From bd042ef1c6131ce1933977a8bde13299708a32dd Mon Sep 17 00:00:00 2001
From: Teng Gao <tegao@microsoft.com>
Date: Mon, 10 May 2021 14:41:11 +0800
Subject: [PATCH 1/2] dump cudaOccDeviceProp

---
 libkineto/src/CudaDeviceProperties.cpp |  3 --
 libkineto/src/CudaDeviceProperties.h   |  4 +++
 libkineto/src/output_json.cpp          | 42 +++++++++++++++++++++++++-
 3 files changed, 45 insertions(+), 4 deletions(-)

diff --git a/libkineto/src/CudaDeviceProperties.cpp b/libkineto/src/CudaDeviceProperties.cpp
index eda37faf0..d5e9252ab 100644
--- a/libkineto/src/CudaDeviceProperties.cpp
+++ b/libkineto/src/CudaDeviceProperties.cpp
@@ -7,10 +7,7 @@
 
 #include "CudaDeviceProperties.h"
 
-#include <vector>
-
 #include <cuda_runtime.h>
-#include <cuda_occupancy.h>
 
 #include "Logger.h"
 
diff --git a/libkineto/src/CudaDeviceProperties.h b/libkineto/src/CudaDeviceProperties.h
index 8156c8b7d..02886b25c 100644
--- a/libkineto/src/CudaDeviceProperties.h
+++ b/libkineto/src/CudaDeviceProperties.h
@@ -7,10 +7,14 @@
 
 #pragma once
 
+#include <vector>
 #include <stdint.h>
+#include <cuda_occupancy.h>
 
 namespace KINETO_NAMESPACE {
 
+const std::vector<cudaOccDeviceProp>& occDeviceProps();
+
 float kernelOccupancy(
     uint32_t deviceId,
     uint16_t registersPerThread,
diff --git a/libkineto/src/output_json.cpp b/libkineto/src/output_json.cpp
index 50577329e..8902ea75f 100644
--- a/libkineto/src/output_json.cpp
+++ b/libkineto/src/output_json.cpp
@@ -33,12 +33,44 @@ static constexpr int kSchemaVersion = 1;
 
 void ChromeTraceLogger::handleTraceStart(
     const std::unordered_map<std::string, std::string>& metadata) {
+  std::string cudaOccDeviceProps = "";
+#ifdef HAS_CUPTI
+  const std::vector<cudaOccDeviceProp>& occProps = KINETO_NAMESPACE::occDeviceProps();
+  if (occProps.size() > 0) {
+    std::ostringstream oss;
+    oss << "[";
+    bool first = true;
+    for (size_t i = 0; i < occProps.size(); i += 1) {
+      const cudaOccDeviceProp& occProp = occProps[i];
+      if (!first) {
+        oss << ", ";
+      }
+      oss << "{";
+      oss << "\"computeMajor\": " << occProp.computeMajor << ", ";
+      oss << "\"computeMinor\": " << occProp.computeMinor << ", ";
+      oss << "\"maxThreadsPerBlock\": " << occProp.maxThreadsPerBlock << ", ";
+      oss << "\"maxThreadsPerMultiprocessor\": " << occProp.maxThreadsPerMultiprocessor << ", ";
+      oss << "\"regsPerBlock\": " << occProp.regsPerBlock << ", ";
+      oss << "\"regsPerMultiprocessor\": " << occProp.regsPerMultiprocessor << ", ";
+      oss << "\"warpSize\": " << occProp.warpSize << ", ";
+      oss << "\"sharedMemPerBlock\": " << occProp.sharedMemPerBlock << ", ";
+      oss << "\"sharedMemPerMultiprocessor\": " << occProp.sharedMemPerMultiprocessor << ", ";
+      oss << "\"numSms\": " << occProp.numSms << ", ";
+      oss << "\"sharedMemPerBlockOptin\": " << occProp.sharedMemPerBlockOptin;
+      oss << "}";
+      first = false;
+    }
+    oss << "]";
+    cudaOccDeviceProps = oss.str();
+  }
+#endif // HAS_CUPTI
+
   traceOf_ << fmt::format(R"JSON(
 {{
   "schemaVersion": {},
   )JSON", kSchemaVersion);
 
-  if (!metadata.empty()) {
+  if (!metadata.empty() || !cudaOccDeviceProps.empty()) {
     traceOf_ << R"JSON(
   "metadata": {
   )JSON";
@@ -50,6 +82,14 @@ void ChromeTraceLogger::handleTraceStart(
       traceOf_ << fmt::format(R"(    "{}": "{}")", kv.first, kv.second);
       first = false;
     }
+    if (!cudaOccDeviceProps.empty()) {
+      if (!first) {
+          traceOf_ << ",\n";
+      }
+      traceOf_ << fmt::format(R"(    "{}": {})", "cudaOccDeviceProps", cudaOccDeviceProps);
+      first = false;
+    }
+
     traceOf_ << R"JSON(
   },
   )JSON";

From 5cee6616214ca2a57e9a10b5128da1ebf7128c92 Mon Sep 17 00:00:00 2001
From: Teng Gao <tegao@microsoft.com>
Date: Tue, 11 May 2021 09:25:16 +0800
Subject: [PATCH 2/2] dump in handleTraceStart

---
 libkineto/src/output_json.cpp | 29 ++++++++++-------------------
 1 file changed, 10 insertions(+), 19 deletions(-)

diff --git a/libkineto/src/output_json.cpp b/libkineto/src/output_json.cpp
index 8902ea75f..7e6bb391b 100644
--- a/libkineto/src/output_json.cpp
+++ b/libkineto/src/output_json.cpp
@@ -33,16 +33,19 @@ static constexpr int kSchemaVersion = 1;
 
 void ChromeTraceLogger::handleTraceStart(
     const std::unordered_map<std::string, std::string>& metadata) {
-  std::string cudaOccDeviceProps = "";
+  traceOf_ << fmt::format(R"JSON(
+{{
+  "schemaVersion": {},
+  )JSON", kSchemaVersion);
+
 #ifdef HAS_CUPTI
   const std::vector<cudaOccDeviceProp>& occProps = KINETO_NAMESPACE::occDeviceProps();
   if (occProps.size() > 0) {
     std::ostringstream oss;
     oss << "[";
-    bool first = true;
     for (size_t i = 0; i < occProps.size(); i += 1) {
       const cudaOccDeviceProp& occProp = occProps[i];
-      if (!first) {
+      if (i > 0) {
         oss << ", ";
       }
       oss << "{";
@@ -58,19 +61,15 @@ void ChromeTraceLogger::handleTraceStart(
       oss << "\"numSms\": " << occProp.numSms << ", ";
       oss << "\"sharedMemPerBlockOptin\": " << occProp.sharedMemPerBlockOptin;
       oss << "}";
-      first = false;
     }
     oss << "]";
-    cudaOccDeviceProps = oss.str();
+    traceOf_ << fmt::format(R"JSON(
+  "cudaOccDeviceProps": {},
+  )JSON", oss.str());
   }
 #endif // HAS_CUPTI
 
-  traceOf_ << fmt::format(R"JSON(
-{{
-  "schemaVersion": {},
-  )JSON", kSchemaVersion);
-
-  if (!metadata.empty() || !cudaOccDeviceProps.empty()) {
+  if (!metadata.empty()) {
     traceOf_ << R"JSON(
   "metadata": {
   )JSON";
@@ -82,14 +81,6 @@ void ChromeTraceLogger::handleTraceStart(
       traceOf_ << fmt::format(R"(    "{}": "{}")", kv.first, kv.second);
       first = false;
     }
-    if (!cudaOccDeviceProps.empty()) {
-      if (!first) {
-          traceOf_ << ",\n";
-      }
-      traceOf_ << fmt::format(R"(    "{}": {})", "cudaOccDeviceProps", cudaOccDeviceProps);
-      first = false;
-    }
-
     traceOf_ << R"JSON(
   },
   )JSON";