Move CUDA-specific occupancy calculation into proper file

-Maintain functionality, just move CalculateOccupancy() and CompareOccupancy() methods from device_description to cuda_gpu_executor
-Remove CUDA requirement in general class device_description

4 files changed, 48 insertions, 49 deletions

diff --git a/tensorflow/stream_executor/cuda/cuda_gpu_executor.cc b/tensorflow/stream_executor/cuda/cuda_gpu_executor.cc
index 39b0696c93..458c0e3030 100644
--- a/tensorflow/stream_executor/cuda/cuda_gpu_executor.cc
+++ b/tensorflow/stream_executor/cuda/cuda_gpu_executor.cc
@@ -490,6 +490,43 @@ void CUDAExecutor::VlogOccupancyInfo(const KernelBase &kernel,
   }
 }
 
+// Compute and return maximum blocks per core (occupancy) based on the
+// device description, some kernel characteristics and the number of threads per
+// block.  If unable to compute occupancy, zero is returned.
+int CalculateOccupancy(const DeviceDescription& device_description,
+                       uint64 registers_per_thread,
+                       uint64 shared_memory_per_block,
+                       const ThreadDim& thread_dims, CUfunction func) {
+  int suggested_blocks = 0;
+  int suggested_threads = 0;
+  CUresult err =
+      cuOccupancyMaxPotentialBlockSize(&suggested_blocks, &suggested_threads,
+                                       func, NULL, shared_memory_per_block, 0);
+  CHECK_EQ(err, CUDA_SUCCESS);
+  return suggested_blocks;
+}
+
+// Compute and return the suggested thread count to acheive ideal occupancy.
+// If the provided thread dimensions match this number, zero is returned.
+int CompareOccupancy(int* initial_blocks,
+                     const DeviceDescription& device_description,
+                     uint64 registers_per_thread,
+                     uint64 shared_memory_per_block,
+                     const ThreadDim& thread_dims, CUfunction func) {
+  int suggested_blocks = 0;
+  int suggested_threads = 0;
+  CUresult err =
+      cuOccupancyMaxPotentialBlockSize(&suggested_blocks, &suggested_threads,
+                                       func, NULL, shared_memory_per_block, 0);
+  CHECK_EQ(err, CUDA_SUCCESS);
+  if (suggested_blocks > *initial_blocks) {
+    *initial_blocks = suggested_blocks;
+    return suggested_threads;
+  } else {
+    return 0;
+  }
+}
+
 void *CUDAExecutor::Allocate(uint64 size) {
   return CUDADriver::DeviceAllocate(context_, size);
 }
diff --git a/tensorflow/stream_executor/cuda/cuda_gpu_executor.h b/tensorflow/stream_executor/cuda/cuda_gpu_executor.h
index 8a954d5461..e8ebbc3220 100644
--- a/tensorflow/stream_executor/cuda/cuda_gpu_executor.h
+++ b/tensorflow/stream_executor/cuda/cuda_gpu_executor.h
@@ -70,6 +70,17 @@ class CUDAExecutor : public internal::StreamExecutorInterface {
               const BlockDim &block_dims, const KernelBase &k,
               const KernelArgsArrayBase &args) override;
 
+  int CalculateOccupancy(const DeviceDescription& device_description,
+                       uint64 registers_per_thread,
+                       uint64 shared_memory_per_block,
+                       const ThreadDim& thread_dims, CUfunction func);
+
+  int CompareOccupancy(int* initial_blocks,
+                     const DeviceDescription& device_description,
+                     uint64 registers_per_thread,
+                     uint64 shared_memory_per_block,
+                     const ThreadDim& thread_dims, CUfunction func);
+
   void *Allocate(uint64 size) override;
 
   void *AllocateSubBuffer(DeviceMemoryBase *mem, uint64 offset_bytes,
diff --git a/tensorflow/stream_executor/device_description.cc b/tensorflow/stream_executor/device_description.cc
index df52ce6cce..726c4adf74 100644
--- a/tensorflow/stream_executor/device_description.cc
+++ b/tensorflow/stream_executor/device_description.cc
@@ -157,36 +157,4 @@ static uint64 RoundDown(uint64 value, uint64 n) {
   return port::MathUtil::FloorOfRatio(value, n) * n;
 }
 
-int CalculateOccupancy(const DeviceDescription& device_description,
-                       uint64 registers_per_thread,
-                       uint64 shared_memory_per_block,
-                       const ThreadDim& thread_dims, CUfunction func) {
-  int suggested_blocks = 0;
-  int suggested_threads = 0;
-  CUresult err =
-      cuOccupancyMaxPotentialBlockSize(&suggested_blocks, &suggested_threads,
-                                       func, NULL, shared_memory_per_block, 0);
-  CHECK_EQ(err, CUDA_SUCCESS);
-  return suggested_blocks;
-}
-
-int CompareOccupancy(int* initial_blocks,
-                     const DeviceDescription& device_description,
-                     uint64 registers_per_thread,
-                     uint64 shared_memory_per_block,
-                     const ThreadDim& thread_dims, CUfunction func) {
-  int suggested_blocks = 0;
-  int suggested_threads = 0;
-  CUresult err =
-      cuOccupancyMaxPotentialBlockSize(&suggested_blocks, &suggested_threads,
-                                       func, NULL, shared_memory_per_block, 0);
-  CHECK_EQ(err, CUDA_SUCCESS);
-  if (suggested_blocks > *initial_blocks) {
-    *initial_blocks = suggested_blocks;
-    return suggested_threads;
-  } else {
-    return 0;
-  }
-}
-
 }  // namespace stream_executor
diff --git a/tensorflow/stream_executor/device_description.h b/tensorflow/stream_executor/device_description.h
index d335b9b875..b15ce31216 100644
--- a/tensorflow/stream_executor/device_description.h
+++ b/tensorflow/stream_executor/device_description.h
@@ -24,7 +24,6 @@ limitations under the License.
 #include <memory>
 #include "tensorflow/stream_executor/platform/port.h"
 
-#include "tensorflow/stream_executor/cuda/cuda_driver.h"
 #include "tensorflow/stream_executor/launch_dim.h"
 #include "tensorflow/stream_executor/platform/port.h"
 
@@ -324,22 +323,6 @@ void CalculateDimensionality(const DeviceDescription &device_description,
                              uint64 element_count, uint64 *threads_per_block,
                              uint64 *block_count);
 
-// Compute and return maximum blocks per core (occupancy) based on the
-// device description, some kernel characteristics and the number of threads per
-// block.  If unable to compute occupancy, zero is returned.
-int CalculateOccupancy(const DeviceDescription& device_description,
-                       uint64 registers_per_thread,
-                       uint64 shared_memory_per_block,
-                       const ThreadDim& thread_dims, CUfunction func);
-
-// Compute and return the suggested thread count to acheive ideal occupancy.
-// If the provided thread dimensions match this number, zero is returned.
-int CompareOccupancy(int* initial_blocks,
-                     const DeviceDescription& device_description,
-                     uint64 registers_per_thread,
-                     uint64 shared_memory_per_block,
-                     const ThreadDim& thread_dims, CUfunction func);
-
 }  // namespace stream_executor
 
 #endif  // TENSORFLOW_STREAM_EXECUTOR_DEVICE_DESCRIPTION_H_