226 lines
8.2 KiB
C++
226 lines
8.2 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 <system/op_boilerplate.h>
|
|
#if NOT_EXCLUDED(OP_multiply)
|
|
|
|
#include <ops/declarable/CustomOperations.h>
|
|
|
|
namespace sd {
|
|
namespace ops {
|
|
|
|
BROADCASTABLE_OP_IMPL(multiply, 0, 0) {
|
|
auto x = INPUT_VARIABLE(0);
|
|
auto y = INPUT_VARIABLE(1);
|
|
auto z = OUTPUT_VARIABLE(0);
|
|
|
|
BROADCAST_CHECK_EMPTY(x,y,z);
|
|
|
|
Nd4jLong* zShapeInfo = nullptr;
|
|
const bool areShapesBroadcastable = ShapeUtils::evalBroadcastShapeInfo(x->getShapeInfo(), y->getShapeInfo(), true, zShapeInfo, block.getWorkspace());
|
|
REQUIRE_TRUE(areShapesBroadcastable, 0, "MULTIPLY OP: the shapes of x %s and y %s are not suitable for broadcast !", ShapeUtils::shapeAsString(x).c_str(), ShapeUtils::shapeAsString(y).c_str());
|
|
|
|
auto tZ = BroadcastHelper::broadcastApply(sd::BroadcastOpsTuple::Multiply(), x, y, z);
|
|
if (tZ == nullptr)
|
|
return ND4J_STATUS_KERNEL_FAILURE;
|
|
else if (tZ != z)
|
|
throw std::runtime_error("multiply: result was replaced");
|
|
|
|
return Status::OK();
|
|
}
|
|
DECLARE_SYN(Mul, multiply);
|
|
|
|
DECLARE_TYPES(multiply) {
|
|
getOpDescriptor()
|
|
->setAllowedInputTypes(0, DataType::ANY)
|
|
->setAllowedInputTypes(1, DataType::ANY)
|
|
->setAllowedOutputTypes(0, DataType::INHERIT);
|
|
}
|
|
|
|
DECLARE_TYPES(multiply_bp) {
|
|
getOpDescriptor()
|
|
->setAllowedInputTypes(DataType::ANY)
|
|
->setAllowedOutputTypes({ALL_FLOATS});
|
|
}
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
CUSTOM_OP_IMPL(multiply_bp, 3, 2, false, 0, 0) {
|
|
auto x = INPUT_VARIABLE(0);
|
|
auto y = INPUT_VARIABLE(1);
|
|
auto dLdz = INPUT_VARIABLE(2);
|
|
|
|
auto dLdx = OUTPUT_VARIABLE(0);
|
|
auto dLdy = OUTPUT_VARIABLE(1);
|
|
|
|
Nd4jLong* dLdzShapeInfo = nullptr;
|
|
const bool areShapesBroadcastable = ShapeUtils::evalBroadcastShapeInfo(x->getShapeInfo(), y->getShapeInfo(), true, dLdzShapeInfo, block.getWorkspace());
|
|
REQUIRE_TRUE(areShapesBroadcastable, 0, "MULTIPLY_BP OP: the shapes of x %s and y %s are not suitable for broadcast !", ShapeUtils::shapeAsString(x).c_str(), ShapeUtils::shapeAsString(y).c_str());
|
|
REQUIRE_TRUE(shape::equalsSoft(dLdz->shapeInfo(), dLdzShapeInfo), 0, "MULTIPLY_BP OP: wrong shape of next epsilon array (dLdOut), expected is %s, but got %s instead !", ShapeUtils::shapeAsString(dLdzShapeInfo).c_str(), ShapeUtils::shapeAsString(dLdz).c_str());
|
|
|
|
const Nd4jLong xLen = x->lengthOf();
|
|
const Nd4jLong yLen = y->lengthOf();
|
|
|
|
if(x->isScalar() && y->isScalar()) { // both are scalars
|
|
y->applyPairwiseTransform(pairwise::Multiply, *dLdz, *dLdx);
|
|
x->applyPairwiseTransform(pairwise::Multiply, *dLdz, *dLdy);
|
|
//dLdx->assign((*y) * (*dLdz));
|
|
//dLdy->assign((*x) * (*dLdz));
|
|
|
|
}
|
|
else if(x->isScalar()) { // x is scalar and y is not
|
|
dLdx->assign((*y * *dLdz).reduceNumber(reduce::Sum));
|
|
dLdz->applyScalarArr(scalar::Multiply, *x, *dLdy);
|
|
//dLdz->applyTrueBroadcast(broadcast::Multiply, x, dLdy, true);
|
|
}
|
|
else if(y->isScalar()) { // y is scalar and x is not
|
|
dLdy->assign((*x * *dLdz).reduceNumber(reduce::Sum));
|
|
dLdz->applyScalarArr(scalar::Multiply, *y, *dLdx);
|
|
}
|
|
else if(x->isSameShape(y)) {
|
|
x->applyPairwiseTransform(pairwise::Multiply, *dLdz, *dLdy);
|
|
y->applyPairwiseTransform(pairwise::Multiply, *dLdz, *dLdx);
|
|
}
|
|
else if (x->isSameShape(dLdz)) {
|
|
|
|
auto yTiled = NDArray(dLdz, false, block.launchContext());
|
|
y->tile(yTiled);
|
|
std::vector<int> axesForY = ShapeUtils::evalBroadcastBackwardAxis(y->getShapeInfo(), dLdz->getShapeInfo());
|
|
|
|
dLdy->assign( (*x * *dLdz).reduceAlongDimension(reduce::Sum, axesForY) );
|
|
yTiled.applyPairwiseTransform(pairwise::Multiply, *dLdz, *dLdx);
|
|
}
|
|
else if (y->isSameShape(dLdz)) {
|
|
|
|
auto xTiled = NDArray(dLdz, false, block.launchContext());
|
|
x->tile(xTiled);
|
|
std::vector<int> axesForX = ShapeUtils::evalBroadcastBackwardAxis(x->getShapeInfo(), dLdz->getShapeInfo());
|
|
|
|
dLdx->assign( (*y * *dLdz).reduceAlongDimension(reduce::Sum, axesForX) );
|
|
xTiled.applyPairwiseTransform(pairwise::Multiply, *dLdz, *dLdy);
|
|
}
|
|
else {
|
|
|
|
auto xTiled = NDArray(dLdz, false, block.launchContext());
|
|
auto yTiled = NDArray(dLdz, false, block.launchContext());
|
|
x->tile(xTiled);
|
|
y->tile(yTiled);
|
|
std::vector<int> axesForX = ShapeUtils::evalBroadcastBackwardAxis(x->getShapeInfo(), dLdz->getShapeInfo());
|
|
std::vector<int> axesForY = ShapeUtils::evalBroadcastBackwardAxis(y->getShapeInfo(), dLdz->getShapeInfo());
|
|
|
|
dLdx->assign( (*y * *dLdz).reduceAlongDimension(reduce::Sum, axesForX) );
|
|
dLdy->assign( (*x * *dLdz).reduceAlongDimension(reduce::Sum, axesForY) );
|
|
}
|
|
|
|
return Status::OK();
|
|
}
|
|
|
|
DECLARE_SHAPE_FN(multiply_bp) {
|
|
|
|
auto xShapeInfo = inputShape->at(0);
|
|
auto yShapeInfo = inputShape->at(1);
|
|
|
|
Nd4jLong *dLdxShapeInfo = nullptr;
|
|
Nd4jLong *dLdyShapeInfo = nullptr;
|
|
|
|
COPY_SHAPE(xShapeInfo, dLdxShapeInfo);
|
|
COPY_SHAPE(yShapeInfo, dLdyShapeInfo);
|
|
|
|
return SHAPELIST(CONSTANT(dLdxShapeInfo), CONSTANT(dLdyShapeInfo));
|
|
}
|
|
/*
|
|
CUSTOM_OP_IMPL(multiply_bp, 3, 2, false, 0, 0) {
|
|
auto x = INPUT_VARIABLE(0);
|
|
auto y = INPUT_VARIABLE(1);
|
|
auto epsNext = INPUT_VARIABLE(2);
|
|
|
|
auto gradX = OUTPUT_VARIABLE(0);
|
|
auto gradY = OUTPUT_VARIABLE(1);
|
|
|
|
auto lambdaX = LAMBDA_TT(_e, _y) {
|
|
return _e * _y;
|
|
};
|
|
|
|
auto lambdaY = LAMBDA_TT(_e, _x) {
|
|
return _e * _x;
|
|
};
|
|
|
|
|
|
if (x->isSameShape(y)) {
|
|
// PWT case case
|
|
|
|
// X gradient
|
|
epsNext->applyPairwiseLambda(y, lambdaX, gradX);
|
|
|
|
// Y gradient
|
|
epsNext->applyPairwiseLambda(x, lambdaY, gradY);
|
|
|
|
} else if (y->isScalar()) {
|
|
// scalar case
|
|
T _y = y->e(0);
|
|
auto lambdaS = LAMBDA_T(_e, _y) {
|
|
return _e * _y;
|
|
};
|
|
|
|
T tmpX = x->template reduceNumber<simdOps::Sum<T>>();
|
|
gradY->assign(tmpX);
|
|
|
|
epsNext->applyLambda(lambdaS, *gradX);
|
|
} else {
|
|
// broadcast case
|
|
|
|
auto preX = x->dup();
|
|
auto preY = y->dup();
|
|
|
|
auto targetShape = epsNext->getShapeAsVector();
|
|
|
|
preX->tileToShape(targetShape);
|
|
preY->tileToShape(targetShape);
|
|
|
|
auto axisX = ShapeUtils::evalBroadcastBackwardAxis(x->shapeInfo(), epsNext->shapeInfo());
|
|
auto axisY = ShapeUtils::evalBroadcastBackwardAxis(y->shapeInfo(), epsNext->shapeInfo());
|
|
|
|
if (axisX.size() > 0) {
|
|
auto sum = preX->template reduceAlongDimension<simdOps::Sum<T>>(axisX);
|
|
gradX->assign(sum);
|
|
delete sum;
|
|
} else
|
|
gradX->assign(preX);
|
|
|
|
if (axisY.size() > 0) {
|
|
auto sum = preY->template reduceAlongDimension<simdOps::Sum<T>>(axisY);
|
|
gradY->assign(sum);
|
|
delete sum;
|
|
} else
|
|
gradY->assign(preY);
|
|
|
|
|
|
delete preX;
|
|
delete preY;
|
|
}
|
|
|
|
return Status::OK();
|
|
}
|
|
*/
|
|
|
|
}
|
|
}
|
|
|
|
#endif |