgs, inputDataFormatTemp, IGCPlatform,
-            profilingTimerResolution, inputShHash, errorMessage);
-        if (!ret) {
-             SetErrorMessage(errorMessage, *pOutputArgs);
-        }
-        return ret;
-  }
+bool TranslateBuildSPMD(const STB_TranslateInputArgs *pInputArgs,
+                        STB_TranslateOutputArgs *pOutputArgs,
+                        TB_DATA_FORMAT inputDataFormatTemp,
+                        const IGC::CPlatform &IGCPlatform,
+                        float profilingTimerResolution,
+                        const ShaderHash& inputShHash) {
 
     // This part of code is a critical-section for threads,
     // due static LLVM object which handles options.
@@ -1372,6 +1331,80 @@ bool TranslateBuild(
     return true;
 }
 
+
+#if defined(IGC_VC_ENABLED)
+bool TranslateBuildESIMD(
+    const STB_TranslateInputArgs* pInputArgs,
+    STB_TranslateOutputArgs* pOutputArgs,
+    TB_DATA_FORMAT inputDataFormatTemp,
+    const IGC::CPlatform& IGCPlatform,
+    float profilingTimerResolution,
+    const ShaderHash& inputShHash) {
+
+    IGC_ASSERT(pInputArgs->pOptions &&
+             (strstr(pInputArgs->pOptions, "-vc-codegen") ||
+              strstr(pInputArgs->pOptions, "-cmc")));
+
+  // Currently, VC compiler effectively uses global variables to store
+  // some configuration information. This may lead to problems
+  // during multi-threaded compilations. The mutex below serializes
+  // the whole compilation process.
+  // This is a temporary measure till a proper re-design is done.
+  const std::lock_guard<std::mutex> lock(llvm_mutex);
+
+  std::error_code Status =
+      vc::translateBuild(pInputArgs, pOutputArgs, inputDataFormatTemp,
+                         IGCPlatform, profilingTimerResolution);
+  if (!Status)
+    return true;
+  return false;
+}
+#endif // defined(IGC_VC_ENABLED)
+
+bool TranslateBuild(
+    const STB_TranslateInputArgs* pInputArgs,
+    STB_TranslateOutputArgs* pOutputArgs,
+    TB_DATA_FORMAT inputDataFormatTemp,
+    const IGC::CPlatform& IGCPlatform,
+    float profilingTimerResolution)
+{
+    ShaderHash inputShHash = ShaderHashOCL(reinterpret_cast<const UINT *>(pInputArgs->pInput),
+                                           pInputArgs->InputSize / 4);
+
+    // on wrong spec constants, vc::translateBuild may fail
+    // so lets dump those early
+    if (pInputArgs->SpecConstantsSize > 0 &&
+        IGC_IS_FLAG_ENABLED(ShaderDumpEnable))
+      WriteSpecConstantsDump(pInputArgs, inputShHash.getAsmHash());
+
+#if defined(IGC_VC_ENABLED)
+    // if VC option was specified, go to ESIMD compilation directly.
+    if (pInputArgs->pOptions && (strstr(pInputArgs->pOptions, "-vc-codegen") ||
+                                 strstr(pInputArgs->pOptions, "-cmc"))) {
+      return TranslateBuildESIMD(pInputArgs, pOutputArgs, inputDataFormatTemp,
+                                 IGCPlatform, profilingTimerResolution,
+                                 inputShHash);
+    }
+#endif // defined(IGC_VC_ENABLED)
+
+    if (inputDataFormatTemp != TB_DATA_FORMAT_SPIR_V) {
+      return TranslateBuildSPMD(pInputArgs, pOutputArgs, inputDataFormatTemp,
+                                IGCPlatform, profilingTimerResolution,
+                                inputShHash);
+    }
+
+    // Recognize if SPIR-V module contains SPMD,ESIMD or SPMD+ESIMD code and compile it.
+    std::string errorMessage;
+    bool ret = VLD::TranslateBuildSPMDAndESIMD(
+        pInputArgs, pOutputArgs, inputDataFormatTemp, IGCPlatform,
+        profilingTimerResolution, inputShHash, errorMessage);
+    if (!ret) {
+      SetErrorMessage(errorMessage, *pOutputArgs);
+    }
+    return ret;
+
+}
+
 bool CIGCTranslationBlock::FreeAllocations(
     STB_TranslateOutputArgs* pOutputArgs)
 {
diff --git a/IGC/AdaptorOCL/ocl_igc_interface/impl/igc_ocl_translation_ctx_impl.h b/IGC/AdaptorOCL/ocl_igc_interface/impl/igc_ocl_translation_ctx_impl.h
@@ -47,6 +47,22 @@ bool TranslateBuild(
   const IGC::CPlatform &platform,
   float profilingTimerResolution);
 
+bool TranslateBuildSPMD(
+  const STB_TranslateInputArgs* pInputArgs,
+  STB_TranslateOutputArgs* pOutputArgs,
+  TB_DATA_FORMAT inputDataFormatTemp,
+  const IGC::CPlatform &platform,
+  float profilingTimerResolution,
+  const ShaderHash& inputShHash);
+
+bool TranslateBuildESIMD(
+  const STB_TranslateInputArgs* pInputArgs,
+  STB_TranslateOutputArgs* pOutputArgs,
+  TB_DATA_FORMAT inputDataFormatTemp,
+  const IGC::CPlatform &platform,
+  float profilingTimerResolution,
+  const ShaderHash& inputShHash);
+
 bool ReadSpecConstantsFromSPIRV(
     std::istream &IS,
     std::vector<std::pair<uint32_t, uint32_t>> &OutSCInfo);
diff --git a/IGC/VISALinkerDriver/VLD.cpp b/IGC/VISALinkerDriver/VLD.cpp
@@ -74,21 +74,24 @@ namespace VLD {
   using namespace TC;
 
 
-// Translates SPIR-V module that consists of SPMD code that invokes ESIMD code.
+// Translates ESIMD and SPMD code in the module.
+// 3 cases are handled:
+// 1. only SPMD code is present
+// 2. only ESIMD code is present
+// 3. ESIMD code is invoked from SPMD code
+//
 // The general flow is:
 // 1. Split input SPIR-V module into SPMD and ESIMD parts
 // 2. Invoke SPMD and ESIMD backends with appropriate SPIR-V modules
-// 3. Extract .visaasm from the output zeBinary
+// 3. If SPMD code invokes ESIMD code, extract .visaasm from the each output zeBinary
 // TODO: 4. Link .visaasm files via vISA interfaces
 //
 // The function signature corresponds to TC::TranslateBuild interface, so that
 // it is easy to pass same arguments to SPMD and VC backends.
 //
 // Assumptions:
-// 1. ZEBinary output format is used.
+// 1. ZEBinary output format is used in SPMD+ESIMD case.
 // TODO: error out if patch token output format is used.
-// 2. Input is in SPIR-V format
-// TODO: error out if other input format is used.
 bool TranslateBuildSPMDAndESIMD(const TC::STB_TranslateInputArgs *pInputArgs,
                                   TC::STB_TranslateOutputArgs *pOutputArgs,
                                   TC::TB_DATA_FORMAT inputDataFormatTemp,
@@ -97,19 +100,54 @@ bool TranslateBuildSPMDAndESIMD(const TC::STB_TranslateInputArgs *pInputArgs,
                                   const ShaderHash &inputShHash,
                                   std::string& errorMessage) {
 
+  IGC_ASSERT(inputDataFormatTemp == TB_DATA_FORMAT_SPIR_V);
+
     // Split ESIMD and SPMD code.
   auto spmd_esimd_programs_or_err = VLD::SplitSPMDAndESIMD(
       pInputArgs->pInput, pInputArgs->InputSize);
 
   if (!spmd_esimd_programs_or_err) {
-    // Caller releases the error string, so we need to make a copy of the error message here.
-        // TODO: pOutputArgs contains field for error string so we can copy it there.
-    // Not done now, as it would require copy-paste code that is avaiable in dllinterfacecompute. Needs to be refactored.
-    errorMessage = "Error while splitting ESIMD and SPMD code: " +
-      llvm::toString(spmd_esimd_programs_or_err.takeError());
-    return false;
+      // Workaround: try to compile on SPMD path if splitting failed.
+      // This is because not all VC opcodes are merged to SPIR-V Tools.
+      return TranslateBuildSPMD(pInputArgs, pOutputArgs, inputDataFormatTemp,
+          IGCPlatform, profilingTimerResolution,
+          inputShHash);
+
+      // TODO: uncomment once above workaround is removed.
+      // Caller releases the error string, so we need to make a copy of the error message here.
+      // TODO: pOutputArgs contains field for error string so we can copy it there.
+      // Not done now, as it would require copy-paste code that is avaiable in dllinterfacecompute. Needs to be refactored.
+      // errorMessage = llvm::toString(spmd_esimd_programs_or_err.takeError());
+      // return false;
   }
 
+  std::string newOptions{pInputArgs->pOptions ? pInputArgs->pOptions : ""};
+  std::string esimdOptions{ newOptions };
+  esimdOptions += " -vc-codegen";
+
+  IGC_ASSERT(!spmd_esimd_programs_or_err->first.empty() || !spmd_esimd_programs_or_err->second.empty());
+  if (spmd_esimd_programs_or_err->first.empty()) {
+#if defined(IGC_VC_ENABLED)
+      // Only ESIMD code detected.
+      STB_TranslateInputArgs newArgs = *pInputArgs;
+      newArgs.pOptions = esimdOptions.data();
+      newArgs.OptionsSize = esimdOptions.size();
+      return TranslateBuildESIMD(&newArgs, pOutputArgs, inputDataFormatTemp,
+                                 IGCPlatform, profilingTimerResolution,
+                                 inputShHash);
+#else // defined(IGC_VC_ENABLED)
+      errorMessage = "ESIMD code detected, but VC not enabled in this build.";
+      return false;
+#endif // defined(IGC_VC_ENABLED)
+  } else if (spmd_esimd_programs_or_err->second.empty()) {
+      // Only SPMD code detected.
+      return TranslateBuildSPMD(pInputArgs, pOutputArgs, inputDataFormatTemp,
+          IGCPlatform, profilingTimerResolution,
+          inputShHash);
+  }
+
+  // SPMD+ESIMD code detected.
+
   if (IGC_IS_FLAG_ENABLED(ShaderDumpEnable)) {
     const char* pOutputFolder = IGC::Debug::GetShaderOutputFolder();
 
@@ -134,27 +172,24 @@ bool TranslateBuildSPMDAndESIMD(const TC::STB_TranslateInputArgs *pInputArgs,
     dump((*spmd_esimd_programs_or_err).second, ".esimd_split.spv");
   }
 
-  std::string newOptions{pInputArgs->pOptions};
-  IGC_ASSERT(newOptions.find(VLD_compilation_enable_option) !=
-             std::string::npos);
-  newOptions.erase(newOptions.find(VLD_compilation_enable_option),
-                   strnlen(VLD_compilation_enable_option, sizeof(VLD_compilation_enable_option)));
-
   auto translateToVISA =
       [&](VLD::ProgramStreamType& program,
-          const char *newOptions) -> decltype(GetVISAAsmFromZEBinary(0,0)) {
+          const std::string& newOptions,
+          decltype(TranslateBuildSPMD) TranslateFunction
+          ) -> decltype(GetVISAAsmFromZEBinary(0,0)) {
         STB_TranslateInputArgs newArgs = *pInputArgs;
 
         TC::STB_TranslateOutputArgs outputArgs;
         CIF::SafeZeroOut(outputArgs);
 
         newArgs.pInput = reinterpret_cast<char*>(program.data());
         newArgs.InputSize = program.size() * sizeof(*program.begin());;
-        newArgs.pOptions = newOptions;
+        newArgs.pOptions = newOptions.data();
+        newArgs.OptionsSize = newOptions.size();
 
         const bool success =
-            TranslateBuild(&newArgs, &outputArgs, inputDataFormatTemp,
-                           IGCPlatform, profilingTimerResolution);
+            TranslateFunction(&newArgs, &outputArgs, inputDataFormatTemp,
+                           IGCPlatform, profilingTimerResolution, inputShHash);
 
         auto outputData = std::unique_ptr<char[]>(outputArgs.pOutput);
         auto errorString = std::unique_ptr<char[]>(outputArgs.pErrorString);
@@ -170,20 +205,22 @@ bool TranslateBuildSPMDAndESIMD(const TC::STB_TranslateInputArgs *pInputArgs,
         return spmdVISAAsmVector;
   };
 
-  std::string esimdOptions{ newOptions };
-  esimdOptions += " -vc-codegen";
-
+#if defined(IGC_VC_ENABLED)
   auto esimdVISA =
-      translateToVISA(spmd_esimd_programs_or_err->second, esimdOptions.data());
+      translateToVISA(spmd_esimd_programs_or_err->second, esimdOptions, TranslateBuildESIMD);
 
   if (!esimdVISA) {
     errorMessage = "VLD: Failed to compile ESIMD part with following error: \n" +
                    llvm::toString(esimdVISA.takeError());
       return false;
   }
+#else // defined(IGC_VC_ENABLED)
+    errorMessage = "Could not compile ESIMD part of SPIR-V module, as VC is not included in this build.";
+    return false;
+#endif // defined(IGC_VC_ENABLED)
 
   auto spmdVISA =
-             translateToVISA(spmd_esimd_programs_or_err->first, newOptions.data());
+             translateToVISA(spmd_esimd_programs_or_err->first, newOptions.data(), TranslateBuildSPMD);
 
   if (!spmdVISA) {
     errorMessage = "VLD: Failed to compile SPMD part with following error: \n" +
diff --git a/IGC/VISALinkerDriver/VLD.hpp b/IGC/VISALinkerDriver/VLD.hpp
@@ -21,9 +21,6 @@ namespace IGC {
 namespace IGC {
 namespace VLD {
 
-  // Option that must be passed to enable VLD path
-  static const char VLD_compilation_enable_option[] = "-spmd_esimd_compile";
-
 bool TranslateBuildSPMDAndESIMD(const TC::STB_TranslateInputArgs* pInputArgs,
                                 TC::STB_TranslateOutputArgs* pOutputArgs,
                                 TC::TB_DATA_FORMAT inputDataFormatTemp,
diff --git a/IGC/VISALinkerDriver/VLD_SPIRVSplitter.cpp b/IGC/VISALinkerDriver/VLD_SPIRVSplitter.cpp
diff --git a/IGC/VISALinkerDriver/VLD_SPIRVSplitter.hpp b/IGC/VISALinkerDriver/VLD_SPIRVSplitter.hpp
