Skip to content
This repository has been archived by the owner on Aug 5, 2022. It is now read-only.

Commit

Permalink
task-mkl2017-layers-reshape-fixes: - ForwardReshapeForward ULT was li…
Browse files Browse the repository at this point in the history 
…mited to MKL2017 engine
  • Loading branch information
@jczaja @ptbuilder
jczaja authored and ptbuilder committed Oct 28, 2016
2 parents b57b3bf + 8ddb521 commit 7c6249d
Show file tree
Hide file tree
Showing 16 changed files with 595 additions and 208 deletions.
38 changes: 34 additions & 4 deletions include/caffe/layers/mkl_layers.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -73,8 +73,8 @@ class MKLConvolutionLayer : public ConvolutionLayer<Dtype> {
const vector<bool>& propagate_down,
const vector<Blob<Dtype>*>& bottom);
// Customized methods
void Init( const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
void Init(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
Expand Down Expand Up @@ -169,6 +169,9 @@ class MKLLRNLayer : public Layer<Dtype> {
virtual void CrossChannelBackward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);

void Init(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

int size_;
int pre_pad_;
Dtype alpha_;
Expand Down Expand Up @@ -205,6 +208,9 @@ class MKLPoolingLayer : public Layer<Dtype> {
virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

void Init(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

virtual inline const char* type() const { return "Pooling"; }
virtual inline int ExactNumBottomBlobs() const { return 1; }
virtual inline int MinTopBlobs() const { return 1; }
Expand Down Expand Up @@ -270,6 +276,12 @@ class MKLReLULayer : public NeuronLayer<Dtype> {
virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

void Init(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

virtual inline const char* type() const { return "ReLU"; }

protected:
Expand All @@ -291,6 +303,8 @@ class MKLReLULayer : public NeuronLayer<Dtype> {
shared_ptr<MKLDiff<Dtype> > bwd_top_diff_;
shared_ptr<MKLDiff<Dtype> > bwd_bottom_diff_;
dnnPrimitive_t reluFwd_, reluBwd_;
vector<size_t> sizes_;
vector<size_t> strides_;
};

template <typename Dtype>
Expand All @@ -301,8 +315,8 @@ class MKLConcatLayer : public Layer<Dtype> {
concatFwd_(static_cast<dnnPrimitive_t>(NULL)),
concatBwd_(static_cast<dnnPrimitive_t>(NULL)),
fwd_top_data_ (new MKLData<Dtype>()),
bwd_top_diff_ (new MKLDiff<Dtype>()) {
}
bwd_top_diff_ (new MKLDiff<Dtype>()),
split_channels_ (NULL) { }
virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);
virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
Expand All @@ -323,6 +337,9 @@ class MKLConcatLayer : public Layer<Dtype> {
const vector<bool>& propagate_down,
const vector<Blob<Dtype>*>& bottom);

void Init(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

private:
dnnPrimitive_t concatFwd_;
dnnPrimitive_t concatBwd_;
Expand Down Expand Up @@ -375,6 +392,9 @@ class MKLBatchNormLayer : public Layer<Dtype> {
virtual void Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom);

void Init(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

// Dtype moving_average_fraction_;
Dtype eps_;
bool use_weight_bias_;
Expand Down Expand Up @@ -410,6 +430,9 @@ class MKLSplitLayer : public Layer<Dtype> {
virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

void Init(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

virtual inline const char* type() const { return "Split"; }
virtual inline int ExactNumBottomBlobs() const { return 1; }
virtual inline int MinTopBlobs() const { return 1; }
Expand All @@ -429,6 +452,8 @@ class MKLSplitLayer : public Layer<Dtype> {
vector<shared_ptr<MKLDiff<Dtype> > > bwd_top_diff;
vector<Dtype> coeffs_;
size_t num_tops;
vector<size_t> sizes_src_;
vector<size_t> strides_src_;
dnnPrimitive_t sumPrimitive;
};

Expand All @@ -446,6 +471,9 @@ class MKLEltwiseLayer : public Layer<Dtype> {
virtual void Reshape(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

void Init(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top);

virtual inline const char* type() const { return "Eltwise"; }
virtual inline int MinBottomBlobs() const { return 2; }
virtual inline int ExactNumTopBlobs() const { return 1; }
Expand All @@ -472,6 +500,8 @@ class MKLEltwiseLayer : public Layer<Dtype> {
vector<Dtype> coeffs_;
Blob<int> max_idx_;
size_t num_bottoms;
int channels_, num_;
int height_, width_;

bool stable_prod_grad_;
};
Expand Down
8 changes: 7 additions & 1 deletion include/caffe/mkl_memory.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -96,11 +96,15 @@ struct MKLMemoryDescriptorBase : PrvMemDescr,
void create_internal_layout(const dnnPrimitive_t primitive,
dnnResourceType_t type);
void create_user_layout(size_t dimension, const size_t size[],
const size_t strides[]);
const size_t strides[],
bool create_conversion_if_possible = true);
void create_layouts(
const dnnPrimitive_t primitive, dnnResourceType_t type,
size_t dimension, const size_t size[], const size_t strides[]);

void remove_internal_layout();
void remove_user_layout();

virtual PrvDescrType get_descr_type() {return PRV_DESCR_MKL2017;}
virtual size_t prv_size() {
return dnnLayoutGetMemorySize<Dtype>(layout_int);
Expand All @@ -112,6 +116,8 @@ struct MKLMemoryDescriptorBase : PrvMemDescr,
virtual void convert_to_prv(void* cpu_ptr);
virtual bool layout_compare(shared_ptr<PrvMemDescr> other);
virtual void convert_from_other(shared_ptr<PrvMemDescr> other);
protected:
void remove_conversions();
protected:
Dtype* internal_ptr;
};
Expand Down
50 changes: 36 additions & 14 deletions include/mkl_dnn_cppwrapper.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -101,28 +101,44 @@ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
{return dnnAllocateBuffer_F64(pPtr, layout);}

TEMPLATE_PREFIX dnnError_t dnnReleaseBuffer(
void *ptr);
void* ptr);
SPEC_PREFIX dnnError_t dnnReleaseBuffer<float>(
void *ptr)
{return dnnReleaseBuffer_F32(ptr);}
void* ptr) {
dnnError_t status = E_SUCCESS;
if( ptr != NULL) {
status = dnnReleaseBuffer_F32(ptr);
}
return status;
}
SPEC_PREFIX dnnError_t dnnReleaseBuffer<double>(
void *ptr)
{return dnnReleaseBuffer_F64(ptr);}
void* ptr) {
dnnError_t status = E_SUCCESS;
if( ptr != NULL) {
status = dnnReleaseBuffer_F64(ptr);
}
return status;
}

TEMPLATE_PREFIX dnnError_t dnnLayoutDelete(
dnnLayout_t& layout);
SPEC_PREFIX dnnError_t dnnLayoutDelete<float>(
dnnLayout_t& layout) {
dnnError_t status = dnnLayoutDelete_F32(layout);
layout = NULL;
dnnError_t status = E_SUCCESS;
if( layout != NULL) {
status = dnnLayoutDelete_F32(layout);
layout = NULL;
}
return status;
}
SPEC_PREFIX dnnError_t dnnLayoutDelete<double>(
dnnLayout_t& layout) {
dnnError_t status = dnnLayoutDelete_F64(layout);
layout = NULL;
dnnError_t status = E_SUCCESS;
if( layout != NULL) {
status = dnnLayoutDelete_F64(layout);
layout = NULL;
}
return status;
}
}

TEMPLATE_PREFIX dnnError_t dnnPrimitiveAttributesCreate(
dnnPrimitiveAttributes_t *attributes);
Expand Down Expand Up @@ -186,14 +202,20 @@ TEMPLATE_PREFIX dnnError_t dnnDelete(
dnnPrimitive_t& primitive);
SPEC_PREFIX dnnError_t dnnDelete<float>(
dnnPrimitive_t& primitive) {
dnnError_t status = dnnDelete_F32(primitive);
primitive = NULL;
dnnError_t status = E_SUCCESS;
if (primitive != NULL) {
status = dnnDelete_F32(primitive);
primitive = NULL;
}
return status;
}
SPEC_PREFIX dnnError_t dnnDelete<double>(
dnnPrimitive_t& primitive) {
dnnError_t status = dnnDelete_F64(primitive);
primitive = NULL;
dnnError_t status = E_SUCCESS;
if (primitive != NULL) {
status = dnnDelete_F64(primitive);
primitive = NULL;
}
return status;
}

Expand Down
11 changes: 8 additions & 3 deletions src/caffe/blob.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -68,19 +68,24 @@ void Blob<Dtype>::Reshape(const vector<int>& shape) {
}
int* shape_data = static_cast<int*>(shape_data_->mutable_cpu_data());
#endif

bool actual_reshaping = false;
for (int i = 0; i < shape.size(); ++i) {
CHECK_GE(shape[i], 0);
if (count_ != 0) {
CHECK_LE(shape[i], INT_MAX / count_) << "blob size exceeds INT_MAX";
}
count_ *= shape[i];
shape_[i] = shape[i];
if (shape_[i] != shape[i]) {
actual_reshaping = true;
shape_[i] = shape[i];
}
#ifndef CPU_ONLY
shape_data[i] = shape[i];
#endif
}
if (count_ > capacity_) {
// We restart sync objects when there was change of shape
// requested count is bgger than current capacity
if ( (actual_reshaping == true) || (count_ > capacity_) ) {
capacity_ = count_;
data_.reset(new SyncedMemory(capacity_ * sizeof(Dtype)));
diff_.reset(new SyncedMemory(capacity_ * sizeof(Dtype)));
Expand Down
2 changes: 1 addition & 1 deletion src/caffe/layers/inner_product_layer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -183,6 +183,6 @@ STUB_GPU(InnerProductLayer);

INSTANTIATE_CLASS(InnerProductLayer);

//REGISTER_LAYER_CLASS(InnerProduct);
// REGISTER_LAYER_CLASS(InnerProduct);

} // namespace caffe
56 changes: 40 additions & 16 deletions src/caffe/layers/mkl_batch_norm_layer.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -47,17 +47,17 @@ namespace caffe {

template <typename Dtype>
MKLBatchNormLayer<Dtype>::~MKLBatchNormLayer() {
if (batchNormFwd != NULL) dnnDelete<Dtype>(batchNormFwd);
if (batchNormBwdData != NULL) dnnDelete<Dtype>(batchNormBwdData);
if (batchNormBwdScaleShift != NULL) dnnDelete<Dtype>(batchNormBwdScaleShift);
dnnDelete<Dtype>(batchNormFwd);
dnnDelete<Dtype>(batchNormBwdData);
dnnDelete<Dtype>(batchNormBwdScaleShift);

dnnLayoutDelete<Dtype>(layout_usr_);
dnnReleaseBuffer<Dtype>(workspace_buffer_);
dnnReleaseBuffer<Dtype>(scaleShift_buffer_);
}

template <typename Dtype>
void MKLBatchNormLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
void MKLBatchNormLayer<Dtype>::Init(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
eps_ = this->layer_param_.batch_norm_param().eps();
use_weight_bias_ = this->layer_param_.batch_norm_param().use_weight_bias();
Expand Down Expand Up @@ -93,24 +93,28 @@ void MKLBatchNormLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
bwd_bottom_diff->name = "bwd_bottom_diff @ " + this->layer_param_.name();
bwd_top_diff->name = "bwd_top_diff @ " + this->layer_param_.name();

// TODO: Make a cleanup routine to avoid
// copy of following code in the Destructor

dnnError_t e;
dnnLayoutDelete<Dtype>(layout_usr_);
e = dnnLayoutCreate<Dtype>(&layout_usr_, dim, sizes, strides);
CHECK_EQ(e, E_SUCCESS);

fwd_bottom_data->create_user_layout(dim, sizes, strides);
fwd_top_data ->create_user_layout(dim, sizes, strides);
bwd_bottom_diff->create_user_layout(dim, sizes, strides);
bwd_top_diff ->create_user_layout(dim, sizes, strides);
fwd_bottom_data->create_user_layout(dim, sizes, strides, false);
fwd_top_data ->create_user_layout(dim, sizes, strides, false);
bwd_bottom_diff->create_user_layout(dim, sizes, strides, false);
bwd_top_diff ->create_user_layout(dim, sizes, strides, false);

workspace_buffer_ = NULL;
scaleShift_buffer_ = NULL;
dnnReleaseBuffer<Dtype>(workspace_buffer_);
dnnReleaseBuffer<Dtype>(scaleShift_buffer_);
// "Lazy" allocation because here we don't know
// what layout is used by neighbours.

// Primitives will be allocated during the first fwd pass
batchNormFwd = NULL;
batchNormBwdData = NULL;
batchNormBwdScaleShift = NULL;
dnnDelete<Dtype>(batchNormFwd);
dnnDelete<Dtype>(batchNormBwdData);
dnnDelete<Dtype>(batchNormBwdScaleShift);

if (use_weight_bias_) {
if ( bias_term_ ) {
Expand Down Expand Up @@ -147,10 +151,23 @@ void MKLBatchNormLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
}
}

template <typename Dtype>
void MKLBatchNormLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
Init(bottom, top);
}

template <typename Dtype>
void MKLBatchNormLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
bool reshaping = true;
if ((num_ == bottom[0]->num()) &&
channels_ == bottom[0]->channels() &&
height_ == bottom[0]->height() &&
width_ == bottom[0]->width()) {
reshaping = false;
}

if (bottom[0] == top[0]) { // in-place computation
temp_.ReshapeLike(*bottom[0]);
} else {
Expand All @@ -160,6 +177,10 @@ void MKLBatchNormLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,
num_ = bottom[0]->num();
top[0]->Reshape(num_, channels_, height_, width_);
}

if (reshaping == true) {
Init(bottom, top);
}
}

template <typename Dtype>
Expand Down Expand Up @@ -302,13 +323,16 @@ void MKLBatchNormLayer<Dtype>::Backward_cpu(
const vector<Blob<Dtype>*>& bottom) {
void *bottom_data = NULL;
if (bottom[0] == top[0]) {
bottom_data = reinterpret_cast<void *>(const_cast<Dtype*>(temp_.cpu_data()));
bottom_data = reinterpret_cast<void *>(
const_cast<Dtype*>(temp_.cpu_data()));
} else {
bottom_data =
reinterpret_cast<void *>(const_cast<Dtype*>(bottom[0]->prv_data()));
reinterpret_cast<void *>(
const_cast<Dtype*>(bottom[0]->prv_data()));
if (NULL == bottom_data)
bottom_data =
reinterpret_cast<void *>(const_cast<Dtype*>(bottom[0]->cpu_data()));
reinterpret_cast<void *>(
const_cast<Dtype*>(bottom[0]->cpu_data()));
}

dnnError_t e;
Expand Down
Loading

0 comments on commit 7c6249d

Please sign in to comment.