285 lines
16 KiB
Plaintext
285 lines
16 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
|
||
|
|
******************************************************************************/
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
#include <op_boilerplate.h>
|
||
|
|
#include <pointercast.h>
|
||
|
|
#include <helpers/TAD.h>
|
||
|
|
#include <loops/grid_strided.legacy>
|
||
|
|
#include <types/float16.h>
|
||
|
|
|
||
|
|
|
||
|
|
#define GRID_WIDTH 19 // number of pointers within single grid row
|
||
|
|
|
||
|
|
#include <ops/ops.h>
|
||
|
|
#include <ops/meta_ops.h>
|
||
|
|
#include <loops/legacy_ops.h>
|
||
|
|
|
||
|
|
template <typename T>
|
||
|
|
__device__ inline static void metaPredicateStridedGeneric(const int opTypeA, const int opNumA, const int opTypeB, const int opNumB,
|
||
|
|
Nd4jLong N, T *dx, Nd4jLong xStride, T *dy, Nd4jLong yStride, T *dz, Nd4jLong zStride, T *extraA, T *extraB, T scalarA, T scalarB
|
||
|
|
) {
|
||
|
|
__shared__ Nd4jPointer params[2];
|
||
|
|
__shared__ T *paramsPtr;
|
||
|
|
if (threadIdx.x == 0) {
|
||
|
|
if (opTypeA == 0) {
|
||
|
|
params[0] = reinterpret_cast<Nd4jPointer *>(&scalarA);
|
||
|
|
}
|
||
|
|
else params[0] = reinterpret_cast<Nd4jPointer *>(extraA);
|
||
|
|
|
||
|
|
if (opTypeB == 0) {
|
||
|
|
params[1] = reinterpret_cast<Nd4jPointer *>(&scalarB);
|
||
|
|
}
|
||
|
|
else params[1] = reinterpret_cast<Nd4jPointer *>(extraB);
|
||
|
|
|
||
|
|
paramsPtr = reinterpret_cast<T *>(params);
|
||
|
|
}
|
||
|
|
__syncthreads();
|
||
|
|
#ifdef __ND4J_EXPERIMENTAL__
|
||
|
|
if (opTypeB == 0) { // SCALAR
|
||
|
|
if (opTypeA == 0) {
|
||
|
|
// double scalar
|
||
|
|
DISPATCH_METAOP(functions::transform::Transform<T>::template transformCuda, PARAMS(N, dx, xStride, paramsPtr, dz, zStride, nullptr, nullptr, nullptr), MetaOp, OPS_A(SCALAR_OPS), OPS_B(SCALAR_OPS));
|
||
|
|
} else if (opTypeA == 1) {
|
||
|
|
// transform
|
||
|
|
DISPATCH_METAOP(functions::transform::Transform<T>::template transformCuda, PARAMS(N, dx, xStride, paramsPtr, dz, zStride, nullptr, nullptr, nullptr), MetaOp, OPS_A(TRANSFORM_OPS), OPS_B(SCALAR_OPS));
|
||
|
|
} else if (opTypeA == 2) {
|
||
|
|
// pwt
|
||
|
|
// this is the most important thing here: its Dup() + Scalar
|
||
|
|
DISPATCH_METAOP(functions::grid::GRID<T>::template transformCuda, PARAMS(N, dx, dy, xStride, yStride, paramsPtr, dz, zStride, nullptr, nullptr, nullptr), InvertedMetaOp, OPS_A(PAIRWISE_TRANSFORM_OPS), OPS_B(SCALAR_OPS));
|
||
|
|
}
|
||
|
|
} else if (opTypeB == 1) { // TRANSFORM
|
||
|
|
if (opTypeA == 0) {
|
||
|
|
DISPATCH_METAOP(functions::transform::Transform<T>::template transformCuda, PARAMS(N, dx, xStride, paramsPtr, dz, zStride, nullptr, nullptr, nullptr), MetaOp, OPS_A(SCALAR_OPS), OPS_B(TRANSFORM_OPS));
|
||
|
|
}
|
||
|
|
} else if (opTypeB == 2) { // PWT
|
||
|
|
if (opTypeA == 0) { // SCALAR
|
||
|
|
|
||
|
|
DISPATCH_METAOP(functions::pairwise_transforms::PairWiseTransform<T>::template transformCuda, PARAMS(N, dx, dy, xStride, yStride, paramsPtr, dz, zStride, nullptr, nullptr, nullptr), MetaOp, OPS_A(SCALAR_OPS), OPS_B(PAIRWISE_TRANSFORM_OPS));
|
||
|
|
} else if (opTypeA == 1) { // TRANSFORM
|
||
|
|
|
||
|
|
DISPATCH_METAOP(functions::grid::GRID<T>::template transformCuda, PARAMS(N, dx, dy, xStride, yStride, paramsPtr, dz, zStride, nullptr, nullptr, nullptr), MetaOp, OPS_A(TRANSFORM_OPS), OPS_B(PAIRWISE_TRANSFORM_OPS));
|
||
|
|
} else if (opTypeA == 2) {
|
||
|
|
|
||
|
|
}
|
||
|
|
} else {
|
||
|
|
if (threadIdx.x == 0 && blockIdx.x)
|
||
|
|
printf("Unknown opTypeB: [%i]\n", opTypeB);
|
||
|
|
}
|
||
|
|
#else
|
||
|
|
if (opTypeA == 2) {
|
||
|
|
if (opTypeB == 0) {
|
||
|
|
// DISPATCH_METAOP(functions::pairwise_transforms::PairWiseTransform<T>::template transformCuda, PARAMS(N, dx, dy, xStride, yStride, paramsPtr, dz, zStride, nullptr, nullptr, nullptr), InvertedMetaOp, OPS_A(PAIRWISE_TRANSFORM_OPS), OPS_B(SCALAR_OPS));
|
||
|
|
// functions::pairwise_transforms::PairWiseTransform<T>::template transformCuda<simdOps::InvertedMetaOp<T, simdOps::Copy<T>, simdOps::Multiply<T>>>(N, dx, dy, xStride, yStride, paramsPtr, dz, zStride, nullptr, nullptr, nullptr);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
#endif
|
||
|
|
}
|
||
|
|
|
||
|
|
template<typename T, typename OpClass>
|
||
|
|
__device__ static inline void invertedMetaPairwiseStridedGeneric(const int opTypeA, const int opTypeB, Nd4jLong N, T *dx, Nd4jLong xStride, T *dy, Nd4jLong yStride, T *dz, Nd4jLong zStride, T *extraA, T *extraB, T scalarA, T scalarB) {
|
||
|
|
__shared__ Nd4jPointer params[2];
|
||
|
|
__shared__ T *paramsPtr;
|
||
|
|
if (threadIdx.x == 0) {
|
||
|
|
if (opTypeA == 0) {
|
||
|
|
params[0] = reinterpret_cast<Nd4jPointer *>(&scalarA);
|
||
|
|
}
|
||
|
|
else params[0] = reinterpret_cast<Nd4jPointer *>(extraA);
|
||
|
|
|
||
|
|
if (opTypeB == 0) {
|
||
|
|
params[1] = reinterpret_cast<Nd4jPointer *>(&scalarB);
|
||
|
|
}
|
||
|
|
else params[1] = reinterpret_cast<Nd4jPointer *>(extraB);
|
||
|
|
|
||
|
|
paramsPtr = reinterpret_cast<T *>(params);
|
||
|
|
}
|
||
|
|
__syncthreads();
|
||
|
|
|
||
|
|
functions::grid::GRIDStrided<T>::template transformCuda<OpClass>(N, dx, dy, xStride, yStride, paramsPtr, dz, zStride, nullptr, nullptr, nullptr);
|
||
|
|
};
|
||
|
|
|
||
|
|
|
||
|
|
template<typename T>
|
||
|
|
__device__ static inline void invertedMetaPairwiseStridedNumericGeneric(const int opTypeA, const int opNumA, const int opTypeB, const int opNumB, Nd4jLong N, T *dx, Nd4jLong xStride, T *dy, Nd4jLong yStride, T *dz, Nd4jLong zStride, T *extraA, T *extraB, T scalarA, T scalarB) {
|
||
|
|
__shared__ Nd4jPointer params[2];
|
||
|
|
__shared__ T *paramsPtr;
|
||
|
|
if (threadIdx.x == 0) {
|
||
|
|
if (opTypeA == 0) {
|
||
|
|
params[0] = reinterpret_cast<Nd4jPointer *>(&scalarA);
|
||
|
|
}
|
||
|
|
else params[0] = reinterpret_cast<Nd4jPointer *>(extraA);
|
||
|
|
|
||
|
|
if (opTypeB == 0) {
|
||
|
|
params[1] = reinterpret_cast<Nd4jPointer *>(&scalarB);
|
||
|
|
}
|
||
|
|
else params[1] = reinterpret_cast<Nd4jPointer *>(extraB);
|
||
|
|
|
||
|
|
paramsPtr = reinterpret_cast<T *>(params);
|
||
|
|
}
|
||
|
|
__syncthreads();
|
||
|
|
|
||
|
|
functions::grid::GRIDStrided<T>::transformCuda(opTypeA, opNumA, opTypeB, opNumB, N, dx, dy, xStride, yStride, paramsPtr, dz, zStride, nullptr, nullptr, nullptr);
|
||
|
|
};
|
||
|
|
|
||
|
|
extern "C" __global__ void invertedMetaPairwiseStridedNumericFloat(const int opTypeA, const int opNumA, const int opTypeB, const int opNumB, Nd4jLong N, float *dx, Nd4jLong xStride, float *dy, Nd4jLong yStride, float *dz, Nd4jLong zStride, float *extraA, float *extraB, float scalarA, float scalarB) {
|
||
|
|
invertedMetaPairwiseStridedNumericGeneric<float>(opTypeA, opNumA, opTypeB, opNumB, N, dx, xStride, dy, yStride, dz, zStride, extraA, extraB, scalarA, scalarB);
|
||
|
|
}
|
||
|
|
|
||
|
|
extern "C" __global__ void invertedMetaPairwiseStridedNumericDouble(const int opTypeA, const int opNumA, const int opTypeB, const int opNumB, Nd4jLong N, double *dx, Nd4jLong xStride, double *dy, Nd4jLong yStride, double *dz, Nd4jLong zStride, double *extraA, double *extraB, double scalarA, double scalarB) {
|
||
|
|
invertedMetaPairwiseStridedNumericGeneric<double>(opTypeA, opNumA, opTypeB, opNumB, N, dx, xStride, dy, yStride, dz, zStride, extraA, extraB, scalarA, scalarB);
|
||
|
|
}
|
||
|
|
|
||
|
|
extern "C" __global__ void invertedMetaPairwiseStridedNumericHalf(const int opTypeA, const int opNumA, const int opTypeB, const int opNumB, Nd4jLong N, float16 *dx, Nd4jLong xStride, float16 *dy, Nd4jLong yStride, float16 *dz, Nd4jLong zStride, float16 *extraA, float16 *extraB, float16 scalarA, float16 scalarB) {
|
||
|
|
invertedMetaPairwiseStridedNumericGeneric<float16>(opTypeA, opNumA, opTypeB, opNumB, N, dx, xStride, dy, yStride, dz, zStride, extraA, extraB, scalarA, scalarB);
|
||
|
|
}
|
||
|
|
|
||
|
|
#ifndef __CLION_IDE__
|
||
|
|
// kernels set for pairwise + scalar based on stride const int opTypeA, const int opTypeB, Nd4jLong N, T *dx, int xStride, T *dy, int yStride, T *dz, int zStride, T *extraA, T *extraB, T scalarA, T scalarB
|
||
|
|
//DISPATCH_KERNEL_META(invertedMetaPairwiseStrided_Pairwise_Scalar_, invertedMetaPairwiseStridedGeneric, float, metaOps::InvertedMetaOp, INPUT(const int opTypeA, const int opTypeB, Nd4jLong N, float *dx, int xStride, float *dy, int yStride, float *dz, int zStride, float *extraA, float *extraB, float scalarA, float scalarB), PARAMS(opTypeA, opTypeB, N, dx, xStride, dy, yStride, dz, zStride, extraA, extraB, scalarA, scalarB), OPS_A(PAIRWISE_TRANSFORM_OPS), OPS_B(SCALAR_OPS))
|
||
|
|
//DISPATCH_KERNEL_META(invertedMetaPairwiseStrided_Pairwise_Scalar_, invertedMetaPairwiseStridedGeneric, double, metaOps::InvertedMetaOp, INPUT(const int opTypeA, const int opTypeB, Nd4jLong N, double *dx, int xStride, double *dy, int yStride, double *dz, int zStride, double *extraA, double *extraB, double scalarA, double scalarB), PARAMS(opTypeA, opTypeB, N, dx, xStride, dy, yStride, dz, zStride, extraA, extraB, scalarA, scalarB), OPS_A(PAIRWISE_TRANSFORM_OPS), OPS_B(SCALAR_OPS))
|
||
|
|
//DISPATCH_KERNEL_META(invertedMetaPairwiseStrided_Pairwise_Scalar_, invertedMetaPairwiseStridedGeneric, float16, metaOps::InvertedMetaOp, INPUT(const int opTypeA, const int opTypeB, Nd4jLong N, float16 *dx, int xStride, float16 *dy, int yStride, float16 *dz, int zStride, float16 *extraA, float16 *extraB, float16 scalarA, float16 scalarB), PARAMS(opTypeA, opTypeB, N, dx, xStride, dy, yStride, dz, zStride, extraA, extraB, scalarA, scalarB), OPS_A(PAIRWISE_TRANSFORM_OPS), OPS_B(SCALAR_OPS))
|
||
|
|
#endif
|
||
|
|
|
||
|
|
|
||
|
|
namespace functions {
|
||
|
|
namespace grid {
|
||
|
|
template <typename T>
|
||
|
|
__device__ __noinline__ T invertedOpExecutorB(const int opTypeA, const int opNumA, const int opTypeB, const int opNumB, T x, T y, T *extras);
|
||
|
|
|
||
|
|
template <typename T>
|
||
|
|
__device__ __noinline__ T execute_2OEF(const int opType, const int opNum, T x, T y, T *extras);
|
||
|
|
|
||
|
|
template <typename T>
|
||
|
|
__device__ __noinline__ T execute_1OEF(const int opType, const int opNum, T x, T *extras);
|
||
|
|
|
||
|
|
|
||
|
|
/**
|
||
|
|
* This method is able to execute various ops that takes 2 operands (x, y) + extras
|
||
|
|
* @tparam T
|
||
|
|
*/
|
||
|
|
template <typename T>
|
||
|
|
__device__ __noinline__ T execute_2OEF(const int opType, const int opNum, T x, T y, T *extras) {
|
||
|
|
T z;
|
||
|
|
switch(opType) {
|
||
|
|
case 2: {
|
||
|
|
EXECUTE_NOE((x, y, extras), OPS_A(PAIRWISE_TRANSFORM_OPS));
|
||
|
|
};
|
||
|
|
break;
|
||
|
|
default: {
|
||
|
|
PRINT_FIRST("Unknown opType provided: [%i]\n", opType);
|
||
|
|
}
|
||
|
|
break;
|
||
|
|
}
|
||
|
|
return z;
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
/**
|
||
|
|
* This method is able to execute various ops that takes 1 operand (x) + extras
|
||
|
|
* @tparam T
|
||
|
|
*/
|
||
|
|
template <typename T>
|
||
|
|
__device__ __noinline__ T execute_1OEF(const int opType, const int opNum, T x, T *extras) {
|
||
|
|
T z;
|
||
|
|
switch(opType) {
|
||
|
|
case 0: {
|
||
|
|
EXECUTE_NOE((x, extras), OPS_A(SCALAR_OPS));
|
||
|
|
}
|
||
|
|
break;
|
||
|
|
default: {
|
||
|
|
PRINT_FIRST("Unknown opType provided: [%i]\n", opType);
|
||
|
|
}
|
||
|
|
break;
|
||
|
|
}
|
||
|
|
|
||
|
|
return z;
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
template <typename T>
|
||
|
|
__device__ __noinline__ T invertedOpExecutorB(const int opTypeA, const int opNumA, const int opTypeB, const int opNumB, T x, T y, T *extras) {
|
||
|
|
// this code is basically InvertedMetaOp, reorganized to suit per-type execution
|
||
|
|
|
||
|
|
Nd4jPointer *wrap = reinterpret_cast<Nd4jPointer *> (extras);
|
||
|
|
T *paramsA = reinterpret_cast<T *> (wrap[0]);
|
||
|
|
T *paramsB = reinterpret_cast<T *> (wrap[1]);
|
||
|
|
T intermediate;
|
||
|
|
|
||
|
|
// Executing first op, opA
|
||
|
|
intermediate = functions::grid::execute_2OEF<T>(opTypeA, opNumA, x, y, paramsA);
|
||
|
|
|
||
|
|
// Executing second op, opB
|
||
|
|
T intermediate2 = functions::grid::execute_1OEF<T>(opTypeB, opNumB, intermediate, paramsB);
|
||
|
|
|
||
|
|
// just returning result now
|
||
|
|
return intermediate2;
|
||
|
|
}
|
||
|
|
|
||
|
|
template<typename T>
|
||
|
|
template<typename OpType>
|
||
|
|
__device__ void GRIDStrided<T>::transformCuda(Nd4jLong n, T *dx, T *dy, Nd4jLong incx, Nd4jLong incy, T *params, T *result, Nd4jLong incz,int *allocationPointer, UnifiedSharedMemory *manager,Nd4jLong *tadOnlyShapeInfo) {
|
||
|
|
int tid = blockIdx.x * blockDim.x + threadIdx.x;
|
||
|
|
|
||
|
|
if (incx == incy && incy == incz && incx == 1) {
|
||
|
|
for (Nd4jLong i = tid; i < n; i += gridDim.x * blockDim.x) {
|
||
|
|
result[i] = OpType::op(dx[i], dy[i], params);
|
||
|
|
}
|
||
|
|
} else {
|
||
|
|
for (Nd4jLong i = tid; i < n; i += gridDim.x * blockDim.x) {
|
||
|
|
result[i * incz] = OpType::op(dx[i * incx], dy[i * incy], params);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
template<typename T>
|
||
|
|
__device__ void GRIDStrided<T>::transformCuda(const int opTypeA, const int opNumA, const int opTypeB, const int opNumB, Nd4jLong n, T *dx, T *dy, Nd4jLong incx, Nd4jLong incy, T *params, T *result, Nd4jLong incz,int *allocationPointer, UnifiedSharedMemory *manager,Nd4jLong *tadOnlyShapeInfo) {
|
||
|
|
int tid = blockIdx.x * blockDim.x + threadIdx.x;
|
||
|
|
|
||
|
|
if (incx == incy && incy == incz && incx == 1) {
|
||
|
|
for (Nd4jLong i = tid; i < n; i += gridDim.x * blockDim.x) {
|
||
|
|
result[i] = functions::grid::invertedOpExecutorB<T>(opTypeA, opNumA, opTypeB, opNumB, dx[i], dy[i], params);
|
||
|
|
}
|
||
|
|
} else {
|
||
|
|
for (Nd4jLong i = tid; i < n; i += gridDim.x * blockDim.x) {
|
||
|
|
result[i * incz] = functions::grid::invertedOpExecutorB<T>(opTypeA, opNumA, opTypeB, opNumB, dx[i * incx], dy[i * incy], params);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
template <>
|
||
|
|
void GRIDStrided<float>::execMetaPredicateStrided(cudaStream_t * stream, Nd4jPointer *extras, const int opTypeA, const int opNumA, const int opTypeB, const int opNumB, Nd4jLong N, float *dx, Nd4jLong xStride, float *dy, Nd4jLong yStride, float *dz, Nd4jLong zStride, float *extraA, float *extraB, float scalarA, float scalarB) {
|
||
|
|
invertedMetaPairwiseStridedNumericFloat<<<128, 1024, 1024, *stream>>>(opTypeA, opNumA, opTypeB, opNumB, N, dx, xStride, dy, yStride, dz, zStride, extraA, extraB, scalarA, scalarB);
|
||
|
|
}
|
||
|
|
|
||
|
|
template <>
|
||
|
|
void GRIDStrided<float16>::execMetaPredicateStrided(cudaStream_t * stream, Nd4jPointer *extras, const int opTypeA, const int opNumA, const int opTypeB, const int opNumB, Nd4jLong N, float16 *dx, Nd4jLong xStride, float16 *dy, Nd4jLong yStride, float16 *dz, Nd4jLong zStride, float16 *extraA, float16 *extraB, float16 scalarA, float16 scalarB) {
|
||
|
|
invertedMetaPairwiseStridedNumericHalf<<<128, 1024, 1024, *stream>>>(opTypeA, opNumA, opTypeB, opNumB, N, dx, xStride, dy, yStride, dz, zStride, extraA, extraB, scalarA, scalarB);
|
||
|
|
}
|
||
|
|
|
||
|
|
template <>
|
||
|
|
void GRIDStrided<double>::execMetaPredicateStrided(cudaStream_t * stream, Nd4jPointer *extras, const int opTypeA, const int opNumA, const int opTypeB, const int opNumB, Nd4jLong N, double *dx, Nd4jLong xStride, double *dy, Nd4jLong yStride, double *dz, Nd4jLong zStride, double *extraA, double *extraB, double scalarA, double scalarB) {
|
||
|
|
invertedMetaPairwiseStridedNumericDouble<<<128, 1024, 1024, *stream>>>(opTypeA, opNumA, opTypeB, opNumB, N, dx, xStride, dy, yStride, dz, zStride, extraA, extraB, scalarA, scalarB);
|
||
|
|
}
|
||
|
|
|
||
|
|
//template class GRID<float>;
|
||
|
|
//template class GRID<float16>;
|
||
|
|
//template class GRID<double>;
|
||
|
|
}
|
||
|
|
}
|