#43 add batch size to benchmarks

gpumat/benchmark.py

@@ -221,12 +221,12 @@ def pad_buffer_probe(
         for _ in range(RECT_N)
     ]
     gst_buffer: Gst.Buffer = info.get_buffer()
+    ts1 = time.time()
     nvds_batch_meta = pyds.gst_buffer_get_nvds_batch_meta(hash(gst_buffer))
     for nvds_frame_meta in nvds_frame_meta_iterator(nvds_batch_meta):
-        ts1 = time.time()
         benchmark_func(gst_buffer, nvds_frame_meta, data)
-        ts2 = time.time()
-        measurements.append((ts2 - ts1) * scale)
+    ts2 = time.time()
+    measurements.append((ts2 - ts1) * scale)
 
     return Gst.PadProbeReturn.OK
 
@@ -253,7 +253,7 @@ def bus_call(bus, message, loop):
 def main(args):
     assert (
         len(args) > 2
-    ), 'Usage: ./benchmark.py <benchmark-name> <cpu|gpu> [n-frames] [output-filename]'
+    ), 'Usage: ./benchmark.py <benchmark-name> <cpu|gpu> [n-frames] [batch-size] [output-filename]'
     benchmark_name = args[1]
     is_gpu = args[2] == 'gpu'
     assert (
@@ -263,18 +263,19 @@ def main(args):
     assert benchmark_func is not None, 'Benchmark not implemented'
 
     output_filename = None
+    n_frames = 1
+    batch_size = 1
     if len(args) > 3:
         n_frames = int(args[3])
-        if len(args) > 4:
-            output_filename = args[4]
-    else:
-        n_frames = 1
+    if len(args) > 4:
+        batch_size = int(args[4])
+    if len(args) > 5:
+        output_filename = args[5]
 
     Gst.init(None)
 
     print("Creating Pipeline")
     pipeline = Gst.Pipeline()
-    is_live = False
 
     print("Creating streammux")
     streammux = Gst.ElementFactory.make("nvstreammux", "streammux")
@@ -331,11 +332,9 @@ def main(args):
 
     source.set_property('num-buffers', n_frames)
 
-    if is_live:
-        streammux.set_property('live-source', 1)
     streammux.set_property('width', 1920)
     streammux.set_property('height', 1080)
-    streammux.set_property('batch-size', 1)
+    streammux.set_property('batch-size', batch_size)
     streammux.set_property('batched-push-timeout', 4000000)
 
     sink.set_property("sync", 0)
@@ -451,12 +450,12 @@ def main(args):
         ('stdev', statistics.stdev(measurements)),
     ]
     for name, val in metrics:
-        print(f'{name}: {val:.3f}')
+        print(f'{name}: {val:.3f} ({val / batch_size:.3f} per frame)')
     device_name = "gpu" if is_gpu else "cpu"
     with open('metrics.csv', 'a') as f:
         f.write(
             ','.join(
-                [benchmark_name, device_name, str(n_frames)]
+                [benchmark_name, device_name, str(n_frames), str(batch_size)]
                 + [f'{val:.3f}' for _, val in metrics]
             )
         )

gpumat/run_benchmarks.sh

@@ -6,6 +6,12 @@ else
     FRAME_NUM=1000
 fi
 
+if [[ -n "${2}" ]]; then
+    BATCH_SIZE="${2}"
+else
+    BATCH_SIZE=1
+fi
+
 if [[ "$(uname -m)" == "aarch64" ]]; then
     DOCKER_IMAGE="savant-deepstream-l4t:0.1.1-6.1.1-base"
 else
@@ -28,7 +34,7 @@ CPU_BENCHMARK_NAMES=(
     "blur-faces"
 )
 
-echo "name,device,frame_num,min,max,mean,median,80%,90%,95%,99%,stdev" >metrics.csv
+echo "name,device,frame_num,batch_size,min,max,mean,median,80%,90%,95%,99%,stdev" >metrics.csv
 
 for BENCHMARK_NAME in "${GPU_BENCHMARK_NAMES[@]}"; do
     echo
@@ -44,7 +50,7 @@ for BENCHMARK_NAME in "${GPU_BENCHMARK_NAMES[@]}"; do
         --workdir /gpumat \
         --entrypoint ./benchmark.py \
         -v "$(pwd):/gpumat" \
-        "${DOCKER_IMAGE}" "${BENCHMARK_NAME}" "gpu" "${FRAME_NUM}"
+        "${DOCKER_IMAGE}" "${BENCHMARK_NAME}" "gpu" "${FRAME_NUM}" "${BATCH_SIZE}"
 done
 
 for BENCHMARK_NAME in "${CPU_BENCHMARK_NAMES[@]}"; do
@@ -61,6 +67,6 @@ for BENCHMARK_NAME in "${CPU_BENCHMARK_NAMES[@]}"; do
         --workdir /gpumat \
         --entrypoint ./benchmark.py \
         -v "$(pwd):/gpumat" \
-        "${DOCKER_IMAGE}" "${BENCHMARK_NAME}" "cpu" "${FRAME_NUM}"
+        "${DOCKER_IMAGE}" "${BENCHMARK_NAME}" "cpu" "${FRAME_NUM}" "${BATCH_SIZE}"
 done
 date