add IMB-RMA-GPU

Author: JuliaRS

Makefile

@@ -68,16 +68,22 @@ IMB-MPI1-GPU:
 	make -C src_cpp -f Makefile TARGET=MPI1 GPU_ENABLE=1
 	@cp src_cpp/IMB-MPI1 ./IMB-MPI1-GPU
 
+IMB-RMA-GPU:
+	make -C src_cpp -f Makefile TARGET=RMA GPU_ENABLE=1
+	@cp src_cpp/IMB-RMA ./IMB-RMA-GPU
+
 
 clean:
 	make -C src_cpp -f Makefile TARGET=MPI1 clean
+	make -C src_cpp -f Makefile TARGET=MPI1 clean GPU_ENABLE=1
 	make -C src_cpp -f Makefile TARGET=NBC clean
 	make -C src_cpp -f Makefile TARGET=RMA clean
+	make -C src_cpp -f Makefile TARGET=RMA clean GPU_ENABLE=1
 	make -C src_cpp -f Makefile TARGET=EXT clean
 	make -C src_cpp -f Makefile TARGET=IO clean
 	make -C src_cpp -f Makefile TARGET=MT clean
 	make -C src_c/P2P -f Makefile TARGET=P2P clean
-	make -C src_cpp -f Makefile TARGET=MPI1 clean GPU_ENABLE=1
 	rm -f IMB-MPI1 IMB-NBC IMB-RMA IMB-EXT IMB-IO IMB-MT IMB-P2P
 	if [ -e IMB-MPI1-GPU ]; then rm -f IMB-MPI1-GPU; fi
+	if [ -e IMB-RMA-GPU ]; then rm -f IMB-RMA-GPU; fi
 	if [ -e IMB-MPI4 ]; then make -C src_cpp -f Makefile TARGET=MPI4 clean; rm -f IMB-MPI4; fi

src_c/IMB_bnames_rma.h

@@ -62,8 +62,10 @@ char *DEFC[] = {
     "Exchange_get",
     "Accumulate",
     "Get_accumulate",
+#ifndef GPU_ENABLE
     "Fetch_and_op",
     "Compare_and_swap",
+#endif //!GPU_ENABLE
     "Truly_passive_put"
 };
 
@@ -88,7 +90,9 @@ char *ALLC[] = {
     "Exchange_get",
     "Accumulate",
     "Get_accumulate",
+#ifndef GPU_ENABLE
     "Fetch_and_op",
     "Compare_and_swap",
+#endif //!GPU_ENABLE
     "Truly_passive_put"
 };

src_c/IMB_g_info.c

@@ -98,7 +98,11 @@ void IMB_general_info() {
 #elif defined NBC
     fprintf(unit, "#    Intel(R) MPI Benchmarks %s, MPI-NBC part\n", VERSION);
 #elif defined RMA
+#ifdef GPU_ENABLE
+    fprintf(unit, "#    Intel(R) MPI Benchmarks %s, RMA part (GPU), Technical preview\n", VERSION);
+#else
     fprintf(unit, "#    Intel(R) MPI Benchmarks %s, MPI-RMA part\n", VERSION);
+#endif //GPU_ENABLE
 #elif defined MPI4
     fprintf(unit, "#    Intel(R) MPI Benchmarks %s, MPI-4 part\n", VERSION);
 #endif

src_c/IMB_mem_manager.c

@@ -70,7 +70,7 @@ For more documentation than found here, see
 #include <limits.h> /* for INT_MAX declaration*/
 #include <stdint.h>
 
-#if defined(GPU_ENABLE) && defined(MPI1)
+#if defined(GPU_ENABLE) && (defined(MPI1) || defined(RMA))
 #include "IMB_gpu_common.h"
 #define IMB_ALLOC(buff, size, where)                                  \
     do {                                                              \

src_c/IMB_output.c

@@ -1323,8 +1323,10 @@ void IMB_help()
         "Exchange_get\n"
         "Accumulate\n"
         "Get_accumulate\n"
+#ifndef GPU_ENABLE
         "Fetch_and_op\n"
         "Compare_and_swap\n"
+#endif //!GPU_ENABLE
         "Truly_passive_put\n"
         "\n");
 

src_c/IMB_parse_name_rma.c

@@ -193,6 +193,7 @@ void IMB_set_bmark(struct Bench* bmark) {
         type = SingleTransfer;
         bmark->access = put;
         bmark->reduction = 1;
+#ifndef GPU_ENABLE
     } else if (!strcmp(bmark->name, "fetch_and_op")) {
         bmark->Benchmark = IMB_rma_fetch_and_op;
         type = SingleElementTransfer;
@@ -203,6 +204,7 @@ void IMB_set_bmark(struct Bench* bmark) {
         type = SingleElementTransfer;
         bmark->access = put;
         bmark->bench_comments = Compare_and_swap_cmt;
+#endif //!GPU_ENABLE
     } else if (!strcmp(bmark->name, "truly_passive_put")) {
         bmark->Benchmark = IMB_rma_passive_put;
         type = SingleTransfer;

src_cpp/RMA/Makefile.RMA.mk

@@ -53,8 +53,42 @@ $(C_SRC_DIR)/IMB_rma_put.c \
 $(C_SRC_DIR)/IMB_cpu_exploit.c \
 $(C_SRC_DIR)/IMB_rma_get.c \
 $(C_SRC_DIR)/IMB_rma_atomic.c
-C_OBJ=$(subst $(C_SRC_DIR),RMA,$(C_SRC:.c=.o))
+ifdef GPU_ENABLE
+override C_SRC += $(C_SRC_DIR)/IMB_gpu_common.c
+override CPPFLAGS += -DGPU_ENABLE
+override LDFLAGS += -ldl
+ifdef CUDA_INCLUDE_DIR
+override C_SRC += $(C_SRC_DIR)/IMB_cuda.c \
+$(C_SRC_DIR)/IMB_cuda_api.c
+override CPPFLAGS += -I${CUDA_INCLUDE_DIR} -DCUDA_INCLUDE_DIR
+endif
+ifdef ZE_INCLUDE_DIR
+override C_SRC += $(C_SRC_DIR)/IMB_ze.c \
+$(C_SRC_DIR)/IMB_ze_api.c
+override CPPFLAGS += -I${ZE_INCLUDE_DIR} -DZE_INCLUDE_DIR
+endif
+ifndef CUDA_INCLUDE_DIR
+ifndef ZE_INCLUDE_DIR
+ifneq (clean, $(filter clean,$(MAKECMDGOALS)))
+$(error CUDA_INCLUDE_DIR and ZE_INCLUDE_DIR are not set) 
+endif
+endif
+endif
+SUBDIR:=GPU
+else
+SUBDIR:=CPU
+endif
+
+C_OBJ=$(subst $(C_SRC_DIR),RMA/$(SUBDIR),$(C_SRC:.c=.o))
 ADDITIONAL_OBJ += $(C_OBJ)
+BECHMARK_SUITE_LOCAL_OBJ=$(subst RMA/,RMA/$(SUBDIR)/,$(BECHMARK_SUITE_LOCAL_SRC:.cpp=.o))
+ADDITIONAL_OBJ += $(BECHMARK_SUITE_LOCAL_OBJ)
 
-RMA/%.o: $(C_SRC_DIR)/%.c
+RMA/$(SUBDIR)/%.o: $(C_SRC_DIR)/%.c $(SUBDIR)
 	$(CC) $(CFLAGS) $(CPPFLAGS) -DRMA -c -o $@ $<
+
+RMA/$(SUBDIR)/%.o: RMA/%.cpp $(SUBDIR)
+	$(CXX) $(CPPFLAGS) $(CXXFLAGS) -DRMA -c -o $@ $<
+
+${SUBDIR}:
+	mkdir -p RMA/${SUBDIR}

src_cpp/RMA/RMA_benchmark.cpp

@@ -259,6 +259,7 @@ BENCHMARK(IMB_rma_get_accumulate, Get_accumulate)
     return true;
 }
 
+#ifndef GPU_ENABLE
 BENCHMARK(IMB_rma_fetch_and_op, Fetch_and_op)
 {
     descr->flags.insert(DEFAULT);
@@ -280,6 +281,7 @@ BENCHMARK(IMB_rma_compare_and_swap, Compare_and_swap)
     descr->comments.push_back("Uses MPI_INT data type");
     return true;
 }
+#endif //!GPU_ENABLE
 
 BENCHMARK(IMB_rma_passive_put, Truly_passive_put)
 {

src_cpp/RMA/RMA_suite.cpp

@@ -325,6 +325,19 @@ template <> bool BenchmarkSuite<BS_RMA>::declare_args(args_parser &parser, std::
                "\n"
                "Default:\n"
                "off\n");
+#ifdef GPU_ENABLE
+   parser.add<string>("mem_alloc_type", "device").set_caption("buffer type").
+           set_description(
+                "The argument after -mem_alloc_type is a one from possible strings,\n"
+                "Specifying that type will be used:\n"
+                "device, host, shared, cpu\n"
+                "\n"
+                "Example:\n"
+                "-mem_alloc_type host\n"
+                "\n"
+                "Default:\n"
+                "device\n");
+#endif //GPU_ENABLE
     parser.set_default_current_group();
     return true;
 }
@@ -502,7 +515,24 @@ template <> bool BenchmarkSuite<BS_RMA>::prepare(const args_parser &parser, cons
     if (parser.get<bool>("msg_pause") == true) {
         c_info.msg_pause = 1;
     }
-         
+
+#ifdef GPU_ENABLE
+    gpu_initialize();
+    // mem_alloc_type
+    string mem_alloc_type = parser.get<string>("mem_alloc_type");
+    if (mem_alloc_type == "cpu") {
+        c_info.mem_alloc_type = MAT_CPU;
+    }
+    else if (mem_alloc_type == "device") {
+        c_info.mem_alloc_type = MAT_DEVICE;
+    }
+    else if (mem_alloc_type == "host") {
+        c_info.mem_alloc_type = MAT_HOST;
+    }
+    else if (mem_alloc_type == "shared") {
+        c_info.mem_alloc_type = MAT_SHARED;
+    }
+#endif //GPU_ENABLE
 #endif
 
 #if BASIC_INPUT_EXPERIMENT == 0