arrayfire · pavanky · Apr 27, 2016 · Apr 26, 2016 · Apr 26, 2016 · Apr 26, 2016
diff --git a/src/backend/cpu/CMakeLists.txt b/src/backend/cpu/CMakeLists.txt
@@ -108,8 +108,7 @@ FILE(GLOB cpu_headers
     "*.h")
 
 FILE(GLOB cpu_sources
-    "*.cpp"
-    "kernel/sort_by_key/*.cpp")
+    "*.cpp")
 
 LIST(SORT cpu_headers)
 LIST(SORT cpu_sources)
@@ -162,13 +161,15 @@ ELSE(${UNIX}) #Windows
     SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
 ENDIF()
 
+INCLUDE("${CMAKE_CURRENT_SOURCE_DIR}/kernel/sort_by_key/CMakeLists.txt")
 
 IF(DEFINED BLAS_SYM_FILE)
   ADD_LIBRARY(afcpu_static STATIC
               ${cpu_headers}
               ${cpu_sources}
               ${backend_headers}
-              ${backend_sources})
+              ${backend_sources}
+              ${SORT_BY_KEY_OBJECTS})
 
   ADD_LIBRARY(afcpu SHARED
               ${c_headers}
@@ -192,14 +193,15 @@ IF(DEFINED BLAS_SYM_FILE)
 
 ELSE(DEFINED BLAS_SYM_FILE)
 
-  ADD_LIBRARY(afcpu SHARED
-              ${cpu_headers}
-              ${cpu_sources}
-              ${backend_headers}
-              ${backend_sources}
-              ${c_headers}
-              ${c_sources}
-              ${cpp_sources})
+ADD_LIBRARY(afcpu SHARED
+            ${cpu_headers}
+            ${cpu_sources}
+            ${backend_headers}
+            ${backend_sources}
+            ${c_headers}
+            ${c_sources}
+            ${cpp_sources}
+            ${SORT_BY_KEY_OBJECTS})
 
 ENDIF(DEFINED BLAS_SYM_FILE)
 

diff --git a/src/backend/cpu/kernel/sort_by_key/CMakeLists.txt b/src/backend/cpu/kernel/sort_by_key/CMakeLists.txt
@@ -0,0 +1,15 @@
+FILE(STRINGS "${CMAKE_CURRENT_SOURCE_DIR}/kernel/sort_by_key/sort_by_key_impl.cpp" FILESTRINGS)
+
+FOREACH(STR ${FILESTRINGS})
+    IF(${STR} MATCHES "// SBK_TYPES")
+        STRING(REPLACE "// SBK_TYPES:" "" TEMP ${STR})
+        STRING(REPLACE " " ";" SBK_TYPES ${TEMP})
+    ENDIF()
+ENDFOREACH()
+
+FOREACH(SBK_TYPE ${SBK_TYPES})
+    ADD_LIBRARY(cpu_sort_by_key_${SBK_TYPE} OBJECT
+        "${CMAKE_CURRENT_SOURCE_DIR}/kernel/sort_by_key/sort_by_key_impl.cpp")
+    SET_TARGET_PROPERTIES(cpu_sort_by_key_${SBK_TYPE} PROPERTIES COMPILE_FLAGS "-DTYPE=${SBK_TYPE}")
+    LIST(APPEND SORT_BY_KEY_OBJECTS $<TARGET_OBJECTS:cpu_sort_by_key_${SBK_TYPE}>)
+ENDFOREACH(SBK_TYPE ${SBK_TYPES})
diff --git a/src/backend/cpu/kernel/sort_by_key/f32.cpp b/src/backend/cpu/kernel/sort_by_key/f32.cpp
diff --git a/src/backend/cpu/kernel/sort_by_key/f64.cpp b/src/backend/cpu/kernel/sort_by_key/f64.cpp
diff --git a/src/backend/cpu/kernel/sort_by_key/s16.cpp b/src/backend/cpu/kernel/sort_by_key/s16.cpp
diff --git a/src/backend/cpu/kernel/sort_by_key/s32.cpp b/src/backend/cpu/kernel/sort_by_key/s32.cpp
diff --git a/src/backend/cpu/kernel/sort_by_key/s64.cpp b/src/backend/cpu/kernel/sort_by_key/s64.cpp
diff --git a/src/backend/cpu/kernel/sort_by_key/b8.cpp → ...u/kernel/sort_by_key/sort_by_key_impl.cpp b/src/backend/cpu/kernel/sort_by_key/b8.cpp → ...u/kernel/sort_by_key/sort_by_key_impl.cpp
@@ -9,11 +9,13 @@
 
 #include <kernel/sort_by_key_impl.hpp>
 
+// SBK_TYPES:float double int uint intl uintl short ushort char uchar
+
 namespace cpu
 {
 namespace kernel
 {
-    INSTANTIATE1(char,true)
-    INSTANTIATE1(char,false)
+    INSTANTIATE1(TYPE,true)
+    INSTANTIATE1(TYPE,false)
 }
 }
diff --git a/src/backend/cpu/kernel/sort_by_key/u16.cpp b/src/backend/cpu/kernel/sort_by_key/u16.cpp
diff --git a/src/backend/cpu/kernel/sort_by_key/u32.cpp b/src/backend/cpu/kernel/sort_by_key/u32.cpp
diff --git a/src/backend/cpu/kernel/sort_by_key/u64.cpp b/src/backend/cpu/kernel/sort_by_key/u64.cpp
diff --git a/src/backend/cpu/kernel/sort_by_key/u8.cpp b/src/backend/cpu/kernel/sort_by_key/u8.cpp
diff --git a/src/backend/cuda/CMakeLists.txt b/src/backend/cuda/CMakeLists.txt
@@ -158,7 +158,6 @@ FILE(GLOB cuda_headers
 FILE(GLOB cuda_sources
     "*.cu"
     "*.cpp"
-    "kernel/sort_by_key/*.cu"
     "kernel/*.cu")
 
 FILE(GLOB jit_sources
@@ -231,6 +230,8 @@ LIST(SORT cpp_sources)
 
 SOURCE_GROUP(api\\cpp\\Sources FILES ${cpp_sources})
 
+INCLUDE("${CMAKE_CURRENT_SOURCE_DIR}/kernel/sort_by_key/CMakeLists.txt")
+
 LIST(LENGTH COMPUTE_VERSIONS COMPUTE_COUNT)
 IF(${COMPUTE_COUNT} EQUAL 1)
     SET(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS} ${CUDA_GENERATE_CODE}")
@@ -361,6 +362,7 @@ MY_CUDA_ADD_LIBRARY(afcuda SHARED
                 ${c_headers}
                 ${c_sources}
                 ${cpp_sources}
+                ${sort_by_key_sources}
                 OPTIONS ${CUDA_GENERATE_CODE})
 
 ADD_DEPENDENCIES(afcuda ${ptx_targets})

diff --git a/src/backend/cuda/kernel/sort_by_key/CMakeLists.txt b/src/backend/cuda/kernel/sort_by_key/CMakeLists.txt
@@ -0,0 +1,29 @@
+FILE(STRINGS "${CMAKE_CURRENT_SOURCE_DIR}/kernel/sort_by_key/sort_by_key_impl.cu.in" FILESTRINGS)
+
+FOREACH(STR ${FILESTRINGS})
+    IF(${STR} MATCHES "// SBK_TYPES")
+        STRING(REPLACE "// SBK_TYPES:" "" TEMP ${STR})
+        STRING(REPLACE " " ";" SBK_TYPES ${TEMP})
+    ELSEIF(${STR} MATCHES "// SBK_DIRS:")
+        STRING(REPLACE "// SBK_DIRS:" "" TEMP ${STR})
+        STRING(REPLACE " " ";" SBK_DIRS ${TEMP})
+    ELSEIF(${STR} MATCHES "// SBK_INSTS:")
+        STRING(REPLACE "// SBK_INSTS:" "" TEMP ${STR})
+        STRING(REPLACE " " ";" SBK_INSTS ${TEMP})
+    ENDIF()
+ENDFOREACH()
+
+FOREACH(SBK_TYPE ${SBK_TYPES})
+    FOREACH(SBK_DIR ${SBK_DIRS})
+        FOREACH(SBK_INST ${SBK_INSTS})
+            CONFIGURE_FILE("${CMAKE_CURRENT_SOURCE_DIR}/kernel/sort_by_key/sort_by_key_impl.cu.in"
+            "${CMAKE_CURRENT_BINARY_DIR}/sort_by_key/sort_by_key_impl_${SBK_TYPE}_${SBK_DIR}_${SBK_INST}.cu")
+        ENDFOREACH(SBK_INST ${SBK_INSTS})
+    ENDFOREACH(SBK_DIR ${SBK_DIRS})
+ENDFOREACH(SBK_TYPE ${SBK_TYPES})
+
+FILE(GLOB sort_by_key_sources
+  "${CMAKE_CURRENT_BINARY_DIR}/sort_by_key/*.cu"
+)
+
+LIST(SORT sort_by_key_sources)
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_f64.cu b/src/backend/cuda/kernel/sort_by_key/ascd_f64.cu
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_s16.cu b/src/backend/cuda/kernel/sort_by_key/ascd_s16.cu
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_s32.cu b/src/backend/cuda/kernel/sort_by_key/ascd_s32.cu
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_s64.cu b/src/backend/cuda/kernel/sort_by_key/ascd_s64.cu