Merge pull request #20706 from ROCmSoftwarePlatform:upstream-staging-stream-executor-pooling-interface

PiperOrigin-RevId: 204805678

2 files changed, 26 insertions, 14 deletions

diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc
index 84916385a8..9e24a4538c 100644
--- a/tensorflow/stream_executor/cuda/cuda_dnn.cc
+++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc
@@ -3603,7 +3603,7 @@ bool CudnnSupport::DoPoolForward(
     const dnn::BatchDescriptor& input_dimensions,
     const DeviceMemory<double>& input_data,
     const dnn::BatchDescriptor& output_dimensions,
-    DeviceMemory<double>* output_data) {
+    DeviceMemory<double>* output_data, ScratchAllocator* workspace_allocator) {
   // Alpha is the scaling factor for input.
   double alpha = 1.0;
   // Beta is the scaling factor for output.
@@ -3628,7 +3628,7 @@ bool CudnnSupport::DoPoolForward(
     const dnn::BatchDescriptor& input_dimensions,
     const DeviceMemory<float>& input_data,
     const dnn::BatchDescriptor& output_dimensions,
-    DeviceMemory<float>* output_data) {
+    DeviceMemory<float>* output_data, ScratchAllocator* workspace_allocator) {
   // Alpha is the scaling factor for input.
   float alpha = 1.0;
   // Beta is the scaling factor for output.
@@ -3653,7 +3653,8 @@ bool CudnnSupport::DoPoolForward(
     const dnn::BatchDescriptor& input_dimensions,
     const DeviceMemory<Eigen::half>& input_data,
     const dnn::BatchDescriptor& output_dimensions,
-    DeviceMemory<Eigen::half>* output_data) {
+    DeviceMemory<Eigen::half>* output_data,
+    ScratchAllocator* workspace_allocator) {
   // Alpha is the scaling factor for input.
   float alpha = 1.0;
   // Beta is the scaling factor for output.
@@ -3679,7 +3680,8 @@ bool CudnnSupport::DoPoolBackward(
     const dnn::BatchDescriptor& output_dimensions,
     const DeviceMemory<double>& output_data,
     const DeviceMemory<double>& input_diff_data,
-    DeviceMemory<double>* output_diff_data) {
+    DeviceMemory<double>* output_diff_data,
+    ScratchAllocator* workspace_allocator) {
   // Alpha is the scaling factor for input.
   double alpha = 1.0;
   // Beta is the scaling factor for output.
@@ -3708,7 +3710,8 @@ bool CudnnSupport::DoPoolBackward(
     const dnn::BatchDescriptor& output_dimensions,
     const DeviceMemory<float>& output_data,
     const DeviceMemory<float>& input_diff_data,
-    DeviceMemory<float>* output_diff_data) {
+    DeviceMemory<float>* output_diff_data,
+    ScratchAllocator* workspace_allocator) {
   // Alpha is the scaling factor for input.
   float alpha = 1.0;
   // Beta is the scaling factor for output.
@@ -3737,7 +3740,8 @@ bool CudnnSupport::DoPoolBackward(
     const dnn::BatchDescriptor& output_dimensions,
     const DeviceMemory<Eigen::half>& output_data,
     const DeviceMemory<Eigen::half>& input_diff_data,
-    DeviceMemory<Eigen::half>* output_diff_data) {
+    DeviceMemory<Eigen::half>* output_diff_data,
+    ScratchAllocator* workspace_allocator) {
   // Alpha is the scaling factor for input.
   float alpha = 1.0;
   // Beta is the scaling factor for output.
@@ -3806,7 +3810,8 @@ bool CudnnSupport::DoNormalizeBackwardWithDimensions(
     const dnn::BatchDescriptor& dimensions, const DeviceMemory<float>& raw_data,
     const DeviceMemory<float>& normalized_data,
     const DeviceMemory<float>& normalized_variable_gradient,
-    DeviceMemory<float>* raw_variable_gradient) {
+    DeviceMemory<float>* raw_variable_gradient,
+    ScratchAllocator* workspace_allocator) {
   // Check for unsupported modes.
   if (normalize_descriptor.wrap_around()) {
     LOG(ERROR) << "CUDA LRN does not support cudnn-around mode";
diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.h b/tensorflow/stream_executor/cuda/cuda_dnn.h
index c924d41cb5..9d88f971bb 100644
--- a/tensorflow/stream_executor/cuda/cuda_dnn.h
+++ b/tensorflow/stream_executor/cuda/cuda_dnn.h
@@ -515,21 +515,24 @@ class CudnnSupport : public dnn::DnnSupport {
                      const dnn::BatchDescriptor& input_dimensions,
                      const DeviceMemory<double>& input_data,
                      const dnn::BatchDescriptor& output_dimensions,
-                     DeviceMemory<double>* output_data) override;
+                     DeviceMemory<double>* output_data,
+                     ScratchAllocator* workspace_allocator) override;
 
   bool DoPoolForward(Stream* stream,
                      const dnn::PoolingDescriptor& pooling_dimensions,
                      const dnn::BatchDescriptor& input_dimensions,
                      const DeviceMemory<float>& input_data,
                      const dnn::BatchDescriptor& output_dimensions,
-                     DeviceMemory<float>* output_data) override;
+                     DeviceMemory<float>* output_data,
+                     ScratchAllocator* workspace_allocator) override;
 
   bool DoPoolForward(Stream* stream,
                      const dnn::PoolingDescriptor& pooling_dimensions,
                      const dnn::BatchDescriptor& input_dimensions,
                      const DeviceMemory<Eigen::half>& input_data,
                      const dnn::BatchDescriptor& output_dimensions,
-                     DeviceMemory<Eigen::half>* output_data) override;
+                     DeviceMemory<Eigen::half>* output_data,
+                     ScratchAllocator* workspace_allocator) override;
 
   bool DoPoolBackward(Stream* stream,
                       const dnn::PoolingDescriptor& pooling_dimensions,
@@ -538,7 +541,8 @@ class CudnnSupport : public dnn::DnnSupport {
                       const dnn::BatchDescriptor& output_dimensions,
                       const DeviceMemory<double>& output_data,
                       const DeviceMemory<double>& input_diff_data,
-                      DeviceMemory<double>* output_diff_data) override;
+                      DeviceMemory<double>* output_diff_data,
+                      ScratchAllocator* workspace_allocator) override;
 
   bool DoPoolBackward(Stream* stream,
                       const dnn::PoolingDescriptor& pooling_dimensions,
@@ -547,7 +551,8 @@ class CudnnSupport : public dnn::DnnSupport {
                       const dnn::BatchDescriptor& output_dimensions,
                       const DeviceMemory<float>& output_data,
                       const DeviceMemory<float>& input_diff_data,
-                      DeviceMemory<float>* output_diff_data) override;
+                      DeviceMemory<float>* output_diff_data,
+                      ScratchAllocator* workspace_allocator) override;
 
   bool DoPoolBackward(Stream* stream,
                       const dnn::PoolingDescriptor& pooling_dimensions,
@@ -556,7 +561,8 @@ class CudnnSupport : public dnn::DnnSupport {
                       const dnn::BatchDescriptor& output_dimensions,
                       const DeviceMemory<Eigen::half>& output_data,
                       const DeviceMemory<Eigen::half>& input_diff_data,
-                      DeviceMemory<Eigen::half>* output_diff_data) override;
+                      DeviceMemory<Eigen::half>* output_diff_data,
+                      ScratchAllocator* workspace_allocator) override;
 
   bool DoNormalize(Stream* stream,
                    const dnn::NormalizeDescriptor& normalize_descriptor,
@@ -575,7 +581,8 @@ class CudnnSupport : public dnn::DnnSupport {
       const DeviceMemory<float>& raw_data,
       const DeviceMemory<float>& normalized_data,
       const DeviceMemory<float>& normalized_variable_gradient,
-      DeviceMemory<float>* raw_variable_gradient) override;
+      DeviceMemory<float>* raw_variable_gradient,
+      ScratchAllocator* workspace_allocator) override;
 
   bool DoDepthConcatenate(
       Stream* stream, port::ArraySlice<dnn::BatchDescriptor> input_dimensions,