diff --git a/llvm/projects/hpvm-tensor-rt/CMakeLists.txt b/llvm/projects/hpvm-tensor-rt/CMakeLists.txt
index e7b6c1e9214236c5735df61d30933dec863efd02..87746c35294cb540998bef0b81a8dd51ad5b6589 100644
--- a/llvm/projects/hpvm-tensor-rt/CMakeLists.txt
+++ b/llvm/projects/hpvm-tensor-rt/CMakeLists.txt
@@ -48,6 +48,7 @@ endif()
target_link_libraries(tensor_cpu_runtime)
+
# Adding rule for the debugging source
add_executable(test_ops dnn_sources/src/test_ops.cc)
target_link_libraries(test_ops tensor_runtime)
@@ -59,12 +60,6 @@ target_link_libraries(fc2_cpu tensor_cpu_runtime)
# Full-Precision versions
-add_executable(fc2_clipped dnn_sources/src/fc2_clipped.cc)
-target_link_libraries(fc2_clipped tensor_runtime)
-
-add_executable(fc3_clipped dnn_sources/src/fc3_clipped.cc)
-target_link_libraries(fc3_clipped tensor_runtime)
-
add_executable(fc4_clipped dnn_sources/src/fc4_clipped.cc)
target_link_libraries(fc4_clipped tensor_runtime)
@@ -78,9 +73,6 @@ target_link_libraries(lenet_keras tensor_runtime)
add_executable(alexnet_cifar10 dnn_sources/src/alexnet_cifar10_front.cc)
target_link_libraries(alexnet_cifar10 tensor_runtime)
-add_executable(alexnet_cifar10_test dnn_sources/src/alexnet_cifar10_test.cc)
-target_link_libraries(alexnet_cifar10_test tensor_runtime)
-
add_executable(alexnet_cifar10_tuner dnn_sources/src/alexnet_cifar10_tuner.cc)
target_link_libraries(alexnet_cifar10_tuner tensor_runtime)
@@ -132,13 +124,6 @@ target_link_libraries(pipeline_GSME tensor_runtime)
#*** Half precision networks
-
-add_executable(fc2_half dnn_sources/src/half/fc2_half.cc)
-target_link_libraries(fc2_half tensor_runtime)
-
-add_executable(fc3_half dnn_sources/src/half/fc3_half.cc)
-target_link_libraries(fc3_half tensor_runtime)
-
add_executable(fc4_half dnn_sources/src/half/fc4_half.cc)
target_link_libraries(fc4_half tensor_runtime)
diff --git a/llvm/projects/hpvm-tensor-rt/dnn_sources/include/utils.h b/llvm/projects/hpvm-tensor-rt/dnn_sources/include/utils.h
index 2aa97bc4d91b616e2fbe2e6c73ef3fa996fa5b35..1d9fed9e2c05eaf35fad010b1daca40eb0828cc7 100644
--- a/llvm/projects/hpvm-tensor-rt/dnn_sources/include/utils.h
+++ b/llvm/projects/hpvm-tensor-rt/dnn_sources/include/utils.h
@@ -88,6 +88,22 @@ void fillWithOnesAndTwos(void* tensor_ptr){
}
+void fillTensorWithVal(void* tensor_ptr, float target_value){
+
+ struct Tensor* tensor = (struct Tensor*) tensor_ptr;
+
+ hpvm_request_tensor(tensor, 0);
+
+ // initialization is specific to the floating point type
+ if(tensor->data_type == CUDNN_DATA_FLOAT){
+ float* data_arr = (float*) tensor->host_data;
+ for(unsigned int i = 0; i < tensor->num_elems; i++){
+ data_arr[i] = target_value;
+ }
+ }
+}
+
+
void fillTensorWithNegOnes(void* tensor_ptr){
struct Tensor* tensor = (struct Tensor*) tensor_ptr;
@@ -223,6 +239,42 @@ void* readInputTensor(const char* file_name, int data_type, int dim1_size, int d
}
+//*** FIXIT: Move this to CPU-only
+struct Tensor* readTrainedWeightsCPU(const char* file_name, int data_type,
+ int dim1_size, int dim2_size,
+ int dim3_size, int dim4_size){
+
+ // FIXIT: Don't assume floating point types
+ int type_size = 4; // NOTE: Assuming floating point tensors
+ long int num_elems = dim1_size * dim2_size * dim3_size * dim4_size;
+ long int size_in_bytes = type_size * dim1_size * dim2_size * dim3_size * dim4_size;
+ float* tensor_data = (float*) malloc(sizeof(float) * num_elems);
+ int file_header_size = 0;
+
+ FILE* file = fopen(file_name, "rb");
+ if(file == NULL){
+ printf("Data file %s is not found. Aborting... \n", file_name);
+ abort();
+ }
+
+ fseek(file, file_header_size, SEEK_CUR); // Skipping the file header
+ size_t bytes_read = fread(tensor_data, 1, size_in_bytes, file);
+
+ printf("size in bytes = %lu, bytes read = %lu \n", size_in_bytes, bytes_read);
+
+ fclose(file);
+
+
+ struct Tensor* weights = (struct Tensor*) create4DTensor(data_type, nchw, dim1_size, dim2_size,
+ dim3_size, dim4_size);
+
+ initTensorData(weights, tensor_data, size_in_bytes);
+ //compareValues(weights, tensor_data, num_elems);
+ free(tensor_data);
+
+ return weights;
+}
+
struct Tensor* readTrainedWeights(const char* file_name, int data_type,
long int dim1_size, long int dim2_size,
@@ -416,7 +468,7 @@ float computeAccuracy2(uint8_t* labels, int num_labels, void* result_ptr, unsign
float accuracy = ((batch_dim - num_errors) * 1.0 / batch_dim * 1.0) * 100.0;
printf("****** Accuracy = %f \n\n", accuracy);
- printf("****** Zero class labels %d \n", num_zeros);
+ //printf("****** Zero class labels %d \n", num_zeros);
FILE* fp = fopen("final_accuracy", "w+");
if(fp != NULL){
@@ -519,6 +571,38 @@ void dumpFinalAccuracy(float accuracy){
}
+
+void dumpAvgPSNR(float avg_psnr){
+
+ FILE* fp = fopen("avg_psnr", "w+");
+ if(fp != NULL){
+ std::ostringstream ss;
+ ss << std::fixed << avg_psnr;
+ std::string print_str = ss.str();
+ fwrite(print_str.c_str(), 1, print_str.length(), fp);
+ }
+
+ fclose(fp);
+}
+
+
+void dumpPSNRStd(float psnr_std){
+
+ FILE* fp = fopen("psnr_std.txt", "w+");
+ if(fp != NULL){
+ std::ostringstream ss;
+ ss << std::fixed << psnr_std;
+ std::string print_str = ss.str();
+ fwrite(print_str.c_str(), 1, print_str.length(), fp);
+ }
+
+ fclose(fp);
+}
+
+
+
+
+
void dumpExecutionAccuracies(){
FILE* fp = fopen("run_accuracies.txt", "w+");
@@ -538,9 +622,27 @@ void dumpExecutionAccuracies(){
}
+float readPSNRFromFile(const char* file_name){
+
+ float psnr;
+ FILE* pFile = fopen(file_name, "r");
+ if(pFile == NULL){
+ printf("ERROR: psnr.txt not found! \n");
+ abort();
+ }
+
+ fscanf(pFile, "%f", &psnr);
+ printf("**** PSNR read = %f \n\n", psnr);
+ return psnr;
+}
+
float computePSNRViolation(void* gold_ptr, void* approx_ptr, float PSNR_threshold){
+
+ PSNR_threshold = readPSNRFromFile("psnr.txt");
+ std::vector<float> psnr_list;
+
struct Tensor* gold_tensor = (struct Tensor*) gold_ptr;
struct Tensor* approx_tensor = (struct Tensor*) approx_ptr;
@@ -548,11 +650,14 @@ float computePSNRViolation(void* gold_ptr, void* approx_ptr, float PSNR_threshol
size_t batch_dim = dim_sizes[0];
size_t image_size = dim_sizes[1] * dim_sizes[2] * dim_sizes[3];
- printf("batch_dim = %d, image_size = %d \n", batch_dim, image_size);
+ printf("batch_dim = %lu, image_size = %lu \n", batch_dim, image_size);
float* gold_data = (float*) gold_tensor->host_data;
float* approx_data = (float*) approx_tensor->host_data;
-
+
+ FILE* fp = fopen("img_psnr.txt", "w+");
+
+ float sum_psnr = 0.0;
int num_errors = 0;
for(size_t i = 0; i < batch_dim; i++){
float mse_sum = 0.0;
@@ -570,17 +675,44 @@ float computePSNRViolation(void* gold_ptr, void* approx_ptr, float PSNR_threshol
}
mse_sum = mse_sum / image_size;
- float psnr = 20 * log10(max_val / sqrt(mse_sum));
+ float psnr = 20 * log10(255 / sqrt(mse_sum));
+ sum_psnr += psnr;
if (psnr < PSNR_threshold)
num_errors += 1;
printf("PSNR value = %f \n", psnr);
- }
+ psnr_list.push_back(psnr);
+ std::ostringstream ss;
+ ss << std::fixed << psnr;
+ std::string print_str = ss.str();
+ fwrite(print_str.c_str(), 1, print_str.length(), fp);
+ fwrite("\n", 1, 1, fp);
+ }
float violation_rate = (num_errors * 1.0) / batch_dim * 100.0;
printf("*** violation_rate= %f \n\n", violation_rate);
+
+ float avg_psnr = sum_psnr / batch_dim;
+ printf("*** avg_psnr = %f \n\n", avg_psnr);
+ dumpAvgPSNR(avg_psnr);
+
+ float success_rate = 100.0 - violation_rate;
+ dumpFinalAccuracy(success_rate);
+
+ fclose(fp);
+
+
+ float var = 0.0;
+ for(size_t i = 0; i < batch_dim; i++){
+ var = var + (psnr_list[i] - avg_psnr) * (psnr_list[i] - avg_psnr);
+ }
+
+ var /= batch_dim;
+ float std = sqrt(var);
+
+ dumpPSNRStd(std);
return violation_rate;
}