From 03425866e7a72c1bf50cef375c75d26460f8c13e Mon Sep 17 00:00:00 2001
From: Yifan Zhao <yifanz16@illinois.edu>
Date: Thu, 25 Mar 2021 03:51:03 -0500
Subject: [PATCH] Fixed bugs in hpvm-c benchmarks
---
.../alexnet2_cifar10/alexnet2_cifar10.cpp | 72 +-
.../alexnet2_cifar10_cudnn.cpp | 54 +-
.../alexnet_cifar10/alexnet_cifar10.cpp | 69 +-
.../alexnet_cifar10/alexnet_cifar10_cudnn.cpp | 48 +-
.../alexnet_imagenet/alexnet_imagenet.cpp | 74 +-
.../alexnet_imagenet_cudnn.cpp | 51 +-
.../benchmarks/lenet_mnist/lenet_mnist.cpp | 48 +-
.../lenet_mnist/lenet_mnist_cudnn.cpp | 53 +-
.../mobilenet_cifar10/mobilenet_cifar10.cpp | 49 +-
.../mobilenet_cifar10_cudnn.cpp | 49 +-
.../resnet18_cifar10/resnet18_cifar10.cpp | 74 +-
.../resnet18_cifar10_cudnn.cpp | 51 +-
.../resnet50_imagenet/resnet50_imagenet.cpp | 64 +-
.../resnet50_imagenet_cudnn.cpp | 48 +-
.../vgg16_cifar10/vgg16_cifar10.cpp | 67 +-
.../vgg16_cifar10/vgg16_cifar10_cudnn.cpp | 60 +-
.../vgg16_cifar100/vgg16_cifar100.cpp | 68 +-
.../vgg16_cifar100/vgg16_cifar100_cudnn.cpp | 51 +-
.../vgg16_imagenet/vgg16_imagenet.cpp | 65 +-
.../vgg16_imagenet/vgg16_imagenet_cudnn.cpp | 48 +-
.../hpvm-c/include/tensorUtils.h | 663 ++++++++----------
21 files changed, 771 insertions(+), 1055 deletions(-)
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet2_cifar10/alexnet2_cifar10.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet2_cifar10/alexnet2_cifar10.cpp
index dd689d202a..35f8188f78 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet2_cifar10/alexnet2_cifar10.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet2_cifar10/alexnet2_cifar10.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -411,10 +405,12 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+const int batch_size = 500, input_size = 5000,
+ batch_count = input_size / batch_size;
+int main() {
std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/alexnet2_cifar10/";
-
+ std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
void *conv2d_1_w =
@@ -458,12 +454,10 @@ int main() {
std::string dense_1_b_path = dir_prefix + std::string("dense_1_b.bin");
void *dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0, 1, 10, 1, 1);
- // void* input = readTrainedWeights(input_path.c_str(), 0,10000,3,32,32);
- // uint8_t* labels = readLabels(labels_path.c_str(),10000);
-
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
+ args->input = input;
+ args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
args->conv2d_1_w_bytes = 0;
args->conv2d_1_b = conv2d_1_b;
@@ -493,45 +487,21 @@ int main() {
args->dense_1_b = dense_1_b;
args->dense_1_b_bytes = 0;
- int batch_size = 500;
- int test_input_size = 5000;
- int batch_count = test_input_size / batch_size;
-
- std::string input_path = dir_prefix + std::string("test_input.bin");
- // void* input = create4DTensor(0,nchw,batch_size,3,32,32);
-
+ __hpvm__init();
startMemTracking();
- startProfiling();
-
- for (int j = 0; j < 1; j++) {
- for (int i = 0; i < batch_count; i++) {
-
- int start = i * batch_size;
- int end = (i + 1) * batch_size;
-
- // copyInputBatch(input_path.c_str(),start,end,3,32,32, input);
-
- void *input =
- readInputBatch(input_path.c_str(), 0, start, end, 3, 32, 32);
-
- args->input = input;
- args->input_bytes = 0;
-
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
- llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
-
- freeBatchMemory();
- }
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 32, 32, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
+ freeBatchMemory();
}
-
- stopProfiling();
__hpvm__cleanup();
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet2_cifar10/alexnet2_cifar10_cudnn.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet2_cifar10/alexnet2_cifar10_cudnn.cpp
index 448db0b4c8..5bcc5b627b 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet2_cifar10/alexnet2_cifar10_cudnn.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet2_cifar10/alexnet2_cifar10_cudnn.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -411,10 +405,17 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+void write_accuracy(float accuracy) {
+ std::ofstream fout("final_accuracy");
+ fout << std::fixed << accuracy;
+}
- std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/alexnet2_cifar10/";
+const int batch_size = 500, input_size = 5000,
+ batch_count = input_size / batch_size;
+int main() {
+ std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/alexnet2_cifar10/";
+ std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
void *conv2d_1_w =
@@ -458,9 +459,10 @@ int main() {
std::string dense_1_b_path = dir_prefix + std::string("dense_1_b.bin");
void *dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0, 1, 10, 1, 1);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
+ args->input = input;
+ args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
args->conv2d_1_w_bytes = 0;
args->conv2d_1_b = conv2d_1_b;
@@ -490,41 +492,25 @@ int main() {
args->dense_1_b = dense_1_b;
args->dense_1_b_bytes = 0;
- int batch_size = 500;
- int test_input_size = 5000;
- int batch_count = test_input_size / batch_size;
-
- std::string input_path = dir_prefix + std::string("test_input.bin");
- void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
-
+ __hpvm__init();
+ float total_accuracy = 0;
startMemTracking();
+#pragma clang loop unroll(disable)
for (int i = 0; i < batch_count; i++) {
-
- int start = i * batch_size;
- int end = (i + 1) * batch_size;
-
+ int start = i * batch_size, end = start + batch_size;
copyInputBatch(input_path.c_str(), start, end, 3, 32, 32, input);
- args->input = input;
- args->input_bytes = 0;
-
void *dfg = __hpvm__launch(0, root, (void *)args);
-
__hpvm__wait(dfg);
-
void *result = static_cast<RootIn *>(args)->r.tensor;
hpvm_request_tensor(result, 0);
uint32_t *labels = readLabelsBatch3(labels_path.c_str(), start, end);
-
- computeAccuracy3(labels, result);
-
- // llvm_hpvm_invokeRtControl2(result, labels);
-
+ float accuracy = computeAccuracy3(labels, result);
+ total_accuracy += accuracy * batch_size;
freeBatchMemory();
}
-
+ write_accuracy(total_accuracy / input_size);
__hpvm__cleanup();
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_cifar10/alexnet_cifar10.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_cifar10/alexnet_cifar10.cpp
index ae5f31b7dc..51e0dd137d 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_cifar10/alexnet_cifar10.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_cifar10/alexnet_cifar10.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -362,12 +356,12 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+const int batch_size = 500, input_size = 5000,
+ batch_count = input_size / batch_size;
+int main() {
std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/alexnet_cifar10/";
-
std::string input_path = dir_prefix + std::string("test_input.bin");
- // void* input = readTrainedWeights(input_path.c_str(), 0,5000,3,32,32);
std::string labels_path = dir_prefix + std::string("test_labels.bin");
uint8_t *labels = readLabels(labels_path.c_str(), 5000);
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
@@ -406,11 +400,10 @@ int main() {
std::string dense_1_b_path = dir_prefix + std::string("dense_1_b.bin");
void *dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0, 1, 10, 1, 1);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
- // args->input = input;
- // args->input_bytes = 0;
+ void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
+ args->input = input;
+ args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
args->conv2d_1_w_bytes = 0;
args->conv2d_1_b = conv2d_1_b;
@@ -436,43 +429,21 @@ int main() {
args->dense_1_b = dense_1_b;
args->dense_1_b_bytes = 0;
- int batch_size = 500;
- int test_input_size = 5000;
- int batch_count = test_input_size / batch_size;
-
+ __hpvm__init();
startMemTracking();
- startProfiling();
-
- for (int j = 0; j < 1; j++) {
- for (int i = 0; i < batch_count; i++) {
-
- int start = i * batch_size;
- int end = (i + 1) * batch_size;
-
- // copyInputBatch(input_path.c_str(),start,end,3,32,32, input);
-
- // Replaced create4DTensor and copyInputBatch with readInputBatch
- void *input =
- readInputBatch(input_path.c_str(), 0, start, end, 3, 32, 32);
-
- args->input = input;
- args->input_bytes = 0;
-
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
- llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
-
- freeBatchMemory();
- }
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 32, 32, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
+ freeBatchMemory();
}
-
- stopProfiling();
__hpvm__cleanup();
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_cifar10/alexnet_cifar10_cudnn.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_cifar10/alexnet_cifar10_cudnn.cpp
index ab4f6a2402..74c5420fd9 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_cifar10/alexnet_cifar10_cudnn.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_cifar10/alexnet_cifar10_cudnn.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -363,12 +357,17 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+void write_accuracy(float accuracy) {
+ std::ofstream fout("final_accuracy");
+ fout << std::fixed << accuracy;
+}
- std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/alexnet_cifar10/";
+const int batch_size = 500, input_size = 5000,
+ batch_count = input_size / batch_size;
+int main() {
+ std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/alexnet_cifar10/";
std::string input_path = dir_prefix + std::string("test_input.bin");
- void *input = readTrainedWeights(input_path.c_str(), 0, 5000, 3, 32, 32);
std::string labels_path = dir_prefix + std::string("test_labels.bin");
uint32_t *labels = readLabels3(labels_path.c_str(), 5000);
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
@@ -407,9 +406,8 @@ int main() {
std::string dense_1_b_path = dir_prefix + std::string("dense_1_b.bin");
void *dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0, 1, 10, 1, 1);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
args->input = input;
args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
@@ -436,14 +434,26 @@ int main() {
args->dense_1_w_bytes = 0;
args->dense_1_b = dense_1_b;
args->dense_1_b_bytes = 0;
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
- void *result = static_cast<RootIn *>(args)->r.tensor;
-
- hpvm_request_tensor(result, 0);
+ __hpvm__init();
+ float total_accuracy = 0;
+ startMemTracking();
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 32, 32, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ uint32_t *labels = readLabelsBatch3(labels_path.c_str(), start, end);
+ float accuracy = computeAccuracy3(labels, result);
+ total_accuracy += accuracy * batch_size;
+ freeBatchMemory();
+ }
+ write_accuracy(total_accuracy / input_size);
__hpvm__cleanup();
- computeAccuracy3(labels, result);
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_imagenet/alexnet_imagenet.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_imagenet/alexnet_imagenet.cpp
index d49c0d2d06..16bcecf939 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_imagenet/alexnet_imagenet.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_imagenet/alexnet_imagenet.cpp
@@ -1,14 +1,8 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <config.h>
#include <hpvm.h>
+#include <string>
#include <tensorTypes.h>
#include <tensorUtils.h>
-#include <config.h>
void var_0_node(void *t1, size_t bytes_t1, void *t2, size_t bytes_t2) {
__hpvm__hint(hpvm::TENSOR_TARGET);
@@ -460,11 +454,11 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
-
- std::string dir_prefix =
- std::string(MODEL_PARAMS_DIR) + "/alexnet_imagenet/";
+const int batch_size = 100, input_size = 5000,
+ batch_count = input_size / batch_size;
+int main() {
+ std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/alexnet_imagenet/";
std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
@@ -515,16 +509,11 @@ int main() {
std::string dense_3_b_path = dir_prefix + std::string("dense_3_b.bin");
void *dense_3_b =
readTrainedWeights(dense_3_b_path.c_str(), 0, 1, 1000, 1, 1);
- // void* input = readTrainedWeights(input_path.c_str(), 0, 1000,3,224,224);
- // uint32_t* labels = readLabels2(labels_path.c_str(),6000);
- // uint32_t* labels = readLabels3(labels_path.c_str(), 1000);
-
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
- // args->input = input;
- // args->input_bytes = 0;
+ void *input = create4DTensor(0, nchw, batch_size, 3, 224, 224);
+ args->input = input;
+ args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
args->conv2d_1_w_bytes = 0;
args->conv2d_1_b = conv2d_1_b;
@@ -558,40 +547,21 @@ int main() {
args->dense_3_b = dense_3_b;
args->dense_3_b_bytes = 0;
- int batch_size = 200;
- int test_input_size = 4000;
- int batch_count = test_input_size / batch_size;
-
+ __hpvm__init();
startMemTracking();
- startProfiling();
-
- for (int j = 0; j < 1; j++) {
- for (int i = 0; i < batch_count; i++) {
-
- int start = i * batch_size;
- int end = (i + 1) * batch_size;
-
- void *input =
- readInputBatch(input_path.c_str(), 0, start, end, 3, 224, 224);
-
- args->input = input;
- args->input_bytes = 0;
-
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
- llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
-
- freeBatchMemory();
- }
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 224, 224, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
+ freeBatchMemory();
}
-
- stopProfiling();
__hpvm__cleanup();
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_imagenet/alexnet_imagenet_cudnn.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_imagenet/alexnet_imagenet_cudnn.cpp
index 4acba95173..5ddd969432 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_imagenet/alexnet_imagenet_cudnn.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/alexnet_imagenet/alexnet_imagenet_cudnn.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -460,8 +454,15 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+void write_accuracy(float accuracy) {
+ std::ofstream fout("final_accuracy");
+ fout << std::fixed << accuracy;
+}
+const int batch_size = 100, input_size = 5000,
+ batch_count = input_size / batch_size;
+
+int main() {
std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/alexnet_imagenet/";
std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
@@ -513,14 +514,9 @@ int main() {
std::string dense_3_b_path = dir_prefix + std::string("dense_3_b.bin");
void *dense_3_b =
readTrainedWeights(dense_3_b_path.c_str(), 0, 1, 1000, 1, 1);
- void *input = readTrainedWeights(input_path.c_str(), 0, 1000, 3, 224, 224);
- // uint32_t* labels = readLabels2(labels_path.c_str(),6000);
-
- uint32_t *labels = readLabels3(labels_path.c_str(), 1000);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 3, 224, 224);
args->input = input;
args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
@@ -556,14 +552,25 @@ int main() {
args->dense_3_b = dense_3_b;
args->dense_3_b_bytes = 0;
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
+ __hpvm__init();
+ float total_accuracy = 0;
+ startMemTracking();
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 224, 224, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ uint32_t *labels = readLabelsBatch3(labels_path.c_str(), start, end);
+ float accuracy = computeAccuracy3(labels, result);
+ total_accuracy += accuracy * batch_size;
+ freeBatchMemory();
+ }
+ write_accuracy(total_accuracy / input_size);
__hpvm__cleanup();
- computeAccuracy3(labels, result);
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/lenet_mnist/lenet_mnist.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/lenet_mnist/lenet_mnist.cpp
index b67d585d01..ee81665ec9 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/lenet_mnist/lenet_mnist.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/lenet_mnist/lenet_mnist.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -264,13 +258,13 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+const int batch_size = 1000, input_size = 5000,
+ batch_count = input_size / batch_size;
+int main() {
std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/lenet_mnist/";
-
std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
-
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
void *conv2d_1_w =
readTrainedWeights(conv2d_1_w_path.c_str(), 0, 32, 1, 5, 5);
@@ -294,15 +288,11 @@ int main() {
readTrainedWeights(dense_2_w_path.c_str(), 0, 1, 1, 1024, 10);
std::string dense_2_b_path = dir_prefix + std::string("dense_2_b.bin");
void *dense_2_b = readTrainedWeights(dense_2_b_path.c_str(), 0, 1, 10, 1, 1);
- // void* input = readTrainedWeights(input_path.c_str(), 0, 5000,1,28,28);
- // uint32_t* labels = readLabels3(labels_path.c_str(), 5000);
-
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
- // args->input = input;
- // args->input_bytes = 0;
+ void *input = create4DTensor(0, nchw, batch_size, 1, 28, 28);
+ args->input = input;
+ args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
args->conv2d_1_w_bytes = 0;
args->conv2d_1_b = conv2d_1_b;
@@ -320,37 +310,21 @@ int main() {
args->dense_2_b = dense_2_b;
args->dense_2_b_bytes = 0;
- int batch_size = 500;
- int test_input_size = 5000;
- int batch_count = test_input_size / batch_size;
-
+ __hpvm__init();
startMemTracking();
- startProfiling();
-
+#pragma clang loop unroll(disable)
for (int i = 0; i < batch_count; i++) {
-
- int start = i * batch_size;
- int end = (i + 1) * batch_size;
-
- void *input = readInputBatch(input_path.c_str(), 0, start, end, 1, 28, 28);
-
- args->input = input;
- args->input_bytes = 0;
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 1, 28, 28, input);
void *dfg = __hpvm__launch(0, root, (void *)args);
-
__hpvm__wait(dfg);
-
void *result = static_cast<RootIn *>(args)->r.tensor;
hpvm_request_tensor(result, 0);
llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
-
freeBatchMemory();
}
-
- stopProfiling();
__hpvm__cleanup();
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/lenet_mnist/lenet_mnist_cudnn.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/lenet_mnist/lenet_mnist_cudnn.cpp
index 2159cfe286..eecc7f5d60 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/lenet_mnist/lenet_mnist_cudnn.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/lenet_mnist/lenet_mnist_cudnn.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -264,13 +258,18 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+void write_accuracy(float accuracy) {
+ std::ofstream fout("final_accuracy");
+ fout << std::fixed << accuracy;
+}
- std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/lenet_mnist/";
+const int batch_size = 1000, input_size = 5000,
+ batch_count = input_size / batch_size;
+int main() {
+ std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/lenet_mnist/";
std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
-
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
void *conv2d_1_w =
readTrainedWeights(conv2d_1_w_path.c_str(), 0, 32, 1, 5, 5);
@@ -294,13 +293,9 @@ int main() {
readTrainedWeights(dense_2_w_path.c_str(), 0, 1, 1, 1024, 10);
std::string dense_2_b_path = dir_prefix + std::string("dense_2_b.bin");
void *dense_2_b = readTrainedWeights(dense_2_b_path.c_str(), 0, 1, 10, 1, 1);
- void *input = readTrainedWeights(input_path.c_str(), 0, 5000, 1, 28, 28);
-
- uint32_t *labels = readLabels3(labels_path.c_str(), 5000);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 1, 28, 28);
args->input = input;
args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
@@ -320,15 +315,25 @@ int main() {
args->dense_2_b = dense_2_b;
args->dense_2_b_bytes = 0;
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
+ __hpvm__init();
+ float total_accuracy = 0;
+ startMemTracking();
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 1, 28, 28, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ uint32_t *labels = readLabelsBatch3(labels_path.c_str(), start, end);
+ float accuracy = computeAccuracy3(labels, result);
+ total_accuracy += accuracy * batch_size;
+ freeBatchMemory();
+ }
+ write_accuracy(total_accuracy / input_size);
__hpvm__cleanup();
- computeAccuracy3(labels, result);
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/mobilenet_cifar10/mobilenet_cifar10.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/mobilenet_cifar10/mobilenet_cifar10.cpp
index a4de282621..58051e0993 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/mobilenet_cifar10/mobilenet_cifar10.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/mobilenet_cifar10/mobilenet_cifar10.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -1965,8 +1959,12 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
+const int batch_size = 500, input_size = 5000,
+ batch_count = input_size / batch_size;
+
int main() {
- std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/mobilenet_cifar10/";
+ std::string dir_prefix =
+ std::string(MODEL_PARAMS_DIR) + "/mobilenet_cifar10/";
std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
@@ -2501,14 +2499,11 @@ int main() {
readTrainedWeights(dense_1_w_path.c_str(), 0, 1, 1, 1024, 10);
std::string dense_1_b_path = dir_prefix + std::string("dense_1_b.bin");
void *dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0, 1, 10, 1, 1);
- // void* input = readTrainedWeights(input_path.c_str(), 0, 5000,3,32,32);
- // uint8_t* labels = readLabels(labels_path.c_str(), 5000);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
- // args->input = input;
- // args->input_bytes = 0;
+ void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
+ args->input = input;
+ args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
args->conv2d_1_w_bytes = 0;
args->batch_normalization_1_gamma = batch_normalization_1_gamma;
@@ -2784,39 +2779,21 @@ int main() {
args->dense_1_b = dense_1_b;
args->dense_1_b_bytes = 0;
- int batch_size = 500;
- int test_input_size = 5000;
- int batch_count = test_input_size / batch_size;
-
- // void* input = create4DTensor(0,nchw,batch_size,3,32,32);
-
+ __hpvm__init();
startMemTracking();
- startProfiling();
-
+#pragma clang loop unroll(disable)
for (int i = 0; i < batch_count; i++) {
-
- int start = i * batch_size;
- int end = (i + 1) * batch_size;
-
- // copyInputBatch(input_path.c_str(),start,end,3,32,32, input);
- void *input = readInputBatch(input_path.c_str(), 0, start, end, 3, 32, 32);
-
- args->input = input;
- args->input_bytes = 0;
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 32, 32, input);
void *dfg = __hpvm__launch(0, root, (void *)args);
-
__hpvm__wait(dfg);
-
void *result = static_cast<RootIn *>(args)->r.tensor;
hpvm_request_tensor(result, 0);
llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
-
freeBatchMemory();
}
- stopProfiling();
__hpvm__cleanup();
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/mobilenet_cifar10/mobilenet_cifar10_cudnn.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/mobilenet_cifar10/mobilenet_cifar10_cudnn.cpp
index 9f4069b34b..482a37d4c4 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/mobilenet_cifar10/mobilenet_cifar10_cudnn.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/mobilenet_cifar10/mobilenet_cifar10_cudnn.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -1965,11 +1959,17 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+void write_accuracy(float accuracy) {
+ std::ofstream fout("final_accuracy");
+ fout << std::fixed << accuracy;
+}
+const int batch_size = 500, input_size = 5000,
+ batch_count = input_size / batch_size;
+
+int main() {
std::string dir_prefix =
std::string(MODEL_PARAMS_DIR) + "/mobilenet_cifar10/";
-
std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
@@ -2503,12 +2503,10 @@ int main() {
readTrainedWeights(dense_1_w_path.c_str(), 0, 1, 1, 1024, 10);
std::string dense_1_b_path = dir_prefix + std::string("dense_1_b.bin");
void *dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0, 1, 10, 1, 1);
- void *input = readTrainedWeights(input_path.c_str(), 0, 5000, 3, 32, 32);
uint32_t *labels = readLabels3(labels_path.c_str(), 5000);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
args->input = input;
args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
@@ -2786,14 +2784,25 @@ int main() {
args->dense_1_b = dense_1_b;
args->dense_1_b_bytes = 0;
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
+ __hpvm__init();
+ float total_accuracy = 0;
+ startMemTracking();
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 32, 32, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ uint32_t *labels = readLabelsBatch3(labels_path.c_str(), start, end);
+ float accuracy = computeAccuracy3(labels, result);
+ total_accuracy += accuracy * batch_size;
+ freeBatchMemory();
+ }
+ write_accuracy(total_accuracy / input_size);
__hpvm__cleanup();
- computeAccuracy3(labels, result);
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet18_cifar10/resnet18_cifar10.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet18_cifar10/resnet18_cifar10.cpp
index 66ab37cd33..a254a62570 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet18_cifar10/resnet18_cifar10.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet18_cifar10/resnet18_cifar10.cpp
@@ -1,11 +1,5 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -1300,14 +1294,13 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
+const int batch_size = 500, input_size = 5000,
+ batch_count = input_size / batch_size;
+
int main() {
std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/resnet18_cifar10/";
-
std::string input_path = dir_prefix + std::string("test_input.bin");
- // void* input = readTrainedWeights(input_path.c_str(), 0,5000,3,32,32);
std::string labels_path = dir_prefix + std::string("test_labels.bin");
- // uint32_t* labels = readLabels3(labels_path.c_str(),5000);
-
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
void *conv2d_1_w =
readTrainedWeights(conv2d_1_w_path.c_str(), 0, 16, 3, 3, 3);
@@ -1439,11 +1432,10 @@ int main() {
std::string dense_1_b_path = dir_prefix + std::string("dense_1_b.bin");
void *dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0, 1, 10, 1, 1);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
- // args->input = input;
- // args->input_bytes = 0;
+ void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
+ args->input = input;
+ args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
args->conv2d_1_w_bytes = 0;
args->conv2d_1_b = conv2d_1_b;
@@ -1533,47 +1525,21 @@ int main() {
args->dense_1_b = dense_1_b;
args->dense_1_b_bytes = 0;
- int batch_size = 500;
- int test_input_size = 5000;
- int batch_count = test_input_size / batch_size;
-
- // NOTE-HASHIM: commented out
- // void* input = create4DTensor(0,nchw,batch_size,3,32,32);
-
+ __hpvm__init();
startMemTracking();
- startProfiling();
-
- for (int j = 0; j < 1; j++) {
- for (int i = 0; i < batch_count; i++) {
-
- int start = i * batch_size;
- int end = (i + 1) * batch_size;
-
- // copyInputBatch(input_path.c_str(),start,end,3,32,32, input);
-
- // NOTE-HASHIM: Commented out above line and line that does create4DTensor
- void *input =
- readInputBatch(input_path.c_str(), 0, start, end, 3, 32, 32);
-
- args->input = input;
- args->input_bytes = 0;
-
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
- llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
-
- printf("RUNNING BATCH = %d \n", i);
-
- freeBatchMemory();
- }
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 32, 32, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
+ freeBatchMemory();
}
-
- stopProfiling();
__hpvm__cleanup();
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet18_cifar10/resnet18_cifar10_cudnn.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet18_cifar10/resnet18_cifar10_cudnn.cpp
index 1b6c98b886..da1ce91ba3 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet18_cifar10/resnet18_cifar10_cudnn.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet18_cifar10/resnet18_cifar10_cudnn.cpp
@@ -1,11 +1,5 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -1226,12 +1220,17 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+void write_accuracy(float accuracy) {
+ std::ofstream fout("final_accuracy");
+ fout << std::fixed << accuracy;
+}
- std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/resnet18_cifar10/";
+const int batch_size = 500, input_size = 5000,
+ batch_count = input_size / batch_size;
+int main() {
+ std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/resnet18_cifar10/";
std::string input_path = dir_prefix + std::string("test_input.bin");
- void *input = readTrainedWeights(input_path.c_str(), 0, 5000, 3, 32, 32);
std::string labels_path = dir_prefix + std::string("test_labels.bin");
uint32_t *labels = readLabels3(labels_path.c_str(), 5000);
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
@@ -1365,9 +1364,8 @@ int main() {
std::string dense_1_b_path = dir_prefix + std::string("dense_1_b.bin");
void *dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0, 1, 10, 1, 1);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
args->input = input;
args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
@@ -1459,16 +1457,25 @@ int main() {
args->dense_1_b = dense_1_b;
args->dense_1_b_bytes = 0;
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
+ __hpvm__init();
+ float total_accuracy = 0;
+ startMemTracking();
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 32, 32, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ uint32_t *labels = readLabelsBatch3(labels_path.c_str(), start, end);
+ float accuracy = computeAccuracy3(labels, result);
+ total_accuracy += accuracy * batch_size;
+ freeBatchMemory();
+ }
+ write_accuracy(total_accuracy / input_size);
__hpvm__cleanup();
-
- computeAccuracy3(labels, result);
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet50_imagenet/resnet50_imagenet.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet50_imagenet/resnet50_imagenet.cpp
index db6b64daa0..a3ece5fede 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet50_imagenet/resnet50_imagenet.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet50_imagenet/resnet50_imagenet.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -5132,8 +5126,10 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+const int batch_size = 25, input_size = 5000,
+ batch_count = input_size / batch_size;
+int main() {
std::string dir_prefix =
std::string(MODEL_PARAMS_DIR) + "/resnet50_imagenet/";
std::string input_path = dir_prefix + std::string("test_input.bin");
@@ -6311,12 +6307,10 @@ int main() {
void *dense_1_b =
readTrainedWeights(dense_1_b_path.c_str(), 0, 1, 1000, 1, 1);
- // void* input = readTrainedWeights(input_path.c_str(), 0,100,3,224,224);
- // uint32_t* labels = readLabelsBatch3(labels_path.c_str(),0,100);
-
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 3, 224, 224);
+ args->input = input;
+ args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
args->conv2d_1_w_bytes = 0;
args->conv2d_1_b = conv2d_1_b;
@@ -6958,39 +6952,21 @@ int main() {
args->dense_1_b = dense_1_b;
args->dense_1_b_bytes = 0;
+ __hpvm__init();
startMemTracking();
- startProfiling();
-
- unsigned int batch_size = 50;
- unsigned int test_input_size = 1000;
- unsigned int batch_count = test_input_size / batch_size;
-
- for (int j = 0; j < 1; j++) {
- for (int i = 0; i < batch_count; i++) {
- int start = i * batch_size;
- int end = (i + 1) * batch_size;
-
- void *input =
- readInputBatch(input_path.c_str(), 0, start, end, 3, 224, 224);
-
- args->input = input;
- args->input_bytes = 0;
-
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
- llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
-
- freeBatchMemory();
- }
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 224, 224, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
+ freeBatchMemory();
}
-
- stopProfiling();
__hpvm__cleanup();
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet50_imagenet/resnet50_imagenet_cudnn.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet50_imagenet/resnet50_imagenet_cudnn.cpp
index c7817caf53..03674b50a5 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet50_imagenet/resnet50_imagenet_cudnn.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/resnet50_imagenet/resnet50_imagenet_cudnn.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -4903,8 +4897,15 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+void write_accuracy(float accuracy) {
+ std::ofstream fout("final_accuracy");
+ fout << std::fixed << accuracy;
+}
+const int batch_size = 50, input_size = 5000,
+ batch_count = input_size / batch_size;
+
+int main() {
std::string dir_prefix =
std::string(MODEL_PARAMS_DIR) + "/resnet50_imagenet/";
std::string input_path = dir_prefix + std::string("test_input.bin");
@@ -6081,12 +6082,10 @@ int main() {
std::string dense_1_b_path = dir_prefix + std::string("dense_1_b.bin");
void *dense_1_b =
readTrainedWeights(dense_1_b_path.c_str(), 0, 1, 1000, 1, 1);
- void *input = readTrainedWeights(input_path.c_str(), 0, 100, 3, 224, 224);
uint32_t *labels = readLabels3(labels_path.c_str(), 100);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 3, 224, 224);
args->input = input;
args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
@@ -6730,14 +6729,25 @@ int main() {
args->dense_1_b = dense_1_b;
args->dense_1_b_bytes = 0;
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
+ __hpvm__init();
+ float total_accuracy = 0;
+ startMemTracking();
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 224, 224, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ uint32_t *labels = readLabelsBatch3(labels_path.c_str(), start, end);
+ float accuracy = computeAccuracy3(labels, result);
+ total_accuracy += accuracy * batch_size;
+ freeBatchMemory();
+ }
+ write_accuracy(total_accuracy / input_size);
__hpvm__cleanup();
- computeAccuracy3(labels, result);
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar10/vgg16_cifar10.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar10/vgg16_cifar10.cpp
index 39c2ffc876..cad22649fd 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar10/vgg16_cifar10.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar10/vgg16_cifar10.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -827,8 +821,10 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+const int batch_size = 500, input_size = 5000,
+ batch_count = input_size / batch_size;
+int main() {
std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/vgg16_cifar10/";
std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
@@ -920,14 +916,11 @@ int main() {
readTrainedWeights(dense_2_w_path.c_str(), 0, 1, 1, 512, 10);
std::string dense_2_b_path = dir_prefix + std::string("dense_2_b.bin");
void *dense_2_b = readTrainedWeights(dense_2_b_path.c_str(), 0, 1, 10, 1, 1);
- // void* input = readTrainedWeights(input_path.c_str(), 0,2000,3,32,32);
- // uint32_t* labels = readLabels3(labels_path.c_str(),2000);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
- // args->input = input;
- // args->input_bytes = 0;
+ void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
+ args->input = input;
+ args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
args->conv2d_1_w_bytes = 0;
args->conv2d_1_b = conv2d_1_b;
@@ -989,41 +982,21 @@ int main() {
args->dense_2_b = dense_2_b;
args->dense_2_b_bytes = 0;
- int batch_size = 500;
- int test_input_size = 5000;
- int batch_count = test_input_size / batch_size;
-
+ __hpvm__init();
startMemTracking();
- startProfiling();
-
- for (int j = 0; j < 1; j++) {
- for (int i = 0; i < batch_count; i++) {
-
- int start = i * batch_size;
- int end = (i + 1) * batch_size;
-
- // copyInputBatch(input_path.c_str(),start,end,3,32,32, input);
- void *input =
- readInputBatch(input_path.c_str(), 0, start, end, 3, 32, 32);
-
- args->input = input;
- args->input_bytes = 0;
-
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
- llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
-
- freeBatchMemory();
- }
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 32, 32, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
+ freeBatchMemory();
}
-
- stopProfiling();
__hpvm__cleanup();
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar10/vgg16_cifar10_cudnn.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar10/vgg16_cifar10_cudnn.cpp
index 2f18dc17c6..6625202828 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar10/vgg16_cifar10_cudnn.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar10/vgg16_cifar10_cudnn.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -827,8 +821,15 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+void write_accuracy(float accuracy) {
+ std::ofstream fout("final_accuracy");
+ fout << std::fixed << accuracy;
+}
+const int batch_size = 500, input_size = 5000,
+ batch_count = input_size / batch_size;
+
+int main() {
std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/vgg16_cifar10/";
std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
@@ -920,12 +921,10 @@ int main() {
readTrainedWeights(dense_2_w_path.c_str(), 0, 1, 1, 512, 10);
std::string dense_2_b_path = dir_prefix + std::string("dense_2_b.bin");
void *dense_2_b = readTrainedWeights(dense_2_b_path.c_str(), 0, 1, 10, 1, 1);
- void *input = readTrainedWeights(input_path.c_str(), 0, 2000, 3, 32, 32);
uint32_t *labels = readLabels3(labels_path.c_str(), 2000);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
args->input = input;
args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
@@ -989,28 +988,25 @@ int main() {
args->dense_2_b = dense_2_b;
args->dense_2_b_bytes = 0;
+ __hpvm__init();
+ float total_accuracy = 0;
startMemTracking();
- startProfiling();
-
- input = readTrainedWeights(input_path.c_str(), 0, 2000, 3, 32, 32);
-
- args->input = input;
- args->input_bytes = 0;
-
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
- computeAccuracy3(labels, result);
-
- freeBatchMemory();
-
- stopProfiling();
-
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 32, 32, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ uint32_t *labels = readLabelsBatch3(labels_path.c_str(), start, end);
+ float accuracy = computeAccuracy3(labels, result);
+ total_accuracy += accuracy * batch_size;
+ freeBatchMemory();
+ }
+ write_accuracy(total_accuracy / input_size);
__hpvm__cleanup();
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar100/vgg16_cifar100.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar100/vgg16_cifar100.cpp
index ce899cd0a2..54417171fb 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar100/vgg16_cifar100.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar100/vgg16_cifar100.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -827,10 +821,11 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+const int batch_size = 500, input_size = 5000,
+ batch_count = input_size / batch_size;
+int main() {
std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/vgg16_cifar100/";
-
std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
@@ -922,14 +917,10 @@ int main() {
std::string dense_2_b_path = dir_prefix + std::string("dense_2_b.bin");
void *dense_2_b = readTrainedWeights(dense_2_b_path.c_str(), 0, 1, 100, 1, 1);
- // void* input = readTrainedWeights(input_path.c_str(), 0,2000,3,32,32);
- // uint32_t* labels = readLabels3(labels_path.c_str(),2000);
-
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
- // args->input = input;
- // args->input_bytes = 0;
+ void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
+ args->input = input;
+ args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
args->conv2d_1_w_bytes = 0;
args->conv2d_1_b = conv2d_1_b;
@@ -991,40 +982,21 @@ int main() {
args->dense_2_b = dense_2_b;
args->dense_2_b_bytes = 0;
- int batch_size = 500;
- int test_input_size = 5000;
- int batch_count = test_input_size / batch_size;
-
+ __hpvm__init();
startMemTracking();
- startProfiling();
-
- for (int j = 0; j < 1; j++) {
- for (int i = 0; i < batch_count; i++) {
-
- int start = i * batch_size;
- int end = (i + 1) * batch_size;
-
- void *input =
- readInputBatch(input_path.c_str(), 0, start, end, 3, 32, 32);
-
- args->input = input;
- args->input_bytes = 0;
-
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
- llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
-
- freeBatchMemory();
- }
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 32, 32, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
+ freeBatchMemory();
}
-
- stopProfiling();
__hpvm__cleanup();
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar100/vgg16_cifar100_cudnn.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar100/vgg16_cifar100_cudnn.cpp
index a3e147cb1a..9f989e3610 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar100/vgg16_cifar100_cudnn.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_cifar100/vgg16_cifar100_cudnn.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -827,10 +821,16 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+void write_accuracy(float accuracy) {
+ std::ofstream fout("final_accuracy");
+ fout << std::fixed << accuracy;
+}
- std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/vgg16_cifar100/";
+const int batch_size = 500, input_size = 5000,
+ batch_count = input_size / batch_size;
+int main() {
+ std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/vgg16_cifar100/";
std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
@@ -922,12 +922,8 @@ int main() {
std::string dense_2_b_path = dir_prefix + std::string("dense_2_b.bin");
void *dense_2_b = readTrainedWeights(dense_2_b_path.c_str(), 0, 1, 100, 1, 1);
- void *input = readTrainedWeights(input_path.c_str(), 0, 2000, 3, 32, 32);
- uint32_t *labels = readLabels3(labels_path.c_str(), 2000);
-
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 3, 32, 32);
args->input = input;
args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
@@ -991,14 +987,25 @@ int main() {
args->dense_2_b = dense_2_b;
args->dense_2_b_bytes = 0;
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
+ __hpvm__init();
+ float total_accuracy = 0;
+ startMemTracking();
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 32, 32, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ uint32_t *labels = readLabelsBatch3(labels_path.c_str(), start, end);
+ float accuracy = computeAccuracy3(labels, result);
+ total_accuracy += accuracy * batch_size;
+ freeBatchMemory();
+ }
+ write_accuracy(total_accuracy / input_size);
__hpvm__cleanup();
- computeAccuracy3(labels, result);
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_imagenet/vgg16_imagenet.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_imagenet/vgg16_imagenet.cpp
index 91af01fe8e..12f7870a15 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_imagenet/vgg16_imagenet.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_imagenet/vgg16_imagenet.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -875,10 +869,11 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+const int batch_size = 10, input_size = 5000,
+ batch_count = input_size / batch_size;
- std::string dir_prefix =
- std::string(MODEL_PARAMS_DIR) + "/vgg16_imagenet/";
+int main() {
+ std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/vgg16_imagenet/";
std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
@@ -978,9 +973,10 @@ int main() {
void *dense_3_b =
readTrainedWeights(dense_3_b_path.c_str(), 0, 1, 1000, 1, 1);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 3, 224, 224);
+ args->input = input;
+ args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
args->conv2d_1_w_bytes = 0;
args->conv2d_1_b = conv2d_1_b;
@@ -1046,40 +1042,21 @@ int main() {
args->dense_3_b = dense_3_b;
args->dense_3_b_bytes = 0;
+ __hpvm__init();
startMemTracking();
- startProfiling();
-
- unsigned int batch_size = 50;
- unsigned int test_input_size = 1000;
- unsigned int batch_count = test_input_size / batch_size;
-
- for (int j = 0; j < 1; j++) {
- for (int i = 0; i < batch_count; i++) {
-
- int start = i * batch_size;
- int end = (i + 1) * batch_size;
-
- void *input =
- readInputBatch(input_path.c_str(), 0, start, end, 3, 224, 224);
-
- args->input = input;
- args->input_bytes = 0;
-
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
- llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
-
- freeBatchMemory();
- }
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 224, 224, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
+ freeBatchMemory();
}
-
- stopProfiling();
__hpvm__cleanup();
-
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_imagenet/vgg16_imagenet_cudnn.cpp b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_imagenet/vgg16_imagenet_cudnn.cpp
index 995502f907..189460c928 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_imagenet/vgg16_imagenet_cudnn.cpp
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/benchmarks/vgg16_imagenet/vgg16_imagenet_cudnn.cpp
@@ -1,10 +1,4 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
+#include <string>
#include <hpvm.h>
#include <tensorTypes.h>
#include <tensorUtils.h>
@@ -875,8 +869,15 @@ typedef struct __attribute__((__packed__)) {
struct ret_t r;
} RootIn;
-int main() {
+void write_accuracy(float accuracy) {
+ std::ofstream fout("final_accuracy");
+ fout << std::fixed << accuracy;
+}
+const int batch_size = 25, input_size = 5000,
+ batch_count = input_size / batch_size;
+
+int main() {
std::string dir_prefix = std::string(MODEL_PARAMS_DIR) + "/vgg16_imagenet/";
std::string input_path = dir_prefix + std::string("test_input.bin");
std::string labels_path = dir_prefix + std::string("test_labels.bin");
@@ -976,12 +977,10 @@ int main() {
std::string dense_3_b_path = dir_prefix + std::string("dense_3_b.bin");
void *dense_3_b =
readTrainedWeights(dense_3_b_path.c_str(), 0, 1, 1000, 1, 1);
- void *input = readTrainedWeights(input_path.c_str(), 0, 100, 3, 224, 224);
uint32_t *labels = readLabelsBatch3(labels_path.c_str(), 0, 100);
- __hpvm__init();
RootIn *args = static_cast<RootIn *>(malloc(sizeof(RootIn)));
-
+ void *input = create4DTensor(0, nchw, batch_size, 3, 224, 224);
args->input = input;
args->input_bytes = 0;
args->conv2d_1_w = conv2d_1_w;
@@ -1049,14 +1048,25 @@ int main() {
args->dense_3_b = dense_3_b;
args->dense_3_b_bytes = 0;
- void *dfg = __hpvm__launch(0, root, (void *)args);
-
- __hpvm__wait(dfg);
-
- void *result = static_cast<RootIn *>(args)->r.tensor;
- hpvm_request_tensor(result, 0);
-
+ __hpvm__init();
+ float total_accuracy = 0;
+ startMemTracking();
+#pragma clang loop unroll(disable)
+ for (int i = 0; i < batch_count; i++) {
+ int start = i * batch_size, end = start + batch_size;
+ copyInputBatch(input_path.c_str(), start, end, 3, 224, 224, input);
+
+ void *dfg = __hpvm__launch(0, root, (void *)args);
+ __hpvm__wait(dfg);
+ void *result = static_cast<RootIn *>(args)->r.tensor;
+ hpvm_request_tensor(result, 0);
+
+ uint32_t *labels = readLabelsBatch3(labels_path.c_str(), start, end);
+ float accuracy = computeAccuracy3(labels, result);
+ total_accuracy += accuracy * batch_size;
+ freeBatchMemory();
+ }
+ write_accuracy(total_accuracy / input_size);
__hpvm__cleanup();
- computeAccuracy3(labels, result);
return 0;
}
diff --git a/hpvm/test/dnn_benchmarks/hpvm-c/include/tensorUtils.h b/hpvm/test/dnn_benchmarks/hpvm-c/include/tensorUtils.h
index 71e1c26872..05d9157a64 100644
--- a/hpvm/test/dnn_benchmarks/hpvm-c/include/tensorUtils.h
+++ b/hpvm/test/dnn_benchmarks/hpvm-c/include/tensorUtils.h
@@ -3,7 +3,6 @@
#ifndef UTILS_HEADER
#define UTILS_HEADER
-
#include <sstream>
#include <vector>
#include <bits/stdc++.h>
@@ -11,15 +10,13 @@
#include <tensor.h>
#include <cmath>
-
std::vector<float> run_accuracies;
+void printTensorInfo(void *tensor_ptr) {
-void printTensorInfo(void* tensor_ptr){
+ struct Tensor *tensor = (struct Tensor *)tensor_ptr;
- struct Tensor* tensor = (struct Tensor*) tensor_ptr;
-
- if(tensor->gpu_data != NULL){
+ if (tensor->gpu_data != NULL) {
printf("Successful cudaMalloc \n");
}
@@ -29,376 +26,354 @@ void printTensorInfo(void* tensor_ptr){
printf("num_elems = %lu \n", tensor->num_elems);
}
-
// FIXIT: Move this to debug.h and include in all files
-void dumpWeightsToFile(char* file_name, void* weights_ptr){
+void dumpWeightsToFile(char *file_name, void *weights_ptr) {
- struct Tensor* weights = (Tensor*) weights_ptr;
+ struct Tensor *weights = (Tensor *)weights_ptr;
// Move data back to host
hpvm_request_tensor(weights, 0);
-
- FILE* fp = fopen(file_name, "wb");
- if(fp == NULL){
- printf("File %s could not be created. Check if directory exists \n", file_name);
+
+ FILE *fp = fopen(file_name, "wb");
+ if (fp == NULL) {
+ printf("File %s could not be created. Check if directory exists \n",
+ file_name);
abort();
}
- //printf("size_in_bytes = %lu \n", weights->size_in_bytes);
- size_t bytes_written = fwrite(weights->host_data, 1, weights->size_in_bytes, fp);
- //printf("bytes_written = %lu \n", bytes_written);
+ // printf("size_in_bytes = %lu \n", weights->size_in_bytes);
+ size_t bytes_written =
+ fwrite(weights->host_data, 1, weights->size_in_bytes, fp);
+ // printf("bytes_written = %lu \n", bytes_written);
fclose(fp);
}
+void fillTensorWithOnes(void *tensor_ptr) {
+ struct Tensor *tensor = (struct Tensor *)tensor_ptr;
-void fillTensorWithOnes(void* tensor_ptr){
-
- 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] = 1.0;
+ 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] = 1.0;
}
}
}
+void fillWithOnesAndTwos(void *tensor_ptr) {
-void fillWithOnesAndTwos(void* tensor_ptr){
+ struct Tensor *tensor = (struct Tensor *)tensor_ptr;
- 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/2; i++){
- data_arr[i] = 1.0;
+ if (tensor->data_type == CUDNN_DATA_FLOAT) {
+ float *data_arr = (float *)tensor->host_data;
+ for (unsigned int i = 0; i < tensor->num_elems / 2; i++) {
+ data_arr[i] = 1.0;
}
- for(unsigned int i = tensor->num_elems/2; i < tensor->num_elems; i++){
- data_arr[i] = 2.0;
+ for (unsigned int i = tensor->num_elems / 2; i < tensor->num_elems; i++) {
+ data_arr[i] = 2.0;
}
-
}
}
+void fillTensorWithVal(void *tensor_ptr, float target_value) {
-void fillTensorWithVal(void* tensor_ptr, float target_value){
+ struct Tensor *tensor = (struct Tensor *)tensor_ptr;
- 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;
+ 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) {
-void fillTensorWithNegOnes(void* tensor_ptr){
+ struct Tensor *tensor = (struct Tensor *)tensor_ptr;
- 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] = -1.0;
+ 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] = -1.0;
}
}
}
+void fillTensorVals(void *tensor_ptr) {
-void fillTensorVals(void* tensor_ptr){
-
- struct Tensor* tensor = (struct Tensor*) tensor_ptr;
+ struct Tensor *tensor = (struct Tensor *)tensor_ptr;
// 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] = i + 1;
+ 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] = i + 1;
}
}
}
+void printTensorValues(void *tensor_ptr) {
-void printTensorValues(void* tensor_ptr){
-
- struct Tensor* tensor = (struct Tensor*) tensor_ptr;
+ struct Tensor *tensor = (struct Tensor *)tensor_ptr;
hpvm_request_tensor(tensor, 0);
-
+
// printing 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++){
- printf("%f,", data_arr[i]);
+ if (tensor->data_type == CUDNN_DATA_FLOAT) {
+ float *data_arr = (float *)tensor->host_data;
+ for (unsigned int i = 0; i < tensor->num_elems; i++) {
+ printf("%f,", data_arr[i]);
}
}
printf("\n");
}
+void printTensorDims(void *tensor_ptr) {
-void printTensorDims(void* tensor_ptr){
-
- struct Tensor* tensor = (struct Tensor*) tensor_ptr;
+ struct Tensor *tensor = (struct Tensor *)tensor_ptr;
printf("Num_elems = %lu \n", tensor->num_elems);
- for (int i = 0; i < tensor->dims.num_dims; i++){
+ for (int i = 0; i < tensor->dims.num_dims; i++) {
printf("dim[%d] = %lu \n", i, tensor->dims.dim_sizes[i]);
}
}
+void compareTensors(void *tensor1_ptr, void *tensor2_ptr) {
-
-void compareTensors(void* tensor1_ptr, void* tensor2_ptr){
-
- struct Tensor* tensor1 = (struct Tensor*) tensor1_ptr;
- struct Tensor* tensor2 = (struct Tensor*) tensor2_ptr;
+ struct Tensor *tensor1 = (struct Tensor *)tensor1_ptr;
+ struct Tensor *tensor2 = (struct Tensor *)tensor2_ptr;
hpvm_request_tensor(tensor1, 0);
hpvm_request_tensor(tensor2, 0);
- float* tensor_data1 = (float*) tensor1->host_data;
- float* tensor_data2 = (float*) tensor2->host_data;
-
- for(unsigned int i = 0; i < tensor1->num_elems; i++){
- if(tensor_data1[i] != tensor_data2[i]){
+ float *tensor_data1 = (float *)tensor1->host_data;
+ float *tensor_data2 = (float *)tensor2->host_data;
+
+ for (unsigned int i = 0; i < tensor1->num_elems; i++) {
+ if (tensor_data1[i] != tensor_data2[i]) {
printf("Tensor data mismatch at index %d \n", i);
abort();
}
}
}
+void compareValues(void *tensor_ptr, float *data, size_t num_elems) {
+ struct Tensor *tensor = (struct Tensor *)tensor_ptr;
-void compareValues(void* tensor_ptr, float* data, size_t num_elems){
-
- struct Tensor* tensor = (struct Tensor*) tensor_ptr;
-
hpvm_request_tensor(tensor, 0);
-
- float* tensor_data = (float*) tensor->host_data;
- for(unsigned int i = 0; i < num_elems; i++){
- if(tensor_data[i] != data[i]){
+
+ float *tensor_data = (float *)tensor->host_data;
+ for (unsigned int i = 0; i < num_elems; i++) {
+ if (tensor_data[i] != data[i]) {
printf("Tensor data mismatch");
abort();
}
}
}
-
-void* readInputTensor(const char* file_name, int data_type, int dim1_size, int dim2_size,
- int dim3_size, int dim4_size){
+void *readInputTensor(const char *file_name, int data_type, int dim1_size,
+ int dim2_size, int dim3_size, int dim4_size) {
int type_size = 4; // NOTE: Assuming floating point tensors
int num_elems = dim1_size * dim2_size * dim3_size * dim4_size;
int size_in_bytes = type_size * dim1_size * dim2_size * dim3_size * dim4_size;
- uint8_t* file_data = (uint8_t*) malloc(sizeof(char) * num_elems);
- float* tensor_data = (float*) malloc(sizeof(float) * num_elems);
+ uint8_t *file_data = (uint8_t *)malloc(sizeof(char) * num_elems);
+ float *tensor_data = (float *)malloc(sizeof(float) * num_elems);
int file_header_size = 16;
-
- FILE* file = fopen(file_name, "rb");
- if(file == NULL){
+
+ 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(file_data, 1, sizeof(uint8_t) * num_elems, file);
fclose(file);
-
- for (size_t i = 0; i < num_elems; ++i){
- tensor_data[i] = (float) file_data[i] / 255.0f;
+
+ for (size_t i = 0; i < num_elems; ++i) {
+ tensor_data[i] = (float)file_data[i] / 255.0f;
}
// NOTE: Using NCHW format
- struct Tensor* input = (struct Tensor*) create4DTensor(data_type, nchw, dim1_size, dim2_size,
- dim3_size, dim4_size);
-
+ struct Tensor *input = (struct Tensor *)create4DTensor(
+ data_type, nchw, dim1_size, dim2_size, dim3_size, dim4_size);
+
initTensorData(input, tensor_data, size_in_bytes);
// compareValues(input, tensor_data, num_elems);
-
- return input;
-}
+ return input;
+}
//*** 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){
+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);
+ 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){
+
+ 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);
-
+
+ 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);
+ // 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,
- long int dim3_size, long int dim4_size){
+struct Tensor *readTrainedWeights(const char *file_name, int data_type,
+ long int dim1_size, long int dim2_size,
+ long int dim3_size, long 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);
+ long int size_in_bytes =
+ type_size * dim1_size * dim2_size * dim3_size * dim4_size;
+ float *tensor_data = (float *)malloc(sizeof(float) * num_elems);
printf("size_in_bytes = %lu \n", size_in_bytes);
-
+
int file_header_size = 0;
-
- FILE* file = fopen(file_name, "rb");
- if(file == NULL){
+
+ 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);
+ // 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);
-
+
+ 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);
+ // compareValues(weights, tensor_data, num_elems);
free(tensor_data);
return weights;
}
-
-
-
-struct Tensor* readInputBatch(const char* file_name, int data_type,
- int start, int end,
- int dim2_size, int dim3_size, int dim4_size){
+struct Tensor *readInputBatch(const char *file_name, long data_type, long start,
+ long end, long dim2_size, long dim3_size,
+ long dim4_size) {
long int dim1_size = end - start;
// FIXIT: Don't assume floating point types
long 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);
- long int file_header_size = type_size * start * dim2_size * dim3_size * dim4_size;
-
- FILE* file = fopen(file_name, "rb");
- if(file == NULL){
+ long int size_in_bytes =
+ type_size * dim1_size * dim2_size * dim3_size * dim4_size;
+ float *tensor_data = (float *)malloc(sizeof(float) * num_elems);
+ long int file_header_size =
+ type_size * start * dim2_size * dim3_size * dim4_size;
+
+ 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_SET); // Skipping the file header
size_t bytes_read = fread(tensor_data, 1, size_in_bytes, file);
fclose(file);
- //printf ("FIXED input BATCH read \n");
-
- struct Tensor* weights = (struct Tensor*) create4DTensor(data_type, nchw, dim1_size, dim2_size,
- dim3_size, dim4_size);
-
+ // printf ("FIXED input BATCH read \n");
+
+ struct Tensor *weights = (struct Tensor *)create4DTensor(
+ data_type, nchw, dim1_size, dim2_size, dim3_size, dim4_size);
+
initTensorData(weights, tensor_data, size_in_bytes);
free(tensor_data);
return weights;
}
+void *copyInputBatch(const char *file_name, long start, long end,
+ long dim2_size, long dim3_size, long dim4_size,
+ void *inputTensor_ptr) {
+ struct Tensor *inputTensor = (struct Tensor *)inputTensor_ptr;
-void* copyInputBatch(const char* file_name,
- int start, int end,
- int dim2_size, int dim3_size, int dim4_size,
- void* inputTensor_ptr){
-
- struct Tensor* inputTensor = (struct Tensor*) inputTensor_ptr;
-
int dim1_size = end - start;
// 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 = type_size * start * dim2_size * dim3_size * dim4_size;
-
- FILE* file = fopen(file_name, "rb");
- if(file == NULL){
+ long int size_in_bytes =
+ type_size * dim1_size * dim2_size * dim3_size * dim4_size;
+ float *tensor_data = (float *)malloc(sizeof(float) * num_elems);
+ long int file_header_size =
+ type_size * start * dim2_size * dim3_size * dim4_size;
+
+ 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_SET); // Skipping the file header
size_t bytes_read = fread(tensor_data, 1, size_in_bytes, file);
fclose(file);
-
-
+
initTensorData(inputTensor, tensor_data, size_in_bytes);
free(tensor_data);
printf("******NOTE: tensor Dims = %d \n", inputTensor->dims.num_dims);
- if(inputTensor->host_data == NULL || inputTensor->gpu_data == NULL)
+ if (inputTensor->host_data == NULL || inputTensor->gpu_data == NULL)
printf("ERROR: NULL data pointers \n");
-
- // Chaning Tensor Placement to HOST
+ // Chaning Tensor Placement to HOST
changeTensorPlacement(inputTensor, HOST);
-
return inputTensor;
}
+uint8_t *readLabels(const char *labels_file, int num_labels) {
-
-uint8_t* readLabels(const char* labels_file, int num_labels){
-
- uint8_t* labels = (uint8_t*) malloc(sizeof(uint8_t) * num_labels);
- FILE* file = fopen(labels_file, "rb");
- if(file == NULL){
+ uint8_t *labels = (uint8_t *)malloc(sizeof(uint8_t) * num_labels);
+ FILE *file = fopen(labels_file, "rb");
+ if (file == NULL) {
printf("Data file %s is not found. Aborting...\n", labels_file);
abort();
}
@@ -406,17 +381,15 @@ uint8_t* readLabels(const char* labels_file, int num_labels){
size_t bytes_read = fread(labels, 1, sizeof(uint8_t) * num_labels, file);
fclose(file);
-
+
return labels;
}
+uint32_t *readLabels3(const char *labels_file, int num_labels) {
-
-uint32_t* readLabels3(const char* labels_file, int num_labels){
-
- uint32_t* labels = (uint32_t*) malloc(sizeof(uint32_t) * num_labels);
- FILE* file = fopen(labels_file, "rb");
- if(file == NULL){
+ uint32_t *labels = (uint32_t *)malloc(sizeof(uint32_t) * num_labels);
+ FILE *file = fopen(labels_file, "rb");
+ if (file == NULL) {
printf("Data file %s is not found. Aborting...\n", labels_file);
abort();
}
@@ -424,264 +397,248 @@ uint32_t* readLabels3(const char* labels_file, int num_labels){
size_t bytes_read = fread(labels, 1, sizeof(uint32_t) * num_labels, file);
fclose(file);
-
+
return labels;
}
-
-uint8_t* readLabelsBatch(const char* labels_file, int start, int end){
+uint8_t *readLabelsBatch(const char *labels_file, int start, int end) {
int num_labels = end - start;
int file_header_size = sizeof(uint8_t) * start;
-
- uint8_t* labels = (uint8_t*) malloc(sizeof(uint8_t) * num_labels);
- FILE* file = fopen(labels_file, "rb");
- if(file == NULL){
+
+ uint8_t *labels = (uint8_t *)malloc(sizeof(uint8_t) * num_labels);
+ FILE *file = fopen(labels_file, "rb");
+ if (file == NULL) {
printf("Data file %s is not found. Aborting...\n", labels_file);
abort();
}
-
+
fseek(file, file_header_size, SEEK_SET); // Skipping the file header
-
- size_t bytes_read = fread(labels, 1, sizeof(uint8_t) * num_labels, file);
+ size_t bytes_read = fread(labels, 1, sizeof(uint8_t) * num_labels, file);
fclose(file);
-
+
// printf("--labels bytes_read = %lu \n", bytes_read);
return labels;
}
-
-uint32_t* readLabelsBatch3(const char* labels_file, int start, int end){
+uint32_t *readLabelsBatch3(const char *labels_file, int start, int end) {
int num_labels = end - start;
int file_header_size = sizeof(uint32_t) * start;
-
- uint32_t* labels = (uint32_t*) malloc(sizeof(uint32_t) * num_labels);
- FILE* file = fopen(labels_file, "rb");
- if(file == NULL){
+
+ uint32_t *labels = (uint32_t *)malloc(sizeof(uint32_t) * num_labels);
+ FILE *file = fopen(labels_file, "rb");
+ if (file == NULL) {
printf("Data file %s is not found. Aborting...\n", labels_file);
abort();
}
-
+
fseek(file, file_header_size, SEEK_SET); // Skipping the file header
-
- size_t bytes_read = fread(labels, 1, sizeof(uint32_t) * num_labels, file);
+ size_t bytes_read = fread(labels, 1, sizeof(uint32_t) * num_labels, file);
fclose(file);
-
+
return labels;
}
+void computeAccuracy(const char *labels_file, int num_labels,
+ void *result_ptr) {
+ struct Tensor *result = (struct Tensor *)result_ptr;
-void computeAccuracy(const char* labels_file, int num_labels, void* result_ptr){
-
- struct Tensor* result = (struct Tensor*) result_ptr;
-
- uint8_t* labels = readLabels(labels_file, num_labels);
+ uint8_t *labels = readLabels(labels_file, num_labels);
size_t batch_dim = result->dims.dim_sizes[0];
size_t channels = result->dims.dim_sizes[1];
- float* data = (float*) result->host_data;
+ float *data = (float *)result->host_data;
int num_errors = 0;
-
- for(int i = 0; i < batch_dim; i++){
+
+ for (int i = 0; i < batch_dim; i++) {
int chosen = 0;
- for (int id = 1; id < 10; ++id){
- if (data[i * channels + chosen] < data[i * channels + id]) chosen = id;
+ for (int id = 1; id < 10; ++id) {
+ if (data[i * channels + chosen] < data[i * channels + id])
+ chosen = id;
}
-
- //printf("chosen = %d, label = %d \n", chosen, labels[i]);
- if(chosen != labels[i])
+
+ // printf("chosen = %d, label = %d \n", chosen, labels[i]);
+ if (chosen != labels[i])
num_errors++;
}
float accuracy = ((batch_dim - num_errors) * 1.0 / batch_dim * 1.0) * 100.0;
printf("****** Accuracy = %f \n\n", accuracy);
-
- FILE* fp = fopen("final_accuracy", "w+");
- if(fp != NULL){
+ FILE *fp = fopen("final_accuracy", "w+");
+ if (fp != NULL) {
std::ostringstream ss;
ss << std::fixed << accuracy;
std::string print_str = ss.str();
-
+
fwrite(print_str.c_str(), 1, print_str.length(), fp);
fclose(fp);
}
-
}
+// NOTE: batch_size and num_classes are Unused arguments
+float computeAccuracy2(uint8_t *labels, int batch_size, void *result_ptr,
+ size_t num_classes = 10) {
+ struct Tensor *result = (struct Tensor *)result_ptr;
-
-// NOTE: batch_size and num_classes are Unused arguments
-float computeAccuracy2(uint8_t* labels, int batch_size,
- void* result_ptr, size_t num_classes = 10){
-
- struct Tensor* result = (struct Tensor*) result_ptr;
-
size_t batch_dim = result->dims.dim_sizes[0];
num_classes = result->dims.dim_sizes[1];
- float* data = (float*) result->host_data;
+ float *data = (float *)result->host_data;
int num_errors = 0;
printf("batch_dim = %lu, channels = %lu \n", batch_dim, num_classes);
-
- for(unsigned int i = 0; i < batch_dim; i++){
-
+
+ for (unsigned int i = 0; i < batch_dim; i++) {
+
int chosen = 0;
- for (int id = 1; id < num_classes; ++id){
- if (data[i * num_classes + chosen] < data[i * num_classes + id]) chosen = id;
+ for (int id = 1; id < num_classes; ++id) {
+ if (data[i * num_classes + chosen] < data[i * num_classes + id])
+ chosen = id;
}
-
- if(chosen != labels[i])
- num_errors++;
+ if (chosen != labels[i])
+ num_errors++;
}
float accuracy = ((batch_dim - num_errors) * 1.0 / batch_dim * 1.0) * 100.0;
printf("****** Accuracy = %f \n\n", accuracy);
- FILE* fp = fopen("final_accuracy", "w+");
- if(fp != NULL){
+ FILE *fp = fopen("final_accuracy", "w+");
+ if (fp != NULL) {
std::ostringstream ss;
ss << std::fixed << accuracy;
std::string print_str = ss.str();
-
+
fwrite(print_str.c_str(), 1, print_str.length(), fp);
}
fclose(fp);
- return accuracy;
+ return accuracy;
}
+float computeAccuracy3(uint32_t *labels, void *result_ptr) {
+ struct Tensor *result = (struct Tensor *)result_ptr;
-float computeAccuracy3(uint32_t* labels, void* result_ptr){
-
- struct Tensor* result = (struct Tensor*) result_ptr;
-
size_t batch_dim = result->dims.dim_sizes[0];
size_t num_classes = result->dims.dim_sizes[1];
- float* data = (float*) result->host_data;
+ float *data = (float *)result->host_data;
int num_errors = 0;
printf("batch_dim = %lu, num_classes = %lu \n", batch_dim, num_classes);
-
- for(int i = 0; i < batch_dim; i++){
-
+
+ for (int i = 0; i < batch_dim; i++) {
+
int chosen = 0;
- for (int id = 1; id < num_classes; ++id){
- if (data[i * num_classes + chosen] < data[i * num_classes + id]) chosen = id;
+ for (int id = 1; id < num_classes; ++id) {
+ if (data[i * num_classes + chosen] < data[i * num_classes + id])
+ chosen = id;
}
-
- if(chosen != labels[i])
+
+ if (chosen != labels[i])
num_errors++;
}
float accuracy = ((batch_dim - num_errors) * 1.0 / batch_dim * 1.0) * 100.0;
printf("****** Accuracy = %f \n\n", accuracy);
- FILE* fp = fopen("final_accuracy", "w+");
- if(fp != NULL){
+ FILE *fp = fopen("final_accuracy", "w+");
+ if (fp != NULL) {
std::ostringstream ss;
ss << std::fixed << accuracy;
std::string print_str = ss.str();
-
+
fwrite(print_str.c_str(), 1, print_str.length(), fp);
}
fclose(fp);
- return accuracy;
+ return accuracy;
}
-
-
-struct ClassProb{
+struct ClassProb {
float prob;
int index;
};
-
-bool descendFloatComp(ClassProb obj1, ClassProb obj2){
+bool descendFloatComp(ClassProb obj1, ClassProb obj2) {
return obj1.prob > obj2.prob;
}
+float computeTop5Accuracy(uint8_t *labels, int num_labels, void *result_ptr,
+ unsigned num_classes = 10) {
+
+ struct Tensor *result = (struct Tensor *)result_ptr;
-float computeTop5Accuracy(uint8_t* labels, int num_labels,
- void* result_ptr, unsigned num_classes = 10){
-
- struct Tensor* result = (struct Tensor*) result_ptr;
-
size_t batch_dim = result->dims.dim_sizes[0];
size_t channels = result->dims.dim_sizes[1];
- float* data = (float*) result->host_data;
+ float *data = (float *)result->host_data;
int num_errors = 0;
printf("batch_dim = %lu, channels = %lu \n", batch_dim, channels);
-
- for(int i = 0; i < num_labels; i++){
+
+ for (int i = 0; i < num_labels; i++) {
std::vector<ClassProb> elem_probs;
- for (int id = 0; id < num_classes; ++id){
+ for (int id = 0; id < num_classes; ++id) {
ClassProb cProb;
cProb.prob = data[i * channels + id];
cProb.index = id;
- elem_probs.push_back(cProb);
+ elem_probs.push_back(cProb);
}
- std:sort(elem_probs.begin(), elem_probs.end(), descendFloatComp);
+ std:
+ sort(elem_probs.begin(), elem_probs.end(), descendFloatComp);
// Check if any of top-5 predictions matches
bool matched = false;
- for(int j = 0; j < 5; j++){
+ for (int j = 0; j < 5; j++) {
ClassProb cProb = elem_probs[j];
- if(cProb.index == labels[i])
+ if (cProb.index == labels[i])
matched = true;
}
- if(!matched)
- num_errors +=1;
+ if (!matched)
+ num_errors += 1;
}
float accuracy = ((batch_dim - num_errors) * 1.0 / batch_dim * 1.0) * 100.0;
printf("****** Accuracy = %f \n\n", accuracy);
- FILE* fp = fopen("final_accuracy", "w+");
- if(fp != NULL){
+ FILE *fp = fopen("final_accuracy", "w+");
+ if (fp != NULL) {
std::ostringstream ss;
ss << std::fixed << accuracy;
std::string print_str = ss.str();
-
+
fwrite(print_str.c_str(), 1, print_str.length(), fp);
}
fclose(fp);
- return accuracy;
+ return accuracy;
}
-
-
-
-void dumpFinalAccuracy(float accuracy){
+void dumpFinalAccuracy(float accuracy) {
printf("\n\n **** Final Accuracy = %f \n", accuracy);
-
- FILE* fp = fopen("final_accuracy", "w+");
- if(fp != NULL){
+
+ FILE *fp = fopen("final_accuracy", "w+");
+ if (fp != NULL) {
std::ostringstream ss;
ss << std::fixed << accuracy;
std::string print_str = ss.str();
-
+
fwrite(print_str.c_str(), 1, print_str.length(), fp);
}
@@ -690,44 +647,37 @@ void dumpFinalAccuracy(float accuracy){
run_accuracies.push_back(accuracy);
}
+void dumpAvgPSNR(float avg_psnr) {
-
-void dumpAvgPSNR(float avg_psnr){
-
- FILE* fp = fopen("avg_psnr", "w+");
- if(fp != NULL){
+ FILE *fp = fopen("avg_psnr", "w+");
+ if (fp != NULL) {
std::ostringstream ss;
ss << std::fixed << avg_psnr;
- std::string print_str = ss.str();
+ std::string print_str = ss.str();
fwrite(print_str.c_str(), 1, print_str.length(), fp);
}
fclose(fp);
}
+void dumpPSNRStd(float psnr_std) {
-void dumpPSNRStd(float psnr_std){
-
- FILE* fp = fopen("psnr_std.txt", "w+");
- if(fp != NULL){
+ FILE *fp = fopen("psnr_std.txt", "w+");
+ if (fp != NULL) {
std::ostringstream ss;
ss << std::fixed << psnr_std;
- std::string print_str = ss.str();
+ std::string print_str = ss.str();
fwrite(print_str.c_str(), 1, print_str.length(), fp);
}
fclose(fp);
}
+void dumpExecutionAccuracies() {
-
-
-
-void dumpExecutionAccuracies(){
-
- FILE* fp = fopen("run_accuracies.txt", "w+");
- if(fp != NULL){
- for (int i = 0; i < run_accuracies.size(); i++){
+ FILE *fp = fopen("run_accuracies.txt", "w+");
+ if (fp != NULL) {
+ for (int i = 0; i < run_accuracies.size(); i++) {
float accuracy = run_accuracies[i];
std::ostringstream ss;
ss << std::fixed << accuracy;
@@ -735,63 +685,60 @@ void dumpExecutionAccuracies(){
fwrite(print_str.c_str(), 1, print_str.length(), fp);
fwrite("\n", 1, 1, fp);
}
-
}
fclose(fp);
}
-
-float readPSNRFromFile(const char* file_name){
+float readPSNRFromFile(const char *file_name) {
float psnr;
- FILE* pFile = fopen(file_name, "r");
- if(pFile == NULL){
+ 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;
+ return psnr;
}
+float computePSNRViolation(void *gold_ptr, void *approx_ptr,
+ float PSNR_threshold) {
-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;
- size_t* dim_sizes = gold_tensor->dims.dim_sizes;
+ struct Tensor *gold_tensor = (struct Tensor *)gold_ptr;
+ struct Tensor *approx_tensor = (struct Tensor *)approx_ptr;
+
+ size_t *dim_sizes = gold_tensor->dims.dim_sizes;
size_t batch_dim = dim_sizes[0];
size_t image_size = dim_sizes[1] * dim_sizes[2] * dim_sizes[3];
-
+
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 *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++){
+ int num_errors = 0;
+ for (size_t i = 0; i < batch_dim; i++) {
float mse_sum = 0.0;
- float max_val = -999999;
+ float max_val = -999999;
size_t offset = i * image_size;
-
- for(size_t j = 0; j < image_size; j++){
+
+ for (size_t j = 0; j < image_size; j++) {
float diff = gold_data[offset + j] - approx_data[offset + j];
float diff_square = diff * diff;
mse_sum += diff_square;
- if(max_val < gold_data[offset + j]){
- max_val = gold_data[offset + j];
- }
+ if (max_val < gold_data[offset + j]) {
+ max_val = gold_data[offset + j];
+ }
}
mse_sum = mse_sum / image_size;
@@ -799,7 +746,7 @@ float computePSNRViolation(void* gold_ptr, void* approx_ptr, float PSNR_threshol
sum_psnr += psnr;
if (psnr < PSNR_threshold)
- num_errors += 1;
+ num_errors += 1;
printf("PSNR value = %f \n", psnr);
psnr_list.push_back(psnr);
@@ -817,39 +764,35 @@ float computePSNRViolation(void* gold_ptr, void* approx_ptr, float PSNR_threshol
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);
+ 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;
-}
+ return violation_rate;
+}
-void dumpOutput(void* output_ptr, const char* file_name){
+void dumpOutput(void *output_ptr, const char *file_name) {
- struct Tensor* out_tensor = (struct Tensor*) output_ptr;
+ struct Tensor *out_tensor = (struct Tensor *)output_ptr;
size_t size_in_bytes = out_tensor->size_in_bytes;
- printf ("** Output size = %lu \n", size_in_bytes);
-
- float* host_data = (float*) out_tensor->host_data;
- FILE* fd = fopen(file_name, "w+");
+ printf("** Output size = %lu \n", size_in_bytes);
+
+ float *host_data = (float *)out_tensor->host_data;
+ FILE *fd = fopen(file_name, "w+");
fwrite(host_data, 1, size_in_bytes, fd);
fclose(fd);
}
-
-
#endif
--
GitLab