clr: Increase kernelArg/managedBuffer size (#1586)

* Increase the buffer to 4MB. That can help kernel launches limited by a deep kernel pipeline

Co-authored-by: JeniferC99 <150404595+JeniferC99@users.noreply.github.com>

projects/clr/rocclr/device/rocm/rocvirtual.cpp

@@ -1801,7 +1801,7 @@ bool VirtualGPU::ManagedBuffer::Create(Device::MemorySegment mem_segment) {
   }
   hsa_agent_t agent = gpu_.dev().getBackendDevice();
   for (auto& it : pool_signal_) {
-    if (HSA_STATUS_SUCCESS != Hsa::signal_create(0, 1, &agent, &it)) {
+    if (HSA_STATUS_SUCCESS != Hsa::signal_create(0, 1, &agent, HSA_AMD_SIGNAL_AMD_GPU_ONLY, &it)) {
       return false;
     }
   }
@@ -1810,8 +1810,7 @@ bool VirtualGPU::ManagedBuffer::Create(Device::MemorySegment mem_segment) {
 
 // ================================================================================================
 address VirtualGPU::ManagedBuffer::Acquire(uint32_t size) {
-  auto alignment = amd::alignUp(256u, gpu_.dev().info().globalMemCacheLineSize_);
-  return Acquire(size, alignment);
+  return Acquire(size, gpu_.dev().info().globalMemCacheLineSize_);
 }
 
 // ================================================================================================

projects/clr/rocclr/device/rocm/rocvirtual.hpp

@@ -196,7 +196,7 @@ class VirtualGPU : public device::VirtualDevice {
   class ManagedBuffer : public amd::EmbeddedObject {
    public:
     //! The number of chunks the arg pool will be divided
-    static constexpr uint32_t kPoolNumSignals = 4;
+    static constexpr uint32_t kPoolNumSignals = 16;
     ManagedBuffer(VirtualGPU& gpu, uint32_t pool_size)
         : gpu_(gpu), pool_size_(pool_size), pool_signal_(kPoolNumSignals) {}
     ~ManagedBuffer();

projects/clr/rocclr/utils/flags.hpp

@@ -108,7 +108,7 @@ release(uint, OPENCL_VERSION, 200,                                            \
         "Force GPU opencl version")                                           \
 release(bool, HSA_LOCAL_MEMORY_ENABLE, true,                                  \
         "Enable HSA device local memory usage")                               \
-release(uint, HSA_KERNARG_POOL_SIZE, 1024 * 1024,                             \
+release(uint, HSA_KERNARG_POOL_SIZE, 4 * 1024 * 1024,                         \
         "Kernarg pool size")                                                  \
 release(bool, GPU_MIPMAP, true,                                               \
         "Enables GPU mipmap extension")                                       \