chore: import upstream snapshot with attribution
This commit is contained in:
wehub-resource-sync
@@ -0,0 +1,558 @@
|
||||
/* ******************************************************************************
|
||||
*
|
||||
*
|
||||
* 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 raver119@gmail.com
|
||||
// @author Yurii Shyrma (iuriish@yahoo.com)
|
||||
//
|
||||
#include <exceptions/cuda_exception.h>
|
||||
#include <execution/LaunchContext.h>
|
||||
#include <helpers/DebugHelper.h>
|
||||
#include <loops/legacy_ops.h>
|
||||
#include <loops/reduce_bool.h>
|
||||
#include <loops/scalar.h>
|
||||
#include <system/op_boilerplate.h>
|
||||
#include <system/common.h>
|
||||
#include <types/types.h>
|
||||
|
||||
using namespace simdOps;
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z, typename OpType>
|
||||
SD_KERNEL SD_INLINE void simpleReduce(
|
||||
const void* x,
|
||||
const sd::LongType* outerXTadShapeInfo,
|
||||
const sd::LongType* innerXTadShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo) {
|
||||
|
||||
functions::reduce::ReduceBoolFunction<X, Z>::template transformCuda<OpType>(
|
||||
x,
|
||||
outerXTadShapeInfo,
|
||||
innerXTadShapeInfo,
|
||||
vreductionBuffer,
|
||||
extraParams,
|
||||
z,
|
||||
zShapeInfo);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z, typename OpType>
|
||||
SD_KERNEL SD_INLINE void simpleScalar(
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
sd::LongType dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
functions::reduce::ReduceBoolFunction<X, Z>::template execScalarCuda<OpType>(
|
||||
x,
|
||||
xShapeInfo,
|
||||
extraParams,
|
||||
z,
|
||||
zShapeInfo,
|
||||
reductionBuffer,
|
||||
tadOnlyShapeInfo);
|
||||
}
|
||||
|
||||
namespace functions {
|
||||
namespace reduce {
|
||||
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_DEVICE SD_INLINE void ReduceBoolFunction<X, Z>::aggregatePartials(
|
||||
void* vsPartials,
|
||||
sd::LongType tid,
|
||||
sd::LongType numItems,
|
||||
void* vextraParams) {
|
||||
|
||||
auto sPartials = reinterpret_cast<Z*>(vsPartials);
|
||||
auto extraParams = reinterpret_cast<X*>(vextraParams);
|
||||
|
||||
sd::LongType floorPow2 = numItems;
|
||||
if (floorPow2 & (floorPow2 - 1)) {
|
||||
while (floorPow2 & (floorPow2 - 1)) {
|
||||
floorPow2 &= (floorPow2 - 1);
|
||||
}
|
||||
if (tid >= floorPow2) {
|
||||
sPartials[tid - floorPow2] =
|
||||
OpType::update(sPartials[tid - floorPow2], sPartials[tid], extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
for (sd::LongType activeThreads = (floorPow2 >> 1); activeThreads; activeThreads >>= 1) {
|
||||
if (tid < activeThreads && (tid + activeThreads) < numItems) {
|
||||
sPartials[tid] = OpType::update(sPartials[tid], sPartials[tid + activeThreads], extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_DEVICE SD_INLINE void ReduceBoolFunction<X, Z>::transformCuda(
|
||||
const void* vx,
|
||||
const sd::LongType* outerXTadShapeInfo,
|
||||
const sd::LongType* innerXTadShapeInfo,
|
||||
void* vreductionBuffer,
|
||||
void* vextraParams,
|
||||
void* vz,
|
||||
const sd::LongType* zShapeInfo) {
|
||||
|
||||
auto x = reinterpret_cast<const X*>(vx);
|
||||
auto z = reinterpret_cast<Z*>(vz);
|
||||
auto extraParams = reinterpret_cast<X*>(vextraParams);
|
||||
|
||||
__shared__ Z sPartials[SD_CUDA_BLOCK_SIZE];
|
||||
__shared__ int tadLen;
|
||||
__shared__ int numTads;
|
||||
__shared__ bool sameOffsets;
|
||||
|
||||
// Cache shape info
|
||||
__shared__ sd::LongType outerRank;
|
||||
__shared__ sd::LongType innerRank;
|
||||
__shared__ sd::LongType zRank;
|
||||
|
||||
__shared__ const sd::LongType* outerShapePtr;
|
||||
__shared__ const sd::LongType* outerStridePtr;
|
||||
|
||||
__shared__ const sd::LongType* innerShapePtr;
|
||||
__shared__ const sd::LongType* innerStridePtr;
|
||||
|
||||
__shared__ const sd::LongType* zShapePtr;
|
||||
__shared__ const sd::LongType* zStridePtr;
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
outerRank = shape::rank(outerXTadShapeInfo);
|
||||
outerShapePtr = shape::shapeOf(outerXTadShapeInfo);
|
||||
outerStridePtr = shape::stride(outerXTadShapeInfo);
|
||||
|
||||
innerRank = shape::rank(innerXTadShapeInfo);
|
||||
innerShapePtr = shape::shapeOf(innerXTadShapeInfo);
|
||||
innerStridePtr = shape::stride(innerXTadShapeInfo);
|
||||
|
||||
zRank = shape::rank(zShapeInfo);
|
||||
zShapePtr = shape::shapeOf(zShapeInfo);
|
||||
zStridePtr = shape::stride(zShapeInfo);
|
||||
|
||||
sameOffsets = shape::haveSameShapeAndStrides(zShapeInfo, outerXTadShapeInfo);
|
||||
tadLen = shape::length(innerXTadShapeInfo);
|
||||
numTads = shape::length(outerXTadShapeInfo);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
sd::LongType coords[SD_MAX_RANK];
|
||||
|
||||
for (sd::LongType r = blockIdx.x; r < numTads; r += gridDim.x) {
|
||||
INDEX2COORDS(r, outerRank, outerShapePtr, coords);
|
||||
|
||||
sd::LongType outerOffset;
|
||||
COORDS2INDEX(outerRank, outerStridePtr, coords, outerOffset);
|
||||
|
||||
sd::LongType zOffset;
|
||||
if (sameOffsets) {
|
||||
zOffset = outerOffset;
|
||||
} else {
|
||||
INDEX2COORDS(r, zRank, zShapePtr, coords);
|
||||
COORDS2INDEX(zRank, zStridePtr, coords, zOffset);
|
||||
}
|
||||
|
||||
const X* xTad = x + outerOffset;
|
||||
sPartials[threadIdx.x] = OpType::startingValue(xTad);
|
||||
|
||||
for (sd::LongType i = threadIdx.x; i < tadLen; i += blockDim.x) {
|
||||
sd::LongType iCoords[SD_MAX_RANK];
|
||||
sd::LongType innerOffset;
|
||||
|
||||
INDEX2COORDS(i, innerRank, innerShapePtr, iCoords);
|
||||
COORDS2INDEX(innerRank, innerStridePtr, iCoords, innerOffset);
|
||||
|
||||
sPartials[threadIdx.x] = OpType::update(
|
||||
sPartials[threadIdx.x],
|
||||
OpType::op(xTad[innerOffset], extraParams),
|
||||
extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
aggregatePartials<OpType>(
|
||||
sPartials,
|
||||
threadIdx.x,
|
||||
sd::math::sd_min<int>(blockDim.x, tadLen),
|
||||
extraParams);
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
z[zOffset] = OpType::postProcess(sPartials[0], tadLen, extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_DEVICE SD_INLINE void ReduceBoolFunction<X, Z>::execScalarCuda(
|
||||
const void* vx,
|
||||
const sd::LongType* xShapeInfo,
|
||||
void* vextraParams,
|
||||
void* vz,
|
||||
const sd::LongType* zShapeInfo,
|
||||
void* vreductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
auto x = reinterpret_cast<const X*>(vx);
|
||||
auto z = reinterpret_cast<Z*>(vz);
|
||||
auto extraParams = reinterpret_cast<X*>(vextraParams);
|
||||
|
||||
__shared__ Z sPartials[SD_CUDA_BLOCK_SIZE];
|
||||
__shared__ sd::LongType length;
|
||||
|
||||
// Cache shape info
|
||||
__shared__ sd::LongType xRank;
|
||||
__shared__ const sd::LongType* xShapePtr;
|
||||
__shared__ const sd::LongType* xStridePtr;
|
||||
auto reductionBuffer = reinterpret_cast<Z *>(vreductionBuffer);
|
||||
|
||||
int tid = blockDim.x * blockIdx.x + threadIdx.x;
|
||||
if (threadIdx.x == 0) {
|
||||
length = shape::length(xShapeInfo);
|
||||
xRank = shape::rank(xShapeInfo);
|
||||
xShapePtr = shape::shapeOf(xShapeInfo);
|
||||
xStridePtr = shape::stride(xShapeInfo);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
sPartials[threadIdx.x] = OpType::startingValue(x);
|
||||
|
||||
sd::LongType gridSize = gridDim.x * blockDim.x;
|
||||
for (sd::LongType i = tid; i < length; i += gridSize) {
|
||||
sd::LongType xCoords[SD_MAX_RANK];
|
||||
sd::LongType xOffset;
|
||||
|
||||
INDEX2COORDS(i, xRank, xShapePtr, xCoords);
|
||||
COORDS2INDEX(xRank, xStridePtr, xCoords, xOffset);
|
||||
|
||||
sPartials[threadIdx.x] = OpType::update(
|
||||
sPartials[threadIdx.x],
|
||||
OpType::op(x[xOffset], extraParams),
|
||||
extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
aggregatePartials<OpType>(
|
||||
sPartials,
|
||||
threadIdx.x,
|
||||
sd::math::sd_min<int>(blockDim.x, length),
|
||||
extraParams);
|
||||
__syncthreads();
|
||||
|
||||
if (gridDim.x > 1) {
|
||||
auto tc = reinterpret_cast<unsigned int*>(vreductionBuffer);
|
||||
__shared__ bool amLast;
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
reductionBuffer[blockIdx.x] = sPartials[0];
|
||||
}
|
||||
__threadfence();
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
unsigned int ticket = atomicInc(&tc[16384], gridDim.x);
|
||||
amLast = (ticket == (gridDim.x - 1));
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
if (amLast) {
|
||||
tc[16384] = 0;
|
||||
sPartials[threadIdx.x] = OpType::startingValue(x);
|
||||
|
||||
for (sd::LongType i = threadIdx.x; i < gridDim.x; i += blockDim.x) {
|
||||
sPartials[threadIdx.x] = OpType::update(
|
||||
sPartials[threadIdx.x],
|
||||
reinterpret_cast<Z*>(vreductionBuffer)[i],
|
||||
extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
aggregatePartials<OpType>(
|
||||
sPartials,
|
||||
threadIdx.x,
|
||||
sd::math::sd_min<int>(gridDim.x, blockDim.x),
|
||||
extraParams);
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
z[0] = OpType::postProcess(sPartials[0], length, extraParams);
|
||||
}
|
||||
}
|
||||
} else {
|
||||
if (threadIdx.x == 0) {
|
||||
auto tc = reinterpret_cast<unsigned int*>(vreductionBuffer);
|
||||
tc[16384] = 0;
|
||||
z[0] = OpType::postProcess(sPartials[0], length, extraParams);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_HOST SD_INLINE void ReduceBoolFunction<X, Z>::intermediate(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const void* x,
|
||||
sd::LongType* dXShapeInfo,
|
||||
sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* z,
|
||||
sd::LongType* dZShapeInfo,
|
||||
sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dims) {
|
||||
|
||||
if (shape::isEmptyConst(hXShapeInfo)) {
|
||||
if (shape::isEmptyConst(hZShapeInfo)) return;
|
||||
|
||||
const auto startingVal =
|
||||
static_cast<Z>(OpType::startingValue(reinterpret_cast<const X*>(x)));
|
||||
|
||||
auto res = cudaMemcpyAsync(
|
||||
sd::LaunchContext::defaultContext()->getScalarPointer(),
|
||||
&startingVal,
|
||||
sizeof(Z),
|
||||
cudaMemcpyHostToDevice,
|
||||
*stream);
|
||||
if (res != 0) {
|
||||
throw sd::cuda_exception::build(
|
||||
"ReduceBoolFunction<X,Z>::intermediate: failed to copy temporary scalar", res);
|
||||
}
|
||||
|
||||
auto ptr = sd::LaunchContext::defaultContext()->getScalarPointer();
|
||||
// scalar assign
|
||||
scalar::ScalarTransform<Z,Z,Z>::executeCudaShaped(
|
||||
launchDims,
|
||||
stream,
|
||||
14,
|
||||
z,
|
||||
dZShapeInfo,
|
||||
hZShapeInfo,
|
||||
z,
|
||||
dZShapeInfo,
|
||||
hZShapeInfo,
|
||||
ptr,
|
||||
nullptr);
|
||||
sd::DebugHelper::checkErrorCode(stream, "reduceBoolDim empty(...) failed");
|
||||
} else {
|
||||
const sd::LongType zRank = shape::rank(hZShapeInfo);
|
||||
const sd::LongType tadRank = shape::rank(hXShapeInfo) - zRank;
|
||||
|
||||
auto outerPack = sd::ConstantShapeHelper::getInstance()
|
||||
.createSubArrShapeInfo(hXShapeInfo, dims, zRank);
|
||||
auto innerPack = sd::ConstantShapeHelper::getInstance()
|
||||
.createSubArrShapeInfo(hXShapeInfo, dims + zRank, tadRank);
|
||||
|
||||
simpleReduce<X, Z, OpType>
|
||||
<<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(
|
||||
x,
|
||||
outerPack->special(),
|
||||
innerPack->special(),
|
||||
extraParams,
|
||||
vreductionBuffer,
|
||||
z,
|
||||
dZShapeInfo);
|
||||
sd::DebugHelper::checkErrorCode(stream, "reduceBoolDim(...) failed");
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_HOST SD_INLINE void ReduceBoolFunction<X, Z>::intermediateScalar(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
const sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo,
|
||||
const sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
sd::LongType dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
if (shape::isEmptyConst(hXShapeInfo)) {
|
||||
if (shape::isEmptyConst(hZShapeInfo)) return;
|
||||
|
||||
const auto startingVal =
|
||||
static_cast<Z>(OpType::startingValue(reinterpret_cast<const X*>(x)));
|
||||
auto res = cudaMemcpyAsync(
|
||||
z,
|
||||
&startingVal,
|
||||
sizeof(Z),
|
||||
cudaMemcpyHostToDevice,
|
||||
*stream);
|
||||
if (res != 0) {
|
||||
throw sd::cuda_exception::build(
|
||||
"ReduceBoolFunction<X,Z>::intermediateScalar: failed to copy resulting scalar", res);
|
||||
}
|
||||
sd::DebugHelper::checkErrorCode(stream, "reduceBoolScalar empty(...) failed");
|
||||
} else {
|
||||
simpleScalar<X, Z, OpType>
|
||||
<<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(
|
||||
x,
|
||||
xShapeInfo,
|
||||
extraParams,
|
||||
z,
|
||||
zShapeInfo,
|
||||
dimension,
|
||||
dimensionLength,
|
||||
reductionBuffer,
|
||||
tadOnlyShapeInfo);
|
||||
sd::DebugHelper::checkErrorCode(stream, "reduceBoolScalar(...) failed");
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Y>
|
||||
SD_HOST SD_INLINE void ReduceBoolFunction<X, Y>::execReduceScalar(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const int opNum,
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
const sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo,
|
||||
const sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
sd::LongType dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
DISPATCH_BY_OPNUM_TT(
|
||||
intermediateScalar,
|
||||
PARAMS(
|
||||
launchDims, stream, x, xShapeInfo, hXShapeInfo, extraParams, z,
|
||||
zShapeInfo, hZShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo),
|
||||
OPS_A(REDUCE_BOOL_OPS));
|
||||
sd::DebugHelper::checkErrorCode(stream, "execReduceScalarFloat(...) failed");
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Y>
|
||||
SD_HOST SD_INLINE void ReduceBoolFunction<X, Y>::execReduce(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const int opNum,
|
||||
const void* x,
|
||||
sd::LongType* dXShapeInfo,
|
||||
sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* z,
|
||||
sd::LongType* dZShapeInfo,
|
||||
sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dims) {
|
||||
|
||||
if (shape::length(hZShapeInfo) == 1) {
|
||||
execReduceScalar(
|
||||
launchDims, stream, opNum,
|
||||
x, dXShapeInfo, hXShapeInfo,
|
||||
extraParams, z,
|
||||
dZShapeInfo, hZShapeInfo,
|
||||
nullptr, 0,
|
||||
vreductionBuffer,
|
||||
nullptr);
|
||||
} else {
|
||||
DISPATCH_BY_OPNUM_TT(
|
||||
intermediate,
|
||||
PARAMS(
|
||||
launchDims, stream, x, dXShapeInfo, hXShapeInfo, extraParams,
|
||||
vreductionBuffer, z, dZShapeInfo, hZShapeInfo, dims),
|
||||
OPS_A(REDUCE_BOOL_OPS));
|
||||
}
|
||||
DEBUG_KERNEL(stream, opNum);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X>
|
||||
SD_DEVICE void initializeShared(X* extraParams, X** sPartials, int sMemSize) {
|
||||
int sPartialsLength = sMemSize / sizeof(X);
|
||||
X* sPartialsDeref = reinterpret_cast<X*>(*sPartials);
|
||||
for (int i = 0; i < sPartialsLength; i++) {
|
||||
sPartialsDeref[i] = extraParams[0];
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
ITERATE_COMBINATIONS(
|
||||
(SD_COMMON_TYPES),
|
||||
(SD_BOOL_TYPES),
|
||||
INSTANT_PROCESS_COMBINATION,
|
||||
functions::reduce::ReduceBoolFunction,
|
||||
::execReduce(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const int opNum,
|
||||
const void* x,
|
||||
sd::LongType* dXShapeInfo,
|
||||
sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* z,
|
||||
sd::LongType* dZShapeInfo,
|
||||
sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dims);
|
||||
);
|
||||
|
||||
ITERATE_COMBINATIONS(
|
||||
(SD_COMMON_TYPES),
|
||||
(SD_BOOL_TYPES),
|
||||
INSTANT_PROCESS_COMBINATION,
|
||||
functions::reduce::ReduceBoolFunction,
|
||||
::execReduceScalar(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const int opNum,
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
const sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo,
|
||||
const sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
sd::LongType dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo);
|
||||
);
|
||||
|
||||
|
||||
} // namespace reduce
|
||||
} // namespace functions
|
||||
@@ -0,0 +1,502 @@
|
||||
/* ******************************************************************************
|
||||
*
|
||||
*
|
||||
* 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 raver119@gmail.com
|
||||
//
|
||||
#include <execution/LaunchContext.h>
|
||||
#include <exceptions/cuda_exception.h>
|
||||
#include <system/op_boilerplate.h>
|
||||
#include <loops/reduce_float.h>
|
||||
#include <loops/scalar.h>
|
||||
#include <loops/legacy_ops.h>
|
||||
#include <helpers/DebugHelper.h>
|
||||
#include <types/types.h>
|
||||
#include <ops/specials_cuda.h>
|
||||
#include <cuda.h>
|
||||
#include <cuda_runtime.h>
|
||||
|
||||
using namespace simdOps;
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z, typename OpType>
|
||||
SD_KERNEL void simpleReduce(
|
||||
const void* x,
|
||||
const sd::LongType* outerXTadShapeInfo,
|
||||
const sd::LongType* innerXTadShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo) {
|
||||
|
||||
functions::reduce::ReduceFloatFunction<X,Z>::template transformCuda<OpType>(
|
||||
x,
|
||||
outerXTadShapeInfo,
|
||||
innerXTadShapeInfo,
|
||||
extraParams,
|
||||
vreductionBuffer,
|
||||
z,
|
||||
zShapeInfo);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z, typename OpType>
|
||||
SD_KERNEL void simpleScalar(
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
long long int dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
functions::reduce::ReduceFloatFunction<X, Z>::template execScalarCuda<OpType>(
|
||||
x,
|
||||
xShapeInfo,
|
||||
extraParams,
|
||||
z,
|
||||
zShapeInfo,
|
||||
reductionBuffer,
|
||||
tadOnlyShapeInfo);
|
||||
}
|
||||
|
||||
namespace functions {
|
||||
namespace reduce {
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_DEVICE void ReduceFloatFunction<X,Z>::aggregatePartials(
|
||||
void* vsPartials,
|
||||
sd::LongType tid,
|
||||
sd::LongType numItems,
|
||||
void* vextraParams) {
|
||||
|
||||
using Y = typename OpType::InterType;
|
||||
auto sPartials = reinterpret_cast<Y*>(vsPartials);
|
||||
auto extraParams = reinterpret_cast<Z*>(vextraParams);
|
||||
|
||||
sd::LongType floorPow2 = numItems;
|
||||
|
||||
if (floorPow2 & (floorPow2 - 1)) {
|
||||
while (floorPow2 & (floorPow2 - 1)) {
|
||||
floorPow2 &= (floorPow2 - 1);
|
||||
}
|
||||
if (tid >= floorPow2) {
|
||||
sPartials[tid - floorPow2] =
|
||||
OpType::update(sPartials[tid - floorPow2], sPartials[tid], extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
for (sd::LongType activeThreads = (floorPow2 >> 1); activeThreads; activeThreads >>= 1) {
|
||||
if (tid < activeThreads && (tid + activeThreads) < numItems) {
|
||||
sPartials[tid] =
|
||||
OpType::update(sPartials[tid], sPartials[tid + activeThreads], extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_DEVICE void ReduceFloatFunction<X,Z>::transformCuda(
|
||||
const void* vx,
|
||||
const sd::LongType* outerXTadShapeInfo,
|
||||
const sd::LongType* innerXTadShapeInfo,
|
||||
void* vextraParams,
|
||||
void* vreductionBuffer,
|
||||
void* vz,
|
||||
const sd::LongType* zShapeInfo) {
|
||||
|
||||
auto x = reinterpret_cast<const X*>(vx);
|
||||
auto z = reinterpret_cast<Z*>(vz);
|
||||
auto extraParams= reinterpret_cast<Z*>(vextraParams);
|
||||
|
||||
__shared__ Z sPartials[SD_CUDA_BLOCK_SIZE];
|
||||
__shared__ sd::LongType tadLen;
|
||||
__shared__ sd::LongType numTads;
|
||||
__shared__ bool sameOffsets;
|
||||
|
||||
// Cache ranks/shape/stride for outer and z
|
||||
__shared__ sd::LongType outerRank;
|
||||
__shared__ const sd::LongType* outerStridePtr;
|
||||
__shared__ const sd::LongType* outerShapePtr;
|
||||
|
||||
__shared__ sd::LongType zRank;
|
||||
__shared__ const sd::LongType* zStridePtr;
|
||||
__shared__ const sd::LongType* zShapePtr;
|
||||
|
||||
// Cache ranks/shape/stride for inner as well if needed
|
||||
__shared__ sd::LongType innerRank;
|
||||
__shared__ const sd::LongType* innerStridePtr;
|
||||
__shared__ const sd::LongType* innerShapePtr;
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
outerRank = shape::rank(outerXTadShapeInfo);
|
||||
outerShapePtr = shape::shapeOf(outerXTadShapeInfo);
|
||||
outerStridePtr = shape::stride(outerXTadShapeInfo);
|
||||
|
||||
zRank = shape::rank(zShapeInfo);
|
||||
zShapePtr = shape::shapeOf(zShapeInfo);
|
||||
zStridePtr = shape::stride(zShapeInfo);
|
||||
|
||||
innerRank = shape::rank(innerXTadShapeInfo);
|
||||
innerShapePtr = shape::shapeOf(innerXTadShapeInfo);
|
||||
innerStridePtr = shape::stride(innerXTadShapeInfo);
|
||||
|
||||
sameOffsets = shape::haveSameShapeAndStrides(zShapeInfo, outerXTadShapeInfo);
|
||||
tadLen = shape::length(innerXTadShapeInfo);
|
||||
numTads = shape::length(outerXTadShapeInfo);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
sd::LongType coords[SD_MAX_RANK];
|
||||
sd::LongType zCoords[SD_MAX_RANK];
|
||||
for (sd::LongType r = blockIdx.x; r < numTads; r += gridDim.x) {
|
||||
INDEX2COORDS(r, outerRank, outerShapePtr, coords);
|
||||
|
||||
sd::LongType outerOffset;
|
||||
COORDS2INDEX(outerRank, outerStridePtr, coords, outerOffset);
|
||||
|
||||
sd::LongType zOffset;
|
||||
if (sameOffsets) {
|
||||
zOffset = outerOffset;
|
||||
} else {
|
||||
INDEX2COORDS(r, zRank, zShapePtr, zCoords);
|
||||
COORDS2INDEX(zRank, zStridePtr, zCoords, zOffset);
|
||||
}
|
||||
|
||||
const X* xTad = x + outerOffset;
|
||||
sPartials[threadIdx.x] = OpType::startingValue(xTad);
|
||||
|
||||
// For the inner dimension
|
||||
for (sd::LongType i = threadIdx.x; i < tadLen; i += blockDim.x) {
|
||||
sd::LongType iCoords[SD_MAX_RANK];
|
||||
sd::LongType innerOffset;
|
||||
|
||||
INDEX2COORDS(i, innerRank, innerShapePtr, iCoords);
|
||||
COORDS2INDEX(innerRank, innerStridePtr, iCoords, innerOffset);
|
||||
|
||||
sPartials[threadIdx.x] = OpType::update(
|
||||
sPartials[threadIdx.x],
|
||||
OpType::op(xTad[innerOffset], extraParams),
|
||||
extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
aggregatePartials<OpType>(
|
||||
sPartials,
|
||||
threadIdx.x,
|
||||
sd::math::sd_min<int>(blockDim.x, tadLen),
|
||||
extraParams);
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
z[zOffset] =
|
||||
OpType::postProcess(sPartials[0], tadLen, extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_DEVICE void ReduceFloatFunction<X, Z>::execScalarCuda(
|
||||
const void* vx,
|
||||
const sd::LongType* xShapeInfo,
|
||||
void* vextraParams,
|
||||
void* vz,
|
||||
const sd::LongType* zShapeInfo,
|
||||
void* vreductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
auto x = reinterpret_cast<const X*>(vx);
|
||||
auto z = reinterpret_cast<Z*>(vz);
|
||||
auto extraParams = reinterpret_cast<Z*>(vextraParams);
|
||||
auto reductionBuffer = reinterpret_cast<Z*>(vreductionBuffer);
|
||||
|
||||
using Y = typename OpType::InterType;
|
||||
|
||||
__shared__ Y sPartials[SD_CUDA_BLOCK_SIZE];
|
||||
__shared__ sd::LongType length;
|
||||
|
||||
// Cache rank/shape/stride
|
||||
__shared__ sd::LongType xRank;
|
||||
__shared__ const sd::LongType* xShapePtr;
|
||||
__shared__ const sd::LongType* xStridePtr;
|
||||
|
||||
int tid = blockDim.x * blockIdx.x + threadIdx.x;
|
||||
if (threadIdx.x == 0) {
|
||||
length = shape::length(xShapeInfo);
|
||||
xRank = shape::rank(xShapeInfo);
|
||||
xShapePtr = shape::shapeOf(xShapeInfo);
|
||||
xStridePtr = shape::stride(xShapeInfo);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
sPartials[threadIdx.x] = OpType::startingValue(x);
|
||||
|
||||
sd::LongType gridSize = gridDim.x * blockDim.x;
|
||||
for (sd::LongType i = tid; i < length; i += gridSize) {
|
||||
sd::LongType xCoords[SD_MAX_RANK];
|
||||
sd::LongType xOffset;
|
||||
|
||||
INDEX2COORDS(i, xRank, xShapePtr, xCoords);
|
||||
COORDS2INDEX(xRank, xStridePtr, xCoords, xOffset);
|
||||
|
||||
if(xOffset < length) {
|
||||
sPartials[threadIdx.x] = OpType::update(
|
||||
sPartials[threadIdx.x],
|
||||
OpType::op(x[xOffset], extraParams),
|
||||
extraParams);
|
||||
}
|
||||
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
aggregatePartials<OpType>(
|
||||
sPartials,
|
||||
threadIdx.x,
|
||||
sd::math::sd_min<int>(blockDim.x, length),
|
||||
extraParams);
|
||||
__syncthreads();
|
||||
|
||||
if (gridDim.x > 1) {
|
||||
auto tc = reinterpret_cast<unsigned int*>(reductionBuffer);
|
||||
__shared__ bool amLast;
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
reductionBuffer[blockIdx.x] = sPartials[0];
|
||||
}
|
||||
__threadfence();
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
unsigned int ticket = atomicInc(&tc[16384], gridDim.x);
|
||||
amLast = (ticket == (gridDim.x - 1));
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
if (amLast) {
|
||||
tc[16384] = 0;
|
||||
sPartials[threadIdx.x] = OpType::startingValue(x);
|
||||
|
||||
for (sd::LongType i = threadIdx.x; i < gridDim.x; i += blockDim.x) {
|
||||
sPartials[threadIdx.x] =
|
||||
OpType::update(sPartials[threadIdx.x],
|
||||
reductionBuffer[i],
|
||||
extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
aggregatePartials<OpType>(
|
||||
sPartials,
|
||||
threadIdx.x,
|
||||
sd::math::sd_min<int>(gridDim.x, blockDim.x),
|
||||
extraParams);
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
z[0] = OpType::postProcess(sPartials[0], length, extraParams);
|
||||
}
|
||||
}
|
||||
} else {
|
||||
if (threadIdx.x == 0) {
|
||||
auto tc = reinterpret_cast<unsigned int*>(reductionBuffer);
|
||||
tc[16384] = 0;
|
||||
z[0] = OpType::postProcess(sPartials[0], length, extraParams);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template<typename OpType>
|
||||
SD_HOST void ReduceFloatFunction<X,Z>::intermediate(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const void* x,
|
||||
const sd::LongType* dXShapeInfo,
|
||||
const sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* z,
|
||||
const sd::LongType* dZShapeInfo,
|
||||
const sd::LongType* hZShapeInfo,
|
||||
const sd::LongType* dims) {
|
||||
|
||||
if (shape::isEmptyConst(hXShapeInfo)) {
|
||||
const auto startingVal = std::is_same<OpType, simdOps::Mean<X,Z>>::value
|
||||
? sd::DataTypeUtils::nanOrZero<Z>()
|
||||
: static_cast<Z>(OpType::startingValue(reinterpret_cast<const X*>(x)));
|
||||
|
||||
auto res = cudaMemcpyAsync(
|
||||
sd::LaunchContext::defaultContext()->getScalarPointer(),
|
||||
&startingVal,
|
||||
sizeof(Z),
|
||||
cudaMemcpyHostToDevice,
|
||||
*stream);
|
||||
if (res != 0) {
|
||||
throw sd::cuda_exception::build(
|
||||
"ReduceFloatFunction<X,Z>::intermediate: failed to copy temporary scalar", res);
|
||||
}
|
||||
|
||||
auto ptr = sd::LaunchContext::defaultContext()->getScalarPointer();
|
||||
// scalar assign
|
||||
functions::scalar::ScalarTransform<Z, Z, Z>::executeCudaShaped(
|
||||
launchDims,
|
||||
stream,
|
||||
14,
|
||||
z,
|
||||
dZShapeInfo,
|
||||
hZShapeInfo,
|
||||
z,
|
||||
dZShapeInfo,
|
||||
hZShapeInfo,
|
||||
ptr,
|
||||
nullptr);
|
||||
} else {
|
||||
const int zRank = shape::rank(hZShapeInfo);
|
||||
const int tadRank = shape::rank(hXShapeInfo) - zRank;
|
||||
|
||||
auto outerPack = sd::ConstantShapeHelper::getInstance().createSubArrShapeInfo(
|
||||
const_cast<sd::LongType*>(hXShapeInfo), const_cast<sd::LongType*>(dims), zRank);
|
||||
auto innerPack = sd::ConstantShapeHelper::getInstance().createSubArrShapeInfo(
|
||||
const_cast<sd::LongType*>(hXShapeInfo), const_cast<sd::LongType*>(dims + zRank), tadRank);
|
||||
|
||||
simpleReduce<X, Z, OpType>
|
||||
<<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(
|
||||
x,
|
||||
reinterpret_cast<const sd::LongType*>(outerPack->special()),
|
||||
reinterpret_cast<const sd::LongType*>(innerPack->special()),
|
||||
extraParams,
|
||||
vreductionBuffer,
|
||||
z,
|
||||
dZShapeInfo);
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template<typename OpType>
|
||||
SD_HOST void ReduceFloatFunction<X,Z>::intermediateScalar(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
const sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* dZShapeInfo,
|
||||
const sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
sd::LongType dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
if (shape::isEmptyConst(hXShapeInfo)) {
|
||||
const auto startingVal = std::is_same<OpType, simdOps::Mean<X,Z>>::value
|
||||
? sd::DataTypeUtils::nanOrZero<Z>()
|
||||
: static_cast<Z>(OpType::startingValue(reinterpret_cast<const X*>(x)));
|
||||
|
||||
auto res = cudaMemcpyAsync(z, &startingVal, sizeof(Z), cudaMemcpyHostToDevice, *stream);
|
||||
if (res != 0) {
|
||||
throw sd::cuda_exception::build(
|
||||
"ReduceFloatFunction<X,Z>::intermediateScalar: failed to copy resulting scalar", res);
|
||||
}
|
||||
} else {
|
||||
simpleScalar<X, Z, OpType><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(
|
||||
x,
|
||||
xShapeInfo,
|
||||
extraParams,
|
||||
z,
|
||||
dZShapeInfo,
|
||||
dimension,
|
||||
dimensionLength,
|
||||
reductionBuffer,
|
||||
tadOnlyShapeInfo);
|
||||
}
|
||||
sd::DebugHelper::checkErrorCode(stream, "ReduceFloatFunction intermediateScalar(...) failed");
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Y>
|
||||
SD_HOST void ReduceFloatFunction<X,Y>::execReduceScalar(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const int opNum,
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
const sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* dZShapeInfo,
|
||||
const sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
long long int dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
DISPATCH_BY_OPNUM_TT(
|
||||
intermediateScalar,
|
||||
PARAMS(
|
||||
launchDims, stream, x, xShapeInfo, hXShapeInfo, extraParams, z, dZShapeInfo, hZShapeInfo, dimension,
|
||||
dimensionLength, reductionBuffer, tadOnlyShapeInfo),
|
||||
OPS_A(REDUCE_FLOAT_OPS));
|
||||
sd::DebugHelper::checkErrorCode(stream, "execReduceScalarFloat(...) failed");
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Y>
|
||||
SD_HOST void ReduceFloatFunction<X,Y>::execReduce(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const int opNum,
|
||||
const void* x,
|
||||
const sd::LongType* dXShapeInfo,
|
||||
const sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* z,
|
||||
const sd::LongType* dZShapeInfo,
|
||||
const sd::LongType* hZShapeInfo,
|
||||
const sd::LongType* dims) {
|
||||
|
||||
if (shape::length(hZShapeInfo) == 1) {
|
||||
ReduceFloatFunction<X,Y>::execReduceScalar(
|
||||
launchDims, stream, opNum, x, dXShapeInfo, hXShapeInfo,
|
||||
extraParams, z, dZShapeInfo, hZShapeInfo,
|
||||
nullptr, 0, vreductionBuffer, nullptr);
|
||||
} else {
|
||||
DISPATCH_BY_OPNUM_TT(
|
||||
intermediate,
|
||||
PARAMS(
|
||||
launchDims, stream, x, dXShapeInfo, hXShapeInfo, extraParams, vreductionBuffer, z, dZShapeInfo,
|
||||
hZShapeInfo, dims),
|
||||
OPS_A(REDUCE_FLOAT_OPS));
|
||||
}
|
||||
sd::DebugHelper::checkErrorCode(stream, "ReduceFloatFunction execReduce(...) failed");
|
||||
}
|
||||
|
||||
} // namespace reduce
|
||||
} // namespace functions
|
||||
@@ -0,0 +1,552 @@
|
||||
/* ******************************************************************************
|
||||
*
|
||||
*
|
||||
* 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 raver119@gmail.com
|
||||
// @author Yurii Shyrma (iuriish@yahoo.com)
|
||||
//
|
||||
#include <exceptions/cuda_exception.h>
|
||||
#include <execution/LaunchContext.h>
|
||||
#include <helpers/DebugHelper.h>
|
||||
#include <loops/legacy_ops.h>
|
||||
#include <loops/reduce_long.h>
|
||||
#include <loops/scalar.h>
|
||||
#include <system/op_boilerplate.h>
|
||||
#include <types/types.h>
|
||||
|
||||
using namespace simdOps;
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z, typename OpType>
|
||||
SD_KERNEL SD_INLINE void simpleReduce(
|
||||
const void* x,
|
||||
const sd::LongType* outerXTadShapeInfo,
|
||||
const sd::LongType* innerXTadShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo) {
|
||||
|
||||
functions::reduce::ReduceLongFunction<X, Z>::template transformCuda<OpType>(
|
||||
x,
|
||||
outerXTadShapeInfo,
|
||||
innerXTadShapeInfo,
|
||||
extraParams,
|
||||
vreductionBuffer,
|
||||
z,
|
||||
zShapeInfo);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z, typename OpType>
|
||||
SD_DEVICE SD_INLINE void reduceScalarGeneric(
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
long long int dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
functions::reduce::ReduceLongFunction<X, Z>::template execScalarCuda<OpType>(
|
||||
x,
|
||||
xShapeInfo,
|
||||
extraParams,
|
||||
z,
|
||||
zShapeInfo,
|
||||
reductionBuffer,
|
||||
tadOnlyShapeInfo);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z, typename OpType>
|
||||
SD_KERNEL SD_INLINE void simpleScalar(
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
sd::LongType dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
reduceScalarGeneric<X, Z, OpType>(
|
||||
x,
|
||||
xShapeInfo,
|
||||
extraParams,
|
||||
z,
|
||||
zShapeInfo,
|
||||
dimension,
|
||||
dimensionLength,
|
||||
reductionBuffer,
|
||||
tadOnlyShapeInfo);
|
||||
}
|
||||
|
||||
namespace functions {
|
||||
namespace reduce {
|
||||
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_DEVICE SD_INLINE void ReduceLongFunction<X, Z>::aggregatePartials(
|
||||
void* vsPartials,
|
||||
sd::LongType tid,
|
||||
sd::LongType numItems,
|
||||
void* vextraParams) {
|
||||
|
||||
auto sPartials = reinterpret_cast<Z*>(vsPartials);
|
||||
auto extraParams = reinterpret_cast<X*>(vextraParams);
|
||||
|
||||
sd::LongType floorPow2 = numItems;
|
||||
if (floorPow2 & (floorPow2 - 1)) {
|
||||
while (floorPow2 & (floorPow2 - 1)) {
|
||||
floorPow2 &= floorPow2 - 1;
|
||||
}
|
||||
if (tid >= floorPow2) {
|
||||
sPartials[tid - floorPow2] =
|
||||
OpType::update(sPartials[tid - floorPow2], sPartials[tid], extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
for (sd::LongType activeThreads = floorPow2 >> 1; activeThreads; activeThreads >>= 1) {
|
||||
if (tid < activeThreads && (tid + activeThreads) < numItems) {
|
||||
sPartials[tid] = OpType::update(sPartials[tid], sPartials[tid + activeThreads], extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_DEVICE SD_INLINE void ReduceLongFunction<X, Z>::transformCuda(
|
||||
const void* vx,
|
||||
const sd::LongType* outerXTadShapeInfo,
|
||||
const sd::LongType* innerXTadShapeInfo,
|
||||
void* vextraParams,
|
||||
void* vreductionBuffer,
|
||||
void* vz,
|
||||
const sd::LongType* zShapeInfo) {
|
||||
|
||||
auto x = reinterpret_cast<const X*>(vx);
|
||||
auto z = reinterpret_cast<Z*>(vz);
|
||||
auto extraParams = reinterpret_cast<X*>(vextraParams);
|
||||
|
||||
__shared__ Z sPartials[SD_CUDA_BLOCK_SIZE];
|
||||
__shared__ int tadLen;
|
||||
__shared__ int numTads;
|
||||
__shared__ bool sameOffsets;
|
||||
|
||||
// Cache shape info for outer/inner and z if needed
|
||||
__shared__ sd::LongType outerRank;
|
||||
__shared__ sd::LongType innerRank;
|
||||
__shared__ sd::LongType zRank;
|
||||
|
||||
__shared__ const sd::LongType* outerShapePtr;
|
||||
__shared__ const sd::LongType* outerStridePtr;
|
||||
|
||||
__shared__ const sd::LongType* innerShapePtr;
|
||||
__shared__ const sd::LongType* innerStridePtr;
|
||||
|
||||
__shared__ const sd::LongType* zShapePtr;
|
||||
__shared__ const sd::LongType* zStridePtr;
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
outerRank = shape::rank(outerXTadShapeInfo);
|
||||
outerShapePtr = shape::shapeOf(outerXTadShapeInfo);
|
||||
outerStridePtr = shape::stride(outerXTadShapeInfo);
|
||||
|
||||
innerRank = shape::rank(innerXTadShapeInfo);
|
||||
innerShapePtr = shape::shapeOf(innerXTadShapeInfo);
|
||||
innerStridePtr = shape::stride(innerXTadShapeInfo);
|
||||
|
||||
zRank = shape::rank(zShapeInfo);
|
||||
zShapePtr = shape::shapeOf(zShapeInfo);
|
||||
zStridePtr = shape::stride(zShapeInfo);
|
||||
|
||||
sameOffsets = shape::haveSameShapeAndStrides(zShapeInfo, outerXTadShapeInfo);
|
||||
tadLen = shape::length(innerXTadShapeInfo);
|
||||
numTads = shape::length(outerXTadShapeInfo);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
sd::LongType coords[SD_MAX_RANK];
|
||||
|
||||
for (sd::LongType r = blockIdx.x; r < numTads; r += gridDim.x) {
|
||||
INDEX2COORDS(r, outerRank, outerShapePtr, coords);
|
||||
|
||||
sd::LongType outerOffset;
|
||||
COORDS2INDEX(outerRank, outerStridePtr, coords, outerOffset);
|
||||
|
||||
sd::LongType zOffset;
|
||||
if (sameOffsets) {
|
||||
zOffset = outerOffset;
|
||||
} else {
|
||||
INDEX2COORDS(r, zRank, zShapePtr, coords);
|
||||
COORDS2INDEX(zRank, zStridePtr, coords, zOffset);
|
||||
}
|
||||
|
||||
const X* xTad = x + outerOffset;
|
||||
sPartials[threadIdx.x] = OpType::startingValue(xTad);
|
||||
|
||||
for (sd::LongType i = threadIdx.x; i < tadLen; i += blockDim.x) {
|
||||
sd::LongType iCoords[SD_MAX_RANK];
|
||||
sd::LongType xOffset;
|
||||
INDEX2COORDS(i, innerRank, innerShapePtr, iCoords);
|
||||
COORDS2INDEX(innerRank, innerStridePtr, iCoords, xOffset);
|
||||
sPartials[threadIdx.x] = OpType::update(
|
||||
sPartials[threadIdx.x],
|
||||
OpType::op(xTad[xOffset], extraParams),
|
||||
extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
aggregatePartials<OpType>(
|
||||
sPartials,
|
||||
threadIdx.x,
|
||||
sd::math::sd_min<int>(blockDim.x, tadLen),
|
||||
extraParams);
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
z[zOffset] = OpType::postProcess(sPartials[0], tadLen, extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_DEVICE SD_INLINE void ReduceLongFunction<X, Z>::execScalarCuda(
|
||||
const void* vx,
|
||||
const sd::LongType* xShapeInfo,
|
||||
void* vextraParams,
|
||||
void* vz,
|
||||
const sd::LongType* zShapeInfo,
|
||||
void* vreductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
auto x = reinterpret_cast<const X*>(vx);
|
||||
auto z = reinterpret_cast<Z*>(vz);
|
||||
auto extraParams = reinterpret_cast<X*>(vextraParams);
|
||||
auto reductionBuffer = reinterpret_cast<Z*>(vreductionBuffer);
|
||||
|
||||
__shared__ Z sPartials[SD_CUDA_BLOCK_SIZE];
|
||||
__shared__ sd::LongType length;
|
||||
|
||||
// Cache shape info
|
||||
__shared__ sd::LongType xRank;
|
||||
__shared__ const sd::LongType* xShapePtr;
|
||||
__shared__ const sd::LongType* xStridePtr;
|
||||
|
||||
int tid = blockDim.x * blockIdx.x + threadIdx.x;
|
||||
if (threadIdx.x == 0) {
|
||||
length = shape::length(xShapeInfo);
|
||||
xRank = shape::rank(xShapeInfo);
|
||||
xShapePtr = shape::shapeOf(xShapeInfo);
|
||||
xStridePtr = shape::stride(xShapeInfo);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
sPartials[threadIdx.x] = OpType::startingValue(x);
|
||||
|
||||
sd::LongType gridSize = gridDim.x * blockDim.x;
|
||||
for (sd::LongType i = tid; i < length; i += gridSize) {
|
||||
sd::LongType xCoords[SD_MAX_RANK];
|
||||
sd::LongType xOffset;
|
||||
|
||||
INDEX2COORDS(i, xRank, xShapePtr, xCoords);
|
||||
COORDS2INDEX(xRank, xStridePtr, xCoords, xOffset);
|
||||
|
||||
sPartials[threadIdx.x] = OpType::update(
|
||||
sPartials[threadIdx.x],
|
||||
OpType::op(x[xOffset], extraParams),
|
||||
extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
aggregatePartials<OpType>(
|
||||
sPartials,
|
||||
threadIdx.x,
|
||||
sd::math::sd_min<int>(blockDim.x, length),
|
||||
extraParams);
|
||||
__syncthreads();
|
||||
|
||||
if (gridDim.x > 1) {
|
||||
auto tc = reinterpret_cast<unsigned int*>(reductionBuffer);
|
||||
__shared__ bool amLast;
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
reductionBuffer[blockIdx.x] = sPartials[0];
|
||||
}
|
||||
__threadfence();
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
unsigned int ticket = atomicInc(&tc[16384], gridDim.x);
|
||||
amLast = (ticket == gridDim.x - 1);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
if (amLast) {
|
||||
tc[16384] = 0;
|
||||
sPartials[threadIdx.x] = OpType::startingValue(x);
|
||||
|
||||
for (sd::LongType i = threadIdx.x; i < gridDim.x; i += blockDim.x) {
|
||||
sPartials[threadIdx.x] =
|
||||
OpType::update(sPartials[threadIdx.x], reductionBuffer[i], extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
aggregatePartials<OpType>(
|
||||
sPartials,
|
||||
threadIdx.x,
|
||||
sd::math::sd_min<int>(gridDim.x, blockDim.x),
|
||||
extraParams);
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
z[0] = OpType::postProcess(sPartials[0], length, extraParams);
|
||||
}
|
||||
}
|
||||
} else {
|
||||
if (threadIdx.x == 0) {
|
||||
auto tc = reinterpret_cast<unsigned int*>(reductionBuffer);
|
||||
tc[16384] = 0;
|
||||
z[0] = OpType::postProcess(sPartials[0], length, extraParams);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_HOST SD_INLINE void ReduceLongFunction<X, Z>::intermediate(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const void* x,
|
||||
sd::LongType* dXShapeInfo,
|
||||
sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* z,
|
||||
sd::LongType* dZShapeInfo,
|
||||
sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dims) {
|
||||
|
||||
if (shape::isEmptyConst(hXShapeInfo)) {
|
||||
// If input is empty, we skip unless z is also empty
|
||||
if (shape::isEmptyConst(hZShapeInfo)) return;
|
||||
|
||||
const auto startingVal = static_cast<Z>(
|
||||
OpType::startingValue(reinterpret_cast<const X*>(x)));
|
||||
auto res = cudaMemcpyAsync(
|
||||
sd::LaunchContext::defaultContext()->getScalarPointer(),
|
||||
&startingVal,
|
||||
sizeof(Z),
|
||||
cudaMemcpyHostToDevice,
|
||||
*stream);
|
||||
if (res != 0) {
|
||||
throw sd::cuda_exception::build(
|
||||
"ReduceLongFunction<X,Z>::intermediate: failed to copy temporary scalar", res);
|
||||
}
|
||||
|
||||
auto ptr = sd::LaunchContext::defaultContext()->getScalarPointer();
|
||||
// scalar assign
|
||||
scalar::ScalarTransform<Z,Z,Z>::executeCudaShaped(
|
||||
launchDims, stream, 14, z, dZShapeInfo, hXShapeInfo,
|
||||
z, dZShapeInfo, hZShapeInfo,
|
||||
ptr, nullptr);
|
||||
} else {
|
||||
sd::LongType zRank = shape::rank(hZShapeInfo);
|
||||
sd::LongType tadRank = shape::rank(hXShapeInfo) - zRank;
|
||||
|
||||
auto outerPack = sd::ConstantShapeHelper::getInstance()
|
||||
.createSubArrShapeInfo(hXShapeInfo, dims, zRank);
|
||||
auto innerPack = sd::ConstantShapeHelper::getInstance()
|
||||
.createSubArrShapeInfo(hXShapeInfo, dims + zRank, tadRank);
|
||||
|
||||
simpleReduce<X, Z, OpType>
|
||||
<<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(
|
||||
x,
|
||||
outerPack->special(),
|
||||
innerPack->special(),
|
||||
extraParams,
|
||||
vreductionBuffer,
|
||||
z,
|
||||
dZShapeInfo);
|
||||
}
|
||||
sd::DebugHelper::checkErrorCode(stream, "ReduceLongFunction intermediate(...) failed");
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Z>
|
||||
template <typename OpType>
|
||||
SD_HOST SD_INLINE void ReduceLongFunction<X, Z>::intermediateScalar(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
const sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo,
|
||||
const sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
sd::LongType dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
if (shape::isEmptyConst(hXShapeInfo)) {
|
||||
if (shape::isEmptyConst(hZShapeInfo)) return;
|
||||
|
||||
const auto startingVal =
|
||||
static_cast<Z>(OpType::startingValue(reinterpret_cast<const X*>(x)));
|
||||
auto res = cudaMemcpyAsync(z, &startingVal, sizeof(Z),
|
||||
cudaMemcpyHostToDevice, *stream);
|
||||
if (res != 0) {
|
||||
throw sd::cuda_exception::build(
|
||||
"ReduceLongFunction<X,Z>::intermediateScalar: failed to copy resulting scalar",
|
||||
res);
|
||||
}
|
||||
} else {
|
||||
simpleScalar<X, Z, OpType>
|
||||
<<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(
|
||||
x,
|
||||
xShapeInfo,
|
||||
extraParams,
|
||||
z,
|
||||
zShapeInfo,
|
||||
dimension,
|
||||
dimensionLength,
|
||||
reductionBuffer,
|
||||
tadOnlyShapeInfo);
|
||||
}
|
||||
sd::DebugHelper::checkErrorCode(stream, "ReduceLongFunction intermediateScalar(...) failed");
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Y>
|
||||
SD_HOST SD_INLINE void ReduceLongFunction<X, Y>::execReduceScalar(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const int opNum,
|
||||
const void* x,
|
||||
sd::LongType* xShapeInfo,
|
||||
sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
sd::LongType* zShapeInfo,
|
||||
sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
sd::LongType dimensionLength,
|
||||
void* reductionBuffer,
|
||||
sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
DISPATCH_BY_OPNUM_TT(
|
||||
intermediateScalar,
|
||||
PARAMS(launchDims, stream, x, xShapeInfo, hXShapeInfo,
|
||||
extraParams, z, zShapeInfo, hZShapeInfo,
|
||||
dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo),
|
||||
OPS_A(REDUCE_LONG_OPS));
|
||||
sd::DebugHelper::checkErrorCode(stream, "ReduceLongFunction execReduceScalar(...) failed");
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename Y>
|
||||
SD_HOST SD_INLINE void ReduceLongFunction<X, Y>::execReduce(dim3 launchDims, cudaStream_t *stream, int opNum, const void *vx,
|
||||
sd::LongType *dXShapeInfo, sd::LongType *hXShapeInfo, void *extraParams,
|
||||
void *vreductionBuffer, void *vz, sd::LongType *dZShapeInfo,
|
||||
sd::LongType *hZShapeInfo, sd::LongType *dims) {
|
||||
|
||||
if (shape::length(hZShapeInfo) == 1) {
|
||||
execReduceScalar(
|
||||
launchDims, stream, opNum,
|
||||
vx, dXShapeInfo, hXShapeInfo,
|
||||
extraParams, vz,
|
||||
dZShapeInfo, hZShapeInfo,
|
||||
nullptr, 0, vreductionBuffer, nullptr);
|
||||
} else {
|
||||
DISPATCH_BY_OPNUM_TT(
|
||||
intermediate,
|
||||
PARAMS(launchDims, stream, vx, dXShapeInfo, hXShapeInfo,
|
||||
extraParams, vreductionBuffer, vz, dZShapeInfo, hZShapeInfo, dims),
|
||||
OPS_A(REDUCE_LONG_OPS));
|
||||
}
|
||||
DEBUG_KERNEL(stream, opNum);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X>
|
||||
SD_DEVICE void initializeShared(X* extraParams, X** sPartials, int sMemSize) {
|
||||
int sPartialsLength = sMemSize / sizeof(X);
|
||||
X* sPartialsDeref = reinterpret_cast<X*>(*sPartials);
|
||||
for (int i = 0; i < sPartialsLength; i++) {
|
||||
sPartialsDeref[i] = extraParams[0];
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
ITERATE_COMBINATIONS(
|
||||
(SD_COMMON_TYPES),
|
||||
(SD_LONG_TYPES),
|
||||
INSTANT_PROCESS_COMBINATION,
|
||||
functions::reduce::ReduceLongFunction,
|
||||
::execReduce(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
int opNum,
|
||||
const void* vx,
|
||||
sd::LongType* dXShapeInfo,
|
||||
sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* vz,
|
||||
sd::LongType* dZShapeInfo,
|
||||
sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dims);
|
||||
);
|
||||
|
||||
ITERATE_COMBINATIONS(
|
||||
(SD_COMMON_TYPES),
|
||||
(SD_LONG_TYPES),
|
||||
INSTANT_PROCESS_COMBINATION,
|
||||
functions::reduce::ReduceLongFunction,
|
||||
::execReduceScalar(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const int opNum,
|
||||
const void* x,
|
||||
sd::LongType* xShapeInfo,
|
||||
sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
sd::LongType* zShapeInfo,
|
||||
sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
sd::LongType dimensionLength,
|
||||
void* reductionBuffer,
|
||||
sd::LongType* tadOnlyShapeInfo);
|
||||
);
|
||||
|
||||
} // namespace reduce
|
||||
} // namespace functions
|
||||
@@ -0,0 +1,553 @@
|
||||
/* ******************************************************************************
|
||||
*
|
||||
*
|
||||
* 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 raver119@gmail.com
|
||||
// @author Yurii Shyrma (iuriish@yahoo.com)
|
||||
//
|
||||
#include <exceptions/cuda_exception.h>
|
||||
#include <execution/LaunchContext.h>
|
||||
#include <helpers/DebugHelper.h>
|
||||
#include <loops/legacy_ops.h>
|
||||
#include <loops/reduce_same.h>
|
||||
#include <loops/scalar.h>
|
||||
#include <system/op_boilerplate.h>
|
||||
#include <types/types.h>
|
||||
|
||||
using namespace simdOps;
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename OpType>
|
||||
SD_KERNEL SD_INLINE void simpleReduce(
|
||||
const void* x,
|
||||
const sd::LongType* outerXTadShapeInfo,
|
||||
const sd::LongType* innerXTadShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo) {
|
||||
|
||||
functions::reduce::ReduceSameFunction<X>::template transformCuda<OpType>(
|
||||
x,
|
||||
outerXTadShapeInfo,
|
||||
innerXTadShapeInfo,
|
||||
extraParams,
|
||||
vreductionBuffer,
|
||||
z,
|
||||
zShapeInfo);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X, typename OpType>
|
||||
SD_KERNEL SD_INLINE void simpleScalar(
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
long long int dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
functions::reduce::ReduceSameFunction<X>::template execScalarCuda<OpType>(
|
||||
x, xShapeInfo, extraParams, z, zShapeInfo, reductionBuffer, tadOnlyShapeInfo);
|
||||
}
|
||||
|
||||
namespace functions {
|
||||
namespace reduce {
|
||||
|
||||
template <typename X>
|
||||
template <typename OpType>
|
||||
SD_DEVICE SD_INLINE void ReduceSameFunction<X>::aggregatePartials(
|
||||
void* vsPartials,
|
||||
sd::LongType tid,
|
||||
sd::LongType numItems,
|
||||
void* vextraParams) {
|
||||
|
||||
auto sPartials = reinterpret_cast<X*>(vsPartials);
|
||||
auto extraParams = reinterpret_cast<X*>(vextraParams);
|
||||
|
||||
sd::LongType floorPow2 = numItems;
|
||||
if (floorPow2 & (floorPow2 - 1)) {
|
||||
while (floorPow2 & (floorPow2 - 1)) {
|
||||
floorPow2 &= floorPow2 - 1;
|
||||
}
|
||||
if (tid >= floorPow2) {
|
||||
sPartials[tid - floorPow2] =
|
||||
OpType::update(sPartials[tid - floorPow2], sPartials[tid], extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
|
||||
for (sd::LongType activeThreads = (floorPow2 >> 1); activeThreads; activeThreads >>= 1) {
|
||||
if (tid < activeThreads && (tid + activeThreads) < numItems) {
|
||||
sPartials[tid] =
|
||||
OpType::update(sPartials[tid], sPartials[tid + activeThreads], extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X>
|
||||
template <typename OpType>
|
||||
SD_DEVICE SD_INLINE void ReduceSameFunction<X>::transformCuda(
|
||||
const void* vx,
|
||||
const sd::LongType* outerXTadShapeInfo,
|
||||
const sd::LongType* innerXTadShapeInfo,
|
||||
void* vextraParams,
|
||||
void* vreductionBuffer,
|
||||
void* vz,
|
||||
const sd::LongType* zShapeInfo) {
|
||||
|
||||
auto x = reinterpret_cast<const X*>(vx);
|
||||
auto z = reinterpret_cast<X*>(vz);
|
||||
auto extraParams = reinterpret_cast<X*>(vextraParams);
|
||||
|
||||
__shared__ X sPartials[SD_CUDA_BLOCK_SIZE];
|
||||
__shared__ int tadLen;
|
||||
__shared__ int numTads;
|
||||
__shared__ bool sameOffsets;
|
||||
|
||||
// Cache shape info for outer/inner and z if needed
|
||||
__shared__ sd::LongType outerRank;
|
||||
__shared__ sd::LongType innerRank;
|
||||
__shared__ sd::LongType zRank;
|
||||
|
||||
__shared__ const sd::LongType* outerShapePtr;
|
||||
__shared__ const sd::LongType* outerStridePtr;
|
||||
|
||||
__shared__ const sd::LongType* innerShapePtr;
|
||||
__shared__ const sd::LongType* innerStridePtr;
|
||||
|
||||
__shared__ const sd::LongType* zShapePtr;
|
||||
__shared__ const sd::LongType* zStridePtr;
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
outerRank = shape::rank(outerXTadShapeInfo);
|
||||
outerShapePtr = shape::shapeOf(outerXTadShapeInfo);
|
||||
outerStridePtr = shape::stride(outerXTadShapeInfo);
|
||||
|
||||
innerRank = shape::rank(innerXTadShapeInfo);
|
||||
innerShapePtr = shape::shapeOf(innerXTadShapeInfo);
|
||||
innerStridePtr = shape::stride(innerXTadShapeInfo);
|
||||
|
||||
zRank = shape::rank(zShapeInfo);
|
||||
zShapePtr = shape::shapeOf(zShapeInfo);
|
||||
zStridePtr = shape::stride(zShapeInfo);
|
||||
|
||||
sameOffsets = shape::haveSameShapeAndStrides(zShapeInfo, outerXTadShapeInfo);
|
||||
tadLen = shape::length(innerXTadShapeInfo);
|
||||
numTads = shape::length(outerXTadShapeInfo);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
sd::LongType coords[SD_MAX_RANK];
|
||||
|
||||
for (sd::LongType r = blockIdx.x; r < numTads; r += gridDim.x) {
|
||||
INDEX2COORDS(r, outerRank, outerShapePtr, coords);
|
||||
|
||||
sd::LongType outerOffset;
|
||||
COORDS2INDEX(outerRank, outerStridePtr, coords, outerOffset);
|
||||
|
||||
sd::LongType zOffset;
|
||||
if (sameOffsets) {
|
||||
zOffset = outerOffset;
|
||||
} else {
|
||||
INDEX2COORDS(r, zRank, zShapePtr, coords);
|
||||
COORDS2INDEX(zRank, zStridePtr, coords, zOffset);
|
||||
}
|
||||
|
||||
const X* xTad = x + outerOffset;
|
||||
sPartials[threadIdx.x] = OpType::startingValue(xTad);
|
||||
|
||||
for (sd::LongType i = threadIdx.x; i < tadLen; i += blockDim.x) {
|
||||
sd::LongType iCoords[SD_MAX_RANK];
|
||||
sd::LongType xOffset;
|
||||
|
||||
INDEX2COORDS(i, innerRank, innerShapePtr, iCoords);
|
||||
COORDS2INDEX(innerRank, innerStridePtr, iCoords, xOffset);
|
||||
|
||||
sPartials[threadIdx.x] = OpType::update(
|
||||
sPartials[threadIdx.x],
|
||||
OpType::op(xTad[xOffset], extraParams),
|
||||
extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
aggregatePartials<OpType>(
|
||||
sPartials,
|
||||
threadIdx.x,
|
||||
sd::math::sd_min<int>(blockDim.x, tadLen),
|
||||
extraParams);
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
z[zOffset] = OpType::postProcess(sPartials[0], tadLen, extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X>
|
||||
SD_DEVICE SD_INLINE void ReduceSameFunction<X>::execScalarCudaLegacy(
|
||||
int opNum,
|
||||
const void* vx,
|
||||
const sd::LongType* xShapeInfo,
|
||||
void* vextraParams,
|
||||
void* vz,
|
||||
const sd::LongType* zShapeInfo,
|
||||
void* vreductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
DISPATCH_BY_OPNUM_T(
|
||||
execScalarCuda,
|
||||
PARAMS(vx, xShapeInfo, vextraParams, vz, zShapeInfo, vreductionBuffer, tadOnlyShapeInfo),
|
||||
REDUCE_SAME_OPS);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X>
|
||||
template <typename OpType>
|
||||
SD_DEVICE SD_INLINE void ReduceSameFunction<X>::execScalarCuda(
|
||||
const void* vx,
|
||||
const sd::LongType* xShapeInfo,
|
||||
void* vextraParams,
|
||||
void* vz,
|
||||
const sd::LongType* zShapeInfo,
|
||||
void* vreductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
auto x = reinterpret_cast<const X*>(vx);
|
||||
auto z = reinterpret_cast<X*>(vz);
|
||||
auto extraParams = reinterpret_cast<X*>(vextraParams);
|
||||
|
||||
__shared__ X sPartials[SD_CUDA_BLOCK_SIZE];
|
||||
__shared__ sd::LongType length;
|
||||
auto reductionBuffer = reinterpret_cast<X *>(vreductionBuffer);
|
||||
|
||||
// Cache shape info
|
||||
__shared__ sd::LongType xRank;
|
||||
__shared__ const sd::LongType* xShapePtr;
|
||||
__shared__ const sd::LongType* xStridePtr;
|
||||
|
||||
int tid = blockDim.x * blockIdx.x + threadIdx.x;
|
||||
if (threadIdx.x == 0) {
|
||||
length = shape::length(xShapeInfo);
|
||||
xRank = shape::rank(xShapeInfo);
|
||||
xShapePtr = shape::shapeOf(xShapeInfo);
|
||||
xStridePtr = shape::stride(xShapeInfo);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
sPartials[threadIdx.x] = OpType::startingValue(x);
|
||||
|
||||
sd::LongType gridSize = gridDim.x * blockDim.x;
|
||||
for (sd::LongType i = tid; i < length; i += gridSize) {
|
||||
sd::LongType xCoords[SD_MAX_RANK];
|
||||
sd::LongType xOffset;
|
||||
|
||||
INDEX2COORDS(i, xRank, xShapePtr, xCoords);
|
||||
COORDS2INDEX(xRank, xStridePtr, xCoords, xOffset);
|
||||
|
||||
sPartials[threadIdx.x] = OpType::update(
|
||||
sPartials[threadIdx.x],
|
||||
OpType::op(x[xOffset], extraParams),
|
||||
extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
aggregatePartials<OpType>(
|
||||
sPartials,
|
||||
threadIdx.x,
|
||||
sd::math::sd_min<int>(blockDim.x, length),
|
||||
extraParams);
|
||||
__syncthreads();
|
||||
|
||||
if (gridDim.x > 1) {
|
||||
auto tc = reinterpret_cast<unsigned int*>(vreductionBuffer);
|
||||
__shared__ bool amLast;
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
reductionBuffer[blockIdx.x] = sPartials[0];
|
||||
}
|
||||
__threadfence();
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
unsigned int ticket = atomicInc(&tc[16384], gridDim.x);
|
||||
amLast = (ticket == (gridDim.x - 1));
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
if (amLast) {
|
||||
tc[16384] = 0;
|
||||
sPartials[threadIdx.x] = OpType::startingValue(x);
|
||||
|
||||
for (sd::LongType i = threadIdx.x; i < gridDim.x; i += blockDim.x) {
|
||||
sPartials[threadIdx.x] =
|
||||
OpType::update(sPartials[threadIdx.x], reinterpret_cast<X*>(vreductionBuffer)[i], extraParams);
|
||||
}
|
||||
__syncthreads();
|
||||
|
||||
aggregatePartials<OpType>(
|
||||
sPartials,
|
||||
threadIdx.x,
|
||||
sd::math::sd_min<int>(gridDim.x, blockDim.x),
|
||||
extraParams);
|
||||
__syncthreads();
|
||||
|
||||
if (threadIdx.x == 0) {
|
||||
z[0] = OpType::postProcess(sPartials[0], length, extraParams);
|
||||
}
|
||||
}
|
||||
} else {
|
||||
if (threadIdx.x == 0) {
|
||||
auto tc = reinterpret_cast<unsigned int*>(vreductionBuffer);
|
||||
tc[16384] = 0;
|
||||
z[0] = OpType::postProcess(sPartials[0], length, extraParams);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X>
|
||||
template <typename OpType>
|
||||
SD_HOST SD_INLINE void ReduceSameFunction<X>::intermediate(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const void* x,
|
||||
const sd::LongType* dXShapeInfo,
|
||||
const sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* z,
|
||||
const sd::LongType* dZShapeInfo,
|
||||
const sd::LongType* hZShapeInfo,
|
||||
const sd::LongType* dims) {
|
||||
|
||||
if (shape::isEmptyConst(hXShapeInfo)) {
|
||||
// If input is empty, skip unless z is also empty
|
||||
if (shape::isEmptyConst(hZShapeInfo)) return;
|
||||
|
||||
const auto startingVal =
|
||||
static_cast<X>(OpType::startingValue(reinterpret_cast<const X*>(x)));
|
||||
auto res = cudaMemcpyAsync(
|
||||
sd::LaunchContext::defaultContext()->getScalarPointer(),
|
||||
&startingVal,
|
||||
sizeof(X),
|
||||
cudaMemcpyHostToDevice,
|
||||
*stream);
|
||||
if (res != 0) {
|
||||
throw sd::cuda_exception::build(
|
||||
"ReduceSameFunction<X>::intermediate: failed to copy temporary scalar", res);
|
||||
}
|
||||
|
||||
auto ptr = sd::LaunchContext::defaultContext()->getScalarPointer();
|
||||
// scalar assign
|
||||
scalar::ScalarTransform<X,X,X>::executeCudaShaped(
|
||||
launchDims,
|
||||
stream,
|
||||
14,
|
||||
z,
|
||||
dZShapeInfo,
|
||||
hXShapeInfo,
|
||||
z,
|
||||
dZShapeInfo,
|
||||
hZShapeInfo,
|
||||
ptr,
|
||||
nullptr);
|
||||
} else {
|
||||
const sd::LongType zRank = shape::rank(hZShapeInfo);
|
||||
const sd::LongType tadRank = shape::rank(hXShapeInfo) - zRank;
|
||||
|
||||
auto outerPack = sd::ConstantShapeHelper::getInstance()
|
||||
.createSubArrShapeInfo(
|
||||
const_cast<sd::LongType*>(hXShapeInfo), const_cast<sd::LongType*>(dims), zRank);
|
||||
auto innerPack = sd::ConstantShapeHelper::getInstance()
|
||||
.createSubArrShapeInfo(
|
||||
const_cast<sd::LongType*>(hXShapeInfo), const_cast<sd::LongType*>(dims + zRank), tadRank);
|
||||
|
||||
simpleReduce<X, OpType>
|
||||
<<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(
|
||||
x,
|
||||
outerPack->special(),
|
||||
innerPack->special(),
|
||||
extraParams,
|
||||
vreductionBuffer,
|
||||
z,
|
||||
dZShapeInfo);
|
||||
sd::DebugHelper::checkErrorCode(stream, "ReduceSameFunction intermediate(...) failed");
|
||||
}
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X>
|
||||
template <typename OpType>
|
||||
SD_HOST SD_INLINE void ReduceSameFunction<X>::intermediateScalar(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
const sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo,
|
||||
const sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
sd::LongType dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
if (shape::isEmptyConst(hXShapeInfo)) {
|
||||
if (shape::isEmptyConst(hZShapeInfo)) return;
|
||||
|
||||
const auto startingVal =
|
||||
static_cast<X>(OpType::startingValue(reinterpret_cast<const X*>(x)));
|
||||
auto res = cudaMemcpyAsync(z, &startingVal, sizeof(X),
|
||||
cudaMemcpyHostToDevice, *stream);
|
||||
if (res != 0) {
|
||||
throw sd::cuda_exception::build(
|
||||
"ReduceSameFunction<X>::intermediateScalar: failed to copy resulting scalar",
|
||||
res);
|
||||
}
|
||||
} else {
|
||||
simpleScalar<X, OpType>
|
||||
<<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(
|
||||
x,
|
||||
xShapeInfo,
|
||||
extraParams,
|
||||
z,
|
||||
zShapeInfo,
|
||||
dimension,
|
||||
dimensionLength,
|
||||
reductionBuffer,
|
||||
tadOnlyShapeInfo);
|
||||
}
|
||||
sd::DebugHelper::checkErrorCode(stream, "ReduceSameFunction intermediateScalar(...) failed");
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X>
|
||||
SD_HOST SD_INLINE void ReduceSameFunction<X>::execReduceScalar(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
int opNum,
|
||||
const void* x,
|
||||
const sd::LongType* xShapeInfo,
|
||||
const sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* z,
|
||||
const sd::LongType* zShapeInfo,
|
||||
const sd::LongType* hZShapeInfo,
|
||||
sd::LongType* dimension,
|
||||
sd::LongType dimensionLength,
|
||||
void* reductionBuffer,
|
||||
const sd::LongType* tadOnlyShapeInfo) {
|
||||
|
||||
DISPATCH_BY_OPNUM_T(
|
||||
intermediateScalar,
|
||||
PARAMS(
|
||||
launchDims, stream, x, xShapeInfo, hXShapeInfo, extraParams, z,
|
||||
zShapeInfo, hZShapeInfo, dimension, dimensionLength, reductionBuffer, tadOnlyShapeInfo),
|
||||
REDUCE_SAME_OPS);
|
||||
sd::DebugHelper::checkErrorCode(stream, "execReduceScalarSame(...) failed");
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X>
|
||||
SD_HOST SD_INLINE void ReduceSameFunction<X>::execReduce(
|
||||
dim3 launchDims,
|
||||
cudaStream_t* stream,
|
||||
const int opNum,
|
||||
const void* x,
|
||||
const sd::LongType* dXShapeInfo,
|
||||
const sd::LongType* hXShapeInfo,
|
||||
void* extraParams,
|
||||
void* vreductionBuffer,
|
||||
void* z,
|
||||
const sd::LongType* dZShapeInfo,
|
||||
const sd::LongType* hZShapeInfo,
|
||||
const sd::LongType* dims) {
|
||||
|
||||
if (shape::length(hZShapeInfo) == 1) {
|
||||
execReduceScalar(
|
||||
launchDims, stream, opNum,
|
||||
x, dXShapeInfo, hXShapeInfo,
|
||||
extraParams, z,
|
||||
dZShapeInfo, hZShapeInfo,
|
||||
nullptr, 0, vreductionBuffer, nullptr);
|
||||
} else {
|
||||
DISPATCH_BY_OPNUM_T(
|
||||
intermediate,
|
||||
PARAMS(
|
||||
launchDims, stream, x, dXShapeInfo, hXShapeInfo, extraParams, vreductionBuffer, z, dZShapeInfo,
|
||||
hZShapeInfo, dims),
|
||||
REDUCE_SAME_OPS);
|
||||
}
|
||||
DEBUG_KERNEL(stream, opNum);
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////
|
||||
template <typename X>
|
||||
SD_DEVICE void initializeShared(X* extraParams, X** sPartials, int sMemSize) {
|
||||
int sPartialsLength = sMemSize / sizeof(X);
|
||||
X* sPartialsDeref = reinterpret_cast<X*>(*sPartials);
|
||||
for (int i = 0; i < sPartialsLength; i++) {
|
||||
sPartialsDeref[i] = extraParams[0];
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
#define INSTANT_PROCESS_SINGLE(a1) \
|
||||
template void functions::reduce::ReduceSameFunction<GET_SECOND(a1)>::execReduce( \
|
||||
dim3 launchDims, \
|
||||
cudaStream_t* stream, \
|
||||
const int opNum, \
|
||||
const void* x, \
|
||||
const sd::LongType* dXShapeInfo, \
|
||||
const sd::LongType* hXShapeInfo, \
|
||||
void* extraParams, \
|
||||
void* vreductionBuffer, \
|
||||
void* z, \
|
||||
const sd::LongType* dZShapeInfo, \
|
||||
const sd::LongType* hZShapeInfo, \
|
||||
const sd::LongType* dims); \
|
||||
\
|
||||
template void functions::reduce::ReduceSameFunction<GET_SECOND(a1)>::execReduceScalar( \
|
||||
dim3 launchDims, \
|
||||
cudaStream_t* stream, \
|
||||
int opNum, \
|
||||
const void* x, \
|
||||
const sd::LongType* xShapeInfo, \
|
||||
const sd::LongType* hXShapeInfo, \
|
||||
void* extraParams, \
|
||||
void* z, \
|
||||
const sd::LongType* zShapeInfo, \
|
||||
const sd::LongType* hZShapeInfo, \
|
||||
sd::LongType* dimension, \
|
||||
sd::LongType dimensionLength, \
|
||||
void* reductionBuffer, \
|
||||
const sd::LongType* tadOnlyShapeInfo);
|
||||
|
||||
ITERATE_LIST((SD_COMMON_TYPES), INSTANT_PROCESS_SINGLE)
|
||||
|
||||
} // namespace reduce
|
||||
} // namespace functions
|
||||
Reference in New Issue
Block a user