Removing ATT buffer size limitation (#534)

* Removing SQTT buffer size limitation

* Update source/lib/rocprofiler-sdk/thread_trace/core.cpp

* Added testing for buffer size. Formatting.

* Add test as unstable

* Increase default buffer size

* Apply suggestions from code review

Co-authored-by: Indic, Vladimir <Vladimir.Indic@amd.com>

* Fix typo from code review

* Update tests/thread-trace/agent.cpp

---------

Co-authored-by: Giovanni <gbaraldi@amd.com>
Co-authored-by: Indic, Vladimir <Vladimir.Indic@amd.com>

source/bin/rocprofv3.py

@@ -769,7 +769,7 @@ For MPI applications (or other job launchers such as SLURM), place rocprofv3 ins
 
     att_options.add_argument(
         "--att-buffer-size",
-        help="Thread trace buffer size. Default 96MB",
+        help="Thread trace buffer size. Default 256MB",
         default=None,
         type=str,
     )

source/lib/rocprofiler-sdk-tool/config.hpp

@@ -139,7 +139,8 @@ struct config : output_config
     int      mpi_rank = get_mpi_rank();
     uint64_t att_param_shader_engine_mask =
         get_env<uint64_t>("ROCPROF_ATT_PARAM_SHADER_ENGINE_MASK", 0x1);
-    uint64_t att_param_buffer_size = get_env<uint64_t>("ROCPROF_ATT_PARAM_BUFFER_SIZE", 0x6000000);
+    // 256MB
+    uint64_t att_param_buffer_size = get_env<uint64_t>("ROCPROF_ATT_PARAM_BUFFER_SIZE", 0x10000000);
     uint64_t att_param_simd_select = get_env<uint64_t>("ROCPROF_ATT_PARAM_SIMD_SELECT", 0xF);
     uint64_t att_param_target_cu   = get_env<uint64_t>("ROCPROF_ATT_PARAM_TARGET_CU", 1);
     uint64_t att_param_perf_ctrl   = get_env<uint64_t>("ROCPROF_ATT_PARAM_PERFCOUNTER_CTRL", 0);

source/lib/rocprofiler-sdk/aql/aql_profile_v2.h

@@ -184,6 +184,14 @@ aqlprofile_get_pmc_info(const aqlprofile_pmc_profile_t* profile,
                         aqlprofile_pmc_info_type_t      attribute,
                         void*                           value);
 
+typedef enum aqlprofile_att_parameter_name_ext_t
+{
+    /**
+     * HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_ATT_BUFFER_SIZE + 1
+     */
+    AQLPROFILE_ATT_PARAMETER_NAME_BUFFER_SIZE_HIGH = 11,
+} aqlprofile_att_parameter_name_ext_t;
+
 // Profile parameter object
 typedef struct
 {

source/lib/rocprofiler-sdk/aql/packet_construct.cpp

@@ -145,7 +145,8 @@ ThreadTraceAQLPacketFactory::ThreadTraceAQLPacketFactory(const hsa::AgentCache&
     uint32_t cu                 = static_cast<uint32_t>(params.target_cu);
     uint32_t shader_engine_mask = static_cast<uint32_t>(params.shader_engine_mask);
     uint32_t simd               = static_cast<uint32_t>(params.simd_select);
-    uint32_t buffer_size        = static_cast<uint32_t>(params.buffer_size);
+    uint32_t buffer_size_lo     = static_cast<uint32_t>(params.buffer_size);
+    uint32_t buffer_size_hi     = static_cast<uint32_t>(params.buffer_size >> 32);
     uint32_t perf_ctrl          = static_cast<uint32_t>(params.perfcounter_ctrl);
 
     aql_params.clear();
@@ -153,7 +154,10 @@ ThreadTraceAQLPacketFactory::ThreadTraceAQLPacketFactory(const hsa::AgentCache&
     aql_params.push_back({HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_COMPUTE_UNIT_TARGET, {cu}});
     aql_params.push_back({HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_SE_MASK, {shader_engine_mask}});
     aql_params.push_back({HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_SIMD_SELECTION, {simd}});
-    aql_params.push_back({HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_ATT_BUFFER_SIZE, {buffer_size}});
+    aql_params.push_back({HSA_VEN_AMD_AQLPROFILE_PARAMETER_NAME_ATT_BUFFER_SIZE, {buffer_size_lo}});
+
+    if(buffer_size_hi != 0) aql_params.push_back({static_cast<hsa_ven_amd_aqlprofile_parameter_name_t>(
+        AQLPROFILE_ATT_PARAMETER_NAME_BUFFER_SIZE_HIGH), {buffer_size_hi}});
 
     if(perf_ctrl != 0 && !params.perfcounters.empty())
     {

source/lib/rocprofiler-sdk/thread_trace/core.cpp

@@ -57,8 +57,7 @@ namespace rocprofiler
 namespace thread_trace
 {
 constexpr size_t   QUEUE_SIZE      = 128;
-constexpr uint64_t MIN_BUFFER_SIZE = 1 << 18;                              // 2 pages per SE
-constexpr uint64_t MAX_BUFFER_SIZE = std::numeric_limits<int32_t>::max();  // aqlprofile limit
+constexpr uint64_t MIN_BUFFER_SIZE = 1 << 20;  // 1MB
 
 struct cbdata_t
 {
@@ -74,15 +73,15 @@ thread_trace_parameter_pack::are_params_valid() const
 {
     if(shader_cb_fn == nullptr)
     {
-        ROCP_WARNING << "Callback cannot be null!";
+        ROCP_CI_LOG(WARNING) << "Callback cannot be null!";
         return false;
     }
 
     if(shader_engine_mask == 0) return false;
 
-    if(buffer_size > MAX_BUFFER_SIZE || buffer_size < MIN_BUFFER_SIZE)
+    if(buffer_size < MIN_BUFFER_SIZE)
     {
-        ROCP_WARNING << "Invalid buffer size: " << buffer_size;
+        ROCP_CI_LOG(WARNING) << "Invalid buffer size: " << buffer_size;
         return false;
     }