Adding support for GFX 11.5 in AQL Profiler (#2340)

* Adding support of AQL Profiler for GFX 11.5

* Removing hard coded value for sa_number

* Adding instance count for WGP block, removing hard coded values.

* Fixed SQ counter block and TD counter block instances

projects/aqlprofile/build.sh

@@ -64,7 +64,7 @@ if [ -z "$PREFIX_PATH" ] ; then PREFIX_PATH=$PACKAGE_ROOT; fi
 if [ -z "$HIP_VDI" ] ; then HIP_VDI=0; fi
 if [ -n "$ROCM_RPATH" ] ; then LD_RUNPATH_FLAG=" -Wl,--enable-new-dtags -Wl,--rpath,${ROCM_RPATH}"; fi
 if [ -z "$TO_CLEAN" ] ; then TO_CLEAN=yes; fi
-if [ -z "$GPU_LIST" ] ; then GPU_LIST="gfx900 gfx906 gfx908 gfx90a gfx942 gfx950 gfx1030 gfx1100 gfx1101 gfx1102 gfx1031 gfx1150 gfx1151"; fi
+if [ -z "$GPU_LIST" ] ; then GPU_LIST="gfx900 gfx906 gfx908 gfx90a gfx942 gfx950 gfx1030 gfx1100 gfx1101 gfx1102 gfx1031 gfx1150 gfx1151 gfx1152 gfx1153"; fi
 
 AQLPROFILE_ROOT=$(cd $AQLPROFILE_ROOT && echo $PWD)
 

projects/aqlprofile/gfxip/gfx11/gfx11_block_table.h

@@ -325,7 +325,7 @@ static const GpuBlockInfo SqCounterBlockInfo = {
     SqCounterBlockNumCounters,
     SqCounterRegAddr,
     gfx11_cntx_prim::sq_select_value,
-    CounterBlockSeAttr | CounterBlockSqAttr | CounterBlockSaAttr,
+    CounterBlockSeAttr | CounterBlockSqAttr | CounterBlockSaAttr | CounterBlockWgpAttr,
     BLOCK_DELAY_NONE};
 // Counter block SX
 static const GpuBlockInfo SxCounterBlockInfo = {
@@ -358,7 +358,7 @@ static const GpuBlockInfo TcpCounterBlockInfo = {
     TcpCounterBlockNumCounters,
     TcpCounterRegAddr,
     gfx11_cntx_prim::select_value_TCP_PERFCOUNTER0_SELECT,
-    CounterBlockDfltAttr | CounterBlockSeAttr | CounterBlockSaAttr,
+    CounterBlockDfltAttr | CounterBlockSeAttr | CounterBlockSaAttr | CounterBlockWgpAttr,
     BLOCK_DELAY_NONE};
 // Counter block GL1A
 static const GpuBlockInfo Gl1aCounterBlockInfo = {

projects/aqlprofile/src/core/aql_profile.cpp

@@ -600,6 +600,7 @@ hsa_ven_amd_aqlprofile_iterate_data(const hsa_ven_amd_aqlprofile_profile_t* prof
     const bool is_concurrent = pm4_factory->IsConcurrent();
     const uint32_t xcc_num = pm4_factory->GetXccNumber();
     const uint32_t se_number = pm4_factory->GetShaderEnginesNumber() / xcc_num;
+    const uint32_t sa_number = pm4_factory->GetShaderArraysNumber();
 
     if (profile->type == HSA_VEN_AMD_AQLPROFILE_EVENT_TYPE_PMC) {
       uint64_t* samples = reinterpret_cast<uint64_t*>(profile->output_buffer.ptr);
@@ -648,11 +649,9 @@ hsa_ven_amd_aqlprofile_iterate_data(const hsa_ven_amd_aqlprofile_profile_t* prof
           if (pm4_factory->GetBlockInfo(p)->attr & CounterBlockSeAttr)
             block_samples_count *= se_number;
           if (pm4_factory->GetBlockInfo(p)->attr & CounterBlockSaAttr)
-            block_samples_count *= 2;
+            block_samples_count *= sa_number;
           if (pm4_factory->GetBlockInfo(p)->attr & CounterBlockWgpAttr)
             block_samples_count *= pm4_factory->GetNumWGPs();
-          if (pm4_factory->GetBlockInfo(p)->attr & CounterBlockSqAttr && pm4_factory->IsGFX11())
-            block_samples_count *= pm4_factory->GetNumWGPs();
 
           for (uint32_t blk = 0; blk < block_samples_count; ++blk) {
             hsa_ven_amd_aqlprofile_info_data_t sample_info;

projects/aqlprofile/src/core/counter_dimensions.hpp

@@ -123,7 +123,11 @@ class EventAttribDimension {
     if (num_aid > 1) dimensions.push_back({"AID", num_aid});
 
     if (workgroup_processor)
+    {
       dimensions.push_back({"WGP", wgp_num});
+      if(bIsGFX11)
+        dimensions.push_back({"INSTANCE", block_instance_count});
+    }
     else
       dimensions.push_back({"INSTANCE", block_instance_count});
 

projects/aqlprofile/src/core/gfx115x_factory.cpp

@@ -50,6 +50,15 @@ Gfx115xFactory::Gfx115xFactory(const AgentInfo* agent_info)
       case Gl2cCounterBlockId:
         block_info->instance_count = 4;
         break;
+      case TcpCounterBlockId:
+        block_info->instance_count = 2;
+        break;
+      case TaCounterBlockId:
+        block_info->instance_count = 2;
+        break;
+      case TdCounterBlockId:
+        block_info->instance_count = 2;
+        break;
       default:
         break;
     }

projects/aqlprofile/src/core/pm4_factory.h

@@ -191,17 +191,16 @@ class Pm4Factory {
 
   virtual size_t GetNumEvents(uint32_t block_name) const {
     size_t se_number = GetShaderEnginesNumber() / GetXccNumber();
+    size_t sa_number = GetShaderArraysNumber();
     size_t block_samples_count = 1;
     auto* block_info = GetBlockInfo(block_name);
 
     if (block_info->attr & CounterBlockSeAttr)
       block_samples_count *= se_number;
     if (block_info->attr & CounterBlockSaAttr)
-      block_samples_count *= 2;
+      block_samples_count *= sa_number;
     if (block_info->attr & CounterBlockWgpAttr)
       block_samples_count *= GetNumWGPs();
-    if ((block_info->attr & CounterBlockSqAttr) && IsGFX11()) // TODO: Move to CounterBlockWgpAttr
-      block_samples_count *= GetNumWGPs();
     return block_samples_count;
   }
 

projects/aqlprofile/src/pm4/pmc_builder.h

@@ -577,7 +577,7 @@ class GpuPmcBuilder : public PmcBuilder, protected Primitives {
         // skip
       } else {
         const uint32_t se_end_index = (block_info->attr & CounterBlockSeAttr) ? se_number_ : 1;
-        const uint32_t sa_end_index = (block_info->attr & CounterBlockSaAttr) ? 2 : 1;
+        const uint32_t sa_end_index = (block_info->attr & CounterBlockSaAttr) ? sarrays_per_se : 1;
         for (uint32_t se_index = 0; se_index < se_end_index; ++se_index)
           for (uint32_t sarray = 0; sarray < sa_end_index; ++sarray) {
             uint32_t grbm_value = Primitives::grbm_broadcast_value();