cavis/libnd4j/include/ops/declarable/helpers/cuda/hamming.cu

95 lines
4.2 KiB
Plaintext
Raw Normal View History

/*******************************************************************************
* 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
//
#include <ops/declarable/helpers/helpers.h>
#include <ops/declarable/helpers/hamming.h>
namespace sd {
namespace ops {
namespace helpers {
template <typename X, typename Z>
static _CUDA_G void _hammingKernel(void *vx, Nd4jLong *xShapeInfo, void *vy, Nd4jLong *yShapeInfo, void *vz, void *reductionBuffer, Nd4jLong length) {
auto x = reinterpret_cast<X*>(vx);
auto y = reinterpret_cast<X*>(vy);
auto z = reinterpret_cast<Z*>(vz);
__shared__ Nd4jLong *shared;
if (threadIdx.x == 0) {
extern __shared__ unsigned char shmem[];
shared = reinterpret_cast<Nd4jLong*>(shmem);
}
__syncthreads();
// we want to nullify temporary memory before accumulating intermediate results
shared[threadIdx.x] = 0;
auto tid = threadIdx.x + blockIdx.x * blockDim.x;
for (Nd4jLong e = tid; e < length; e += blockDim.x * gridDim.x) {
auto _x = static_cast<unsigned long long>(x[shape::getIndexOffset(e, xShapeInfo)]);
auto _y = static_cast<unsigned long long>(y[shape::getIndexOffset(e, yShapeInfo)]);
// we save intermediate result into shared memory
shared[threadIdx.x] += __popcll(_x ^ _y);
}
__syncthreads();
// now we accumulate values
auto numItems = sd::math::nd4j_min<Nd4jLong>(blockDim.x, length);
auto floorPow2 = numItems;
if (floorPow2 & (floorPow2 - 1)) {
while (floorPow2 & (floorPow2 - 1))
floorPow2 &= floorPow2 - 1;
if (threadIdx.x >= floorPow2)
shared[threadIdx.x - floorPow2] = shared[threadIdx.x - floorPow2] + shared[threadIdx.x];
__syncthreads();
}
__syncthreads();
for (Nd4jLong activeThreads = floorPow2 >> 1; activeThreads; activeThreads >>= 1) {
if (threadIdx.x < activeThreads && threadIdx.x + activeThreads < numItems)
shared[threadIdx.x] = shared[threadIdx.x] + shared[threadIdx.x + activeThreads];
__syncthreads();
}
__syncthreads();
// FIXME: do we really want atomicAdd on global memory here
// and store them to output
if (threadIdx.x == 0 && shared[0] > 0)
sd::math::atomics::nd4j_atomicAdd<Z>(&z[0], static_cast<Z>(shared[threadIdx.x]));
}
template <typename X, typename Z>
static void _hamming(LaunchContext *context, NDArray &x, NDArray &y, NDArray &z) {
_hammingKernel<X, Z><<<256, 256, 256 * sizeof(Nd4jLong) + 256, *context->getCudaStream()>>>(x.specialBuffer(), x.specialShapeInfo(), y.specialBuffer(), y.specialShapeInfo(), z.specialBuffer(), nullptr, x.lengthOf());
}
void hamming(LaunchContext *context, NDArray &x, NDArray &y, NDArray &output) {
NDArray::prepareSpecialUse({&output}, {&x, &y});
BUILD_DOUBLE_SELECTOR(x.dataType(), output.dataType(), _hamming, (context, x, y, output), INTEGER_TYPES, INDEXING_TYPES);
NDArray::registerSpecialUse({&output}, {&x, &y});
}
}
}
}