From b3804505e202700f1403c4076e5e25f2e1864716 Mon Sep 17 00:00:00 2001
From: Tao Li <litaotju@live.cn>
Date: Sat, 18 Jan 2020 16:50:05 +0800
Subject: [PATCH] Fix mshadow broken basic_stream example. The NewTensor and
 NewStream API signature has changed the sample was not updated.

---
 3rdparty/mshadow/guide/Makefile        |  5 +++++
 3rdparty/mshadow/guide/basic_stream.cu | 27 ++++++++++++++------------
 3rdparty/mshadow/make/mshadow.mk       |  1 +
 3 files changed, 21 insertions(+), 12 deletions(-)

diff --git a/3rdparty/mshadow/guide/Makefile b/3rdparty/mshadow/guide/Makefile
index bad7a8e94b1d..38d153e77cd7 100644
--- a/3rdparty/mshadow/guide/Makefile
+++ b/3rdparty/mshadow/guide/Makefile
@@ -12,7 +12,12 @@ export NVCCFLAGS = -O3 --use_fast_math -ccbin $(CXX) $(MSHADOW_NVCCFLAGS)
 BIN = basic defop
 OBJ =
 CUOBJ =
+ifeq ($(USE_CUDA), 1)
+CUBIN = basic_stream
+else
 CUBIN =
+endif
+
 .PHONY: clean all
 
 all: $(BIN) $(OBJ) $(CUBIN) $(CUOBJ)
diff --git a/3rdparty/mshadow/guide/basic_stream.cu b/3rdparty/mshadow/guide/basic_stream.cu
index 930fc54710f7..d7dda7564787 100644
--- a/3rdparty/mshadow/guide/basic_stream.cu
+++ b/3rdparty/mshadow/guide/basic_stream.cu
@@ -9,20 +9,22 @@ int main(void) {
   // intialize tensor engine before using tensor operation, needed for CuBLAS
   InitTensorEngine<gpu>();
   // create a 2 x 5 tensor, from existing space
-  Stream<gpu> *sm1 = NewStream<gpu>();
-  Stream<gpu> *sm2 = NewStream<gpu>();
-  Tensor<gpu, 2, float> ts1 = NewTensor<gpu, float>(Shape2(2, 5), 0.0f, sm1);
-  Tensor<gpu, 2, float> ts2 = NewTensor<gpu, float>(Shape2(2, 5), 0.0f, sm2);
-  ts1 = 1; // Should use stream 0.
-  ts2 = 2; // Should use stream 1. Can run in parallel with stream 0.
-  Tensor<gpu, 2> res = NewTensor<gpu, float>(Shape2(2, 2), 0.0f);
-  res.stream_ = NewStream<gpu>();
-  res = dot(ts1, ts2.T()); //Should use stream 2.
+  Stream<gpu> *sm1 = NewStream<gpu>(0);
+  Stream<gpu> *sm2 = NewStream<gpu>(0);
+  Tensor<gpu, 2, float> ts1 =
+      NewTensor<gpu, float>(Shape2(2, 5), 0.0f, false, sm1);
+  Tensor<gpu, 2, float> ts2 =
+      NewTensor<gpu, float>(Shape2(2, 5), 0.0f, false, sm2);
+  ts1 = 1; // Should use stream 1.
+  ts2 = 2; // Should use stream 2. Can run in parallel with stream 1.
+  Stream<gpu> *sm3 = NewStream<gpu>(0);
+  Tensor<gpu, 2> res = NewTensor<gpu, float>(Shape2(2, 2), 0.0f, false, sm3);
+  res = dot(ts1, ts2.T()); // Should use stream 3.
 
   Tensor<cpu, 2> cpu_res = NewTensor<cpu, float>(Shape2(2, 2), 0.0f);
-  Copy(cpu_res, res); // default stream, should be 0.
-  for (index_t i = 0; i < cpu_res.size(0); ++i){
-    for (index_t j = 0; j < cpu_res.size(1); ++j){
+  Copy(cpu_res, res, sm3);
+  for (index_t i = 0; i < cpu_res.size(0); ++i) {
+    for (index_t j = 0; j < cpu_res.size(1); ++j) {
       printf("%.2f ", cpu_res[i][j]);
     }
     printf("\n");
@@ -30,6 +32,7 @@ int main(void) {
   // shutdown tensor enigne after usage
   DeleteStream(sm1);
   DeleteStream(sm2);
+  DeleteStream(sm3);
   ShutdownTensorEngine<gpu>();
   return 0;
 }
diff --git a/3rdparty/mshadow/make/mshadow.mk b/3rdparty/mshadow/make/mshadow.mk
index 86155eaaadcf..3694772cd35c 100644
--- a/3rdparty/mshadow/make/mshadow.mk
+++ b/3rdparty/mshadow/make/mshadow.mk
@@ -64,6 +64,7 @@ endif
 ifeq ($(USE_CUDA), 0)
 	MSHADOW_CFLAGS += -DMSHADOW_USE_CUDA=0
 else
+	MSHADOW_CFLAGS += -DMSHADOW_USE_CUDA=1
 	MSHADOW_LDFLAGS += -lcudart -lcublas -lcurand -lcusolver
 endif
 ifneq ($(USE_CUDA_PATH), NONE)