From 60a9c8b33c6da9e64f7e5ef3178c8275895d519e Mon Sep 17 00:00:00 2001
From: Shehzan Mohammed <shehzan@arrayfire.com>
Date: Tue, 26 Apr 2016 17:46:42 -0400
Subject: [PATCH 1/4] Fix compiler warning in tests

---
 test/gloh_nonfree.cpp          | 22 ----------------------
 test/orb.cpp                   | 12 ------------
 test/sift_nonfree.cpp          | 22 ----------------------
 test/transform_coordinates.cpp |  4 ++--
 4 files changed, 2 insertions(+), 58 deletions(-)

diff --git a/test/gloh_nonfree.cpp b/test/gloh_nonfree.cpp
index 5794051152..558acabe25 100644
--- a/test/gloh_nonfree.cpp
+++ b/test/gloh_nonfree.cpp
@@ -77,18 +77,6 @@ static void array_to_feat_desc(vector<feat_desc_t>& feat, float* x, float* y, fl
     }
 }
 
-static void array_to_feat(vector<feat_t>& feat, float *x, float *y, float *score, float *ori, float *size, unsigned nfeat)
-{
-    feat.resize(nfeat);
-    for (unsigned i = 0; i < feat.size(); i++) {
-        feat[i].f[0] = x[i];
-        feat[i].f[1] = y[i];
-        feat[i].f[2] = score[i];
-        feat[i].f[3] = ori[i];
-        feat[i].f[4] = size[i];
-    }
-}
-
 static void split_feat_desc(vector<feat_desc_t>& fd, vector<feat_t>& f, vector<desc_t>& d)
 {
     f.resize(fd.size());
@@ -104,16 +92,6 @@ static void split_feat_desc(vector<feat_desc_t>& fd, vector<feat_t>& f, vector<d
     }
 }
 
-static unsigned popcount(unsigned x)
-{
-    x = x - ((x >> 1) & 0x55555555);
-    x = (x & 0x33333333) + ((x >> 2) & 0x33333333);
-    x = (x + (x >> 4)) & 0x0F0F0F0F;
-    x = x + (x >> 8);
-    x = x + (x >> 16);
-    return x & 0x0000003F;
-}
-
 static bool compareEuclidean(dim_t desc_len, dim_t ndesc, float *cpu, float *gpu, float unit_thr = 1.f, float euc_thr = 1.f)
 {
     bool ret = true;
diff --git a/test/orb.cpp b/test/orb.cpp
index 1266f20eb6..28e56a1132 100644
--- a/test/orb.cpp
+++ b/test/orb.cpp
@@ -76,18 +76,6 @@ static void array_to_feat_desc(vector<feat_desc_t>& feat, float* x, float* y, fl
     }
 }
 
-static void array_to_feat(vector<feat_t>& feat, float *x, float *y, float *score, float *ori, float *size, unsigned nfeat)
-{
-    feat.resize(nfeat);
-    for (unsigned i = 0; i < feat.size(); i++) {
-        feat[i].f[0] = x[i];
-        feat[i].f[1] = y[i];
-        feat[i].f[2] = score[i];
-        feat[i].f[3] = ori[i];
-        feat[i].f[4] = size[i];
-    }
-}
-
 static void split_feat_desc(vector<feat_desc_t>& fd, vector<feat_t>& f, vector<desc_t>& d)
 {
     f.resize(fd.size());
diff --git a/test/sift_nonfree.cpp b/test/sift_nonfree.cpp
index 6776c18a86..f6dca7ba16 100644
--- a/test/sift_nonfree.cpp
+++ b/test/sift_nonfree.cpp
@@ -76,18 +76,6 @@ static void array_to_feat_desc(vector<feat_desc_t>& feat, float* x, float* y, fl
     }
 }
 
-static void array_to_feat(vector<feat_t>& feat, float *x, float *y, float *score, float *ori, float *size, unsigned nfeat)
-{
-    feat.resize(nfeat);
-    for (unsigned i = 0; i < feat.size(); i++) {
-        feat[i].f[0] = x[i];
-        feat[i].f[1] = y[i];
-        feat[i].f[2] = score[i];
-        feat[i].f[3] = ori[i];
-        feat[i].f[4] = size[i];
-    }
-}
-
 static void split_feat_desc(vector<feat_desc_t>& fd, vector<feat_t>& f, vector<desc_t>& d)
 {
     f.resize(fd.size());
@@ -103,16 +91,6 @@ static void split_feat_desc(vector<feat_desc_t>& fd, vector<feat_t>& f, vector<d
     }
 }
 
-static unsigned popcount(unsigned x)
-{
-    x = x - ((x >> 1) & 0x55555555);
-    x = (x & 0x33333333) + ((x >> 2) & 0x33333333);
-    x = (x + (x >> 4)) & 0x0F0F0F0F;
-    x = x + (x >> 8);
-    x = x + (x >> 16);
-    return x & 0x0000003F;
-}
-
 static bool compareEuclidean(dim_t desc_len, dim_t ndesc, float *cpu, float *gpu, float unit_thr = 1.f, float euc_thr = 1.f)
 {
     bool ret = true;
diff --git a/test/transform_coordinates.cpp b/test/transform_coordinates.cpp
index 7f1ac4e893..dc8598121e 100644
--- a/test/transform_coordinates.cpp
+++ b/test/transform_coordinates.cpp
@@ -47,7 +47,7 @@ void transformCoordinatesTest(string pTestFile)
     af_array outArray = 0;
     ASSERT_EQ(AF_SUCCESS, af_create_array(&tfArray, &(in[0].front()), inDims[0].ndims(), inDims[0].get(), (af_dtype)af::dtype_traits<T>::af_type));
 
-    size_t nTests = in.size();
+    int nTests = in.size();
 
     for (int test = 1; test < nTests; test++) {
         dim_t d0 = (dim_t)in[test][0];
@@ -63,7 +63,7 @@ void transformCoordinatesTest(string pTestFile)
 
         const float thr = 1.f;
 
-        for (size_t elIter = 0; elIter < outEl; elIter++) {
+        for (dim_t elIter = 0; elIter < outEl; elIter++) {
             ASSERT_LE(fabs(outData[elIter] - gold[test-1][elIter]), thr) << "at: " << elIter << std::endl;
         }
 

From b8c506d14fb504b24abdb8410559b236776a4029 Mon Sep 17 00:00:00 2001
From: Shehzan Mohammed <shehzan@arrayfire.com>
Date: Tue, 26 Apr 2016 18:01:58 -0400
Subject: [PATCH 2/4] CPU: Compile sort_by_key instantiations from single file
 into objects

---
 src/backend/cpu/CMakeLists.txt                | 24 ++++++++++---------
 .../cpu/kernel/sort_by_key/CMakeLists.txt     | 15 ++++++++++++
 src/backend/cpu/kernel/sort_by_key/f32.cpp    | 19 ---------------
 src/backend/cpu/kernel/sort_by_key/f64.cpp    | 19 ---------------
 src/backend/cpu/kernel/sort_by_key/s16.cpp    | 19 ---------------
 src/backend/cpu/kernel/sort_by_key/s32.cpp    | 19 ---------------
 src/backend/cpu/kernel/sort_by_key/s64.cpp    | 19 ---------------
 .../{b8.cpp => sort_by_key_impl.cpp}          |  6 +++--
 src/backend/cpu/kernel/sort_by_key/u16.cpp    | 19 ---------------
 src/backend/cpu/kernel/sort_by_key/u32.cpp    | 19 ---------------
 src/backend/cpu/kernel/sort_by_key/u64.cpp    | 19 ---------------
 src/backend/cpu/kernel/sort_by_key/u8.cpp     | 19 ---------------
 12 files changed, 32 insertions(+), 184 deletions(-)
 create mode 100644 src/backend/cpu/kernel/sort_by_key/CMakeLists.txt
 delete mode 100644 src/backend/cpu/kernel/sort_by_key/f32.cpp
 delete mode 100644 src/backend/cpu/kernel/sort_by_key/f64.cpp
 delete mode 100644 src/backend/cpu/kernel/sort_by_key/s16.cpp
 delete mode 100644 src/backend/cpu/kernel/sort_by_key/s32.cpp
 delete mode 100644 src/backend/cpu/kernel/sort_by_key/s64.cpp
 rename src/backend/cpu/kernel/sort_by_key/{b8.cpp => sort_by_key_impl.cpp} (76%)
 delete mode 100644 src/backend/cpu/kernel/sort_by_key/u16.cpp
 delete mode 100644 src/backend/cpu/kernel/sort_by_key/u32.cpp
 delete mode 100644 src/backend/cpu/kernel/sort_by_key/u64.cpp
 delete mode 100644 src/backend/cpu/kernel/sort_by_key/u8.cpp

diff --git a/src/backend/cpu/CMakeLists.txt b/src/backend/cpu/CMakeLists.txt
index ea572693e1..0863266c4f 100644
--- 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
new file mode 100644
index 0000000000..017bb90bb6
--- /dev/null
+++ 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
deleted file mode 100644
index 11d8139957..0000000000
--- a/src/backend/cpu/kernel/sort_by_key/f32.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cpu
-{
-namespace kernel
-{
-    INSTANTIATE1(float,true)
-    INSTANTIATE1(float,false)
-}
-}
diff --git a/src/backend/cpu/kernel/sort_by_key/f64.cpp b/src/backend/cpu/kernel/sort_by_key/f64.cpp
deleted file mode 100644
index 21746d773a..0000000000
--- a/src/backend/cpu/kernel/sort_by_key/f64.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cpu
-{
-namespace kernel
-{
-    INSTANTIATE1(double,true)
-    INSTANTIATE1(double,false)
-}
-}
diff --git a/src/backend/cpu/kernel/sort_by_key/s16.cpp b/src/backend/cpu/kernel/sort_by_key/s16.cpp
deleted file mode 100644
index 50b718d04c..0000000000
--- a/src/backend/cpu/kernel/sort_by_key/s16.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cpu
-{
-namespace kernel
-{
-    INSTANTIATE1(short,true)
-    INSTANTIATE1(short,false)
-}
-}
diff --git a/src/backend/cpu/kernel/sort_by_key/s32.cpp b/src/backend/cpu/kernel/sort_by_key/s32.cpp
deleted file mode 100644
index c50efcdc26..0000000000
--- a/src/backend/cpu/kernel/sort_by_key/s32.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cpu
-{
-namespace kernel
-{
-    INSTANTIATE1(int,true)
-    INSTANTIATE1(int,false)
-}
-}
diff --git a/src/backend/cpu/kernel/sort_by_key/s64.cpp b/src/backend/cpu/kernel/sort_by_key/s64.cpp
deleted file mode 100644
index 82946f820e..0000000000
--- a/src/backend/cpu/kernel/sort_by_key/s64.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cpu
-{
-namespace kernel
-{
-    INSTANTIATE1(intl,true)
-    INSTANTIATE1(intl,false)
-}
-}
diff --git a/src/backend/cpu/kernel/sort_by_key/b8.cpp b/src/backend/cpu/kernel/sort_by_key/sort_by_key_impl.cpp
similarity index 76%
rename from src/backend/cpu/kernel/sort_by_key/b8.cpp
rename to src/backend/cpu/kernel/sort_by_key/sort_by_key_impl.cpp
index 855e7a93ca..fbdedfde39 100644
--- a/src/backend/cpu/kernel/sort_by_key/b8.cpp
+++ b/src/backend/cpu/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
deleted file mode 100644
index feedd1d56e..0000000000
--- a/src/backend/cpu/kernel/sort_by_key/u16.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cpu
-{
-namespace kernel
-{
-    INSTANTIATE1(ushort,true)
-    INSTANTIATE1(ushort,false)
-}
-}
diff --git a/src/backend/cpu/kernel/sort_by_key/u32.cpp b/src/backend/cpu/kernel/sort_by_key/u32.cpp
deleted file mode 100644
index cd514af19a..0000000000
--- a/src/backend/cpu/kernel/sort_by_key/u32.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cpu
-{
-namespace kernel
-{
-    INSTANTIATE1(uint,true)
-    INSTANTIATE1(uint,false)
-}
-}
diff --git a/src/backend/cpu/kernel/sort_by_key/u64.cpp b/src/backend/cpu/kernel/sort_by_key/u64.cpp
deleted file mode 100644
index ec955b3de7..0000000000
--- a/src/backend/cpu/kernel/sort_by_key/u64.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cpu
-{
-namespace kernel
-{
-    INSTANTIATE1(uintl,true)
-    INSTANTIATE1(uintl,false)
-}
-}
diff --git a/src/backend/cpu/kernel/sort_by_key/u8.cpp b/src/backend/cpu/kernel/sort_by_key/u8.cpp
deleted file mode 100644
index fd58cbfaa1..0000000000
--- a/src/backend/cpu/kernel/sort_by_key/u8.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cpu
-{
-namespace kernel
-{
-    INSTANTIATE1(uchar,true)
-    INSTANTIATE1(uchar,false)
-}
-}

From c354015282c322ff9418f2d4b1814d5f0263951c Mon Sep 17 00:00:00 2001
From: Shehzan Mohammed <shehzan@arrayfire.com>
Date: Tue, 26 Apr 2016 17:47:30 -0400
Subject: [PATCH 3/4] CUDA: Generate sort_by_key instantiations using CMake

---
 src/backend/cuda/CMakeLists.txt               |  4 ++-
 .../cuda/kernel/sort_by_key/CMakeLists.txt    | 29 +++++++++++++++++++
 .../cuda/kernel/sort_by_key/ascd_f64.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/ascd_s16.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/ascd_s32.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/ascd_s64.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/ascd_s8.cu        | 18 ------------
 .../cuda/kernel/sort_by_key/ascd_u16.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/ascd_u32.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/ascd_u64.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/ascd_u8.cu        | 18 ------------
 .../cuda/kernel/sort_by_key/desc_f32.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/desc_f64.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/desc_s16.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/desc_s32.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/desc_s64.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/desc_s8.cu        | 18 ------------
 .../cuda/kernel/sort_by_key/desc_u16.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/desc_u32.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/desc_u64.cu       | 18 ------------
 .../cuda/kernel/sort_by_key/desc_u8.cu        | 18 ------------
 .../{ascd_f32.cu => sort_by_key_impl.cu.in}   |  8 ++++-
 src/backend/cuda/kernel/sort_by_key_impl.hpp  |  9 ++++--
 23 files changed, 45 insertions(+), 347 deletions(-)
 create mode 100644 src/backend/cuda/kernel/sort_by_key/CMakeLists.txt
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/ascd_f64.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/ascd_s16.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/ascd_s32.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/ascd_s64.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/ascd_s8.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/ascd_u16.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/ascd_u32.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/ascd_u64.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/ascd_u8.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/desc_f32.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/desc_f64.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/desc_s16.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/desc_s32.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/desc_s64.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/desc_s8.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/desc_u16.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/desc_u32.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/desc_u64.cu
 delete mode 100644 src/backend/cuda/kernel/sort_by_key/desc_u8.cu
 rename src/backend/cuda/kernel/sort_by_key/{ascd_f32.cu => sort_by_key_impl.cu.in} (57%)

diff --git a/src/backend/cuda/CMakeLists.txt b/src/backend/cuda/CMakeLists.txt
index 4efb42764a..d1727ffd40 100644
--- 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 ${ptx_targets})
diff --git a/src/backend/cuda/kernel/sort_by_key/CMakeLists.txt b/src/backend/cuda/kernel/sort_by_key/CMakeLists.txt
new file mode 100644
index 0000000000..a9143f282d
--- /dev/null
+++ b/src/backend/cuda/kernel/sort_by_key/CMakeLists.txt
@@ -0,0 +1,29 @@
+FILE(STRINGS "${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
deleted file mode 100644
index ba19ec447c..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/ascd_f64.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(double, true)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_s16.cu b/src/backend/cuda/kernel/sort_by_key/ascd_s16.cu
deleted file mode 100644
index 1be6e540ca..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/ascd_s16.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(short, true)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_s32.cu b/src/backend/cuda/kernel/sort_by_key/ascd_s32.cu
deleted file mode 100644
index 8cee7c9b49..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/ascd_s32.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(int, true)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_s64.cu b/src/backend/cuda/kernel/sort_by_key/ascd_s64.cu
deleted file mode 100644
index 0e5a7c81a2..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/ascd_s64.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(intl, true)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_s8.cu b/src/backend/cuda/kernel/sort_by_key/ascd_s8.cu
deleted file mode 100644
index 81ed32952f..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/ascd_s8.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(char, true)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_u16.cu b/src/backend/cuda/kernel/sort_by_key/ascd_u16.cu
deleted file mode 100644
index e232c08376..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/ascd_u16.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(ushort, true)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_u32.cu b/src/backend/cuda/kernel/sort_by_key/ascd_u32.cu
deleted file mode 100644
index 34a4580936..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/ascd_u32.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(uint, true)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_u64.cu b/src/backend/cuda/kernel/sort_by_key/ascd_u64.cu
deleted file mode 100644
index fc576e7f99..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/ascd_u64.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(uintl, true)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_u8.cu b/src/backend/cuda/kernel/sort_by_key/ascd_u8.cu
deleted file mode 100644
index ed8454d53e..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/ascd_u8.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(uchar, true)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/desc_f32.cu b/src/backend/cuda/kernel/sort_by_key/desc_f32.cu
deleted file mode 100644
index 73459ac033..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/desc_f32.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(float, false)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/desc_f64.cu b/src/backend/cuda/kernel/sort_by_key/desc_f64.cu
deleted file mode 100644
index be0536b1e3..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/desc_f64.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(double, false)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/desc_s16.cu b/src/backend/cuda/kernel/sort_by_key/desc_s16.cu
deleted file mode 100644
index 0fc3b50827..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/desc_s16.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(short, false)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/desc_s32.cu b/src/backend/cuda/kernel/sort_by_key/desc_s32.cu
deleted file mode 100644
index cfda29c7de..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/desc_s32.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(int, false)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/desc_s64.cu b/src/backend/cuda/kernel/sort_by_key/desc_s64.cu
deleted file mode 100644
index b334a91a99..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/desc_s64.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(intl, false)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/desc_s8.cu b/src/backend/cuda/kernel/sort_by_key/desc_s8.cu
deleted file mode 100644
index f02d5ce2fe..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/desc_s8.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(char, false)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/desc_u16.cu b/src/backend/cuda/kernel/sort_by_key/desc_u16.cu
deleted file mode 100644
index 9b0a77cb25..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/desc_u16.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(ushort, false)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/desc_u32.cu b/src/backend/cuda/kernel/sort_by_key/desc_u32.cu
deleted file mode 100644
index 1d02aec848..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/desc_u32.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(uint, false)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/desc_u64.cu b/src/backend/cuda/kernel/sort_by_key/desc_u64.cu
deleted file mode 100644
index 597bd2c1b4..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/desc_u64.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(uintl, false)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/desc_u8.cu b/src/backend/cuda/kernel/sort_by_key/desc_u8.cu
deleted file mode 100644
index 4f55479604..0000000000
--- a/src/backend/cuda/kernel/sort_by_key/desc_u8.cu
+++ /dev/null
@@ -1,18 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace cuda
-{
-namespace kernel
-{
-    INSTANTIATE1(uchar, false)
-}
-}
diff --git a/src/backend/cuda/kernel/sort_by_key/ascd_f32.cu b/src/backend/cuda/kernel/sort_by_key/sort_by_key_impl.cu.in
similarity index 57%
rename from src/backend/cuda/kernel/sort_by_key/ascd_f32.cu
rename to src/backend/cuda/kernel/sort_by_key/sort_by_key_impl.cu.in
index 284e8b4938..94168df1de 100644
--- a/src/backend/cuda/kernel/sort_by_key/ascd_f32.cu
+++ b/src/backend/cuda/kernel/sort_by_key/sort_by_key_impl.cu.in
@@ -9,10 +9,16 @@
 
 #include <kernel/sort_by_key_impl.hpp>
 
+// This file instantiates sort_by_key as separate object files from CMake
+// The 3 lines below are read by CMake to determenine the instantiations
+// SBK_TYPES:float double int uint intl uintl short ushort char uchar
+// SBK_DIRS:true false
+// SBK_INSTS:0 1
+
 namespace cuda
 {
 namespace kernel
 {
-    INSTANTIATE1(float, true)
+    INSTANTIATE@SBK_INST@(@SBK_TYPE@, @SBK_DIR@)
 }
 }
diff --git a/src/backend/cuda/kernel/sort_by_key_impl.hpp b/src/backend/cuda/kernel/sort_by_key_impl.hpp
index bcc2fefeb4..c035e3ae07 100644
--- a/src/backend/cuda/kernel/sort_by_key_impl.hpp
+++ b/src/backend/cuda/kernel/sort_by_key_impl.hpp
@@ -198,18 +198,21 @@ namespace cuda
     template void sortByKeyBatched<Tk, Tv, dr, 2>(Param<Tk> okey, Param<Tv> oval);      \
     template void sortByKeyBatched<Tk, Tv, dr, 3>(Param<Tk> okey, Param<Tv> oval);      \
 
-#define INSTANTIATE1(Tk    , dr) \
+#define INSTANTIATE0(Tk    , dr) \
     INSTANTIATE(Tk, float  , dr) \
     INSTANTIATE(Tk, double , dr) \
     INSTANTIATE(Tk, cfloat , dr) \
     INSTANTIATE(Tk, cdouble, dr) \
+    INSTANTIATE(Tk, char   , dr) \
+    INSTANTIATE(Tk, uchar  , dr) \
+
+#define INSTANTIATE1(Tk    , dr) \
     INSTANTIATE(Tk, int    , dr) \
     INSTANTIATE(Tk, uint   , dr) \
     INSTANTIATE(Tk, short  , dr) \
     INSTANTIATE(Tk, ushort , dr) \
-    INSTANTIATE(Tk, char   , dr) \
-    INSTANTIATE(Tk, uchar  , dr) \
     INSTANTIATE(Tk, intl   , dr) \
     INSTANTIATE(Tk, uintl  , dr)
+
     }
 }

From da69c30a3556cc7909d5193c8e8dcbb68591ab81 Mon Sep 17 00:00:00 2001
From: Shehzan Mohammed <shehzan@arrayfire.com>
Date: Tue, 26 Apr 2016 18:11:16 -0400
Subject: [PATCH 4/4] OpenCL: Compile sort_by_key instantiations from single
 file into objects

---
 src/backend/opencl/CMakeLists.txt             | 11 +++++++----
 .../opencl/kernel/sort_by_key/CMakeLists.txt  | 19 +++++++++++++++++++
 src/backend/opencl/kernel/sort_by_key/f32.cpp | 19 -------------------
 src/backend/opencl/kernel/sort_by_key/f64.cpp | 19 -------------------
 src/backend/opencl/kernel/sort_by_key/s16.cpp | 19 -------------------
 src/backend/opencl/kernel/sort_by_key/s32.cpp | 19 -------------------
 src/backend/opencl/kernel/sort_by_key/s64.cpp | 19 -------------------
 .../{b8.cpp => sort_by_key_impl.cpp}          |  6 ++++--
 src/backend/opencl/kernel/sort_by_key/u16.cpp | 19 -------------------
 src/backend/opencl/kernel/sort_by_key/u32.cpp | 19 -------------------
 src/backend/opencl/kernel/sort_by_key/u64.cpp | 19 -------------------
 src/backend/opencl/kernel/sort_by_key/u8.cpp  | 19 -------------------
 12 files changed, 30 insertions(+), 177 deletions(-)
 create mode 100644 src/backend/opencl/kernel/sort_by_key/CMakeLists.txt
 delete mode 100644 src/backend/opencl/kernel/sort_by_key/f32.cpp
 delete mode 100644 src/backend/opencl/kernel/sort_by_key/f64.cpp
 delete mode 100644 src/backend/opencl/kernel/sort_by_key/s16.cpp
 delete mode 100644 src/backend/opencl/kernel/sort_by_key/s32.cpp
 delete mode 100644 src/backend/opencl/kernel/sort_by_key/s64.cpp
 rename src/backend/opencl/kernel/sort_by_key/{b8.cpp => sort_by_key_impl.cpp} (76%)
 delete mode 100644 src/backend/opencl/kernel/sort_by_key/u16.cpp
 delete mode 100644 src/backend/opencl/kernel/sort_by_key/u32.cpp
 delete mode 100644 src/backend/opencl/kernel/sort_by_key/u64.cpp
 delete mode 100644 src/backend/opencl/kernel/sort_by_key/u8.cpp

diff --git a/src/backend/opencl/CMakeLists.txt b/src/backend/opencl/CMakeLists.txt
index 71247ce3da..9e4918a5ff 100644
--- a/src/backend/opencl/CMakeLists.txt
+++ b/src/backend/opencl/CMakeLists.txt
@@ -126,8 +126,7 @@ FILE(GLOB opencl_kernels
     "kernel/*.cl")
 
 FILE(GLOB kernel_sources
-     "kernel/*.cpp"
-     "kernel/sort_by_key/*.cpp")
+     "kernel/*.cpp")
 
 FILE(GLOB conv_ker_headers
     "kernel/convolve/*.hpp")
@@ -237,6 +236,8 @@ IF(UNIX)
     SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fPIC -pthread -Wno-comment")
 ENDIF()
 
+INCLUDE("${CMAKE_CURRENT_SOURCE_DIR}/kernel/sort_by_key/CMakeLists.txt")
+
 IF(DEFINED BLAS_SYM_FILE)
 
     ADD_LIBRARY(afopencl_static STATIC
@@ -253,7 +254,8 @@ IF(DEFINED BLAS_SYM_FILE)
                 ${backend_headers}
                 ${backend_sources}
                 ${magma_sources}
-                ${magma_headers})
+                ${magma_headers}
+                ${SORT_BY_KEY_OBJECTS})
 
     ADD_LIBRARY(afopencl SHARED
                ${c_headers}
@@ -296,7 +298,8 @@ ELSE(DEFINED BLAS_SYM_FILE)
                 ${c_sources}
                 ${cpp_sources}
                 ${magma_sources}
-                ${magma_headers})
+                ${magma_headers}
+                ${SORT_BY_KEY_OBJECTS})
 
 ENDIF()
 
diff --git a/src/backend/opencl/kernel/sort_by_key/CMakeLists.txt b/src/backend/opencl/kernel/sort_by_key/CMakeLists.txt
new file mode 100644
index 0000000000..760fe6b634
--- /dev/null
+++ b/src/backend/opencl/kernel/sort_by_key/CMakeLists.txt
@@ -0,0 +1,19 @@
+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 ${SBK_TYPES})
+    ADD_LIBRARY(opencl_sort_by_key_${SBK_TYPE} OBJECT
+        "${CMAKE_CURRENT_SOURCE_DIR}/kernel/sort_by_key/sort_by_key_impl.cpp")
+    ADD_DEPENDENCIES(opencl_sort_by_key_${SBK_TYPE} ${cl_kernel_targets})
+    IF(FORGE_FOUND AND NOT USE_SYSTEM_FORGE)
+        ADD_DEPENDENCIES(opencl_sort_by_key_${SBK_TYPE} forge)
+    ENDIF()
+    SET_TARGET_PROPERTIES(opencl_sort_by_key_${SBK_TYPE} PROPERTIES COMPILE_FLAGS "-DTYPE=${SBK_TYPE}")
+    LIST(APPEND SORT_BY_KEY_OBJECTS $<TARGET_OBJECTS:opencl_sort_by_key_${SBK_TYPE}>)
+ENDFOREACH(SBK_TYPE ${SBK_TYPES})
diff --git a/src/backend/opencl/kernel/sort_by_key/f32.cpp b/src/backend/opencl/kernel/sort_by_key/f32.cpp
deleted file mode 100644
index a1e9ae5f1f..0000000000
--- a/src/backend/opencl/kernel/sort_by_key/f32.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace opencl
-{
-namespace kernel
-{
-    INSTANTIATE1(float,true)
-    INSTANTIATE1(float,false)
-}
-}
diff --git a/src/backend/opencl/kernel/sort_by_key/f64.cpp b/src/backend/opencl/kernel/sort_by_key/f64.cpp
deleted file mode 100644
index 7fb7a79bd8..0000000000
--- a/src/backend/opencl/kernel/sort_by_key/f64.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace opencl
-{
-namespace kernel
-{
-    INSTANTIATE1(double,true)
-    INSTANTIATE1(double,false)
-}
-}
diff --git a/src/backend/opencl/kernel/sort_by_key/s16.cpp b/src/backend/opencl/kernel/sort_by_key/s16.cpp
deleted file mode 100644
index 491ea0e3a2..0000000000
--- a/src/backend/opencl/kernel/sort_by_key/s16.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace opencl
-{
-namespace kernel
-{
-    INSTANTIATE1(short,true)
-    INSTANTIATE1(short,false)
-}
-}
diff --git a/src/backend/opencl/kernel/sort_by_key/s32.cpp b/src/backend/opencl/kernel/sort_by_key/s32.cpp
deleted file mode 100644
index 67ba20e7dd..0000000000
--- a/src/backend/opencl/kernel/sort_by_key/s32.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace opencl
-{
-namespace kernel
-{
-    INSTANTIATE1(int,true)
-    INSTANTIATE1(int,false)
-}
-}
diff --git a/src/backend/opencl/kernel/sort_by_key/s64.cpp b/src/backend/opencl/kernel/sort_by_key/s64.cpp
deleted file mode 100644
index a48f36ee47..0000000000
--- a/src/backend/opencl/kernel/sort_by_key/s64.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace opencl
-{
-namespace kernel
-{
-    INSTANTIATE1(intl,true)
-    INSTANTIATE1(intl,false)
-}
-}
diff --git a/src/backend/opencl/kernel/sort_by_key/b8.cpp b/src/backend/opencl/kernel/sort_by_key/sort_by_key_impl.cpp
similarity index 76%
rename from src/backend/opencl/kernel/sort_by_key/b8.cpp
rename to src/backend/opencl/kernel/sort_by_key/sort_by_key_impl.cpp
index ad0d7f48ae..bf4c96bbb2 100644
--- a/src/backend/opencl/kernel/sort_by_key/b8.cpp
+++ b/src/backend/opencl/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 opencl
 {
 namespace kernel
 {
-    INSTANTIATE1(char,true)
-    INSTANTIATE1(char,false)
+    INSTANTIATE1(TYPE,true)
+    INSTANTIATE1(TYPE,false)
 }
 }
diff --git a/src/backend/opencl/kernel/sort_by_key/u16.cpp b/src/backend/opencl/kernel/sort_by_key/u16.cpp
deleted file mode 100644
index 36678d0a42..0000000000
--- a/src/backend/opencl/kernel/sort_by_key/u16.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace opencl
-{
-namespace kernel
-{
-    INSTANTIATE1(ushort,true)
-    INSTANTIATE1(ushort,false)
-}
-}
diff --git a/src/backend/opencl/kernel/sort_by_key/u32.cpp b/src/backend/opencl/kernel/sort_by_key/u32.cpp
deleted file mode 100644
index f1e4b5322f..0000000000
--- a/src/backend/opencl/kernel/sort_by_key/u32.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace opencl
-{
-namespace kernel
-{
-    INSTANTIATE1(uint,true)
-    INSTANTIATE1(uint,false)
-}
-}
diff --git a/src/backend/opencl/kernel/sort_by_key/u64.cpp b/src/backend/opencl/kernel/sort_by_key/u64.cpp
deleted file mode 100644
index 0a6f5b0c4f..0000000000
--- a/src/backend/opencl/kernel/sort_by_key/u64.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace opencl
-{
-namespace kernel
-{
-    INSTANTIATE1(uintl,true)
-    INSTANTIATE1(uintl,false)
-}
-}
diff --git a/src/backend/opencl/kernel/sort_by_key/u8.cpp b/src/backend/opencl/kernel/sort_by_key/u8.cpp
deleted file mode 100644
index 45af011a86..0000000000
--- a/src/backend/opencl/kernel/sort_by_key/u8.cpp
+++ /dev/null
@@ -1,19 +0,0 @@
-/*******************************************************
- * Copyright (c) 2014, ArrayFire
- * All rights reserved.
- *
- * This file is distributed under 3-clause BSD license.
- * The complete license agreement can be obtained at:
- * http://arrayfire.com/licenses/BSD-3-Clause
- ********************************************************/
-
-#include <kernel/sort_by_key_impl.hpp>
-
-namespace opencl
-{
-namespace kernel
-{
-    INSTANTIATE1(uchar,true)
-    INSTANTIATE1(uchar,false)
-}
-}