Files
2026-07-13 12:47:05 +08:00

166 lines
6.4 KiB
C++

/*
* ******************************************************************************
* *
* *
* * 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.
* *
* * See the NOTICE file distributed with this work for additional
* * information regarding copyright ownership.
* * 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 Oleh Semeniv (oleg.semeniv@gmail.com)
//
#include <execution/Threads.h>
#include <math/platformmath.h>
#include <math/templatemath.h>
#include <ops/declarable/helpers/updatersHelpers.h>
#if NOT_EXCLUDED(OP_adadelta_updater)
namespace sd {
namespace ops {
namespace helpers {
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
template <typename T>
static void adaDeltaUpdater_(NDArray& gradient, NDArray& initStateMsg, NDArray& initStateMsdx,
NDArray& update, NDArray& stateMsg, NDArray& stateMsdx, const double dRho,
const double dEpsilon) {
// Cache shape information
const auto gradientShapeInfo = gradient.shapeInfo();
const auto updateShapeInfo = update.shapeInfo();
const auto initStateMsgShapeInfo = initStateMsg.shapeInfo();
const auto stateMsgShapeInfo = stateMsg.shapeInfo();
const auto initStateMsdxShapeInfo = initStateMsdx.shapeInfo();
const auto stateMsdxShapeInfo = stateMsdx.shapeInfo();
const auto gradRank = shape::rank(gradientShapeInfo);
const auto* gradShape = shape::shapeOf(gradientShapeInfo);
const auto* gradStride = shape::stride(gradientShapeInfo);
const auto* updateStride = shape::stride(updateShapeInfo);
const auto* initStateMsgStride = shape::stride(initStateMsgShapeInfo);
const auto* stateMsgStride = shape::stride(stateMsgShapeInfo);
const auto* initStateMsdxStride = shape::stride(initStateMsdxShapeInfo);
const auto* stateMsdxStride = shape::stride(stateMsdxShapeInfo);
const T* grad = gradient.bufferAsT<T>();
const T* initMsg = initStateMsg.bufferAsT<T>();
const T* initMsdx = initStateMsdx.bufferAsT<T>();
T* up = update.bufferAsT<T>();
T* stMsg = stateMsg.bufferAsT<T>();
T* stMsdx = stateMsdx.bufferAsT<T>();
const T rho = static_cast<T>(dRho);
T epsilon = static_cast<T>(dEpsilon);
//fp16 to prevent underflow
if(epsilon == 0.0) {
epsilon = static_cast<T>(1e-7);
}
const T rhoT = (1 - rho);
bool bEws1 = 1 == gradient.ews() && 1 == update.ews() && 1 == stateMsg.ews() && 1 == initStateMsg.ews() &&
1 == stateMsdx.ews() && 1 == initStateMsdx.ews();
bool bSameOrdering = gradient.ordering() == update.ordering() && update.ordering() == stateMsdx.ordering() &&
stateMsdx.ordering() == initStateMsdx.ordering() &&
stateMsdx.ordering() == initStateMsg.ordering() &&
stateMsg.ordering() == initStateMsg.ordering();
if (bEws1 && bSameOrdering) {
auto func = PRAGMA_THREADS_FOR {
for (auto i = start; i < stop; i++) {
stMsg[i] = rho * initMsg[i] + grad[i] * grad[i] * rhoT;
up[i] =
grad[i] * (sd::math::sd_sqrt<T, T>(initMsdx[i] + epsilon) / sd::math::sd_sqrt<T, T>(stMsg[i] + epsilon));
stMsdx[i] = rho * initMsdx[i] + up[i] * up[i] * rhoT;
}
};
samediff::Threads::parallel_for(func, 0, gradient.lengthOf(), 1);
return;
}
bool bXZsame = shape::haveSameShapeAndStrides(gradientShapeInfo, updateShapeInfo);
bool bXInMsgSame = shape::haveSameShapeAndStrides(gradientShapeInfo, initStateMsgShapeInfo);
bool bXStMsgSame = shape::haveSameShapeAndStrides(gradientShapeInfo, stateMsgShapeInfo);
bool bXInMsdxSame = shape::haveSameShapeAndStrides(gradientShapeInfo, initStateMsdxShapeInfo);
bool bXStMsdxSame = shape::haveSameShapeAndStrides(gradientShapeInfo, stateMsdxShapeInfo);
auto func = PRAGMA_THREADS_FOR {
sd::LongType coords[SD_MAX_RANK];
for (sd::LongType i = start; i < stop; i++) {
INDEX2COORDS(i, gradRank, gradShape, coords);
sd::LongType xOffset;
COORDS2INDEX(gradRank, gradStride, coords, xOffset);
sd::LongType zOffset;
if (bXZsame) {
zOffset = xOffset;
} else {
COORDS2INDEX(gradRank, updateStride, coords, zOffset);
}
sd::LongType initMsgOffset;
if (bXInMsgSame) {
initMsgOffset = xOffset;
} else {
COORDS2INDEX(gradRank, initStateMsgStride, coords, initMsgOffset);
}
sd::LongType stMsgOffset;
if (bXStMsgSame) {
stMsgOffset = xOffset;
} else {
COORDS2INDEX(gradRank, stateMsgStride, coords, stMsgOffset);
}
sd::LongType initMsdxOffset;
if (bXInMsdxSame) {
initMsdxOffset = xOffset;
} else {
COORDS2INDEX(gradRank, initStateMsdxStride, coords, initMsdxOffset);
}
sd::LongType stMsdxOffset;
if (bXStMsdxSame) {
stMsdxOffset = xOffset;
} else {
COORDS2INDEX(gradRank, stateMsdxStride, coords, stMsdxOffset);
}
stMsg[stMsgOffset] = rho * initMsg[initMsgOffset] + grad[xOffset] * grad[xOffset] * rhoT;
up[zOffset] = grad[xOffset] * (sd::math::sd_sqrt<T, T>(initMsdx[initMsdxOffset] + epsilon) /
sd::math::sd_sqrt<T, T>(stMsg[stMsgOffset] + epsilon));
stMsdx[stMsdxOffset] = rho * initMsdx[initMsdxOffset] + up[zOffset] * up[zOffset] * rhoT;
}
};
samediff::Threads::parallel_for(func, 0, gradient.lengthOf(), 1);
return;
}
void updaterAdaDelta(sd::LaunchContext* context, NDArray& gradient, NDArray& initStateMsg,
NDArray& initStateMsdx, NDArray& update, NDArray& stateMsg, NDArray& stateMsdx,
const double dRho, const double dEpsilon) {
BUILD_SINGLE_SELECTOR(gradient.dataType(), adaDeltaUpdater_,
(gradient, initStateMsg, initStateMsdx, update, stateMsg, stateMsdx, dRho, dEpsilon),
SD_FLOAT_TYPES);
}
} // namespace helpers
} // namespace ops
} // namespace sd
#endif