diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/Makefile b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/Makefile
index 76f4b4791780d468f206aba64a0e66f4d4fc6f2a..86e93cb809336f51b18eaeb110b18dd666367e56 100644
--- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/Makefile
+++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/Makefile
@@ -1,6 +1,6 @@
DNN_BENCHMARK_ROOT = $(LLVM_SRC_ROOT)/test/VISC/DNN_Benchmarks
# NOTE: can configure build directory
-HPVM_BUILD_DIR = $(LLVM_SRC_ROOT)/../build_new/
+HPVM_BUILD_DIR = $(LLVM_SRC_ROOT)/../build_fresh/
CC = $(HPVM_BUILD_DIR)/bin/clang++
OPT = $(HPVM_BUILD_DIR)/bin/opt
@@ -23,7 +23,7 @@ CCFLAGS += -DDEVICE=CUDNN_TARGET
LINKER_FLAGS = -lpthread -lcudart -lcurand -lcudnn -lcublas -lOpenCL
HPVM_LIB_DIR = $(HPVM_BUILD_DIR)/lib
-HPVM_LIB_DIR = /home/hsharif3/Gitlab/hpvm/build_new/lib
+#HPVM_LIB_DIR = /home/hsharif3/Gitlab/hpvm/build_new/lib
VISC_OPTFLAGS = -load $(HPVM_LIB_DIR)/LLVMBuildDFG.so -load $(HPVM_LIB_DIR)/LLVMInPlaceDFGAnalysis.so -load $(HPVM_LIB_DIR)/LLVMDFG2LLVM_CUDNN.so -load $(HPVM_LIB_DIR)/LLVMDFG2LLVM_X86.so -load $(HPVM_LIB_DIR)/LLVMClearDFG.so -inplace -dfg2llvm-cudnn -dfg2llvm-x86 -clearDFG
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/src/alexnet.cpp b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/src/alexnet.cpp
index a8cab99d52acae5b88816da8d51539331a5590f2..ef6fda1d3194a97f4ec48f7ef8040744e24c7fa1 100644
--- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/src/alexnet.cpp
+++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/src/alexnet.cpp
@@ -1,538 +1,446 @@
-#include <iostream>
-#include <cstdio>
-#include <cstring>
-#include <cinttypes>
-#include <visc.h>
-#include <tensorTypes.h>
-#include <tensorUtils.h>
-
-using namespace std;
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <sys/stat.h>
+#include <cstring>
+#include <visc.h>
+#include <tensorTypes.h>
+#include <tensorUtils.h>
+
+void var_0_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
+
+ void *r = __visc__tensor_convolution(t1, t2, 5, 5, 1, 1);
+ __visc__return(2, r, (size_t) 0);
+}
+void var_1_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
+ void *r = __visc__tensor_add(t1, t2);
+ __visc__return(2, r, (size_t) 0);
+}
-/* DNN Layer 1 **/
-void tensorConvNode1(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+void var_2_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
- void *r = __visc__tensor_convolution(t1, t2, 5, 5, 1, 1);
- __visc__return(2, r, (size_t) 0);
+ void* r = __visc__tensor_tanh(t1);
+ __visc__return(2, r, (size_t) 0);
}
+void var_3_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
-void tensorAddNode1(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
+ void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
+ __visc__return(2, r, (size_t) 0);
}
+void var_4_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
-void tensorPoolNode1(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_convolution(t1, t2, 2, 2, 1, 1);
+ __visc__return(2, r, (size_t) 0);
}
-void tensorTanhNode1(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+void var_5_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_add(t1, t2);
+ __visc__return(2, r, (size_t) 0);
}
-/** End of Layer 1 **/
+void var_6_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
+ void* r = __visc__tensor_tanh(t1);
+ __visc__return(2, r, (size_t) 0);
+}
+void var_7_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
-/* DNN Layer 2 **/
-void tensorConvNode2(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+ void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
+ __visc__return(2, r, (size_t) 0);
+}
+
+void var_8_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- // X * W = t2 * t1
- void *r = __visc__tensor_convolution(t1, t2, 2, 2, 1, 1);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
+ __visc__return(2, r, (size_t) 0);
}
+void var_9_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
-void tensorAddNode2(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+ void *r = __visc__tensor_add(t1, t2);
+ __visc__return(2, r, (size_t) 0);
+}
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
+void var_10_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
+
+ void* r = __visc__tensor_tanh(t1);
+ __visc__return(2, r, (size_t) 0);
}
-void tensorPoolNode2(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+void var_11_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
+ __visc__return(2, r, (size_t) 0);
}
-void tensorTanhNode2(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+void var_12_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_add(t1, t2);
+ __visc__return(2, r, (size_t) 0);
}
-/** End of Layer 2 **/
+void var_13_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
+ void* r = __visc__tensor_tanh(t1);
+ __visc__return(2, r, (size_t) 0);
+}
-/* DNN Layer 3 **/
-void tensorConvNode3(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+void var_14_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- // X * W = t2 * t1
- void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
+ __visc__return(2, r, (size_t) 0);
}
-// FIXIT: Think if we can reuse the add, tanh, and pool nodes?
-void tensorAddNode3(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+void var_15_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_add(t1, t2);
+ __visc__return(2, r, (size_t) 0);
}
-void tensorTanhNode3(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+void var_16_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
+ void* r = __visc__tensor_tanh(t1);
+ __visc__return(2, r, (size_t) 0);
}
-/** End of Layer 3 **/
+void var_17_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
+ void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
+ __visc__return(2, r, (size_t) 0);
+}
-/* DNN Layer 4 **/
-void tensorConvNode4(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+void var_18_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- // X * W = t2 * t1
- void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_mul(t1, t2);
+ __visc__return(2, r, (size_t) 0);
}
-// FIXIT: Think if we can reuse the add, tanh, and pool nodes?
-void tensorAddNode4(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+void var_19_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_add(t1, t2);
+ __visc__return(2, r, (size_t) 0);
}
-void tensorTanhNode4(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+void var_20_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
+ void* r = __visc__tensor_softmax(t1);
+ __visc__return(2, r, (size_t) 0);
}
-/** End of Layer 4 **/
+void root(void* input, size_t input_bytes,
+ void* conv2d_1_w, size_t conv2d_1_w_bytes,
+ void* conv2d_1_b, size_t conv2d_1_b_bytes,
+ void* conv2d_2_w, size_t conv2d_2_w_bytes,
+ void* conv2d_2_b, size_t conv2d_2_b_bytes,
+ void* conv2d_3_w, size_t conv2d_3_w_bytes,
+ void* conv2d_3_b, size_t conv2d_3_b_bytes,
+ void* conv2d_4_w, size_t conv2d_4_w_bytes,
+ void* conv2d_4_b, size_t conv2d_4_b_bytes,
+ void* conv2d_5_w, size_t conv2d_5_w_bytes,
+ void* conv2d_5_b, size_t conv2d_5_b_bytes,
+ void* dense_1_w, size_t dense_1_w_bytes,
+ void* dense_1_b, size_t dense_1_b_bytes){
+ __visc__hint(visc::CPU_TARGET);
+ __visc__attributes(13, input, conv2d_1_w, conv2d_1_b, conv2d_2_w, conv2d_2_b, conv2d_3_w, conv2d_3_b, conv2d_4_w, conv2d_4_b, conv2d_5_w, conv2d_5_b, dense_1_w, dense_1_b, 0);
-/*** DNN Layer 5 ****/
-void tensorConvNode5(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
- void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
- __visc__return(2, r, (size_t) 0);
-}
+ void* var_0 = __visc__createNodeND(0, var_0_node);
+ __visc__bindIn(var_0, 0, 0, 0);
+ __visc__bindIn(var_0, 1, 1, 0);
+ __visc__bindIn(var_0, 2, 2, 0);
+ __visc__bindIn(var_0, 3, 3, 0);
-void tensorAddNode5(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+ void* var_1 = __visc__createNodeND(0, var_1_node);
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
+ __visc__edge(var_0, var_1, 1, 0, 0, 0);
+ __visc__edge(var_0, var_1, 1, 1, 1, 0);
+ __visc__bindIn(var_1, 4, 2, 0);
+ __visc__bindIn(var_1, 5, 3, 0);
-void tensorPoolNode5(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+ void* var_2 = __visc__createNodeND(0, var_2_node);
- void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
- __visc__return(2, r, (size_t) 0);
-}
+ __visc__edge(var_1, var_2, 1, 0, 0, 0);
+ __visc__edge(var_1, var_2, 1, 1, 1, 0);
-void tensorTanhNode5(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+ void* var_3 = __visc__createNodeND(0, var_3_node);
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
-}
+ __visc__edge(var_2, var_3, 1, 0, 0, 0);
+ __visc__edge(var_2, var_3, 1, 1, 1, 0);
-/** End of Layer 5 **/
+ void* var_4 = __visc__createNodeND(0, var_4_node);
+ __visc__edge(var_3, var_4, 1, 0, 0, 0);
+ __visc__edge(var_3, var_4, 1, 1, 1, 0);
+ __visc__bindIn(var_4, 6, 2, 0);
+ __visc__bindIn(var_4, 7, 3, 0);
-/***** DNN Layer 6 - FC layer ****/
-void tensorMulNode6(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+ void* var_5 = __visc__createNodeND(0, var_5_node);
- // X * W = t2 * t1
- void *r = __visc__tensor_mul(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
+ __visc__edge(var_4, var_5, 1, 0, 0, 0);
+ __visc__edge(var_4, var_5, 1, 1, 1, 0);
+ __visc__bindIn(var_5, 8, 2, 0);
+ __visc__bindIn(var_5, 9, 3, 0);
-void tensorAddNode6(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+ void* var_6 = __visc__createNodeND(0, var_6_node);
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
+ __visc__edge(var_5, var_6, 1, 0, 0, 0);
+ __visc__edge(var_5, var_6, 1, 1, 1, 0);
+ void* var_7 = __visc__createNodeND(0, var_7_node);
-/** End of Layer 6 **/
-
-
-
-
-void root(void *x, size_t x_bytes,
- void *conv1_w, size_t conv1_w_bytes,
- void *conv1_b, size_t conv1_b_bytes,
- void *conv2_w, size_t conv2_w_bytes,
- void *conv2_b, size_t conv2_b_bytes,
- void *conv3_w, size_t conv3_w_bytes,
- void *conv3_b, size_t conv3_b_bytes,
- void *conv4_w, size_t conv4_w_bytes,
- void *conv4_b, size_t conv4_b_bytes,
- void *conv5_w, size_t conv5_w_bytes,
- void *conv5_b, size_t conv5_b_bytes,
- void *fc1_w, size_t fc1_w_bytes,
- void *fc1_b, size_t fc1_b_bytes){
-
- __visc__hint(visc::CPU_TARGET);
- __visc__attributes(13, x, conv1_w, conv1_b, conv2_w, conv2_b, conv3_w, conv3_b,
- conv4_w, conv4_b, conv5_w, conv5_b, fc1_w, fc1_b, 0);
-
- // Conv1 Nodes
- void *nodeConv1 = __visc__createNodeND(0, tensorConvNode1);
- void *nodeAdd1 = __visc__createNodeND(0, tensorAddNode1);
- void *nodeTanh1 = __visc__createNodeND(0, tensorTanhNode1);
- void *nodePool1 = __visc__createNodeND(0, tensorPoolNode1);
- // Conv2 Nodes
- void *nodeConv2 = __visc__createNodeND(0, tensorConvNode2);
- void *nodeAdd2 = __visc__createNodeND(0, tensorAddNode2);
- void *nodeTanh2 = __visc__createNodeND(0, tensorTanhNode2);
- void *nodePool2 = __visc__createNodeND(0, tensorPoolNode2);
- // Conv3 Nodes
- void *nodeConv3 = __visc__createNodeND(0, tensorConvNode3);
- void *nodeAdd3 = __visc__createNodeND(0, tensorAddNode3);
- void *nodeTanh3 = __visc__createNodeND(0, tensorTanhNode3);
- // Conv4 Nodes
- void *nodeConv4 = __visc__createNodeND(0, tensorConvNode4);
- void *nodeAdd4 = __visc__createNodeND(0, tensorAddNode4);
- void *nodeTanh4 = __visc__createNodeND(0, tensorTanhNode4);
- // Conv5 Nodes
- void *nodeConv5 = __visc__createNodeND(0, tensorConvNode5);
- void *nodeAdd5 = __visc__createNodeND(0, tensorAddNode5);
- void *nodeTanh5 = __visc__createNodeND(0, tensorTanhNode5);
- void *nodePool5 = __visc__createNodeND(0, tensorPoolNode5);
- // FC1 Nodes
- void *nodeMul6 = __visc__createNodeND(0, tensorMulNode6);
- void *nodeAdd6 = __visc__createNodeND(0, tensorAddNode6);
-
- //***** Conv Layer 1 *******/
- // node, src, dst, stream
- __visc__bindIn(nodeConv1, 0, 0, 0);
- __visc__bindIn(nodeConv1, 1, 1, 0);
- __visc__bindIn(nodeConv1, 2, 2, 0);
- __visc__bindIn(nodeConv1, 3, 3, 0);
-
- // node, node, type, src, dst, stream
- __visc__edge(nodeConv1, nodeAdd1, 1, 0, 0, 0);
- __visc__edge(nodeConv1, nodeAdd1, 1, 1, 1, 0);
-
- // parent_index, dest_index, bind_type
- __visc__bindIn(nodeAdd1, 4, 2, 0);
- __visc__bindIn(nodeAdd1, 5, 3, 0);
-
- // node, node, type, src, dst, stream
- __visc__edge(nodeAdd1, nodeTanh1, 1, 0, 0, 0);
- __visc__edge(nodeAdd1, nodeTanh1, 1, 1, 1, 0);
-
- // node, node, type, src, dst, stream
- __visc__edge(nodeTanh1, nodePool1, 1, 0, 0, 0);
- __visc__edge(nodeTanh1, nodePool1, 1, 1, 1, 0);
-
-
- /**** Conv Layer 2 ****/
- // ConvOp2
- __visc__bindIn(nodeConv2, 6, 2, 0);
- __visc__bindIn(nodeConv2, 7, 3, 0);
-
- __visc__edge(nodePool1, nodeConv2, 1, 0, 0, 0);
- __visc__edge(nodePool1, nodeConv2, 1, 1, 1, 0);
-
- // AddOp2
- __visc__bindIn(nodeAdd2, 8, 2, 0);
- __visc__bindIn(nodeAdd2, 9, 3, 0);
-
- __visc__edge(nodeConv2, nodeAdd2, 1, 0, 0, 0);
- __visc__edge(nodeConv2, nodeAdd2, 1, 1, 1, 0);
-
- // TanhOp2
- __visc__edge(nodeAdd2, nodeTanh2, 1, 0, 0, 0);
- __visc__edge(nodeAdd2, nodeTanh2, 1, 1, 1, 0);
-
- // PoolOp2
- __visc__edge(nodeTanh2, nodePool2, 1, 0, 0, 0);
- __visc__edge(nodeTanh2, nodePool2, 1, 1, 1, 0);
-
-
- /**** Conv Layer 3 ****/
- // ConvOp3
- __visc__bindIn(nodeConv3, 10, 2, 0);
- __visc__bindIn(nodeConv3, 11, 3, 0);
-
- __visc__edge(nodePool2, nodeConv3, 1, 0, 0, 0);
- __visc__edge(nodePool2, nodeConv3, 1, 1, 1, 0);
-
- // AddOp3
- __visc__bindIn(nodeAdd3, 12, 2, 0);
- __visc__bindIn(nodeAdd3, 13, 3, 0);
-
- __visc__edge(nodeConv3, nodeAdd3, 1, 0, 0, 0);
- __visc__edge(nodeConv3, nodeAdd3, 1, 1, 1, 0);
-
- // TanhOp3
- __visc__edge(nodeAdd3, nodeTanh3, 1, 0, 0, 0);
- __visc__edge(nodeAdd3, nodeTanh3, 1, 1, 1, 0);
-
-
- /**** Conv Layer 4 ****/
- // ConvOp4
- __visc__bindIn(nodeConv4, 14, 2, 0);
- __visc__bindIn(nodeConv4, 15, 3, 0);
-
- __visc__edge(nodeTanh3, nodeConv4, 1, 0, 0, 0);
- __visc__edge(nodeTanh3, nodeConv4, 1, 1, 1, 0);
-
- // AddOp4
- __visc__bindIn(nodeAdd4, 16, 2, 0);
- __visc__bindIn(nodeAdd4, 17, 3, 0);
-
- __visc__edge(nodeConv4, nodeAdd4, 1, 0, 0, 0);
- __visc__edge(nodeConv4, nodeAdd4, 1, 1, 1, 0);
-
- // TanhOp4
- __visc__edge(nodeAdd4, nodeTanh4, 1, 0, 0, 0);
- __visc__edge(nodeAdd4, nodeTanh4, 1, 1, 1, 0);
-
-
- /**** Conv Layer 5 ****/
- // ConvOp5
- __visc__bindIn(nodeConv5, 18, 2, 0);
- __visc__bindIn(nodeConv5, 19, 3, 0);
-
- __visc__edge(nodeTanh4, nodeConv5, 1, 0, 0, 0);
- __visc__edge(nodeTanh4, nodeConv5, 1, 1, 1, 0);
-
- // AddOp5
- __visc__bindIn(nodeAdd5, 20, 2, 0);
- __visc__bindIn(nodeAdd5, 21, 3, 0);
-
- __visc__edge(nodeConv5, nodeAdd5, 1, 0, 0, 0);
- __visc__edge(nodeConv5, nodeAdd5, 1, 1, 1, 0);
+ __visc__edge(var_6, var_7, 1, 0, 0, 0);
+ __visc__edge(var_6, var_7, 1, 1, 1, 0);
- // TanhOp5
- __visc__edge(nodeAdd5, nodeTanh5, 1, 0, 0, 0);
- __visc__edge(nodeAdd5, nodeTanh5, 1, 1, 1, 0);
+ void* var_8 = __visc__createNodeND(0, var_8_node);
- // PoolOp5
- __visc__edge(nodeTanh5, nodePool5, 1, 0, 0, 0);
- __visc__edge(nodeTanh5, nodePool5, 1, 1, 1, 0);
+ __visc__edge(var_7, var_8, 1, 0, 0, 0);
+ __visc__edge(var_7, var_8, 1, 1, 1, 0);
+ __visc__bindIn(var_8, 10, 2, 0);
+ __visc__bindIn(var_8, 11, 3, 0);
-
+ void* var_9 = __visc__createNodeND(0, var_9_node);
- /**** FC Layer 1 ****/
- // MulOp6
- __visc__bindIn(nodeMul6, 22, 2, 0);
- __visc__bindIn(nodeMul6, 23, 3, 0);
-
- __visc__edge(nodePool5, nodeMul6, 1, 0, 0, 0);
- __visc__edge(nodePool5, nodeMul6, 1, 1, 1, 0);
+ __visc__edge(var_8, var_9, 1, 0, 0, 0);
+ __visc__edge(var_8, var_9, 1, 1, 1, 0);
+ __visc__bindIn(var_9, 12, 2, 0);
+ __visc__bindIn(var_9, 13, 3, 0);
- // AddOp6
- __visc__bindIn(nodeAdd6, 24, 2, 0);
- __visc__bindIn(nodeAdd6, 25, 3, 0);
-
- __visc__edge(nodeMul6, nodeAdd6, 1, 0, 0, 0);
- __visc__edge(nodeMul6, nodeAdd6, 1, 1, 1, 0);
+ void* var_10 = __visc__createNodeND(0, var_10_node);
- /***** Output Binding ****/
- __visc__bindOut(nodeAdd6, 0, 0, 0);
- __visc__bindOut(nodeAdd6, 1, 1, 0);
+ __visc__edge(var_9, var_10, 1, 0, 0, 0);
+ __visc__edge(var_9, var_10, 1, 1, 1, 0);
-}
+ void* var_11 = __visc__createNodeND(0, var_11_node);
+
+ __visc__edge(var_10, var_11, 1, 0, 0, 0);
+ __visc__edge(var_10, var_11, 1, 1, 1, 0);
+ __visc__bindIn(var_11, 14, 2, 0);
+ __visc__bindIn(var_11, 15, 3, 0);
+
+ void* var_12 = __visc__createNodeND(0, var_12_node);
+
+ __visc__edge(var_11, var_12, 1, 0, 0, 0);
+ __visc__edge(var_11, var_12, 1, 1, 1, 0);
+ __visc__bindIn(var_12, 16, 2, 0);
+ __visc__bindIn(var_12, 17, 3, 0);
+ void* var_13 = __visc__createNodeND(0, var_13_node);
+
+ __visc__edge(var_12, var_13, 1, 0, 0, 0);
+ __visc__edge(var_12, var_13, 1, 1, 1, 0);
+
+ void* var_14 = __visc__createNodeND(0, var_14_node);
+
+ __visc__edge(var_13, var_14, 1, 0, 0, 0);
+ __visc__edge(var_13, var_14, 1, 1, 1, 0);
+ __visc__bindIn(var_14, 18, 2, 0);
+ __visc__bindIn(var_14, 19, 3, 0);
+
+ void* var_15 = __visc__createNodeND(0, var_15_node);
+
+ __visc__edge(var_14, var_15, 1, 0, 0, 0);
+ __visc__edge(var_14, var_15, 1, 1, 1, 0);
+ __visc__bindIn(var_15, 20, 2, 0);
+ __visc__bindIn(var_15, 21, 3, 0);
+
+ void* var_16 = __visc__createNodeND(0, var_16_node);
+
+ __visc__edge(var_15, var_16, 1, 0, 0, 0);
+ __visc__edge(var_15, var_16, 1, 1, 1, 0);
+
+ void* var_17 = __visc__createNodeND(0, var_17_node);
+
+ __visc__edge(var_16, var_17, 1, 0, 0, 0);
+ __visc__edge(var_16, var_17, 1, 1, 1, 0);
+
+ void* var_18 = __visc__createNodeND(0, var_18_node);
+
+ __visc__edge(var_17, var_18, 1, 0, 0, 0);
+ __visc__edge(var_17, var_18, 1, 1, 1, 0);
+ __visc__bindIn(var_18, 22, 2, 0);
+ __visc__bindIn(var_18, 23, 3, 0);
+
+ void* var_19 = __visc__createNodeND(0, var_19_node);
+
+ __visc__edge(var_18, var_19, 1, 0, 0, 0);
+ __visc__edge(var_18, var_19, 1, 1, 1, 0);
+ __visc__bindIn(var_19, 24, 2, 0);
+ __visc__bindIn(var_19, 25, 3, 0);
+
+ void* var_20 = __visc__createNodeND(0, var_20_node);
+
+ __visc__edge(var_19, var_20, 1, 0, 0, 0);
+ __visc__edge(var_19, var_20, 1, 1, 1, 0);
+
+ __visc__bindOut(var_20, 0, 0, 0);
+ __visc__bindOut(var_20, 1, 1, 0);
+
+}
-// Return type for the nodes
struct ret_t {
- void *tensor;
- size_t bytes;
-};
+ void* tensor;
+ size_t bytes;
+};
typedef struct __attribute__((__packed__)) {
- void *x;
- size_t x_bytes;
- // 1st Layer parameters
- void* conv1_w;
- size_t conv1_w_bytes;
- void* conv1_b;
- size_t conv1_b_bytes;
- // 2nd Layer parameters
- void* conv2_w;
- size_t conv2_w_bytes;
- void* conv2_b;
- size_t conv2_b_bytes;
- // 3rd Layer parameters
- void* conv3_w;
- size_t conv3_w_bytes;
- void* conv3_b;
- size_t conv3_b_bytes;
- // 4th Layer parameters
- void* conv4_w;
- size_t conv4_w_bytes;
- void* conv4_b;
- size_t conv4_b_bytes;
- // 5th Layer parameters
- void* conv5_w;
- size_t conv5_w_bytes;
- void* conv5_b;
- size_t conv5_b_bytes;
- // 6th Layer parameters
- void* fc1_w;
- size_t fc1_w_bytes;
- void* fc1_b;
- size_t fc1_b_bytes;
-
- struct ret_t r;
+ void* input;
+ size_t input_bytes;
+ void* conv2d_1_w;
+ size_t conv2d_1_w_bytes;
+ void* conv2d_1_b;
+ size_t conv2d_1_b_bytes;
+ void* conv2d_2_w;
+ size_t conv2d_2_w_bytes;
+ void* conv2d_2_b;
+ size_t conv2d_2_b_bytes;
+ void* conv2d_3_w;
+ size_t conv2d_3_w_bytes;
+ void* conv2d_3_b;
+ size_t conv2d_3_b_bytes;
+ void* conv2d_4_w;
+ size_t conv2d_4_w_bytes;
+ void* conv2d_4_b;
+ size_t conv2d_4_b_bytes;
+ void* conv2d_5_w;
+ size_t conv2d_5_w_bytes;
+ void* conv2d_5_b;
+ size_t conv2d_5_b_bytes;
+ void* dense_1_w;
+ size_t dense_1_w_bytes;
+ void* dense_1_b;
+ size_t dense_1_b_bytes;
+
+ struct ret_t r;
}
RootIn;
-int main() {
-
- int test_batch_size = 5000;
- std::string prefix = "../../../../../../projects/hpvm-tensor-rt/model_params";
-
- std::string input_data_path = prefix + std::string("/alexnet_cifar10/norm_cifar_input.bin");
- std::string labels_path = prefix + std::string("/alexnet_cifar10/test_labels.bin");
-
- std::string conv1_w_path = prefix + std::string("/alexnet_cifar10/conv1.bin");
- std::string conv1_b_path = prefix + std::string("/alexnet_cifar10/conv1_bias.bin");
-
- std::string conv2_w_path = prefix + std::string("/alexnet_cifar10/conv2.bin");
- std::string conv2_b_path = prefix + std::string("/alexnet_cifar10/conv2_bias.bin");
-
- std::string conv3_w_path = prefix + std::string("/alexnet_cifar10/conv3.bin");
- std::string conv3_b_path = prefix + std::string("/alexnet_cifar10/conv3_bias.bin");
-
- std::string conv4_w_path = prefix + std::string("/alexnet_cifar10/conv4.bin");
- std::string conv4_b_path = prefix + std::string("/alexnet_cifar10/conv4_bias.bin");
-
- std::string conv5_w_path = prefix + std::string("/alexnet_cifar10/conv5.bin");
- std::string conv5_b_path = prefix + std::string("/alexnet_cifar10/conv5_bias.bin");
-
- std::string fc1_w_path = prefix + std::string("/alexnet_cifar10/fc1.bin");
- std::string fc1_b_path = prefix + std::string("/alexnet_cifar10/fc1_bias.bin");
-
- printf("Reading Input Data from = %s \n", input_data_path.c_str());
-
- uint8_t* labels = readLabels(labels_path.c_str(), test_batch_size);
- void* x = readTrainedWeights(input_data_path.c_str(), float_type,
- test_batch_size, 3, 32, 32);
-
- // Loading convolution weights
- void* conv1_w = readTrainedWeights(conv1_w_path.c_str(), float_type, 64, 3, 11, 11);
- void* conv1_b = readTrainedWeights(conv1_b_path.c_str(), float_type, 1, 64, 1, 1);
- void* conv2_w = readTrainedWeights(conv2_w_path.c_str(), float_type, 192, 64, 5, 5);
- void* conv2_b = readTrainedWeights(conv2_b_path.c_str(), float_type, 1, 192, 1, 1);
- void* conv3_w = readTrainedWeights(conv3_w_path.c_str(), float_type, 384, 192, 3, 3);
- void* conv3_b = readTrainedWeights(conv3_b_path.c_str(), float_type, 1, 384, 1, 1);
- void* conv4_w = readTrainedWeights(conv4_w_path.c_str(), float_type, 256, 384, 3, 3);
- void* conv4_b = readTrainedWeights(conv4_b_path.c_str(), float_type, 1, 256, 1, 1);
- void* conv5_w = readTrainedWeights(conv5_w_path.c_str(), float_type, 256, 256, 3, 3);
- void* conv5_b = readTrainedWeights(conv5_b_path.c_str(), float_type, 1, 256, 1, 1);
-
- // Loading FC layer weights
- void* fc1_w = readTrainedWeights(fc1_w_path.c_str(), float_type, 1, 1, 4096, 10);
- void* fc1_b = readTrainedWeights(fc1_b_path.c_str(), float_type, 1, 10, 1, 1);
-
- __visc__init();
-
- RootIn* args = static_cast<RootIn*>(malloc(sizeof(RootIn)));
- args->x = x;
- args->x_bytes = 0;
- // Conv Layers params
- args->conv1_w = conv1_w;
- args->conv1_w_bytes = 0;
- args->conv1_b = conv1_b;
- args->conv1_b_bytes = 0;
-
- args->conv2_w = conv2_w;
- args->conv2_w_bytes = 0;
- args->conv2_b = conv2_b;
- args->conv2_b_bytes = 0;
-
- args->conv3_w = conv3_w;
- args->conv3_w_bytes = 0;
- args->conv3_b = conv3_b;
- args->conv3_b_bytes = 0;
-
- args->conv4_w = conv4_w;
- args->conv4_w_bytes = 0;
- args->conv4_b = conv4_b;
- args->conv4_b_bytes = 0;
-
- args->conv5_w = conv5_w;
- args->conv5_w_bytes = 0;
- args->conv5_b = conv5_b;
- args->conv5_b_bytes = 0;
-
- // FC Layers params
- args->fc1_w = fc1_w;
- args->fc1_w_bytes = 0;
- args->fc1_b = fc1_b;
- args->fc1_b_bytes = 0;
-
- void *dfg = __visc__launch(0, root, (void *)args);
-
- __visc__wait(dfg);
-
- // FIXME: Value returned in the wrong index!!
- //void *r = static_cast<RootIn*>(args)->r.tensor;
- void *result = static_cast<RootIn*>(args)->x;
- hpvm_request_tensor(result, 0);
-
- __visc__cleanup();
-
- computeAccuracy2(labels, test_batch_size, result);
-
- return 0;
-}
-
-
+int main(){
+
+ std::string dir_prefix = std::string("../../../../../../projects/hpvm-tensor-rt/model_params/alexnet_cifar10_front/");
+ //std::string dir_prefix = std::string("../../../../../../projects/hpvm-tensor-rt/model_params/alexnet_cifar10/");
+
+ std::string input_path = dir_prefix + std::string("input.bin");
+ void* input = readTrainedWeights(input_path.c_str(), 0,10000,3,32,32);
+ std::string labels_path = dir_prefix + std::string("labels.bin");
+ uint8_t* labels = readLabels(labels_path.c_str(),10000);
+ std::string conv2d_1_w_path = dir_prefix + std::string("conv0.bin");
+ void* conv2d_1_w = readTrainedWeights(conv2d_1_w_path.c_str(), 0,64,3,11,11);
+ std::string conv2d_1_b_path = dir_prefix + std::string("conv_bias0.bin");
+ void* conv2d_1_b = readTrainedWeights(conv2d_1_b_path.c_str(), 0,1,64,1,1);
+ std::string conv2d_2_w_path = dir_prefix + std::string("conv3.bin");
+ void* conv2d_2_w = readTrainedWeights(conv2d_2_w_path.c_str(), 0,192,64,5,5);
+ std::string conv2d_2_b_path = dir_prefix + std::string("conv_bias3.bin");
+ void* conv2d_2_b = readTrainedWeights(conv2d_2_b_path.c_str(), 0,1,192,1,1);
+ std::string conv2d_3_w_path = dir_prefix + std::string("conv6.bin");
+ void* conv2d_3_w = readTrainedWeights(conv2d_3_w_path.c_str(), 0,384,192,3,3);
+ std::string conv2d_3_b_path = dir_prefix + std::string("conv_bias6.bin");
+ void* conv2d_3_b = readTrainedWeights(conv2d_3_b_path.c_str(), 0,1,384,1,1);
+ std::string conv2d_4_w_path = dir_prefix + std::string("conv7.bin");
+ void* conv2d_4_w = readTrainedWeights(conv2d_4_w_path.c_str(), 0,256,384,3,3);
+ std::string conv2d_4_b_path = dir_prefix + std::string("conv_bias7.bin");
+ void* conv2d_4_b = readTrainedWeights(conv2d_4_b_path.c_str(), 0,1,256,1,1);
+ std::string conv2d_5_w_path = dir_prefix + std::string("conv8.bin");
+ void* conv2d_5_w = readTrainedWeights(conv2d_5_w_path.c_str(), 0,256,256,3,3);
+ std::string conv2d_5_b_path = dir_prefix + std::string("conv_bias8.bin");
+ void* conv2d_5_b = readTrainedWeights(conv2d_5_b_path.c_str(), 0,1,256,1,1);
+ std::string dense_1_w_path = dir_prefix + std::string("fc12.bin");
+ void* dense_1_w = readTrainedWeights(dense_1_w_path.c_str(), 0,1,1,4096,10);
+ std::string dense_1_b_path = dir_prefix + std::string("fc_bias12.bin");
+ void* dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0,1,10,1,1);
+
+
+ __visc__init();
+ RootIn* args = static_cast<RootIn*>(malloc(sizeof(RootIn)));
+
+ 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;
+ args->conv2d_1_b_bytes = 0;
+ args->conv2d_2_w = conv2d_2_w;
+ args->conv2d_2_w_bytes = 0;
+ args->conv2d_2_b = conv2d_2_b;
+ args->conv2d_2_b_bytes = 0;
+ args->conv2d_3_w = conv2d_3_w;
+ args->conv2d_3_w_bytes = 0;
+ args->conv2d_3_b = conv2d_3_b;
+ args->conv2d_3_b_bytes = 0;
+ args->conv2d_4_w = conv2d_4_w;
+ args->conv2d_4_w_bytes = 0;
+ args->conv2d_4_b = conv2d_4_b;
+ args->conv2d_4_b_bytes = 0;
+ args->conv2d_5_w = conv2d_5_w;
+ args->conv2d_5_w_bytes = 0;
+ args->conv2d_5_b = conv2d_5_b;
+ args->conv2d_5_b_bytes = 0;
+ args->dense_1_w = dense_1_w;
+ args->dense_1_w_bytes = 0;
+ args->dense_1_b = dense_1_b;
+ args->dense_1_b_bytes = 0;
+
+ void* dfg = __visc__launch(0, root, (void*) args);
+
+ __visc__wait(dfg);
+
+ void *result = static_cast<RootIn*>(args)->input;
+ hpvm_request_tensor(result, 0);
+
+ __visc__cleanup();
+ computeAccuracy2(labels, 10000, result);
+ return 0;
+
+}
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet_front/Makefile b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet_front/Makefile
deleted file mode 100644
index 86e93cb809336f51b18eaeb110b18dd666367e56..0000000000000000000000000000000000000000
--- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet_front/Makefile
+++ /dev/null
@@ -1,57 +0,0 @@
-DNN_BENCHMARK_ROOT = $(LLVM_SRC_ROOT)/test/VISC/DNN_Benchmarks
-# NOTE: can configure build directory
-HPVM_BUILD_DIR = $(LLVM_SRC_ROOT)/../build_fresh/
-
-CC = $(HPVM_BUILD_DIR)/bin/clang++
-OPT = $(HPVM_BUILD_DIR)/bin/opt
-LLVM_DIS = $(HPVM_BUILD_DIR)/bin/llvm-dis
-LLVM_LINK = $(HPVM_BUILD_DIR)/bin/llvm-link
-LLVM_INCLUDE_DIR = $(LLVM_SRC_ROOT)/include
-
-
-SRC_DIR = src
-BUILD_DIR = build
-APP = alexnet
-
-TENSOR_INCLUDE_DIR = $(DNN_BENCHMARK_ROOT)/common/include
-TENSOR_RT_INCLUDE_DIR = $(LLVM_SRC_ROOT)/projects/hpvm-tensor-rt/tensor_runtime/include
-TENSOR_LIB_DIR = $(LLVM_SRC_ROOT)/projects/hpvm-tensor-rt/lib/libtensor_runtime.a
-TENSOR_AUTOTUNER_DIR = $(LLVM_SRC_ROOT)/projects/hpvm-tensor-rt/lib/libtensor_autotuner.a
-
-CC_FLAGS = -I $(LLVM_INCLUDE_DIR) -I $(TENSOR_INCLUDE_DIR) -I $(TENSOR_RT_INCLUDE_DIR) -I $(CUDA_INCLUDE_PATH) -fno-exceptions -ffast-math -std=c++11 -O3
-CCFLAGS += -DDEVICE=CUDNN_TARGET
-LINKER_FLAGS = -lpthread -lcudart -lcurand -lcudnn -lcublas -lOpenCL
-
-HPVM_LIB_DIR = $(HPVM_BUILD_DIR)/lib
-#HPVM_LIB_DIR = /home/hsharif3/Gitlab/hpvm/build_new/lib
-
-
-VISC_OPTFLAGS = -load $(HPVM_LIB_DIR)/LLVMBuildDFG.so -load $(HPVM_LIB_DIR)/LLVMInPlaceDFGAnalysis.so -load $(HPVM_LIB_DIR)/LLVMDFG2LLVM_CUDNN.so -load $(HPVM_LIB_DIR)/LLVMDFG2LLVM_X86.so -load $(HPVM_LIB_DIR)/LLVMClearDFG.so -inplace -dfg2llvm-cudnn -dfg2llvm-x86 -clearDFG
-
-TARGET = $(BUILD_DIR)/$(APP).opt.bc
-SOURCES = $(SRC_DIR)/$(APP).cpp
-VISC_RT_PATH = $(LLVM_SRC_ROOT)/../build/projects/visc-rt/visc-rt.ll
-
-#OBJS = $(BUILD_DIR)/$(wildcabrd *.ll)
-.PRECIOUS: $(BUILD_DIR)/$(APP).ll $(BUILD_DIR)/$(APP).visc.ll
-default: $(BUILD_DIR) $(TARGET)
-
-
-$(BUILD_DIR)/%.ll: $(SRC_DIR)/%.cpp
- $(CC) $(CC_FLAGS) -emit-llvm -S -o $@ $<
-
-#-visc-timers-gen
-$(BUILD_DIR)/%.visc.ll: $(BUILD_DIR)/%.ll
- $(OPT) -load LLVMGenVISC.so -genvisc -globaldce $< -S -o $@
-
-$(BUILD_DIR)/%.opt.bc: $(BUILD_DIR)/%.visc.ll
- $(OPT) $(VISC_OPTFLAGS) $< -o $@
- $(LLVM_LINK) $@ $(VISC_RT_PATH) -o $(BUILD_DIR)/$(APP)_linked.bc
- $(CC) $(BUILD_DIR)/$(APP)_linked.bc $(TENSOR_LIB_DIR) -o $(BUILD_DIR)/$(APP)_linked $(LINKER_FLAGS)
- $(CC) $(BUILD_DIR)/$(APP)_linked.bc $(TENSOR_AUTOTUNER_DIR) -o $(BUILD_DIR)/$(APP)_tune $(LINKER_FLAGS)
-
-$(BUILD_DIR):
- mkdir -p $@
-
-clean:
- rm -rf $(BUILD_DIR)
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet_front/src/alexnet.cpp b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet_front/src/alexnet.cpp
deleted file mode 100644
index ef6fda1d3194a97f4ec48f7ef8040744e24c7fa1..0000000000000000000000000000000000000000
--- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet_front/src/alexnet.cpp
+++ /dev/null
@@ -1,446 +0,0 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
-#include <visc.h>
-#include <tensorTypes.h>
-#include <tensorUtils.h>
-
-void var_0_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_convolution(t1, t2, 5, 5, 1, 1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_1_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_2_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_3_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_4_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_convolution(t1, t2, 2, 2, 1, 1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_5_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_6_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_7_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_8_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_9_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_10_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_11_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_12_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_13_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_14_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_15_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_16_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_17_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_18_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_mul(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_19_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_20_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_softmax(t1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void root(void* input, size_t input_bytes,
- void* conv2d_1_w, size_t conv2d_1_w_bytes,
- void* conv2d_1_b, size_t conv2d_1_b_bytes,
- void* conv2d_2_w, size_t conv2d_2_w_bytes,
- void* conv2d_2_b, size_t conv2d_2_b_bytes,
- void* conv2d_3_w, size_t conv2d_3_w_bytes,
- void* conv2d_3_b, size_t conv2d_3_b_bytes,
- void* conv2d_4_w, size_t conv2d_4_w_bytes,
- void* conv2d_4_b, size_t conv2d_4_b_bytes,
- void* conv2d_5_w, size_t conv2d_5_w_bytes,
- void* conv2d_5_b, size_t conv2d_5_b_bytes,
- void* dense_1_w, size_t dense_1_w_bytes,
- void* dense_1_b, size_t dense_1_b_bytes){
-
-
- __visc__hint(visc::CPU_TARGET);
- __visc__attributes(13, input, conv2d_1_w, conv2d_1_b, conv2d_2_w, conv2d_2_b, conv2d_3_w, conv2d_3_b, conv2d_4_w, conv2d_4_b, conv2d_5_w, conv2d_5_b, dense_1_w, dense_1_b, 0);
-
-
- void* var_0 = __visc__createNodeND(0, var_0_node);
-
- __visc__bindIn(var_0, 0, 0, 0);
- __visc__bindIn(var_0, 1, 1, 0);
- __visc__bindIn(var_0, 2, 2, 0);
- __visc__bindIn(var_0, 3, 3, 0);
-
- void* var_1 = __visc__createNodeND(0, var_1_node);
-
- __visc__edge(var_0, var_1, 1, 0, 0, 0);
- __visc__edge(var_0, var_1, 1, 1, 1, 0);
- __visc__bindIn(var_1, 4, 2, 0);
- __visc__bindIn(var_1, 5, 3, 0);
-
- void* var_2 = __visc__createNodeND(0, var_2_node);
-
- __visc__edge(var_1, var_2, 1, 0, 0, 0);
- __visc__edge(var_1, var_2, 1, 1, 1, 0);
-
- void* var_3 = __visc__createNodeND(0, var_3_node);
-
- __visc__edge(var_2, var_3, 1, 0, 0, 0);
- __visc__edge(var_2, var_3, 1, 1, 1, 0);
-
- void* var_4 = __visc__createNodeND(0, var_4_node);
-
- __visc__edge(var_3, var_4, 1, 0, 0, 0);
- __visc__edge(var_3, var_4, 1, 1, 1, 0);
- __visc__bindIn(var_4, 6, 2, 0);
- __visc__bindIn(var_4, 7, 3, 0);
-
- void* var_5 = __visc__createNodeND(0, var_5_node);
-
- __visc__edge(var_4, var_5, 1, 0, 0, 0);
- __visc__edge(var_4, var_5, 1, 1, 1, 0);
- __visc__bindIn(var_5, 8, 2, 0);
- __visc__bindIn(var_5, 9, 3, 0);
-
- void* var_6 = __visc__createNodeND(0, var_6_node);
-
- __visc__edge(var_5, var_6, 1, 0, 0, 0);
- __visc__edge(var_5, var_6, 1, 1, 1, 0);
-
- void* var_7 = __visc__createNodeND(0, var_7_node);
-
- __visc__edge(var_6, var_7, 1, 0, 0, 0);
- __visc__edge(var_6, var_7, 1, 1, 1, 0);
-
- void* var_8 = __visc__createNodeND(0, var_8_node);
-
- __visc__edge(var_7, var_8, 1, 0, 0, 0);
- __visc__edge(var_7, var_8, 1, 1, 1, 0);
- __visc__bindIn(var_8, 10, 2, 0);
- __visc__bindIn(var_8, 11, 3, 0);
-
- void* var_9 = __visc__createNodeND(0, var_9_node);
-
- __visc__edge(var_8, var_9, 1, 0, 0, 0);
- __visc__edge(var_8, var_9, 1, 1, 1, 0);
- __visc__bindIn(var_9, 12, 2, 0);
- __visc__bindIn(var_9, 13, 3, 0);
-
- void* var_10 = __visc__createNodeND(0, var_10_node);
-
- __visc__edge(var_9, var_10, 1, 0, 0, 0);
- __visc__edge(var_9, var_10, 1, 1, 1, 0);
-
- void* var_11 = __visc__createNodeND(0, var_11_node);
-
- __visc__edge(var_10, var_11, 1, 0, 0, 0);
- __visc__edge(var_10, var_11, 1, 1, 1, 0);
- __visc__bindIn(var_11, 14, 2, 0);
- __visc__bindIn(var_11, 15, 3, 0);
-
- void* var_12 = __visc__createNodeND(0, var_12_node);
-
- __visc__edge(var_11, var_12, 1, 0, 0, 0);
- __visc__edge(var_11, var_12, 1, 1, 1, 0);
- __visc__bindIn(var_12, 16, 2, 0);
- __visc__bindIn(var_12, 17, 3, 0);
-
- void* var_13 = __visc__createNodeND(0, var_13_node);
-
- __visc__edge(var_12, var_13, 1, 0, 0, 0);
- __visc__edge(var_12, var_13, 1, 1, 1, 0);
-
- void* var_14 = __visc__createNodeND(0, var_14_node);
-
- __visc__edge(var_13, var_14, 1, 0, 0, 0);
- __visc__edge(var_13, var_14, 1, 1, 1, 0);
- __visc__bindIn(var_14, 18, 2, 0);
- __visc__bindIn(var_14, 19, 3, 0);
-
- void* var_15 = __visc__createNodeND(0, var_15_node);
-
- __visc__edge(var_14, var_15, 1, 0, 0, 0);
- __visc__edge(var_14, var_15, 1, 1, 1, 0);
- __visc__bindIn(var_15, 20, 2, 0);
- __visc__bindIn(var_15, 21, 3, 0);
-
- void* var_16 = __visc__createNodeND(0, var_16_node);
-
- __visc__edge(var_15, var_16, 1, 0, 0, 0);
- __visc__edge(var_15, var_16, 1, 1, 1, 0);
-
- void* var_17 = __visc__createNodeND(0, var_17_node);
-
- __visc__edge(var_16, var_17, 1, 0, 0, 0);
- __visc__edge(var_16, var_17, 1, 1, 1, 0);
-
- void* var_18 = __visc__createNodeND(0, var_18_node);
-
- __visc__edge(var_17, var_18, 1, 0, 0, 0);
- __visc__edge(var_17, var_18, 1, 1, 1, 0);
- __visc__bindIn(var_18, 22, 2, 0);
- __visc__bindIn(var_18, 23, 3, 0);
-
- void* var_19 = __visc__createNodeND(0, var_19_node);
-
- __visc__edge(var_18, var_19, 1, 0, 0, 0);
- __visc__edge(var_18, var_19, 1, 1, 1, 0);
- __visc__bindIn(var_19, 24, 2, 0);
- __visc__bindIn(var_19, 25, 3, 0);
-
- void* var_20 = __visc__createNodeND(0, var_20_node);
-
- __visc__edge(var_19, var_20, 1, 0, 0, 0);
- __visc__edge(var_19, var_20, 1, 1, 1, 0);
-
- __visc__bindOut(var_20, 0, 0, 0);
- __visc__bindOut(var_20, 1, 1, 0);
-
-}
-
-struct ret_t {
- void* tensor;
- size_t bytes;
-};
-
-typedef struct __attribute__((__packed__)) {
- void* input;
- size_t input_bytes;
- void* conv2d_1_w;
- size_t conv2d_1_w_bytes;
- void* conv2d_1_b;
- size_t conv2d_1_b_bytes;
- void* conv2d_2_w;
- size_t conv2d_2_w_bytes;
- void* conv2d_2_b;
- size_t conv2d_2_b_bytes;
- void* conv2d_3_w;
- size_t conv2d_3_w_bytes;
- void* conv2d_3_b;
- size_t conv2d_3_b_bytes;
- void* conv2d_4_w;
- size_t conv2d_4_w_bytes;
- void* conv2d_4_b;
- size_t conv2d_4_b_bytes;
- void* conv2d_5_w;
- size_t conv2d_5_w_bytes;
- void* conv2d_5_b;
- size_t conv2d_5_b_bytes;
- void* dense_1_w;
- size_t dense_1_w_bytes;
- void* dense_1_b;
- size_t dense_1_b_bytes;
-
- struct ret_t r;
-}
-RootIn;
-
-int main(){
-
- std::string dir_prefix = std::string("../../../../../../projects/hpvm-tensor-rt/model_params/alexnet_cifar10_front/");
- //std::string dir_prefix = std::string("../../../../../../projects/hpvm-tensor-rt/model_params/alexnet_cifar10/");
-
- std::string input_path = dir_prefix + std::string("input.bin");
- void* input = readTrainedWeights(input_path.c_str(), 0,10000,3,32,32);
- std::string labels_path = dir_prefix + std::string("labels.bin");
- uint8_t* labels = readLabels(labels_path.c_str(),10000);
- std::string conv2d_1_w_path = dir_prefix + std::string("conv0.bin");
- void* conv2d_1_w = readTrainedWeights(conv2d_1_w_path.c_str(), 0,64,3,11,11);
- std::string conv2d_1_b_path = dir_prefix + std::string("conv_bias0.bin");
- void* conv2d_1_b = readTrainedWeights(conv2d_1_b_path.c_str(), 0,1,64,1,1);
- std::string conv2d_2_w_path = dir_prefix + std::string("conv3.bin");
- void* conv2d_2_w = readTrainedWeights(conv2d_2_w_path.c_str(), 0,192,64,5,5);
- std::string conv2d_2_b_path = dir_prefix + std::string("conv_bias3.bin");
- void* conv2d_2_b = readTrainedWeights(conv2d_2_b_path.c_str(), 0,1,192,1,1);
- std::string conv2d_3_w_path = dir_prefix + std::string("conv6.bin");
- void* conv2d_3_w = readTrainedWeights(conv2d_3_w_path.c_str(), 0,384,192,3,3);
- std::string conv2d_3_b_path = dir_prefix + std::string("conv_bias6.bin");
- void* conv2d_3_b = readTrainedWeights(conv2d_3_b_path.c_str(), 0,1,384,1,1);
- std::string conv2d_4_w_path = dir_prefix + std::string("conv7.bin");
- void* conv2d_4_w = readTrainedWeights(conv2d_4_w_path.c_str(), 0,256,384,3,3);
- std::string conv2d_4_b_path = dir_prefix + std::string("conv_bias7.bin");
- void* conv2d_4_b = readTrainedWeights(conv2d_4_b_path.c_str(), 0,1,256,1,1);
- std::string conv2d_5_w_path = dir_prefix + std::string("conv8.bin");
- void* conv2d_5_w = readTrainedWeights(conv2d_5_w_path.c_str(), 0,256,256,3,3);
- std::string conv2d_5_b_path = dir_prefix + std::string("conv_bias8.bin");
- void* conv2d_5_b = readTrainedWeights(conv2d_5_b_path.c_str(), 0,1,256,1,1);
- std::string dense_1_w_path = dir_prefix + std::string("fc12.bin");
- void* dense_1_w = readTrainedWeights(dense_1_w_path.c_str(), 0,1,1,4096,10);
- std::string dense_1_b_path = dir_prefix + std::string("fc_bias12.bin");
- void* dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0,1,10,1,1);
-
-
- __visc__init();
- RootIn* args = static_cast<RootIn*>(malloc(sizeof(RootIn)));
-
- 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;
- args->conv2d_1_b_bytes = 0;
- args->conv2d_2_w = conv2d_2_w;
- args->conv2d_2_w_bytes = 0;
- args->conv2d_2_b = conv2d_2_b;
- args->conv2d_2_b_bytes = 0;
- args->conv2d_3_w = conv2d_3_w;
- args->conv2d_3_w_bytes = 0;
- args->conv2d_3_b = conv2d_3_b;
- args->conv2d_3_b_bytes = 0;
- args->conv2d_4_w = conv2d_4_w;
- args->conv2d_4_w_bytes = 0;
- args->conv2d_4_b = conv2d_4_b;
- args->conv2d_4_b_bytes = 0;
- args->conv2d_5_w = conv2d_5_w;
- args->conv2d_5_w_bytes = 0;
- args->conv2d_5_b = conv2d_5_b;
- args->conv2d_5_b_bytes = 0;
- args->dense_1_w = dense_1_w;
- args->dense_1_w_bytes = 0;
- args->dense_1_b = dense_1_b;
- args->dense_1_b_bytes = 0;
-
- void* dfg = __visc__launch(0, root, (void*) args);
-
- __visc__wait(dfg);
-
- void *result = static_cast<RootIn*>(args)->input;
- hpvm_request_tensor(result, 0);
-
- __visc__cleanup();
- computeAccuracy2(labels, 10000, result);
- return 0;
-
-}
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/fc_mnist_3/Makefile b/llvm/test/VISC/DNN_Benchmarks/benchmarks/fc_mnist_3/Makefile
deleted file mode 100644
index 17e0c11fec7b49617c2ad5a68cdc5cb3e394df7a..0000000000000000000000000000000000000000
--- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/fc_mnist_3/Makefile
+++ /dev/null
@@ -1,36 +0,0 @@
-DNN_BENCHMARK_ROOT = $(LLVM_SRC_ROOT)/test/VISC/DNN_Benchmarks
-CC = $(LLVM_SRC_ROOT)/../build/bin/clang++
-OPT = $(LLVM_SRC_ROOT)/../build/bin/opt
-LLVM_INCLUDE_DIR = $(LLVM_SRC_ROOT)/include
-
-SRC_DIR = src
-BUILD_DIR = build
-APP = fc_mnist_3
-
-TENSOR_INCLUDE_DIR = $(DNN_BENCHMARK_ROOT)/common/include
-
-CC_FLAGS = -I $(LLVM_INCLUDE_DIR) -I $(TENSOR_INCLUDE_DIR) -fno-exceptions -ffast-math -std=c++11 -O3
-
-VISC_OPTFLAGS = -load LLVMBuildDFG.so -load LLVMInPlaceDFGAnalysis.so -load LLVMDFG2LLVM_CUDNN.so -load LLVMDFG2LLVM_X86.so -load LLVMClearDFG.so -inplace -dfg2llvm-cudnn -dfg2llvm-x86 -clearDFG
-
-TARGET = $(BUILD_DIR)/$(APP).opt.ll
-SOURCES = $(SRC_DIR)/$(APP).cpp
-#OBJS = $(BUILD_DIR)/$(wildcabrd *.ll)
-.PRECIOUS: $(BUILD_DIR)/$(APP).ll $(BUILD_DIR)/$(APP).visc.ll
-default: $(BUILD_DIR) $(TARGET)
-
-
-$(BUILD_DIR)/%.ll: $(SRC_DIR)/%.cpp
- $(CC) $(CC_FLAGS) -emit-llvm -S -o $@ $<
-
-$(BUILD_DIR)/%.visc.ll: $(BUILD_DIR)/%.ll
- $(OPT) -load LLVMGenVISC.so -genvisc -globaldce -visc-timers-gen $< -S -o $@
-
-$(BUILD_DIR)/%.opt.ll: $(BUILD_DIR)/%.visc.ll
- $(OPT) $(VISC_OPTFLAGS) $< -o $@
-
-$(BUILD_DIR):
- mkdir -p $@
-
-clean:
- rm -rf $(BUILD_DIR)
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/fc_mnist_3/src/fc_mnist_3.cpp b/llvm/test/VISC/DNN_Benchmarks/benchmarks/fc_mnist_3/src/fc_mnist_3.cpp
deleted file mode 100644
index d1848504744af2897a1e741be676fa1338764ea1..0000000000000000000000000000000000000000
--- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/fc_mnist_3/src/fc_mnist_3.cpp
+++ /dev/null
@@ -1,305 +0,0 @@
-#include <iostream>
-#include <cstdio>
-#include <cstring>
-#include <cinttypes>
-#include <visc.h>
-#include <tensorTypes.h>
-#include <tensorUtils.h>
-
-using namespace std;
-
-void L1_FC(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::DEVICE);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_mul(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void L1_Bias(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::DEVICE);
- __visc__attributes(2, t1, t2, 0);
-
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void L1_Act(void *t, size_t bytest) {
- __visc__hint(visc::DEVICE);
- __visc__attributes(1, t, 0);
-
- void* r = __visc__tensor_relu(t);
- __visc__return(2, r, (size_t) 0);
-}
-
-void L2_FC(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::DEVICE);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_mul(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void L2_Bias(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::DEVICE);
- __visc__attributes(2, t1, t2, 0);
-
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void L2_Act(void *t, size_t bytest) {
- __visc__hint(visc::DEVICE);
- __visc__attributes(1, t, 0);
-
- void* r = __visc__tensor_relu(t);
- __visc__return(2, r, (size_t) 0);
-}
-
-void L3_FC(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::DEVICE);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_mul(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void L3_Bias(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::DEVICE);
- __visc__attributes(2, t1, t2, 0);
-
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void L3_Act(void *t, size_t bytest) {
- __visc__hint(visc::DEVICE);
- __visc__attributes(1, t, 0);
-
- void* r = __visc__tensor_relu(t);
- __visc__return(2, r, (size_t) 0);
-}
-
-void L4_FC(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::DEVICE);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_mul(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void L4_Bias(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::DEVICE);
- __visc__attributes(2, t1, t2, 0);
-
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void L4_SoftMax(void *t, size_t bytest) {
- __visc__hint(visc::DEVICE);
- __visc__attributes(1, t, 0);
-
- void* r = __visc__tensor_softmax(t);
- __visc__return(2, r, (size_t) 0);
-}
-
-void root(void* input, size_t bytes_input, // 0
- void* fc1_weights, size_t bytes_fc1_weights, // 2
- void* fc1_bias, size_t bytes_fc1_bias, // 4
- void* fc2_weights, size_t bytes_fc2_weights, // 6
- void* fc2_bias, size_t bytes_fc2_bias, // 8
- void* fc3_weights, size_t bytes_fc3_weights, // 10
- void* fc3_bias, size_t bytes_fc3_bias, // 12
- void* fc4_weights, size_t bytes_fc4_weights, // 14
- void* fc4_bias, size_t bytes_fc4_bias // 16
- ) {
-
- __visc__hint(visc::CPU_TARGET);
- __visc__attributes(9, input, fc1_weights, fc1_bias, fc2_weights, fc2_bias, fc3_weights, fc3_bias, fc4_weights, fc4_bias, 0);
-
- void *node_L1_FC = __visc__createNodeND(0, L1_FC);
- void *node_L1_Bias = __visc__createNodeND(0, L1_Bias);
- void *node_L1_Act = __visc__createNodeND(0, L1_Act);
- void *node_L2_FC = __visc__createNodeND(0, L2_FC);
- void *node_L2_Bias = __visc__createNodeND(0, L2_Bias);
- void *node_L2_Act = __visc__createNodeND(0, L2_Act);
- void *node_L3_FC = __visc__createNodeND(0, L3_FC);
- void *node_L3_Bias = __visc__createNodeND(0, L3_Bias);
- void *node_L3_Act = __visc__createNodeND(0, L3_Act);
- void *node_L4_FC = __visc__createNodeND(0, L4_FC);
- void *node_L4_Bias = __visc__createNodeND(0, L4_Bias);
- void *node_L4_Softmax = __visc__createNodeND(0, L4_SoftMax);
-
-
- // Layer 1
- __visc__bindIn(node_L1_FC, 0, 0, 0); // input
- __visc__bindIn(node_L1_FC, 1, 1, 0); // bytes_input
- __visc__bindIn(node_L1_FC, 2, 2, 0); // fc1_weights
- __visc__bindIn(node_L1_FC, 3, 3, 0); // bytes
-
- __visc__edge(node_L1_FC, node_L1_Bias, 1, 0, 0, 0); // fc1_out
- __visc__edge(node_L1_FC, node_L1_Bias, 1, 1, 1, 0); // bytes
- __visc__bindIn(node_L1_Bias, 4, 2, 0); // fc1_bias
- __visc__bindIn(node_L1_Bias, 5, 3, 0); // bytes
-
- __visc__edge(node_L1_Bias, node_L1_Act, 1, 0, 0, 0); // fc1_bias_out
- __visc__edge(node_L1_Bias, node_L1_Act, 1, 1, 1, 0); // bytes
-
- // Layer 2
- __visc__edge(node_L1_Act, node_L2_FC, 1, 0, 0, 0); // fc1_act_out
- __visc__edge(node_L1_Act, node_L2_FC, 1, 1, 1, 0); // bytes
- __visc__bindIn(node_L2_FC, 6, 2, 0); // fc2_weights
- __visc__bindIn(node_L2_FC, 7, 3, 0); // bytes
-
- __visc__edge(node_L2_FC, node_L2_Bias, 1, 0, 0, 0); // fc2_out
- __visc__edge(node_L2_FC, node_L2_Bias, 1, 1, 1, 0); // bytes
- __visc__bindIn(node_L2_Bias, 8, 2, 0); // fc2_bias
- __visc__bindIn(node_L2_Bias, 9, 3, 0); // bytes
-
- __visc__edge(node_L2_Bias, node_L2_Act, 1, 0, 0, 0); // fc2_bias_out
- __visc__edge(node_L2_Bias, node_L2_Act, 1, 1, 1, 0); // bytes
-
- // Layer 3
- __visc__edge(node_L2_Act, node_L3_FC, 1, 0, 0, 0); // fc2_act_out
- __visc__edge(node_L2_Act, node_L3_FC, 1, 1, 1, 0); // bytes
- __visc__bindIn(node_L3_FC, 10, 2, 0); // fc3_weights
- __visc__bindIn(node_L3_FC, 11, 3, 0); // bytes
-
- __visc__edge(node_L3_FC, node_L3_Bias, 1, 0, 0, 0); // fc3_out
- __visc__edge(node_L3_FC, node_L3_Bias, 1, 1, 1, 0); // bytes
- __visc__bindIn(node_L3_Bias, 12, 2, 0); // fc3_bias
- __visc__bindIn(node_L3_Bias, 13, 3, 0); // bytes
-
- __visc__edge(node_L3_Bias, node_L3_Act, 1, 0, 0, 0); // fc3_bias_out
- __visc__edge(node_L3_Bias, node_L3_Act, 1, 1, 1, 0); // bytes
-
- // Layer 4
- __visc__edge(node_L3_Act, node_L4_FC, 1, 0, 0, 0); // fc3_act_out
- __visc__edge(node_L3_Act, node_L4_FC, 1, 1, 1, 0); // bytes
- __visc__bindIn(node_L4_FC, 14, 2, 0); // fc4_weights
- __visc__bindIn(node_L4_FC, 15, 3, 0); // bytes
-
- __visc__edge(node_L4_FC, node_L4_Bias, 1, 0, 0, 0); // fc4_out
- __visc__edge(node_L4_FC, node_L4_Bias, 1, 1, 1, 0); // bytes
- __visc__bindIn(node_L4_Bias, 16, 2, 0); // fc4_bias
- __visc__bindIn(node_L4_Bias, 17, 3, 0); // bytes
-
- __visc__edge(node_L4_Bias, node_L4_Softmax, 1, 0, 0, 0); // fc3_bias_out
- __visc__edge(node_L4_Bias, node_L4_Softmax, 1, 1, 1, 0); // bytes
-
- __visc__bindOut(node_L4_Softmax, 0, 0, 0); // softmax output
- __visc__bindOut(node_L4_Softmax, 1, 1, 0); // bytes
-
-}
-
-
-// Return type for the nodes
-struct ret_t {
- void *tensor;
- size_t bytes;
-};
-
-typedef struct __attribute__((__packed__)) {
- void *input;
- size_t bytes_input;
- void *fc1_weights;
- size_t bytes_fc1_weights;
- void *fc1_bias;
- size_t bytes_fc1_bias;
- void *fc2_weights;
- size_t bytes_fc2_weights;
- void *fc2_bias;
- size_t bytes_fc2_bias;
- void *fc3_weights;
- size_t bytes_fc3_weights;
- void *fc3_bias;
- size_t bytes_fc3_bias;
- void *fc4_weights;
- size_t bytes_fc4_weights;
- void *fc4_bias;
- size_t bytes_fc4_bias;
-
- struct ret_t r;
-}
-RootIn;
-
-void packLaunchArguments(RootIn* args,
- void* input,
- void* fc1_weights,
- void* fc1_bias,
- void* fc2_weights,
- void* fc2_bias,
- void* fc3_weights,
- void* fc3_bias,
- void* fc4_weights,
- void* fc4_bias) {
- args->input = input;
- args->bytes_input = 0;
- args->fc1_weights = fc1_weights;
- args->bytes_fc1_weights = 0;
- args->fc1_bias = fc1_bias;
- args->bytes_fc1_bias = 0;
- args->fc2_weights = fc2_weights;
- args->bytes_fc2_weights = 0;
- args->fc2_bias = fc2_bias;
- args->bytes_fc2_bias = 0;
- args->fc3_weights = fc3_weights;
- args->bytes_fc3_weights = 0;
- args->fc3_bias = fc3_bias;
- args->bytes_fc3_bias = 0;
- args->fc4_weights = fc4_weights;
- args->bytes_fc4_weights = 0;
- args->fc4_bias = fc4_bias;
- args->bytes_fc4_bias = 0;
-
-}
-
-int main() {
-
- int test_batch_size = 10000;
- void* input = readInputTensor("t10k-images-idx3-ubyte", float_type,
- test_batch_size, 1, 28, 28);
- void* fc1_weights = readTrainedWeights("../model_params/FC_network1/fc1.bin",
- float_type, 1, 1, 784, 1000);
- void* fc1_bias = readTrainedWeights("../model_params/FC_network1/fc1_bias.bin",
- float_type, 1, 1000, 1, 1);
- void* fc2_weights = readTrainedWeights("../model_params/FC_network1/fc2.bin",
- float_type, 1, 1, 1000, 500);
- void* fc2_bias = readTrainedWeights("../model_params/FC_network1/fc2_bias.bin",
- float_type, 1, 500, 1, 1);
- void* fc3_weights = readTrainedWeights("../model_params/FC_network1/fc3.bin",
- float_type, 1, 1, 500, 200);
- void* fc3_bias = readTrainedWeights("../model_params/FC_network1/fc3_bias.bin",
- float_type, 1, 200, 1, 1);
- void* fc4_weights = readTrainedWeights("../model_params/FC_network1/fc4.bin",
- float_type, 1, 1, 200, 10);
- void* fc4_bias = readTrainedWeights("../model_params/FC_network1/fc4_bias.bin",
- float_type, 1, 10, 1, 1);
-
- __visc__init();
-
- RootIn* args = static_cast<RootIn*>(malloc(sizeof(RootIn)));
- packLaunchArguments(args,
- input,
- fc1_weights,
- fc1_bias,
- fc2_weights,
- fc2_bias,
- fc3_weights,
- fc3_bias,
- fc4_weights,
- fc4_bias);
-
- void *dfg = __visc__launch(0, root, (void *)args);
-
- __visc__wait(dfg);
-
- void *r = args->r.tensor;
- hpvm_request_tensor(r, visc::CPU_TARGET);
-
- __visc__cleanup();
- return 0;
-}
-
-
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet/Makefile b/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet/Makefile
index 479b0a2caf914e314b991577ada8d7dc20abff66..aeb8cd6bb1606c496258fcd92f28e1b3d0f10566 100644
--- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet/Makefile
+++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet/Makefile
@@ -1,8 +1,11 @@
DNN_BENCHMARK_ROOT = $(LLVM_SRC_ROOT)/test/VISC/DNN_Benchmarks
-CC = $(LLVM_SRC_ROOT)/../build/bin/clang++
-OPT = $(LLVM_SRC_ROOT)/../build/bin/opt
-LLVM_DIS = $(LLVM_SRC_ROOT)/../build/bin/llvm-dis
-LLVM_LINK = $(LLVM_SRC_ROOT)/../build/bin/llvm-link
+# NOTE: can configure build directory
+HPVM_BUILD_DIR = $(LLVM_SRC_ROOT)/../build_hpvm/
+
+CC = $(HPVM_BUILD_DIR)/bin/clang++
+OPT = $(HPVM_BUILD_DIR)/bin/opt
+LLVM_DIS = $(HPVM_BUILD_DIR)/bin/llvm-dis
+LLVM_LINK = $(HPVM_BUILD_DIR)/bin/llvm-link
LLVM_INCLUDE_DIR = $(LLVM_SRC_ROOT)/include
SRC_DIR = src
@@ -18,7 +21,15 @@ CC_FLAGS = -I $(LLVM_INCLUDE_DIR) -I $(TENSOR_INCLUDE_DIR) -I $(TENSOR_RT_INCLUD
CCFLAGS += -DDEVICE=CUDNN_TARGET
LINKER_FLAGS = -lpthread -lcudart -lcurand -lcudnn -lcublas -lOpenCL
-VISC_OPTFLAGS = -load LLVMBuildDFG.so -load LLVMInPlaceDFGAnalysis.so -load LLVMDFG2LLVM_CUDNN.so -load LLVMDFG2LLVM_X86.so -load LLVMClearDFG.so -inplace -dfg2llvm-cudnn -dfg2llvm-x86 -clearDFG
+HPVM_LIB_DIR = $(HPVM_BUILD_DIR)/lib
+
+
+VISC_OPTFLAGS = -load $(HPVM_LIB_DIR)/LLVMBuildDFG.so -load $(HPVM_LIB_DIR)/LLVMInPlaceDFGAnalysis.so -load $(HPVM_LIB_DIR)/LLVMDFG2LLVM_CUDNN.so -load $(HPVM_LIB_DIR)/LLVMDFG2LLVM_X86.so -load $(HPVM_LIB_DIR)/LLVMClearDFG.so -inplace -dfg2llvm-cudnn -dfg2llvm-x86 -clearDFG
+
+
+VISC_OPTFLAGS2 = -load $(HPVM_LIB_DIR)/LLVMBuildDFG.so -load $(HPVM_LIB_DIR)/LLVMInPlaceDFGAnalysis.so -load $(HPVM_LIB_DIR)/LLVMDFG2LLVM_PROMISE.so -load $(HPVM_LIB_DIR)/LLVMDFG2LLVM_X86.so -load $(HPVM_LIB_DIR)/LLVMFuseHPVMTensorNodes.so -load $(HPVM_LIB_DIR)/LLVMClearDFG.so -inplace -hpvm-fuse -dfg2llvm-promise -dfg2llvm-x86 -clearDFG
+
+
TARGET = $(BUILD_DIR)/$(APP).opt.bc
SOURCES = $(SRC_DIR)/$(APP).cpp
@@ -30,17 +41,20 @@ default: $(BUILD_DIR) $(TARGET)
$(BUILD_DIR)/%.ll: $(SRC_DIR)/%.cpp
- $(CC) $(CC_FLAGS) -emit-llvm -S -o $@ $<
-
-#-visc-timers-gen
-$(BUILD_DIR)/%.visc.ll: $(BUILD_DIR)/%.ll
- $(OPT) -load LLVMGenVISC.so -genvisc -globaldce $< -S -o $@
-
-$(BUILD_DIR)/%.opt.bc: $(BUILD_DIR)/%.visc.ll
- $(OPT) $(VISC_OPTFLAGS) $< -o $@
- $(LLVM_LINK) $@ $(VISC_RT_PATH) -o $(BUILD_DIR)/lenet_linked.bc
- $(CC) $(BUILD_DIR)/lenet_linked.bc $(TENSOR_LIB_DIR) -o $(BUILD_DIR)/lenet_linked $(LINKER_FLAGS)
- $(CC) $(BUILD_DIR)/lenet_linked.bc $(TENSOR_AUTOTUNER_DIR) -o $(BUILD_DIR)/lenet_tune $(LINKER_FLAGS)
+ $(CC) $(CC_FLAGS) -emit-llvm src/$(APP).cpp -S -o $(BUILD_DIR)/$(APP).ll
+ $(CC) $(CC_FLAGS) -emit-llvm src/$(APP)_promise.cpp -S -o $(BUILD_DIR)/$(APP)_promise.ll
+
+
+$(BUILD_DIR)/%.opt.bc: $(BUILD_DIR)/%.ll
+ $(OPT) -load LLVMGenVISC.so -genvisc -globaldce $(BUILD_DIR)/$(APP).ll -S -o $(BUILD_DIR)/$(APP).visc.ll
+ $(OPT) -load LLVMGenVISC.so -genvisc -globaldce $(BUILD_DIR)/$(APP)_promise.ll -S -o $(BUILD_DIR)/$(APP)_promise.visc.ll
+ $(OPT) $(VISC_OPTFLAGS) $(BUILD_DIR)/$(APP).visc.ll -o $(BUILD_DIR)/$(APP)_cudnn.bc
+ $(OPT) $(VISC_OPTFLAGS2) $(BUILD_DIR)/$(APP)_promise.visc.ll -o $(BUILD_DIR)/$(APP)_promise.bc
+ $(LLVM_LINK) $(BUILD_DIR)/$(APP)_cudnn.bc $(VISC_RT_PATH) -o $(BUILD_DIR)/$(APP)_cudnn_linked.bc
+ $(LLVM_LINK) $(BUILD_DIR)/$(APP)_promise.bc $(VISC_RT_PATH) -o $(BUILD_DIR)/$(APP)_promise_linked.bc
+ $(CC) $(BUILD_DIR)/$(BUILD_DIR)/$(APP)_cudnn_linked.bc $(TENSOR_LIB_DIR) -o $(BUILD_DIR)/$(APP)_cudnn_linked $(LINKER_FLAGS)
+ $(CC) $(BUILD_DIR)/$(BUILD_DIR)/$(APP)_promise_linked.bc $(TENSOR_LIB_DIR) -o $(BUILD_DIR)/$(APP)_promise_linked $(LINKER_FLAGS)
+ $(CC) $(BUILD_DIR)/$(BUILD_DIR)/$(APP)_cudnn_linked.bc $(TENSOR_AUTOTUNER_DIR) -o $(BUILD_DIR)/lenet_tune $(LINKER_FLAGS)
$(BUILD_DIR):
mkdir -p $@
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet/src/lenet.cpp b/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet/src/lenet.cpp
index 7698e1511b4dbc8f15eea3386ec98b338178fd86..1746fc13dc4809f8c3d806fa144903fac50f3315 100644
--- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet/src/lenet.cpp
+++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet/src/lenet.cpp
@@ -1,382 +1,373 @@
-#include <iostream>
-#include <cstdio>
-#include <cstring>
-#include <cinttypes>
-#include <visc.h>
-#include <tensorTypes.h>
-#include <tensorUtils.h>
-
-using namespace std;
-
-/* DNN Layer 1 **/
-void tensorConvNode1(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- // X * W = t2 * t1
- void *r = __visc__tensor_convolution(t1, t2, 2, 2, 1, 1);
- __visc__return(2, r, (size_t) 0);
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <sys/stat.h>
+#include <cstring>
+#include <visc.h>
+#include <tensorTypes.h>
+#include <tensorUtils.h>
+
+void var_0_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
+
+ void *r = __visc__tensor_convolution(t1, t2, 2, 2, 1, 1);
+ __visc__return(2, r, (size_t) 0);
}
+void var_1_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
-void tensorAddNode1(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+ void *r = __visc__tensor_add(t1, t2);
+ __visc__return(2, r, (size_t) 0);
+}
+
+void var_2_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
+ void* r = __visc__tensor_relu(t1);
+ __visc__return(2, r, (size_t) 0);
}
-void tensorPoolNode1(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+void var_3_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
- void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
- __visc__return(2, r, (size_t) 0);
+ void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
+ __visc__return(2, r, (size_t) 0);
}
-void tensorTanhNode1(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+void var_4_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_convolution(t1, t2, 2, 2, 1, 1);
+ __visc__return(2, r, (size_t) 0);
}
-/** End of Layer 1 **/
+void var_5_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
+
+ void *r = __visc__tensor_add(t1, t2);
+ __visc__return(2, r, (size_t) 0);
+}
+void var_6_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
+ void* r = __visc__tensor_relu(t1);
+ __visc__return(2, r, (size_t) 0);
+}
-/* DNN Layer 2 **/
-void tensorConvNode2(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+void var_7_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- // X * W = t2 * t1
- void *r = __visc__tensor_convolution(t1, t2, 2, 2, 1, 1);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_convolution(t1, t2, 1, 1, 2, 2);
+ __visc__return(2, r, (size_t) 0);
}
+void var_8_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
+
+ void *r = __visc__tensor_add(t1, t2);
+ __visc__return(2, r, (size_t) 0);
+}
-void tensorAddNode2(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+void var_9_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
+ void* r = __visc__tensor_relu(t1);
+ __visc__return(2, r, (size_t) 0);
}
-void tensorPoolNode2(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+void var_10_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_mul(t1, t2);
+ __visc__return(2, r, (size_t) 0);
}
-void tensorTanhNode2(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+void var_11_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_add(t1, t2);
+ __visc__return(2, r, (size_t) 0);
}
-/** End of Layer 2 **/
+void var_12_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
+ void* r = __visc__tensor_relu(t1);
+ __visc__return(2, r, (size_t) 0);
+}
-/***** DNN Layer3 ****/
-void tensorMulNode3(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+void var_13_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- // X * W = t2 * t1
- void *r = __visc__tensor_mul(t1, t2);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_mul(t1, t2);
+ __visc__return(2, r, (size_t) 0);
}
-void tensorAddNode3(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
+void var_14_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(2, t1, t2, 0);
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
+ void *r = __visc__tensor_add(t1, t2);
+ __visc__return(2, r, (size_t) 0);
}
-void tensorTanhNode3(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+void var_15_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
+ void* r = __visc__tensor_relu(t1);
+ __visc__return(2, r, (size_t) 0);
}
-/** End of Layer 3 **/
+void var_16_node(void* t1, size_t bytes_t1) {
+ __visc__hint(visc::CUDNN_TARGET);
+ __visc__attributes(1, t1, 0);
+ void* r = __visc__tensor_softmax(t1);
+ __visc__return(2, r, (size_t) 0);
+}
+void root(void* input, size_t input_bytes,
+ void* conv2d_1_w, size_t conv2d_1_w_bytes,
+ void* conv2d_1_b, size_t conv2d_1_b_bytes,
+ void* conv2d_2_w, size_t conv2d_2_w_bytes,
+ void* conv2d_2_b, size_t conv2d_2_b_bytes,
+ void* conv2d_3_w, size_t conv2d_3_w_bytes,
+ void* conv2d_3_b, size_t conv2d_3_b_bytes,
+ void* dense_1_w, size_t dense_1_w_bytes,
+ void* dense_1_b, size_t dense_1_b_bytes,
+ void* dense_2_w, size_t dense_2_w_bytes,
+ void* dense_2_b, size_t dense_2_b_bytes){
-/***** DNN Layer4 ****/
-void tensorMulNode4(void *t1, size_t bytes1, void *t2, size_t bytes2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
- // X * W = t2 * t1
- void *r = __visc__tensor_mul(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
+ __visc__hint(visc::CPU_TARGET);
+ __visc__attributes(11, input, conv2d_1_w, conv2d_1_b, conv2d_2_w, conv2d_2_b, conv2d_3_w, conv2d_3_b, dense_1_w, dense_1_b, dense_2_w, dense_2_b, 0);
-void tensorAddNode4(void *t1, size_t bytest1, void *t2, size_t bytest2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
- void* r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
+ void* var_0 = __visc__createNodeND(0, var_0_node);
-void tensorTanhNode4(void *t1, size_t bytest1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
+ __visc__bindIn(var_0, 0, 0, 0);
+ __visc__bindIn(var_0, 1, 1, 0);
+ __visc__bindIn(var_0, 2, 2, 0);
+ __visc__bindIn(var_0, 3, 3, 0);
- void* r = __visc__tensor_tanh(t1);
- __visc__return(2, r, (size_t) 0);
-}
+ void* var_1 = __visc__createNodeND(0, var_1_node);
-/** End of Layer 4 **/
-
-
-
-void root(void *x, size_t x_bytes,
- void *conv1_w, size_t conv1_w_bytes,
- void *conv1_b, size_t conv1_b_bytes,
- void *conv2_w, size_t conv2_w_bytes,
- void *conv2_b, size_t conv2_b_bytes,
- void *fc1_w, size_t fc1_w_bytes,
- void *fc1_b, size_t fc1_b_bytes,
- void *fc2_w, size_t fc2_w_bytes,
- void *fc2_b, size_t fc2_b_bytes){
-
- __visc__hint(visc::CPU_TARGET);
- __visc__attributes(5, x, conv1_w, conv1_b, conv2_w, conv2_b, 0);
-
- // Conv1 Nodes
- void *nodeConv1 = __visc__createNodeND(0, tensorConvNode1);
- void *nodeAdd1 = __visc__createNodeND(0, tensorAddNode1);
- void *nodePool1 = __visc__createNodeND(0, tensorPoolNode1);
- void *nodeTanh1 = __visc__createNodeND(0, tensorTanhNode1);
- // Conv2 Nodes
- void *nodeConv2 = __visc__createNodeND(0, tensorConvNode2);
- void *nodeAdd2 = __visc__createNodeND(0, tensorAddNode2);
- void *nodePool2 = __visc__createNodeND(0, tensorPoolNode2);
- void *nodeTanh2 = __visc__createNodeND(0, tensorTanhNode2);
- // FC1 Nodes
- void *nodeMul3 = __visc__createNodeND(0, tensorMulNode3);
- void *nodeAdd3 = __visc__createNodeND(0, tensorAddNode3);
- void *nodeTanh3 = __visc__createNodeND(0, tensorTanhNode3);
- // FC2 Nodes
- void *nodeMul4 = __visc__createNodeND(0, tensorMulNode4);
- void *nodeAdd4 = __visc__createNodeND(0, tensorAddNode4);
- void *nodeTanh4 = __visc__createNodeND(0, tensorTanhNode4);
-
-
- //***** Conv Layer 1 *******/
- // node, src, dst, stream
- __visc__bindIn(nodeConv1, 0, 0, 0);
- __visc__bindIn(nodeConv1, 1, 1, 0);
- __visc__bindIn(nodeConv1, 2, 2, 0);
- __visc__bindIn(nodeConv1, 3, 3, 0);
-
- // node, node, type, src, dst, stream
- __visc__edge(nodeConv1, nodeAdd1, 1, 0, 0, 0);
- __visc__edge(nodeConv1, nodeAdd1, 1, 1, 1, 0);
-
- // parent_index, dest_index, bind_type
- __visc__bindIn(nodeAdd1, 4, 2, 0);
- __visc__bindIn(nodeAdd1, 5, 3, 0);
-
- // node, node, type, src, dst, stream
- __visc__edge(nodeAdd1, nodePool1, 1, 0, 0, 0);
- __visc__edge(nodeAdd1, nodePool1, 1, 1, 1, 0);
-
- // node, node, type, src, dst, stream
- __visc__edge(nodePool1, nodeTanh1, 1, 0, 0, 0);
- __visc__edge(nodePool1, nodeTanh1, 1, 1, 1, 0);
-
-
- /**** Conv Layer 2 ****/
- // ConvOp2
- __visc__bindIn(nodeConv2, 6, 2, 0);
- __visc__bindIn(nodeConv2, 7, 3, 0);
-
- __visc__edge(nodeTanh1, nodeConv2, 1, 0, 0, 0);
- __visc__edge(nodeTanh1, nodeConv2, 1, 1, 1, 0);
-
- // AddOp2
- __visc__bindIn(nodeAdd2, 8, 2, 0);
- __visc__bindIn(nodeAdd2, 9, 3, 0);
-
- __visc__edge(nodeConv2, nodeAdd2, 1, 0, 0, 0);
- __visc__edge(nodeConv2, nodeAdd2, 1, 1, 1, 0);
-
- // PoolOp2
- __visc__edge(nodeAdd2, nodePool2, 1, 0, 0, 0);
- __visc__edge(nodeAdd2, nodePool2, 1, 1, 1, 0);
-
- // TanhOp2
- __visc__edge(nodePool2, nodeTanh2, 1, 0, 0, 0);
- __visc__edge(nodePool2, nodeTanh2, 1, 1, 1, 0);
-
-
- /**** FC Layer 1 ****/
- // MulOp3
- __visc__bindIn(nodeMul3, 10, 2, 0);
- __visc__bindIn(nodeMul3, 11, 3, 0);
-
- __visc__edge(nodeTanh2, nodeMul3, 1, 0, 0, 0);
- __visc__edge(nodeTanh2, nodeMul3, 1, 1, 1, 0);
-
- // AddOp3
- __visc__bindIn(nodeAdd3, 12, 2, 0);
- __visc__bindIn(nodeAdd3, 13, 3, 0);
-
- __visc__edge(nodeMul3, nodeAdd3, 1, 0, 0, 0);
- __visc__edge(nodeMul3, nodeAdd3, 1, 1, 1, 0);
-
- // TanhOp3
- __visc__edge(nodeAdd3, nodeTanh3, 1, 0, 0, 0);
- __visc__edge(nodeAdd3, nodeTanh3, 1, 1, 1, 0);
-
-
- /**** FC Layer 2 ****/
- // MulOp4
- __visc__bindIn(nodeMul4, 14, 2, 0);
- __visc__bindIn(nodeMul4, 15, 3, 0);
-
- __visc__edge(nodeTanh3, nodeMul4, 1, 0, 0, 0);
- __visc__edge(nodeTanh3, nodeMul4, 1, 1, 1, 0);
-
- // AddOp4
- __visc__bindIn(nodeAdd4, 16, 2, 0);
- __visc__bindIn(nodeAdd4, 17, 3, 0);
-
- __visc__edge(nodeMul4, nodeAdd4, 1, 0, 0, 0);
- __visc__edge(nodeMul4, nodeAdd4, 1, 1, 1, 0);
-
- // TanhOp4
- __visc__edge(nodeAdd4, nodeTanh4, 1, 0, 0, 0);
- __visc__edge(nodeAdd4, nodeTanh4, 1, 1, 1, 0);
-
-
-
- /***** Output Binding ****/
- __visc__bindOut(nodeTanh4, 0, 0, 0);
- __visc__bindOut(nodeTanh4, 1, 1, 0);
+ __visc__edge(var_0, var_1, 1, 0, 0, 0);
+ __visc__edge(var_0, var_1, 1, 1, 1, 0);
+ __visc__bindIn(var_1, 4, 2, 0);
+ __visc__bindIn(var_1, 5, 3, 0);
-}
+ void* var_2 = __visc__createNodeND(0, var_2_node);
+
+ __visc__edge(var_1, var_2, 1, 0, 0, 0);
+ __visc__edge(var_1, var_2, 1, 1, 1, 0);
+
+ void* var_3 = __visc__createNodeND(0, var_3_node);
+
+ __visc__edge(var_2, var_3, 1, 0, 0, 0);
+ __visc__edge(var_2, var_3, 1, 1, 1, 0);
+
+ void* var_4 = __visc__createNodeND(0, var_4_node);
+
+ __visc__edge(var_3, var_4, 1, 0, 0, 0);
+ __visc__edge(var_3, var_4, 1, 1, 1, 0);
+ __visc__bindIn(var_4, 6, 2, 0);
+ __visc__bindIn(var_4, 7, 3, 0);
+
+ void* var_5 = __visc__createNodeND(0, var_5_node);
+
+ __visc__edge(var_4, var_5, 1, 0, 0, 0);
+ __visc__edge(var_4, var_5, 1, 1, 1, 0);
+ __visc__bindIn(var_5, 8, 2, 0);
+ __visc__bindIn(var_5, 9, 3, 0);
+
+ void* var_6 = __visc__createNodeND(0, var_6_node);
+
+ __visc__edge(var_5, var_6, 1, 0, 0, 0);
+ __visc__edge(var_5, var_6, 1, 1, 1, 0);
+
+ void* var_7 = __visc__createNodeND(0, var_7_node);
+
+ __visc__edge(var_6, var_7, 1, 0, 0, 0);
+ __visc__edge(var_6, var_7, 1, 1, 1, 0);
+ __visc__bindIn(var_7, 10, 2, 0);
+ __visc__bindIn(var_7, 11, 3, 0);
+ void* var_8 = __visc__createNodeND(0, var_8_node);
+
+ __visc__edge(var_7, var_8, 1, 0, 0, 0);
+ __visc__edge(var_7, var_8, 1, 1, 1, 0);
+ __visc__bindIn(var_8, 12, 2, 0);
+ __visc__bindIn(var_8, 13, 3, 0);
+
+ void* var_9 = __visc__createNodeND(0, var_9_node);
+
+ __visc__edge(var_8, var_9, 1, 0, 0, 0);
+ __visc__edge(var_8, var_9, 1, 1, 1, 0);
+
+ void* var_10 = __visc__createNodeND(0, var_10_node);
+
+ __visc__edge(var_9, var_10, 1, 0, 0, 0);
+ __visc__edge(var_9, var_10, 1, 1, 1, 0);
+ __visc__bindIn(var_10, 14, 2, 0);
+ __visc__bindIn(var_10, 15, 3, 0);
+
+ void* var_11 = __visc__createNodeND(0, var_11_node);
+
+ __visc__edge(var_10, var_11, 1, 0, 0, 0);
+ __visc__edge(var_10, var_11, 1, 1, 1, 0);
+ __visc__bindIn(var_11, 16, 2, 0);
+ __visc__bindIn(var_11, 17, 3, 0);
+
+ void* var_12 = __visc__createNodeND(0, var_12_node);
+
+ __visc__edge(var_11, var_12, 1, 0, 0, 0);
+ __visc__edge(var_11, var_12, 1, 1, 1, 0);
+
+ void* var_13 = __visc__createNodeND(0, var_13_node);
+
+ __visc__edge(var_12, var_13, 1, 0, 0, 0);
+ __visc__edge(var_12, var_13, 1, 1, 1, 0);
+ __visc__bindIn(var_13, 18, 2, 0);
+ __visc__bindIn(var_13, 19, 3, 0);
+
+ void* var_14 = __visc__createNodeND(0, var_14_node);
+
+ __visc__edge(var_13, var_14, 1, 0, 0, 0);
+ __visc__edge(var_13, var_14, 1, 1, 1, 0);
+ __visc__bindIn(var_14, 20, 2, 0);
+ __visc__bindIn(var_14, 21, 3, 0);
+
+ void* var_15 = __visc__createNodeND(0, var_15_node);
+
+ __visc__edge(var_14, var_15, 1, 0, 0, 0);
+ __visc__edge(var_14, var_15, 1, 1, 1, 0);
+
+ void* var_16 = __visc__createNodeND(0, var_16_node);
+
+ __visc__edge(var_15, var_16, 1, 0, 0, 0);
+ __visc__edge(var_15, var_16, 1, 1, 1, 0);
+
+ __visc__bindOut(var_16, 0, 0, 0);
+ __visc__bindOut(var_16, 1, 1, 0);
+
+}
-// Return type for the nodes
struct ret_t {
- void *tensor;
- size_t bytes;
-};
+ void* tensor;
+ size_t bytes;
+};
typedef struct __attribute__((__packed__)) {
- void *x;
- size_t x_bytes;
- // 1st Layer parameters
- void* conv1_w;
- size_t conv1_w_bytes;
- void* conv1_b;
- size_t conv1_b_bytes;
- // 2nd Layer parameters
- void* conv2_w;
- size_t conv2_w_bytes;
- void* conv2_b;
- size_t conv2_b_bytes;
- // 3rd Layer parameters
- void* fc1_w;
- size_t fc1_w_bytes;
- void* fc1_b;
- size_t fc1_b_bytes;
- // 4th Layer parameters
- void* fc2_w;
- size_t fc2_w_bytes;
- void* fc2_b;
- size_t fc2_b_bytes;
-
- struct ret_t r;
+ void* input;
+ size_t input_bytes;
+ void* conv2d_1_w;
+ size_t conv2d_1_w_bytes;
+ void* conv2d_1_b;
+ size_t conv2d_1_b_bytes;
+ void* conv2d_2_w;
+ size_t conv2d_2_w_bytes;
+ void* conv2d_2_b;
+ size_t conv2d_2_b_bytes;
+ void* conv2d_3_w;
+ size_t conv2d_3_w_bytes;
+ void* conv2d_3_b;
+ size_t conv2d_3_b_bytes;
+ void* dense_1_w;
+ size_t dense_1_w_bytes;
+ void* dense_1_b;
+ size_t dense_1_b_bytes;
+ void* dense_2_w;
+ size_t dense_2_w_bytes;
+ void* dense_2_b;
+ size_t dense_2_b_bytes;
+
+ struct ret_t r;
}
RootIn;
-int main() {
-
- int test_batch_size = 10000;
- std::string prefix = "../../../../../../projects/hpvm-tensor-rt/model_params";
- std::string input_data_path = prefix + std::string("/FC_network2/mnist_float_input.bin");
- std::string labels_path = "../../../../../../projects/hpvm-tensor-rt/model_params/lenet_params/datasets/t10k-labels-idx1-ubyte";
- std::string conv1_w_path = prefix + std::string("/lenet_keras/conv1.bin");
- std::string conv1_b_path = prefix + std::string("/lenet_keras/conv1_bias.bin");
- std::string conv2_w_path = prefix + std::string("/lenet_keras/conv2.bin");
- std::string conv2_b_path = prefix + std::string("/lenet_keras/conv2_bias.bin");
- std::string fc1_w_path = prefix + std::string("/lenet_keras/fc1.bin");
- std::string fc1_b_path = prefix + std::string("/lenet_keras/fc1_bias.bin");
- std::string fc2_w_path = prefix + std::string("/lenet_keras/fc2.bin");
- std::string fc2_b_path = prefix + std::string("/lenet_keras/fc2_bias.bin");
-
-
- printf("Reading Input Data from = %s \n", input_data_path.c_str());
-
- uint8_t* labels = readLabels(labels_path.c_str(), test_batch_size);
- void* x = readTrainedWeights(input_data_path.c_str(), float_type,
- test_batch_size, 1, 28, 28);
- void* conv1_w = readTrainedWeights(conv1_w_path.c_str(), float_type, 32, 1, 5, 5);
- void* conv1_b = readTrainedWeights(conv1_b_path.c_str(), float_type, 1, 32, 1, 1);
- void* conv2_w = readTrainedWeights(conv2_w_path.c_str(), float_type, 64, 32, 5, 5);
- void* conv2_b = readTrainedWeights(conv2_b_path.c_str(), float_type, 1, 64, 1, 1);
- void* fc1_w = readTrainedWeights(fc1_w_path.c_str(), float_type, 1, 1, 7*7*64, 1024);
- void* fc1_b = readTrainedWeights(fc1_b_path.c_str(), float_type, 1, 1024, 1, 1);
- void* fc2_w = readTrainedWeights(fc2_w_path.c_str(), float_type, 1, 1, 1024, 10);
- void* fc2_b = readTrainedWeights(fc2_b_path.c_str(), float_type, 1, 10, 1, 1);
-
- __visc__init();
-
- RootIn* args = static_cast<RootIn*>(malloc(sizeof(RootIn)));
- args->x = x;
- args->x_bytes = 0;
- // Conv Layers params
- args->conv1_w = conv1_w;
- args->conv1_w_bytes = 0;
- args->conv1_b = conv1_b;
- args->conv1_b_bytes = 0;
- args->conv2_w = conv2_w;
- args->conv2_w_bytes = 0;
- args->conv2_b = conv2_b;
- args->conv2_b_bytes = 0;
- // FC Layers params
- args->fc1_w = fc1_w;
- args->fc1_w_bytes = 0;
- args->fc1_b = fc1_b;
- args->fc1_b_bytes = 0;
- args->fc2_w = fc2_w;
- args->fc2_w_bytes = 0;
- args->fc2_b = fc2_b;
- args->fc2_b_bytes = 0;
-
- void *dfg = __visc__launch(0, root, (void *)args);
-
- __visc__wait(dfg);
-
- // FIXME: Value returned in the wrong index!!
- //void *r = static_cast<RootIn*>(args)->r.tensor;
- void *result = static_cast<RootIn*>(args)->x;
- hpvm_request_tensor(result, 0);
-
- __visc__cleanup();
-
- computeAccuracy2(labels, test_batch_size, result);
-
- return 0;
-}
-
-
+int main(){
+
+std::string dir_prefix = std::string("../../../../../../projects/hpvm-tensor-rt/model_params/lenet_hpvm/");
+std::string input_path = dir_prefix + std::string("input.bin");
+void* input = readTrainedWeights(input_path.c_str(), 0,10000,1,28,28);
+std::string labels_path = dir_prefix + std::string("labels.bin");
+uint8_t* labels = readLabels(labels_path.c_str(),10000);
+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);
+std::string conv2d_1_b_path = dir_prefix + std::string("conv2d_1_b.bin");
+void* conv2d_1_b = readTrainedWeights(conv2d_1_b_path.c_str(), 0,1,32,1,1);
+std::string conv2d_2_w_path = dir_prefix + std::string("conv2d_2_w.bin");
+void* conv2d_2_w = readTrainedWeights(conv2d_2_w_path.c_str(), 0,64,32,5,5);
+std::string conv2d_2_b_path = dir_prefix + std::string("conv2d_2_b.bin");
+void* conv2d_2_b = readTrainedWeights(conv2d_2_b_path.c_str(), 0,1,64,1,1);
+std::string conv2d_3_w_path = dir_prefix + std::string("conv2d_3_w.bin");
+void* conv2d_3_w = readTrainedWeights(conv2d_3_w_path.c_str(), 0,64,64,3,3);
+std::string conv2d_3_b_path = dir_prefix + std::string("conv2d_3_b.bin");
+void* conv2d_3_b = readTrainedWeights(conv2d_3_b_path.c_str(), 0,1,64,1,1);
+std::string dense_1_w_path = dir_prefix + std::string("dense_1_w.bin");
+void* dense_1_w = readTrainedWeights(dense_1_w_path.c_str(), 0,1,1,3136,1024);
+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(), 0,1,1024,1,1);
+std::string dense_2_w_path = dir_prefix + std::string("dense_2_w.bin");
+void* dense_2_w = 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);
+
+__visc__init();
+RootIn* args = static_cast<RootIn*>(malloc(sizeof(RootIn)));
+
+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;
+args->conv2d_1_b_bytes = 0;
+args->conv2d_2_w = conv2d_2_w;
+args->conv2d_2_w_bytes = 0;
+args->conv2d_2_b = conv2d_2_b;
+args->conv2d_2_b_bytes = 0;
+args->conv2d_3_w = conv2d_3_w;
+args->conv2d_3_w_bytes = 0;
+args->conv2d_3_b = conv2d_3_b;
+args->conv2d_3_b_bytes = 0;
+args->dense_1_w = dense_1_w;
+args->dense_1_w_bytes = 0;
+args->dense_1_b = dense_1_b;
+args->dense_1_b_bytes = 0;
+args->dense_2_w = dense_2_w;
+args->dense_2_w_bytes = 0;
+args->dense_2_b = dense_2_b;
+args->dense_2_b_bytes = 0;
+
+void* dfg = __visc__launch(0, root, (void*) args);
+
+__visc__wait(dfg);
+
+void *result = static_cast<RootIn*>(args)->input;
+hpvm_request_tensor(result, 0);
+
+__visc__cleanup();
+ computeAccuracy2(labels, 10000, result);
+return 0;
+
+}
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet_front/src/lenet_promise.cpp b/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet/src/lenet_promise.cpp
similarity index 100%
rename from llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet_front/src/lenet_promise.cpp
rename to llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet/src/lenet_promise.cpp
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet_front/Makefile b/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet_front/Makefile
deleted file mode 100644
index aeb8cd6bb1606c496258fcd92f28e1b3d0f10566..0000000000000000000000000000000000000000
--- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet_front/Makefile
+++ /dev/null
@@ -1,63 +0,0 @@
-DNN_BENCHMARK_ROOT = $(LLVM_SRC_ROOT)/test/VISC/DNN_Benchmarks
-# NOTE: can configure build directory
-HPVM_BUILD_DIR = $(LLVM_SRC_ROOT)/../build_hpvm/
-
-CC = $(HPVM_BUILD_DIR)/bin/clang++
-OPT = $(HPVM_BUILD_DIR)/bin/opt
-LLVM_DIS = $(HPVM_BUILD_DIR)/bin/llvm-dis
-LLVM_LINK = $(HPVM_BUILD_DIR)/bin/llvm-link
-LLVM_INCLUDE_DIR = $(LLVM_SRC_ROOT)/include
-
-SRC_DIR = src
-BUILD_DIR = build
-APP = lenet
-
-TENSOR_INCLUDE_DIR = $(DNN_BENCHMARK_ROOT)/common/include
-TENSOR_RT_INCLUDE_DIR = $(LLVM_SRC_ROOT)/projects/hpvm-tensor-rt/tensor_runtime/include
-TENSOR_LIB_DIR = $(LLVM_SRC_ROOT)/projects/hpvm-tensor-rt/lib/libtensor_runtime.a
-TENSOR_AUTOTUNER_DIR = $(LLVM_SRC_ROOT)/projects/hpvm-tensor-rt/lib/libtensor_autotuner.a
-
-CC_FLAGS = -I $(LLVM_INCLUDE_DIR) -I $(TENSOR_INCLUDE_DIR) -I $(TENSOR_RT_INCLUDE_DIR) -I $(CUDA_INCLUDE_PATH) -fno-exceptions -ffast-math -std=c++11 -O3
-CCFLAGS += -DDEVICE=CUDNN_TARGET
-LINKER_FLAGS = -lpthread -lcudart -lcurand -lcudnn -lcublas -lOpenCL
-
-HPVM_LIB_DIR = $(HPVM_BUILD_DIR)/lib
-
-
-VISC_OPTFLAGS = -load $(HPVM_LIB_DIR)/LLVMBuildDFG.so -load $(HPVM_LIB_DIR)/LLVMInPlaceDFGAnalysis.so -load $(HPVM_LIB_DIR)/LLVMDFG2LLVM_CUDNN.so -load $(HPVM_LIB_DIR)/LLVMDFG2LLVM_X86.so -load $(HPVM_LIB_DIR)/LLVMClearDFG.so -inplace -dfg2llvm-cudnn -dfg2llvm-x86 -clearDFG
-
-
-VISC_OPTFLAGS2 = -load $(HPVM_LIB_DIR)/LLVMBuildDFG.so -load $(HPVM_LIB_DIR)/LLVMInPlaceDFGAnalysis.so -load $(HPVM_LIB_DIR)/LLVMDFG2LLVM_PROMISE.so -load $(HPVM_LIB_DIR)/LLVMDFG2LLVM_X86.so -load $(HPVM_LIB_DIR)/LLVMFuseHPVMTensorNodes.so -load $(HPVM_LIB_DIR)/LLVMClearDFG.so -inplace -hpvm-fuse -dfg2llvm-promise -dfg2llvm-x86 -clearDFG
-
-
-
-TARGET = $(BUILD_DIR)/$(APP).opt.bc
-SOURCES = $(SRC_DIR)/$(APP).cpp
-VISC_RT_PATH = $(LLVM_SRC_ROOT)/../build/projects/visc-rt/visc-rt.ll
-
-#OBJS = $(BUILD_DIR)/$(wildcabrd *.ll)
-.PRECIOUS: $(BUILD_DIR)/$(APP).ll $(BUILD_DIR)/$(APP).visc.ll
-default: $(BUILD_DIR) $(TARGET)
-
-
-$(BUILD_DIR)/%.ll: $(SRC_DIR)/%.cpp
- $(CC) $(CC_FLAGS) -emit-llvm src/$(APP).cpp -S -o $(BUILD_DIR)/$(APP).ll
- $(CC) $(CC_FLAGS) -emit-llvm src/$(APP)_promise.cpp -S -o $(BUILD_DIR)/$(APP)_promise.ll
-
-
-$(BUILD_DIR)/%.opt.bc: $(BUILD_DIR)/%.ll
- $(OPT) -load LLVMGenVISC.so -genvisc -globaldce $(BUILD_DIR)/$(APP).ll -S -o $(BUILD_DIR)/$(APP).visc.ll
- $(OPT) -load LLVMGenVISC.so -genvisc -globaldce $(BUILD_DIR)/$(APP)_promise.ll -S -o $(BUILD_DIR)/$(APP)_promise.visc.ll
- $(OPT) $(VISC_OPTFLAGS) $(BUILD_DIR)/$(APP).visc.ll -o $(BUILD_DIR)/$(APP)_cudnn.bc
- $(OPT) $(VISC_OPTFLAGS2) $(BUILD_DIR)/$(APP)_promise.visc.ll -o $(BUILD_DIR)/$(APP)_promise.bc
- $(LLVM_LINK) $(BUILD_DIR)/$(APP)_cudnn.bc $(VISC_RT_PATH) -o $(BUILD_DIR)/$(APP)_cudnn_linked.bc
- $(LLVM_LINK) $(BUILD_DIR)/$(APP)_promise.bc $(VISC_RT_PATH) -o $(BUILD_DIR)/$(APP)_promise_linked.bc
- $(CC) $(BUILD_DIR)/$(BUILD_DIR)/$(APP)_cudnn_linked.bc $(TENSOR_LIB_DIR) -o $(BUILD_DIR)/$(APP)_cudnn_linked $(LINKER_FLAGS)
- $(CC) $(BUILD_DIR)/$(BUILD_DIR)/$(APP)_promise_linked.bc $(TENSOR_LIB_DIR) -o $(BUILD_DIR)/$(APP)_promise_linked $(LINKER_FLAGS)
- $(CC) $(BUILD_DIR)/$(BUILD_DIR)/$(APP)_cudnn_linked.bc $(TENSOR_AUTOTUNER_DIR) -o $(BUILD_DIR)/lenet_tune $(LINKER_FLAGS)
-
-$(BUILD_DIR):
- mkdir -p $@
-
-clean:
- rm -rf $(BUILD_DIR)
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet_front/src/lenet.cpp b/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet_front/src/lenet.cpp
deleted file mode 100644
index 1746fc13dc4809f8c3d806fa144903fac50f3315..0000000000000000000000000000000000000000
--- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/lenet_front/src/lenet.cpp
+++ /dev/null
@@ -1,373 +0,0 @@
-
-#include <stdio.h>
-#include <stdlib.h>
-#include <unistd.h>
-#include <fcntl.h>
-#include <sys/stat.h>
-#include <cstring>
-#include <visc.h>
-#include <tensorTypes.h>
-#include <tensorUtils.h>
-
-void var_0_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_convolution(t1, t2, 2, 2, 1, 1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_1_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_2_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_relu(t1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_3_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_4_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_convolution(t1, t2, 2, 2, 1, 1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_5_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_6_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_relu(t1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_7_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_convolution(t1, t2, 1, 1, 2, 2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_8_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_9_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_relu(t1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_10_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_mul(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_11_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_12_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_relu(t1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_13_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_mul(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_14_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(2, t1, t2, 0);
-
- void *r = __visc__tensor_add(t1, t2);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_15_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_relu(t1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void var_16_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
- __visc__attributes(1, t1, 0);
-
- void* r = __visc__tensor_softmax(t1);
- __visc__return(2, r, (size_t) 0);
-}
-
-void root(void* input, size_t input_bytes,
- void* conv2d_1_w, size_t conv2d_1_w_bytes,
- void* conv2d_1_b, size_t conv2d_1_b_bytes,
- void* conv2d_2_w, size_t conv2d_2_w_bytes,
- void* conv2d_2_b, size_t conv2d_2_b_bytes,
- void* conv2d_3_w, size_t conv2d_3_w_bytes,
- void* conv2d_3_b, size_t conv2d_3_b_bytes,
- void* dense_1_w, size_t dense_1_w_bytes,
- void* dense_1_b, size_t dense_1_b_bytes,
- void* dense_2_w, size_t dense_2_w_bytes,
- void* dense_2_b, size_t dense_2_b_bytes){
-
-
- __visc__hint(visc::CPU_TARGET);
- __visc__attributes(11, input, conv2d_1_w, conv2d_1_b, conv2d_2_w, conv2d_2_b, conv2d_3_w, conv2d_3_b, dense_1_w, dense_1_b, dense_2_w, dense_2_b, 0);
-
-
- void* var_0 = __visc__createNodeND(0, var_0_node);
-
- __visc__bindIn(var_0, 0, 0, 0);
- __visc__bindIn(var_0, 1, 1, 0);
- __visc__bindIn(var_0, 2, 2, 0);
- __visc__bindIn(var_0, 3, 3, 0);
-
- void* var_1 = __visc__createNodeND(0, var_1_node);
-
- __visc__edge(var_0, var_1, 1, 0, 0, 0);
- __visc__edge(var_0, var_1, 1, 1, 1, 0);
- __visc__bindIn(var_1, 4, 2, 0);
- __visc__bindIn(var_1, 5, 3, 0);
-
- void* var_2 = __visc__createNodeND(0, var_2_node);
-
- __visc__edge(var_1, var_2, 1, 0, 0, 0);
- __visc__edge(var_1, var_2, 1, 1, 1, 0);
-
- void* var_3 = __visc__createNodeND(0, var_3_node);
-
- __visc__edge(var_2, var_3, 1, 0, 0, 0);
- __visc__edge(var_2, var_3, 1, 1, 1, 0);
-
- void* var_4 = __visc__createNodeND(0, var_4_node);
-
- __visc__edge(var_3, var_4, 1, 0, 0, 0);
- __visc__edge(var_3, var_4, 1, 1, 1, 0);
- __visc__bindIn(var_4, 6, 2, 0);
- __visc__bindIn(var_4, 7, 3, 0);
-
- void* var_5 = __visc__createNodeND(0, var_5_node);
-
- __visc__edge(var_4, var_5, 1, 0, 0, 0);
- __visc__edge(var_4, var_5, 1, 1, 1, 0);
- __visc__bindIn(var_5, 8, 2, 0);
- __visc__bindIn(var_5, 9, 3, 0);
-
- void* var_6 = __visc__createNodeND(0, var_6_node);
-
- __visc__edge(var_5, var_6, 1, 0, 0, 0);
- __visc__edge(var_5, var_6, 1, 1, 1, 0);
-
- void* var_7 = __visc__createNodeND(0, var_7_node);
-
- __visc__edge(var_6, var_7, 1, 0, 0, 0);
- __visc__edge(var_6, var_7, 1, 1, 1, 0);
- __visc__bindIn(var_7, 10, 2, 0);
- __visc__bindIn(var_7, 11, 3, 0);
-
- void* var_8 = __visc__createNodeND(0, var_8_node);
-
- __visc__edge(var_7, var_8, 1, 0, 0, 0);
- __visc__edge(var_7, var_8, 1, 1, 1, 0);
- __visc__bindIn(var_8, 12, 2, 0);
- __visc__bindIn(var_8, 13, 3, 0);
-
- void* var_9 = __visc__createNodeND(0, var_9_node);
-
- __visc__edge(var_8, var_9, 1, 0, 0, 0);
- __visc__edge(var_8, var_9, 1, 1, 1, 0);
-
- void* var_10 = __visc__createNodeND(0, var_10_node);
-
- __visc__edge(var_9, var_10, 1, 0, 0, 0);
- __visc__edge(var_9, var_10, 1, 1, 1, 0);
- __visc__bindIn(var_10, 14, 2, 0);
- __visc__bindIn(var_10, 15, 3, 0);
-
- void* var_11 = __visc__createNodeND(0, var_11_node);
-
- __visc__edge(var_10, var_11, 1, 0, 0, 0);
- __visc__edge(var_10, var_11, 1, 1, 1, 0);
- __visc__bindIn(var_11, 16, 2, 0);
- __visc__bindIn(var_11, 17, 3, 0);
-
- void* var_12 = __visc__createNodeND(0, var_12_node);
-
- __visc__edge(var_11, var_12, 1, 0, 0, 0);
- __visc__edge(var_11, var_12, 1, 1, 1, 0);
-
- void* var_13 = __visc__createNodeND(0, var_13_node);
-
- __visc__edge(var_12, var_13, 1, 0, 0, 0);
- __visc__edge(var_12, var_13, 1, 1, 1, 0);
- __visc__bindIn(var_13, 18, 2, 0);
- __visc__bindIn(var_13, 19, 3, 0);
-
- void* var_14 = __visc__createNodeND(0, var_14_node);
-
- __visc__edge(var_13, var_14, 1, 0, 0, 0);
- __visc__edge(var_13, var_14, 1, 1, 1, 0);
- __visc__bindIn(var_14, 20, 2, 0);
- __visc__bindIn(var_14, 21, 3, 0);
-
- void* var_15 = __visc__createNodeND(0, var_15_node);
-
- __visc__edge(var_14, var_15, 1, 0, 0, 0);
- __visc__edge(var_14, var_15, 1, 1, 1, 0);
-
- void* var_16 = __visc__createNodeND(0, var_16_node);
-
- __visc__edge(var_15, var_16, 1, 0, 0, 0);
- __visc__edge(var_15, var_16, 1, 1, 1, 0);
-
- __visc__bindOut(var_16, 0, 0, 0);
- __visc__bindOut(var_16, 1, 1, 0);
-
-}
-
-struct ret_t {
- void* tensor;
- size_t bytes;
-};
-
-typedef struct __attribute__((__packed__)) {
- void* input;
- size_t input_bytes;
- void* conv2d_1_w;
- size_t conv2d_1_w_bytes;
- void* conv2d_1_b;
- size_t conv2d_1_b_bytes;
- void* conv2d_2_w;
- size_t conv2d_2_w_bytes;
- void* conv2d_2_b;
- size_t conv2d_2_b_bytes;
- void* conv2d_3_w;
- size_t conv2d_3_w_bytes;
- void* conv2d_3_b;
- size_t conv2d_3_b_bytes;
- void* dense_1_w;
- size_t dense_1_w_bytes;
- void* dense_1_b;
- size_t dense_1_b_bytes;
- void* dense_2_w;
- size_t dense_2_w_bytes;
- void* dense_2_b;
- size_t dense_2_b_bytes;
-
- struct ret_t r;
-}
-RootIn;
-
-int main(){
-
-std::string dir_prefix = std::string("../../../../../../projects/hpvm-tensor-rt/model_params/lenet_hpvm/");
-std::string input_path = dir_prefix + std::string("input.bin");
-void* input = readTrainedWeights(input_path.c_str(), 0,10000,1,28,28);
-std::string labels_path = dir_prefix + std::string("labels.bin");
-uint8_t* labels = readLabels(labels_path.c_str(),10000);
-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);
-std::string conv2d_1_b_path = dir_prefix + std::string("conv2d_1_b.bin");
-void* conv2d_1_b = readTrainedWeights(conv2d_1_b_path.c_str(), 0,1,32,1,1);
-std::string conv2d_2_w_path = dir_prefix + std::string("conv2d_2_w.bin");
-void* conv2d_2_w = readTrainedWeights(conv2d_2_w_path.c_str(), 0,64,32,5,5);
-std::string conv2d_2_b_path = dir_prefix + std::string("conv2d_2_b.bin");
-void* conv2d_2_b = readTrainedWeights(conv2d_2_b_path.c_str(), 0,1,64,1,1);
-std::string conv2d_3_w_path = dir_prefix + std::string("conv2d_3_w.bin");
-void* conv2d_3_w = readTrainedWeights(conv2d_3_w_path.c_str(), 0,64,64,3,3);
-std::string conv2d_3_b_path = dir_prefix + std::string("conv2d_3_b.bin");
-void* conv2d_3_b = readTrainedWeights(conv2d_3_b_path.c_str(), 0,1,64,1,1);
-std::string dense_1_w_path = dir_prefix + std::string("dense_1_w.bin");
-void* dense_1_w = readTrainedWeights(dense_1_w_path.c_str(), 0,1,1,3136,1024);
-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(), 0,1,1024,1,1);
-std::string dense_2_w_path = dir_prefix + std::string("dense_2_w.bin");
-void* dense_2_w = 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);
-
-__visc__init();
-RootIn* args = static_cast<RootIn*>(malloc(sizeof(RootIn)));
-
-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;
-args->conv2d_1_b_bytes = 0;
-args->conv2d_2_w = conv2d_2_w;
-args->conv2d_2_w_bytes = 0;
-args->conv2d_2_b = conv2d_2_b;
-args->conv2d_2_b_bytes = 0;
-args->conv2d_3_w = conv2d_3_w;
-args->conv2d_3_w_bytes = 0;
-args->conv2d_3_b = conv2d_3_b;
-args->conv2d_3_b_bytes = 0;
-args->dense_1_w = dense_1_w;
-args->dense_1_w_bytes = 0;
-args->dense_1_b = dense_1_b;
-args->dense_1_b_bytes = 0;
-args->dense_2_w = dense_2_w;
-args->dense_2_w_bytes = 0;
-args->dense_2_b = dense_2_b;
-args->dense_2_b_bytes = 0;
-
-void* dfg = __visc__launch(0, root, (void*) args);
-
-__visc__wait(dfg);
-
-void *result = static_cast<RootIn*>(args)->input;
-hpvm_request_tensor(result, 0);
-
-__visc__cleanup();
- computeAccuracy2(labels, 10000, result);
-return 0;
-
-}
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/resnet18_front/Makefile b/llvm/test/VISC/DNN_Benchmarks/benchmarks/resnet18/Makefile
similarity index 100%
rename from llvm/test/VISC/DNN_Benchmarks/benchmarks/resnet18_front/Makefile
rename to llvm/test/VISC/DNN_Benchmarks/benchmarks/resnet18/Makefile
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/resnet18_front/src/resnet18.cpp b/llvm/test/VISC/DNN_Benchmarks/benchmarks/resnet18/src/resnet18.cpp
similarity index 100%
rename from llvm/test/VISC/DNN_Benchmarks/benchmarks/resnet18_front/src/resnet18.cpp
rename to llvm/test/VISC/DNN_Benchmarks/benchmarks/resnet18/src/resnet18.cpp
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/Makefile b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/Makefile
similarity index 100%
rename from llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/Makefile
rename to llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/Makefile
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/final_accuracy b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/final_accuracy
similarity index 100%
rename from llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/final_accuracy
rename to llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/final_accuracy
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/vgg16_cifar10.ll b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/vgg16_cifar10.ll
similarity index 100%
rename from llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/vgg16_cifar10.ll
rename to llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/vgg16_cifar10.ll
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/vgg16_cifar10.opt.bc b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/vgg16_cifar10.opt.bc
similarity index 100%
rename from llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/vgg16_cifar10.opt.bc
rename to llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/vgg16_cifar10.opt.bc
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/vgg16_cifar10.visc.ll b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/vgg16_cifar10.visc.ll
similarity index 100%
rename from llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/vgg16_cifar10.visc.ll
rename to llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/vgg16_cifar10.visc.ll
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/vgg16_cifar10_linked b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/vgg16_cifar10_linked
similarity index 100%
rename from llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/vgg16_cifar10_linked
rename to llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/vgg16_cifar10_linked
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/vgg16_cifar10_linked.bc b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/vgg16_cifar10_linked.bc
similarity index 100%
rename from llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/vgg16_cifar10_linked.bc
rename to llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/vgg16_cifar10_linked.bc
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/vgg16_cifar10_tune b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/vgg16_cifar10_tune
similarity index 100%
rename from llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/build/vgg16_cifar10_tune
rename to llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/build/vgg16_cifar10_tune
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/src/vgg16_cifar10.cpp b/llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/src/vgg16_cifar10.cpp
similarity index 100%
rename from llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10_front/src/vgg16_cifar10.cpp
rename to llvm/test/VISC/DNN_Benchmarks/benchmarks/vgg16_cifar10/src/vgg16_cifar10.cpp