added shufflenet, fixed profiling issue (#18)

* added shufflenet, fixed profiling issue

* minor fixes

bin/injectfault.py

@@ -289,7 +289,7 @@ def checkValues(key, val, var1 = None,var2 = None,var3 = None,var4 = None):
   ##BEHROOZ: Add max number of target locations
   elif key == "fi_max_multiple":
     assert isinstance(val, int)==True, key+" must be an integer in input.yaml"
-    assert int(val) >1, key+" must be greater than one in input.yaml"
+    assert int(val) >0, key+" must be greater than zero in input.yaml"
     assert int(val) <=int(fi_max_multiple_default), key+" must be smaller than or equal to "+str(fi_max_multiple_default)+ " in input.yaml"
   ##==============================================================
 

llvm_passes/core/Utils.cpp

@@ -57,6 +57,17 @@ Instruction *getTermInstofFunction(Function *func) {
   return ret;
 }
 
+void getAllTermInstofFunction(Function *func, std::set<Instruction*> &exitinsts) {
+
+  for (auto i = inst_begin(func); i != inst_end(func); i++) {	
+    Instruction* ret = &*i;	
+
+    if (isa<ReturnInst>(ret) || isa<ResumeInst>(ret) ||
+         isa<UnreachableInst>(ret))
+	    exitinsts.insert(ret);
+  }
+}
+
 void getProgramExitInsts(Module &M, std::set<Instruction*> &exitinsts) {
   for (Module::iterator m_it = M.begin(); m_it != M.end(); ++m_it) {
     if (!m_it->isDeclaration()) {
@@ -76,7 +87,7 @@ void getProgramExitInsts(Module &M, std::set<Instruction*> &exitinsts) {
   }
 
   Function* mainfunc = M.getFunction("main");
-  exitinsts.insert(getTermInstofFunction(mainfunc));
+  getAllTermInstofFunction(mainfunc, exitinsts);
 }
 
 Instruction *getInsertPtrforRegsofInst(Value *reg, Instruction *inst) {

runtime_lib/FaultInjectionLib.c

@@ -130,7 +130,7 @@ void _parseLLFIConfigFile() {
     //==============================================================	
     /*BEHROOZ: Add multiple corrupted regs*/
     } else if (strcmp(option, "fi_max_multiple") == 0){
-        assert(atoll(value) > 1 && "invalid fi_max_multiple in config file");
+        assert(atoll(value) > 0 && "invalid fi_max_multiple in config file");
     	config.fi_max_multiple = atoi(value);
     } else if (strcmp(option, "fi_next_cycle") == 0){
     	assert(atoll(value) > 0 && "invalid fi_next_cycle in config file");

runtime_lib/ProfilingLib.c

@@ -5,7 +5,7 @@
 
 #include "Utils.h"
 
-static long long opcodecount[OPCODE_CYCLE_ARRAY_LEN] = {0};
+static long long unsigned opcodecount[OPCODE_CYCLE_ARRAY_LEN] = {0};
 
 void doProfiling(int opcode) {
   assert(opcodecount[opcode] >= 0 && 
@@ -27,7 +27,7 @@ void endProfiling() {
   getOpcodeExecCycleArray(OPCODE_CYCLE_ARRAY_LEN, opcode_cycle_arr);
 
   unsigned i = 0;
-  long long total_cycle = 0;
+  long long unsigned total_cycle = 0;
   for (i = 0; i < 100; ++i) {
     assert(total_cycle >= 0 && 
             "total dynamic instruction cycle too large to be handled by llfi");

sample_programs/shufflenetv2_10/compile.sh

@@ -0,0 +1,13 @@
+printf "\n[Compile Script]: Convert TF model to LLVM IR\n"
+onnx-mlir --EmitLLVMIR  --instrument-onnx-ops="ALL" --InstrumentBeforeOp --InstrumentAfterOp $1.onnx
+mlir-translate -mlir-to-llvmir $1.onnx.mlir > model.mlir.ll
+
+printf "\n[Compile Script]: Compile main driver program and link to TF model in LLVM IR\n"
+clang++ -DONNX_ML=1 image.c -o main.ll -O0 -S -emit-llvm -lonnx_proto -lprotobuf -I$ONNX_MLIR_SRC/include 
+llvm-link -o model.ll -S main.ll model.mlir.ll
+
+printf "\n[Compile Script]: Generate model.exe \n"
+/home/llvm-project/build/bin/llc -filetype=obj -o model.o model.ll -O0 --relocation-model=pic
+clang++ -o model.exe model.o -L/home/LLTFI/LLTFI/build/bin/../runtime_lib -lllfi-rt -lpthread -L /Debug/lib -Wl,-rpath /home/LLTFI/LLTFI/build/bin/../runtime_lib -I$ONNX_MLIR_SRC/include -O0 -lonnx_proto -lprotobuf -lcruntime -ljson-c
+
+printf "\n[Compile Script]: Compilation complete\n"

sample_programs/shufflenetv2_10/image.c

@@ -0,0 +1,235 @@
+/*
+ * image.c - Sample program for Shufflenet
+ *
+ *
+ *
+ */
+
+#include <stdio.h>
+#include <OnnxMlirRuntime.h>
+#include "json-c/json.h"
+#include <string.h>
+#include <assert.h>
+#include <stdlib.h>
+#include "onnx/onnx_pb.h"
+#include <fstream>
+#include <vector>
+
+#define NUM_INPUTS 1 // When using the same image.c file for a different model, specify the number of inputs here depending on the model
+
+using namespace std;
+
+extern "C" {
+  OMTensorList *run_main_graph(OMTensorList *);
+}
+
+void export_layer_output_to_json(OMTensorList *, char*, char*);
+
+int main(int argc, char *argv[]) {
+
+    //Input pointers needed for the model
+    char *inp[NUM_INPUTS];
+    char* savefilename = "layeroutput.txt";
+    char* output_seq = NULL;
+    vector<void*> heapAllocs;
+
+    unsigned int numArguments = NUM_INPUTS+2; 
+    if (argc == numArguments) {
+    	for (int i = 0; i < NUM_INPUTS; i++){
+    		inp[i] = argv[i+1];
+    	}
+	output_seq = argv[NUM_INPUTS+1];
+    } else {
+        printf("Must supply the path to an image file.\n");
+    }
+
+    OMTensor *img_list[NUM_INPUTS];
+    for (int i = 0; i < NUM_INPUTS; i++) {
+	    onnx::TensorProto input;
+	    std::ifstream in(inp[i], std::ios_base::binary);
+            input.ParseFromIstream(&in);
+
+	    auto d = input.dims();
+
+	    // When using the same image.c file for a different model, Change "float" to the type of input data that the model expects
+	    float* input_data {reinterpret_cast<float *>(const_cast<char*>(input.raw_data().data()))};
+	    float* heap_input_data;
+	    long *in_shape = (long*)malloc(d.size()*sizeof(long));
+	    heapAllocs.push_back((void*)in_shape);
+
+	    int64_t totalSize = 1; 
+            std::cout<< d.size()<< std::endl;
+	    for(int j =0; j < d.size(); j++){
+		    in_shape[j] = d[j];
+		    totalSize *= d[j];
+		    std::cout << "in_shape[i] " << in_shape[j] << endl;
+	    }
+
+	    heap_input_data = (float*)malloc(sizeof(float) * totalSize);
+	    memcpy(heap_input_data, input_data, totalSize * sizeof(float));
+
+
+	    //When using the same image.c file for a different model, Change ONNX_TYPE depending on the input type
+	    OMTensor *x1 = omTensorCreate(heap_input_data, in_shape, d.size(), ONNX_TYPE_FLOAT);
+	    img_list[i] = x1;
+	    heapAllocs.push_back((void*)heap_input_data);
+    }
+
+    OMTensorList *graph_input = omTensorListCreate(img_list, 4);
+
+    // Call the compiled onnx model function.
+    OMTensorList *outputList = run_main_graph(graph_input);
+
+    // Export layer outputs to a JSON file
+    export_layer_output_to_json(outputList, savefilename, output_seq);
+
+    for (void* ptr : heapAllocs) {
+	   free(ptr);
+    }
+
+  return 0;
+}
+
+// Function to split a string
+char** str_split(char* a_str, const char a_delim, int* len)
+{
+    char** result    = 0;
+    size_t count     = 0;
+    char* tmp        = a_str;
+    char* last_comma = 0;
+    char delim[2];
+    delim[0] = a_delim;
+    delim[1] = 0;
+    *len = 0;
+
+    /* Count how many elements will be extracted. */
+    while (*tmp)
+    {
+        if (a_delim == *tmp)
+        {
+            count++;
+            last_comma = tmp;
+        }
+        tmp++;
+    }
+
+    /* Add space for trailing token. */
+    count += last_comma < (a_str + strlen(a_str) - 1);
+
+    /* Add space for terminating null string so caller
+       knows where the list of returned strings ends. */
+    count++;
+
+    result = (char**)malloc(sizeof(char*) * count);
+
+    if (result)
+    {
+        size_t idx  = 0;
+        char* token = strtok(a_str, delim);
+
+        while (token)
+        {
+            assert(idx < count);
+            *(result + idx++) = strdup(token);
+            token = strtok(0, delim);
+
+            *len = (*len) + 1;
+        }
+        assert(idx == count - 1);
+    }
+
+    return result;
+}
+
+// Convert a list of char**to int*
+int* convert_to_int(char** list, int count)
+{
+	int* result = (int*) malloc(count*sizeof(int));
+
+	for (int i = 0 ; i < count; i++)
+	{
+		result[i] = atoi(list[i]);
+	}
+
+	return result;
+}
+
+// Turn this on for debugging JSON creater.
+// #define DEBUG_MSG(...) printf(__VA_ARGS__)
+#define DEBUG_MSG(...)
+
+//Function to export layer outputs to JSON format.
+void export_layer_output_to_json(OMTensorList *outputList, char* savefile, char* expected_op_seq)
+{
+
+	int count = 0;
+	char** tokens = str_split(expected_op_seq, ',', &count);
+	int* layer_seq = convert_to_int(tokens, count);
+
+	// Global JSON object
+	json_object* jobj = json_object_new_object();
+
+	for (int64_t i = 0; i < omTensorListGetSize(outputList); i++) {
+
+		// JSON object for this layer
+		json_object* jobj_layer = json_object_new_object();
+
+        	DEBUG_MSG("Reading output of layer %lu\n", i);
+
+        	OMTensor *omt = omTensorListGetOmtByIndex(outputList, i);
+
+		// Get properties of the tensor that you want to export to the JSON file
+		int64_t rank = omTensorGetRank(omt);
+		int64_t *shape = omTensorGetShape(omt);
+		int64_t numElements = (int64_t) (omTensorGetNumElems(omt) / shape[0]);
+		float *dataBuf = (float *)omTensorGetDataPtr(omt);
+		int64_t bufferSize = omTensorGetBufferSize(omt);
+
+		DEBUG_MSG("Rank: %lu \nNumber of elements: %lu \n", rank, numElements);
+		DEBUG_MSG("Shape: ");
+		for (int64_t j = 0; j < rank; j++){
+			DEBUG_MSG("%lu,  ", shape[j]);
+		}
+		DEBUG_MSG("\n");
+
+		DEBUG_MSG("Buffer Size: %lu\n", bufferSize);
+
+		json_object *JLayerId = json_object_new_int(layer_seq[i]);	
+		json_object *JRank = json_object_new_int(rank);	
+		json_object *JNumElements = json_object_new_int(numElements);
+
+		json_object_object_add(jobj_layer, "Layer Id", JLayerId);	
+		json_object_object_add(jobj_layer, "Rank", JRank);	
+		json_object_object_add(jobj_layer, "Number of Elements", JNumElements);	
+
+
+		json_object* JShape = json_object_new_array();
+		for (int64_t j = 0; j < rank; j++) {
+			json_object* temp = json_object_new_int(shape[j]);
+			json_object_array_add(JShape, temp);
+		}
+
+
+		json_object* JData = json_object_new_array();
+		for (int64_t j = 0; j < numElements; j++) {
+			json_object* temp = json_object_new_double(dataBuf[j]);
+			json_object_array_add(JData, temp);
+		}
+
+		json_object_object_add(jobj_layer, "Shape", JShape);	
+		json_object_object_add(jobj_layer, "Data", JData);	
+
+		char str[5];
+		sprintf(str, "%ld", i);
+		json_object_object_add(jobj, str, jobj_layer);
+    	}
+
+    // Free heap memory
+    free(tokens);
+    free(layer_seq);
+
+	char* val = (char*) json_object_get_string(jobj);	
+	FILE* save = fopen(savefile, "w+");
+	fputs(val, save);
+	fclose(save);
+}

sample_programs/shufflenetv2_10/input.yaml

@@ -0,0 +1,30 @@
+compileOption:
+    instSelMethod:
+      - customInstselector:
+          include:
+            - CustomTensorOperator
+          options:
+            - -layerNo=0
+            - -layerName=all
+
+    regSelMethod: regloc
+    regloc: dstreg
+
+    includeInjectionTrace:
+        - forward
+
+    tracingPropagation: False # trace dynamic instruction values.
+
+    tracingPropagationOption:
+        maxTrace: 250 # max number of instructions to trace during fault injection run
+        debugTrace: False
+        mlTrace: False # enable for tracing ML programs
+        generateCDFG: True
+
+runOption:
+    - run:
+        numOfRuns: 1000
+        fi_type: bitflip
+        window_len_multiple_startindex: 1
+        window_len_multiple_endindex: 500
+        fi_max_multiple: 1

sample_programs/shufflenetv2_10/runllfi.sh

@@ -0,0 +1,4 @@
+rm -rf llfi*
+$LLFI_BUILD_ROOT/bin/instrument --readable -L $ONNX_MLIR_BUILD/Debug/lib -lcruntime -ljson-c -lprotobuf -lonnx_proto model.ll 
+$LLFI_BUILD_ROOT/bin/profile ./llfi/model-profiling.exe input_0.pb 0
+$LLFI_BUILD_ROOT/bin/injectfault ./llfi/model-faultinjection.exe input_0.pb 0