Fail build program in shared system USM + statefull access case OCL

Signed-off-by: Kamil Kopryk <kamil.kopryk@intel.com>
Related-To: NEO-6075

opencl/source/program/build.cpp

@@ -34,8 +34,7 @@ cl_int Program::build(
     const char *buildOptions,
     bool enableCaching) {
     cl_int retVal = CL_SUCCESS;
-    std::string internalOptions;
-    initInternalOptions(internalOptions);
+
     auto defaultClDevice = deviceVector[0];
     UNRECOVERABLE_IF(defaultClDevice == nullptr);
     auto &defaultDevice = defaultClDevice->getDevice();
@@ -69,6 +68,9 @@ cl_int Program::build(
             } else if (this->createdFrom != CreatedFrom::BINARY) {
                 options = "";
             }
+            std::string internalOptions;
+            initInternalOptions(internalOptions);
+
             extractInternalOptions(options, internalOptions);
             applyAdditionalOptions(internalOptions);
 
@@ -166,6 +168,10 @@ cl_int Program::build(
             phaseReached[clDevice->getRootDeviceIndex()] = BuildPhase::BinaryProcessing;
         }
 
+        if (containsStatefulAccess(defaultDevice.getRootDeviceIndex()) && forceToStatelessNeeded() && !isBuiltIn) {
+            retVal = CL_BUILD_PROGRAM_FAILURE;
+        }
+
         if (retVal != CL_SUCCESS) {
             break;
         }

opencl/source/program/program.cpp

@@ -66,6 +66,7 @@ Program::Program(Context *context, bool isBuiltIn, const ClDeviceVector &clDevic
     kernelDebugEnabled = clDevices[0]->isDebuggerActive();
 }
 void Program::initInternalOptions(std::string &internalOptions) const {
+
     auto pClDevice = clDevices[0];
     auto force32BitAddressess = pClDevice->getSharedDeviceInfo().force32BitAddressess;
     internalOptions = getOclVersionCompilerInternalOption(pClDevice->getEnabledClVersion());
@@ -74,7 +75,7 @@ void Program::initInternalOptions(std::string &internalOptions) const {
         CompilerOptions::concatenateAppend(internalOptions, CompilerOptions::arch32bit);
     }
 
-    if ((isBuiltIn && is32bit) || pClDevice->areSharedSystemAllocationsAllowed() ||
+    if ((isBuiltIn && is32bit) || forceToStatelessNeeded() ||
         DebugManager.flags.DisableStatelessToStatefulOptimization.get()) {
         CompilerOptions::concatenateAppend(internalOptions, CompilerOptions::greaterThan4gbBuffersRequired);
     }
@@ -136,6 +137,20 @@ Program::~Program() {
     }
 }
 
+bool Program::forceToStatelessNeeded() const {
+    auto preferStateful = false;
+    if (auto it = options.find(NEO::CompilerOptions::smallerThan4gbBuffersOnly.data()); it != std::string::npos) {
+        preferStateful = true;
+    }
+
+    if (DebugManager.flags.UseSmallerThan4gbBuffersOnly.get() != -1) {
+        preferStateful = static_cast<bool>(DebugManager.flags.UseSmallerThan4gbBuffersOnly.get());
+    }
+
+    auto forceStateless = !preferStateful && clDevices[0]->areSharedSystemAllocationsAllowed();
+    return forceStateless;
+}
+
 cl_int Program::createProgramFromBinary(
     const void *pBinary,
     size_t binarySize, ClDevice &clDevice) {
@@ -489,6 +504,22 @@ cl_int Program::packDeviceBinary(ClDevice &clDevice) {
     return CL_SUCCESS;
 }
 
+bool Program::containsStatefulAccess(uint32_t rootDeviceIndex) const {
+    auto &buildInfo = buildInfos[rootDeviceIndex];
+    for (const auto &kernelInfo : buildInfo.kernelInfoArray) {
+        for (const auto &arg : kernelInfo->kernelDescriptor.payloadMappings.explicitArgs) {
+            auto isStatefulAccess = arg.is<ArgDescriptor::ArgTPointer>() &&
+                                    (isValidOffset(arg.as<ArgDescPointer>().bindless) ||
+                                     isValidOffset(arg.as<ArgDescPointer>().bindful));
+            if (isStatefulAccess) {
+                return true;
+            }
+        }
+    }
+
+    return false;
+}
+
 void Program::setBuildStatus(cl_build_status status) {
     for (auto &deviceBuildInfo : deviceBuildInfos) {
         deviceBuildInfo.second.buildStatus = status;

opencl/source/program/program.h

@@ -283,6 +283,10 @@ class Program : public BaseObject<_cl_program> {
     }
 
   protected:
+    bool forceToStatelessNeeded() const;
+
+    MOCKABLE_VIRTUAL bool containsStatefulAccess(uint32_t rootDeviceIndex) const;
+
     MOCKABLE_VIRTUAL cl_int createProgramFromBinary(const void *pBinary, size_t binarySize, ClDevice &clDevice);
 
     cl_int packDeviceBinary(ClDevice &clDevice);
@@ -366,6 +370,7 @@ class Program : public BaseObject<_cl_program> {
 
     bool isBuiltIn = false;
     bool kernelDebugEnabled = false;
+    bool containsStatefulAccesses = false;
     uint32_t maxRootDeviceIndex = std::numeric_limits<uint32_t>::max();
     std::mutex lockMutex;
     uint32_t exposedKernels = 0;