add a compiler flag to use int64 as tensor size

.gitmodules

@@ -1,6 +1,6 @@
 [submodule "3rdparty/mshadow"]
 	path = 3rdparty/mshadow
-	url = https://github.com/dmlc/mshadow.git
+	url = https://github.com/apeforest/mshadow.git
 [submodule "3rdparty/dmlc-core"]
 	path = 3rdparty/dmlc-core
 	url = https://github.com/dmlc/dmlc-core.git

Makefile

@@ -188,6 +188,11 @@ ifeq ($(USE_OPERATOR_TUNING), 1)
 	CFLAGS += -DMXNET_USE_OPERATOR_TUNING=1
 endif
 
+ifeq ($(USE_INT64_TENSOR_SIZE), 1)
+   CFLAGS += -DMSHADOW_INT64_TENSOR_SIZE=1
+else
+   CFLAGS += -DMSHADOW_INT64_TENSOR_SIZE=0
+endif
 # verify existence of separate lapack library when using blas/openblas/atlas
 # switch off lapack support in case it can't be found
 # issue covered with this

make/config.mk

@@ -214,6 +214,10 @@ EXTRA_OPERATORS =
 # Create C++ interface package
 USE_CPP_PACKAGE = 0
 
+# Use int64_t type for tensor size.
+# This will cause performance degradation reported in issue #14496
+USE_INT64_TENSOR_SIZE = 0
+
 #----------------------------
 # plugins
 #----------------------------

make/crosscompile.jetson.mk

@@ -191,6 +191,10 @@ EXTRA_OPERATORS =
 # Create C++ interface package
 USE_CPP_PACKAGE = 0
 
+# Use int64_t type for tensor size.
+# This will cause performance degradation reported in issue #14496
+USE_INT64_TENSOR_SIZE = 0
+
 #----------------------------
 # plugins
 #----------------------------

make/osx.mk

@@ -135,6 +135,10 @@ EXTRA_OPERATORS =
 # Create C++ interface package
 USE_CPP_PACKAGE = 0
 
+# Use int64_t type for tensor size.
+# This will cause performance degradation reported in issue #14496
+USE_INT64_TENSOR_SIZE = 0
+
 #----------------------------
 # plugins
 #----------------------------

src/operator/convolution_v1-inl.h

@@ -336,7 +336,7 @@ class ConvolutionV1Op : public Operator {
     // param_.workspace is in elements of sizeof(DType)
     // if param_.workspace is set to zero the nstep_ equals ishape[0] (batch)
     nstep_ = std::max<index_t>(
-        std::min(static_cast<index_t>(param_.workspace) /
+        std::min<index_t>(param_.workspace /
           (shape_colunit_.Size() + shape_dstunit_.Size()), ishape[0]),
       1);
 

src/operator/nn/deconvolution-inl.h

@@ -460,7 +460,7 @@ class DeconvolutionOp {
                                      oshape[2] * oshape[3]);
     // See convolution for workspace calculations. nstep_ will be the effective batch size
     nstep_ = std::max<index_t>(
-        std::min(static_cast<index_t>(param_.workspace) /
+        std::min<index_t>(param_.workspace /
           (shape_colunit_.Size() + shape_dstunit_.Size()), ishape[0]),
       1);