wehub-resource-sync
230 lines
10 KiB
Plaintext
230 lines
10 KiB
Plaintext
/* ******************************************************************************
|
|
*
|
|
*
|
|
* 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 <helpers/PointersManager.h>
|
|
#include <math/platformmath.h>
|
|
#include <math/templatemath.h>
|
|
#include <ops/declarable/helpers/updatersHelpers.h>
|
|
#include <system/op_boilerplate.h>
|
|
|
|
#include "execution/cuda/LaunchDims.h"
|
|
#include "helpers/DebugHelper.h"
|
|
|
|
|
|
namespace sd {
|
|
namespace ops {
|
|
namespace helpers {
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
template <typename T>
|
|
SD_KERNEL void amsGradUpdaterCuda(const void* vx, const LongType* xShapeInfo, const void* vinv,
|
|
const LongType* invShapeInfo, const void* vinm, const LongType* inmShapeInfo,
|
|
const void* vinh, const LongType* inhShapeInfo, void* vz,
|
|
const LongType* zShapeInfo, void* vstV, const LongType* stvShapeInfo,
|
|
void* vstM, const LongType* stmShapeInfo, void* vstH,
|
|
const LongType* sthShapeInfo, const T lr, const T beta1, const T beta2,
|
|
const T epsilon, const T iteration) {
|
|
const auto grad = reinterpret_cast<const T*>(vx);
|
|
const auto initV = reinterpret_cast<const T*>(vinv);
|
|
const auto initM = reinterpret_cast<const T*>(vinm);
|
|
const auto initH = reinterpret_cast<const T*>(vinh);
|
|
auto up = reinterpret_cast<T*>(vz);
|
|
auto stV = reinterpret_cast<T*>(vstV);
|
|
auto stM = reinterpret_cast<T*>(vstM);
|
|
auto stH = reinterpret_cast<T*>(vstH);
|
|
|
|
__shared__ LongType xLen, xRank, zRank, invRank, inmRank, inhRank, stvRank, stmRank, sthRank;
|
|
__shared__ T mbeta1, mbeta2, epsilonT;
|
|
__shared__ bool bOrdering, bXZsame, bXInUSame, bXStUSame, bXInMSame, bXStMSame, bXInHSame, bXStHSame;
|
|
__shared__ LongType *sharedMem;
|
|
__shared__ const LongType *xShape, *zShape, *invShape, *inmShape, *inhShape, *stvShape, *stmShape, *sthShape;
|
|
__shared__ const LongType *xStride, *zStride, *invStride, *inmStride, *inhStride, *stvStride, *stmStride, *sthStride;
|
|
|
|
if (threadIdx.x == 0) {
|
|
extern __shared__ unsigned char shmem[];
|
|
sharedMem = reinterpret_cast<LongType*>(shmem);
|
|
|
|
xLen = shape::length(xShapeInfo);
|
|
|
|
epsilonT = lr * math::sd_sqrt<T, T>(1.0 - math::sd_pow<T, T, T>(beta2, (iteration + 1))) /
|
|
(1.0 - math::sd_pow<T, T, T>(beta1, (iteration + 1)));
|
|
|
|
if (math::sd_isnan(epsilonT) || 0 == epsilonT || math::sd_isinf(epsilonT)) epsilonT = epsilon;
|
|
|
|
mbeta1 = (1 - beta1);
|
|
mbeta2 = (1 - beta2);
|
|
|
|
xRank = shape::rank(xShapeInfo);
|
|
zRank = shape::rank(zShapeInfo);
|
|
invRank = shape::rank(invShapeInfo);
|
|
inmRank = shape::rank(inmShapeInfo);
|
|
inhRank = shape::rank(inhShapeInfo);
|
|
stvRank = shape::rank(stvShapeInfo);
|
|
stmRank = shape::rank(stmShapeInfo);
|
|
sthRank = shape::rank(sthShapeInfo);
|
|
|
|
xShape = shape::shapeOf(xShapeInfo);
|
|
xStride = shape::stride(xShapeInfo);
|
|
zShape = shape::shapeOf(zShapeInfo);
|
|
zStride = shape::stride(zShapeInfo);
|
|
invShape = shape::shapeOf(invShapeInfo);
|
|
invStride = shape::stride(invShapeInfo);
|
|
inmShape = shape::shapeOf(inmShapeInfo);
|
|
inmStride = shape::stride(inmShapeInfo);
|
|
inhShape = shape::shapeOf(inhShapeInfo);
|
|
inhStride = shape::stride(inhShapeInfo);
|
|
stvShape = shape::shapeOf(stvShapeInfo);
|
|
stvStride = shape::stride(stvShapeInfo);
|
|
stmShape = shape::shapeOf(stmShapeInfo);
|
|
stmStride = shape::stride(stmShapeInfo);
|
|
sthShape = shape::shapeOf(sthShapeInfo);
|
|
sthStride = shape::stride(sthShapeInfo);
|
|
|
|
bOrdering = shape::order(xShapeInfo) == shape::order(zShapeInfo) &&
|
|
shape::order(zShapeInfo) == shape::order(stmShapeInfo) &&
|
|
shape::order(stmShapeInfo) == shape::order(inmShapeInfo) &&
|
|
shape::order(inmShapeInfo) == shape::order(stvShapeInfo) &&
|
|
shape::order(stvShapeInfo) == shape::order(invShapeInfo) &&
|
|
shape::order(invShapeInfo) == shape::order(sthShapeInfo) &&
|
|
shape::order(sthShapeInfo) == shape::order(inhShapeInfo);
|
|
|
|
bXZsame = shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo);
|
|
bXInUSame = shape::haveSameShapeAndStrides(xShapeInfo, invShapeInfo);
|
|
bXStUSame = shape::haveSameShapeAndStrides(xShapeInfo, stvShapeInfo);
|
|
bXInMSame = shape::haveSameShapeAndStrides(xShapeInfo, inmShapeInfo);
|
|
bXStMSame = shape::haveSameShapeAndStrides(xShapeInfo, stmShapeInfo);
|
|
bXInHSame = shape::haveSameShapeAndStrides(xShapeInfo, inhShapeInfo);
|
|
bXStHSame = shape::haveSameShapeAndStrides(xShapeInfo, sthShapeInfo);
|
|
}
|
|
__syncthreads();
|
|
|
|
LongType coords[SD_MAX_RANK];
|
|
|
|
for (LongType i = blockIdx.x * blockDim.x + threadIdx.x; i < xLen; i += gridDim.x * blockDim.x) {
|
|
LongType xOffset = i, zOffset = i, initMOffset = i, initVOffset = i, initHOffset = i, stMOffset = i, stVOffset = i,
|
|
stHOffset = i;
|
|
|
|
if (!bOrdering) {
|
|
INDEX2COORDS(i, xRank, xShape, coords);
|
|
COORDS2INDEX(xRank, xStride, coords, xOffset);
|
|
|
|
if (bXZsame) {
|
|
zOffset = xOffset;
|
|
} else {
|
|
COORDS2INDEX(zRank, zStride, coords, zOffset);
|
|
}
|
|
|
|
if (bXInMSame) {
|
|
initMOffset = xOffset;
|
|
} else {
|
|
COORDS2INDEX(inmRank, inmStride, coords, initMOffset);
|
|
}
|
|
|
|
if (bXStMSame) {
|
|
stMOffset = xOffset;
|
|
} else {
|
|
COORDS2INDEX(stmRank, stmStride, coords, stMOffset);
|
|
}
|
|
|
|
if (bXInUSame) {
|
|
initVOffset = xOffset;
|
|
} else {
|
|
COORDS2INDEX(invRank, invStride, coords, initVOffset);
|
|
}
|
|
|
|
if (bXStUSame) {
|
|
stVOffset = xOffset;
|
|
} else {
|
|
COORDS2INDEX(stvRank, stvStride, coords, stVOffset);
|
|
}
|
|
|
|
if (bXInHSame) {
|
|
initHOffset = xOffset;
|
|
} else {
|
|
COORDS2INDEX(inhRank, inhStride, coords, initHOffset);
|
|
}
|
|
|
|
if (bXStHSame) {
|
|
stHOffset = xOffset;
|
|
} else {
|
|
COORDS2INDEX(sthRank, sthStride, coords, stHOffset);
|
|
}
|
|
}
|
|
|
|
stM[stMOffset] = beta1 * initM[initMOffset] + grad[xOffset] * mbeta1;
|
|
stV[stVOffset] = beta2 * initV[initVOffset] + grad[xOffset] * grad[xOffset] * mbeta2;
|
|
stH[stHOffset] = math::sd_max(initH[initHOffset], stV[stVOffset]);
|
|
up[zOffset] = epsilonT * stM[stMOffset] / (math::sd_sqrt<T, T>(stH[stHOffset]) + epsilon);
|
|
}
|
|
}
|
|
///////////////////////////////////////////////////////////////////
|
|
template <typename T>
|
|
void amsGradUpdaterCudaLauncher(const int blocksPerGrid, const int threadsPerBlock, const int sharedMemory,
|
|
const cudaStream_t* stream, const void* vx, const LongType* xShapeInfo,
|
|
const void* vinv, const LongType* invShapeInfo, const void* vinm,
|
|
const LongType* inmShapeInfo, const void* vinh, const LongType* inhShapeInfo,
|
|
void* vz, const LongType* zShapeInfo, void* vstV, const LongType* stvShapeInfo,
|
|
void* vstM, const LongType* stmShapeInfo, void* vstH,
|
|
const LongType* sthShapeInfo, const double dLr, const double dBeta1,
|
|
const double dBeta2, const double dEpsilon, const int nIteration) {
|
|
const T lr = static_cast<T>(dLr);
|
|
const T beta1 = static_cast<T>(dBeta1);
|
|
const T beta2 = static_cast<T>(dBeta2);
|
|
T epsilon = static_cast<T>(dEpsilon);
|
|
//fp16 to prevent underflow
|
|
if(epsilon == 0.0) {
|
|
epsilon = static_cast<T>(1e-7);
|
|
}
|
|
const T iteration = static_cast<T>(nIteration);
|
|
|
|
amsGradUpdaterCuda<T><<<blocksPerGrid, threadsPerBlock, sharedMemory, *stream>>>(
|
|
vx, xShapeInfo, vinv, invShapeInfo, vinm, inmShapeInfo, vinh, inhShapeInfo, vz, zShapeInfo, vstV, stvShapeInfo,
|
|
vstM, stmShapeInfo, vstH, sthShapeInfo, lr, beta1, beta2, epsilon, iteration);
|
|
sd::DebugHelper::checkErrorCode(const_cast<cudaStream_t *>(stream), "amsGradUpdaterCudaLauncher failed");
|
|
|
|
}
|
|
|
|
///////////////////////////////////////////////////////////////////
|
|
void updaterAmsGrad(LaunchContext* context, NDArray& gradient, NDArray& initStateV,
|
|
NDArray& initStateM, NDArray& initStateH, NDArray& update, NDArray& stateV,
|
|
NDArray& stateM, NDArray& stateH, const double dLr, const double dBeta1, const double dBeta2,
|
|
const double dEpsilon, const int nIteration) {
|
|
PointersManager manager(context, "amsGradUpdater");
|
|
dim3 launchDims = updaterDims(gradient.lengthOf());
|
|
NDArray::prepareSpecialUse({&update, &stateV, &stateM, &stateH}, {&gradient, &initStateV, &initStateM, &initStateH});
|
|
BUILD_SINGLE_SELECTOR(
|
|
gradient.dataType(), amsGradUpdaterCudaLauncher,
|
|
(launchDims.y, launchDims.x,launchDims.z, context->getCudaStream(), gradient.specialBuffer(), gradient.specialShapeInfo(),
|
|
initStateV.specialBuffer(), initStateV.specialShapeInfo(), initStateM.specialBuffer(),
|
|
initStateM.specialShapeInfo(), initStateH.specialBuffer(), initStateH.specialShapeInfo(), update.specialBuffer(),
|
|
update.specialShapeInfo(), stateV.specialBuffer(), stateV.specialShapeInfo(), stateM.specialBuffer(),
|
|
stateM.specialShapeInfo(), stateH.specialBuffer(), stateH.specialShapeInfo(), dLr, dBeta1, dBeta2, dEpsilon,
|
|
nIteration),
|
|
SD_FLOAT_TYPES);
|
|
NDArray::registerSpecialUse({&update, &stateV, &stateM, &stateH}, {&gradient, &initStateV, &initStateM, &initStateH});
|
|
|
|
manager.synchronize();
|
|
}
|
|
|
|
} // namespace helpers
|
|
} // namespace ops
|
|
} // namespace sd
|