[WIP] CUDA error checking/debugging

.clang-format

@@ -3,7 +3,7 @@ BasedOnStyle: Google
 ---
 Language:               Cpp
 Cpp11BracedListStyle:   true
-Standard:               Cpp11
+Standard:               c++11
 DerivePointerAlignment: false
 PointerAlignment:       Right
 ---

CMakeLists.txt

@@ -77,6 +77,7 @@ list(APPEND CMAKE_MODULE_PATH ${CMAKE_SOURCE_DIR}/cmake)
 include(cpplint)
 include(glog)
 include(googletest)
+include(gtest_utils)
 include(pybind11)
 include(cub)
 

cmake/gtest_utils.cmake

@@ -0,0 +1,64 @@
+enable_testing()
+
+function(k2_add_gtest)
+  cmake_parse_arguments(
+      ARGS                  # prefix of output variables
+      "EXCLUDE_FROM_ALL"    # list of names of the boolean arguments (only defined ones will be true)
+      ""                    # list of names of mono-valued arguments
+      "SOURCES;DEPENDS"     # list of names of multi-valued arguments (output variables are lists)
+      ${ARGN}               # arguments of the function to parse, here we take the all original ones
+  ) # remaining unparsed arguments can be found in ARGS_UNPARSED_ARGUMENTS
+
+  set(dependencies ${ARGS_DEPENDS})
+
+  list(LENGTH ARGS_UNPARSED_ARGUMENTS other_args_size)
+  if(other_args_size EQUAL 1)
+    set(name ${ARGS_UNPARSED_ARGUMENTS})
+  else()
+    message(FATAL_ERROR "too many or not enough unnamed arguments")
+  endif()
+
+  if(ARGS_SOURCES)
+    set(sources ${ARGS_SOURCES})
+  else()
+    message(FATAL_ERROR "you must specify at least one source for the test")
+  endif()
+
+  if(ARGS_EXCLUDE_FROM_ALL)
+    set(EXCLUDE_FROM_ALL "EXCLUDE_FROM_ALL")
+  endif()
+
+  string(REGEX REPLACE "^.*\\." "" tgt ${name})
+  add_executable(${tgt} ${sources} ${EXCLUDE_FROM_ALL})
+
+  target_link_libraries(${tgt} PUBLIC ${dependencies} gtest)
+  target_link_libraries(${tgt} PRIVATE gtest_main)
+
+  add_test(NAME ${name} COMMAND $<TARGET_FILE:${tgt}>)
+
+  set(${CMAKE_PROJECT_NAME}_all_tests "${${CMAKE_PROJECT_NAME}_all_tests};${tgt}" CACHE STRING "" FORCE)
+endfunction()
+
+# directly give the suite name source files,
+function(k2_add_gtests test_set_name)
+  cmake_parse_arguments(
+      ARGS                      # prefix of output variables
+      "EXCLUDE_FROM_ALL"        # list of names of the boolean arguments (only defined ones will be true)
+      ""                        # list of names of mono-valued arguments
+      "SOURCES;DEPENDS"         # each source share the same args in this api, and each's a sub-gtest
+      ${ARGN}                   # arguments of the function to parse, here we take the all original ones
+  ) # remaining unparsed arguments can be found in ARGS_UNPARSED_ARGUMENTS
+
+  list(LENGTH ARGS_UNPARSED_ARGUMENTS other_args_size)
+  if(NOT other_args_size EQUAL 0)
+    message(FATAL_ERROR "too many or not enough unnamed arguments")
+  endif()
+
+  foreach(source ${ARGS_SOURCES})
+    string(REGEX REPLACE "\\..*$" "" name ${source})
+    k2_add_gtest("${test_set_name}.${name}"
+        SOURCES ${source}
+        DEPENDS ${ARGS_DEPENDS}
+    )
+  endforeach()
+endfunction()

k2/csrc/CMakeLists.txt

@@ -1,3 +1,5 @@
+include_directories(${CMAKE_SOURCE_DIR})
+
 # please sort the source files alphabetically
 add_library(fsa
   arcsort.cc

k2/csrc/cuda/CMakeLists.txt

@@ -1,28 +1,13 @@
-add_library(context context.cu)
-target_include_directories(context PUBLIC ${CMAKE_SOURCE_DIR})
-target_link_libraries(context PUBLIC cub)
-target_link_libraries(context PUBLIC glog)
+set(CUDA_SOURCE
+    context.cu
+)
 
-function(k2_add_cuda_test name)
-  add_executable(${name} "${name}.cu")
-  target_link_libraries(${name}
-    PRIVATE
-      context
-      gtest
-      gtest_main
-  )
-  add_test(NAME "Test.Cuda.${name}"
-    COMMAND
-      $<TARGET_FILE:${name}>
-  )
-endfunction()
+add_library(k2_cuda ${CUDA_SOURCE})
+target_include_directories(k2_cuda PUBLIC ${CMAKE_SOURCE_DIR})
+target_link_libraries(k2_cuda PUBLIC cub)
+target_link_libraries(k2_cuda PUBLIC glog)
 
-# please sort the source files alphabetically
-set(cuda_tests
-  ops_test
-  utils_test
+k2_add_gtests(Test.Cuda
+    SOURCES debug_test.cu utils_test.cu #ops_test.cu
+    DEPENDS k2_cuda cub glog
 )
-
-foreach(name IN LISTS cuda_tests)
-  k2_add_cuda_test(${name})
-endforeach()

k2/csrc/cuda/arch.h

@@ -0,0 +1,32 @@
+// k2/csrc/cuda/arch.h
+
+// Copyright (c) 2020 Xiaomi Corporation ( authors: Meixu Song )
+
+// See ../../LICENSE for clarification regarding multiple authors
+#ifndef K2_CSRC_CUDA_ARCH_H_
+#define K2_CSRC_CUDA_ARCH_H_
+
+namespace k2 {
+
+/**
+ * @def K2_PTX_ARCH
+ *
+ * @brief
+ *  K2_PTX_ARCH get the target PTX version through the active compiler.
+ *
+ * @details
+ *  - for host, it's 0.
+ *  - for device, it's `__CUDA_ARCH__`, which indicates the
+ *                compute compatibility, should >= 200.
+ */
+#ifndef K2_PTX_ARCH
+  #ifndef __CUDA_ARCH__
+    #define K2_PTX_ARCH 0
+  #else
+    #define K2_PTX_ARCH __CUDA_ARCH__
+  #endif
+#endif
+
+}  // namespace k2
+
+#endif  // K2_CSRC_CUDA_ARCH_H_

k2/csrc/cuda/context.cu

@@ -9,6 +9,7 @@
 #include <cstdlib>
 
 #include "k2/csrc/cuda/context.h"
+#include "k2/csrc/cuda/debug.h"
 #include "k2/csrc/cuda/error.h"
 
 static constexpr size_t kAlignment = 64;