diff --git a/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp b/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
index 7ea0c1dce23cf94385df3089c499338bec281b64..c54dd9ef3bbf6fbd8e75cbb4bc488cb215e580ce 100644
--- a/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
@@ -857,7 +857,7 @@ errs() << "TensorII: " << *TensorII << "\n";
Args.push_back(TensorII->getOperand(7));
// Create wrapper API runtime function call
- Constant* wrapper_tensorGroupConvolution;
+ Constant* wrapper_tensorGroupConvolution =
M->getOrInsertFunction(StringRef("wrapper_tensorGroupConvolution"),
RtM->getFunction(StringRef("wrapper_tensorGroupConvolution"))->getFunctionType());
CallInst* CI = CallInst::Create(wrapper_tensorGroupConvolution,
@@ -895,9 +895,9 @@ errs() << "TensorII: " << *TensorII << "\n";
Args.push_back(TensorII->getOperand(3));
Args.push_back(TensorII->getOperand(4));
Args.push_back(TensorII->getOperand(5));
-
+
// Create wrapper API runtime function call
- Constant* wrapper_tensorBatchNorm;
+ Constant* wrapper_tensorBatchNorm =
M->getOrInsertFunction(StringRef("wrapper_tensorBatchNorm"),
RtM->getFunction(StringRef("wrapper_tensorBatchNorm"))->getFunctionType());
CallInst* CI = CallInst::Create(wrapper_tensorBatchNorm,
diff --git a/llvm/projects/hpvm-tensor-rt/bin/install_runtime.sh b/llvm/projects/hpvm-tensor-rt/bin/install_runtime.sh
index cde03bd6d0ffa9969c785e17fe2f708c75396158..33a54cd0de626113e5cf11e2f6a6928d4fa384eb 100644
--- a/llvm/projects/hpvm-tensor-rt/bin/install_runtime.sh
+++ b/llvm/projects/hpvm-tensor-rt/bin/install_runtime.sh
@@ -3,11 +3,9 @@
export HPVM_TENSOR_RT_HOME=/home/hsharif3/Gitlab/hpvm/llvm/projects/hpvm-tensor-rt/
export PATH=/home/hsharif3/Gitlab/hpvm/build/bin/:$PATH
-clang++ -emit-llvm -c ${HPVM_TENSOR_RT_HOME}/tensor_runtime/include/tensor_signatures.cc -o ${HPVM_TENSOR_RT_HOME}/lib/tensor_runtime.bc
+clang++ -I/software/cuda-9.1/include -emit-llvm -c ${HPVM_TENSOR_RT_HOME}/tensor_runtime/include/tensor_signatures.cc -o ${HPVM_TENSOR_RT_HOME}/lib/tensor_runtime.bc
llvm-dis --version
llvm-dis ${HPVM_TENSOR_RT_HOME}/lib/tensor_runtime.bc
-cp ${HPVM_TENSOR_RT_HOME}/build/libtensor_runtime.a ${HPVM_TENSOR_RT_HOME}/lib/libtensor_runtime.a
-cp ${HPVM_TENSOR_RT_HOME}/build_autotuner/libtensor_runtime.a ${HPVM_TENSOR_RT_HOME}/lib/libtensor_autotuner.a
diff --git a/llvm/projects/hpvm-tensor-rt/build_pldi/table_generator.py b/llvm/projects/hpvm-tensor-rt/build_pldi/table_generator.py
deleted file mode 100644
index e3b94082f5be7b83a1598625afd5ef05a0472506..0000000000000000000000000000000000000000
--- a/llvm/projects/hpvm-tensor-rt/build_pldi/table_generator.py
+++ /dev/null
@@ -1,304 +0,0 @@
-import glob
-import os
-import subprocess
-import shutil
-import sys
-
-from collections import defaultdict
-
-'''
-FORMAT
-
-** LayerName NumOpsInLayer <cols>
-OpName Col1Val Col2Val ...
-
-** Conv1 1 h2f_time h2f_energy fp32_time fp32_energy f2h_time f2h_energy fp16_perf_time fp16_perf_energy fp16_time fp16_energy
-Conv1 51.8808 97.2844 319.582 601.966 12.81 18.758 388.092 650.649 340.037 590.664
-
-'''
-
-class TableGenerator:
-
- __ops_header_delimiter = "#"
- __table_header_delimter = "**"
- __time_col_name = "time"
- __energy_col_name = "energy"
-
- '''
- Stores all precision conversions used.
- '''
- precision_conversions = frozenset(["h2f", "f2h"])
-
- def __init__(self, dir_path, iters, profiler_binary_name):
- '''
- Args:
- dir_path: Path of directory containing network binaries
- iters: Number of iterations to run each binary for
- profiler_binary_name: Name of offline profiler binary to run
- '''
- self.__dir_path = dir_path
-
- # Name of the actual directory
- self.__network_name = os.path.split(dir_path)[-1]
-
- self.__iters = iters
- self.__profiler_binary_name = profiler_binary_name
-
- # Path to results directory
- self.__results_dir_path = "%s_results" % self.__dir_path
-
- # Outputted table file
- self.__table_filename = "%s_tensors.txt" % self.__network_name
-
- # Nested default dictionary of default dicts
- self.__table = self.__build_nested_default_dict()
-
-
- def generate_table(self):
- '''
- Generates a table file called <network_name>_tensors.txt in the following
- steps:
- 1. Runs the offline profiler against the inputted binaries to generate
- results files
- 2. Builds an internal table storing all data from the parsed results files
- the offline profiler generated
- 3. Writes the internal table to <network_name>_tensors.txt file and uses the
- <network_name>_ops.txt file as a guideline in terms of row order
- '''
- self.__run_inputted_binaries()
- self.__build_internal_table()
- self.__output_table_to_file()
-
-
- def __run_inputted_binaries(self):
- '''
- Invokes the profiler to run all appropriate binaries (must start with the network
- name) in the inputted directory. Result files generated by the profiler are
- stored in the results file directory and are named <binary_name>.txt. These results
- files are then parsed in a later step to generate the table
- '''
- if not os.path.isdir(self.__dir_path):
- print("ERROR: Directory %s not found" % self.__dir_path)
- exit(1)
-
- try:
- os.mkdir(self.__results_dir_path)
- except OSError:
- if os.path.isdir(self.__results_dir_path):
- print("Directory already exists. Clearing directory.")
- for old_file in glob.glob(os.path.join(self.__results_dir_path, "*")):
- os.remove(old_file)
- else:
- print("ERROR: Directory doesn't exist but failed to create dir")
-
- for binary_name in os.listdir(self.__dir_path):
- binary_path = os.path.join(self.__dir_path, binary_name)
-
- if not self.__should_execute_file(binary_path):
- continue
-
- output_file = os.path.join(self.__results_dir_path, binary_name + ".txt")
- # No stdout/stderr piping needed for now
- subprocess.Popen([profiler_binary_name, binary_path, str(self.__iters), \
- output_file]).communicate()
-
-
- def __build_internal_table(self):
- '''
- Iterates through each results file generated by the runs of the offline
- profiler and stores the data in a dictionary in the following format:
- [operation name][approximation type OR conversion type][time/energy]
- '''
- for results_file_name in os.listdir(self.__results_dir_path):
- # Ignore if it's not a results file
- if results_file_name == self.__table_filename or \
- not results_file_name.startswith(self.__network_name):
- continue
-
- approx_type = self.__get_approximation_type(results_file_name)
- results_file = open(os.path.join(self.__results_dir_path, results_file_name), "r")
-
- for line in results_file:
- line = line.strip()
- op_name, total_time, total_energy = self.__parse_tensor_operation_line(line)
-
- # If the current operation is f2h or h2f
- if any(op_name.endswith(prec_conv) for prec_conv in TableGenerator.precision_conversions):
- # Get the original operation name (without the f2h/h2f) and the conversion type
- orig_op_name, conversion_type = self.__get_original_operation_name(op_name)
-
- if orig_op_name not in self.__table:
- print("ERROR: Conversion found but original %s is not in the table" % orig_op_name)
- exit(1)
-
- # Store f2h and h2f as columns in the row belonging to the original operation
- self.__table[orig_op_name][conversion_type][TableGenerator.__time_col_name] = total_time
- self.__table[orig_op_name][conversion_type][TableGenerator.__energy_col_name] = total_energy
-
- # Create a new row in the dictionary
- else:
- self.__table[op_name][approx_type][TableGenerator.__time_col_name] = total_time
- self.__table[op_name][approx_type][TableGenerator.__energy_col_name] = total_energy
- results_file.close()
-
-
- def __output_table_to_file(self):
- '''
- Outputs the internally stored table to a file using the <network_name>_ops.txt file as
- a guideline in the following steps:
- 1. Opens the ops file and the file to output the table to
- 2. Reads a line from the ops file (guaranteed to be the layers/NML header)
- 3. For each operation in the layer (or 1 operation if the "layer" is a NML), we store the
- time and the energy
- '''
- table_file_path = os.path.join(self.__results_dir_path, self.__table_filename)
- soc_operations_file_name = os.path.join("/", "home", "nvidia", "soc_simulator", \
- "%s_cifar10" % self.__network_name, "%s_ops.txt" % self.__network_name)
-
- soc_operations_file = open(soc_operations_file_name, "r")
- table_file = open(table_file_path, "w")
-
- curr_line = soc_operations_file.readline().strip()
-
- while curr_line:
- # First line is always the layers line (#layer_name,num_ops)
- layer_name, num_ops = self.__parse_layer_info_line(curr_line)
-
- # List of strings, where each string is a row corresponding to an operation
- # in the layer
- ops_in_layer = []
-
- # Stores a list of elements in the header, which will be joined into a string
- # The header is only generated for the first operation in the layer
- # CRITICAL ASSUMPTION: All operations within a layer have the same # columns
- # or everything breaks bc the header is per layer, not per operation
- header = [TableGenerator.__table_header_delimter, layer_name, str(num_ops)]
-
- # Iterate through all operations within the layer
- for op_in_layer_count in range(num_ops):
- # Contains the operation name
- curr_line = soc_operations_file.readline().strip()
-
- # Stores a list of elements that will be joined to make up a row
- curr_op = [curr_line]
- operation_data = self.__table[curr_line]
-
- # Iterate through time/energy data for each approximation type corresponding
- # to the current operation
- for approx_type in operation_data:
- op_time = operation_data[approx_type][TableGenerator.__time_col_name]
- op_energy = operation_data[approx_type][TableGenerator.__energy_col_name]
-
- curr_op.append(op_time)
- curr_op.append(op_energy)
-
- if op_in_layer_count == 0:
- header.append("%s_time" % approx_type)
- header.append("%s_energy" % approx_type)
-
- ops_in_layer.append(' '.join(curr_op))
-
- # Getting all operation rows and then writing everything because
- # calls to write() are slow (memory vs time tradeoff)
- table_file.write("%s\n%s\n" % (' '.join(header), '\n'.join(ops_in_layer)))
-
- curr_line = soc_operations_file.readline().strip()
-
-
- def __should_execute_file(self, file_path):
- '''
- Checks if the file at the given file path is a binary that should be run
- by the profiler. Must exist, be a binary, and must start with the network
- name as per our naming standards.
-
- Args:
- file_path: Path of the file to check
- '''
- return os.path.isfile(file_path) and os.access(file_path, os.X_OK) and \
- file_path.find(self.__network_name) != -1
-
-
- def __get_approximation_type(self, results_filename):
- '''
- Parses a given results filename for the approximation type.
- Format assumption: <network_name>_<approx_type>.txt
-
- Args:
- results_filename: Name of results file
-
- Returns:
- the approximation technique (ex: fp16)
- '''
- approx_type_start_ind = results_filename.find(self.__network_name) \
- + len(self.__network_name) + 1 # + 1 to account for _ delimiter
- approx_type_end_ind = results_filename.find(".txt")
- return results_filename[approx_type_start_ind : approx_type_end_ind]
-
-
- def __parse_tensor_operation_line(self, tensor_op_line):
- '''
- Parses a tensor operation line (within a output file from the offline
- profiler for the operation name, the total time used, and the total
- energy used
-
- Args:
- tensor_op_line: Tensor operation line from output file
-
- Returns:
- operation name
- total time used
- total energy used
- '''
- line_as_list = tensor_op_line.split(",")
- return line_as_list[0], line_as_list[1], line_as_list[2]
-
-
- def __build_nested_default_dict(self):
- '''
- Builds a nested default dictionary with an arbitrary number of levels
- '''
- return defaultdict(self.__build_nested_default_dict)
-
- def __get_original_operation_name(self, op_name):
- '''
- Parses an operation name containing _<conversion type> for the original
- operation name.
- Format assumption: <original_op_name>_<conversion type>
-
- Args:
- op_name: Name of the operation
-
- Returns:
- the original operation name
- '''
- underscore_ind = op_name.find("_")
- return op_name[ : underscore_ind], op_name[underscore_ind + 1 : ]
-
-
- def __parse_layer_info_line(self, layer_info_line): #layer_name,num_ops
- '''
- Parses a layer header (from the original ops.txt file) into the layer name
- and the number of operations
- Assumed format: #layer_name,num_ops
-
- Args:
- layer_info_line: Line at the beginning of each layer in the ops file
-
- Returns:
- layer name
- number of ops in the layer
- '''
- comma_ind = layer_info_line.find(",")
- return layer_info_line[layer_info_line.find(TableGenerator.__ops_header_delimiter) + 1 : comma_ind], \
- int(layer_info_line[comma_ind + 1 : ])
-
-
-if __name__ == "__main__":
- if len(sys.argv) != 4:
- print("python table_generator.py <binary dir path> <num itrs> <profiler bin path>")
- exit(1)
- binary_dir_path = sys.argv[1]
- num_iters = int(sys.argv[2])
- profiler_binary_name = sys.argv[3]
- table_gen = TableGenerator(binary_dir_path, num_iters, profiler_binary_name)
- table_gen.generate_table()
diff --git a/llvm/projects/hpvm-tensor-rt/code_autogenerators/benchmark_testing_automator.py b/llvm/projects/hpvm-tensor-rt/code_autogenerators/benchmark_testing_automator.py
index 5f5c28032d721dcf1e77ab52407a165c0251deb2..f1f00f4e285fbf487fee03bfee72dbe1a84ea55a 100644
--- a/llvm/projects/hpvm-tensor-rt/code_autogenerators/benchmark_testing_automator.py
+++ b/llvm/projects/hpvm-tensor-rt/code_autogenerators/benchmark_testing_automator.py
@@ -27,17 +27,37 @@ def parse_binary_output(proc_output):
return avg_time
-# Input: a list of tuples of benchmark names
-# Can change to input a file containing benchmarks to run
-def run_benchmarks(builds_dir, output_filename, should_print_bin_output = True):
- output_file = open(output_filename, "w")
+def get_sorted_binaries(builds_dir):
+ # dict of network names to lists of binaries
+ # list of binaries should be in sorted order (can do that when we run the benchmarks)
+ network_bins = defaultdict(list)
for bin_name in os.listdir(builds_dir):
if bin_name.find("profiling") == -1:
continue
- output_file.write("%s: %s\n" % (bin_name, \
+ network_name = bin_name[ : bin_name.rfind("_")]
+ network_bins[network_name].append(bin_name)
+ return network_bins
+
+# Input: a list of tuples of benchmark names
+# Can change to input a file containing benchmarks to run
+def run_benchmarks(sorted_bins, builds_dir, output_filename, should_print_bin_output = False):
+ def get_knob_id(bin_name):
+ return int(bin_name[bin_name.rfind("_") + 1 : ])
+
+ output_file = open(output_filename, "w", buffering = 0)
+ for network_name in sorted_bins:
+ # Sort the binaries in order by knob id
+ sorted_bins[network_name].sort(key = get_knob_id)
+ print("--------------------------------------")
+ print(network_name)
+ # Go through all binaries
+ for bin_name in sorted_bins[network_name]:
+ print(bin_name)
+ output_file.write("%s results\n" % bin_name)
+ output_file.write("%s: %s\n" % (bin_name, \
parse_binary_output(run_benchmark(os.path.join(builds_dir, bin_name), \
should_print_bin_output))))
- print(bin_name)
+ print("--------------------------------------\n")
output_file.close()
@@ -48,4 +68,5 @@ if __name__ == "__main__":
print("Usage: python online_benchmark_testing_automator.py <builds dir> <outputs_file_name>")
exit(1)
print("Output file name: %s" % sys.argv[2])
- run_benchmarks(sys.argv[1], sys.argv[2])
+ sorted_bins = get_sorted_binaries(sys.argv[1])
+ run_benchmarks(sorted_bins, sys.argv[1], sys.argv[2])
diff --git a/llvm/projects/hpvm-tensor-rt/lib/tensor_runtime.ll b/llvm/projects/hpvm-tensor-rt/lib/tensor_runtime.ll
index 3e48a094b89ac506cf50f712a0d60b1bac95f75d..89c8da90f8ab740062bd84cdd365baa67311a7a4 100644
--- a/llvm/projects/hpvm-tensor-rt/lib/tensor_runtime.ll
+++ b/llvm/projects/hpvm-tensor-rt/lib/tensor_runtime.ll
@@ -8,8 +8,8 @@ define void @_Z13dummyFunctionv() #0 {
entry:
%initRT = alloca i8*, align 8
%cleanRT = alloca i8*, align 8
- %initApproxhpvmRT = alloca i8*, align 8
- %cleaApproxhpvmRT = alloca i8*, align 8
+ %initApproxRT = alloca i8*, align 8
+ %cleanApproxRT = alloca i8*, align 8
%initRTController = alloca i8*, align 8
%cleanRTController = alloca i8*, align 8
%request_tensorPtr = alloca i8*, align 8
@@ -44,17 +44,18 @@ entry:
%ConvLayer = alloca i8*, align 8
%FCLayer = alloca i8*, align 8
%ConvLayer2 = alloca i8*, align 8
+ %ConvLayer3 = alloca i8*, align 8
%FCLayer2 = alloca i8*, align 8
%AddWrapper = alloca i8*, align 8
%ReluWrapper = alloca i8*, align 8
%TanhWrapper = alloca i8*, align 8
%BatchNormWrapper = alloca i8*, align 8
%PoolingWrapper = alloca i8*, align 8
- %SoftmaxWrapper = alloca i8*, align 8
+ %softmaxWrapper = alloca i8*, align 8
store i8* bitcast (void (i32)* @llvm_hpvm_initTensorRt to i8*), i8** %initRT, align 8
store i8* bitcast (void ()* @llvm_hpvm_cleanupTensorRt to i8*), i8** %cleanRT, align 8
- store i8* bitcast (void (i32)* @llvm_hpvm_initApproxhpvmRt to i8*), i8** %initApproxhpvmRT, align 8
- store i8* bitcast (void ()* @llvm_hpvm_cleanupApproxhpvmRt to i8*), i8** %cleaApproxhpvmRT, align 8
+ store i8* bitcast (void (i32)* @llvm_hpvm_initApproxhpvmRt to i8*), i8** %initApproxRT, align 8
+ store i8* bitcast (void ()* @llvm_hpvm_cleanupApproxhpvmRt to i8*), i8** %cleanApproxRT, align 8
store i8* bitcast (void (i8*, i8*)* @llvm_hpvm_initializeRuntimeController to i8*), i8** %initRTController, align 8
store i8* bitcast (void ()* @llvm_hpvm_clearRuntimeController to i8*), i8** %cleanRTController, align 8
store i8* bitcast (void (i8*, i32)* @hpvm_request_tensor to i8*), i8** %request_tensorPtr, align 8
@@ -89,13 +90,14 @@ entry:
store i8* bitcast (i8* (i8*, float, float, i8*, float, float, i8*, float, float, i32, i32, i32, i32, i32, i32, i32, float, float, i32)* @ConvLayer_PROMISE to i8*), i8** %ConvLayer, align 8
store i8* bitcast (i8* (i8*, float, float, i8*, float, float, i8*, float, float, i32, float, float, i32)* @FCLayer_PROMISE to i8*), i8** %FCLayer, align 8
store i8* bitcast (i8* (i8*, i8*, i8*, i8*, i32, i32, i32, i32, i32, i32, i32, float, float)* @wrapper_ConvLayer to i8*), i8** %ConvLayer2, align 8
+ store i8* bitcast (i8* (i8*, i8*, i8*, i32, i32, i32, i32, i32, i32)* @wrapper_tensorGroupConvolution to i8*), i8** %ConvLayer3, align 8
store i8* bitcast (i8* (i8*, i8*, i8*, i8*, i32, float, float)* @wrapper_FCLayer to i8*), i8** %FCLayer2, align 8
store i8* bitcast (i8* (i8*, i8*, i8*)* @wrapper_tensorAdd to i8*), i8** %AddWrapper, align 8
store i8* bitcast (i8* (i8*, i8*)* @wrapper_tensorRelu to i8*), i8** %ReluWrapper, align 8
store i8* bitcast (i8* (i8*, i8*)* @wrapper_tensorTanh to i8*), i8** %TanhWrapper, align 8
store i8* bitcast (i8* (i8*, i8*, i8*, i8*, i8*, i8*, double)* @wrapper_tensorBatchNorm to i8*), i8** %BatchNormWrapper, align 8
store i8* bitcast (i8* (i8*, i8*, i32, i32, i32, i32, i32, i32, i32)* @wrapper_tensorPooling to i8*), i8** %PoolingWrapper, align 8
- store i8* bitcast (i8* (i8*, i8*)* @wrapper_tensorSoftmax to i8*), i8** %SoftmaxWrapper, align 8
+ store i8* bitcast (i8* (i8*, i8*)* @wrapper_tensorSoftmax to i8*), i8** %softmaxWrapper, align 8
ret void
}
@@ -175,6 +177,8 @@ declare i8* @FCLayer_PROMISE(i8*, float, float, i8*, float, float, i8*, float, f
declare i8* @wrapper_ConvLayer(i8*, i8*, i8*, i8*, i32, i32, i32, i32, i32, i32, i32, float, float) #1
+declare i8* @wrapper_tensorGroupConvolution(i8*, i8*, i8*, i32, i32, i32, i32, i32, i32) #1
+
declare i8* @wrapper_FCLayer(i8*, i8*, i8*, i8*, i32, float, float) #1
declare i8* @wrapper_tensorAdd(i8*, i8*, i8*) #1
diff --git a/llvm/projects/hpvm-tensor-rt/opentuner/data/alexnet/op_cost.txt b/llvm/projects/hpvm-tensor-rt/opentuner/data/alexnet/op_cost.txt
new file mode 100644
index 0000000000000000000000000000000000000000..04336fca2708d5e5d78849e1c12014f5ddbd1ad7
--- /dev/null
+++ b/llvm/projects/hpvm-tensor-rt/opentuner/data/alexnet/op_cost.txt
@@ -0,0 +1,6 @@
+11894784.000000
+39321600.000000
+21233664.000000
+28311552.000000
+18874368.000000
+20480.000000
diff --git a/llvm/projects/hpvm-tensor-rt/opentuner/data/alexnet2/op_cost.txt b/llvm/projects/hpvm-tensor-rt/opentuner/data/alexnet2/op_cost.txt
new file mode 100644
index 0000000000000000000000000000000000000000..5a5722f202dde469dca94c71dd9c5fc1cd7aa32b
--- /dev/null
+++ b/llvm/projects/hpvm-tensor-rt/opentuner/data/alexnet2/op_cost.txt
@@ -0,0 +1,7 @@
+88473.601562
+943718.375000
+471859.187500
+943718.375000
+471859.187500
+943718.375000
+2048.000000
diff --git a/llvm/projects/hpvm-tensor-rt/opentuner/data/lenet/op_cost.txt b/llvm/projects/hpvm-tensor-rt/opentuner/data/lenet/op_cost.txt
new file mode 100644
index 0000000000000000000000000000000000000000..74b1b668e2f27f3ddb77dcac7fff9890c70a6f02
--- /dev/null
+++ b/llvm/projects/hpvm-tensor-rt/opentuner/data/lenet/op_cost.txt
@@ -0,0 +1,4 @@
+62720.000000
+1003520.000000
+321126.406250
+1024.000000
diff --git a/llvm/projects/hpvm-tensor-rt/opentuner/data/mobilenet/op_cost.txt b/llvm/projects/hpvm-tensor-rt/opentuner/data/mobilenet/op_cost.txt
new file mode 100644
index 0000000000000000000000000000000000000000..673e704b7e37e19c090e98799189a4411bad9f7c
--- /dev/null
+++ b/llvm/projects/hpvm-tensor-rt/opentuner/data/mobilenet/op_cost.txt
@@ -0,0 +1,28 @@
+88473.601562
+29491.199219
+209715.203125
+14745.599609
+209715.203125
+29491.199219
+419430.406250
+7372.799805
+209715.203125
+14745.599609
+419430.406250
+3686.399902
+209715.203125
+7372.799805
+419430.406250
+7372.799805
+419430.406250
+7372.799805
+419430.406250
+7372.799805
+419430.406250
+7372.799805
+419430.406250
+1843.199951
+209715.203125
+3686.399902
+419430.406250
+1024.000000
diff --git a/llvm/projects/hpvm-tensor-rt/opentuner/data/mobilenet_shallow/op_cost.txt b/llvm/projects/hpvm-tensor-rt/opentuner/data/mobilenet_shallow/op_cost.txt
new file mode 100644
index 0000000000000000000000000000000000000000..7266441905a08c1ef1796dec8ee6c05660998378
--- /dev/null
+++ b/llvm/projects/hpvm-tensor-rt/opentuner/data/mobilenet_shallow/op_cost.txt
@@ -0,0 +1,8 @@
+265420.812500
+629145.625000
+629145.625000
+1258291.250000
+629145.625000
+1258291.250000
+629145.625000
+6144.000000
diff --git a/llvm/projects/hpvm-tensor-rt/opentuner/data/resnet/op_cost.txt b/llvm/projects/hpvm-tensor-rt/opentuner/data/resnet/op_cost.txt
new file mode 100644
index 0000000000000000000000000000000000000000..fdba070cfc5eac559c8384306993fb52a1eb2e04
--- /dev/null
+++ b/llvm/projects/hpvm-tensor-rt/opentuner/data/resnet/op_cost.txt
@@ -0,0 +1,22 @@
+44236.800781
+235929.593750
+235929.593750
+235929.593750
+235929.593750
+235929.593750
+235929.593750
+117964.796875
+235929.593750
+13107.200195
+235929.593750
+235929.593750
+235929.593750
+235929.593750
+117964.796875
+235929.593750
+13107.200195
+235929.593750
+235929.593750
+235929.593750
+235929.593750
+64.000000
diff --git a/llvm/projects/hpvm-tensor-rt/opentuner/data/vgg16_cifar10/op_cost.txt b/llvm/projects/hpvm-tensor-rt/opentuner/data/vgg16_cifar10/op_cost.txt
new file mode 100644
index 0000000000000000000000000000000000000000..5f58ebcc043915d28cf874a1f67e5b2637db1dfc
--- /dev/null
+++ b/llvm/projects/hpvm-tensor-rt/opentuner/data/vgg16_cifar10/op_cost.txt
@@ -0,0 +1,15 @@
+88473.601562
+1887436.750000
+943718.375000
+1887436.750000
+943718.375000
+1887436.750000
+1887436.750000
+943718.375000
+1887436.750000
+1887436.750000
+471859.187500
+471859.187500
+471859.187500
+13107.200195
+256.000000
diff --git a/llvm/projects/hpvm-tensor-rt/opentuner/data/vgg16_cifar100/op_cost.txt b/llvm/projects/hpvm-tensor-rt/opentuner/data/vgg16_cifar100/op_cost.txt
new file mode 100644
index 0000000000000000000000000000000000000000..8c6daad2e2902e3ac821d99ebbe12e21b6428cc7
--- /dev/null
+++ b/llvm/projects/hpvm-tensor-rt/opentuner/data/vgg16_cifar100/op_cost.txt
@@ -0,0 +1,15 @@
+884736.000000
+18874368.000000
+9437184.000000
+18874368.000000
+9437184.000000
+18874368.000000
+18874368.000000
+9437184.000000
+18874368.000000
+18874368.000000
+4718592.000000
+4718592.000000
+4718592.000000
+131072.000000
+25600.000000
diff --git a/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/configuration.h b/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/configuration.h
index b482cef5377e0f879b43f06a7ebbfbe01b39be09..14dc8f20f2111e85e82630cdbcc0c695a39c5ecd 100644
--- a/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/configuration.h
+++ b/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/configuration.h
@@ -72,7 +72,9 @@ public:
FP32,
FP16,
PERFORATION,
-// INPUT_SAMPLING,
+ INPUT_SAMPLING,
+ REDUCTION_SAMPLING,
+// ADDITIONAL_APPROXIMATION_METHOD
APPROX_END
};
@@ -91,6 +93,15 @@ public:
POOL_MEAN,
POOL_MIN,
SOFTMAX,
+ FFT,
+ REDUCE,
+ PROJECTIVE_T,
+ MAP1,
+ MAP2,
+ MAP3,
+// STENCIL,
+// COSINE_T,
+// ADDITIONAL_TENSOR_OPERATION
TENSOR_OP_END
};
@@ -269,6 +280,24 @@ void GPUNodeConfiguration::print() {
case TENSOR_OP::SOFTMAX :
DEBUG("softmax");
break;
+ case TENSOR_OP::FFT :
+ DEBUG("fft");
+ break;
+ case TENSOR_OP::REDUCE :
+ DEBUG("reduce");
+ break;
+ case TENSOR_OP::PROJECTIVE_T :
+ DEBUG("projectiveT");
+ break;
+ case TENSOR_OP::MAP1 :
+ DEBUG("map1");
+ break;
+ case TENSOR_OP::MAP2 :
+ DEBUG("map2");
+ break;
+ case TENSOR_OP::MAP3 :
+ DEBUG("map3");
+ break;
default :
ERROR("Unknown tensor operation.");
break;
@@ -288,6 +317,12 @@ void GPUNodeConfiguration::print() {
case APPROX::PERFORATION :
DEBUG("perf");
break;
+ case APPROX::INPUT_SAMPLING :
+ DEBUG("input_samp");
+ break;
+ case APPROX::REDUCTION_SAMPLING :
+ DEBUG("red_samp");
+ break;
default:
ERROR("Unknown approximation option");
break;
diff --git a/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/hpvm-rt-controller.h b/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/hpvm-rt-controller.h
index 911f42b955a72cb756aadc1fc78231187ef3394e..21c6df7f1749e891dba257bbb1933c3beefb8c4f 100644
--- a/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/hpvm-rt-controller.h
+++ b/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/hpvm-rt-controller.h
@@ -735,6 +735,30 @@ void RuntimeController::readConfigurationFile(const char *str) {
DEBUG ("Found softmax operation\n");
NodeConf->pushNewTensorOperation(GPUNodeConfiguration::TENSOR_OP::SOFTMAX);
idx++;
+ } else if (tokens[idx] == "fft") {
+ DEBUG ("Found fft operation\n");
+ NodeConf->pushNewTensorOperation(GPUNodeConfiguration::TENSOR_OP::FFT);
+ idx++;
+ } else if (tokens[idx] == "reduce") {
+ DEBUG ("Found reduce operation\n");
+ NodeConf->pushNewTensorOperation(GPUNodeConfiguration::TENSOR_OP::REDUCE);
+ idx++;
+ } else if (tokens[idx] == "projectiveT") {
+ DEBUG ("Found projectiveT operation\n");
+ NodeConf->pushNewTensorOperation(GPUNodeConfiguration::TENSOR_OP::PROJECTIVE_T);
+ idx++;
+ } else if (tokens[idx] == "map1") {
+ DEBUG ("Found map1 operation\n");
+ NodeConf->pushNewTensorOperation(GPUNodeConfiguration::TENSOR_OP::MAP1);
+ idx++;
+ } else if (tokens[idx] == "map2") {
+ DEBUG ("Found map2 operation\n");
+ NodeConf->pushNewTensorOperation(GPUNodeConfiguration::TENSOR_OP::MAP2);
+ idx++;
+ } else if (tokens[idx] == "map3") {
+ DEBUG ("Found map3 operation\n");
+ NodeConf->pushNewTensorOperation(GPUNodeConfiguration::TENSOR_OP::MAP3);
+ idx++;
} else /*Not a new operation. This means an approximation option*/
if (tokens[idx] == "fp32") {
DEBUG("Found fp32 option\n");
@@ -756,6 +780,18 @@ void RuntimeController::readConfigurationFile(const char *str) {
DEBUG("perf parameter: %d\n", perf);
NodeConf->pushNewApproximationChoiceForOperation(GPUNodeConfiguration::APPROX::PERFORATION, perf);
idx += 2;
+ } else if (tokens[idx] == "input_samp") {
+ DEBUG("Found input_samp option\n");
+ int input_samp = std::stoi(tokens[idx+1]);
+ DEBUG("input_samp parameter: %d\n", input_samp);
+ NodeConf->pushNewApproximationChoiceForOperation(GPUNodeConfiguration::APPROX::INPUT_SAMPLING, input_samp);
+ idx += 2;
+ } else if (tokens[idx] == "red_samp") {
+ DEBUG("Found red_samp option\n");
+ int red_samp = std::stoi(tokens[idx+1]);
+ DEBUG("red_samp parameter: %d\n", red_samp);
+ NodeConf->pushNewApproximationChoiceForOperation(GPUNodeConfiguration::APPROX::REDUCTION_SAMPLING, red_samp);
+ idx += 2;
}
// TODO: other approximation options handled here
diff --git a/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/tensor_utils.cu b/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/tensor_utils.cu
index 13ef262bac301e35d7c3e6306a3706dfe79a68a2..282a0cbb68de4f033b46cdc5c4a8ad69aa1f20c0 100644
--- a/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/tensor_utils.cu
+++ b/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/tensor_utils.cu
@@ -329,12 +329,18 @@ extern "C"{
void initTensorData(void* tensor_ptr, void* data_ptr, size_t size_in_bytes){
Tensor* tensor = (Tensor*) tensor_ptr;
-
- if(tensor->size_in_bytes != size_in_bytes){
+
+ size_t host_size_in_bytes = tensor->num_elems * 4;
+ //if(tensor->size_in_bytes != size_in_bytes){
+ if(host_size_in_bytes != size_in_bytes){
ERROR("The destination and source sizes don't match");
}
std::memcpy(tensor->host_data, data_ptr, size_in_bytes);
+
+ changeTensorPlacement(tensor, HOST);
+
+ tensor->cur_type = float_type;
}
@@ -428,7 +434,7 @@ extern "C"{
-bool ONLINE_PROFILING = false;
+bool ONLINE_PROFILING = false; // true;
void convertToFP16(struct Tensor* tensor){
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/Makefile b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/Makefile
index 4e762ea9894405bb375f518b65c209b4129d9f70..83b4dc9431ee84051def8a0f6850e7f2c194f033 100644
--- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/Makefile
+++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/Makefile
@@ -1,5 +1,6 @@
DNN_BENCHMARK_ROOT = $(LLVM_SRC_ROOT)/test/VISC/DNN_Benchmarks
# NOTE: can configure build directory
+#HPVM_BUILD_DIR = $(LLVM_SRC_ROOT)/../build_hpvm/
HPVM_BUILD_DIR = $(LLVM_BUILD_ROOT)
CC = $(HPVM_BUILD_DIR)/bin/clang++
@@ -15,9 +16,10 @@ 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
+PROFILER_LIB_DIR = $(LLVM_SRC_ROOT)/projects/gpu_profiler/lib/libgpu_profiler.a
+SOC_SIMULATOR_LIB_DIR = $(LLVM_SRC_ROOT)/projects/soc_simulator/lib/libpromise_profiler.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
@@ -58,15 +60,17 @@ $(BUILD_DIR)/%.opt.bc: $(BUILD_DIR)/%.ll
$(OPT) -load LLVMGenVISC.so -genvisc -globaldce $(BUILD_DIR)/$(APP)_promise.ll -S -o $(BUILD_DIR)/$(APP)_promise.visc.ll
$(OPT) -load LLVMGenVISC.so -genvisc -globaldce $(BUILD_DIR)/$(APP)_loop.ll -S -o $(BUILD_DIR)/$(APP)_loop.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
- $(OPT) $(VISC_OPTFLAGS) $(BUILD_DIR)/$(APP)_loop.visc.ll -o $(BUILD_DIR)/$(APP)_loop.bc
+ #$(OPT) $(VISC_OPTFLAGS2) $(BUILD_DIR)/$(APP)_promise.visc.ll -o $(BUILD_DIR)/$(APP)_promise.bc
+ $(OPT) $(VISC_OPTFLAGS3) $(BUILD_DIR)/$(APP)_promise.visc.ll -o $(BUILD_DIR)/$(APP)_wrapperapi.bc
+ $(OPT) $(VISC_OPTFLAGS3) $(BUILD_DIR)/$(APP)_loop.visc.ll -o $(BUILD_DIR)/$(APP)_loop_wrapperapi.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
- $(LLVM_LINK) $(BUILD_DIR)/$(APP)_loop.bc $(VISC_RT_PATH) -o $(BUILD_DIR)/$(APP)_loop_linked.bc
- $(CC) $(BUILD_DIR)/$(APP)_cudnn_linked.bc $(TENSOR_LIB_DIR) -o $(BUILD_DIR)/$(APP)_cudnn_linked $(LINKER_FLAGS)
- $(CC) $(BUILD_DIR)/$(APP)_promise_linked.bc $(TENSOR_LIB_DIR) -o $(BUILD_DIR)/$(APP)_promise_linked $(LINKER_FLAGS)
- $(CC) $(BUILD_DIR)/$(APP)_loop_linked.bc $(TENSOR_LIB_DIR) -o $(BUILD_DIR)/$(APP)_loop_linked $(LINKER_FLAGS)
- #$(CC) $(BUILD_DIR)/$(APP)_cudnn_linked.bc $(TENSOR_AUTOTUNER_DIR) -o $(BUILD_DIR)/lenet_tune $(LINKER_FLAGS)
+ #$(LLVM_LINK) $(BUILD_DIR)/$(APP)_promise.bc $(VISC_RT_PATH) -o $(BUILD_DIR)/$(APP)_promise_linked.bc
+ $(LLVM_LINK) $(BUILD_DIR)/$(APP)_wrapperapi.bc $(VISC_RT_PATH) -o $(BUILD_DIR)/$(APP)_wrapperapi_linked.bc
+ $(LLVM_LINK) $(BUILD_DIR)/$(APP)_loop_wrapperapi.bc $(VISC_RT_PATH) -o $(BUILD_DIR)/$(APP)_loop_wrapperapi_linked.bc
+ $(CC) $(BUILD_DIR)/$(APP)_cudnn_linked.bc $(TENSOR_LIB_DIR) $(PROFILER_LIB_DIR) $(SOC_SIMULATOR_LIB_DIR) -o $(BUILD_DIR)/$(APP)_cudnn_linked $(LINKER_FLAGS)
+ #$(CC) $(BUILD_DIR)/$(APP)_promise_linked.bc $(TENSOR_LIB_DIR) $(PROFILER_LIB_DIR) $(SOC_SIMULATOR_LIB_DIR) -o $(BUILD_DIR)/$(APP)_promise_linked $(LINKER_FLAGS)
+ $(CC) $(BUILD_DIR)/$(APP)_wrapperapi_linked.bc $(TENSOR_LIB_DIR) $(PROFILER_LIB_DIR) $(SOC_SIMULATOR_LIB_DIR) -o $(BUILD_DIR)/$(APP)_wrapperapi_linked $(LINKER_FLAGS)
+ $(CC) $(BUILD_DIR)/$(APP)_loop_wrapperapi_linked.bc $(TENSOR_LIB_DIR) $(PROFILER_LIB_DIR) $(SOC_SIMULATOR_LIB_DIR) -o $(BUILD_DIR)/$(APP)_loop_wrapperapi_linked $(LINKER_FLAGS)
$(BUILD_DIR):
mkdir -p $@
diff --git a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/src/alexnet_loop.cpp b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/src/alexnet_loop.cpp
index ee07bdd8f9901f1582d5f7642a2a86c099397a14..d92bc0c45d1115620d529aea4636ece8d3d62127 100644
--- a/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/src/alexnet_loop.cpp
+++ b/llvm/test/VISC/DNN_Benchmarks/benchmarks/alexnet/src/alexnet_loop.cpp
@@ -9,8 +9,10 @@
#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__hint(visc::PROMISE_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 5, 5, 1, 1);
@@ -18,7 +20,7 @@ void var_0_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
}
void var_1_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
@@ -26,7 +28,7 @@ void var_1_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
}
void var_2_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_tanh(t1);
@@ -34,7 +36,7 @@ void var_2_node(void* t1, size_t bytes_t1) {
}
void var_3_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
@@ -42,7 +44,7 @@ void var_3_node(void* t1, size_t bytes_t1) {
}
void var_4_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 2, 2, 1, 1);
@@ -50,7 +52,7 @@ void var_4_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
}
void var_5_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
@@ -58,7 +60,7 @@ void var_5_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
}
void var_6_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_tanh(t1);
@@ -66,7 +68,7 @@ void var_6_node(void* t1, size_t bytes_t1) {
}
void var_7_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
@@ -74,7 +76,7 @@ void var_7_node(void* t1, size_t bytes_t1) {
}
void var_8_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
@@ -82,7 +84,7 @@ void var_8_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
}
void var_9_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
@@ -90,7 +92,7 @@ void var_9_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
}
void var_10_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_tanh(t1);
@@ -98,7 +100,7 @@ void var_10_node(void* t1, size_t bytes_t1) {
}
void var_11_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
@@ -106,7 +108,7 @@ void var_11_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
}
void var_12_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
@@ -114,7 +116,7 @@ void var_12_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
}
void var_13_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_tanh(t1);
@@ -122,7 +124,7 @@ void var_13_node(void* t1, size_t bytes_t1) {
}
void var_14_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_convolution(t1, t2, 1, 1, 1, 1);
@@ -130,7 +132,7 @@ void var_14_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
}
void var_15_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
@@ -138,7 +140,7 @@ void var_15_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
}
void var_16_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_tanh(t1);
@@ -146,7 +148,7 @@ void var_16_node(void* t1, size_t bytes_t1) {
}
void var_17_node(void* t1, size_t bytes_t1) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(1, t1, 0);
void* r = __visc__tensor_pool_max(t1, 2, 2, 0, 0, 2, 2);
@@ -154,7 +156,7 @@ void var_17_node(void* t1, size_t bytes_t1) {
}
void var_18_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_mul(t1, t2);
@@ -162,7 +164,7 @@ void var_18_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
}
void var_19_node(void* t1, size_t bytes_t1, void* t2, size_t bytes_t2) {
- __visc__hint(visc::CUDNN_TARGET);
+ __visc__hint(visc::PROMISE_TARGET);
__visc__attributes(2, t1, t2, 0);
void *r = __visc__tensor_add(t1, t2);
@@ -177,6 +179,8 @@ void var_20_node(void* t1, size_t bytes_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,
@@ -371,9 +375,10 @@ int main(){
std::string dir_prefix = std::string("../../../../../../projects/hpvm-tensor-rt/model_params/alexnet_cifar10_test/");
-
+ std::string input_path = dir_prefix + std::string("input.bin");
+ //void* input = readTrainedWeights(input_path.c_str(), 0,5000,3,32,32);
std::string labels_path = dir_prefix + std::string("labels32.bin");
- //uint8_t* labels = readLabels(labels_path.c_str(),10000);
+ uint8_t* labels = readLabels(labels_path.c_str(),5000);
std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
void* conv2d_1_w = readTrainedWeights(conv2d_1_w_path.c_str(), 0,64,3,11,11);
std::string conv2d_1_b_path = dir_prefix + std::string("conv2d_1_b.bin");
@@ -404,6 +409,8 @@ int main(){
__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;
@@ -429,48 +436,38 @@ int main(){
args->dense_1_b = dense_1_b;
args->dense_1_b_bytes = 0;
- int batch_size = 500;
- int test_input_size = 10000;
- int batch_count = test_input_size / batch_size;
-
- std::string input_path = dir_prefix + std::string("input.bin");
+ int batch_size = 500;
+ int test_input_size = 10000;
+ int batch_count = test_input_size / batch_size;
+
void* input = create4DTensor(0,nchw,batch_size,3,32,32);
-
startMemTracking();
- for (int i = 0; i < batch_count; i++){
+ startProfiling();
- int start = i * batch_size;
- int end = (i + 1) * batch_size;
+ for (int i = 0; i < batch_count; i++){
+
+ int start = i * batch_size;
+ int end = (i + 1) * batch_size;
copyInputBatch(input_path.c_str(),start,end,3,32,32, input);
-
- args->input = input;
+
+ args->input = input;
args->input_bytes = 0;
-
- //void* input = readInputBatch(input_path.c_str(),0,start,end,3,32,32);
-
- void* dfg = __visc__launch(0, root, (void*) args);
+
+ void* dfg = __visc__launch(0, root, (void*) args);
__visc__wait(dfg);
+
+ void *result = static_cast<RootIn*>(args)->input;
+ hpvm_request_tensor(result, 0);
+
+ llvm_hpvm_invokeRtControl(result, labels_path.c_str(), start, end);
- void *result = static_cast<RootIn*>(args)->input;
- hpvm_request_tensor(result, 0);
-
-
- uint32_t* labels = readLabelsBatch3(labels_path.c_str(),start,end);
-
- computeAccuracy3(labels, result);
-
- llvm_hpvm_invokeRtControl2(result, labels);
-
freeBatchMemory();
}
-
-
- __visc__cleanup();
-
-
+ stopProfiling();
+ __visc__cleanup();
+
return 0;
-
-}
+}