From 2be0b1e612252b3ee44153e6e7c923af72c81170 Mon Sep 17 00:00:00 2001
From: German Andryeyev <German.Andryeyev@amd.com>
Date: Thu, 7 Apr 2022 17:41:43 -0400
Subject: [PATCH] SWDEV-307184 - Report 1 for unused dimensions

Remove assert for kernel arg size, because COv5 reports a value
bigger than the actual usage in the most of cases

Change-Id: I8e15bc45a9e21b58a5894f9977511ca84408ce61
---
 rocclr/device/pal/palkernel.cpp   | 15 +++++++++------
 rocclr/device/rocm/rocvirtual.cpp | 13 ++++++++-----
 2 files changed, 17 insertions(+), 11 deletions(-)

diff --git a/rocclr/device/pal/palkernel.cpp b/rocclr/device/pal/palkernel.cpp
index 9c4729c71f..e39f87b371 100644
--- a/rocclr/device/pal/palkernel.cpp
+++ b/rocclr/device/pal/palkernel.cpp
@@ -374,12 +374,16 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(VirtualGPU& gpu, const
         if (sizes.dimensions() >= 2) {
           WriteAqlArgAt(hidden_arguments, static_cast<uint32_t>(global[1] / local[1]),
                         it.size_, it.offset_);
+        } else {
+          WriteAqlArgAt(hidden_arguments, static_cast<uint32_t>(1), it.size_, it.offset_);
         }
         break;
       case amd::KernelParameterDescriptor::HiddenBlockCountZ:
         if (sizes.dimensions() >= 3) {
           WriteAqlArgAt(hidden_arguments, static_cast<uint32_t>(global[2] / local[2]),
                         it.size_, it.offset_);
+        } else {
+          WriteAqlArgAt(hidden_arguments, static_cast<uint32_t>(1), it.size_, it.offset_);
         }
         break;
       case amd::KernelParameterDescriptor::HiddenGroupSizeX:
@@ -388,11 +392,15 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(VirtualGPU& gpu, const
       case amd::KernelParameterDescriptor::HiddenGroupSizeY:
         if (sizes.dimensions() >= 2) {
           WriteAqlArgAt(hidden_arguments, static_cast<uint16_t>(local[1]), it.size_, it.offset_);
+        } else {
+          WriteAqlArgAt(hidden_arguments, static_cast<uint16_t>(1), it.size_, it.offset_);
         }
-        break;
+      break;
       case amd::KernelParameterDescriptor::HiddenGroupSizeZ:
         if (sizes.dimensions() >= 3) {
           WriteAqlArgAt(hidden_arguments, static_cast<uint16_t>(local[2]), it.size_, it.offset_);
+        } else {
+          WriteAqlArgAt(hidden_arguments, static_cast<uint16_t>(1), it.size_, it.offset_);
         }
         break;
       case amd::KernelParameterDescriptor::HiddenRemainderX:
@@ -437,11 +445,6 @@ hsa_kernel_dispatch_packet_t* HSAILKernel::loadArguments(VirtualGPU& gpu, const
     memcpy(aqlArgBuf, parameters, argsBufferSize());
   }
 
-  // Note: In a case of structs the size won't match,
-  // since HSAIL compiler expects a reference...
-  assert(argsBufferSize() <= signature.paramsSize() &&
-         "A mismatch of sizes of arguments between compiler and runtime!");
-
   // hsa_kernel_dispatch_packet_t disp;
   hsa_kernel_dispatch_packet_t* hsaDisp =
       reinterpret_cast<hsa_kernel_dispatch_packet_t*>(gpu.cb(0)->SysMemCopy());
diff --git a/rocclr/device/rocm/rocvirtual.cpp b/rocclr/device/rocm/rocvirtual.cpp
index 1d4b38e570..8fe12dd8e6 100644
--- a/rocclr/device/rocm/rocvirtual.cpp
+++ b/rocclr/device/rocm/rocvirtual.cpp
@@ -2840,12 +2840,16 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
           if (sizes.dimensions() >= 2) {
             WriteAqlArgAt(hidden_arguments, static_cast<uint32_t>(newGlobalSize[1] / local[1]),
                           it.size_, it.offset_);
+          } else {
+            WriteAqlArgAt(hidden_arguments, static_cast<uint32_t>(1), it.size_, it.offset_);
           }
           break;
         case amd::KernelParameterDescriptor::HiddenBlockCountZ:
           if (sizes.dimensions() >= 3) {
             WriteAqlArgAt(hidden_arguments, static_cast<uint32_t>(newGlobalSize[2] / local[2]),
                           it.size_, it.offset_);
+          } else {
+            WriteAqlArgAt(hidden_arguments, static_cast<uint32_t>(1), it.size_, it.offset_);
           }
           break;
         case amd::KernelParameterDescriptor::HiddenGroupSizeX:
@@ -2854,11 +2858,15 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
         case amd::KernelParameterDescriptor::HiddenGroupSizeY:
           if (sizes.dimensions() >= 2) {
             WriteAqlArgAt(hidden_arguments, static_cast<uint16_t>(local[1]), it.size_, it.offset_);
+          } else {
+            WriteAqlArgAt(hidden_arguments, static_cast<uint16_t>(1), it.size_, it.offset_);
           }
           break;
         case amd::KernelParameterDescriptor::HiddenGroupSizeZ:
           if (sizes.dimensions() >= 3) {
             WriteAqlArgAt(hidden_arguments, static_cast<uint16_t>(local[2]), it.size_, it.offset_);
+          } else {
+            WriteAqlArgAt(hidden_arguments, static_cast<uint16_t>(1), it.size_, it.offset_);
           }
           break;
         case amd::KernelParameterDescriptor::HiddenRemainderX:
@@ -2907,11 +2915,6 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
       memcpy(argBuffer, parameters, gpuKernel.KernargSegmentByteSize());
     }
 
-    // Note: In a case of structs the size won't match,
-    // since HSAIL compiler expects a reference...
-    assert(gpuKernel.KernargSegmentByteSize() <= signature.paramsSize() &&
-      "A mismatch of sizes of arguments between compiler and runtime!");
-
     // Check for group memory overflow
     //! @todo Check should be in HSA - here we should have at most an assert
     assert(roc_device_.info().localMemSizePerCU_ > 0);