add primitive single matmul bench

bench_primitive_single_matmul.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+set -e
+set -x
+
+if [ -z "${WORKLOAD_BENCH_ROOT_PATH}" ]; then
+  echo "WORKLOAD_BENCH_ROOT_PATH unset, by default set to ${PWD}" 
+  WORKLOAD_BENCH_ROOT_PATH=${PWD}
+fi
+
+# collecting oneDNN primitive performance
+_DNNL_GRAPH_DISABLE_COMPILER_BACKEND=1 ONEDNN_GRAPH_VERBOSE=1 bash ${WORKLOAD_BENCH_ROOT_PATH}/run_benchdnn_graph.sh MLP > ${WORKLOAD_BENCH_ROOT_PATH}/primitive.log
+
+# aggregate data
+python ${WORKLOAD_BENCH_ROOT_PATH}/scripts/onednn_verbose_parser.py --log_dir ${WORKLOAD_BENCH_ROOT_PATH}
+echo "Performance data dumped to ${WORKLOAD_BENCH_ROOT_PATH}/primitive_single_matmul.csv"

scripts/onednn_verbose_parser.py

@@ -0,0 +1,46 @@
+import argparse
+import re
+import csv
+
+def main():
+    args_parser = argparse.ArgumentParser(description='onednn verbose parser')
+    args_parser.add_argument('--log_dir', help='path to input files')
+    args = args_parser.parse_args()
+
+    # as order of "M, K, N"
+    mlp_shapes = [[13, 512, 256, 128], [479, 1024, 1024, 512, 256, 1]]
+    batch_sizes = [32, 64, 128, 256, 512]
+    matmul_shapes = []
+    for mlp_shape in mlp_shapes:
+        for i in range(len(mlp_shape) - 1):
+            for bs in batch_sizes:
+                matmul_shapes.append(", ".join([str(bs), str(mlp_shape[i]), str(mlp_shape[i + 1])]))
+
+    execution_time_dict = {"f32": {}, "int8": {}}
+    with open(args.log_dir + "/primitive.log", "r") as file:
+        lines = file.readlines()
+        for line in lines:
+            if line.startswith("onednn_graph_verbose,exec,cpu"):
+                verbose_seg = line.split(",")
+                time = float(verbose_seg[-1])
+                shapes = verbose_seg[7]
+                result = re.findall(r"([0-9]+)x([0-9]+)", shapes)
+                dtype = "int8" if "u8" in shapes else "f32"
+                matmul_shape = ", ".join([str(result[0][0]), str(result[0][1]), str(result[1][1])])
+                if matmul_shape in execution_time_dict[dtype]:
+                    execution_time_dict[dtype][matmul_shape].append(time)
+                else:
+                    execution_time_dict[dtype][matmul_shape] = [time]
+
+    with open(args.log_dir + "/primitive_single_matmul.csv", "w+") as f:
+        writer = csv.writer(f)
+        writer.writerow(["data type", "m, k, n", "time (ms)"])
+        for dtype in ["f32", "int8"]:
+            for shape in matmul_shapes:
+                length = len(execution_time_dict[dtype][shape]) // 10 * 9
+                avg_time = sum(execution_time_dict[dtype][shape][-length:]) / length
+                writer.writerow([dtype, shape, avg_time])
+
+
+if __name__ == "__main__":
+    main()