Fall back when sparse arrays are passed to MKLDNN-enabled operators (a…

…pache#11664) * softmax_fallbach * Fallback Amend This is the final rectify for fallback problem(functions call) * Lint amend * test_try * Patch for test fail * Pooling amend * Delete non_rectified_operation_test * fallback_normal * Fixed_dispatch * activation-amend * activation second * activation backward * activate_try * activation_debug * Act change. * test_random * mkldnn choice * format_modify * rebase
anirudh2290 · Sep 19, 2018 · 9e942d5 · 9e942d5
1 parent 4f0743f
commit 9e942d5
Show file tree

Hide file tree

Showing 9 changed files with 198 additions and 165 deletions.
diff --git a/src/operator/nn/activation.cc b/src/operator/nn/activation.cc
@@ -31,6 +31,8 @@
 #include "./mkldnn/mkldnn_base-inl.h"
 #include "./mkldnn/mkldnn_ops-inl.h"
 #endif  // MXNET_USE_MKLDNN
+#include "../operator_common.h"
+#include "../../common/utils.h"
 
 namespace mxnet {
 namespace op {
@@ -101,74 +103,35 @@ void ActivationGradComputeExCPU(const nnvm::NodeAttrs& attrs,
 }
 #endif
 
+#if MXNET_USE_MKLDNN == 1
 inline static bool ActivationStorageType(const nnvm::NodeAttrs& attrs,
                                          const int dev_mask,
                                          DispatchMode* dispatch_mode,
                                          std::vector<int> *in_attrs,
                                          std::vector<int> *out_attrs) {
   CHECK_EQ(in_attrs->size(), 1);
   CHECK_EQ(out_attrs->size(), 1);
-  bool ret = ElemwiseStorageType<1, 1, false, false, false>(attrs, dev_mask,
-                                                            dispatch_mode,
-                                                            in_attrs, out_attrs);
-#if MXNET_USE_MKLDNN == 1
   const ActivationParam& param = nnvm::get<ActivationParam>(attrs.parsed);
-  if (dev_mask == mshadow::cpu::kDevMask && SupportMKLDNNAct(param)) {
-    *dispatch_mode = DispatchMode::kFComputeEx;
-  }
-  if (dev_mask == mshadow::cpu::kDevMask && !MKLDNNEnvSet()) {
-    *dispatch_mode = DispatchMode::kFComputeFallback;
-    return ret;
-  }
-#endif
-  return ret;
+  return MKLDNNStorageType(attrs, dev_mask, SupportMKLDNNAct(param),
+                           dispatch_mode, in_attrs, out_attrs);
 }
 
 inline static bool BackwardActStorageType(const nnvm::NodeAttrs& attrs,
                                           const int dev_mask,
                                           DispatchMode* dispatch_mode,
                                           std::vector<int> *in_attrs,
                                           std::vector<int> *out_attrs) {
-  bool ret = false;
   const ActivationParam& param = nnvm::get<ActivationParam>(attrs.parsed);
-#if (MXNET_USE_CUDNN == 1 || MXNET_USE_MKLDNN == 1)
   if (param.act_type != activation::kReLU) {
     CHECK_EQ(in_attrs->size(), 3U);
-    ret = ElemwiseStorageType<3, 1, false, false, false>(attrs, dev_mask,
-                                                         dispatch_mode,
-                                                         in_attrs, out_attrs);
   } else {
     // for ReLU activation, the backward pass only needs ograd and output
     CHECK_EQ(in_attrs->size(), 2U);
-    ret = ElemwiseStorageType<2, 1, false, false, false>(attrs, dev_mask,
-                                                         dispatch_mode,
-                                                         in_attrs, out_attrs);
-  }
-#else
-  if (param.act_type == activation::kSoftSign) {
-    CHECK_EQ(in_attrs->size(), 3U);
-    ret = ElemwiseStorageType<3, 1, false, false, false>(attrs, dev_mask,
-                                                         dispatch_mode,
-                                                         in_attrs, out_attrs);
-  } else {
-    CHECK_EQ(in_attrs->size(), 2U);
-    ret = ElemwiseStorageType<2, 1, false, false, false>(attrs, dev_mask,
-                                                         dispatch_mode,
-                                                         in_attrs, out_attrs);
   }
-#endif
-  CHECK_EQ(out_attrs->size(), 1U);
-#if MXNET_USE_MKLDNN == 1
-  if (dev_mask == mshadow::cpu::kDevMask && SupportMKLDNNAct(param)) {
-    *dispatch_mode = DispatchMode::kFComputeEx;
-  }
-  if (dev_mask == mshadow::cpu::kDevMask && !MKLDNNEnvSet()) {
-    *dispatch_mode = DispatchMode::kFComputeFallback;
-    return ret;
-  }
-#endif
-  return ret;
+  return MKLDNNStorageType(attrs, dev_mask, SupportMKLDNNAct(param),
+                           dispatch_mode, in_attrs, out_attrs);
 }
+#endif
 
 MXNET_OPERATOR_REGISTER_UNARY(Activation)
 .describe(R"code(Applies an activation function element-wise to the input.
@@ -183,7 +146,9 @@ The following activation functions are supported:
 
 )code" ADD_FILELINE)
 .set_attr_parser(ParamParser<ActivationParam>)
+#if MXNET_USE_MKLDNN == 1
 .set_attr<FInferStorageType>("FInferStorageType", ActivationStorageType)
+#endif
 .set_attr<nnvm::FListOutputNames>("FListOutputNames",
     [](const NodeAttrs& attrs) {
     return std::vector<std::string>{"output"};
@@ -204,7 +169,9 @@ NNVM_REGISTER_OP(_backward_Activation)
   })
 .set_num_outputs(1)
 .set_attr<nnvm::TIsBackward>("TIsBackward", true)
+#if MXNET_USE_MKLDNN == 1
 .set_attr<FInferStorageType>("FInferStorageType", BackwardActStorageType)
+#endif
 .set_attr<nnvm::FInferShape>("FInferShape", ElemwiseShape<3, 1>)
 .set_attr<nnvm::FInferType>("FInferType", ElemwiseType<3, 1>)
 .set_attr<nnvm::FInplaceOption>("FInplaceOption", [](const NodeAttrs& attrs){

diff --git a/src/operator/nn/batch_norm.cc b/src/operator/nn/batch_norm.cc
@@ -27,6 +27,7 @@
 #include "batch_norm-inl.h"
 #include <nnvm/op_attr_types.h>
 #include "../elemwise_op_common.h"
+#include "../operator_common.h"
 #if MXNET_USE_MKLDNN == 1
 #include "./mkldnn/mkldnn_batch_norm-inl.h"
 #endif
@@ -544,7 +545,7 @@ Both *mean* and *var* returns a scalar by treating the input as a vector.
 
 Assume the input has size *k* on axis 1, then both ``gamma`` and ``beta``
 have shape *(k,)*. If ``output_mean_var`` is set to be true, then outputs both ``data_mean`` and
-the inverse of ``data_var``, which are needed for the backward pass. Note that gradient of these 
+the inverse of ``data_var``, which are needed for the backward pass. Note that gradient of these
 two outputs are blocked.
 
 Besides the inputs and the outputs, this operator accepts two auxiliary

diff --git a/src/operator/nn/convolution.cc b/src/operator/nn/convolution.cc
@@ -26,11 +26,14 @@
 
 #include "./convolution-inl.h"
 #include "../elemwise_op_common.h"
-#include "./mkldnn/mkldnn_ops-inl.h"
-#include "./mkldnn/mkldnn_base-inl.h"
+#include "../operator_common.h"
 #if MXNET_USE_NNPACK == 1
 #include "../nnpack/nnpack_pooling-inl.h"
 #endif  // MXNET_USE_NNPACK
+#if MXNET_USE_MKLDNN == 1
+#include "./mkldnn/mkldnn_base-inl.h"
+#include "./mkldnn/mkldnn_ops-inl.h"
+#endif  // MXNET_USE_MKLDNN
 
 namespace mxnet {
 namespace op {
@@ -288,27 +291,19 @@ static bool ConvolutionType(const nnvm::NodeAttrs& attrs,
   return true;
 }
 
+#if MXNET_USE_MKLDNN == 1
 inline static bool ConvStorageType(const nnvm::NodeAttrs& attrs,
                                    const int dev_mask,
                                    DispatchMode* dispatch_mode,
-                                   std::vector<int> *in_attrs,
-                                   std::vector<int> *out_attrs) {
+                                   std::vector<int>* in_attrs,
+                                   std::vector<int>* out_attrs) {
   const ConvolutionParam& param = nnvm::get<ConvolutionParam>(attrs.parsed);
   uint32_t in_expected = param.no_bias ? 2 : 3;
   CHECK_EQ(in_attrs->size(), in_expected);
   CHECK_EQ(out_attrs->size(), 1);
 
-  DispatchMode wanted_mode;
-#if MXNET_USE_MKLDNN == 1
-  if (dev_mask == mshadow::cpu::kDevMask && !MKLDNNEnvSet())
-    wanted_mode = DispatchMode::kFComputeFallback;
-  else if (dev_mask == mshadow::cpu::kDevMask)
-    wanted_mode = DispatchMode::kFComputeEx;
-  else
-#endif
-    wanted_mode = DispatchMode::kFCompute;
-  return storage_type_assign(out_attrs, mxnet::kDefaultStorage,
-                             dispatch_mode, wanted_mode);
+  return MKLDNNStorageType(attrs, dev_mask, true, dispatch_mode, in_attrs,
+                           out_attrs);
 }
 
 inline static bool BackwardConvStorageType(const nnvm::NodeAttrs& attrs,
@@ -322,18 +317,10 @@ inline static bool BackwardConvStorageType(const nnvm::NodeAttrs& attrs,
   CHECK_EQ(in_attrs->size(), in_expected);
   CHECK_EQ(out_attrs->size(), out_expected);
 
-  DispatchMode wanted_mode;
-#if MXNET_USE_MKLDNN == 1
-  if (dev_mask == mshadow::cpu::kDevMask && !MKLDNNEnvSet())
-    wanted_mode = DispatchMode::kFComputeFallback;
-  else if (dev_mask == mshadow::cpu::kDevMask)
-    wanted_mode = DispatchMode::kFComputeEx;
-  else
-#endif
-    wanted_mode = DispatchMode::kFCompute;
-  return storage_type_assign(out_attrs, mxnet::kDefaultStorage,
-                             dispatch_mode, wanted_mode);
+  return MKLDNNStorageType(attrs, dev_mask, true, dispatch_mode, in_attrs,
+                           out_attrs);
 }
+#endif
 
 void ConvolutionParamParser(nnvm::NodeAttrs* attrs) {
   using namespace mshadow;
@@ -492,7 +479,9 @@ There are other options to tune the performance.
 })
 .set_attr<nnvm::FInferShape>("FInferShape", ConvolutionShape)
 .set_attr<nnvm::FInferType>("FInferType", ConvolutionType)
+#if MXNET_USE_MKLDNN == 1
 .set_attr<FInferStorageType>("FInferStorageType", ConvStorageType)
+#endif
 .set_attr<FCompute>("FCompute<cpu>", ConvolutionCompute<cpu>)
 #if MXNET_USE_MKLDNN == 1
 .set_attr<FComputeEx>("FComputeEx<cpu>", ConvolutionComputeExCPU)
@@ -512,7 +501,9 @@ NNVM_REGISTER_OP(_backward_Convolution)
   return params.no_bias ? 2 : 3;
 })
 .set_attr<nnvm::TIsBackward>("TIsBackward", true)
+#if MXNET_USE_MKLDNN == 1
 .set_attr<FInferStorageType>("FInferStorageType", BackwardConvStorageType)
+#endif
 .set_attr<FResourceRequest>("FResourceRequest", [](const NodeAttrs& n) {
   return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
 })

diff --git a/src/operator/nn/deconvolution.cc b/src/operator/nn/deconvolution.cc
@@ -25,8 +25,12 @@
 */
 
 #include "./deconvolution-inl.h"
+#include "../operator_common.h"
+#include "../../common/utils.h"
+#if MXNET_USE_MKLDNN == 1
 #include "./mkldnn/mkldnn_ops-inl.h"
 #include "./mkldnn/mkldnn_base-inl.h"
+#endif
 
 namespace mxnet {
 namespace op {
@@ -256,6 +260,7 @@ static bool DeconvolutionType(const nnvm::NodeAttrs& attrs,
   return true;
 }
 
+#if MXNET_USE_MKLDNN == 1
 inline static bool DeconvStorageType(const nnvm::NodeAttrs& attrs,
                                      const int dev_mask,
                                      DispatchMode* dispatch_mode,
@@ -266,17 +271,8 @@ inline static bool DeconvStorageType(const nnvm::NodeAttrs& attrs,
   CHECK_EQ(in_attrs->size(), in_expected);
   CHECK_EQ(out_attrs->size(), 1);
 
-  DispatchMode wanted_mode;
-#if MXNET_USE_MKLDNN == 1
-  if (dev_mask == mshadow::cpu::kDevMask && !MKLDNNEnvSet())
-    wanted_mode = DispatchMode::kFComputeFallback;
-  else if (dev_mask == mshadow::cpu::kDevMask)
-    wanted_mode = DispatchMode::kFComputeEx;
-  else
-#endif
-    wanted_mode = DispatchMode::kFCompute;
-  return storage_type_assign(out_attrs, mxnet::kDefaultStorage,
-                             dispatch_mode, wanted_mode);
+  return MKLDNNStorageType(attrs, dev_mask, true, dispatch_mode, in_attrs,
+                           out_attrs);
 }
 
 inline static bool BackwardDeconvStorageType(const nnvm::NodeAttrs& attrs,
@@ -289,20 +285,10 @@ inline static bool BackwardDeconvStorageType(const nnvm::NodeAttrs& attrs,
   CHECK_EQ(in_attrs->size(), param.no_bias ? 3U : 4U);
   CHECK_EQ(out_attrs->size(), out_expected);
 
-  DispatchMode wanted_mode;
-#if MXNET_USE_MKLDNN == 1
-  if (dev_mask == mshadow::cpu::kDevMask && !MKLDNNEnvSet())
-    wanted_mode = DispatchMode::kFComputeFallback;
-  else if (dev_mask == mshadow::cpu::kDevMask)
-    wanted_mode = DispatchMode::kFComputeEx;
-  else
-#endif
-    wanted_mode = DispatchMode::kFCompute;
-  return storage_type_assign(out_attrs, mxnet::kDefaultStorage,
-                             dispatch_mode, wanted_mode);
+  return MKLDNNStorageType(attrs, dev_mask, true, dispatch_mode, in_attrs,
+                           out_attrs);
 }
 
-#if MXNET_USE_MKLDNN == 1
 static void DeconvolutionComputeExCPU(const nnvm::NodeAttrs& attrs,
                                       const OpContext& ctx,
                                       const std::vector<NDArray>& inputs,
@@ -419,7 +405,9 @@ NNVM_REGISTER_OP(Deconvolution)
 })
 .set_attr<nnvm::FInferShape>("FInferShape", DeconvolutionShape)
 .set_attr<nnvm::FInferType>("FInferType", DeconvolutionType)
+#if MXNET_USE_MKLDNN == 1
 .set_attr<FInferStorageType>("FInferStorageType", DeconvStorageType)
+#endif
 .set_attr<FResourceRequest>("FResourceRequest", [](const NodeAttrs& n) {
   return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
 })
@@ -440,7 +428,9 @@ NNVM_REGISTER_OP(_backward_Deconvolution)
   return params.no_bias ? 2 : 3;
 })
 .set_attr<nnvm::TIsBackward>("TIsBackward", true)
+#if MXNET_USE_MKLDNN == 1
 .set_attr<FInferStorageType>("FInferStorageType", BackwardDeconvStorageType)
+#endif
 .set_attr<FResourceRequest>("FResourceRequest", [](const NodeAttrs& n) {
   return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
 })

diff --git a/src/operator/nn/lrn.cc b/src/operator/nn/lrn.cc
@@ -28,6 +28,7 @@
 #include "../operator_common.h"
 #if MXNET_USE_MKLDNN == 1
 #include "./mkldnn/mkldnn_lrn-inl.h"
+#include "./mkldnn/mkldnn_base-inl.h"
 #endif
 
 namespace mxnet {
@@ -81,24 +82,16 @@ struct LRNGrad {
   }
 };
 
+#if MXNET_USE_MKLDNN == 1
 bool LRNForwardInferStorageType(const nnvm::NodeAttrs& attrs,
                                 const int dev_mask,
                                 DispatchMode* dispatch_mode,
                                 std::vector<int> *in_attrs,
                                 std::vector<int> *out_attrs) {
   CHECK(!in_attrs->empty());
-#if MXNET_USE_MKLDNN == 1
-  if (dev_mask == mshadow::cpu::kDevMask && !MKLDNNEnvSet()) {
-    return storage_type_assign(out_attrs, mxnet::kDefaultStorage,
-                        dispatch_mode, DispatchMode::kFComputeFallback);
-  } else if (dev_mask == mshadow::cpu::kDevMask) {
-    return storage_type_assign(out_attrs, mxnet::kDefaultStorage,
-                        dispatch_mode, DispatchMode::kFComputeEx);
-  }
-#endif
-  storage_type_assign(out_attrs, mxnet::kDefaultStorage,
-                      dispatch_mode, DispatchMode::kFCompute);
-  return true;
+
+  return MKLDNNStorageType(attrs, dev_mask, true, dispatch_mode, in_attrs,
+                           out_attrs);
 }
 
 bool LRNBackwardInferStorageType(const nnvm::NodeAttrs& attrs,
@@ -107,20 +100,11 @@ bool LRNBackwardInferStorageType(const nnvm::NodeAttrs& attrs,
                                  std::vector<int> *in_attrs,
                                  std::vector<int> *out_attrs) {
   CHECK(!in_attrs->empty());
-#if MXNET_USE_MKLDNN == 1
-  if (dev_mask == mshadow::cpu::kDevMask && !MKLDNNEnvSet()) {
-    return storage_type_assign(out_attrs, mxnet::kDefaultStorage,
-                        dispatch_mode, DispatchMode::kFComputeFallback);
-  } else if (dev_mask == mshadow::cpu::kDevMask) {
-    return storage_type_assign(out_attrs, mxnet::kDefaultStorage,
-                        dispatch_mode, DispatchMode::kFComputeEx);
-  }
-#endif
-  return storage_type_assign(out_attrs, mxnet::kDefaultStorage,
-                      dispatch_mode, DispatchMode::kFCompute);
+
+  return MKLDNNStorageType(attrs, dev_mask, true, dispatch_mode, in_attrs,
+                           out_attrs);
 }
 
-#if MXNET_USE_MKLDNN == 1
 void LRNComputeExCPU(const nnvm::NodeAttrs &attrs,
                      const OpContext &ctx,
                      const std::vector<NDArray> &inputs,
@@ -183,7 +167,9 @@ number of kernels in the layer.
 .set_attr_parser(ParamParser<LRNParam>)
 .set_attr<nnvm::FInferShape>("FInferShape", LRNShape)
 .set_attr<nnvm::FInferType>("FInferType", LRNType)
+#if MXNET_USE_MKLDNN == 1
 .set_attr<FInferStorageType>("FInferStorageType", LRNForwardInferStorageType)
+#endif
 .set_attr<nnvm::FListInputNames>("FListInputNames",
     [](const NodeAttrs& attrs) {
   return std::vector<std::string>{"data"};
@@ -203,7 +189,9 @@ number of kernels in the layer.
 NNVM_REGISTER_OP(_backward_LRN)
 .set_num_outputs(1)
 .set_attr_parser(ParamParser<LRNParam>)
+#if MXNET_USE_MKLDNN == 1
 .set_attr<FInferStorageType>("FInferStorageType", LRNBackwardInferStorageType)
+#endif
 .set_attr<nnvm::TIsBackward>("TIsBackward", true)
 #if MXNET_USE_MKLDNN == 1
 .set_attr<FComputeEx>("FComputeEx<cpu>", LRNGradComputeExCPU)

diff --git a/src/operator/nn/mkldnn/mkldnn_base.cc b/src/operator/nn/mkldnn/mkldnn_base.cc
@@ -536,7 +536,9 @@ bool MKLDNNStorageType(const nnvm::NodeAttrs &attrs,
 
   DispatchMode wanted_mode;
 #if MXNET_USE_MKLDNN == 1
-  if (dev_mask == mshadow::cpu::kDevMask && support_mkldnn)
+  if (dev_mask == mshadow::cpu::kDevMask && !MKLDNNEnvSet())
+    wanted_mode = DispatchMode::kFComputeFallback;
+  else if (dev_mask == mshadow::cpu::kDevMask && support_mkldnn)
     wanted_mode = DispatchMode::kFComputeEx;
   else
 #endif