From 4636377cf4bf40544144933f0d646b1b890e5f21 Mon Sep 17 00:00:00 2001
From: Robert Maynard <rmaynard@nvidia.com>
Date: Thu, 27 May 2021 08:45:25 -0400
Subject: [PATCH 01/11] Remove unneeded public dependencies from `cuml`

The C api library like the C++ api doesn't need consumers to
link to faiss, so remove it from public API
---
 cpp/CMakeLists.txt | 1 -
 1 file changed, 1 deletion(-)

diff --git a/cpp/CMakeLists.txt b/cpp/CMakeLists.txt
index 1da22c3fa6..e68efb54f3 100644
--- a/cpp/CMakeLists.txt
+++ b/cpp/CMakeLists.txt
@@ -375,7 +375,6 @@ if(BUILD_CUML_C_LIBRARY)
   target_link_libraries(${CUML_C_TARGET}
     PUBLIC
       ${CUML_CPP_TARGET}
-      FAISS::FAISS
   )
 
 endif()

From f9deab8dee151e9e72f57ab858b730734ede0d2d Mon Sep 17 00:00:00 2001
From: Robert Maynard <rmaynard@nvidia.com>
Date: Thu, 27 May 2021 08:46:47 -0400
Subject: [PATCH 02/11] Correctly make sure the C language is enabled when
 needed

The tests for the cuml C library need to ensure that the C language
is enabled, as that isn't done by default since the `cuml` library
uses C++ internally with `extern "C"` blocks.
---
 cpp/test/CMakeLists.txt | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/cpp/test/CMakeLists.txt b/cpp/test/CMakeLists.txt
index 040df7b461..bcebb4bb6e 100644
--- a/cpp/test/CMakeLists.txt
+++ b/cpp/test/CMakeLists.txt
@@ -228,6 +228,9 @@ endif()
 # - build C-API test library -------------------------------------------------
 
 if(BUILD_CUML_C_LIBRARY)
+
+  enable_language(C)
+
   add_library(${CUML_C_TEST_TARGET} SHARED
     c_api/dbscan_api_test.c
     c_api/glm_api_test.c

From db47776faf1b71c2c47f859715c63b9464a27fbd Mon Sep 17 00:00:00 2001
From: Robert Maynard <rmaynard@nvidia.com>
Date: Thu, 27 May 2021 09:35:34 -0400
Subject: [PATCH 03/11] Correct dependencies for `cuml` C target

The C api library like the C++ api doesn't need consumers to
link to faiss, so remove it from public API, but it needs to be
private.

In addition correctly specify where to find the FAISS headers
---
 cpp/CMakeLists.txt                   | 2 ++
 cpp/cmake/thirdparty/get_faiss.cmake | 2 +-
 2 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/cpp/CMakeLists.txt b/cpp/CMakeLists.txt
index e68efb54f3..50289129ab 100644
--- a/cpp/CMakeLists.txt
+++ b/cpp/CMakeLists.txt
@@ -375,6 +375,8 @@ if(BUILD_CUML_C_LIBRARY)
   target_link_libraries(${CUML_C_TARGET}
     PUBLIC
       ${CUML_CPP_TARGET}
+    PRIVATE
+      FAISS::FAISS
   )
 
 endif()
diff --git a/cpp/cmake/thirdparty/get_faiss.cmake b/cpp/cmake/thirdparty/get_faiss.cmake
index 4991990e37..626af5196c 100644
--- a/cpp/cmake/thirdparty/get_faiss.cmake
+++ b/cpp/cmake/thirdparty/get_faiss.cmake
@@ -39,7 +39,7 @@ function(find_and_configure_faiss)
     )
 
     if(FAISS_ADDED)
-      set(FAISS_GPU_HEADERS ${FAISS_SOURCE_DIR} PARENT_SCOPE)
+      target_include_directories(faiss INTERFACE $<BUILD_INTERFACE:${FAISS_SOURCE_DIR}>)
       add_library(FAISS::FAISS ALIAS faiss)
     endif()
 

From efa36ca62a4d06317433e3c94b2639e815e19b3e Mon Sep 17 00:00:00 2001
From: Robert Maynard <rmaynard@nvidia.com>
Date: Thu, 27 May 2021 09:36:39 -0400
Subject: [PATCH 04/11] Reenable benchmarks

---
 cpp/CMakeLists.txt | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/cpp/CMakeLists.txt b/cpp/CMakeLists.txt
index 50289129ab..273e76019c 100644
--- a/cpp/CMakeLists.txt
+++ b/cpp/CMakeLists.txt
@@ -449,6 +449,13 @@ rapids_export(BUILD cuml
     FINAL_CODE_BLOCK code_string
     )
 
+##############################################################################
+# - build benchmark executable -----------------------------------------------
+
+if(BUILD_CUML_BENCH)
+  add_subdirectory(bench)
+endif(BUILD_CUML_BENCH)
+
 ##############################################################################
 # - doxygen targets ----------------------------------------------------------
 

From 60d2b823173dc90e5b83446df09037b090eda593 Mon Sep 17 00:00:00 2001
From: Robert Maynard <rmaynard@nvidia.com>
Date: Thu, 27 May 2021 10:49:44 -0400
Subject: [PATCH 05/11] Correct dependencies for `CUML_C_TEST_TARGET` target

---
 cpp/test/CMakeLists.txt | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/cpp/test/CMakeLists.txt b/cpp/test/CMakeLists.txt
index bcebb4bb6e..3e89543763 100644
--- a/cpp/test/CMakeLists.txt
+++ b/cpp/test/CMakeLists.txt
@@ -239,8 +239,6 @@ if(BUILD_CUML_C_LIBRARY)
     c_api/svm_api_test.c
   )
 
-  set_target_properties(${CUML_C_TEST_TARGET} PROPERTIES LINKER_LANGUAGE C)
-
-  target_link_libraries(${CUML_C_TEST_TARGET} PUBLIC cuml_c)
+  target_link_libraries(${CUML_C_TEST_TARGET} PUBLIC ${CUML_C_TARGET})
 
 endif(BUILD_CUML_C_LIBRARY)

From c1aa33abff4dc1465883179cf20f4034a2fa0d35 Mon Sep 17 00:00:00 2001
From: Robert Maynard <robertjmaynard@gmail.com>
Date: Thu, 27 May 2021 11:43:02 -0400
Subject: [PATCH 06/11] Update cpp/CMakeLists.txt

Co-authored-by: Dante Gama Dessavre <dante.gamadessavre@gmail.com>
---
 cpp/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/cpp/CMakeLists.txt b/cpp/CMakeLists.txt
index 273e76019c..386a8f6f4c 100644
--- a/cpp/CMakeLists.txt
+++ b/cpp/CMakeLists.txt
@@ -452,7 +452,7 @@ rapids_export(BUILD cuml
 ##############################################################################
 # - build benchmark executable -----------------------------------------------
 
-if(BUILD_CUML_BENCH)
+if(BUILD_CUML_BENCH OR BUILD_CUML_PRIMS_BENCH)
   add_subdirectory(bench)
 endif(BUILD_CUML_BENCH)
 

From 318f5ea43bd7f45b91464977a6892792248e949b Mon Sep 17 00:00:00 2001
From: Robert Maynard <rmaynard@nvidia.com>
Date: Thu, 27 May 2021 14:18:57 -0400
Subject: [PATCH 07/11] Cleanly build all cuml benchmarks

---
 cpp/bench/CMakeLists.txt            | 1 +
 cpp/bench/prims/add.cu              | 2 +-
 cpp/bench/prims/distance_common.cuh | 2 +-
 cpp/bench/prims/fused_l2_nn.cu      | 2 +-
 cpp/bench/prims/gram_matrix.cu      | 2 +-
 cpp/bench/prims/make_blobs.cu       | 2 +-
 cpp/bench/prims/map_then_reduce.cu  | 2 +-
 cpp/bench/prims/matrix_vector_op.cu | 2 +-
 cpp/bench/prims/permute.cu          | 2 +-
 cpp/bench/prims/reduce.cu           | 2 +-
 cpp/bench/prims/rng.cu              | 2 +-
 11 files changed, 11 insertions(+), 10 deletions(-)

diff --git a/cpp/bench/CMakeLists.txt b/cpp/bench/CMakeLists.txt
index 4d6d64d72a..08d7c1bcf8 100644
--- a/cpp/bench/CMakeLists.txt
+++ b/cpp/bench/CMakeLists.txt
@@ -89,5 +89,6 @@ if(BUILD_CUML_PRIMS_BENCH)
   target_include_directories(${PRIMS_BENCH_TARGET}
     PRIVATE
       $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../src_prims>
+      $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}>
   )
 endif(BUILD_CUML_PRIMS_BENCH)
diff --git a/cpp/bench/prims/add.cu b/cpp/bench/prims/add.cu
index cfe30a5b96..b28c3595e0 100644
--- a/cpp/bench/prims/add.cu
+++ b/cpp/bench/prims/add.cu
@@ -16,7 +16,7 @@
 
 #include <raft/linalg/add.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include "../common/ml_benchmark.hpp"
+#include "common/ml_benchmark.hpp"
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/distance_common.cuh b/cpp/bench/prims/distance_common.cuh
index de51183c68..f83fb791c8 100644
--- a/cpp/bench/prims/distance_common.cuh
+++ b/cpp/bench/prims/distance_common.cuh
@@ -17,7 +17,7 @@
 #include <raft/cudart_utils.h>
 #include <raft/distance/distance.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include "../common/ml_benchmark.hpp"
+#include "common/ml_benchmark.hpp"
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/fused_l2_nn.cu b/cpp/bench/prims/fused_l2_nn.cu
index d84e27003d..f38cca1adb 100644
--- a/cpp/bench/prims/fused_l2_nn.cu
+++ b/cpp/bench/prims/fused_l2_nn.cu
@@ -20,7 +20,7 @@
 #include <raft/linalg/norm.cuh>
 #include <raft/mr/device/allocator.hpp>
 #include <raft/random/rng.cuh>
-#include "../common/ml_benchmark.hpp"
+#include "common/ml_benchmark.hpp"
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/gram_matrix.cu b/cpp/bench/prims/gram_matrix.cu
index d3e18ca44c..e925446b49 100644
--- a/cpp/bench/prims/gram_matrix.cu
+++ b/cpp/bench/prims/gram_matrix.cu
@@ -23,7 +23,7 @@
 #include <sstream>
 #include <string>
 #include <vector>
-#include "../common/ml_benchmark.hpp"
+#include "common/ml_benchmark.hpp"
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/make_blobs.cu b/cpp/bench/prims/make_blobs.cu
index b7285065da..346e6b162e 100644
--- a/cpp/bench/prims/make_blobs.cu
+++ b/cpp/bench/prims/make_blobs.cu
@@ -16,7 +16,7 @@
 
 #include <raft/mr/device/allocator.hpp>
 #include <random/make_blobs.cuh>
-#include "../common/ml_benchmark.hpp"
+#include "common/ml_benchmark.hpp"
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/map_then_reduce.cu b/cpp/bench/prims/map_then_reduce.cu
index fd2be4d58c..8a1917d6d0 100644
--- a/cpp/bench/prims/map_then_reduce.cu
+++ b/cpp/bench/prims/map_then_reduce.cu
@@ -16,7 +16,7 @@
 
 #include <raft/linalg/map_then_reduce.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include "../common/ml_benchmark.hpp"
+#include "common/ml_benchmark.hpp"
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/matrix_vector_op.cu b/cpp/bench/prims/matrix_vector_op.cu
index 448f3be88f..787deecfeb 100644
--- a/cpp/bench/prims/matrix_vector_op.cu
+++ b/cpp/bench/prims/matrix_vector_op.cu
@@ -16,7 +16,7 @@
 
 #include <raft/linalg/matrix_vector_op.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include "../common/ml_benchmark.hpp"
+#include "common/ml_benchmark.hpp"
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/permute.cu b/cpp/bench/prims/permute.cu
index 35190eddca..6246aee8e7 100644
--- a/cpp/bench/prims/permute.cu
+++ b/cpp/bench/prims/permute.cu
@@ -18,7 +18,7 @@
 #include <raft/mr/device/allocator.hpp>
 #include <raft/random/rng.cuh>
 #include <random/permute.cuh>
-#include "../common/ml_benchmark.hpp"
+#include "common/ml_benchmark.hpp"
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/reduce.cu b/cpp/bench/prims/reduce.cu
index 30b7fdba7d..58cc4f2a7f 100644
--- a/cpp/bench/prims/reduce.cu
+++ b/cpp/bench/prims/reduce.cu
@@ -16,7 +16,7 @@
 
 #include <raft/linalg/reduce.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include "../common/ml_benchmark.hpp"
+#include "common/ml_benchmark.hpp"
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/rng.cu b/cpp/bench/prims/rng.cu
index 082a8163a9..f90bf773d3 100644
--- a/cpp/bench/prims/rng.cu
+++ b/cpp/bench/prims/rng.cu
@@ -17,7 +17,7 @@
 #include <raft/cudart_utils.h>
 #include <raft/mr/device/allocator.hpp>
 #include <raft/random/rng.cuh>
-#include "../common/ml_benchmark.hpp"
+#include "common/ml_benchmark.hpp"
 
 namespace MLCommon {
 namespace Bench {

From 803b4db23bf90fc26fb301d726f71978cee1fd58 Mon Sep 17 00:00:00 2001
From: Robert Maynard <rmaynard@nvidia.com>
Date: Thu, 27 May 2021 14:28:02 -0400
Subject: [PATCH 08/11] Correct style issues found by CI

---
 cpp/bench/prims/add.cu              | 2 +-
 cpp/bench/prims/distance_common.cuh | 2 +-
 cpp/bench/prims/fused_l2_nn.cu      | 2 +-
 cpp/bench/prims/gram_matrix.cu      | 2 +-
 cpp/bench/prims/make_blobs.cu       | 2 +-
 cpp/bench/prims/map_then_reduce.cu  | 2 +-
 cpp/bench/prims/matrix_vector_op.cu | 2 +-
 cpp/bench/prims/permute.cu          | 2 +-
 cpp/bench/prims/reduce.cu           | 2 +-
 cpp/bench/prims/rng.cu              | 2 +-
 10 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/cpp/bench/prims/add.cu b/cpp/bench/prims/add.cu
index b28c3595e0..b43021c7d1 100644
--- a/cpp/bench/prims/add.cu
+++ b/cpp/bench/prims/add.cu
@@ -16,7 +16,7 @@
 
 #include <raft/linalg/add.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include "common/ml_benchmark.hpp"
+#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/distance_common.cuh b/cpp/bench/prims/distance_common.cuh
index f83fb791c8..3b86025754 100644
--- a/cpp/bench/prims/distance_common.cuh
+++ b/cpp/bench/prims/distance_common.cuh
@@ -17,7 +17,7 @@
 #include <raft/cudart_utils.h>
 #include <raft/distance/distance.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include "common/ml_benchmark.hpp"
+#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/fused_l2_nn.cu b/cpp/bench/prims/fused_l2_nn.cu
index f38cca1adb..f51ad76078 100644
--- a/cpp/bench/prims/fused_l2_nn.cu
+++ b/cpp/bench/prims/fused_l2_nn.cu
@@ -20,7 +20,7 @@
 #include <raft/linalg/norm.cuh>
 #include <raft/mr/device/allocator.hpp>
 #include <raft/random/rng.cuh>
-#include "common/ml_benchmark.hpp"
+#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/gram_matrix.cu b/cpp/bench/prims/gram_matrix.cu
index e925446b49..df9019dc46 100644
--- a/cpp/bench/prims/gram_matrix.cu
+++ b/cpp/bench/prims/gram_matrix.cu
@@ -23,7 +23,7 @@
 #include <sstream>
 #include <string>
 #include <vector>
-#include "common/ml_benchmark.hpp"
+#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/make_blobs.cu b/cpp/bench/prims/make_blobs.cu
index 346e6b162e..464993582b 100644
--- a/cpp/bench/prims/make_blobs.cu
+++ b/cpp/bench/prims/make_blobs.cu
@@ -16,7 +16,7 @@
 
 #include <raft/mr/device/allocator.hpp>
 #include <random/make_blobs.cuh>
-#include "common/ml_benchmark.hpp"
+#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/map_then_reduce.cu b/cpp/bench/prims/map_then_reduce.cu
index 8a1917d6d0..619151a21f 100644
--- a/cpp/bench/prims/map_then_reduce.cu
+++ b/cpp/bench/prims/map_then_reduce.cu
@@ -16,7 +16,7 @@
 
 #include <raft/linalg/map_then_reduce.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include "common/ml_benchmark.hpp"
+#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/matrix_vector_op.cu b/cpp/bench/prims/matrix_vector_op.cu
index 787deecfeb..b0a1ea257c 100644
--- a/cpp/bench/prims/matrix_vector_op.cu
+++ b/cpp/bench/prims/matrix_vector_op.cu
@@ -16,7 +16,7 @@
 
 #include <raft/linalg/matrix_vector_op.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include "common/ml_benchmark.hpp"
+#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/permute.cu b/cpp/bench/prims/permute.cu
index 6246aee8e7..58a8e68892 100644
--- a/cpp/bench/prims/permute.cu
+++ b/cpp/bench/prims/permute.cu
@@ -18,7 +18,7 @@
 #include <raft/mr/device/allocator.hpp>
 #include <raft/random/rng.cuh>
 #include <random/permute.cuh>
-#include "common/ml_benchmark.hpp"
+#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/reduce.cu b/cpp/bench/prims/reduce.cu
index 58cc4f2a7f..7c54f83722 100644
--- a/cpp/bench/prims/reduce.cu
+++ b/cpp/bench/prims/reduce.cu
@@ -16,7 +16,7 @@
 
 #include <raft/linalg/reduce.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include "common/ml_benchmark.hpp"
+#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/rng.cu b/cpp/bench/prims/rng.cu
index f90bf773d3..34bd09820a 100644
--- a/cpp/bench/prims/rng.cu
+++ b/cpp/bench/prims/rng.cu
@@ -17,7 +17,7 @@
 #include <raft/cudart_utils.h>
 #include <raft/mr/device/allocator.hpp>
 #include <raft/random/rng.cuh>
-#include "common/ml_benchmark.hpp"
+#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {

From c03fc65b61856fa98fdefc35747c849ab019b68d Mon Sep 17 00:00:00 2001
From: Robert Maynard <rmaynard@nvidia.com>
Date: Thu, 27 May 2021 16:54:10 -0400
Subject: [PATCH 09/11] Correct include order issues found by CI

---
 cpp/bench/prims/add.cu              | 2 +-
 cpp/bench/prims/distance_common.cuh | 2 +-
 cpp/bench/prims/fused_l2_nn.cu      | 4 ++--
 cpp/bench/prims/gram_matrix.cu      | 2 +-
 cpp/bench/prims/make_blobs.cu       | 2 +-
 cpp/bench/prims/map_then_reduce.cu  | 2 +-
 cpp/bench/prims/matrix_vector_op.cu | 2 +-
 cpp/bench/prims/permute.cu          | 2 +-
 cpp/bench/prims/reduce.cu           | 2 +-
 cpp/bench/prims/rng.cu              | 2 +-
 10 files changed, 11 insertions(+), 11 deletions(-)

diff --git a/cpp/bench/prims/add.cu b/cpp/bench/prims/add.cu
index b43021c7d1..c4dd15dd3d 100644
--- a/cpp/bench/prims/add.cu
+++ b/cpp/bench/prims/add.cu
@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
+#include <common/ml_benchmark.hpp>
 #include <raft/linalg/add.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/distance_common.cuh b/cpp/bench/prims/distance_common.cuh
index 3b86025754..7b2a579644 100644
--- a/cpp/bench/prims/distance_common.cuh
+++ b/cpp/bench/prims/distance_common.cuh
@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
+#include <common/ml_benchmark.hpp>
 #include <raft/cudart_utils.h>
 #include <raft/distance/distance.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/fused_l2_nn.cu b/cpp/bench/prims/fused_l2_nn.cu
index f51ad76078..a7145354a2 100644
--- a/cpp/bench/prims/fused_l2_nn.cu
+++ b/cpp/bench/prims/fused_l2_nn.cu
@@ -14,13 +14,13 @@
  * limitations under the License.
  */
 
-#include <raft/cudart_utils.h>
+#include <common/ml_benchmark.hpp>
 #include <limits>
+#include <raft/cudart_utils.h>
 #include <raft/distance/fused_l2_nn.cuh>
 #include <raft/linalg/norm.cuh>
 #include <raft/mr/device/allocator.hpp>
 #include <raft/random/rng.cuh>
-#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/gram_matrix.cu b/cpp/bench/prims/gram_matrix.cu
index df9019dc46..a86c08bc6a 100644
--- a/cpp/bench/prims/gram_matrix.cu
+++ b/cpp/bench/prims/gram_matrix.cu
@@ -14,6 +14,7 @@
  * limitations under the License.
  */
 
+#include <common/ml_benchmark.hpp>
 #include <cuml/matrix/kernelparams.h>
 #include <matrix/grammatrix.cuh>
 #include <matrix/kernelfactory.cuh>
@@ -23,7 +24,6 @@
 #include <sstream>
 #include <string>
 #include <vector>
-#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/make_blobs.cu b/cpp/bench/prims/make_blobs.cu
index 464993582b..3ba6e88235 100644
--- a/cpp/bench/prims/make_blobs.cu
+++ b/cpp/bench/prims/make_blobs.cu
@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
+#include <common/ml_benchmark.hpp>
 #include <raft/mr/device/allocator.hpp>
 #include <random/make_blobs.cuh>
-#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/map_then_reduce.cu b/cpp/bench/prims/map_then_reduce.cu
index 619151a21f..23b9790d24 100644
--- a/cpp/bench/prims/map_then_reduce.cu
+++ b/cpp/bench/prims/map_then_reduce.cu
@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
+#include <common/ml_benchmark.hpp>
 #include <raft/linalg/map_then_reduce.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/matrix_vector_op.cu b/cpp/bench/prims/matrix_vector_op.cu
index b0a1ea257c..bab86b504e 100644
--- a/cpp/bench/prims/matrix_vector_op.cu
+++ b/cpp/bench/prims/matrix_vector_op.cu
@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
+#include <common/ml_benchmark.hpp>
 #include <raft/linalg/matrix_vector_op.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/permute.cu b/cpp/bench/prims/permute.cu
index 58a8e68892..f6672d4137 100644
--- a/cpp/bench/prims/permute.cu
+++ b/cpp/bench/prims/permute.cu
@@ -14,11 +14,11 @@
  * limitations under the License.
  */
 
+#include <common/ml_benchmark.hpp>
 #include <raft/cudart_utils.h>
 #include <raft/mr/device/allocator.hpp>
 #include <raft/random/rng.cuh>
 #include <random/permute.cuh>
-#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/reduce.cu b/cpp/bench/prims/reduce.cu
index 7c54f83722..b016eeddb2 100644
--- a/cpp/bench/prims/reduce.cu
+++ b/cpp/bench/prims/reduce.cu
@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
+#include <common/ml_benchmark.hpp>
 #include <raft/linalg/reduce.cuh>
 #include <raft/mr/device/allocator.hpp>
-#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {
diff --git a/cpp/bench/prims/rng.cu b/cpp/bench/prims/rng.cu
index 34bd09820a..2fd7914a22 100644
--- a/cpp/bench/prims/rng.cu
+++ b/cpp/bench/prims/rng.cu
@@ -14,10 +14,10 @@
  * limitations under the License.
  */
 
+#include <common/ml_benchmark.hpp>
 #include <raft/cudart_utils.h>
 #include <raft/mr/device/allocator.hpp>
 #include <raft/random/rng.cuh>
-#include <common/ml_benchmark.hpp>
 
 namespace MLCommon {
 namespace Bench {

From a832efac8f577cefebf16869506833dc284bebe7 Mon Sep 17 00:00:00 2001
From: Robert Maynard <rmaynard@nvidia.com>
Date: Fri, 28 May 2021 08:03:18 -0400
Subject: [PATCH 10/11] Correct include order issues found by CI

---
 cpp/bench/prims/distance_common.cuh | 2 +-
 cpp/bench/prims/fused_l2_nn.cu      | 2 +-
 cpp/bench/prims/gram_matrix.cu      | 2 +-
 cpp/bench/prims/permute.cu          | 2 +-
 cpp/bench/prims/rng.cu              | 2 +-
 5 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/cpp/bench/prims/distance_common.cuh b/cpp/bench/prims/distance_common.cuh
index 7b2a579644..dbcb94da58 100644
--- a/cpp/bench/prims/distance_common.cuh
+++ b/cpp/bench/prims/distance_common.cuh
@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#include <common/ml_benchmark.hpp>
 #include <raft/cudart_utils.h>
+#include <common/ml_benchmark.hpp>
 #include <raft/distance/distance.cuh>
 #include <raft/mr/device/allocator.hpp>
 
diff --git a/cpp/bench/prims/fused_l2_nn.cu b/cpp/bench/prims/fused_l2_nn.cu
index a7145354a2..b5e5dde80b 100644
--- a/cpp/bench/prims/fused_l2_nn.cu
+++ b/cpp/bench/prims/fused_l2_nn.cu
@@ -14,9 +14,9 @@
  * limitations under the License.
  */
 
+#include <raft/cudart_utils.h>
 #include <common/ml_benchmark.hpp>
 #include <limits>
-#include <raft/cudart_utils.h>
 #include <raft/distance/fused_l2_nn.cuh>
 #include <raft/linalg/norm.cuh>
 #include <raft/mr/device/allocator.hpp>
diff --git a/cpp/bench/prims/gram_matrix.cu b/cpp/bench/prims/gram_matrix.cu
index a86c08bc6a..46ef9c94e9 100644
--- a/cpp/bench/prims/gram_matrix.cu
+++ b/cpp/bench/prims/gram_matrix.cu
@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#include <common/ml_benchmark.hpp>
 #include <cuml/matrix/kernelparams.h>
+#include <common/ml_benchmark.hpp>
 #include <matrix/grammatrix.cuh>
 #include <matrix/kernelfactory.cuh>
 #include <memory>
diff --git a/cpp/bench/prims/permute.cu b/cpp/bench/prims/permute.cu
index f6672d4137..a7e85db46a 100644
--- a/cpp/bench/prims/permute.cu
+++ b/cpp/bench/prims/permute.cu
@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#include <common/ml_benchmark.hpp>
 #include <raft/cudart_utils.h>
+#include <common/ml_benchmark.hpp>
 #include <raft/mr/device/allocator.hpp>
 #include <raft/random/rng.cuh>
 #include <random/permute.cuh>
diff --git a/cpp/bench/prims/rng.cu b/cpp/bench/prims/rng.cu
index 2fd7914a22..22cd35abd0 100644
--- a/cpp/bench/prims/rng.cu
+++ b/cpp/bench/prims/rng.cu
@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#include <common/ml_benchmark.hpp>
 #include <raft/cudart_utils.h>
+#include <common/ml_benchmark.hpp>
 #include <raft/mr/device/allocator.hpp>
 #include <raft/random/rng.cuh>
 

From f565ff3157f6504248f57dd04597d9cf3573884c Mon Sep 17 00:00:00 2001
From: Robert Maynard <rmaynard@nvidia.com>
Date: Tue, 1 Jun 2021 16:39:20 -0400
Subject: [PATCH 11/11] Correct missing cuml dependency of prims benchmarks

---
 cpp/bench/CMakeLists.txt | 1 +
 1 file changed, 1 insertion(+)

diff --git a/cpp/bench/CMakeLists.txt b/cpp/bench/CMakeLists.txt
index 08d7c1bcf8..8d8986abfd 100644
--- a/cpp/bench/CMakeLists.txt
+++ b/cpp/bench/CMakeLists.txt
@@ -81,6 +81,7 @@ if(BUILD_CUML_PRIMS_BENCH)
 
   target_link_libraries(${PRIMS_BENCH_TARGET}
     PUBLIC
+      cuml
       CUDA::cublas
       benchmark::benchmark
       raft::raft