From f58df363b4ae81446f47085e15732d681b3f799d Mon Sep 17 00:00:00 2001
From: Hayden Laccabue <hlaccabu@amd.com>
Date: Fri, 13 Sep 2024 14:47:10 -0700
Subject: [PATCH] Creating initial support for NPU3 no-op tests throughput and
 latency

Signed-off-by: Hayden Laccabue <hlaccabu@amd.com>
---
 src/runtime_src/core/common/query_requests.h  |  5 +-
 .../tools/common/tests/TestNPULatency.cpp     | 48 ++++++++++++++++---
 .../tools/common/tests/TestNPUThroughput.cpp  | 48 ++++++++++++++++---
 3 files changed, 88 insertions(+), 13 deletions(-)

diff --git a/src/runtime_src/core/common/query_requests.h b/src/runtime_src/core/common/query_requests.h
index bb7da05f800..bbdf6d29887 100644
--- a/src/runtime_src/core/common/query_requests.h
+++ b/src/runtime_src/core/common/query_requests.h
@@ -568,7 +568,8 @@ struct sequence_name : request
     tct_one_column,
     tct_all_column,
     gemm_int8,
-    aie_reconfig_overhead
+    aie_reconfig_overhead,
+    nop
   };
 
   static std::string
@@ -585,6 +586,8 @@ struct sequence_name : request
         return "gemm_int8";
       case type::aie_reconfig_overhead:
         return "aie_reconfig_overhead";
+      case type::nop:
+        return "nop";
     }
     return "unknown";
   }
diff --git a/src/runtime_src/core/tools/common/tests/TestNPULatency.cpp b/src/runtime_src/core/tools/common/tests/TestNPULatency.cpp
index 58ff7c3c813..8a0e2719003 100644
--- a/src/runtime_src/core/tools/common/tests/TestNPULatency.cpp
+++ b/src/runtime_src/core/tools/common/tests/TestNPULatency.cpp
@@ -9,7 +9,11 @@
 #include "xrt/xrt_device.h"
 #include "xrt/xrt_hw_context.h"
 #include "xrt/xrt_kernel.h"
+#include "experimental/xrt_elf.h"
+#include "experimental/xrt_ext.h"
+#include "experimental/xrt_module.h"
 namespace XBU = XBUtilities;
+namespace xq = xrt_core::query;
 
 #include <filesystem>
 
@@ -70,14 +74,46 @@ TestNPULatency::run(std::shared_ptr<xrt_core::device> dev)
 
   xrt::hw_context hwctx;
   xrt::kernel testker;
-  try {
+  if(xrt_core::device_query<xq::rom_vbnv>(dev).find("umq") != std::string::npos) {
+    //Elf flow
+    const auto elf_name = xrt_core::device_query<xrt_core::query::sequence_name>(dev, xrt_core::query::sequence_name::type::nop);
+    auto elf_path = findPlatformFile(elf_name, ptree);
+    if (!std::filesystem::exists(elf_path))
+      return ptree;
+
+    logger(ptree, "Elf", elf_path);
+
+    xrt::elf elf;
+    try {
+      elf = xrt::elf(elf_path);
+    }
+    catch (const std::runtime_error& ex) {
+      logger(ptree, "Error", ex.what());
+      ptree.put("status", test_token_failed);
+      return ptree;
+    }
+
+    xrt::module mod{elf};
+    try{
+      hwctx = xrt::hw_context{working_dev, xclbin.get_uuid()};
+      testker = xrt::ext::kernel{hwctx, mod, kernelName};
+    }
+    catch (const std::exception& ex){
+      logger(ptree, "Error", ex.what());
+      ptree.put("status", test_token_failed);
+      return ptree;
+    }
+  }
+  else {
+    try {
     hwctx = xrt::hw_context(working_dev, xclbin.get_uuid());
     testker = xrt::kernel(hwctx, kernelName);
-  }
-  catch (const std::exception& ex){
-    logger(ptree, "Error", ex.what());
-    ptree.put("status", test_token_failed);
-    return ptree;
+    }
+    catch (const std::exception& ex){
+      logger(ptree, "Error", ex.what());
+      ptree.put("status", test_token_failed);
+      return ptree;
+    }
   }
 
   //Create BOs, the values are not initialized as they are not really used by this special test running on the device
diff --git a/src/runtime_src/core/tools/common/tests/TestNPUThroughput.cpp b/src/runtime_src/core/tools/common/tests/TestNPUThroughput.cpp
index ceffe31986a..c862bd121b8 100644
--- a/src/runtime_src/core/tools/common/tests/TestNPUThroughput.cpp
+++ b/src/runtime_src/core/tools/common/tests/TestNPUThroughput.cpp
@@ -9,7 +9,11 @@
 #include "xrt/xrt_device.h"
 #include "xrt/xrt_hw_context.h"
 #include "xrt/xrt_kernel.h"
+#include "experimental/xrt_elf.h"
+#include "experimental/xrt_ext.h"
+#include "experimental/xrt_module.h"
 namespace XBU = XBUtilities;
+namespace xq = xrt_core::query;
 
 #include <filesystem>
 
@@ -69,14 +73,46 @@ TestNPUThroughput::run(std::shared_ptr<xrt_core::device> dev)
 
   xrt::hw_context hwctx;
   xrt::kernel testker;
-  try {
+  if(xrt_core::device_query<xq::rom_vbnv>(dev).find("umq") != std::string::npos) {
+    //Elf flow
+    const auto elf_name = xrt_core::device_query<xrt_core::query::sequence_name>(dev, xrt_core::query::sequence_name::type::nop);
+    auto elf_path = findPlatformFile(elf_name, ptree);
+    if (!std::filesystem::exists(elf_path))
+      return ptree;
+      
+    logger(ptree, "Elf", elf_path);
+
+    xrt::elf elf;
+    try {
+      elf = xrt::elf(elf_path);
+    }
+    catch (const std::runtime_error& ex) {
+      logger(ptree, "Error", ex.what());
+      ptree.put("status", test_token_failed);
+      return ptree;
+    }
+
+    xrt::module mod{elf};
+    try{
+      hwctx = xrt::hw_context{working_dev, xclbin.get_uuid()};
+      testker = xrt::ext::kernel{hwctx, mod, kernelName};
+    }
+    catch (const std::exception& ex){
+      logger(ptree, "Error", ex.what());
+      ptree.put("status", test_token_failed);
+      return ptree;
+    }
+  }
+  else {
+    try {
     hwctx = xrt::hw_context(working_dev, xclbin.get_uuid());
     testker = xrt::kernel(hwctx, kernelName);
-  }
-  catch (const std::exception& ex){
-    logger(ptree, "Error", ex.what());
-    ptree.put("status", test_token_failed);
-    return ptree;
+    }
+    catch (const std::exception& ex){
+      logger(ptree, "Error", ex.what());
+      ptree.put("status", test_token_failed);
+      return ptree;
+    }
   }
 
   //Create BOs, the values are not initialized as they are not really used by this special test running on the device