Merge changes I1f0ec9d3,I55e0cf22
* changes:
Lint cleanup benchmark_kelvin.cc
Fix carry over benchmark mismatch count
diff --git a/benchmarks/benchmark_kelvin.cc b/benchmarks/benchmark_kelvin.cc
index fdec54a..079f567 100644
--- a/benchmarks/benchmark_kelvin.cc
+++ b/benchmarks/benchmark_kelvin.cc
@@ -16,10 +16,10 @@
#include <memory>
-#include "crt/kelvin.h"
-#include "crt/log.h"
#include "benchmarks/benchmark.h"
#include "benchmarks/cycle_count.h"
+#include "crt/kelvin.h"
+#include "crt/log.h"
#include "tensorflow/lite/micro/micro_interpreter.h"
#include "tensorflow/lite/micro/micro_log.h"
#include "tensorflow/lite/micro/micro_mutable_op_resolver.h"
@@ -34,20 +34,23 @@
// In order to include the model data generate from Bazel, include the header
// using the name passed as a macro.
-#define MODEL_HEADER_DIRECTORY BENCHMARK_PATH/
+#define MODEL_HEADER_DIRECTORY BENCHMARK_PATH
#define MODEL_HEADER_TYPE _model.h
-#define MODEL_HEADER STR(MODEL_HEADER_DIRECTORY BENCHMARK_NAME MODEL_HEADER_TYPE)
+#define MODEL_HEADER \
+ STR(MODEL_HEADER_DIRECTORY BENCHMARK_NAME MODEL_HEADER_TYPE)
#include MODEL_HEADER
#if (TEST_DATA_INPUT == 1)
#define TEST_DATA_INPUT_HEADER_TYPE _input.h
-#define TEST_DATA_INPUT_HEADER STR(MODEL_HEADER_DIRECTORY BENCHMARK_NAME TEST_DATA_INPUT_HEADER_TYPE)
+#define TEST_DATA_INPUT_HEADER \
+ STR(MODEL_HEADER_DIRECTORY BENCHMARK_NAME TEST_DATA_INPUT_HEADER_TYPE)
#include TEST_DATA_INPUT_HEADER
#endif
#if (TEST_DATA_OUTPUT == 1)
#define TEST_DATA_OUTPUT_HEADER_TYPE _output.h
-#define TEST_DATA_OUTPUT_HEADER STR(MODEL_HEADER_DIRECTORY BENCHMARK_NAME TEST_DATA_OUTPUT_HEADER_TYPE)
+#define TEST_DATA_OUTPUT_HEADER \
+ STR(MODEL_HEADER_DIRECTORY BENCHMARK_NAME TEST_DATA_OUTPUT_HEADER_TYPE)
#include TEST_DATA_OUTPUT_HEADER
#endif
@@ -59,16 +62,19 @@
#endif
uint8_t g_tensor_arena[kTensorArenaSize] __attribute__((aligned(64)));
-__attribute__((section(".model_output_header"))) BenchmarkOutputHeader output_header = {
- .return_code = 0, // Set by kelvin_start based on return value in main.
+__attribute__((
+ section(".model_output_header"))) BenchmarkOutputHeader output_header = {
+ .return_code = 0, // Set by kelvin_start based on return value in main.
.iterations = 0,
.cycles = 0,
.mismatch_count = 0,
};
-// This includes all ops currently used in the Kelvin model suite. More can be added.
+// This includes all ops currently used in the Kelvin model suite. More can be
+// added.
constexpr int kAllOpsNum = 28;
-std::unique_ptr<tflite::MicroMutableOpResolver<kAllOpsNum>> GetAllOpsResolver() {
+std::unique_ptr<tflite::MicroMutableOpResolver<kAllOpsNum>>
+GetAllOpsResolver() {
tflite::MicroMutableOpResolver<kAllOpsNum> resolver;
resolver.AddAveragePool2D();
resolver.AddMaxPool2D();
@@ -110,28 +116,32 @@
constexpr int kSuccess = 0;
constexpr int kAllocatonFailed = -1;
constexpr int kInvokeFailed = -2;
-} // namespace
+} // namespace
-
-int main(int argc, char **argv) {
- std::unique_ptr<tflite::MicroMutableOpResolver<kAllOpsNum>> resolver = GetAllOpsResolver();
+int main(int argc, char** argv) {
+ std::unique_ptr<tflite::MicroMutableOpResolver<kAllOpsNum>> resolver =
+ GetAllOpsResolver();
const auto* model = tflite::GetModel(g_benchmark_model_data);
uint8_t variable_arena[2048];
- tflite::MicroAllocator *variable_allocator =
+ tflite::MicroAllocator* variable_allocator =
tflite::MicroAllocator::Create(variable_arena, 1024);
- tflite::MicroResourceVariables *resource_variables =
+ tflite::MicroResourceVariables* resource_variables =
tflite::MicroResourceVariables::Create(variable_allocator, 20);
#if (PROFILE == 1)
tflite::MicroProfiler profiler;
- std::unique_ptr<tflite::MicroInterpreter> interpreter = std::make_unique<tflite::MicroInterpreter>(
- model, *resolver.get(), g_tensor_arena, kTensorArenaSize, resource_variables, &profiler);
+ std::unique_ptr<tflite::MicroInterpreter> interpreter =
+ std::make_unique<tflite::MicroInterpreter>(
+ model, *resolver.get(), g_tensor_arena, kTensorArenaSize,
+ resource_variables, &profiler);
// For a profiled model, just run a single iteration
const int iterations = 1;
#else
- std::unique_ptr<tflite::MicroInterpreter> interpreter = std::make_unique<tflite::MicroInterpreter>(
- model, *resolver.get(), g_tensor_arena, kTensorArenaSize, resource_variables);
+ std::unique_ptr<tflite::MicroInterpreter> interpreter =
+ std::make_unique<tflite::MicroInterpreter>(
+ model, *resolver.get(), g_tensor_arena, kTensorArenaSize,
+ resource_variables);
const int iterations = ITERATIONS;
#endif
@@ -142,7 +152,8 @@
TfLiteTensor* input = interpreter->input(0);
#if (TEST_DATA_INPUT == 1)
- memcpy(tflite::GetTensorData<uint8_t>(input), g_benchmark_input, input->bytes);
+ memcpy(tflite::GetTensorData<uint8_t>(input), g_benchmark_input,
+ input->bytes);
#else
memset(tflite::GetTensorData<uint8_t>(input), 0, input->bytes);
#endif
@@ -157,15 +168,17 @@
// TODO(michaelbrooks): Possibly set/verify test data?
for (int i = 0; i < iterations; ++i) {
#if (TEST_DATA_INPUT == 1)
- memcpy(tflite::GetTensorData<uint8_t>(input), g_benchmark_input, input->bytes);
+ memcpy(tflite::GetTensorData<uint8_t>(input), g_benchmark_input,
+ input->bytes);
#else
- memset(tflite::GetTensorData<uint8_t>(input), 0, input->bytes);
+ memset(tflite::GetTensorData<uint8_t>(input), 0, input->bytes);
#endif
interpreter->Invoke();
}
uint64_t end = mcycle_read();
uint64_t num_cycles = end - begin;
+ output_header.mismatch_count = 0;
#if (TEST_DATA_OUTPUT == 1)
TfLiteTensor* output = interpreter->output(0);
int mismatch_count = 0;
diff --git a/benchmarks/benchmark_smc.c b/benchmarks/benchmark_smc.c
index 7edaf46..757ffe8 100644
--- a/benchmarks/benchmark_smc.c
+++ b/benchmarks/benchmark_smc.c
@@ -140,13 +140,15 @@
uint64_t average_cycles = udiv64_slow(cycles, iterations, NULL);
uint64_t wall_time_us = timer_finish - timer_start;
uint64_t average_wall_time_us = udiv64_slow(wall_time_us, iterations, NULL);
- uint32_t mismatch_count = output_header_ptr->mismatch_count;
LOG_INFO("Iterations: %d", iterations);
_print64("Total Cycles", cycles);
_print64("Average Cycles per Iteration", average_cycles);
_print64("Wall time (us)", wall_time_us);
_print64("Wall time per Iteration (us)", average_wall_time_us);
+#if (TEST_DATA_OUTPUT == 1)
+ uint32_t mismatch_count = output_header_ptr->mismatch_count;
LOG_INFO("Mismatch count: %d", mismatch_count);
+#endif
LOG_INFO("========== End Benchmark ==========");
while (true) {
wait_for_interrupt();
diff --git a/benchmarks/benchmarks.bzl b/benchmarks/benchmarks.bzl
index c941f22..c622c99 100644
--- a/benchmarks/benchmarks.bzl
+++ b/benchmarks/benchmarks.bzl
@@ -25,7 +25,7 @@
test_data_output = None,
profile = False,
kelvin_binary_info = None,
- benchmark_path = "benchmarks",
+ benchmark_path = "benchmarks/",
hw_test_size = "medium",
hw_test_tags = [],
iss_test_size = "small",
@@ -103,7 +103,7 @@
test_data = None,
profile = False,
kelvin_binary_info = None,
- benchmark_path = "benchmarks",
+ benchmark_path = "benchmarks/",
**kwargs):
_kelvin_benchmark_device(
name = name,
@@ -124,7 +124,7 @@
test_data = None,
profile = False,
kelvin_binary_info = None,
- benchmark_path = "benchmarks",
+ benchmark_path = "benchmarks/",
**kwargs):
_kelvin_benchmark_device(
name = name,
@@ -145,7 +145,7 @@
test_data = None,
profile = False,
kelvin_binary_info = None,
- benchmark_path = "benchmarks",
+ benchmark_path = "benchmarks/",
**kwargs):
kelvin_benchmark_asic(
name = "{}_asic".format(name),
@@ -188,7 +188,7 @@
test_data_output = None,
profile = False,
kelvin_binary_info = None,
- benchmark_path = "benchmarks",
+ benchmark_path = "benchmarks/",
arena_size_bytes = 1536 * 1024, # 1.5MB
tags = [],
**kwargs):
@@ -199,7 +199,10 @@
"@kelvin_sw//benchmarks:benchmark_smc.c",
"@kelvin_sw//benchmarks:benchmark.h",
],
- copts = ["-DBENCHMARK_NAME={}".format(name)],
+ copts = [
+ "-DBENCHMARK_NAME={}".format(name),
+ "-DTEST_DATA_OUTPUT={}".format(1 if test_data_output else 0),
+ ],
per_device_deps = {
device_type: device_deps("smc").get(device_type),
},
