314 lines
15 KiB
Plaintext
314 lines
15 KiB
Plaintext
/*******************************************************************************
|
|
* 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 <system/op_boilerplate.h>
|
|
#include <loops/broadcasting.h>
|
|
#include <loops/legacy_ops.h>
|
|
#include <types/types.h>
|
|
#include <system/Environment.h>
|
|
#include <cuda.h>
|
|
#include <cuda_runtime.h>
|
|
#include <string>
|
|
#include <stdexcept>
|
|
#include <helpers/StringUtils.h>
|
|
#include <ops/specials_cuda.h>
|
|
|
|
using namespace simdOps;
|
|
|
|
template<typename X, typename Y, typename Z, typename OpClass>
|
|
static __global__ void broadcastSimple(
|
|
void const* x,
|
|
Nd4jLong const* xShapeInfo,
|
|
void const* y,
|
|
Nd4jLong const* yShapeInfo,
|
|
void *z,
|
|
Nd4jLong const* zShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength, Nd4jLong const* tadOnlyShapeInfo, Nd4jLong const* tadOffsets, Nd4jLong const* tadOnlyShapeInfoZ, Nd4jLong const* tadOffsetsZ) {
|
|
|
|
functions::broadcast::Broadcast<X,Y,Z>::template transformCuda<OpClass>(x,xShapeInfo,y,yShapeInfo,z,zShapeInfo,dimension,dimensionLength,tadOnlyShapeInfo,tadOffsets,tadOnlyShapeInfoZ,tadOffsetsZ);
|
|
}
|
|
|
|
template<typename X, typename Y, typename Z, typename OpClass>
|
|
static __global__ void broadcastSimple(const void const* x, const Nd4jLong const* xShapeInfo,
|
|
const void const* y, const Nd4jLong const* yShapeInfo,
|
|
void *z, const Nd4jLong const* zShapeInfo ) {
|
|
|
|
functions::broadcast::Broadcast<X,Y,Z>::template transformCuda<OpClass>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo);
|
|
}
|
|
|
|
|
|
template<typename X, typename Y, typename Z, typename OpClass>
|
|
static __global__ void broadcastInverseSimple(
|
|
void const* x,
|
|
Nd4jLong const* xShapeInfo,
|
|
void const* y,
|
|
Nd4jLong const* yShapeInfo,
|
|
void *z,
|
|
Nd4jLong const* zShapeInfo,
|
|
int *dimension,
|
|
int dimensionLength, Nd4jLong const* tadOnlyShapeInfo, Nd4jLong const* tadOffsets, Nd4jLong const* tadOnlyShapeInfoZ, Nd4jLong const* tadOffsetsZ) {
|
|
|
|
functions::broadcast::Broadcast<X,Y,Z>::template transformInverseCuda<OpClass>(x,xShapeInfo,y,yShapeInfo,z,zShapeInfo,dimension,dimensionLength,tadOnlyShapeInfo,tadOffsets,tadOnlyShapeInfoZ,tadOffsetsZ);
|
|
}
|
|
|
|
|
|
namespace functions {
|
|
namespace broadcast {
|
|
|
|
static Nd4jLong __device__ __noinline__ getIndexOffset(Nd4jLong index, const Nd4jLong *shapeInfo) {
|
|
return shape::getIndexOffset(index, shapeInfo);
|
|
}
|
|
|
|
static Nd4jLong __device__ __noinline__ length(const Nd4jLong *shapeInfo) {
|
|
return shape::length(shapeInfo);
|
|
}
|
|
|
|
template<typename X, typename Y, typename Z>
|
|
template <typename OpClass>
|
|
__host__ void Broadcast<X,Y,Z>::intermediateBroadcast(dim3 launchDims, cudaStream_t *stream, void const* x, Nd4jLong const* xShapeInfo, void const* y, Nd4jLong const* yShapeInfo, void* z, Nd4jLong const* zShapeInfo, int *dimension, int dimensionLength, Nd4jLong const* tadOnlyShapeInfo, Nd4jLong const* tadOffsets, Nd4jLong const* tadOnlyShapeInfoZ, Nd4jLong const* tadOffsetsZ) {
|
|
broadcastSimple<X, Y, Z, OpClass><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ);
|
|
}
|
|
|
|
template<typename X, typename Y, typename Z>
|
|
template <typename OpClass>
|
|
__host__ void Broadcast<X,Y,Z>::intermediateBroadcast(dim3 launchDims, cudaStream_t *stream, const void *x, const Nd4jLong *xShapeInfo, const void *y, const Nd4jLong *yShapeInfo, void *z, const Nd4jLong *zShapeInfo) {
|
|
broadcastSimple<X, Y, Z, OpClass><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo);
|
|
}
|
|
|
|
template<typename X, typename Y, typename Z>
|
|
__host__ void Broadcast<X,Y,Z>::execBroadcast(dim3 launchDims, cudaStream_t *stream, int opNum, void const* x, Nd4jLong const* xShapeInfo, void const* y, Nd4jLong const* yShapeInfo, void *z, Nd4jLong const* zShapeInfo, int *dimension, int dimensionLength, Nd4jLong const* tadOnlyShapeInfo, Nd4jLong const* tadOffsets, Nd4jLong const* tadOnlyShapeInfoZ, Nd4jLong const* tadOffsetsZ) {
|
|
DISPATCH_BY_OPNUM_TTT(intermediateBroadcast, PARAMS(launchDims, stream, x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ), OPS_A(BROADCAST_OPS))
|
|
|
|
DEBUG_KERNEL(stream, opNum);
|
|
}
|
|
|
|
template<typename X, typename Y, typename Z>
|
|
__host__ void Broadcast<X,Y,Z>::execBroadcast(dim3 launchDims, cudaStream_t *stream, const int opNum, const void *x, const Nd4jLong *xShapeInfo, const void *y, const Nd4jLong *yShapeInfo, void *z, const Nd4jLong const* zShapeInfo) {
|
|
DISPATCH_BY_OPNUM_TTT(intermediateBroadcast, PARAMS(launchDims, stream, x, xShapeInfo, y, yShapeInfo, z, zShapeInfo), OPS_A(BROADCAST_OPS))
|
|
|
|
DEBUG_KERNEL(stream, opNum);
|
|
}
|
|
|
|
template<typename X, typename Y, typename Z>
|
|
template <typename OpClass>
|
|
__host__ void Broadcast<X,Y,Z>::intermediateInverseBroadcast(dim3 launchDims, cudaStream_t *stream, void const* x, Nd4jLong const* xShapeInfo, void const* y, Nd4jLong const* yShapeInfo, void *z, Nd4jLong const* zShapeInfo, int *dimension, int dimensionLength, Nd4jLong const* tadOnlyShapeInfo, Nd4jLong const* tadOffsets, Nd4jLong const* tadOnlyShapeInfoZ, Nd4jLong const* tadOffsetsZ) {
|
|
broadcastInverseSimple<X, Y, Z, OpClass><<<launchDims.x, launchDims.y, launchDims.z, *stream>>>(x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ);
|
|
}
|
|
|
|
template<typename X, typename Y, typename Z>
|
|
__host__ void Broadcast<X,Y,Z>::execInverseBroadcast(dim3 launchDims, cudaStream_t *stream, int opNum, void const* x, Nd4jLong const* xShapeInfo, void const* y, Nd4jLong const* yShapeInfo, void *z, Nd4jLong const* zShapeInfo, int *dimension, int dimensionLength, Nd4jLong const* tadOnlyShapeInfo, Nd4jLong const* tadOffsets, Nd4jLong const* tadOnlyShapeInfoZ, Nd4jLong const* tadOffsetsZ) {
|
|
DISPATCH_BY_OPNUM_TTT(intermediateInverseBroadcast, PARAMS(launchDims, stream, x, xShapeInfo, y, yShapeInfo, z, zShapeInfo, dimension, dimensionLength, tadOnlyShapeInfo, tadOffsets, tadOnlyShapeInfoZ, tadOffsetsZ), OPS_A(BROADCAST_OPS))
|
|
|
|
DEBUG_KERNEL(stream, opNum);
|
|
}
|
|
|
|
template<typename X, typename Y, typename Z>
|
|
template <typename OpType>
|
|
__device__ void Broadcast<X,Y,Z>::transformInverseCuda(
|
|
void const* vx, Nd4jLong const* xShapeInfo,
|
|
void const* vy, Nd4jLong const* yShapeInfo,
|
|
void* vz, Nd4jLong const* zShapeInfo,
|
|
int *dimension, int dimensionLength,
|
|
Nd4jLong const* tadOnlyShapeInfo, Nd4jLong const* tadOffsets, Nd4jLong const* tadOnlyShapeInfoZ, Nd4jLong const* tadOffsetsZ) {
|
|
|
|
if (tadOnlyShapeInfoZ == nullptr) {
|
|
tadOnlyShapeInfoZ = tadOnlyShapeInfo;
|
|
tadOffsetsZ = tadOffsets;
|
|
}
|
|
|
|
auto x = reinterpret_cast<X const*>(vx);
|
|
auto y = reinterpret_cast<Y const*>(vy);
|
|
auto z = reinterpret_cast<Z*>(vz);
|
|
|
|
//decompose in to several sub tads after
|
|
//moving all dimensions (in sorted order)
|
|
//to the back.
|
|
//permuted version of the x shape info for setting up the tad problem
|
|
__shared__ Nd4jLong tadLength;
|
|
__shared__ Nd4jLong tadEWS;
|
|
__shared__ int numTads;
|
|
__shared__ Nd4jLong xEWS;
|
|
__shared__ Nd4jLong zEWS;
|
|
|
|
|
|
if (threadIdx.x == 0) {
|
|
tadLength = length(tadOnlyShapeInfo);
|
|
tadEWS = shape::elementWiseStride(tadOnlyShapeInfo);
|
|
numTads = length(yShapeInfo) / tadLength;
|
|
xEWS = shape::elementWiseStride(xShapeInfo);
|
|
zEWS = shape::elementWiseStride(tadOnlyShapeInfoZ);
|
|
}
|
|
__syncthreads();
|
|
|
|
auto xOrder = shape::order(xShapeInfo);
|
|
auto yOrder = shape::order(tadOnlyShapeInfo);
|
|
auto zOrder = shape::order(tadOnlyShapeInfoZ);
|
|
|
|
for (int r = blockIdx.x; r < numTads; r += gridDim.x) {
|
|
|
|
|
|
auto rY = y + tadOffsets[r];
|
|
auto rZ = z + tadOffsetsZ[r];
|
|
|
|
|
|
if(tadEWS > 0 && zEWS > 0 && xEWS > 0 && dimensionLength == 1 && xOrder == yOrder && xOrder == zOrder) {
|
|
for (int i = threadIdx.x; i < tadLength; i+= blockDim.x)
|
|
rZ[i * zEWS] = OpType::op(x[i * xEWS], rY[i * tadEWS]);
|
|
}
|
|
else {
|
|
// it is expected that x and z tads and y array all have the same length
|
|
for (Nd4jLong i = threadIdx.x; i < tadLength; i+= blockDim.x) {
|
|
auto xOffset = getIndexOffset(i, xShapeInfo);
|
|
auto yOffset = getIndexOffset(i, tadOnlyShapeInfo);
|
|
auto zOffset = getIndexOffset(i, tadOnlyShapeInfoZ);
|
|
rZ[zOffset] = OpType::op(x[xOffset], rY[yOffset]);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
template<typename X, typename Y, typename Z>
|
|
template <typename OpType>
|
|
__device__ void Broadcast<X,Y,Z>::transformCuda(
|
|
void const* vx, Nd4jLong const* xShapeInfo,
|
|
void const* vy, Nd4jLong const* yShapeInfo,
|
|
void *vz, Nd4jLong const* zShapeInfo,
|
|
int *dimension, int dimensionLength,
|
|
Nd4jLong const* tadOnlyShapeInfo, Nd4jLong const* tadOffsets, Nd4jLong const* tadOnlyShapeInfoZ, Nd4jLong const* tadOffsetsZ) {
|
|
|
|
if (tadOnlyShapeInfoZ == nullptr) {
|
|
tadOnlyShapeInfoZ = tadOnlyShapeInfo;
|
|
tadOffsetsZ = tadOffsets;
|
|
}
|
|
|
|
auto x = reinterpret_cast<X const*>(vx);
|
|
auto y = reinterpret_cast<Y const*>(vy);
|
|
auto z = reinterpret_cast<Z*>(vz);
|
|
|
|
//decompose in to several sub tads after
|
|
//moving all dimensions (in sorted order)
|
|
//to the back.
|
|
//permuted version of the x shape info for setting up the tad problem
|
|
__shared__ Nd4jLong tadLength;
|
|
__shared__ Nd4jLong tadEWS;
|
|
__shared__ int numTads;
|
|
__shared__ Nd4jLong yEWS;
|
|
__shared__ Nd4jLong zEWS;
|
|
|
|
if (threadIdx.x == 0) {
|
|
tadLength = length(tadOnlyShapeInfo);
|
|
tadEWS = shape::elementWiseStride(tadOnlyShapeInfo);
|
|
numTads = length(xShapeInfo) / tadLength;
|
|
yEWS = shape::elementWiseStride(yShapeInfo);
|
|
zEWS = shape::elementWiseStride(tadOnlyShapeInfoZ);
|
|
}
|
|
__syncthreads();
|
|
|
|
auto xOrder = shape::order(tadOnlyShapeInfo);
|
|
auto yOrder = shape::order(yShapeInfo);
|
|
auto zOrder = shape::order(tadOnlyShapeInfoZ);
|
|
|
|
for (int r = blockIdx.x; r < numTads; r += gridDim.x) {
|
|
|
|
auto rX = x + tadOffsets[r];
|
|
auto rZ = z + tadOffsetsZ[r];
|
|
|
|
|
|
if(tadEWS > 0 && zEWS > 0 && yEWS > 0 && xOrder == yOrder && xOrder == zOrder) {
|
|
for (int i = threadIdx.x; i < tadLength; i+= blockDim.x)
|
|
rZ[i * zEWS] = OpType::op(rX[i * tadEWS], y[i * yEWS]);
|
|
}
|
|
else {
|
|
// it is expected that x and z tads and y array all have the same length
|
|
for (Nd4jLong i = threadIdx.x; i < tadLength; i+= blockDim.x) {
|
|
|
|
auto xOffset = getIndexOffset(i, tadOnlyShapeInfo);
|
|
auto yOffset = getIndexOffset(i, yShapeInfo);
|
|
auto zOffset = getIndexOffset(i, tadOnlyShapeInfoZ);
|
|
rZ[zOffset] = OpType::op(rX[xOffset], y[yOffset]);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
////////////////////////////////////////////////////////////////////////
|
|
template<typename X, typename Y, typename Z>
|
|
template <typename OpType>
|
|
__device__ void Broadcast<X,Y,Z>::transformCuda(
|
|
const void *vx, const Nd4jLong *xShapeInfo,
|
|
const void *vy, const Nd4jLong *yShapeInfo,
|
|
void *vz, const Nd4jLong *zShapeInfo) {
|
|
|
|
const X* x = reinterpret_cast<const X*>(vx);
|
|
const Y* y = reinterpret_cast<const Y*>(vy);
|
|
Z* z = reinterpret_cast<Z*>(vz);
|
|
|
|
__shared__ Nd4jLong zLen;
|
|
__shared__ int rank;
|
|
__shared__ bool xzSameOffsets, yzSameOffsets;
|
|
|
|
if (threadIdx.x == 0) {
|
|
|
|
zLen = shape::length(zShapeInfo);
|
|
rank = shape::rank(zShapeInfo);
|
|
|
|
xzSameOffsets = shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo);
|
|
yzSameOffsets = shape::haveSameShapeAndStrides(yShapeInfo, zShapeInfo);
|
|
}
|
|
__syncthreads();
|
|
|
|
|
|
const auto tid = blockIdx.x * blockDim.x + threadIdx.x;
|
|
|
|
int xCoords[MAX_RANK], yCoords[MAX_RANK], zCoords[MAX_RANK];
|
|
|
|
for (int i = tid; i < zLen; i += blockDim.x * gridDim.x) {
|
|
|
|
shape::index2coords(i, zShapeInfo, zCoords);
|
|
|
|
for (uint j = 0; j < rank; ++j) {
|
|
xCoords[j] = shape::sizeAt(xShapeInfo, j) == 1 ? 0 : zCoords[j];
|
|
yCoords[j] = shape::sizeAt(yShapeInfo, j) == 1 ? 0 : zCoords[j];
|
|
}
|
|
|
|
const auto zOffset = shape::getOffset(zShapeInfo, zCoords);
|
|
const auto xOffset = xzSameOffsets ? zOffset : shape::getOffset(xShapeInfo, xCoords);
|
|
const auto yOffset = yzSameOffsets ? zOffset : shape::getOffset(yShapeInfo, yCoords);
|
|
|
|
z[zOffset] = OpType::op(x[xOffset], y[yOffset]);
|
|
}
|
|
}
|
|
|
|
/*
|
|
BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT Broadcast, , PAIRWISE_TYPES_0);
|
|
BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT Broadcast, , PAIRWISE_TYPES_1);
|
|
BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT Broadcast, , PAIRWISE_TYPES_2);
|
|
BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT Broadcast, , PAIRWISE_TYPES_3);
|
|
BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT Broadcast, , PAIRWISE_TYPES_4);
|
|
BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT Broadcast, , PAIRWISE_TYPES_5);
|
|
BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT Broadcast, , PAIRWISE_TYPES_6);
|
|
BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT Broadcast, , PAIRWISE_TYPES_7);
|
|
BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT Broadcast, , PAIRWISE_TYPES_8);
|
|
BUILD_PAIRWISE_TEMPLATE(template class ND4J_EXPORT Broadcast, , PAIRWISE_TYPES_9);
|
|
*/
|
|
}
|
|
} |