Merge pull request #724 from CHIP-SPV/fix-hipcc-path

some additional 1.1 fixes (on main)

CMakeLists.txt

@@ -284,7 +284,7 @@ endif()
 string(TOUPPER "${CMAKE_BUILD_TYPE}" uppercase_CMAKE_BUILD_TYPE)
 
 if(uppercase_CMAKE_BUILD_TYPE STREQUAL "DEBUG")
- add_compile_options(-Wall -gdwarf-4 -O1)
+ add_compile_options(-Wall -O1)
  list(APPEND CHIP_SPV_DEFINITIONS SPDLOG_ACTIVE_LEVEL=SPDLOG_LEVEL_TRACE)
 else()
  list(APPEND CHIP_SPV_DEFINITIONS SPDLOG_ACTIVE_LEVEL=SPDLOG_LEVEL_WARN)

docs/release_notes/chipStar_1.1.rst

@@ -62,6 +62,24 @@ speeding up various workloads significantly. Workloads that do not exploit
 parallelism but enqueue a lot of very small kernels (in the 10's of microseconds
 range) may also benefit as the barrier itself could dominate the execution time.
 
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+Improved portability to other platforms & devices
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The portability of chipStar has been improved & tested with ARM Mali G52 GPU device.
+
+To build chipStar that works with ARM Mali G52:
+
+1) build LLVM and SPIRV-LLVM-Translator as described in the README.md
+
+2) build chipStar with -DCHIP_MALI_GPU_WORKAROUNDS=ON cmake option
+
+There are some limitations. Kernels that require subgroups will not work,
+because chipStar currently implements subgroups with the ``cl_intel_subgroups``
+extension.
+
+
+
 ==============
 Minor Features
 ==============
@@ -76,6 +94,11 @@ Minor Features
 
 * Level Zero command lists are now recycled (#665).
 
+* Improved support for fast-math compilation. Note that this requires using
+ Clang's -ffast-math flag, not CUDA's -use-fast-math flag.
+
+
+
 ==============
 Major Bugfixes
 ==============

scripts/check.py

@@ -86,17 +86,24 @@ def run_cmd(cmd):
 
 cmd = f"{modules} {env_vars} ./hipInfo"
 out, _ = run_cmd(cmd)
-texture_support = "maxTexture1DLinear:" in out and 0 < int(out.split("maxTexture1DLinear:")[1].split("\n")[0].strip())
+texture_support = 0 < int(out.split("maxTexture1DLinear:")[1].split("\n")[0].strip())
+double_support = 0 < int(out.split("arch.hasDoubles:")[1].split("\n")[0].strip())
+if double_support:
+ double_cmd = ""
+else:
+ double_cmd = "|[Dd]ouble"
 if not texture_support:
  texture_cmd = "|[Tt]ex"
 else:
  texture_cmd = ""
 
+
+
 if args.categories:
- cmd_deviceFunc = f"{modules} {env_vars} ctest --output-on-failure --timeout {args.timeout} --repeat until-fail:{args.num_tries} -j 100 -E \"`cat ./test_lists/{args.device_type}_{args.backend}_failed_{level0_cmd_list}tests.txt`{texture_cmd}\" -R deviceFunc -O checkpy_{args.device_type}_{args.backend}_device.txt"
- cmd_graph = f"{modules} {env_vars} ctest --output-on-failure --timeout {args.timeout} --repeat until-fail:{args.num_tries} -j 100 -E \"`cat ./test_lists/{args.device_type}_{args.backend}_failed_{level0_cmd_list}tests.txt`{texture_cmd}\" -R \"[Gg]raph\" -O checkpy_{args.device_type}_{args.backend}_graph.txt"
- cmd_single = f"{modules} {env_vars} ctest --output-on-failure --timeout {args.timeout} --repeat until-fail:{args.num_tries} -j 1 -E \"`cat ./test_lists/{args.device_type}_{args.backend}_failed_{level0_cmd_list}tests.txt`{texture_cmd}\" -R \"`cat ./test_lists/non_parallel_tests.txt`\" -O checkpy_{args.device_type}_{args.backend}_single.txt"
- cmd_other = f"{modules} {env_vars} ctest --output-on-failure --timeout {args.timeout} --repeat until-fail:{args.num_tries} -j {args.num_threads} -E \"`cat ./test_lists/{args.device_type}_{args.backend}_failed_{level0_cmd_list}tests.txt`{texture_cmd}|deviceFunc|[Gg]raph|`cat ./test_lists/non_parallel_tests.txt`\" -O checkpy_{args.device_type}_{args.backend}_other.txt"
+ cmd_deviceFunc = f"{modules} {env_vars} ctest --output-on-failure --timeout {args.timeout} --repeat until-fail:{args.num_tries} -j 100 -E \"`cat ./test_lists/{args.device_type}_{args.backend}_failed_{level0_cmd_list}tests.txt`{texture_cmd}{double_cmd}\" -R deviceFunc -O checkpy_{args.device_type}_{args.backend}_device.txt"
+ cmd_graph = f"{modules} {env_vars} ctest --output-on-failure --timeout {args.timeout} --repeat until-fail:{args.num_tries} -j 100 -E \"`cat ./test_lists/{args.device_type}_{args.backend}_failed_{level0_cmd_list}tests.txt`{texture_cmd}{double_cmd}\" -R \"[Gg]raph\" -O checkpy_{args.device_type}_{args.backend}_graph.txt"
+ cmd_single = f"{modules} {env_vars} ctest --output-on-failure --timeout {args.timeout} --repeat until-fail:{args.num_tries} -j 1 -E \"`cat ./test_lists/{args.device_type}_{args.backend}_failed_{level0_cmd_list}tests.txt`{texture_cmd}{double_cmd}\" -R \"`cat ./test_lists/non_parallel_tests.txt`\" -O checkpy_{args.device_type}_{args.backend}_single.txt"
+ cmd_other = f"{modules} {env_vars} ctest --output-on-failure --timeout {args.timeout} --repeat until-fail:{args.num_tries} -j {args.num_threads} -E \"`cat ./test_lists/{args.device_type}_{args.backend}_failed_{level0_cmd_list}tests.txt`{texture_cmd}{double_cmd}|deviceFunc|[Gg]raph|`cat ./test_lists/non_parallel_tests.txt`\" -O checkpy_{args.device_type}_{args.backend}_other.txt"
  if(args.dry_run):
  print(cmd_deviceFunc)
  print(cmd_graph)

scripts/unit_tests.sh

@@ -2,8 +2,15 @@
 
 set -e
 
+# read the file /opt/actions-runner/num-threads.txt and set the number of threads to the value in the file
+# if the file does not exist, set the number of threads to 24
+if [ -f "/opt/actions-runner/num-threads.txt" ]; then
+ num_threads=$(cat /opt/actions-runner/num-threads.txt)
+else
+ num_threads=24
+fi
+
 num_tries=1
-num_threads=24
 timeout=200
 
 # Check if at least one argument is provided
@@ -23,13 +30,13 @@ build_type=$(echo "$1" | tr '[:lower:]' '[:upper:]')
 
 if [ "$2" == "llvm-15" ]; then
  LLVM=llvm-15
- CLANG=clang/clang15-spirv-omp
+ CLANG=llvm/15.0/dynamic
 elif [ "$2" == "llvm-16" ]; then
  LLVM=llvm-16
- CLANG=clang/clang16-spirv-omp
+ CLANG=llvm/16.0/dynamic
 elif [ "$2" == "llvm-17" ]; then
  LLVM=llvm-17
- CLANG=clang/clang17-spirv-omp
+ CLANG=llvm/17.0/dynamic
 else
  echo "$2"
  echo "Invalid 2nd argument. Use either 'llvm-15', 'llvm-16' or 'llvm-17'."
@@ -95,10 +102,9 @@ export IGC_EnableDPEmulation=1
 export OverrideDefaultFP64Settings=1
 export CHIP_LOGLEVEL=err
 export POCL_KERNEL_CACHE=0
-export CHIP_L0_COLLECT_EVENTS_TIMEOUT=30
 
 # Use OpenCL for building/test discovery to prevent Level Zero from being used in multi-thread/multi-process environment
-module load $CLANG intel/opencl # leave intel/opencl loaded otherwise hip_sycl_interop samples segfault upon exit
+module load $CLANG opencl/dgpu # leave intel/opencl loaded otherwise hip_sycl_interop samples segfault upon exit
 
 output=$(clinfo -l 2>&1 | grep "Platform #0")
 echo $output
@@ -142,6 +148,8 @@ else
  # ../scripts/compile_libceed.sh ${CHIPSTAR_INSTALL_DIR}
 fi
 
+module unload opencl/dgpu
+
 # module load HIP/hipBLAS/main/release # for libCEED NOTE: Must be after build step otherwise it will cause link issues.
 
 # Test Level Zero Regular Cmd Lists iGPU

src/Utils.cc

@@ -27,6 +27,11 @@
 #include <fstream>
 #include <random>
 
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+#include <link.h>
+
 bool isConvertibleToInt(const std::string &str) {
  try {
  std::stoi(str);
@@ -132,23 +137,41 @@ std::optional<std::string> readFromFile(const fs::path Path) {
  return std::nullopt;
 }
 
+static int dlIterateCallback(struct dl_phdr_info *Info,
+ size_t Size, void *Data) {
+ std::string *Res = static_cast<std::string *>(Data);
+ std::string DlName(Info->dlpi_name);
+ size_t Pos = DlName.find("/libCHIP.so");
+ if (Pos == std::string::npos)
+ return 0;
+
+ DlName.erase(Pos);
+ Res->assign(DlName);
+ return 1;
+}
+
 std::optional<fs::path> getHIPCCPath() {
  static std::once_flag Flag;
  static std::optional<fs::path> HIPCCPath;
 
- std::call_once(Flag, []() {
+ std::string LibCHIPPath("/dev/null");
+ dl_iterate_phdr(dlIterateCallback, static_cast<void*>(&LibCHIPPath));
+
+ std::call_once(Flag, [&]() {
  for (const auto &ExeCand : {
+ fs::path(LibCHIPPath) / "bin/hipcc",
 #if !CHIP_DEBUG_BUILD
  fs::path(CHIP_INSTALL_DIR) / "bin/hipcc",
 #endif
-  fs::path(CHIP_BUILD_DIR) / "bin/hipcc"
+ fs::path(CHIP_BUILD_DIR) / "bin/hipcc"
  })
  if (canExecuteHipcc(ExeCand)) {
  HIPCCPath = ExeCand;
  return;
  }
  });
 
+ logDebug("HIPCC path: {}", HIPCCPath->c_str());
  return HIPCCPath;
 }