This is an automated email from the ASF dual-hosted git repository.
marcoabreu pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git
The following commit(s) were added to refs/heads/master by this push:
new 148a50a Revert "[MXNET-351] Fix a bug in the MKLDNN integration."
(#10718)
148a50a is described below
commit 148a50acedb9a066f10ae4af3cd325e3b9fc5a61
Author: Marco de Abreu <marcoab...@users.noreply.github.com>
AuthorDate: Fri Apr 27 07:45:21 2018 -0700
Revert "[MXNET-351] Fix a bug in the MKLDNN integration." (#10718)
* Revert "test Jenkins."
This reverts commit 1d33fc798231f27d8f7c19e517955d91890faf3a.
* Revert "test jenkins."
This reverts commit 652a8953c44585f831f4e220c8ae82d402be1576.
* Revert "Fix Jenkins."
This reverts commit e5549995b9bac8194826503985b4e3292d271cc8.
* Revert "Fix jenkins."
This reverts commit 608583158557a40668705a5c090dd0f9fe7a89e8.
* Revert "test in jenkins."
This reverts commit 07819ad72268370b459b42f4f7b6a011c5dfcdfe.
* Revert "Reorganize test code."
This reverts commit 9e90090f0b64906d191d1e3fa99802c8be2ba5cf.
* Revert "add comments for the test code."
This reverts commit 99bd17f917a8e1ba533567a543ed8f5bd147e60f.
* Revert "Add more test cases."
This reverts commit 2bd669172526307a94943daf8f4d409962e44c64.
* Revert "Fix lint error."
This reverts commit 9d1b9dc89f9a59b8ca3ee982786f64210a697085.
* Revert "Unit tests on GetMKLDNNDataReorder."
This reverts commit 7d6e8e0b3adb4ff039a7585e01b19f88498a9f89.
* Revert "Fix a minor bug."
This reverts commit deeeea42e4c1590b3ca16b028056295116c6ae9e.
* Revert "Add C++ unit test for NDArray."
This reverts commit 29cd215d29594583733cf50c19957775278f13df.
* Revert "avoid unnecessary message."
This reverts commit 3713ead5133d076ebfcd76413d965999cd380e08.
* Revert "Handle all cases for GetMKLDNNDataReorder"
This reverts commit a1cb6a6c152731067622aff101a8b66b383698ec.
* Revert "Update comments."
This reverts commit d62f138d7fc27f775befa7c61072caff0ceac718.
* Revert "Fix a bug in GetMKLDNNData()."
This reverts commit 76a102989af0c4b635083437b9ef133435710fd8.
* Revert "test inference multiple times."
This reverts commit 0a664ef842f5c3ecc9400af2e37921b2a81b49fb.
---
Jenkinsfile | 5 +-
src/ndarray/ndarray.cc | 48 ++----
src/operator/nn/mkldnn/mkldnn_base-inl.h | 31 +---
src/operator/nn/mkldnn/mkldnn_base.cc | 12 +-
tests/cpp/operator/mkldnn.cc | 248 ---------------------------
tests/python/gpu/test_gluon_model_zoo_gpu.py | 19 +-
6 files changed, 34 insertions(+), 329 deletions(-)
diff --git a/Jenkinsfile b/Jenkinsfile
index 1402e8e..5601c52 100644
--- a/Jenkinsfile
+++ b/Jenkinsfile
@@ -26,7 +26,7 @@ mx_lib = 'lib/libmxnet.so, lib/libmxnet.a,
3rdparty/dmlc-core/libdmlc.a, 3rdpart
mx_dist_lib = 'lib/libmxnet.so, lib/libmxnet.a, 3rdparty/dmlc-core/libdmlc.a,
3rdparty/nnvm/lib/libnnvm.a, 3rdparty/ps-lite/build/libps.a,
deps/lib/libprotobuf-lite.a, deps/lib/libzmq.a'
// mxnet cmake libraries, in cmake builds we do not produce a libnvvm static
library by default.
mx_cmake_lib = 'build/libmxnet.so, build/libmxnet.a,
build/3rdparty/dmlc-core/libdmlc.a, build/tests/mxnet_unit_tests,
build/3rdparty/openmp/runtime/src/libomp.so'
-mx_cmake_mkldnn_lib = 'build/libmxnet.so, build/libmxnet.a,
build/3rdparty/dmlc-core/libdmlc.a, build/tests/mxnet_unit_tests,
build/3rdparty/openmp/runtime/src/libomp.so, build/libmkldnn.so.0'
+mx_cmake_mkldnn_lib = 'build/libmxnet.so, build/libmxnet.a,
build/3rdparty/dmlc-core/libdmlc.a, build/tests/mxnet_unit_tests,
build/3rdparty/openmp/runtime/src/libomp.so,
build/3rdparty/mkldnn/src/libmkldnn.so,
build/3rdparty/mkldnn/src/libmkldnn.so.0'
mx_mkldnn_lib = 'lib/libmxnet.so, lib/libmxnet.a, lib/libiomp5.so,
lib/libmkldnn.so.0, lib/libmklml_intel.so, 3rdparty/dmlc-core/libdmlc.a,
3rdparty/nnvm/lib/libnnvm.a'
// command to start a docker container
docker_run = 'tests/ci_build/ci_build.sh'
@@ -288,7 +288,6 @@ try {
timeout(time: max_time, unit: 'MINUTES') {
init_git()
sh "ci/build.py --platform ubuntu_gpu /work/runtime_functions.sh
build_ubuntu_gpu_cmake_mkldnn" //build_cuda
- sh "cp build/3rdparty/mkldnn/src/libmkldnn.so.0 build/"
pack_lib('cmake_mkldnn_gpu', mx_cmake_mkldnn_lib)
}
}
@@ -569,7 +568,7 @@ try {
ws('workspace/ut-cpp-gpu') {
timeout(time: max_time, unit: 'MINUTES') {
init_git()
- unpack_lib('cmake_mkldnn_gpu', mx_cmake_mkldnn_lib)
+ unpack_lib('cmake_gpu', mx_cmake_lib)
sh "ci/build.py --nvidiadocker --platform ubuntu_gpu
/work/runtime_functions.sh unittest_ubuntu_gpu_cpp"
}
}
diff --git a/src/ndarray/ndarray.cc b/src/ndarray/ndarray.cc
index 82de094..b428c2c 100644
--- a/src/ndarray/ndarray.cc
+++ b/src/ndarray/ndarray.cc
@@ -485,8 +485,8 @@ const mkldnn::memory *NDArray::GetMKLDNNData(
}
const mkldnn::memory *NDArray::GetMKLDNNDataReorder(
- const mkldnn::memory::primitive_desc &new_pd) const {
- if (new_pd.get_size() != shape().Size() * GetTypeSize(dtype_)) {
+ const mkldnn::memory::primitive_desc &desc) const {
+ if (desc.get_size() != shape().Size() * GetTypeSize(dtype_)) {
LOG(FATAL) << "The size of NDArray doesn't match the requested MKLDNN
memory desc";
return nullptr;
}
@@ -495,41 +495,24 @@ const mkldnn::memory *NDArray::GetMKLDNNDataReorder(
const mkldnn::memory *mem = GetMKLDNNData();
// If the memory descriptor matches, it's easy.
MKLDNNStream *stream = MKLDNNStream::Get();
- if (mem->get_primitive_desc() == new_pd) {
- return GetMKLDNNExact(mem, new_pd);
+ if (mem->get_primitive_desc() == desc) {
+ return GetMKLDNNExact(mem, desc);
}
- mkldnn::memory::primitive_desc _pd = new_pd;
- mkldnn::memory::desc desc1 = mem->get_primitive_desc().desc();
- mkldnn::memory::desc desc2 = _pd.desc();
+ mkldnn::memory::primitive_desc _desc = desc;
// Now we need to determine if we should reorder the memory.
// If both use the default formats, we think we don't need to reorder.
+ mkldnn::memory::desc desc1 = mem->get_primitive_desc().desc();
+ mkldnn::memory::desc desc2 = _desc.desc();
if (desc1.data.format == GetDefaultFormat(desc1) &&
desc2.data.format == GetDefaultFormat(desc2)) {
- mkldnn_mem_ptr ret(new mkldnn::memory(new_pd, mem->get_data_handle()));
+ mkldnn_mem_ptr ret(new mkldnn::memory(desc, mem->get_data_handle()));
stream->RegisterMem(ret);
return ret.get();
- } else if (same_shape(desc1, desc2)) {
- // If they have the same shape, we can reorder data directly.
- mkldnn::memory *ret = TmpMemMgr::Get()->Alloc(new_pd);
+ } else {
+ mkldnn::memory *ret = TmpMemMgr::Get()->Alloc(desc);
stream->RegisterPrim(mkldnn::reorder(*mem, *ret));
return ret;
- } else {
- // If they have different shapes, we need to reshape the array first.
- // Since this method will only be used inside an operator, we can call
- // MKLDNNDataReshape to reshape an array.
- TShape required_shape(desc2.data.ndims);
- for (int i = 0; i < desc2.data.ndims; i++)
- required_shape[i] = desc2.data.dims[i];
- NDArray reshaped = MKLDNNDataReshape(required_shape);
- const mkldnn::memory *ret = reshaped.GetMKLDNNData();
- if (ret->get_primitive_desc() == new_pd) {
- return GetMKLDNNExact(ret, new_pd);
- } else {
- mkldnn::memory *ret2 = TmpMemMgr::Get()->Alloc(new_pd);
- stream->RegisterPrim(mkldnn::reorder(*ret, *ret2));
- return ret2;
- }
}
}
@@ -583,15 +566,10 @@ void NDArray::MKLDNNDataReorderAsync(const
mkldnn::memory::primitive_desc &desc)
const mkldnn::memory *NDArray::GetMKLDNNData() const {
CHECK(storage_type() == kDefaultStorage);
- if (IsMKLDNNData()) {
- // If this array uses MKLDNN layout, we have to make sure it's not a view.
- // Otherwise, we'll have to change the layout inside the array.
+ // If this array uses MKLDNN layout, we have to make sure it's not a view.
+ // Otherwise, we'll have to change the layout inside the array.
+ if (IsMKLDNNData())
CHECK(!IsView());
- MKLDNNStream::Get()->RegisterMem(ptr_->mkl_mem_->GetMem());
- // If this array uses MKLDNN format, we should return now. Otherwise,
- // SetMKLMem may mess up mkl_mem_.
- return ptr_->mkl_mem_->GetRaw();
- }
ptr_->SetMKLMem(IsView() ? ptr_->storage_shape : shape_, dtype_);
MKLDNNStream::Get()->RegisterMem(ptr_->mkl_mem_->GetMem());
if (IsView()) {
diff --git a/src/operator/nn/mkldnn/mkldnn_base-inl.h
b/src/operator/nn/mkldnn/mkldnn_base-inl.h
index f47a051..489351e 100644
--- a/src/operator/nn/mkldnn/mkldnn_base-inl.h
+++ b/src/operator/nn/mkldnn/mkldnn_base-inl.h
@@ -272,12 +272,13 @@ class MKLDNNStream {
std::vector<std::shared_ptr<const mkldnn::memory> > mem_holder;
public:
- static MKLDNNStream *Get();
-
- void RegisterPrim(const mkldnn::primitive &prim) {
- net.push_back(prim);
+ static MKLDNNStream *Get() {
+ static thread_local MKLDNNStream stream;
+ return &stream;
}
+ void RegisterPrim(const mkldnn::primitive &prim) { net.push_back(prim); }
+
void RegisterMem(std::shared_ptr<const mkldnn::memory> mem) {
mem_holder.push_back(mem);
}
@@ -286,16 +287,10 @@ class MKLDNNStream {
return !net.empty();
}
- void Submit(bool cleanup = true) {
- if (!net.empty()) {
+ void Submit() {
+ if (!net.empty())
mkldnn::stream(mkldnn::stream::kind::eager).submit(net).wait();
- net.clear();
- }
- if (cleanup)
- Cleanup();
- }
-
- void Cleanup() {
+ net.clear();
mem_holder.clear();
TmpMemMgr::Get()->Reset();
}
@@ -353,16 +348,6 @@ inline bool same_shape(const TShape &shape, const
mkldnn_dims_t dims, int ndims)
return true;
}
-inline bool same_shape(const mkldnn::memory::desc &desc1,
- const mkldnn::memory::desc &desc2) {
- if (desc1.data.ndims != desc2.data.ndims)
- return false;
- for (int i = 0; i < desc1.data.ndims; i++)
- if (desc1.data.dims[i] != desc2.data.dims[i])
- return false;
- return true;
-}
-
inline bool same_shape(const TShape &shape, int dtype,
const mkldnn::memory::desc &desc) {
return same_shape(shape, desc.data.dims, desc.data.ndims)
diff --git a/src/operator/nn/mkldnn/mkldnn_base.cc
b/src/operator/nn/mkldnn/mkldnn_base.cc
index c0e1ee6..df37db5 100644
--- a/src/operator/nn/mkldnn/mkldnn_base.cc
+++ b/src/operator/nn/mkldnn/mkldnn_base.cc
@@ -25,11 +25,6 @@
namespace mxnet {
-MKLDNNStream *MKLDNNStream::Get() {
- static thread_local MKLDNNStream stream;
- return &stream;
-}
-
void *AlignMem(void *mem, size_t size, size_t alignment, size_t *space) {
if (size > *space)
return nullptr;
@@ -62,11 +57,8 @@ mkldnn::memory *TmpMemMgr::Alloc(const
mkldnn::memory::primitive_desc &pd) {
this->curr_mem = static_cast<char *>(mem) + pd.get_size();
return ret.get();
} else {
- // If curr_mem has been initialized and we still reach here. It means
- // the current allocated memory isn't enough.
- if (this->curr_mem)
- LOG(WARNING) << "Allocate " << pd.get_size()
- << " bytes with malloc directly";
+ LOG(WARNING) << "Allocate " << pd.get_size()
+ << " bytes with malloc directly";
mkldnn_mem_ptr ret(new mkldnn::memory(pd));
MKLDNNStream::Get()->RegisterMem(ret);
return ret.get();
diff --git a/tests/cpp/operator/mkldnn.cc b/tests/cpp/operator/mkldnn.cc
index 58ad894..9d4b982 100644
--- a/tests/cpp/operator/mkldnn.cc
+++ b/tests/cpp/operator/mkldnn.cc
@@ -28,8 +28,6 @@
#include "gtest/gtest.h"
#include "../../src/operator/nn/mkldnn/mkldnn_base-inl.h"
-using namespace mxnet;
-
#if __GNUC__ >= 5
bool test_mem_align(void *mem, size_t size, size_t alignment, size_t space) {
void *ret1, *ret2;
@@ -86,250 +84,4 @@ TEST(MKLDNN_UTIL_FUNC, MemFormat) {
CHECK_EQ(mkldnn_nchw, 5);
CHECK_EQ(mkldnn_oihw, 12);
}
-
-// Init arrays with the default layout.
-static void InitArray(NDArray *arr) {
- const TBlob &blob = arr->data();
- mshadow::default_real_t *data = blob.dptr<mshadow::default_real_t>();
- size_t size = blob.Size();
- for (size_t i = 0; i < size; i++)
- data[i] = i;
-}
-
-// Init arrays with the specified layout.
-static void InitMKLDNNArray(NDArray *arr, const mkldnn::memory::primitive_desc
&pd) {
- const TBlob &blob = arr->data();
- mshadow::default_real_t *data = blob.dptr<mshadow::default_real_t>();
- size_t size = blob.Size();
- for (size_t i = 0; i < size; i++)
- data[i] = i;
- arr->MKLDNNDataReorderAsync(pd);
- arr->WaitToRead();
-}
-
-static void VerifyDefMem(const mkldnn::memory &mem) {
- mkldnn::memory::primitive_desc pd = mem.get_primitive_desc();
- mshadow::default_real_t *data
- = static_cast<mshadow::default_real_t *>(mem.get_data_handle());
- size_t size = pd.get_size() / sizeof(mshadow::default_real_t);
- size_t num_same = 0;
- for (size_t i = 0; i < size; i++)
- num_same += data[i] == static_cast<mshadow::default_real_t>(i);
- EXPECT_EQ(num_same, size);
-}
-
-static void VerifyMem(const mkldnn::memory &mem) {
- mkldnn::memory::primitive_desc pd = mem.get_primitive_desc();
-
- if (pd.desc().data.format == GetDefaultFormat(pd.desc())) {
- VerifyDefMem(mem);
- } else {
- mkldnn::memory::dims dims(pd.desc().data.ndims);
- for (size_t i = 0; i < dims.size(); i++)
- dims[i] = pd.desc().data.dims[i];
- mkldnn::memory::desc desc{dims,
-
static_cast<mkldnn::memory::data_type>(pd.desc().data.data_type),
-
static_cast<mkldnn::memory::format>(GetDefaultFormat(pd.desc()))};
- mkldnn::memory::primitive_desc new_pd(desc,
CpuEngine::Get()->get_engine());
- mkldnn::memory new_mem(new_pd);
-
- std::vector<mkldnn::primitive> net;
- net.push_back(mkldnn::reorder(mem, new_mem));
- mkldnn::stream(mkldnn::stream::kind::eager).submit(net).wait();
- VerifyDefMem(new_mem);
- }
-}
-
-static mkldnn::memory::primitive_desc GetMemPD(const TShape s, int dtype,
- mkldnn::memory::format format) {
- mkldnn::memory::dims dims(s.ndim());
- for (size_t i = 0; i < dims.size(); i++)
- dims[i] = s[i];
- mkldnn::memory::desc desc{dims, get_mkldnn_type(dtype), format};
- return mkldnn::memory::primitive_desc(desc, CpuEngine::Get()->get_engine());
-}
-
-// This function gets special MKLDNN formats without knowing the specific
-// hardware configuration. Certainly, it potentially misses some format if
-// it's specific for certain array shapes. It covers at least one special
format
-// for each of the formats: nchw, oihw, goihw.
-// To test the logic of the code in NDArray, these formats should be enough.
-static std::vector<mkldnn::memory::format> GetMKLDNNFormat(size_t num_dims,
int dtype) {
- if (num_dims == 4) {
- mkldnn::memory::dims data_dims{1, 3, 224, 224};
- mkldnn::memory::desc data_md{data_dims, get_mkldnn_type(dtype),
- mkldnn::memory::format::any};
- mkldnn::memory::dims weight_dims{96, 3, 11, 11};
- mkldnn::memory::desc weight_md{weight_dims, get_mkldnn_type(dtype),
- mkldnn::memory::format::any};
- mkldnn::memory::dims output_dims{1, 96, 54, 54};
- mkldnn::memory::desc out_md{output_dims, get_mkldnn_type(dtype),
- mkldnn::memory::format::any};
- mkldnn::memory::dims strides{4, 4};
- mkldnn::memory::dims padding{0, 0};
-
- mkldnn::convolution_forward::desc desc(mkldnn::prop_kind::forward_training,
-
mkldnn::algorithm::convolution_direct,
- data_md, weight_md, out_md, strides,
- padding, padding,
mkldnn::padding_kind::zero);
- mkldnn::convolution_forward::primitive_desc pd(desc,
CpuEngine::Get()->get_engine());
- std::vector<mkldnn::memory::format> ret(2);
- ret[0] =
static_cast<mkldnn::memory::format>(pd.dst_primitive_desc().desc().data.format);
- ret[1] =
static_cast<mkldnn::memory::format>(pd.weights_primitive_desc().desc().data.format);
- printf("format: %d, %d\n", ret[0], ret[1]);
- return ret;
- } else if (num_dims == 5) {
- mkldnn::memory::dims data_dims{1, 32, 112, 112};
- mkldnn::memory::desc data_md{data_dims, get_mkldnn_type(dtype),
- mkldnn::memory::format::any};
- mkldnn::memory::dims weight_dims{32, 1, 1, 3, 3};
- mkldnn::memory::desc weight_md{weight_dims, get_mkldnn_type(dtype),
- mkldnn::memory::format::any};
- mkldnn::memory::dims output_dims{1, 32, 112, 112};
- mkldnn::memory::desc out_md{output_dims, get_mkldnn_type(dtype),
- mkldnn::memory::format::any};
- mkldnn::memory::dims strides{1, 1};
- mkldnn::memory::dims padding{1, 1};
-
- mkldnn::convolution_forward::desc desc(mkldnn::prop_kind::forward_training,
-
mkldnn::algorithm::convolution_direct,
- data_md, weight_md, out_md, strides,
- padding, padding,
mkldnn::padding_kind::zero);
- mkldnn::convolution_forward::primitive_desc pd(desc,
CpuEngine::Get()->get_engine());
- std::vector<mkldnn::memory::format> ret(1);
- ret[0] =
static_cast<mkldnn::memory::format>(pd.weights_primitive_desc().desc().data.format);
- printf("format: %d\n", ret[0]);
- return ret;
- } else {
- return std::vector<mkldnn::memory::format>();
- }
-}
-
-struct TestArrayShapes {
- std::vector<TShape> shapes;
- std::vector<mkldnn::memory::primitive_desc> pds;
-};
-
-static TestArrayShapes GetTestArrayShapes() {
- int dtype = mshadow::DataType<mshadow::default_real_t>::kFlag;
- std::vector<TShape> shapes;
- std::vector<mkldnn::memory::primitive_desc> pds;
- {
- // 1D
- TShape s(1);
- s[0] = 279936;
- shapes.push_back(s);
- pds.push_back(GetMemPD(s, dtype, mkldnn::memory::format::x));
- s[0] = 34848;
- shapes.push_back(s);
- pds.push_back(GetMemPD(s, dtype, mkldnn::memory::format::x));
- }
- {
- // 2D
- TShape s(2);
- s[0] = 96;
- s[1] = 2916;
- shapes.push_back(s);
- pds.push_back(GetMemPD(s, dtype, mkldnn::memory::format::nc));
- s[0] = 96;
- s[1] = 363;
- shapes.push_back(s);
- pds.push_back(GetMemPD(s, dtype, mkldnn::memory::format::nc));
- }
- {
- // 4D
- TShape s1(4);
- s1[0] = 1; s1[1] = 96; s1[2] = 54; s1[3] = 54;
- shapes.push_back(s1);
- pds.push_back(GetMemPD(s1, dtype, mkldnn::memory::format::nchw));
-
- TShape s2(4);
- s2[0] = 96; s2[1] = 3; s2[2] = 11; s2[3] = 11;
- shapes.push_back(s2);
- pds.push_back(GetMemPD(s2, dtype, mkldnn::memory::format::oihw));
-
- std::vector<mkldnn::memory::format> formats = GetMKLDNNFormat(4, dtype);
- pds.push_back(GetMemPD(s1, dtype, formats[0]));
- pds.push_back(GetMemPD(s2, dtype, formats[1]));
- }
- {
- // 5D
- TShape s(5);
- s[0] = 96; s[1] = 1; s[2] = 3; s[3] = 11; s[4] = 11;
- shapes.push_back(s);
- pds.push_back(GetMemPD(s, dtype, mkldnn::memory::format::goihw));
-
- std::vector<mkldnn::memory::format> formats = GetMKLDNNFormat(5, dtype);
- pds.push_back(GetMemPD(s, dtype, formats[0]));
- }
-
- TestArrayShapes ret;
- ret.shapes = shapes;
- ret.pds = pds;
- return ret;
-}
-
-TEST(MKLDNN_NDArray, GetDataReorder) {
- TestArrayShapes tas = GetTestArrayShapes();
- std::vector<TShape> shapes = tas.shapes;
- std::vector<mkldnn::memory::primitive_desc> pds = tas.pds;
-
-
- // Reorder from the default to any other layout.
- for (auto s : shapes) {
- NDArray arr(s, Context());
- InitArray(&arr);
- for (auto pd : pds) {
- if (s.Size() == pd.get_size() / sizeof(mshadow::default_real_t)) {
- const mkldnn::memory *mem = arr.GetMKLDNNDataReorder(pd);
- printf("reorder from (");
- for (size_t i = 0; i < s.ndim(); i++)
- printf("%ld, ", s[i]);
- printf(") to (");
- for (int i = 0; i < pd.desc().data.ndims; i++)
- printf("%d, ", pd.desc().data.dims[i]);
- printf("), format: %d\n", pd.desc().data.format);
- MKLDNNStream::Get()->Submit(false);
- VerifyMem(*mem);
- MKLDNNStream::Get()->Cleanup();
- }
- }
- }
-
- // Reorder from a special layout to another layout.
- for (auto s : shapes) {
- for (auto from_pd : pds) {
- if (from_pd.get_size() / sizeof(mshadow::default_real_t) == s.Size()) {
- NDArray arr(s, Context());
- // There is possibility that the dimensions of an NDArray doesn't match
- // with the MKLDNN memory inside.
- printf("Init array (");
- for (size_t i = 0; i < s.ndim(); i++)
- printf("%ld, ", s[i]);
- printf(") with MKLDNN memory (");
- for (int i = 0; i < from_pd.desc().data.ndims; i++)
- printf("%d, ", from_pd.desc().data.dims[i]);
- printf("), format: %d\n", from_pd.desc().data.format);
- InitMKLDNNArray(&arr, from_pd);
- for (auto to_pd : pds) {
- if (to_pd.get_size() / sizeof(mshadow::default_real_t) == s.Size()) {
- const mkldnn::memory *mem = arr.GetMKLDNNDataReorder(to_pd);
- printf("reorder from (");
- for (size_t i = 0; i < s.ndim(); i++)
- printf("%ld, ", s[i]);
- printf("), format: %d to (",
-
arr.GetMKLDNNData()->get_primitive_desc().desc().data.format);
- for (int i = 0; i < to_pd.desc().data.ndims; i++)
- printf("%d, ", to_pd.desc().data.dims[i]);
- printf("), format: %d\n", to_pd.desc().data.format);
- MKLDNNStream::Get()->Submit(false);
- VerifyMem(*mem);
- MKLDNNStream::Get()->Cleanup();
- }
- }
- }
- }
- }
-}
-
#endif
diff --git a/tests/python/gpu/test_gluon_model_zoo_gpu.py
b/tests/python/gpu/test_gluon_model_zoo_gpu.py
index 273ad3d..378a822 100644
--- a/tests/python/gpu/test_gluon_model_zoo_gpu.py
+++ b/tests/python/gpu/test_gluon_model_zoo_gpu.py
@@ -81,16 +81,15 @@ def test_inference():
gpu_param = gpu_params.get(k)
gpu_param.set_data(cpu_param.data().as_in_context(mx.gpu()))
- for i in range(5):
- # Run inference.
- with autograd.record(train_mode=False):
- cpu_out = cpu_model(mx.nd.array(data, ctx=mx.cpu()))
- gpu_out = gpu_model(gpu_data)
- out = cpu_out.asnumpy()
- max_val = np.max(np.abs(out))
- gpu_max_val = np.max(np.abs(gpu_out.asnumpy()))
- eprint(model_name + ": CPU " + str(max_val) + ", GPU " +
str(gpu_max_val))
- assert_almost_equal(out / max_val, gpu_out.asnumpy() / max_val,
rtol=1e-3, atol=1e-3)
+ # Run inference.
+ with autograd.record(train_mode=False):
+ cpu_out = cpu_model(mx.nd.array(data, ctx=mx.cpu()))
+ gpu_out = gpu_model(gpu_data)
+ out = cpu_out.asnumpy()
+ max_val = np.max(np.abs(out))
+ gpu_max_val = np.max(np.abs(gpu_out.asnumpy()))
+ eprint(model_name + ": CPU " + str(max_val) + ", GPU " +
str(gpu_max_val))
+ assert_almost_equal(out / max_val, gpu_out.asnumpy() / max_val,
rtol=1e-3, atol=1e-3)
def get_nn_model(name):
if "densenet" in name:
--
To stop receiving notification emails like this one, please contact
marcoab...@apache.org.
