SWDEV-282397 - Alloc scratch memory on the current queue

Device enqueue has an option to execute scheduler on the current
queue and it's enabled by default. Make sure scratch is allocated
on the current queue for that case. Add max vgpr tracking per
program to adjust scratch size accordingly.

Change-Id: I2a6d796913a4551a1e7f343a2465d589eec60d8a

rocclr/device/pal/palprogram.cpp

@@ -181,6 +181,7 @@ HSAILProgram::HSAILProgram(Device& device, amd::Program& owner)
       codeSegGpu_(nullptr),
       codeSegment_(nullptr),
       maxScratchRegs_(0),
+      maxVgprs_(0),
       executable_(nullptr),
       loaderContext_(this) {
   assert(device.isOnline());
@@ -194,6 +195,7 @@ HSAILProgram::HSAILProgram(NullDevice& device, amd::Program& owner)
       codeSegGpu_(nullptr),
       codeSegment_(nullptr),
       maxScratchRegs_(0),
+      maxVgprs_(0),
       executable_(nullptr),
       loaderContext_(this) {
   assert(!device.isOnline());
@@ -317,6 +319,7 @@ bool HSAILProgram::setKernels(amd::option::Options* options, void* binary, size_
       // with dynamic parallelism, since runtime doesn't know which child kernel will be called
       maxScratchRegs_ =
           std::max(static_cast<uint>(aKernel->workGroupInfo()->scratchRegs_), maxScratchRegs_);
+      maxVgprs_ = std::max(static_cast<uint>(aKernel->workGroupInfo()->usedVGPRs_), maxVgprs_);
     }
     // Allocate kernel table for device enqueuing
     if (!isNull() && dynamicParallelism && !allocKernelTable()) {
@@ -781,6 +784,7 @@ bool LightningProgram::setKernels(amd::option::Options* options, void* binary, s
     // with dynamic parallelism, since runtime doesn't know which child kernel will be called
     maxScratchRegs_ =
         std::max(static_cast<uint>(kernel->workGroupInfo()->scratchRegs_), maxScratchRegs_);
+    maxVgprs_ = std::max(static_cast<uint>(kernel->workGroupInfo()->usedVGPRs_), maxVgprs_);
   }
   DestroySegmentCpuAccess();
 #endif  // defined(USE_COMGR_LIBRARY)