merge amd-mainline into rocm-rel-5.3

Change-Id: Ia2120bb38526f8cd864762faba4b24f9d0458217
ROCm · Aug 22, 2022 · 4b32c52 · 4b32c52
2 parents 3b6d104 + dfdd3b8
commit 4b32c52
Show file tree

Hide file tree

Showing 26 changed files with 3,194 additions and 3,224 deletions.
diff --git a/README.md b/README.md
diff --git a/bin/hipify-perl b/bin/hipify-perl
diff --git a/doc/markdown/CUDA_Driver_API_functions_supported_by_HIP.md b/doc/markdown/CUDA_Driver_API_functions_supported_by_HIP.md
diff --git a/doc/markdown/CUDA_RTC_API_supported_by_HIP.md b/doc/markdown/CUDA_RTC_API_supported_by_HIP.md
@@ -1,4 +1,4 @@
-# CUDA_RTC_API_supported_by_HIP API supported by HIP
+# CUDA RTC API supported by HIP
 
 ## **1. RTC Data types**
 

diff --git a/doc/markdown/CUDA_Runtime_API_functions_supported_by_HIP.md b/doc/markdown/CUDA_Runtime_API_functions_supported_by_HIP.md
diff --git a/src/CUDA2HIP_Driver_API_functions.cpp b/src/CUDA2HIP_Driver_API_functions.cpp
diff --git a/src/CUDA2HIP_Driver_API_types.cpp b/src/CUDA2HIP_Driver_API_types.cpp
diff --git a/src/CUDA2HIP_Perl.cpp b/src/CUDA2HIP_Perl.cpp
@@ -345,7 +345,7 @@ namespace perl {
   void generateExperimentalSubstitutions(unique_ptr<ostream> &streamPtr) {
     *streamPtr.get() << endl << sub << sExperimentalSubstitutions << " {" << endl;
     for (int i = 0; i < NUM_CONV_TYPES; ++i) {
-      if (i == CONV_INCLUDE_CUDA_MAIN_H || i == CONV_INCLUDE) {
+      if (i == CONV_INCLUDE_CUDA_MAIN_H || i == CONV_INCLUDE_CUDA_MAIN_V2_H || i == CONV_INCLUDE) {
         for (auto &ma : CUDA_INCLUDE_MAP) {
           if (!Statistics::isHipExperimental(ma.second)) continue;
           if (i == ma.second.type) {
@@ -373,7 +373,7 @@ namespace perl {
     bool bTranslateToRoc = TranslateToRoc;
     TranslateToRoc = true;
     for (int i = 0; i < NUM_CONV_TYPES; ++i) {
-      if (i == CONV_INCLUDE_CUDA_MAIN_H) {
+      if (i == CONV_INCLUDE_CUDA_MAIN_H || i == CONV_INCLUDE_CUDA_MAIN_V2_H) {
         for (auto &ma : CUDA_INCLUDE_MAP) {
           if (i == ma.second.type) {
             string sCUDA = ma.first.str();
@@ -400,7 +400,7 @@ namespace perl {
   void generateSimpleSubstitutions(unique_ptr<ostream> &streamPtr) {
     *streamPtr.get() << endl << sub << sSimpleSubstitutions << " {" << endl;
     for (int i = 0; i < NUM_CONV_TYPES; ++i) {
-      if (i == CONV_INCLUDE_CUDA_MAIN_H || i == CONV_INCLUDE) {
+      if (i == CONV_INCLUDE_CUDA_MAIN_H || i == CONV_INCLUDE_CUDA_MAIN_V2_H || i == CONV_INCLUDE) {
         for (auto &ma : CUDA_INCLUDE_MAP) {
           if (Statistics::isUnsupported(ma.second) || Statistics::isHipExperimental(ma.second)) continue;
           if (i == ma.second.type) {

diff --git a/src/CUDA2HIP_Runtime_API_functions.cpp b/src/CUDA2HIP_Runtime_API_functions.cpp
diff --git a/src/CUDA2HIP_Runtime_API_types.cpp b/src/CUDA2HIP_Runtime_API_types.cpp
diff --git a/src/Statistics.cpp b/src/Statistics.cpp
@@ -537,6 +537,7 @@ std::string Statistics::getHipVersion(const hipVersions& ver) {
     case HIP_5010: return "5.1.0";
     case HIP_5011: return "5.1.1";
     case HIP_5020: return "5.2.0";
+    case HIP_5030: return "5.3.0";
   }
   return "";
 }

diff --git a/src/Statistics.h b/src/Statistics.h
@@ -289,7 +289,8 @@ enum hipVersions {
   HIP_5010 = 5010,
   HIP_5011 = 5011,
   HIP_5020 = 5020,
-  HIP_LATEST = HIP_5020,
+  HIP_5030 = 5030,
+  HIP_LATEST = HIP_5030,
 };
 
 struct cudaAPIversions {

diff --git a/tests/lit.cfg b/tests/lit.cfg
@@ -55,6 +55,9 @@ if config.cuda_version_major < 9:
     config.excludes.append('cuSPARSE_07.cu')
     config.excludes.append('benchmark_curand_kernel.cpp')
     config.excludes.append('reinterpret_cast.cu')
+    config.excludes.append('cub_01.cu')
+    config.excludes.append('cub_02.cu')
+    config.excludes.append('cub_03.cu')
 if config.cuda_version_major < 10:
     config.excludes.append('cuSPARSE_08.cu')
     config.excludes.append('cuSPARSE_09.cu')
@@ -76,6 +79,9 @@ if config.cuda_version_major > 10:
     config.excludes.append('cuSPARSE_02.cu')
     config.excludes.append('cuSPARSE_12.cu')
 
+if config.cuda_version_major < 11 or (config.cuda_version_major == 11 and config.cuda_version_minor < 1):
+    config.excludes.append('runtime_functions_11010.cu')
+
 # name: The name of this test suite.
 config.name = 'hipify'
 

diff --git a/tests/unit_tests/libraries/cuDNN/cudnn_convolution_forward.cu b/tests/unit_tests/libraries/cuDNN/cudnn_convolution_forward.cu
@@ -1,4 +1,4 @@
-// RUN: %run_test hipify "%s" "%t" --skip-excluded-preprocessor-conditional-blocks %hipify_args %clang_args
+// RUN: %run_test hipify "%s" "%t" %hipify_args 2 --hip-kernel-execution-syntax --skip-excluded-preprocessor-conditional-blocks %clang_args
 
 #include <iomanip>
 #include <iostream>

diff --git a/tests/unit_tests/synthetic/driver_defines.cu b/tests/unit_tests/synthetic/driver_defines.cu
@@ -10,13 +10,6 @@ int main() {
   // CHECK: #define __HIPCC__;
   #define __CUDACC__;
 
-#if CUDA_VERSION > 7050
-  // CHECK: int DEVICE_CPU = hipCpuDeviceId;
-  // CHECK-NEXT: int DEVICE_INVALID = hipInvalidDeviceId;
-  int DEVICE_CPU = CU_DEVICE_CPU;
-  int DEVICE_INVALID = CU_DEVICE_INVALID;
-#endif
-
   // CHECK: int IPC_HANDLE_SIZE = HIP_IPC_HANDLE_SIZE;
   int IPC_HANDLE_SIZE = CU_IPC_HANDLE_SIZE;
 
@@ -37,10 +30,6 @@ int main() {
   int MEMHOSTALLOC_WRITECOMBINED = CU_MEMHOSTALLOC_WRITECOMBINED;
   int MEMHOSTREGISTER_PORTABLE = CU_MEMHOSTREGISTER_PORTABLE;
   int MEMHOSTREGISTER_DEVICEMAP = CU_MEMHOSTREGISTER_DEVICEMAP;
-#if CUDA_VERSION > 7000
-  // CHECK: int MEMHOSTREGISTER_IOMEMORY = hipHostRegisterIoMemory;
-  int MEMHOSTREGISTER_IOMEMORY = CU_MEMHOSTREGISTER_IOMEMORY;
-#endif
 
   // CHECK: int TRSA_OVERRIDE_FORMAT = HIP_TRSA_OVERRIDE_FORMAT;
   // CHECK-NEXT: int TRSF_NORMALIZED_COORDINATES = HIP_TRSF_NORMALIZED_COORDINATES;
@@ -60,15 +49,27 @@ int main() {
   int ARRAY3D_CUBEMAP = CUDA_ARRAY3D_CUBEMAP;
   int ARRAY3D_TEXTURE_GATHER = CUDA_ARRAY3D_TEXTURE_GATHER;
 
-#if CUDA_VERSION > 8000
+  // CHECK: hipStream_t STREAM_PER_THREAD = hipStreamPerThread;
+  CUstream STREAM_PER_THREAD = CU_STREAM_PER_THREAD;
+
+#if CUDA_VERSION >= 7050
+  // CHECK: int MEMHOSTREGISTER_IOMEMORY = hipHostRegisterIoMemory;
+  int MEMHOSTREGISTER_IOMEMORY = CU_MEMHOSTREGISTER_IOMEMORY;
+#endif
+
+#if CUDA_VERSION >= 8000
+  // CHECK: int DEVICE_CPU = hipCpuDeviceId;
+  // CHECK-NEXT: int DEVICE_INVALID = hipInvalidDeviceId;
+  int DEVICE_CPU = CU_DEVICE_CPU;
+  int DEVICE_INVALID = CU_DEVICE_INVALID;
+#endif
+
+#if CUDA_VERSION >= 9000
   // CHECK: int COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_PRE_LAUNCH_SYNC = hipCooperativeLaunchMultiDeviceNoPreSync;
   // CHECK-NEXT: int COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_POST_LAUNCH_SYNC = hipCooperativeLaunchMultiDeviceNoPostSync;
   int COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_PRE_LAUNCH_SYNC = CUDA_COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_PRE_LAUNCH_SYNC;
   int COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_POST_LAUNCH_SYNC = CUDA_COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_POST_LAUNCH_SYNC;
 #endif
 
-  // CHECK: hipStream_t STREAM_PER_THREAD = hipStreamPerThread;
-  CUstream STREAM_PER_THREAD = CU_STREAM_PER_THREAD;
-
   return 0;
 }