ROCm 3.9.0 updates

src/CMakeLists.txt

@@ -393,9 +393,9 @@ set (CPACK_COMPONENTS_ALL_IN_ONE_PACKAGE 1)
 set (CPACK_COMPONENTS_ALL binary dev)
 
 set ( CPACK_PACKAGE_NAME "hsa-rocr-dev" )
-set ( CPACK_PACKAGE_VENDOR "AMD" )
+set ( CPACK_PACKAGE_VENDOR "Advanced Micro Devices, Inc." )
 set ( CPACK_PACKAGE_VERSION ${PACKAGE_VERSION_STRING} )
-set ( CPACK_PACKAGE_CONTACT "Advanced Micro Devices Inc." )
+set ( CPACK_PACKAGE_CONTACT "Advanced Micro Devices, Inc." )
 set ( CPACK_PACKAGE_DESCRIPTION_SUMMARY "AMD Heterogeneous System Architecture HSA - Linux HSA Runtime for Boltzmann (ROCm) platforms" )
 set ( CPACK_RESOURCE_FILE_LICENSE "${CMAKE_CURRENT_SOURCE_DIR}/LICENSE.md" )
 
@@ -404,6 +404,7 @@ configure_file(${CMAKE_CURRENT_SOURCE_DIR}/DEBIAN/post_install DEBIAN/postinst @
 configure_file(${CMAKE_CURRENT_SOURCE_DIR}/DEBIAN/pre_remove DEBIAN/prerm @ONLY)
 
 # Debian package specific variables
+set ( CPACK_DEBIAN_FILE_NAME "DEB-DEFAULT" )
 set ( CPACK_DEBIAN_PACKAGE_DEPENDS "hsakmt-roct" )
 set ( CPACK_DEBIAN_PACKAGE_HOMEPAGE "https://github.com/RadeonOpenCompute/ROCR-Runtime" )
 set ( CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "DEBIAN/postinst;DEBIAN/prerm" )
@@ -415,6 +416,7 @@ set ( CPACK_DEBIAN_PACKAGE_REPLACES "hsa-ext-rocr-dev" )
 set ( CPACK_DEBIAN_PACKAGE_CONFLICTS "hsa-ext-rocr-dev" )
 
 ## RPM package specific variables
+set ( CPACK_RPM_FILE_NAME "RPM-DEFAULT" )
 set ( CPACK_RPM_PACKAGE_DEPENDS "hsakmt-roct" )
 
 # Declare that this package will replace functionality provided by hsa-ext-rocr-dev package

src/core/inc/amd_gpu_agent.h

@@ -502,7 +502,8 @@ class GpuAgent : public GpuAgentInt {
   lazy_ptr<core::Blit>& GetPcieBlit(const core::Agent& dst_agent, const core::Agent& src_agent);
 
   // Bind the Blit object that will drive the copy operation
-  lazy_ptr<core::Blit>& GetBlitObject(const core::Agent& dst_agent, const core::Agent& src_agent);
+  lazy_ptr<core::Blit>& GetBlitObject(const core::Agent& dst_agent, const core::Agent& src_agent,
+                                      const size_t size);
   // @brief Alternative aperture base address. Only on KV.
   uintptr_t ape1_base_;
 

src/core/runtime/amd_gpu_agent.cpp

@@ -645,7 +645,7 @@ hsa_status_t GpuAgent::DmaCopy(void* dst, core::Agent& dst_agent,
                                std::vector<core::Signal*>& dep_signals,
                                core::Signal& out_signal) {
   // Bind the Blit object that will drive this copy operation
-  lazy_ptr<core::Blit>& blit = GetBlitObject(dst_agent, src_agent);
+  lazy_ptr<core::Blit>& blit = GetBlitObject(dst_agent, src_agent, size);
 
   if (profiling_enabled()) {
     // Track the agent so we could translate the resulting timestamp to system
@@ -1188,7 +1188,7 @@ uint64_t GpuAgent::TranslateTime(uint64_t tick) {
   uint64_t system_tick = 0;
   double ratio = double(t1_.SystemClockCounter - t0_.SystemClockCounter) /
       double(t1_.GPUClockCounter - t0_.GPUClockCounter);
-  system_tick = uint64_t(ratio * double(int64_t(tick - t1_.GPUClockCounter))) + t1_.SystemClockCounter;
+  system_tick = uint64_t(int64_t(ratio * double(int64_t(tick - t1_.GPUClockCounter)))) + t1_.SystemClockCounter;
 
   // tick predates HSA startup - extrapolate with fixed clock ratio
   if (tick < t0_.GPUClockCounter) {
@@ -1354,7 +1354,7 @@ lazy_ptr<core::Blit>& GpuAgent::GetPcieBlit(const core::Agent& dst_agent,
 }
 
 lazy_ptr<core::Blit>& GpuAgent::GetBlitObject(const core::Agent& dst_agent,
-                                              const core::Agent& src_agent) {
+                                              const core::Agent& src_agent, const size_t size) {
   // At this point it is guaranteed that one of
   // the two devices is a GPU, potentially both
   assert(((src_agent.device_type() == core::Agent::kAmdGpuDevice) ||
@@ -1363,6 +1363,11 @@ lazy_ptr<core::Blit>& GpuAgent::GetBlitObject(const core::Agent& dst_agent,
 
   // Determine if Src and Dst devices are same
   if ((src_agent.public_handle().handle) == (dst_agent.public_handle().handle)) {
+    // If the copy is very small then cache flush overheads can dominate.
+    // Choose a (potentially) SDMA enabled engine to avoid cache flushing.
+    if (size < core::Runtime::runtime_singleton_->flag().force_sdma_size()) {
+      return blits_[BlitDevToHost];
+    }
     return blits_[BlitDevToDev];
   }
 

src/core/runtime/hsa.cpp

@@ -1888,6 +1888,10 @@ static std::string ConvertOldTargetNameToNew(
     NewName = "amdgcn-amd-amdhsa--gfx1011";
   else if (OldName == "AMD:AMDGPU:10:1:2")
     NewName = "amdgcn-amd-amdhsa--gfx1012";
+  else if (OldName == "AMD:AMDGPU:10:3:0")
+    NewName = "amdgcn-amd-amdhsa--gfx1030";
+  else if (OldName == "AMD:AMDGPU:10:3:1")
+    NewName = "amdgcn-amd-amdhsa--gfx1031";
   else
     assert(false && "Unhandled target");
 

src/core/runtime/isa.cpp

@@ -236,7 +236,10 @@ const IsaRegistry::IsaMap IsaRegistry::GetSupportedIsas() {
   ISAREG_ENTRY_GEN(10, 1, 1, true,  false)
   ISAREG_ENTRY_GEN(10, 1, 2, false, false)
   ISAREG_ENTRY_GEN(10, 1, 2, true,  false)
-
+  ISAREG_ENTRY_GEN(10, 3, 0, false, false)
+  ISAREG_ENTRY_GEN(10, 3, 0, true,  false)
+  ISAREG_ENTRY_GEN(10, 3, 1, false, false)
+  ISAREG_ENTRY_GEN(10, 3, 1, true,  false)
   return supported_isas;
 }
 

src/core/util/flag.h

@@ -121,6 +121,9 @@ class Flag {
 
     var = os::GetEnvVar("HSA_LOADER_ENABLE_MMAP_URI");
     loader_enable_mmap_uri_ = (var == "1") ? true : false;
+
+    var = os::GetEnvVar("HSA_FORCE_SDMA_SIZE");
+    force_sdma_size_ = var.empty() ? 1024 * 1024 : atoi(var.c_str());
   }
 
   bool check_flat_scratch() const { return check_flat_scratch_; }
@@ -165,6 +168,8 @@ class Flag {
 
   bool loader_enable_mmap_uri() const { return loader_enable_mmap_uri_; }
 
+  size_t force_sdma_size() const { return force_sdma_size_; }
+
  private:
   bool check_flat_scratch_;
   bool enable_vm_fault_message_;
@@ -193,6 +198,8 @@ class Flag {
 
   std::string tools_lib_names_;
 
+  size_t force_sdma_size_;
+
   DISALLOW_COPY_AND_ASSIGN(Flag);
 };
 

src/image/blit_kernel.cpp

@@ -83,6 +83,8 @@ extern uint8_t ocl_blit_object_gfx908[];
 extern uint8_t ocl_blit_object_gfx1010[];
 extern uint8_t ocl_blit_object_gfx1011[];
 extern uint8_t ocl_blit_object_gfx1012[];
+extern uint8_t ocl_blit_object_gfx1030[];
+extern uint8_t ocl_blit_object_gfx1031[];
 
 // Arguments inserted by OCL compiler, all zero here.
 struct OCLHiddenArgs {
@@ -1005,6 +1007,12 @@ hsa_status_t BlitKernel::GetPatchedBlitObject(const char* agent_name,
   case 1012:
     *blit_code_object = ocl_blit_object_gfx1012;
     break;
+  case 1030:
+    *blit_code_object = ocl_blit_object_gfx1030;
+    break;
+  case 1031:
+    *blit_code_object = ocl_blit_object_gfx1031;
+    break;
   default:
     return HSA_STATUS_ERROR_INVALID_ISA_NAME;
   }

src/image/blit_src/CMakeLists.txt

@@ -75,7 +75,7 @@ set( XNACK_DEVS ${XNACK_DEVS} CACHE STRING "XNACK targets" FORCE )
 
 # Determine the target devices if not specified
 if (NOT DEFINED TARGET_DEVICES)
-  set (TARGET_DEVICES "gfx700;gfx701;gfx702;gfx801;gfx802;gfx803;gfx900;gfx902;gfx904;gfx906;gfx908;gfx1010;gfx1011;gfx1012")
+  set (TARGET_DEVICES "gfx700;gfx701;gfx702;gfx801;gfx802;gfx803;gfx900;gfx902;gfx904;gfx906;gfx908;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031")
 endif()
 set( TARGET_DEVICES ${TARGET_DEVICES} CACHE STRING "Build targets" FORCE )
 

src/image/device_info.cpp

@@ -181,6 +181,7 @@ uint32_t DevIDToAddrLibFamily(uint32_t dev_id) {
         switch (minor_ver) {
           case 0:
           case 1:  // Navi
+          case 3:
             switch (step) {
               case 0:
               case 1: