[HIP] Hip updates (#677)

libocca · May 4, 2023 · be1f7f3 · be1f7f3
1 parent cbdd074
commit be1f7f3
Show file tree

Hide file tree

Showing 5 changed files with 29 additions and 20 deletions.
diff --git a/cmake/FindOpenCLWrapper.cmake b/cmake/FindOpenCLWrapper.cmake
@@ -16,6 +16,7 @@ if(NOT OpenCL_FOUND)
       ENV NVHPC_ROOT
       ENV SYCL_ROOT
       /usr/local/cuda
+      /opt/rocm/
       /opt/rocm/opencl
       /opt/intel/oneapi/compiler/latest/linux
     PATH_SUFFIXES
@@ -32,6 +33,7 @@ if(NOT OpenCL_FOUND)
       ENV NVHPC_ROOT
       ENV SYCL_ROOT
       /usr/local/cuda
+      /opt/rocm/
       /opt/rocm/opencl
       /opt/intel/oneapi/compiler/latest/linux
     PATH_SUFFIXES

diff --git a/scripts/build/shellTools.sh b/scripts/build/shellTools.sh
@@ -82,8 +82,8 @@ function defaultIncludePath {
     mergedPaths+=":$CPLUS_INCLUDE_PATH"
     mergedPaths+=":$C_INCLUDE_PATH"
     mergedPaths+=":$INCLUDEPATH"
-    mergedPaths+=":/opt/rocm*/opencl/include"
     mergedPaths+=":/opt/rocm*/include"
+    mergedPaths+=":/opt/rocm*/opencl/include"
     mergedPaths+=":/usr/local/cuda*/include"
     mergedPaths+=":/Developer/NVIDIA/CUDA*/include"
     mergedPaths+=":/usr/local/cuda*/targets/*/include/"
@@ -99,8 +99,8 @@ function defaultLibraryPath {
     mergedPaths+=":$OCCA_LIBRARY_PATH"
     mergedPaths+=":$LD_LIBRARY_PATH"
     mergedPaths+=":$DYLD_LIBRARY_PATH"
-    mergedPaths+=":/opt/rocm*/opencl/lib/*"
     mergedPaths+=":/opt/rocm*/lib"
+    mergedPaths+=":/opt/rocm*/opencl/lib/*"
     mergedPaths+=":/usr/local/cuda*/lib*"
     mergedPaths+=":/usr/local/cuda*/lib*/stubs"
     mergedPaths+=":/opt/cuda*/lib*"

diff --git a/src/occa/internal/modes/hip/device.cpp b/src/occa/internal/modes/hip/device.cpp
@@ -72,7 +72,9 @@ namespace occa {
       if (startsWith(arch, "sm_")) {
         archFlag = " -arch=" + arch;
       } else if (startsWith(arch, "gfx")) {
-#if HIP_VERSION >= 305
+#if HIP_VERSION >= 502
+        archFlag = " --offload-arch=" + arch;
+#elif HIP_VERSION >= 305
         archFlag = " --amdgpu-target=" + arch;
 #else
         archFlag = " -t " + arch;
@@ -243,10 +245,12 @@ namespace occa {
       );
 
       if (hipccCompilerFlags.find("-arch=sm") == std::string::npos &&
-#if HIP_VERSION >= 305
-          hipccCompilerFlags.find("-t gfx") == std::string::npos
-#else
+#if HIP_VERSION >= 502
+          hipccCompilerFlags.find("--offload-arch=gfx") == std::string::npos
+#elif HIP_VERSION >= 305
           hipccCompilerFlags.find("--amdgpu-target=gfx") == std::string::npos
+#else
+          hipccCompilerFlags.find("-t gfx") == std::string::npos
 #endif
           ) {
         kernelProps["hipcc_compiler_flags"] += " ";
@@ -320,7 +324,7 @@ namespace occa {
       } else if (verbose) {
           io::stdout << "Output:\n\n" << commandOutput << "\n";
       }
-      
+
       io::sync(binaryFilename);
     }
 

diff --git a/src/occa/internal/modes/hip/kernel.cpp b/src/occa/internal/modes/hip/kernel.cpp
@@ -78,28 +78,30 @@ namespace occa {
         vArgs.resize(args);
       }
 
-      // HIP expects kernel arguments to be byte-aligned so we add padding to arguments
+      // HIP expects kernel arguments to be aligned
       char *dataPtr = (char*) &(vArgs[0]);
-      int padding = 0;
+      int offset = 0;
       for (int i = 0; i < args; ++i) {
         const kernelArgData &arg = arguments[i];
         const udim_t argSize = arg.size();
 
-        size_t bytes;
-        if ((padding + argSize) <= sizeof(void*)) {
-          bytes = argSize;
-          padding = sizeof(void*) - padding - argSize;
+        offset += offset % std::min(static_cast<size_t>(argSize),
+                                    sizeof(void*)); //align
+
+        const void* val;
+        if (!arg.isPointer()) {
+          val = arg.ptr();
         } else {
-          bytes = sizeof(void*);
-          dataPtr += padding;
-          padding = 0;
+          val = &(arg.value.value.ptr);
         }
 
-        ::memcpy(dataPtr, &arg.value.value.int64_, bytes);
-        dataPtr += bytes;
+        if (val) {
+          ::memcpy(dataPtr+offset, val, argSize);
+        }
+        offset += argSize;
       }
 
-      size_t size = vArgs.size() * sizeof(vArgs[0]);
+      size_t size = offset;
       void* config[] = {
         (void*) HIP_LAUNCH_PARAM_BUFFER_POINTER, &(vArgs[0]),
         (void*) HIP_LAUNCH_PARAM_BUFFER_SIZE, &size,

diff --git a/tests/run_examples b/tests/run_examples
@@ -56,6 +56,7 @@ for mode in $("${OCCA_DIR}/bin/occa" modes); do
         Serial) device="{mode: 'Serial'}";;
         OpenMP) device="{mode: 'OpenMP'}";;
         CUDA)   device="{mode: 'CUDA', device_id: 0}";;
+        HIP)    device="{mode: 'HIP', device_id: 0}";;
         OpenCL) device="{mode: 'OpenCL', platform_id: 0, device_id: 0}";;
         Metal)  device="{mode: 'Metal', device_id: 0}";;
         dpcpp) device="{mode: 'dpcpp', platform_id: 0, device_id: 0}";;
@@ -66,7 +67,7 @@ for mode in $("${OCCA_DIR}/bin/occa" modes); do
 
         # Filters
         case "${example_dir}" in
-            
+
             cpp/02_for_loops)
                 if [[ "${mode}" == OpenCL ]]; then
                     continue