Make global workitems multiple of local workitems (#1976)

HIP requires global work items in multiple of local work items. If it is not it is not guaranteed to generate correct results all the time.
Fixes #1977
Fixes #1644
MIGraphX CI has moved to rocm-5.6 which doesn't require hipRTC workarounds

.github/workflows/ci.yaml

@@ -323,6 +323,7 @@ jobs:
       env:
         GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
       run: |
+        set +x
         gh extension install actions/gh-actions-cache --pin v1.0.1
         gh actions-cache delete ${{ steps.ccache_restore.outputs.cache-matched-key }} --confirm
 
@@ -439,6 +440,7 @@ jobs:
       env:
         GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
       run: |
+        set +x
         gh extension install actions/gh-actions-cache
         gh actions-cache delete ${{ steps.ccache_restore_fpga.outputs.cache-matched-key }} --confirm
       continue-on-error: true

Jenkinsfile

@@ -15,13 +15,11 @@ def rocmtestnode(Map conf) {
         def compiler = bconf.get("compiler", "/opt/rocm/llvm/bin/clang++")
         def flags = bconf.get("flags", "")
         def gpu_debug = bconf.get("gpu_debug", "0")
-        def hiprtc_workarounds = bconf.get("hiprtc_workarounds", "0")
         def cmd = """
             ulimit -c unlimited
             echo "leak:dnnl::impl::malloc" > suppressions.txt
             export LSAN_OPTIONS="suppressions=\$(pwd)/suppressions.txt"
             export MIGRAPHX_GPU_DEBUG=${gpu_debug}
-            export MIGRAPHX_ENABLE_HIPRTC_WORKAROUNDS=${hiprtc_workarounds}
             export CXX=${compiler}
             export CXXFLAGS='-Werror'
             env
@@ -109,7 +107,7 @@ rocmtest clang_debug: rocmnode('cdna') { cmake_build ->
     stage('hipRTC Debug') {
         def sanitizers = "undefined"
         def debug_flags = "-g -O2 -fsanitize=${sanitizers} -fno-sanitize-recover=${sanitizers}"
-        cmake_build(flags: "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_PYTHON=Off -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}' -DCMAKE_C_FLAGS_DEBUG='${debug_flags}' -DMIGRAPHX_USE_HIPRTC=On", gpu_debug: true, hiprtc_workarounds: true)
+        cmake_build(flags: "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_PYTHON=Off -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}' -DCMAKE_C_FLAGS_DEBUG='${debug_flags}' -DMIGRAPHX_USE_HIPRTC=On", gpu_debug: true)
     }
 }, clang_release: rocmnode('cdna') { cmake_build ->
     stage('Hip Clang Release') {

src/targets/gpu/compile_hip_code_object.cpp

@@ -135,14 +135,13 @@ compute_global_for(context& ctx, std::size_t n, std::size_t over)
     std::size_t max_global = ctx.get_current_device().get_cu_count() *
                              ctx.get_current_device().get_max_workitems_per_cu();
     return [n, over, max_global](std::size_t local) {
-        std::size_t num_elements = n;
+        // hip require global workitems multiple of local workitems. It may degrade performance.
+        // [TODO]: consider adding "fno-hip-uniform-block" flag when it becomes available.
+        // https://reviews.llvm.org/D155213
+        std::size_t num_elements = ((n + local - 1) / local) * local;
         std::size_t groups       = (num_elements + local - 1) / local;
         std::size_t max_blocks   = max_global / local;
         std::size_t nglobal      = std::min(max_blocks * over, groups) * local;
-#ifdef MIGRAPHX_USE_HIPRTC
-        if(enabled(MIGRAPHX_ENABLE_HIPRTC_WORKAROUNDS{}))
-            num_elements = ((num_elements + local - 1) / local) * local;
-#endif
         return std::min(nglobal, num_elements);
     };
 }