Merge branch 'aomp-epsdb' into aomp-epsdb-mainline

Change-Id: I95a6ab1882780ef432c210ced4b65b4b9110eabe

examples/tools/ompt/README.md

@@ -0,0 +1,54 @@
+OMPT target support: Examples to demonstrate how a tool would use the OMPT target APIs
+=======================================================================================
+
+The examples simulate how a tool is expected to use OMPT target
+support. The tool would register callbacks and call OMPT runtime entry
+points to start and stop device tracing, if required. The tool would
+have an OpenMP thread call these runtime entry points to control
+device tracing. When certain events occur, the OpenMP runtime would
+invoke the event callbacks so that the tool can establish the event
+context. If device tracing has been requested, the OpenMP runtime
+would collect and manage trace records in buffers. When a buffer fills
+up or if an OpenMP thread requests explicit flushing of trace records,
+an OpenMP runtime helper thread would invoke a buffer-completion
+callback. The buffer-completion callback is implemented by the tool
+and would typically traverse the trace records returned as part of the
+callback. Once the trace records are returned, they can be correlated
+to the context established earlier through the event callbacks.
+
+Here are the steps:
+(1) The tool has to define a function called ompt_start_tool with
+C-linkage and the appropriate signature as defined by the OpenMP
+spec. This function provides 2 function pointers as part of the
+returned object, one for an initialization function and the other for
+a finalization function.
+
+(2) The tool has to define the initialization and the finalization
+functions referred to above. The initialization function is invoked by
+the OpenMP runtime with an input lookup parameter. Typically, the
+initialization function would use the lookup parameter to obtain a
+handle to the function ompt_set_callback that is implemented by the
+OpenMP runtime. Using this handle, the tool can then register
+callbacks. In our examples for OMPT target, some common callbacks
+registered include device initialization, data transfer operations,
+and target submit.
+
+(3) The device initialize callback, implemented by the tool, is
+invoked by the OpenMP device plugin runtime during device
+initialization with a lookup parameter. This callback would look up
+entry points (such as ompt_start_trace) for device tracing so that the
+tool can control the regions that should be traced.
+
+(4) The ompt_start_trace entry point expects 2 function pointers, one
+for an allocation function that will be invoked by the OpenMP runtime
+for allocating space for trace record buffers. The other one is a
+buffer-completion callback function that will be invoked by an OpenMP
+runtime helper thread for returning trace records to the tool. The
+tool is expected to use the entry point, ompt_get_record_ompt, to
+inspect a trace record at a given cursor and the entry point,
+ompt_advance_buffer_cursor, to traverse the returned trace records.
+
+(5) If device tracing is desired, calls to entry points,
+ompt_set_trace_ompt, ompt_start_trace, ompt_flush_trace, and
+ompt_stop_trace will be injected into the OpenMP program by the tool
+to control the type and region of tracing.

examples/tools/ompt/veccopy-ompt-target-tracing/Makefile

@@ -0,0 +1,129 @@
+#-----------------------------------------------------------------------
+#
+#  Makefile: Cuda clang demo Makefile for both amdgcn and nvptx targets.
+#            amdgcn GPU targets begin with "gfx". nvptx targets begin
+#            with sm_.  Example: To build and run on k4000 do this:
+#
+#            export AOMP_GPU=sm_30
+#            make run
+#
+#  Run "make help" to see other options for this Makefile
+
+TESTNAME = veccopy-ompt-target-tracing
+TESTSRC  = veccopy-ompt-target-tracing.c
+
+UNAMEP = $(shell uname -m)
+AOMP_CPUTARGET = $(UNAMEP)-pc-linux-gnu
+ifeq ($(UNAMEP),ppc64le)
+  AOMP_CPUTARGET = ppc64le-linux-gnu
+endif
+
+# --- Standard Makefile check for AOMP installation ---
+ifeq ("$(wildcard $(AOMP))","")
+  ifneq ($(AOMP),)
+    $(warning AOMP not found at $(AOMP))
+  endif
+  AOMP = $(HOME)/rocm/aomp
+  ifeq ("$(wildcard $(AOMP))","")
+    $(warning AOMP not found at $(AOMP))
+    AOMP = /usr/lib/aomp
+    ifeq ("$(wildcard $(AOMP))","")
+      $(warning AOMP not found at $(AOMP))
+      $(error Please install AOMP or correctly set env-var AOMP)
+    endif
+  endif
+endif
+# --- End Standard Makefile check for AOMP installation ---
+INSTALLED_GPU  = $(shell $(AOMP)/bin/mygpu -d gfx900)# Default AOMP_GPU is gfx900 which is vega
+AOMP_GPU       ?= $(INSTALLED_GPU)
+CC              = $(AOMP)/bin/clang
+
+ifeq (sm_,$(findstring sm_,$(AOMP_GPU)))
+  AOMP_GPUTARGET = nvptx64-nvidia-cuda
+else
+  AOMP_GPUTARGET = amdgcn-amd-amdhsa
+endif
+
+# Sorry, clang openmp requires these complex options
+CFLAGS = -O3 -target $(AOMP_CPUTARGET) -fopenmp -fopenmp-targets=$(AOMP_GPUTARGET) -Xopenmp-target=$(AOMP_GPUTARGET) -march=$(AOMP_GPU)
+
+ifeq ($(OFFLOAD_DEBUG),1)
+  $(info    DEBUG Mode ON)
+  CCENV  = env LIBRARY_PATH=$(AOMP)/lib-debug
+  RUNENV = LIBOMPTARGET_DEBUG=1
+endif
+
+ifeq ($(VERBOSE),1)
+  $(info    Compilation VERBOSE Mode ON)
+  CFLAGS += -v
+endif
+
+ifeq ($(TEMPS),1)
+  $(info    Compilation and linking save-temp Mode ON)
+  CFLAGS += -save-temps 
+endif
+
+ifeq (sm_,$(findstring sm_,$(AOMP_GPU)))
+  CUDA   ?= /usr/local/cuda
+  LFLAGS += -L$(CUDA)/targets/$(UNAMEP)-linux/lib -lcudart
+endif
+
+CFLAGS += $(EXTRA_CFLAGS)
+
+# ----- Demo compile and link in one step, no object code saved
+$(TESTNAME): $(TESTSRC)
+	$(CCENV) $(CC) $(CFLAGS) $(LFLAGS) $^ -o $@
+
+run: $(TESTNAME)
+	$(RUNENV) ./$(TESTNAME)
+
+#  ----   Demo compile and link in two steps, object saved
+$(TESTNAME).o: $(TESTSRC)
+	$(CCENV) $(CC) -c $(CFLAGS) $^ -o $@
+
+obin:	$(TESTNAME).o
+	$(CCENV) $(CC) $(CFLAGS) $(LFLAGS) $^ -o $@
+
+run_obin: obin
+	$(RUNENV) ./obin
+
+help:
+	@echo
+	@echo "Source[s]:		$(TESTSRC)"
+	@echo "Application binary:    	$(TESTNAME)"
+	@echo "Target GPU:		$(AOMP_GPU)"
+	@echo "Target triple:		$(AOMP_GPUTARGET)"
+	@echo "AOMP compiler: 		$(CC)"
+	@echo "Compile flags:		$(CFLAGS)"
+ifeq (sm_,$(findstring sm_,$(AOMP_GPU)))
+	@echo "CUDA installation:	$(CUDA)"
+endif
+	@echo
+	@echo "This Makefile supports these targets:"
+	@echo
+	@echo " make			// Builds $(TESTNAME) "
+	@echo " make run		// Executes $(TESTNAME) "
+	@echo
+	@echo " make $(TESTNAME).o		// build object file "
+	@echo " make obin		// Link object file to build binary "
+	@echo " make run_obin		// Execute obin "
+	@echo
+	@echo " make clean"
+	@echo " make help"
+	@echo
+	@echo "Environment variables used by this Makefile:"
+	@echo "  AOMP_GPU=<GPU>       Target GPU, e.g sm_30, default=gfx900. To build for"
+	@echo "                       Nvidia GPUs, set AOMP_GPU=sm_60 or appropriate sm_"
+	@echo "  AOMP=<dir>           AOMP install dir, default=/usr/lib/aomp"
+	@echo "  EXTRA_CFLAGS=<args>  extra arguments for compiler"
+	@echo "  OFFLOAD_DEBUG=n      if n=1, compile and run in Debug mode"
+	@echo "  VERBOSE=n            if n=1, add verbose output"
+	@echo "  TEMPS=1              do not delete intermediate files"
+ifeq (sm_,$(findstring sm_,$(AOMP_GPU)))
+	@echo "  CUDA=<dir>           CUDA install dir, default=/usr/local/cuda"
+endif
+	@echo
+
+# Cleanup anything this makefile can create
+clean:
+	rm -f $(TESTNAME) obin *.i *.ii *.bc *.lk a.out-* *.ll *.s *.o *.cubin

examples/tools/ompt/veccopy-ompt-target-tracing/README

@@ -0,0 +1,10 @@
+An illustration of how a tool would use OMPT target support for a
+simple vector copy OpenMP program.
+
+To compile and run:
+make run
+
+For help:
+make help
+
+Example output is in example_run.log.