From 838b42153bffdb5cbc1209ea0c409859f5be9c0f Mon Sep 17 00:00:00 2001
From: Evgeny <evgeny.shcherbakov@amd.com>
Date: Mon, 13 Nov 2017 15:40:39 -0600
Subject: [PATCH] Adding trace data output file, trace_data.sqtt

---
 test/ctrl/tool.cpp | 45 +++++++++++++++++++++++++++++++++------------
 test/input.xml     |  2 +-
 test/metrics.xml   |  4 ++--
 3 files changed, 36 insertions(+), 15 deletions(-)

diff --git a/test/ctrl/tool.cpp b/test/ctrl/tool.cpp
index 570a738916..a849b7c892 100644
--- a/test/ctrl/tool.cpp
+++ b/test/ctrl/tool.cpp
@@ -1,7 +1,11 @@
+#include <fcntl.h>
 #include <hsa.h>
 #include <pthread.h>
 #include <stdio.h>
 #include <stdlib.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+
 #include <iostream>
 #include <map>
 #include <vector>
@@ -49,9 +53,11 @@ hsa_status_t trace_data_cb(
   hsa_ven_amd_aqlprofile_info_data_t* info_data,
   void* data)
 {
+  FILE* file = reinterpret_cast<FILE*>(data);
   hsa_status_t status = HSA_STATUS_SUCCESS;
   if (info_type == HSA_VEN_AMD_AQLPROFILE_INFO_SQTT_DATA) {
-    printf("    data ptr (%p), size(%u)\n", info_data->sqtt_data.ptr, info_data->sqtt_data.size);
+    fprintf(file, "    data ptr (%p), size(%u)\n", info_data->sqtt_data.ptr, info_data->sqtt_data.size);
+
   } else status = HSA_STATUS_ERROR;
   return status;
 }
@@ -68,23 +74,38 @@ void print_info(FILE* file, const rocprofiler_info_t* info, const unsigned info_
         fprintf(file, "(%lu)\n", p->data.result_int64);
         break;
       case ROCPROFILER_BYTES: {
-        fprintf(file, "(\n");
         if (p->data.result_bytes.copy) {
-          fprintf(file, "    system memory copy\n");
-          const char* ptr = reinterpret_cast<const char*>(p->data.result_bytes.ptr);
           uint64_t size = 0;
+
+          const char* ptr = reinterpret_cast<const char*>(p->data.result_bytes.ptr);
           for (unsigned i = 0; i < p->data.result_bytes.instance_count; ++i) {
-            size = *reinterpret_cast<const uint64_t*>(ptr);
-            const char* data = ptr + sizeof(size);
-            fprintf(file, "    data (%p), size (%lu)\n", data, size);
-            size = align_size(size, sizeof(uint64_t));
-            ptr = data + size;
+            uint64_t chunk_size = *reinterpret_cast<const uint64_t*>(ptr);
+            const char* data = ptr + sizeof(uint64_t);
+            chunk_size = align_size(chunk_size, sizeof(uint64_t));
+            ptr = data + chunk_size;
+            size += chunk_size;
+          }
+          fprintf(file, "size(%lu)\n", size);
+          if (size > p->data.result_bytes.size) {
+            fprintf(stderr, "SQTT data size is out of the result buffer size\n");
+            exit(1);
+          }
+
+          const int fd = open("trace_data.sqtt", O_RDWR|O_CREAT|O_TRUNC, 0640);
+          if (fd == -1) {
+            perror("open 'trace_data.sqtt'");
+            exit(1);
+          }
+          const size_t write_size = write(fd, p->data.result_bytes.ptr, size);
+          if (write_size < size) {
+            fprintf(stderr, "SQTT data size(%lu) write failed, returned size(%zd)\n", size, write_size);
+            exit(1);
           }
         } else {
-          fprintf(file, "    local memory buffer\n");
-          rocprofiler_iterate_trace_data(context, trace_data_cb, NULL);
+          fprintf(file, "iterate GPU local memory (\n");
+          rocprofiler_iterate_trace_data(context, trace_data_cb, reinterpret_cast<void*>(file));
+          fprintf(file, "  )\n");
         }
-        fprintf(file, "  )\n");
         break;
       }
       default:
diff --git a/test/input.xml b/test/input.xml
index 7bfd9b81e0..2d4924c277 100644
--- a/test/input.xml
+++ b/test/input.xml
@@ -1,4 +1,4 @@
-<metric name=SQ_CYCLES,SQ_WAVES,SQ_INSTS_SMEM,SQ_INSTS_VALU,TA_FLAT_WRITE_WAVEFRONTS[0],TA_FLAT_WRITE_WAVEFRONTS[1],CPC_ALWAYS_COUNT,CPC_ME1_STALL_WAIT_ON_RCIU_READ,GPU_BUSY,MEM_BUSY_AVR,SFETCH_INSTS,VALU_INSTS,FETCH_SIZE,VWRITE_INSTS ></metric>
+<metric name=SQ_CYCLES,SQ_WAVES,SQ_INSTS_SMEM,SQ_INSTS_VALU,TA_FLAT_WRITE_WAVEFRONTS[0],TA_FLAT_WRITE_WAVEFRONTS[1],CPC_ALWAYS_COUNT,CPC_ME1_STALL_WAIT_ON_RCIU_READ,GPU_BUSY,MEM_BUSY,SFETCH_INSTS,VALU_INSTS,FETCH_SIZE,VWRITE_INSTS ></metric>
 
 <trace name=SQTT copy=true >
   <parameters
diff --git a/test/metrics.xml b/test/metrics.xml
index 137e86af22..6bd37d507d 100644
--- a/test/metrics.xml
+++ b/test/metrics.xml
@@ -123,9 +123,9 @@
   # The percentage of time GPU was busy.
   <metric name=GPU_BUSY expr=100*GRBM_GUI_ACTIVE/GRBM_COUNT ></metric>
 
-  # MEM_BUSY_AVR, percentage
+  # MEM_BUSY, percentage
   # The percentage of GPUTime the memory unit is active. The result includes the stall time (MemUnitStalled). This is measured with all extra fetches and writes and any cache or memory effects taken into account. Value range: 0% to 100% (fetch-bound).
-  <metric name=MEM_BUSY_AVR expr=100*TA_BUSY_avr/GRBM_GUI_ACTIVE ></metric>
+  <metric name=MEM_BUSY expr=100*TA_BUSY_avr/GRBM_GUI_ACTIVE ></metric>
 
   # VWRITE_INSTS
   # The average number of vector write instructions to the video memory executed per work-item (affected by flow control). Excludes FLAT instructions that write to video memory.