First implementation of hw counters monitoring for Arm.

src/samples/perf_event/perf_event.hpp

@@ -83,7 +83,7 @@ namespace samples {
 			// Here we define special configuration for each group
 			if (group == MEM_SAMPLE) { // Memory
 				// Config for latency monitoring
-				// fds[0].hw.config1    = samples::minimum_latency;
+				fds[0].hw.config1    = samples::minimum_latency;
 				fds[0].hw.precise_ip = 2;
 			}
 
@@ -125,13 +125,13 @@ namespace samples {
 
 				if (fds[i].fd == -1) {
 					if (errno == EOPNOTSUPP) {
-						err(EXIT_FAILURE, "Cannot attach event %s. Event is not supported. %s", fds[i].name,
-						    pfm_strerror(errno));
+						err(EXIT_FAILURE, "Cannot attach event %s in CPU %d. Event is not supported. %s", fds[i].name,
+						    cpu, pfm_strerror(errno));
 					} else if (fds[i].hw.precise_ip) {
-						err(EXIT_FAILURE, "Cannot attach event %s: precise mode may not be supported. %s",
-						    fds[group].name, pfm_strerror(errno));
+						err(EXIT_FAILURE, "Cannot attach event %s in CPU %d: precise mode may not be supported. %s",
+						    fds[group].name, cpu, pfm_strerror(errno));
 					}
-					err(EXIT_FAILURE, "Cannot attach event %s. %s", fds[i].name, pfm_strerror(errno));
+					err(EXIT_FAILURE, "Cannot attach event %s in CPU %d. %s", fds[i].name, cpu, pfm_strerror(errno));
 				} else {
 					if (VERBOSE_LVL >= VERB_LVL4) {
 						std::cout << "Event " << fds[i].name << " successfully opened." << '\n';
@@ -400,9 +400,7 @@ namespace samples {
 					total_supported_events++;
 				}
 
-				if (pmu_info.is_dfl && pmu_info.num_cntrs > MAX_HW_COUNTERS) {
-					MAX_HW_COUNTERS = pmu_info.num_cntrs;
-				}
+				MAX_HW_COUNTERS = std::max(MAX_HW_COUNTERS, pmu_info.num_cntrs);
 
 				if constexpr (NUM_GROUPS > MAX_BROADWELL_CTRS) { // Computed at compile time
 					if (pmu == PFM_PMU_INTEL_BDW || pmu == PFM_PMU_INTEL_BDW_EP) {

src/samples/samples.hpp

@@ -17,9 +17,9 @@
 
 namespace samples {
 	enum sample_type_t {
-		MEM_SAMPLE = 0,
+		INS_SAMPLE = 0,
+		MEM_SAMPLE,
 		REQ_SAMPLE,
-		INS_SAMPLE,
 #ifndef JUST_INS
 		FP_SCALAR_SAMPLE, // multiply by 1 to get flops
 		FP_128B_D_SAMPLE, // multiply by 2 to get flops
@@ -139,9 +139,9 @@ namespace samples {
 
 		size_t map_size;
 
-		const std::array<const char * const, NUM_GROUPS> events = { "MEM_TRANS_RETIRED:LATENCY_ABOVE_THRESHOLD",
-			                                                        "OFFCORE_REQUESTS:ALL_DATA_RD",
-			                                                        "INST_RETIRED"
+		const std::array<const char * const, NUM_GROUPS> events = { "perf::INSTRUCTIONS",
+			                                                        "perf::CACHE-REFERENCES",
+			                                                        "perf::PERF_COUNT_HW_CACHE_NODE"
 #ifndef JUST_INS
 			                                                        ,
 			                                                        "FP_ARITH:SCALAR_DOUBLE:SCALAR_SINGLE",