Included code to calculate value of maxThreadsPerMultiprocessor property

Change-Id: Ie7cad7442f36a7163e715048de5a309febc28664


[ROCm/hip commit: 6e641df649]

projects/hip/CMakeLists.txt

@@ -114,8 +114,8 @@ if(HIP_PLATFORM STREQUAL "hcc")
     set(CMAKE_C_COMPILER   "${HCC_HOME}/bin/hcc")
 
     # Set HIP_HCC so we know this is HIP compile, some files are shared with HCC (staging_buffer).
-    set(CMAKE_CXX_FLAGS " -hc -I${HCC_HOME}/include -I${HSA_PATH}/include -stdlib=libc++ -DHIP_HCC")
-    set(CMAKE_C_FLAGS   " -hc -I${HCC_HOME}/include -I${HSA_PATH}/include -stdlib=libc++ -DHIP_HCC")
+    set(CMAKE_CXX_FLAGS " -hc -I${HCC_HOME}/include -I${HSA_PATH}/include -I/opt/rocm/libhsakmt/include/libhsakmt -stdlib=libc++ -DHIP_HCC")
+    set(CMAKE_C_FLAGS   " -hc -I${HCC_HOME}/include -I${HSA_PATH}/include -I/opt/rocm/libhsakmt/include/libhsakmt -stdlib=libc++ -DHIP_HCC")
 
     set(SOURCE_FILES src/device_util.cpp
                      src/hip_hcc.cpp 

projects/hip/bin/hipcc

@@ -81,7 +81,7 @@ if ($HIP_PLATFORM eq "hcc") {
     $HIPLDFLAGS .= " -Wl,--defsym=_binary_kernel_spir_end=0 -Wl,--defsym=_binary_kernel_spir_start=0 -Wl,--defsym=_binary_kernel_cl_start=0 -Wl,--defsym=_binary_kernel_cl_end=0";
     # Satisfy HCC dependencies
     $HIPLDFLAGS .= " -lc++abi";
-	$HIPLDFLAGS .= " -L$HSA_PATH/lib -L$ROCM_PATH/lib -lhsa-runtime64 -lhc_am";
+	$HIPLDFLAGS .= " -L$HSA_PATH/lib -L$ROCM_PATH/lib -lhsa-runtime64 -lhc_am -lhsakmt";
 
     # Add trace marker library:
     # TODO - once we cleanly separate the HIP API headers from HIP library headers this logic should move to CMakebuild option - apps do not need to see the marker library.

projects/hip/src/hip_hcc.cpp

@@ -42,6 +42,7 @@ THE SOFTWARE.
 #include "hip_runtime.h"
 #include "hcc_detail/hip_hcc.h"
 #include "hsa_ext_amd.h"
+#include "hsakmt.h"
 
 // TODO, re-org header order.
 extern const char *ihipErrorString(hipError_t hip_error);
@@ -720,17 +721,18 @@ hipError_t ihipDevice_t::getProperties(hipDeviceProp_t* prop)
     prop->computeMode = 0;
 
     // Get Max Threads Per Multiprocessor
-/*
+
     HsaSystemProperties props;
     hsaKmtReleaseSystemProperties();
     if(HSAKMT_STATUS_SUCCESS == hsaKmtAcquireSystemProperties(&props)) {
         HsaNodeProperties node_prop = {0};
         if(HSAKMT_STATUS_SUCCESS == hsaKmtGetNodeProperties(node, &node_prop)) {
             uint32_t waves_per_cu = node_prop.MaxWavesPerSIMD;
-            prop-> maxThreadsPerMultiProcessor = prop->warpsize*waves_per_cu;
+            uint32_t simd_per_cu = node_prop.NumSIMDPerCU;
+            prop-> maxThreadsPerMultiProcessor = prop->warpSize*waves_per_cu*simd_per_cu;
         }
     }
-*/
+
 
     // Get memory properties
     err = hsa_agent_iterate_regions(_hsa_agent, get_region_info, prop);