6 files changed, 76 insertions, 0 deletions
diff --git a/BS/baselines/cpu/run-perf.sh b/BS/baselines/cpu/run-perf.sh
new file mode 100755
index 0000000..a35e0fc
--- /dev/null
+++ b/BS/baselines/cpu/run-perf.sh
@@ -0,0 +1,6 @@
+#!/bin/zsh
+
+make -B NUMA=1
+
+OMP_NUM_THREADS=1 perf stat record -o t1.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./bs_omp $((2**29)) 16777216 4 4
+OMP_NUM_THREADS=4 perf stat record -o t4.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./bs_omp $((2**29)) 16777216 4 4
diff --git a/GEMV/baselines/cpu/run-perf.sh b/GEMV/baselines/cpu/run-perf.sh
new file mode 100755
index 0000000..c4274c0
--- /dev/null
+++ b/GEMV/baselines/cpu/run-perf.sh
@@ -0,0 +1,6 @@
+#!/bin/zsh
+
+make -B NUMA=1
+
+OMP_NUM_THREADS=1 perf stat record -o t1.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./gemv 4 4 4
+OMP_NUM_THREADS=4 perf stat record -o t4.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./gemv 4 4 4
diff --git a/TRNS/baselines/cpu/run-perf.sh b/TRNS/baselines/cpu/run-perf.sh
new file mode 100755
index 0000000..08bff53
--- /dev/null
+++ b/TRNS/baselines/cpu/run-perf.sh
@@ -0,0 +1,6 @@
+#!/bin/zsh
+
+make -B NUMA=1
+
+perf stat record -o t1.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./trns -w 0 -r 20 -p 2048 -o 2048 -m 16 -n 8 -t 1 -a 4 -c 4
+perf stat record -o t4.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./trns -w 0 -r 20 -p 2048 -o 2048 -m 16 -n 8 -t 4 -a 4 -c 4
diff --git a/TS/baselines/cpu/run-perf.sh b/TS/baselines/cpu/run-perf.sh
new file mode 100755
index 0000000..453b64b
--- /dev/null
+++ b/TS/baselines/cpu/run-perf.sh
@@ -0,0 +1,8 @@
+#!/bin/zsh
+
+make -B NUMA=1
+
+for i in $(seq 1 20); do
+	OMP_NUM_THREADS=1 perf stat record -o t1.${i}.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./streamp_openmp inputs/randomlist10M.txt 256 4 4
+	OMP_NUM_THREADS=4 perf stat record -o t4.${i}.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./streamp_openmp inputs/randomlist10M.txt 256 4 4
+done
diff --git a/VA/baselines/cpu/run-perf.sh b/VA/baselines/cpu/run-perf.sh
new file mode 100755
index 0000000..33cb7b5
--- /dev/null
+++ b/VA/baselines/cpu/run-perf.sh
@@ -0,0 +1,6 @@
+#!/bin/zsh
+
+make -B NUMA=1
+
+perf stat record -o t1.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./va -a 4 -b 4 -c 4 -t 1 -e 20 -w 0 -i 167772160
+perf stat record -o t4.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./va -a 4 -b 4 -c 4 -t 4 -e 20 -w 0 -i 167772160
diff --git a/perf-events.txt b/perf-events.txt
new file mode 100644
index 0000000..ab57ab2
--- /dev/null
+++ b/perf-events.txt
@@ -0,0 +1,44 @@
+cache-misses # NMPO
+cache-references
+
+cpu-cycles # NMPO
+instructions # NMPO
+
+page-faults
+
+mem-loads
+mem-loads-aux
+mem-stores
+
+branch-misses # NMPO
+branch-instructions # NMPO
+branch-load-misses # NMPO
+branch-loads # NMPO
+
+l1d_pend_miss.pending # mccalpin2023hpc <https://link.springer.com/chapter/10.1007/978-3-031-40843-4_30>
+l1d_pend_miss.pending_cycles
+
+offcore_requests.all_requests
+offcore_requests.data_rd
+offcore_requests.demand_data_rd
+
+offcore_requests_outstanding.data_rd # mccalpin2023hpc
+offcore_requests_outstanding.cycles_with_data_rd
+offcore_requests_outstanding.cycles_with_demand_data_rd
+offcore_requests_outstanding.demand_data_rd # mccalpin2023hpc
+
+L1-dcache-loads # NMPO
+L1-dcache-load-misses # NMPO
+L1-dcache-stores # NMPO
+L1-icache-load-misses # NMPO
+
+LLC-loads # NMPO
+LLC-load-misses
+LLC-stores # NMPO
+LLC-store-misses # NMPO
+
+l2_lines_out.useless_hwpf
+l2_lines_out.non_silent
+l2_lines_out.silent
+l2_request.all
+l2_request.miss