P4 to Git Change 1472384 by lmoriche@lmoriche_opencl_dev2 on 2017/10/19 03:29:10

SWDEV-126884 - OCL 2.1 Platform APIs
	- Fix OCL/ROCm OpenCL C version

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpucompiler.cpp#157 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudevice.cpp#578 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuresource.cpp#240 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#65 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocdevice.cpp#70 edit
... //depot/stg/opencl/drivers/opencl/runtime/runtimedefs#46 edit


[ROCm/clr commit: 46dcd95add]

projects/clr/rocclr/runtime/device/gpu/gpucompiler.cpp

@@ -138,6 +138,7 @@ bool NullProgram::compileImpl(const std::string& src,
         opts << " -cl-std=CL1.1";
         break;
       case OpenCL20:
+      case OpenCL21:
       default:
       case OpenCL12:
         opts << " -cl-std=CL1.2";

projects/clr/rocclr/runtime/device/gpu/gpudevice.cpp

@@ -216,7 +216,7 @@ bool NullDevice::create(CALtarget target) {
     compiler_ = aclCompilerInit(&opts, NULL);
   }
 
-  if (settings().hsail_ || (settings().oclVersion_ == OpenCL20)) {
+  if (settings().hsail_ || (settings().oclVersion_ >= OpenCL20)) {
     // Runtime doesn't know what local size could be on the real board
     info_.maxGlobalVariableSize_ = static_cast<size_t>(512 * Mi);
 
@@ -487,7 +487,7 @@ void NullDevice::fillDeviceInfo(const CALdeviceattribs& calAttr, const gslMemInf
   ::snprintf(info_.driverVersion_, sizeof(info_.driverVersion_) - 1, AMD_BUILD_STRING);
 
   info_.profile_ = "FULL_PROFILE";
-  if (settings().oclVersion_ == OpenCL20) {
+  if (settings().oclVersion_ >= OpenCL20) {
     info_.version_ = "OpenCL 2.0 " AMD_PLATFORM_INFO;
     info_.oclcVersion_ = "OpenCL C 2.0 ";
     info_.spirVersions_ = "1.2";
@@ -958,7 +958,7 @@ bool Device::create(CALuint ordinal, CALuint numOfDevices) {
     compiler_ = aclCompilerInit(&opts, NULL);
   }
 
-  if (settings().hsail_ || (settings().oclVersion_ == OpenCL20)) {
+  if (settings().hsail_ || (settings().oclVersion_ >= OpenCL20)) {
     if (NULL == hsaCompiler_) {
       const char* library = getenv("HSA_COMPILER_LIBRARY");
       aclCompilerOptions opts = {
@@ -1068,7 +1068,7 @@ bool Device::initializeHeapResources() {
     if (settings().ciPlus_) {
       const char* CL20extraBlits = NULL;
       const char* ocl20 = NULL;
-      if (settings().oclVersion_ == OpenCL20) {
+      if (settings().oclVersion_ >= OpenCL20) {
         CL20extraBlits = SchedulerSourceCode;
         ocl20 = "-cl-std=CL2.0";
       }

projects/clr/rocclr/runtime/device/gpu/gpuresource.cpp

@@ -924,7 +924,7 @@ bool Resource::create(MemoryType memType, CreateParams* params) {
     }
   }
 
-  if ((dev().settings().hsail_ || (dev().settings().oclVersion_ == OpenCL20)) && !cal()->buffer_) {
+  if ((dev().settings().hsail_ || (dev().settings().oclVersion_ >= OpenCL20)) && !cal()->buffer_) {
     hwSrd_ = dev().srds().allocSrdSlot(reinterpret_cast<address*>(&hwState_));
     if (0 == hwSrd_) {
       return false;
@@ -1032,7 +1032,7 @@ void Resource::free() {
   }
 
   // Free SRD for images
-  if ((dev().settings().hsail_ || (dev().settings().oclVersion_ == OpenCL20)) && !cal()->buffer_) {
+  if ((dev().settings().hsail_ || (dev().settings().oclVersion_ >= OpenCL20)) && !cal()->buffer_) {
     dev().srds().freeSrdSlot(hwSrd_);
   }
 }

projects/clr/rocclr/runtime/device/pal/paldevice.cpp

@@ -439,7 +439,7 @@ void NullDevice::fillDeviceInfo(const Pal::DeviceProperties& palProp,
              );
 
   info_.profile_ = "FULL_PROFILE";
-  if (settings().oclVersion_ == OpenCL20) {
+  if (settings().oclVersion_ >= OpenCL20) {
     info_.version_ = "OpenCL 2.0 " AMD_PLATFORM_INFO;
     info_.oclcVersion_ = "OpenCL C 2.0 ";
     info_.spirVersions_ = "1.2";
@@ -2151,7 +2151,7 @@ bool Device::createBlitProgram() {
   const char* ocl20 = nullptr;
 #if !defined(WITH_LIGHTNING_COMPILER)
   std::string sch = SchedulerSourceCode;
-  if (settings().oclVersion_ == OpenCL20) {
+  if (settings().oclVersion_ >= OpenCL20) {
     size_t loc = sch.find("%s");
     sch.replace(loc, 2, iDev()->GetDispatchKernelSource());
     scheduler = sch.c_str();

projects/clr/rocclr/runtime/device/rocm/rocdevice.cpp

@@ -37,6 +37,7 @@
 #endif  // WITHOUT_HSA_BACKEND
 
 #define OPENCL_VERSION_STR XSTR(OPENCL_MAJOR) "." XSTR(OPENCL_MINOR)
+#define OPENCL_C_VERSION_STR XSTR(OPENCL_C_MAJOR) "." XSTR(OPENCL_C_MINOR)
 
 #ifndef WITHOUT_HSA_BACKEND
 namespace device {
@@ -111,7 +112,7 @@ bool NullDevice::create(const AMDDeviceInfo& deviceInfo) {
   info_.extensions_ = getExtensionString();
   info_.maxWorkGroupSize_ = hsaSettings->maxWorkGroupSize_;
   ::strcpy(info_.vendor_, "Advanced Micro Devices, Inc.");
-  info_.oclcVersion_ = "OpenCL C " IF(IS_LIGHTNING, OPENCL_VERSION_STR, "1.2") " ";
+  info_.oclcVersion_ = "OpenCL C " IF(IS_LIGHTNING, OPENCL_C_VERSION_STR, "1.2") " ";
   strcpy(info_.driverVersion_, "1.0 Provisional (hsa)");
   info_.version_ = "OpenCL " OPENCL_VERSION_STR " ";
   return true;
@@ -952,7 +953,7 @@ bool Device::populateOCLDeviceConstants() {
   info_.addressBits_ = LP64_SWITCH(32, 64);
   info_.maxSamplers_ = 16;
   info_.bufferFromImageSupport_ = CL_FALSE;
-  info_.oclcVersion_ = "OpenCL C " IF(IS_LIGHTNING, OPENCL_VERSION_STR, "1.2") " ";
+  info_.oclcVersion_ = "OpenCL C " IF(IS_LIGHTNING, OPENCL_C_VERSION_STR, "1.2") " ";
 
   uint16_t major, minor;
   if (hsa_agent_get_info(_bkendDevice, HSA_AGENT_INFO_VERSION_MAJOR, &major) !=