update for gfx9 and 3.1

Change-Id: I0c218f8a2dcdad1c8deb44770d8a64ccd95a92fb

test/app/intercept_test.cpp

@@ -228,7 +228,7 @@ hsa_status_t dispatch_callback(const rocprofiler_callback_data_t* callback_data,
 
 unsigned metrics_input(rocprofiler_feature_t** ret) {
   // Profiling feature objects
-  const unsigned feature_count = 9;
+  const unsigned feature_count = 6;
   rocprofiler_feature_t* features = new rocprofiler_feature_t[feature_count];
   memset(features, 0, feature_count * sizeof(rocprofiler_feature_t));
 
@@ -245,12 +245,12 @@ unsigned metrics_input(rocprofiler_feature_t** ret) {
   features[4].name = "SQ_INSTS_VALU";
   features[5].kind = ROCPROFILER_FEATURE_KIND_METRIC;
   features[5].name = "VALUInsts";
-  features[6].kind = ROCPROFILER_FEATURE_KIND_METRIC;
-  features[6].name = "TCC_HIT_sum";
-  features[7].kind = ROCPROFILER_FEATURE_KIND_METRIC;
-  features[7].name = "TCC_MISS_sum";
-  features[8].kind = ROCPROFILER_FEATURE_KIND_METRIC;
-  features[8].name = "WRITE_SIZE";
+//  features[6].kind = ROCPROFILER_FEATURE_KIND_METRIC;
+//  features[6].name = "TCC_HIT_sum";
+//  features[7].kind = ROCPROFILER_FEATURE_KIND_METRIC;
+//  features[7].name = "TCC_MISS_sum";
+//  features[8].kind = ROCPROFILER_FEATURE_KIND_METRIC;
+//  features[8].name = "WRITE_SIZE";
 
   *ret = features;
   return feature_count;

test/app/standalone_test.cpp

@@ -82,7 +82,7 @@ int main() {
   rocprofiler_properties_t properties;
 
   // Profiling feature objects
-  const unsigned feature_count = 9;
+  const unsigned feature_count = 6;
   rocprofiler_feature_t feature[feature_count];
   // PMC events
   memset(feature, 0, sizeof(feature));
@@ -98,12 +98,12 @@ int main() {
   feature[4].name = "SQ_INSTS_VALU";
   feature[5].kind = ROCPROFILER_FEATURE_KIND_METRIC;
   feature[5].name = "VALUInsts";
-  feature[6].kind = ROCPROFILER_FEATURE_KIND_METRIC;
-  feature[6].name = "TCC_HIT_sum";
-  feature[7].kind = ROCPROFILER_FEATURE_KIND_METRIC;
-  feature[7].name = "TCC_MISS_sum";
-  feature[8].kind = ROCPROFILER_FEATURE_KIND_METRIC;
-  feature[8].name = "WRITE_SIZE";
+//  feature[6].kind = ROCPROFILER_FEATURE_KIND_METRIC;
+//  feature[6].name = "TCC_HIT_sum";
+//  feature[7].kind = ROCPROFILER_FEATURE_KIND_METRIC;
+//  feature[7].name = "TCC_MISS_sum";
+//  feature[8].kind = ROCPROFILER_FEATURE_KIND_METRIC;
+//  feature[8].name = "WRITE_SIZE";
 //  feature[8].kind = ROCPROFILER_FEATURE_KIND_METRIC;
 //  feature[8].name = "TCC_EA_WRREQ_sum";
 //  feature[9].kind = ROCPROFILER_FEATURE_KIND_METRIC;

test/app/test.cpp

@@ -20,9 +20,13 @@ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 THE SOFTWARE.
 *******************************************************************************/
 
+#include <dirent.h>
 #include <hsa.h>
+#include <hsakmt.h>
 #include <stdio.h>
 #include <string.h>
+
+#include <fstream>
 #include <iostream>
 #include <thread>
 
@@ -31,6 +35,64 @@ THE SOFTWARE.
 #include "dummy_kernel/dummy_kernel.h"
 #include "simple_convolution/simple_convolution.h"
 
+int get_gpu_node_id() {
+  int gpu_node = - 1;
+
+#if 0
+  // find a valid gpu node from /sys/class/kfd/kfd/topology/nodes
+  std::string path = "/sys/class/kfd/kfd/topology/nodes";
+  DIR *dir;
+  struct dirent *ent;
+
+  if ((dir = opendir(path.c_str())) != NULL) {
+    while ((ent = readdir(dir)) != NULL) {
+
+      std::string dir = ent->d_name;
+
+      if (dir.find_first_not_of("0123456789") == std::string::npos) {
+        std::string file = path + "/" + ent->d_name + "/gpu_id";
+        std::ifstream infile(file);
+        int id;
+
+        infile >> id;
+        if (id != 0) {
+          gpu_node = atoi(ent->d_name);
+          break;
+        }
+      }
+    }
+    closedir (dir);
+  }
+#else
+  HsaSystemProperties m_SystemProperties;
+  memset(&m_SystemProperties, 0, sizeof(m_SystemProperties));
+
+  HSAKMT_STATUS status = hsaKmtAcquireSystemProperties(&m_SystemProperties);
+  if (status != HSAKMT_STATUS_SUCCESS) {
+    std::cerr << "Error in hsaKmtAcquireSystemProperties"<< std::endl;
+    return 1;
+  }
+
+  // tranverse all CPU and GPU nodes and break when a GPU node is found
+  for (unsigned i = 0; i < m_SystemProperties.NumNodes; ++i) {
+    HsaNodeProperties nodeProperties;
+    memset(&nodeProperties, 0, sizeof(HsaNodeProperties));
+
+    status = hsaKmtGetNodeProperties(i, &nodeProperties);
+    if (status != HSAKMT_STATUS_SUCCESS) {
+      std::cerr << "Error in hsaKmtAcquireSystemProperties"<< std::endl;
+      break;
+    } else if(nodeProperties.NumFComputeCores) {
+      gpu_node = i;
+      break;
+    }
+  }
+#endif
+
+  printf ("GPU node id(%d)\n", gpu_node);
+  return gpu_node;
+}
+
 void thread_fun(const int kiter, const int diter, const uint32_t agents_number) {
   const AgentInfo* agent_info[agents_number];
   hsa_queue_t* queue[agents_number];
@@ -65,12 +127,31 @@ int main(int argc, char** argv) {
   const char* diter_s = getenv("ROCP_DITER");
   const char* agents_s = getenv("ROCP_AGENTS");
   const char* thrs_s = getenv("ROCP_THRS");
+  const char* spm_enabled = getenv("ROCP_SPM");
+  int gpu_node_id = -1;
 
   const int kiter = (kiter_s != NULL) ? atol(kiter_s) : 1;
   const int diter = (diter_s != NULL) ? atol(diter_s) : 1;
   const uint32_t agents_number = (agents_s != NULL) ? (uint32_t)atol(agents_s) : 1;
   const int thrs = (thrs_s != NULL) ? atol(thrs_s) : 1;
 
+  if (spm_enabled != NULL) {
+    if (hsa_init() != HSA_STATUS_SUCCESS) {
+      std::cerr << "Error in hsa_init()" << std::endl;
+      return 1;
+    }
+    gpu_node_id = get_gpu_node_id();
+    if (gpu_node_id == -1) {
+      std::cerr << "Error in get_gpu_node_id()" << std::endl;
+      return 1;
+    }
+    HSAKMT_STATUS status = hsaKmtEnableDebugTrap(gpu_node_id, INVALID_QUEUEID);
+    if (status != HSAKMT_STATUS_SUCCESS) {
+      std::cerr << "Error in enabling debug trap" << std::endl;
+      return 1;
+    }
+  }
+
   TestHsa::HsaInstantiate();
 
   std::vector<std::thread> t(thrs);
@@ -81,6 +162,18 @@ int main(int argc, char** argv) {
     t[n].join();
   }
 
+  if (spm_enabled != NULL) {
+    if (gpu_node_id == -1) {
+      std::cerr << "Invalid GPU node id" << std::endl;
+      return 1;
+    }
+    HSAKMT_STATUS status = hsaKmtDisableDebugTrap(gpu_node_id);
+    if (status != HSAKMT_STATUS_SUCCESS) {
+      std::cerr << "Error in disabling debug" << std::endl;
+      return 1;
+    }
+  }
+
   TestHsa::HsaShutdown();
   return 0;
 }