422 lines
20 KiB
C++
422 lines
20 KiB
C++
/*******************************************************************************
|
|
* Copyright (c) 2015-2018 Skymind, Inc.
|
|
*
|
|
* This program and the accompanying materials are made available under the
|
|
* terms of the Apache License, Version 2.0 which is available at
|
|
* https://www.apache.org/licenses/LICENSE-2.0.
|
|
*
|
|
* Unless required by applicable law or agreed to in writing, software
|
|
* distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
|
|
* WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
|
|
* License for the specific language governing permissions and limitations
|
|
* under the License.
|
|
*
|
|
* SPDX-License-Identifier: Apache-2.0
|
|
******************************************************************************/
|
|
|
|
//
|
|
// @author raver119@gmail.com
|
|
// @author Yurii Shyrma (iuriish@yahoo.com)
|
|
//
|
|
|
|
#include <ops/declarable/helpers/lrn.h>
|
|
#include <Status.h>
|
|
#include <ConstantTadHelper.h>
|
|
|
|
namespace nd4j {
|
|
namespace ops {
|
|
namespace helpers {
|
|
|
|
#ifdef HAVE_MKLDNN
|
|
using namespace mkldnn;
|
|
|
|
static void getMKLDNNMemoryDescLrn(const NDArray* src, const NDArray* diff_src, const NDArray* dst,
|
|
mkldnn::memory::desc* lrn_src_md, mkldnn::memory::desc* lrn_diff_src_md, mkldnn::memory::desc* lrn_dst_md,
|
|
mkldnn::memory::desc* user_src_md, mkldnn::memory::desc* user_diff_src_md, mkldnn::memory::desc* user_dst_md, int axis) {
|
|
const Nd4jLong* shape = src->getShapeInfo();
|
|
long rank = shape[0];
|
|
long dim1 = axis; // MKL-DNN supports only 1 axis, which has to be the "channel" one
|
|
long dim2 = axis >= 2 ? 1 : 2;
|
|
long dim3 = axis >= 3 ? 2 : 3;
|
|
mkldnn::memory::dims lrn_src_tz = { (int)shape[1], (int)shape[dim1 + 1], rank > 2 ? (int)shape[dim2 + 1] : 1, rank > 3 ? (int)shape[dim3 + 1] : 1};
|
|
|
|
auto type = mkldnn::memory::data_type::f32;
|
|
auto format = axis == 1 ? mkldnn::memory::format::nchw : mkldnn::memory::format::nhwc;
|
|
auto supposed_to_be_any_format = format; // doesn't work with "any"
|
|
|
|
if (src != nullptr && src->getBuffer() != nullptr && lrn_src_md != nullptr) {
|
|
*lrn_src_md = mkldnn::memory::desc({ lrn_src_tz }, type, supposed_to_be_any_format);
|
|
*user_src_md = mkldnn::memory::desc({ lrn_src_tz }, type, format);
|
|
user_src_md->data.format = mkldnn_blocked;
|
|
user_src_md->data.layout_desc.blocking.strides[0][0] = src->stridesOf()[0];
|
|
user_src_md->data.layout_desc.blocking.strides[0][1] = src->stridesOf()[dim1];
|
|
user_src_md->data.layout_desc.blocking.strides[0][2] = rank > 2 ? src->stridesOf()[dim2] : 1;
|
|
user_src_md->data.layout_desc.blocking.strides[0][3] = rank > 3 ? src->stridesOf()[dim3] : 1;
|
|
}
|
|
|
|
if (diff_src != nullptr && diff_src->getBuffer() != nullptr && lrn_diff_src_md != nullptr) {
|
|
*lrn_diff_src_md = mkldnn::memory::desc({ lrn_src_tz }, type, supposed_to_be_any_format);
|
|
*user_diff_src_md = mkldnn::memory::desc({ lrn_src_tz }, type, format);
|
|
user_diff_src_md->data.format = mkldnn_blocked;
|
|
user_diff_src_md->data.layout_desc.blocking.strides[0][0] = diff_src->stridesOf()[0];
|
|
user_diff_src_md->data.layout_desc.blocking.strides[0][1] = diff_src->stridesOf()[dim1];
|
|
user_diff_src_md->data.layout_desc.blocking.strides[0][2] = rank > 2 ? diff_src->stridesOf()[dim2] : 1;
|
|
user_diff_src_md->data.layout_desc.blocking.strides[0][3] = rank > 3 ? diff_src->stridesOf()[dim3] : 1;
|
|
}
|
|
|
|
if (dst != nullptr && dst->getBuffer() != nullptr && lrn_dst_md != nullptr) {
|
|
*lrn_dst_md = mkldnn::memory::desc({ lrn_src_tz }, type, supposed_to_be_any_format);
|
|
*user_dst_md = mkldnn::memory::desc({ lrn_src_tz }, type, format);
|
|
user_dst_md->data.format = mkldnn_blocked;
|
|
user_dst_md->data.layout_desc.blocking.strides[0][0] = dst->stridesOf()[0];
|
|
user_dst_md->data.layout_desc.blocking.strides[0][1] = dst->stridesOf()[dim1];
|
|
user_dst_md->data.layout_desc.blocking.strides[0][2] = rank > 2 ? dst->stridesOf()[dim2] : 1;
|
|
user_dst_md->data.layout_desc.blocking.strides[0][3] = rank > 3 ? dst->stridesOf()[dim3] : 1;
|
|
}
|
|
}
|
|
#endif
|
|
|
|
template <typename T>
|
|
static int lrnFunctor_(nd4j::graph::Context& block, NDArray* input, NDArray* output, int depth, float bias, float alpha, float beta) {
|
|
|
|
#ifdef HAVE_MKLDNN
|
|
if (block.isUseMKLDNN() && nd4j::MKLDNNStream::isSupported({input, output})) {
|
|
std::vector<nd4j::MKLDNNStream>& streams = block.getMKLDNNStreams();
|
|
if (streams.empty()) {
|
|
streams.push_back(MKLDNNStream("lrn"));
|
|
}
|
|
|
|
if (streams[0].checkAndReset({input}, {output}, {(float)bias, (float)alpha, (float)beta}, {depth})) {
|
|
mkldnn_memory_desc_t empty;
|
|
mkldnn::memory::desc lrn_src_md(empty), lrn_dst_md(empty), user_src_md(empty), user_dst_md(empty);
|
|
|
|
getMKLDNNMemoryDescLrn(input, nullptr, output, &lrn_src_md, nullptr, &lrn_dst_md, &user_src_md, nullptr, &user_dst_md, input->rankOf() - 1);
|
|
|
|
auto lrn_desc = lrn_forward::desc(prop_kind::forward_inference, lrn_across_channels, lrn_src_md, (2 * depth + 1), alpha * (2 * depth + 1), beta, bias);
|
|
|
|
auto engine = streams[0].getEngine();
|
|
auto lrn_prim_desc = lrn_forward::primitive_desc(lrn_desc, engine);
|
|
auto user_src_memory = mkldnn::memory({user_src_md, engine}, input->buffer());
|
|
auto user_dst_memory = mkldnn::memory({user_dst_md, engine}, output->buffer());
|
|
|
|
auto lrn_src_memory = user_src_memory;
|
|
streams[0].addMemory(user_src_memory);
|
|
if (mkldnn::memory::primitive_desc(lrn_prim_desc.src_primitive_desc())
|
|
!= user_src_memory.get_primitive_desc()) {
|
|
lrn_src_memory = mkldnn::memory(lrn_prim_desc.src_primitive_desc());
|
|
streams[0].addMemory(lrn_src_memory);
|
|
streams[0].addOperation(reorder(user_src_memory, lrn_src_memory));
|
|
}
|
|
|
|
auto lrn_dst_memory = user_dst_memory;
|
|
streams[0].addMemory(user_dst_memory);
|
|
if (mkldnn::memory::primitive_desc(lrn_prim_desc.dst_primitive_desc())
|
|
!= user_dst_memory.get_primitive_desc()) {
|
|
lrn_dst_memory = mkldnn::memory(lrn_prim_desc.dst_primitive_desc());
|
|
streams[0].addMemory(lrn_dst_memory);
|
|
}
|
|
|
|
streams[0].addOperation(lrn_forward(lrn_prim_desc, lrn_src_memory, lrn_dst_memory));
|
|
|
|
if (mkldnn::memory::primitive_desc(lrn_prim_desc.dst_primitive_desc())
|
|
!= user_dst_memory.get_primitive_desc()) {
|
|
streams[0].addOperation(reorder(lrn_dst_memory, user_dst_memory));
|
|
}
|
|
}
|
|
|
|
streams[0].submitAndWait();
|
|
return ND4J_STATUS_OK;
|
|
}
|
|
#endif
|
|
nd4j_debug("MKL-DNN is not used for lrn!\n", 0);
|
|
|
|
const int rank = input->rankOf();
|
|
|
|
TadPack inTadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(input->getShapeInfo(), {rank - 1});
|
|
TadPack outTadPack;
|
|
|
|
if(shape::haveSameShapeAndStrides(input->getShapeInfo(), output->getShapeInfo()))
|
|
outTadPack = inTadPack;
|
|
else
|
|
outTadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(output->getShapeInfo(), {rank - 1});
|
|
|
|
const Nd4jLong numOfTads = inTadPack.numberOfTads();
|
|
const Nd4jLong tadLen = input->sizeAt(-1);
|
|
|
|
const Nd4jLong* inTadOffsets = inTadPack.primaryOffsets();
|
|
const Nd4jLong* outTadOffsets = outTadPack.primaryOffsets();
|
|
|
|
const Nd4jLong inTadEws = shape::elementWiseStride(inTadPack.primaryShapeInfo());
|
|
const Nd4jLong outTadEws = shape::elementWiseStride(outTadPack.primaryShapeInfo());
|
|
|
|
const T* inBuff = reinterpret_cast<T*>(input->getBuffer());
|
|
T* outBuff = reinterpret_cast<T*>(output->getBuffer());
|
|
|
|
const T tbias = static_cast<T>(bias);
|
|
const T tbeta = static_cast<T>(beta);
|
|
const T talpha = static_cast<T>(alpha);
|
|
|
|
if(inTadEws == 1 && outTadEws == 1) {
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_SIMD
|
|
for (uint i = 0; i < numOfTads; ++i) {
|
|
const T* x = inBuff + inTadOffsets[i];
|
|
T* y = outBuff + outTadOffsets[i];
|
|
|
|
T prev = 0;
|
|
|
|
// calculate squared sum of elements per each j-th element range [j - depth, j + depth + 1]
|
|
// we store each squared sum in corresponding element of y array
|
|
for (uint j = 0; j < tadLen; ++j) {
|
|
const uint begin = nd4j::math::nd4j_max<int>(0, j - depth);
|
|
const uint last = depth + j + 1;
|
|
const uint end = nd4j::math::nd4j_min<int>(last, tadLen);
|
|
|
|
if (j == 0) {
|
|
for (uint s = begin; s < end; ++s)
|
|
prev = prev + x[s] * x[s];
|
|
y[j] = prev;
|
|
}
|
|
else if (begin == 0 && last <= tadLen)
|
|
y[j] = prev + x[end - 1] * x[end - 1];
|
|
else if (begin > 0 && last <= tadLen)
|
|
y[j] = prev + x[end - 1] * x[end - 1] - x[begin - 1] * x[begin - 1];
|
|
else if (begin > 0 && last > tadLen)
|
|
y[j] = prev - x[begin - 1] * x[begin - 1];
|
|
else
|
|
y[j] = prev;
|
|
|
|
if(j != 0)
|
|
prev = y[j];
|
|
|
|
y[j] = x[j] / nd4j::math::nd4j_pow<T, T, T>(tbias + alpha * prev, tbeta);
|
|
}
|
|
}
|
|
}
|
|
else {
|
|
PRAGMA_OMP_PARALLEL_FOR_SIMD
|
|
for (uint i = 0; i < numOfTads; ++i) {
|
|
const T* x = inBuff + inTadOffsets[i];
|
|
T* y = outBuff + outTadOffsets[i];
|
|
|
|
T prev = 0;
|
|
|
|
// calculate squared sum of elements per each j-th element range [j - depth, j + depth + 1]
|
|
// we store each squared sum in corresponding element of y array
|
|
for (uint j = 0; j < tadLen; ++j) {
|
|
const uint begin = nd4j::math::nd4j_max<int>(0, j - depth);
|
|
const uint last = depth + j + 1;
|
|
const uint end = nd4j::math::nd4j_min<int>(last, tadLen);
|
|
|
|
if (j == 0) {
|
|
for (uint s = begin; s < end; ++s)
|
|
prev = prev + x[s*inTadEws] * x[s*inTadEws];
|
|
y[j*outTadEws] = prev;
|
|
}
|
|
else if (begin == 0 && last <= tadLen)
|
|
y[j*outTadEws] = prev + x[(end - 1)*inTadEws] * x[(end - 1)*inTadEws];
|
|
else if (begin > 0 && last <= tadLen)
|
|
y[j*outTadEws] = prev + x[(end - 1)*inTadEws] * x[(end - 1)*inTadEws] - x[(begin - 1)*inTadEws] * x[(begin - 1)*inTadEws];
|
|
else if (begin > 0 && last > tadLen)
|
|
y[j*outTadEws] = prev - x[(begin - 1)*inTadEws] * x[(begin - 1)*inTadEws];
|
|
else
|
|
y[j*outTadEws] = prev;
|
|
|
|
if(j != 0)
|
|
prev = y[j*outTadEws];
|
|
|
|
y[j*outTadEws] = x[j*inTadEws] / nd4j::math::nd4j_pow<T, T, T>(tbias + alpha * prev, tbeta);
|
|
}
|
|
}
|
|
}
|
|
return Status::OK();
|
|
}
|
|
|
|
BUILD_SINGLE_TEMPLATE(template int lrnFunctor_, (nd4j::graph::Context& block, NDArray* input, NDArray* output, int depth, float bias, float alpha, float beta), FLOAT_TYPES);
|
|
|
|
int lrnFunctor(nd4j::graph::Context& block, NDArray* input, NDArray* output, int depth, double bias, double alpha, double beta) {
|
|
BUILD_SINGLE_SELECTOR(input->dataType(), return lrnFunctor_, (block, input, output, depth, bias, alpha, beta), FLOAT_TYPES);
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
template <typename X, typename Y>
|
|
static void lrnBP_(const NDArray& input, const NDArray& gradO, NDArray& gradI, const int depth, const float bias, const float alpha, const float beta) {
|
|
|
|
const int rank = input.rankOf();
|
|
|
|
TadPack inTadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(input.getShapeInfo(), {rank - 1});
|
|
TadPack gradITadPack;
|
|
|
|
if(shape::haveSameShapeAndStrides(input.getShapeInfo(), gradI.getShapeInfo()))
|
|
gradITadPack = inTadPack;
|
|
else
|
|
gradITadPack = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(gradI.getShapeInfo(), {rank - 1});
|
|
|
|
const Nd4jLong numOfTads = inTadPack.numberOfTads();
|
|
const Nd4jLong tadLen = input.sizeAt(-1);
|
|
|
|
const Nd4jLong* inTadOffsets = inTadPack.primaryOffsets();
|
|
const Nd4jLong* gradITadOffsets = gradITadPack.primaryOffsets();
|
|
|
|
const Nd4jLong inTadEws = shape::elementWiseStride(inTadPack.primaryShapeInfo());
|
|
const Nd4jLong gradITadEws = shape::elementWiseStride(gradITadPack.primaryShapeInfo());
|
|
|
|
const X* inBuff = reinterpret_cast<X*>(input.getBuffer());
|
|
Y* gradIBuff = reinterpret_cast<Y*>(gradI.getBuffer());
|
|
|
|
const Y tbias = static_cast<Y>(bias);
|
|
const Y tbeta = static_cast<Y>(beta);
|
|
const Y talpha = static_cast<Y>(alpha);
|
|
const Y coeff = talpha * tbeta;
|
|
|
|
if(inTadEws == 1 && gradITadEws == 1) {
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_SIMD
|
|
for (uint i = 0; i < numOfTads; ++i) {
|
|
const X* x = inBuff + inTadOffsets[i];
|
|
Y* y = gradIBuff + gradITadOffsets[i];
|
|
|
|
// this loop calculates squared sum of elements per each j-th element range [j - depth, j + depth + 1]
|
|
// we store each squared sum in corresponding element of y array
|
|
for (uint j = 0; j < tadLen; ++j) {
|
|
const uint begin = nd4j::math::nd4j_max<int>(0, j - depth);
|
|
const uint last = depth + j + 1;
|
|
const uint end = nd4j::math::nd4j_min<int>(last, tadLen);
|
|
|
|
if (j == 0) {
|
|
y[0] = 0;
|
|
for (uint s = begin; s < end; ++s)
|
|
y[0] = y[0] + x[s] * x[s];
|
|
}
|
|
else if (begin == 0 && last <= tadLen)
|
|
y[j] = y[j - 1] + x[end - 1] * x[end - 1];
|
|
else if (begin > 0 && last <= tadLen)
|
|
y[j] = y[j - 1] + x[end - 1] * x[end - 1] - x[begin - 1] * x[begin - 1];
|
|
else if (begin > 0 && last > tadLen)
|
|
y[j] = y[j - 1] - x[begin - 1] * x[begin - 1];
|
|
else
|
|
y[j] = y[j - 1];
|
|
}
|
|
|
|
Y* factor = new Y[tadLen];
|
|
|
|
Y prev = 0;
|
|
// second loop calculates derivatives using information gained in first loop above
|
|
for (uint j = 0; j < tadLen; ++j) {
|
|
const uint begin = nd4j::math::nd4j_max<int>(0, j - depth);
|
|
const uint last = depth + j + 1;
|
|
const uint end = nd4j::math::nd4j_min<int>(last, tadLen);
|
|
|
|
Y init = tbias + talpha * y[j];
|
|
|
|
if (j == 0) {
|
|
for (uint s = begin; s < end; ++s) {
|
|
factor[s] = nd4j::math::nd4j_pow<Y, Y, Y>(tbias + talpha * y[s], -tbeta - 1);
|
|
prev = prev + x[s] * factor[s];
|
|
}
|
|
y[0] = prev;
|
|
}
|
|
else if(begin == 0 && last <= tadLen) {
|
|
factor[end - 1] = nd4j::math::nd4j_pow<Y, Y, Y>(tbias + talpha * y[end - 1], -tbeta - 1);
|
|
y[j] = prev + x[end - 1] * factor[end - 1];
|
|
}
|
|
else if (begin > 0 && last <= tadLen) {
|
|
factor[end - 1] = nd4j::math::nd4j_pow<Y, Y, Y>(tbias + talpha * y[end - 1], -tbeta - 1);
|
|
y[j] = prev + x[end - 1] * factor[end - 1] - x[begin - 1] * factor[begin - 1];
|
|
}
|
|
else if (begin > 0 && last > tadLen)
|
|
y[j] = prev - x[begin - 1] * factor[begin - 1];
|
|
else
|
|
y[j] = prev;
|
|
|
|
if(j != 0)
|
|
prev = y[j];
|
|
|
|
y[j] = factor[j] * init - 2 * x[j] * coeff * prev;
|
|
}
|
|
|
|
delete []factor;
|
|
}
|
|
}
|
|
else {
|
|
|
|
PRAGMA_OMP_PARALLEL_FOR_SIMD
|
|
for (uint i = 0; i < numOfTads; ++i) {
|
|
const X* x = inBuff + inTadOffsets[i];
|
|
Y* y = gradIBuff + gradITadOffsets[i];
|
|
|
|
// this loop calculates squared sum of elements per each j-th element range [j - depth, j + depth + 1]
|
|
// we store each squared sum in corresponding element of y array
|
|
for (uint j = 0; j < tadLen; ++j) {
|
|
const uint begin = nd4j::math::nd4j_max<int>(0, j - depth);
|
|
const uint last = depth + j + 1;
|
|
const uint end = nd4j::math::nd4j_min<int>(last, tadLen);
|
|
|
|
if (j == 0) {
|
|
y[0] = 0;
|
|
for (uint s = begin; s < end; ++s)
|
|
y[0] = y[0] + x[s*inTadEws] * x[s*inTadEws];
|
|
}
|
|
else if (begin == 0 && last <= tadLen)
|
|
y[j*gradITadEws] = y[(j - 1)*gradITadEws] + x[(end - 1)*inTadEws] * x[(end - 1)*inTadEws];
|
|
else if (begin > 0 && last <= tadLen)
|
|
y[j*gradITadEws] = y[(j - 1)*gradITadEws] + x[(end - 1)*inTadEws] * x[(end - 1)*inTadEws] - x[(begin - 1)*inTadEws] * x[(begin - 1)*inTadEws];
|
|
else if (begin > 0 && last > tadLen)
|
|
y[j*gradITadEws] = y[(j - 1)*gradITadEws] - x[(begin - 1)*inTadEws] * x[(begin - 1)*inTadEws];
|
|
else
|
|
y[j*gradITadEws] = y[(j - 1)*gradITadEws];
|
|
}
|
|
|
|
Y* factor = new Y[tadLen];
|
|
|
|
Y prev = 0;
|
|
// second loop calculates derivatives using information gained in first loop above
|
|
for (uint j = 0; j < tadLen; ++j) {
|
|
const uint begin = nd4j::math::nd4j_max<int>(0, j - depth);
|
|
const uint last = depth + j + 1;
|
|
const uint end = nd4j::math::nd4j_min<int>(last, tadLen);
|
|
|
|
Y init = tbias + talpha * y[j*gradITadEws];
|
|
|
|
if (j == 0) {
|
|
for (uint s = begin; s < end; ++s) {
|
|
factor[s] = nd4j::math::nd4j_pow<Y, Y, Y>(tbias + talpha * y[s*gradITadEws], -tbeta - 1);
|
|
prev = prev + x[s*inTadEws] * factor[s];
|
|
}
|
|
y[0] = prev;
|
|
}
|
|
else if(begin == 0 && last <= tadLen) {
|
|
factor[end - 1] = nd4j::math::nd4j_pow<Y, Y, Y>(tbias + talpha * y[(end - 1)*gradITadEws], -tbeta - 1);
|
|
y[j*gradITadEws] = prev + x[(end - 1)*inTadEws] * factor[end - 1];
|
|
}
|
|
else if (begin > 0 && last <= tadLen) {
|
|
factor[end - 1] = nd4j::math::nd4j_pow<Y, Y, Y>(tbias + talpha * y[(end - 1)*gradITadEws], -tbeta - 1);
|
|
y[j*gradITadEws] = prev + x[(end - 1)*inTadEws] * factor[end - 1] - x[(begin - 1)*inTadEws] * factor[begin - 1];
|
|
}
|
|
else if (begin > 0 && last > tadLen)
|
|
y[j*gradITadEws] = prev - x[(begin - 1)*inTadEws] * factor[begin - 1];
|
|
else
|
|
y[j*gradITadEws] = prev;
|
|
|
|
if(j != 0)
|
|
prev = y[j*gradITadEws];
|
|
|
|
y[j*gradITadEws] = factor[j] * init - 2 * x[j*inTadEws] * coeff * prev;
|
|
}
|
|
|
|
delete []factor;
|
|
}
|
|
}
|
|
gradI *= gradO;
|
|
}
|
|
|
|
BUILD_DOUBLE_TEMPLATE(template void lrnBP_, (const NDArray& input, const NDArray& gradO, NDArray& gradI, const int depth, const float bias, const float alpha, const float beta), LIBND4J_TYPES, FLOAT_TYPES);
|
|
|
|
void lrnBP(nd4j::graph::Context& block, const NDArray& input, const NDArray& gradO, NDArray& gradI, const int depth, const float bias, const float alpha, const float beta) {
|
|
BUILD_DOUBLE_SELECTOR(input.dataType(), gradO.dataType(), lrnBP_, (input, gradO, gradI, depth, bias, alpha, beta), LIBND4J_TYPES, FLOAT_TYPES);
|
|
}
|
|
|
|
}
|
|
}
|
|
}
|