Fix carry over benchmark mismatch count

Benchmark mismatch count should reset per test, and in the device
test the mismatch count should not be reported.

Bug: 345095360
Change-Id: I55e0cf229d314bc2f4c7910c39a7d06d9856f14b

benchmarks/benchmark_kelvin.cc

diff --git a/benchmarks/benchmark_kelvin.cc b/benchmarks/benchmark_kelvin.cc
index fdec54a..9c3eabb 100644
--- a/benchmarks/benchmark_kelvin.cc
+++ b/benchmarks/benchmark_kelvin.cc
@@ -166,6 +166,7 @@
   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;

benchmarks/benchmark_smc.c

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();

benchmarks/benchmarks.bzl

diff --git a/benchmarks/benchmarks.bzl b/benchmarks/benchmarks.bzl
index c941f22..8ad5f0a 100644
--- a/benchmarks/benchmarks.bzl
+++ b/benchmarks/benchmarks.bzl
@@ -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),
         },