Fix RBT to allow Fine grain regions to be reported by GPU devices

cmake_modules/utils.cmake

@@ -99,8 +99,11 @@ function( parse_version VERSION_STRING )
   endif ()
 
   # Add Build Info from Jenkins
-  set ( ROCM_BUILD_ID "DevBld" CACHE STRING "Jenkins Build Id" )
-   
+  set ( ROCM_BUILD_ID "DevBld" CACHE STRING "Local Build Id" FORCE )
+  if(DEFINED ENV{ROCM_BUILD_ID})
+    set ( ROCM_BUILD_ID $ENV{ROCM_BUILD_ID} CACHE STRING "Jenkins Build Id" FORCE )
+  endif()
+
   # Update Version Patch to include Number of Commits and hash of HEAD
   set ( VERSION_PATCH "${PATCH}.${COMMIT_CNT}-${ROCM_BUILD_ID}-${COMMIT_HASH}" PARENT_SCOPE )
 

rocm_bandwidth_test.cpp

@@ -789,13 +789,14 @@ RocmBandwidthTest::RocmBandwidthTest(int argc, char** argv) : BaseTest() {
   // Initialize version of the test
   version_.major_id = 2;
   version_.minor_id = 3;
-  version_.step_id = 3;
+  version_.step_id = 4;
   version_.reserved = 0;
 
   bw_iter_cnt_ = getenv("ROCM_BW_ITER_CNT");
   bw_default_run_ = getenv("ROCM_BW_DEFAULT_RUN");
   bw_blocking_run_ = getenv("ROCR_BW_RUN_BLOCKING");
-  skip_fine_grain_ = getenv("ROCM_SKIP_FINE_GRAINED_POOL");
+  skip_cpu_fine_grain_ = getenv("ROCM_SKIP_CPU_FINE_GRAINED_POOL");
+  skip_gpu_coarse_grain_ = getenv("ROCM_SKIP_GPU_COARSE_GRAINED_POOL");
 
   if (bw_iter_cnt_ != NULL) {
     int32_t num = atoi(bw_iter_cnt_);

rocm_bandwidth_test.hpp

@@ -488,7 +488,8 @@ class RocmBandwidthTest : public BaseTest {
   
   // Env key to determine if Fine-grained or
   // Coarse-grained pool should be filtered out
-  char* skip_fine_grain_;
+  char* skip_cpu_fine_grain_;
+  char* skip_gpu_coarse_grain_;
   
   // Env key to determine if the run should block
   // or actively wait on completion signal

rocm_bandwidth_test_topology.cpp

@@ -106,10 +106,11 @@ hsa_status_t MemPoolInfo(hsa_amd_memory_pool_t pool, void* data) {
   }
 
   // Consult user request and add either fine-grained or
-  // coarse-grained memory pools if agent is CPU
+  // coarse-grained memory pools if agent is CPU. Default
+  // is to skip coarse-grained memory pools
   agent_info_t& agent_info = asyncDrvr->agent_list_.back();
   if (agent_info.device_type_ == HSA_DEVICE_TYPE_CPU) {
-    if (asyncDrvr->skip_fine_grain_ != NULL) {
+    if (asyncDrvr->skip_cpu_fine_grain_ != NULL) {
       if (is_fine_grained == true) {
         return HSA_STATUS_SUCCESS;
       }
@@ -119,9 +120,21 @@ hsa_status_t MemPoolInfo(hsa_amd_memory_pool_t pool, void* data) {
       }
     }
   }
-  // hsa_device_type_t device_type;
-  // status = hsa_agent_get_info(agent, HSA_AGENT_INFO_DEVICE, &device_type);
-  // ErrorCheck(status);
+
+  // Consult user request and add either fine-grained or
+  // coarse-grained memory pools if agent is GPU. Default
+  // is to skip fine-grained memory pools
+  if (agent_info.device_type_ == HSA_DEVICE_TYPE_GPU) {
+    if (asyncDrvr->skip_gpu_coarse_grain_ != NULL) {
+      if (is_fine_grained == false) {
+        return HSA_STATUS_SUCCESS;
+      }
+    } else {
+      if (is_fine_grained == true) {
+        return HSA_STATUS_SUCCESS;
+      }
+    }
+  }
 
   // Create an instance of agent_pool_info and add it to the list
   pool_info_t pool_info(agent, asyncDrvr->agent_index_, pool,
@@ -231,7 +244,7 @@ void RocmBandwidthTest::PopulateAccessMatrix() {
       uint32_t path;
       path = (access == HSA_AMD_MEMORY_POOL_ACCESS_NEVER_ALLOWED) ? 0 : 1;
       direct_access_matrix_[(src_dev_idx * agent_index_) + dst_dev_idx] = path;
-
+      
       if ((src_dev_type == HSA_DEVICE_TYPE_CPU) &&
           (dst_dev_type == HSA_DEVICE_TYPE_GPU) &&
           (access == HSA_AMD_MEMORY_POOL_ACCESS_NEVER_ALLOWED)) {

rocm_bandwidth_test_trans.cpp

@@ -130,7 +130,7 @@ bool RocmBandwidthTest::FilterCpuPool(uint32_t req_type,
 
   // If env to skip fine grain is NULL it means
   // we should filter out coarse-grain pools
-  if (skip_fine_grain_ == NULL) {
+  if (skip_cpu_fine_grain_ == NULL) {
     return (fine_grained == false);
   }