P4 to Git Change 1568521 by gandryey@gera-w8 on 2018/06/14 17:43:52

SWDEV-79445 - OCL generic changes and code clean-up
	- Change LDS setup to account the size, since LC forces 4 bytes for LDS offsets always

	http://ocltc.amd.com/reviews/r/15197/

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_execute.cpp#28 edit
... //depot/stg/opencl/drivers/opencl/api/opencl/amdocl/cl_program.cpp#49 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpublit.cpp#130 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpukernel.cpp#327 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palblit.cpp#25 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palkernel.cpp#56 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#109 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocblit.hpp#11 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rockernel.cpp#38 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.cpp#88 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/kernel.cpp#34 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/kernel.hpp#25 edit

rocclr/runtime/device/gpu/gpublit.cpp

@@ -955,13 +955,14 @@ static void setArgument(amd::Kernel* kernel, size_t index, size_t size, const vo
   const amd::KernelParameterDescriptor& desc = kernel->signature().at(index);
 
   void* param = kernel->parameters().values() + desc.offset_;
-  assert((desc.type_ == T_POINTER || value != NULL || desc.size_ == 0) &&
-         "not a valid local mem arg");
+  assert((desc.type_ == T_POINTER || value != NULL ||
+    (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL)) &&
+    "not a valid local mem arg");
 
   uint32_t uint32_value = 0;
   uint64_t uint64_value = 0;
 
-  if (desc.type_ == T_POINTER && desc.size_ != 0) {
+  if (desc.type_ == T_POINTER && (desc.addressQualifier_ != CL_KERNEL_ARG_ADDRESS_LOCAL)) {
     if ((value == NULL) || (static_cast<const cl_mem*>(value) == NULL)) {
       LP64_SWITCH(uint32_value, uint64_value) = 0;
       reinterpret_cast<Memory**>(kernel->parameters().values() +
@@ -978,26 +979,24 @@ static void setArgument(amd::Kernel* kernel, size_t index, size_t size, const vo
     assert(false && "No sampler support in blit manager! Use internal samplers!");
   } else
     switch (desc.size_) {
-      case 1:
-        uint32_value = *static_cast<const uint8_t*>(value);
-        break;
-      case 2:
-        uint32_value = *static_cast<const uint16_t*>(value);
-        break;
       case 4:
-        uint32_value = *static_cast<const uint32_t*>(value);
+        if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
+          uint32_value = size;
+        } else {
+          uint32_value = *static_cast<const uint32_t*>(value);
+        }
         break;
       case 8:
-        uint64_value = *static_cast<const uint64_t*>(value);
+        if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
+          uint64_value = size;
+        } else {
+          uint64_value = *static_cast<const uint64_t*>(value);
+        }
         break;
       default:
         break;
     }
-
   switch (desc.size_) {
-    case 0 /*local mem*/:
-      *static_cast<size_t*>(param) = size;
-      break;
     case sizeof(uint32_t):
       *static_cast<uint32_t*>(param) = uint32_value;
       break;

rocclr/runtime/device/gpu/gpukernel.cpp

@@ -378,7 +378,7 @@ size_t KernelArg::size(bool gpuLayer) const {
       return (gpuLayer) ? 0 : sizeof(cl_mem);
     case PointerLocal:
     case PointerHwLocal:
-      return (gpuLayer) ? sizeof(uint32_t) * size_ : 0;
+      return (gpuLayer) ? sizeof(uint32_t) * size_ : sizeof(cl_mem);
     case PointerPrivate:
     case PointerHwPrivate:
       return (gpuLayer) ? sizeof(uint32_t) * size_ : 0;
@@ -2991,7 +2991,7 @@ void HSAILKernel::initArgList(const aclArgData* aclArg) {
 
     // Make a check if it is local or global
     if (desc.addressQualifier_ == CL_KERNEL_ARG_ADDRESS_LOCAL) {
-      desc.size_ = 0;
+      desc.size_ = sizeof(cl_mem);
     } else {
       desc.size_ = GetOclSize(aclArg);
     }
@@ -3000,10 +3000,7 @@ void HSAILKernel::initArgList(const aclArgData* aclArg) {
     // in multidevice config abstraction layer has a single signature
     // and CPU sends the paramaters as they are allocated in memory
     size_t size = desc.size_;
-    if (size == 0) {
-      // Local memory for CPU
-      size = sizeof(cl_mem);
-    }
+
     offset = amd::alignUp(offset, std::min(size, size_t(16)));
     desc.offset_ = offset;
     offset += amd::alignUp(size, sizeof(uint32_t));
@@ -3527,8 +3524,12 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(
         else {
           assert((arg->addrQual_ == HSAIL_ADDRESS_LOCAL) && "Unsupported address type");
           ldsAddress = amd::alignUp(ldsAddress, arg->alignment_);
-          WriteAqlArg(&aqlArgBuf, &ldsAddress, sizeof(size_t));
-          ldsAddress += *reinterpret_cast<const size_t*>(paramaddr);
+          WriteAqlArg(&aqlArgBuf, &ldsAddress, desc.size_);
+          if (desc.size_ == 8) {
+            ldsAddress += *reinterpret_cast<const uint64_t*>(paramaddr);
+          } else {
+            ldsAddress += *reinterpret_cast<const uint32_t*>(paramaddr);
+          }
         }
         break;
       case HSAIL_ARGTYPE_VALUE: