Bugfix: memsetAsync uses wrong default stream

Signed-off-by: Peter Han <fujun.han@iluvatar.ai>
pdh930105 · Mar 23, 2021 · 92393b2 · 92393b2
1 parent 50bf00e
commit 92393b2
Show file tree

Hide file tree

Showing 8 changed files with 16 additions and 16 deletions.
diff --git a/include/cutlass/conv/device/implicit_gemm_convolution.h b/include/cutlass/conv/device/implicit_gemm_convolution.h
@@ -244,7 +244,7 @@ class ImplicitGemmConvolution {
     void *workspace = nullptr, 
     cudaStream_t stream = nullptr) {
 
-    Status status = initialize(args, workspace);
+    Status status = initialize(args, workspace, stream);
 
     if (status == Status::kSuccess) {
       status = run(stream);

diff --git a/include/cutlass/gemm/device/gemm.h b/include/cutlass/gemm/device/gemm.h
@@ -482,7 +482,7 @@ class Gemm {
     void *workspace = nullptr, 
     cudaStream_t stream = nullptr) {
 
-    Status status = initialize(args, workspace);
+    Status status = initialize(args, workspace, stream);
 
     if (status == Status::kSuccess) {
       status = run(stream);
@@ -673,7 +673,7 @@ class Gemm<ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_,
   /// Initializes GEMM state from arguments.
   Status initialize(Arguments const &args, void *workspace = nullptr, cudaStream_t stream = nullptr) {
 
-    return underlying_operator_.initialize(to_underlying_arguments(args), workspace);
+    return underlying_operator_.initialize(to_underlying_arguments(args), workspace, stream);
   }
 
   /// Lightweight update given a subset of arguments
@@ -699,7 +699,7 @@ class Gemm<ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_,
     void *workspace = nullptr, 
     cudaStream_t stream = nullptr) {
 
-    Status status = initialize(args, workspace);
+    Status status = initialize(args, workspace, stream);
 
     if (status == Status::kSuccess) {
       status = run(stream);

diff --git a/include/cutlass/gemm/device/gemm_array.h b/include/cutlass/gemm/device/gemm_array.h
@@ -473,7 +473,7 @@ class GemmArray {
     void *workspace = nullptr, 
     cudaStream_t stream = nullptr) {
 
-    Status status = initialize(args, workspace);
+    Status status = initialize(args, workspace, stream);
 
     if (status == Status::kSuccess) {
       status = run(stream);
@@ -700,7 +700,7 @@ class GemmArray<
   /// Initializes GEMM state from arguments.
   Status initialize(Arguments const &args, void *workspace = nullptr, cudaStream_t stream = nullptr) {
 
-    return underlying_operator_.initialize(to_underlying_arguments(args), workspace);
+    return underlying_operator_.initialize(to_underlying_arguments(args), workspace, stream);
   }
 
   /// Lightweight update given a subset of arguments
@@ -726,7 +726,7 @@ class GemmArray<
     void *workspace = nullptr, 
     cudaStream_t stream = nullptr) {
 
-    Status status = initialize(args, workspace);
+    Status status = initialize(args, workspace, stream);
 
     if (status == Status::kSuccess) {
       status = run(stream);

diff --git a/include/cutlass/gemm/device/gemm_batched.h b/include/cutlass/gemm/device/gemm_batched.h
@@ -451,7 +451,7 @@ class GemmBatched {
     void *workspace = nullptr, 
     cudaStream_t stream = nullptr) {
 
-    Status status = initialize(args, workspace);
+    Status status = initialize(args, workspace, stream);
 
     if (status == Status::kSuccess) {
       status = run(stream);
@@ -666,7 +666,7 @@ class GemmBatched<
   /// Initializes GEMM state from arguments.
   Status initialize(Arguments const &args, void *workspace = nullptr, cudaStream_t stream = nullptr) {
 
-    return underlying_operator_.initialize(to_underlying_arguments(args), workspace);
+    return underlying_operator_.initialize(to_underlying_arguments(args), workspace, stream);
   }
 
   /// Lightweight update given a subset of arguments
@@ -692,7 +692,7 @@ class GemmBatched<
     void *workspace = nullptr, 
     cudaStream_t stream = nullptr) {
 
-    Status status = initialize(args, workspace);
+    Status status = initialize(args, workspace, stream);
 
     if (status == Status::kSuccess) {
       status = run(stream);

diff --git a/include/cutlass/gemm/device/gemm_complex.h b/include/cutlass/gemm/device/gemm_complex.h
@@ -465,7 +465,7 @@ class GemmComplex {
     void *workspace = nullptr, 
     cudaStream_t stream = nullptr) {
 
-    Status status = initialize(args, workspace);
+    Status status = initialize(args, workspace, stream);
 
     if (status == Status::kSuccess) {
       status = run(stream);
@@ -674,7 +674,7 @@ class GemmComplex<
   /// Initializes GEMM state from arguments.
   Status initialize(Arguments const &args, void *workspace = nullptr, cudaStream_t stream = nullptr) {
 
-    return underlying_operator_.initialize(to_underlying_arguments(args), workspace);
+    return underlying_operator_.initialize(to_underlying_arguments(args), workspace, stream);
   }
 
   /// Lightweight update given a subset of arguments
@@ -700,7 +700,7 @@ class GemmComplex<
     void *workspace = nullptr, 
     cudaStream_t stream = nullptr) {
 
-    Status status = initialize(args, workspace);
+    Status status = initialize(args, workspace, stream);
 
     if (status == Status::kSuccess) {
       status = run(stream);

diff --git a/include/cutlass/gemm/device/gemm_sparse.h b/include/cutlass/gemm/device/gemm_sparse.h
@@ -498,7 +498,7 @@ class SparseGemm {
     void *workspace = nullptr, 
     cudaStream_t stream = nullptr) {
 
-    Status status = initialize(args, workspace);
+    Status status = initialize(args, workspace, stream);
 
     if (status == Status::kSuccess) {
       status = run(stream);

diff --git a/include/cutlass/gemm/device/gemm_universal_base.h b/include/cutlass/gemm/device/gemm_universal_base.h
@@ -347,7 +347,7 @@ class GemmUniversalBase {
       return Status::kErrorWorkspaceNull;
     }
 
-    params_.update(args, workspace);
+    params_.update(args, workspace, stream);
 
     return Status::kSuccess;
   }

diff --git a/include/cutlass/reduction/device/reduce_split_k.h b/include/cutlass/reduction/device/reduce_split_k.h
@@ -197,7 +197,7 @@ class ReduceSplitK {
     void *workspace = nullptr, 
     cudaStream_t stream = nullptr) {
 
-    Status status = initialize(args, workspace);
+    Status status = initialize(args, workspace, stream);
 
     if (status == Status::kSuccess) {
       status = run(stream);