added support for two more MLPerfTiny benchmarks: Image Classification

.gitignore

@@ -7,3 +7,8 @@ benchmark_sources/third_party/*
 benchmark_sources/tflite-micro
 benchmark_sources/tensorflow/*
 benchmark_sources/signal/*
+
+benchmark_sources/models/cifar10_resnet8/dataset/*
+benchmark_sources/models/vww/dataset/*
+
+__pycache__/

README.md

@@ -2,7 +2,7 @@
 
 ### Overview
 
-This repository contains the system for running TinyML Benchmarks on a system with the CV32E40X scalar core and Vicuna 2.0 on the CV-X-IF interface.  It currently supports two benchmarks from the [MLPerfTiny Suite](https://github.com/mlcommons/tiny), anomaly detection (toycar) and keyword spotting (aww).
+This repository contains the system for running TinyML Benchmarks on a system with the CV32E40X scalar core and Vicuna 2.0 on the CV-X-IF interface.  It currently supports four benchmarks from the [MLPerfTiny Suite](https://github.com/mlcommons/tiny), anomaly detection (toycar), keyword spotting (aww), image classification (resnet8), and visual wake words (vww).
 
 Due to the limitations of CMake support for native and cross-compiliation within the same project, two separate CMake projects are used.  The first, defined in the **/build_model** directory, uses Verilator version v5.030 to create a executable Verilator model of Vicuna for the selected configuration.  The second, defined in **/build_benchmarks**, uses the LLVM version 18.1.4 and [Tensorflow Lite for Microcontrollers](https://github.com/tensorflow/tflite-micro) to compile the benchmarks for the selected configuration of Vicuna.  Each compiled benchmark is registered with CTest.
 

benchmark_sources/CMakeLists.txt

@@ -516,18 +516,24 @@ target_link_libraries(tflm PRIVATE
 
 add_subdirectory(Int8/toycar)
 add_subdirectory(Int8/aww)
+add_subdirectory(Int8/resnet8)
+add_subdirectory(Int8/vww)
 
 if(RISCV_F)
     if(RISCV_ZFH)
         if(FH_OPS)
             add_subdirectory(FP16/toycar_16_bit_ops)
             add_subdirectory(FP16/aww_16_bit_ops)
+            add_subdirectory(FP16/resnet8_16_bit_ops)
+            add_subdirectory(FP16/vww_16_bit_ops)
         else()
             add_subdirectory(FP16/toycar_32_bit_ops)
         endif()
     else()
       add_subdirectory(FP32/toycar)
       add_subdirectory(FP32/aww)
+      add_subdirectory(FP32/resnet8)
+      add_subdirectory(FP32/vww)
     endif()
 endif()
 

benchmark_sources/FP16/resnet8_16_bit_ops/CMakeLists.txt

@@ -0,0 +1 @@
+add_Benchmark(resnet8_fp16 ${CMAKE_CURRENT_SOURCE_DIR} ${BUILD_DIR}/benchmark_sources/FP16/resnet8)

benchmark_sources/FP16/resnet8_16_bit_ops/resnet8_fp16.cpp

@@ -0,0 +1,105 @@
+#include <cstdarg>
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+
+#include "resnet8_fp16_data/resnet8_fp16_input_data.h"
+#include "resnet8_fp16_data/resnet8_fp16_model_data.h"
+#include "resnet8_fp16_data/resnet8_fp16_model_settings.h"
+#include "resnet8_fp16_data/resnet8_fp16_output_data_ref.h"
+
+#include "tensorflow/lite/micro/tflite_bridge/micro_error_reporter.h"
+#include "tensorflow/lite/micro/micro_interpreter.h"
+#include "tensorflow/lite/micro/micro_mutable_op_resolver.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+
+extern "C" {
+#include "runtime.h"
+#include "uart.h"
+#include "terminate_benchmark.h"
+}
+
+constexpr size_t tensor_arena_size = 256 * 1024 * 4;
+alignas(16) uint8_t tensor_arena[tensor_arena_size];
+
+int run_test()
+{
+    //tflite::MicroErrorReporter micro_error_reporter;
+    //tflite::ErrorReporter *error_reporter = &micro_error_reporter;
+
+    const tflite::Model *model = tflite::GetModel(resnet8_fp16_model_data);
+
+    static tflite::MicroMutableOpResolver<8> resolver;
+    resolver.AddFullyConnected();
+    resolver.AddConv2D();
+    resolver.AddDepthwiseConv2D();
+    resolver.AddAveragePool2D();
+    resolver.AddReshape();
+    resolver.AddSoftmax();
+    resolver.AddAdd();
+    resolver.AddDequantize();
+
+    tflite::MicroInterpreter interpreter(model, resolver, tensor_arena, tensor_arena_size);
+
+    if (interpreter.AllocateTensors() != kTfLiteOk)
+    {
+        //TF_LITE_REPORT_ERROR(error_reporter, "ERROR: In AllocateTensors().");
+        return -1;
+    }
+
+    //for (size_t i = 0; i < resnet8_fp16_data_sample_cnt; i++)
+    for (size_t i = 0; i < 1; i++)
+    {
+        memcpy(interpreter.input(0)->data.f, (float *)resnet8_fp16_input_data[i], resnet8_fp16_input_data_len[i] * 2); //Bytes to copy is 4*len
+
+        if (interpreter.Invoke() != kTfLiteOk)
+        {
+            //TF_LITE_REPORT_ERROR(error_reporter, "ERROR: In Invoke().");
+            return -1;
+        }
+
+        int8_t top_index = 0;
+        _Float16* output_ptr = (_Float16*)interpreter.output(0)->data.f;
+        for (size_t j = 0; j < resnet8_fp16_model_label_cnt; j++)
+        {
+            if (output_ptr[j] > output_ptr[top_index])
+            {
+                top_index = j;
+            }
+        }
+
+        if (top_index != resnet8_fp16_output_data_ref[i])
+        {
+            //uart_printf("ERROR: at #%d, top_index %d resnet8_fp16_output_data_ref %d \n", i, top_index, resnet8_fp16_output_data_ref[i]);
+            return -1;
+        }
+        else
+        {
+            //uart_printf("Sample #%d pass, top_index %d matches ref %d \n", i, top_index, resnet8_fp16_output_data_ref[i]);
+        }
+    }
+    return 0;
+}
+
+int main(int argc, char *argv[])
+{
+    int ret = run_test();
+    if (ret != 0)
+    {
+        #if defined(PRINT_OUTPUTS)
+        uart_printf("Test Failed!\n");
+        #endif 
+        benchmark_failure();
+
+    }
+    else
+    {
+        #if defined(PRINT_OUTPUTS)
+        uart_printf("Test Success!\n");
+        #endif
+        benchmark_success(); 
+    }
+
+    return ret;
+}