Merge pull request #52 from RadeonOpenCompute/refactorValidationSignal

Refactor validation signal treatment in a device agnostic manner

rocm_bandwidth_test.cpp

@@ -641,17 +641,25 @@ void RocmBandwidthTest::RunCopyBenchmark(async_trans_t& trans) {
       }
     }
 
-    // Get Cpu min and mean times for copy
-    // Push them into the Cpu time list
-    trans.cpu_min_time_.push_back(GetMinTime(cpu_time));
-    trans.cpu_avg_time_.push_back(GetMeanTime(cpu_time));
-
+    // Collecting Cpu time. Capture verify failures if any
+    // Get min and mean copy times and collect them into Cpu
+    // time list
+    double min_time = 0;
+    double mean_time = 0;
+    if (print_cpu_time_) {
+      min_time = (verify) ? GetMinTime(cpu_time) : VALIDATE_COPY_OP_FAILURE;
+      mean_time = (verify) ? GetMeanTime(cpu_time) : VALIDATE_COPY_OP_FAILURE;
+      trans.cpu_min_time_.push_back(min_time);
+      trans.cpu_avg_time_.push_back(mean_time);
+    }
+
+    // Collecting Gpu time. Capture verify failures if any
+    // Get min and mean copy times and collect them into Gpu
+    // time list
     if (print_cpu_time_ == false) {
       if (trans.copy.uses_gpu_) {
-        // Get Gpu min and mean copy times
-        // Push them into the Gpu time list
-        double min_time = (verify) ? GetMinTime(gpu_time) : VALIDATE_COPY_OP_FAILURE;
-        double mean_time = (verify) ? GetMeanTime(gpu_time) : VALIDATE_COPY_OP_FAILURE;
+        min_time = (verify) ? GetMinTime(gpu_time) : VALIDATE_COPY_OP_FAILURE;
+        mean_time = (verify) ? GetMeanTime(gpu_time) : VALIDATE_COPY_OP_FAILURE;
         trans.gpu_min_time_.push_back(min_time);
         trans.gpu_avg_time_.push_back(mean_time);
       }
@@ -796,7 +804,7 @@ RocmBandwidthTest::RocmBandwidthTest(int argc, char** argv) : BaseTest() {
   // Initialize version of the test
   version_.major_id = 2;
   version_.minor_id = 3;
-  version_.step_id = 7;
+  version_.step_id = 9;
   version_.reserved = 0;
 
   bw_iter_cnt_ = getenv("ROCM_BW_ITER_CNT");

rocm_bandwidth_test_parse.cpp

@@ -147,9 +147,7 @@ void RocmBandwidthTest::ValidateCopyUnidirFlags(uint32_t copy_mask,
   // It is illegal to specify Latency and another
   // secondary flag that affects a copy operation
   if ((copy_ctrl_mask & DEV_COPY_LATENCY) &&
-       ((copy_ctrl_mask & USR_BUFFER_INIT) ||
-        (copy_ctrl_mask & CPU_VISIBLE_TIME) ||
-        (copy_ctrl_mask & VALIDATE_COPY_OP))) {
+      (copy_ctrl_mask & VALIDATE_COPY_OP)) {
     PrintHelpScreen();
     exit(0);
   }

rocm_bandwidth_test_print.cpp

@@ -70,13 +70,10 @@ void RocmBandwidthTest::PrintHelpScreen() {
   std::cout << std::endl;
   
   std::cout << "\t NOTE: Mixing following options is illegal/unsupported" << std::endl;
-  std::cout << "\t\t Case 1: rocm_bandwidth_test -a or -A with -c" << std::endl;
-  std::cout << "\t\t Case 2: rocm_bandwidth_test -b or -A with -m" << std::endl;
-  std::cout << "\t\t Case 3: rocm_bandwidth_test -b or -A with -l" << std::endl;
-  std::cout << "\t\t Case 4: rocm_bandwidth_test -b or -A with -v" << std::endl;
-  std::cout << "\t\t Case 5: rocm_bandwidth_test -a or -s x -d y with -l and -c" << std::endl;
-  std::cout << "\t\t Case 6: rocm_bandwidth_test -a or -s x -d y with -l and -m" << std::endl;
-  std::cout << "\t\t Case 7: rocm_bandwidth_test -a or -s x -d y with -l and -v" << std::endl;
+  std::cout << "\t\t Case 1: rocm_bandwidth_test -a with {lm}{1,}" << std::endl;
+  std::cout << "\t\t Case 2: rocm_bandwidth_test -b with {clv}{1,}" << std::endl;
+  std::cout << "\t\t Case 3: rocm_bandwidth_test -A with {clmv}{1,}" << std::endl;
+  std::cout << "\t\t Case 4: rocm_bandwidth_test -s x -d y with {lmv}{2,}" << std::endl;
   std::cout << std::endl;
 
   std::cout << std::endl;
@@ -152,11 +149,19 @@ void RocmBandwidthTest::PrintTopology() {
     
     if (HSA_DEVICE_TYPE_CPU == node.agent.device_type_) {
       std::cout << "  Device Type:                            CPU" << std::endl;
+      std::cout.width(format);
+      std::cout << "";
+      std::cout.width(format);
+      std::cout << "  Device Name:                            " << node.agent.name_ << std::endl;
     } else if (HSA_DEVICE_TYPE_GPU == node.agent.device_type_) {
       std::cout << "  Device Type:                            GPU" << std::endl;
       std::cout.width(format);
       std::cout << "";
       std::cout.width(format);
+      std::cout << "  Device Name:                            " << node.agent.name_ << std::endl;
+      std::cout.width(format);
+      std::cout << "";
+      std::cout.width(format);
       std::cout << "  Device  BDF:                            " << node.agent.bdf_id_ << std::endl;
     }
 

rocm_bandwidth_test_trans.cpp

@@ -418,44 +418,45 @@ void RocmBandwidthTest::ComputeCopyTime(async_trans_t& trans) {
       data_size += data_size;
     }
 
-    // Copy operation does not involve a Gpu device
-    // Divide bandwidth with 10^9 to get size in GigaBytes (10^9)
-    if (trans.copy.uses_gpu_ != true) {
-      avg_time = trans.cpu_avg_time_[idx];
-      min_time = trans.cpu_min_time_[idx];
-      avg_bandwidth = (double)data_size / avg_time / 1000 / 1000 / 1000;
-      peak_bandwidth = (double)data_size / min_time / 1000 / 1000 / 1000;
-    } else {
-      if (print_cpu_time_) {
+    // Get time taken by copy operation
+    if ((print_cpu_time_) ||
+        (trans.copy.uses_gpu_ != true)) {
       avg_time = trans.cpu_avg_time_[idx];
       min_time = trans.cpu_min_time_[idx];
     } else {
-        avg_time = trans.gpu_avg_time_[idx] / sys_freq;
-        min_time = trans.gpu_min_time_[idx] / sys_freq;
-      }
-      avg_bandwidth = (double)data_size / avg_time / 1000 / 1000 / 1000;
-      peak_bandwidth = (double)data_size / min_time / 1000 / 1000 / 1000;
+      avg_time = trans.gpu_avg_time_[idx];
+      min_time = trans.gpu_min_time_[idx];
     }
 
-    trans.min_time_.push_back(min_time);
-    trans.avg_time_.push_back(avg_time);
+    // Determine if there was a validation failure
+    // @note: Value is set to VALIDATE_COPY_OP_FAILURE
+    // if copy transaction wa validated and it failed
+    hsa_status_t verify_status = HSA_STATUS_ERROR;
+    if ((avg_time != VALIDATE_COPY_OP_FAILURE) &&
+        (min_time != VALIDATE_COPY_OP_FAILURE)) {
+      verify_status = HSA_STATUS_SUCCESS;
+    }
 
-    // Check validation failures as that signal is
-    // captured via Min and Avg time values. If there
-    // is a failure propagate that value as computed
-    // bandwidth
-    if (validate_) {
-      avg_time = trans.gpu_avg_time_[idx];
-      min_time = trans.gpu_min_time_[idx];
-      if ((avg_time == VALIDATE_COPY_OP_FAILURE) &&
-          (min_time == VALIDATE_COPY_OP_FAILURE)) {
-        trans.avg_bandwidth_.push_back(avg_time);
-        trans.peak_bandwidth_.push_back(min_time);
-        continue;  
+    // Adjust Gpu time if there is no validation error
+    if ((trans.copy.uses_gpu_) &&
+        (print_cpu_time_ == false) &&
+        (verify_status == HSA_STATUS_SUCCESS)) {
+      avg_time = avg_time / sys_freq;
+      min_time = min_time / sys_freq;
     }
+
+    // Compute bandwidth - divide bandwidth with
+    // 10^9 not 1024^3 to get size in GigaBytes
+    // @note: For validation failures bandwidth
+    // is encoded by VALIDATE_COPY_OP_FAILURE
+    if (verify_status == HSA_STATUS_SUCCESS) {
+      avg_bandwidth = (double)data_size / avg_time / 1000 / 1000 / 1000;
+      peak_bandwidth = (double)data_size / min_time / 1000 / 1000 / 1000;
     }
 
     // Update computed bandwidth for the transaction
+    trans.min_time_.push_back(min_time);
+    trans.avg_time_.push_back(avg_time);
     trans.avg_bandwidth_.push_back(avg_bandwidth);
     trans.peak_bandwidth_.push_back(peak_bandwidth);
   }