raver119 c969b724bb [WIP] more CUDA stuff (#57)
* initial commit

Signed-off-by: raver119 <raver119@gmail.com>

* Added gradcheck test for dynamic_partition_bp op.

* - implementation of dilation op (cpu and cuda)

Signed-off-by: Yurii <yurii@skymind.io>

* Fixed broadcast_dynamic_shape 1D case and tests.

* Fixed usage of default integer arguments.

* Fixed dynamic_partition_bp op and tests.

* Eliminated test with grad check for dynamic_partition_bp op.

* start working on cuda svd - porting available corresponding api from cuSOLVER library

Signed-off-by: Yurii <yurii@skymind.io>

* provide prelu_bp

Signed-off-by: Yurii <yurii@skymind.io>

* - provide gruCell_bp (old version ??)

Signed-off-by: Yurii <yurii@skymind.io>

* - polishing cumsum_bp and cumprod_bp tests

Signed-off-by: Yurii <yurii@skymind.io>

* provide sparseSoftmaxCrossEntropyWithLogits and sparseSoftmaxCrossEntropyWithLogits_grad

Signed-off-by: Yurii <yurii@skymind.io>

* Fixed atomicMul with float input/output

* implementation of cuda kernel for triu_bp operation

Signed-off-by: Yurii <yurii@skymind.io>

* Refactored lup helper to add parrallel computing.

* cusolver libraries

Signed-off-by: raver119 <raver119@gmail.com>

* uncomment cuSolver APIs in svd.cu

Signed-off-by: Yurii <yurii@skymind.io>

* cusolver var

Signed-off-by: raver119 <raver119@gmail.com>

* - further work on cuSolver svd

Signed-off-by: Yurii <yurii@skymind.io>

* Implement usage of cuda solver to LUP decomposition.

* - correct naames in lup functions

Signed-off-by: Yurii <yurii@skymind.io>

* correct svdQR cuda

Signed-off-by: Yurii <yurii@skymind.io>

* - provide transpositions of input matrices in case of c order in svdCudaQR

Signed-off-by: Yurii <yurii@skymind.io>

* Fixed implementation issues with LUP usign cuda solver.

* Implementation of matrix_determinant helper with cuda kernels. Working revision.

* Implemented log_matrix_determinant helper with cuda kernels.

* - implementation of batched cuda svd

Signed-off-by: Yurii <yurii@skymind.io>

* Refactored cholesky helper and implementation of cuda solver cholesky batch.

* - implementation of cuda kernel for tile bp

Signed-off-by: Yurii <yurii@skymind.io>

* Implementation of cholesky and logdet with cuda kernels.

* - implementation of cuda kernel for sru_bidirectional

Signed-off-by: Yurii <yurii@skymind.io>

* Fixed cholesky helper.

* Cholesky op helper implementation. Working double-based cublas implementation.

* bad import excluded

Signed-off-by: raver119 <raver119@gmail.com>

* Finished with cuda implementation of cholesky helper and tests.

* - implementation of cuda kernel for sru_bidirectional_backprop operation

Signed-off-by: Yurii <yurii@skymind.io>

* Implementation of matrix_inverse op helper with cuda kernels. The first revision.

* - start working on gruCell_bp

Signed-off-by: Yurii <yurii@skymind.io>

* Implementation of matrix_inverse helper.

* - further work on new gruCell_bp

Signed-off-by: Yurii <yurii@skymind.io>

* cuBLAS related fixes

Signed-off-by: raver119 <raver119@gmail.com>

* calculateOutputShapes() now passes device buffers as well

Signed-off-by: raver119 <raver119@gmail.com>

* special concat/average/accumulate init host pointers now

Signed-off-by: raver119 <raver119@gmail.com>

* few more tweaks

Signed-off-by: raver119 <raver119@gmail.com>

* additional CudaDataBufferFactory signatures certain for data types

Signed-off-by: raver119 <raver119@gmail.com>

* cuSolver host buffer

Signed-off-by: raver119 <raver119@gmail.com>

* buffer to buffer memcpy host ptr allocation

Signed-off-by: raver119 <raver119@gmail.com>
2019-07-20 23:05:21 +10:00

79 lines
3.1 KiB
C++

/*******************************************************************************
* 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 GS <sgazeos@gmail.com>
//
#include <ops/declarable/helpers/bds.h>
#include <Status.h>
namespace nd4j {
namespace ops {
namespace helpers {
Nd4jStatus bdsFunctor(nd4j::LaunchContext * context, NDArray* x_shape, NDArray* y_shape, NDArray* output) {
if (x_shape->lengthOf() == 1 || y_shape->lengthOf() == 1) {// except case
// lenght are equals
if (x_shape->lengthOf() == y_shape->lengthOf()) {
auto greater = (x_shape->e<Nd4jLong>(0) < y_shape->e<Nd4jLong>(0) ? y_shape : x_shape);
output->assign(greater);
}
else {
auto lesser = (x_shape->lengthOf() == 1 ? x_shape : y_shape);
auto greater = (x_shape->lengthOf() == 1 ? y_shape : x_shape);
output->assign(greater);
auto lastG = greater->lengthOf() - 1;
auto lastL = lesser->lengthOf() - 1;
if (greater->e<Nd4jLong>(lastG) < lesser->e<Nd4jLong>(lastL))
output->p(lastG, lesser->e(lastL));
}
}
else {
//int e = 0, x = 0, y = 0;
Nd4jLong xLen = x_shape->lengthOf();
Nd4jLong yLen = y_shape->lengthOf();
Nd4jLong zLen = output->lengthOf();
Nd4jLong borderLen = nd4j::math::nd4j_min(xLen, yLen);
for (Nd4jLong e = 0; e < zLen; e++) {
Nd4jLong val;
if (e < borderLen) {
val = nd4j::math::nd4j_max(x_shape->e<Nd4jLong>(e), y_shape->e<Nd4jLong>(e));
} else if (e < xLen) {
val = nd4j::math::nd4j_max(x_shape->e<Nd4jLong>(e), y_shape->e<Nd4jLong>(yLen - 1));
} else {
val = nd4j::math::nd4j_max(x_shape->e<Nd4jLong>(xLen - 1), y_shape->e<Nd4jLong>(e));
}
// if (e)
// if (val != output->e<Nd4jLong>(e - 1)) {
// nd4j_printf(
// "broadcast_dynamic_shape: Input shapes should be compatible, but %lld and %lld were given.\n",
// val, output->e<Nd4jLong>(e - 1));
// return Status::CODE(ND4J_STATUS_VALIDATION, "broadcast_dynamic_shape: BDS validation failed!");
// }
output->p(e, val);
}
}
return Status::OK();
}
}
}
}