11.22

docs/deep_dive/arch.rst

@@ -18,9 +18,13 @@ This document is intended for developers who want to understand the architecture
   process logic for specific input data with strict mathematical
   description, guiding the source code.
 
+- The :ref:`operator-test` introduces the relevant content of 
+  operator testing.
+
 .. toctree::
   :maxdepth: 2
 
   system
   infer_flow
   math_formalization
+  operator_test

docs/deep_dive/operator_test.rst

@@ -0,0 +1,65 @@
+
+.. _operator-test:
+
+*************
+Operator Test
+*************
+
+.. contents::
+
+Test Method
+-----------
+
+First generate test data with test_ops.py, then verify the 
+generated data with test_op.cc, execute the following code 
+to test(test_op.cc is compiled with Makefile):
+
+.. code-block::
+
+    python tests/test_ops.py
+    ./test_op 0/1/2 (0 is cpu, 1 is formal, and 2 is gpu)
+
+Test Results
+------------
+
+The following are the test results of the operator on different devices:
+
++-------------------+-------------------------+
+|     operator      |          results        |
++===================+=========================+
+| broadcast_sub     | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| broadcast_add     | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| broadcast_mul     | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| broadcast_max     | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| broadcast_div     | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| broadcast_greater | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| max_pool2d        | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| dense             | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| sum               | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| max               | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| slice_like        | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| tile              | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| repeat            | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| concatenate       | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| transpose         | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| take              | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| get_valid_counts  | cpu & formal & gpu pass |
++-------------------+-------------------------+
+| strided_slice     | cpu & formal & gpu pass |
++-------------------+-------------------------+

tests/Makefile

@@ -10,8 +10,8 @@ test_model_formal: test_model.cc
 test_model_gpu: test_model.cc
 	g++ -o test_model_gpu test_model.cc -I../include -L../build/cpu/ -L../build/gpu/ -lcvm_runtime_cuda  -lcudart -lcuda --std=c++11 -pthread -fopenmp -I./ -ldl -g -DUSE_GPU=1 -DCVM_PROFILING -fsigned-char
 
-test_op:
-	g++ -o test_op test_op.cc -I../include -L../build/gpu -lcvm_runtime_cuda -lcuda -lcudart --std=c++11 -pthread -fopenmp -I../ -ldl  -DCVM_PROFILING -fsigned-char -DUSE_GPU
+test_op : test_op.cc
+	g++ -o test_op test_op.cc -I../include -L../build/ -lcvm_runtime --std=c++11 -pthread -fopenmp -I../ -ldl -g -fsigned-char
 
 clean:
 	rm -f test_model_cpu test_model_formal test_model_gpu test_op

tests/deprecated/test_ops.py

@@ -649,7 +649,7 @@ def verify_get_valid_counts():
     def get_valid_counts(data, score_threshold):
         np_data = np.array(data)
         dshp = np_data.shape
-        if len(dshp) != 3 or (dshp[2] <= 2):
+        if len(dshp) != 3 or (dshp[2] < 2):
             raise ValueError("data shape error: " + str(dshp))
         batch_size, num_anchor, elem_length = dshp
         np_out1 = np.zeros(shape=(batch_size,))

tests/test_op.cc

@@ -50,11 +50,11 @@ struct CVMOpParam {
   std::string attrs;
 };
 
-#ifndef DEVICE
-#define DEVICE 0
-#endif 
+// #ifndef DEVICE
+// #define DEVICE 1
+// #endif 
 
-int ctx = APIDevTypeMap.at(DEVICE);
+int ctx;
 
 int device_id = 0;
 /*
@@ -128,6 +128,9 @@ std::function<void()> get_func(
   arg_ptr->arg_tcodes.push_back(kHandle);
 
 
+  // cvm.runtime.{version}.{op_name}
+  // cvm.runtime.cpu.conv2d
+  // cvm.runtime.cuda.dense
   auto op = param.func_name;
   int device_type = static_cast<int>(ctx);
   std::string module_name = "cvm.runtime.";
@@ -472,7 +475,7 @@ void load_input(int num_inputs, string case_path, vector<vector<uint64_t>>& tsha
       CVMArrayFree(cpu_tensor);
     }
 }
-const string CASE_DIR = "/data/ops_generator";
+const string CASE_DIR = "/home/remloveh/cvm-runtime/data/ops_generator";
 //const string CASE_DIR = "/data/zkh";
 
 void test_op(string op_name) {
@@ -659,36 +662,38 @@ void test_op(string op_name) {
     // }
   }
 }
-int main() {
-  test_op("max_pool2d"); // formal & cpu pass
-  test_op("upsampling"); // formal & cpu pass
-  test_op("dense");  // formal & cpu pass
+int main(int argc, char* argv[]) {
+  ctx = APIDevTypeMap.at(stoi(argv[1]));
+  // test_op("max_pool2d"); // formal & cpu pass
+  //test_op("upsampling"); // formal & cpu pass
+  // test_op("dense");  // formal & cpu pass
   test_op("conv2d"); // formal & cpu pass
-  test_op("sum");  // formal & cpu pass
-  test_op("max");  // formal & cpu pass
-  test_op("slice_like"); // formal & cpu pass
-  test_op("tile"); // formal & cpu pass
-  test_op("repeat"); // formal & cpu pass
-  test_op("get_valid_counts");  // formal & CPU pass
-
-  test_op("strided_slice"); // formal & cpu pass
-  test_op("concatenate");// formal & cpu pass
-  test_op("transpose");// formal & cpu pass
-  test_op("take"); // formal & cpu pass
-  //test_op("clip"); // no test case
-  //test_op("cvm_clip"); // no test case
-  //test_op("cvm_right_shift");  // no test case
-  test_op("elemwise_add"); // formal & cpu pass
+  // test_op("sum");  // formal & cpu pass
+  // test_op("max");  // formal & cpu pass
+  // test_op("slice_like"); // formal & cpu pass
+  // test_op("tile"); // formal & cpu pass
+  // test_op("repeat"); // formal & cpu pass
+  // test_op("get_valid_counts");  // formal & CPU pass
+
+  // test_op("strided_slice"); // formal & cpu pass
+  // test_op("concatenate");// formal & cpu pass
+  // test_op("transpose");// formal & cpu pass
+  // test_op("take"); // formal & cpu pass
+  // test_op("clip"); // no test case
+  // test_op("cvm_clip"); // no test case
+  // test_op("cvm_right_shift");  // no test case
+  //test_op("elemwise_add"); // formal & cpu pass
   //test_op("elemwise_sub"); // no test case
   //test_op("where"); // no test case
-  test_op("non_max_suppression");  // formal & cpu pass
-  test_op("broadcast_sub"); // formal & cpu pass
-  test_op("broadcast_add");  // formal & cpu pass
-  test_op("broadcast_mul");  // formal & cpu pass
-  test_op("broadcast_max");  // formal & cpu pass
-  test_op("broadcast_div");  // formal & cpu pass
-  test_op("broadcast_greater");  // formal & cpu pass
+  //test_op("non_max_suppression");  // formal & cpu pass
+  // test_op("broadcast_sub"); // formal & cpu pass
+  // test_op("broadcast_add");  // formal & cpu pass
+  // test_op("broadcast_mul");  // formal & cpu pass
+  // test_op("broadcast_max");  // formal & cpu pass
+  // test_op("broadcast_div");  // formal & cpu pass
+  // test_op("broadcast_greater");  // formal & cpu pass
 
+  cout << "test device: "<< DeviceName(static_cast<int>(ctx)) << endl;
   cout << "all tests finished" << endl;
   return 0;
 }