cavis/libnd4j/include/array/impl/InteropDataBuffer.cpp
raver119 29e8e09db6
String changes (#3)
* initial commit

* additional data types & tensor type

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

* next step

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

* missing include

* sparse_to_dense

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

* few more tests files

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

* draft

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

* numeric sparse_to_dense

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

* comment

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

* string sparse_to_dense version

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

* CUDA DataBuffer expand

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

* few tweaks for CUDA build

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

* shape fn for string_split

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

* one more comment

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

* string_split indices

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

* next step

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

* test passes

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

* few rearrangements for databuffer implementations

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

* DataBuffer: move inline methods to common implementations

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

* add native DataBuffer to Nd4j presets

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

* DataBuffer creation

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

* use DataBuffer for allocation

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

* cpu databuffer as deallocatable

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

* DataBuffer setters for bufers

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

* couple of wrappers

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

* DataBuffers being passed around

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

* Bunch of ByteBuffer-related signatures gone

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

* - few more Nd4j signatures removed
- minor fix for bfloat16

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

* nullptr pointer is still a pointer, but 0 as address :)

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

* one special test

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

* empty string array init

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

* one more test in cpp

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

* memcpy instead of databuffer swap

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

* special InteropDataBuffer for front-end languages

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

* few tweaks for java

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

* pointer/indexer actualization

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

* CustomOp returns list for inputArumgents and outputArguments instead of array

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

* redundant call

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

* print_variable op

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

* - view handling (but wrong one)
- print_variable java wrapper

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

* one more test

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

* - empty arrays handling

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

* - deserialization works now

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

* minor fix

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

* meh

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

* one more fix

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

* initial cuda commit

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

* print_variable message validation

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

* CUDA views

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

* CUDA special buffer size

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

* minor update to match master changes

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

* - consider arrays always actual on device for CUDA
- additional PrintVariable constructor
- CudaUtf8Buffer now allocates host buffer by default

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

* meh

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

* - print_variable now allows print from device

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

* InteropDataBuffer data type fix

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

* ...

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

* disable some debug messages

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

* master pulled in

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

* couple of new methods for DataBuffer interop

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

* java side

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

* offsetted constructor

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

* new CUDA deallocator

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

* CUDA backend torn apart

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

* CUDA backend torn apart 2

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

* CUDA backend torn apart 3

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

* - few new tests
- few new methods for DataBuffer management

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

* few more tests + few more tweaks

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

* two failing tests

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

* one more test

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

* two failing tests pass

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

* now we pass DataBuffer to legacy ops too

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

* Native DataBuffer for legacy ops, Java side

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

* CPU java side update

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

* CUDA java side update

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

* no more prepare/register action on java side

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

* NDArray::prepare/register use now accepts vectors

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

* InteropDataBuffer now has few more convenience methods

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

* java bindings update

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

* tick device in NativeOps

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

* Corrected usage of OpaqueBuffer for tests.

* Corrected usage of OpaqueBuffer for java tests.

* NativeOpsTests fixes.

* print_variable now returns scalar

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

* one more test

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

* compat_string_split fix for CUDA

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

* - CUDA execScalar fix
- CUDA lazyAllocateHostPointer now checks java indexer/pointer instead of native pointer

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

* legacy ops DataBuffer migration prototype

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

* ignore device shapeinfo coming from java

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

* minor fix

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

* minor transformAny fix

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

* minor tweak for lazy host allocation

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

* - DataBuffer::memcpy method
- bitcast now uses memcpy

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

* - IndexReduce CUDA dimension buffer fix

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

* views for CPU and CUDA

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

* less spam

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

* optional memory init

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

* async memset

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

* - SummaryStats CUDA fix
- DataBuffer.sameUnderlyingData() impl
- execBroadcast fix

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

* - reduce3All fix
switch to CUDA 10 temporarily

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

* CUDA version

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

* proper memory deallocator registration

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

* HOST_ONLY workspace allocation

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

* temp commit

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

* few conflicts resolved

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

* few minor fixes

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

* one more minor fix

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

* NDArray permute should operate on JVM primitives

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

* - create InteropDataBuffer for shapes as well
- update pointers after view creation in Java

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

* - addressPointer temporary moved to C++

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

* CUDA: don't account offset twice

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

* CUDA: DataBuffer pointer constructor updated

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

* CUDA NDArray.unsafeDuplication() simplified

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

* CUDA minor workspace-related fixes

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

* CPU DataBuffer.reallocate()

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

* print_affinity op

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

* print_affinity java side

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

* CUDA more tweaks for data locality

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

* - compat_string_split tweak
- CudaUtf8Buffer update

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

* INDArray.close() mechanic restored

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

* one more test fixed

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

* - CUDA DataBuffer.reallocate() updated
- cudaMemcpy (synchronous) restored

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

* one last fix

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

* bad import removed

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

* another small fix

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

* one special test

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

* fix bad databuffer size

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

* release primaryBuffer on replace

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

* higher timeout

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

* disable timeouts

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

* dbCreateView now validates offset and length of a view

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

* additional validation for dbExpand

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

* restore timeout back again

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

* smaller distribution for rng test to prevent timeouts

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

* CUDA DataBuffer::memcpy now copies to device all the time

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

* OpaqueDataBuffer now contains all required methods for interop

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

* some javadoc

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

* GC on failed allocations

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

* minoe memcpu tweak

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

* one more bitcast test

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

* - NDArray::deviceId() propagation
- special multi-threaded test for data locality checks

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

* DataBuffer additional syncStream

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

* DataBuffer additional syncStream

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

* one ignored test

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

* skip host alloc for empty arrays

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

* ByteBuffer support is back

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

* DataBuffer::memcpy minor fix

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

* few minor prelu/bp tweaks

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

* nullify-related fixes

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

* PReLU fixes (#157)

Signed-off-by: Alex Black <blacka101@gmail.com>

* Build fixed

* Fix tests

* one more ByteBuffer signature restored

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

* nd4j-jdbc-hsql profiles fix

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

* nd4j-jdbc-hsql profiles fix

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

* PReLU weight init fix

Signed-off-by: Alex Black <blacka101@gmail.com>

* Small PReLU fix

Signed-off-by: Alex Black <blacka101@gmail.com>

* - INDArray.migrate() reactivated
- DataBuffer::setDeviceId(...) added
- InteropDataBuffer Z syncToDevice added for views

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

* missed file

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

* Small tweak

Signed-off-by: Alex Black <blacka101@gmail.com>

* cuda 10.2

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

* minor fix

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

Co-authored-by: shugeo <sgazeos@gmail.com>
Co-authored-by: Alex Black <blacka101@gmail.com>
Co-authored-by: Alexander Stoyakin <alexander.stoyakin@gmail.com>
2020-01-04 13:27:50 +03:00

147 lines
5.0 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 raver119@gmail.com
//
#include <array/InteropDataBuffer.h>
#include <array/DataTypeUtils.h>
#include <execution/AffinityManager.h>
#include <helpers/logger.h>
namespace nd4j {
InteropDataBuffer::InteropDataBuffer(InteropDataBuffer &dataBuffer, uint64_t length, uint64_t offset) {
_dataBuffer = dataBuffer.getDataBuffer();
// offset is always absolute to the original buffer
_offset = offset;
if (_offset + length > _dataBuffer->getLenInBytes()) {
throw std::runtime_error("offset + length is higher than original length");
}
}
InteropDataBuffer::InteropDataBuffer(std::shared_ptr<DataBuffer> databuffer) {
_dataBuffer = databuffer;
}
InteropDataBuffer::InteropDataBuffer(size_t elements, nd4j::DataType dtype, bool allocateBoth) {
if (elements == 0) {
_dataBuffer = std::make_shared<DataBuffer>();
_dataBuffer->setDataType(dtype);
} else {
_dataBuffer = std::make_shared<DataBuffer>(elements, dtype, nullptr, allocateBoth);
}
}
std::shared_ptr<DataBuffer> InteropDataBuffer::getDataBuffer() const {
return _dataBuffer;
}
std::shared_ptr<DataBuffer> InteropDataBuffer::dataBuffer() {
return _dataBuffer;
}
void* InteropDataBuffer::primary() const {
return reinterpret_cast<int8_t *>(_dataBuffer->primary()) + _offset;
}
void* InteropDataBuffer::special() const {
return reinterpret_cast<int8_t *>(_dataBuffer->special()) + _offset;
}
void InteropDataBuffer::setPrimary(void* ptr, size_t length) {
_dataBuffer->setPrimaryBuffer(ptr, length);
}
void InteropDataBuffer::setSpecial(void* ptr, size_t length) {
_dataBuffer->setSpecialBuffer(ptr, length);
}
uint64_t InteropDataBuffer::offset() const {
return _offset;
}
void InteropDataBuffer::setOffset(uint64_t offset) {
_offset = offset;
}
int InteropDataBuffer::deviceId() const {
return _dataBuffer->deviceId();
}
void InteropDataBuffer::registerSpecialUse(const std::vector<const InteropDataBuffer*>& writeList, const std::vector<const InteropDataBuffer*>& readList) {
for (const auto &v:writeList) {
if (v == nullptr)
continue;
v->getDataBuffer()->writeSpecial();
}
}
void InteropDataBuffer::prepareSpecialUse(const std::vector<const InteropDataBuffer*>& writeList, const std::vector<const InteropDataBuffer*>& readList, bool synchronizeWritables) {
auto currentDeviceId = nd4j::AffinityManager::currentDeviceId();
for (const auto &v:readList) {
if (v == nullptr)
continue;
if (v->getDataBuffer()->deviceId() != currentDeviceId)
v->getDataBuffer()->migrate();
v->getDataBuffer()->syncToSpecial();
}
// we don't tick write list, only ensure the same device affinity
for (const auto &v:writeList) {
if (v == nullptr)
continue;
// special case for legacy ops - views can be updated on host side, thus original array can be not updated
if (!v->getDataBuffer()->isSpecialActual())
v->getDataBuffer()->syncToSpecial();
if (v->getDataBuffer()->deviceId() != currentDeviceId)
v->getDataBuffer()->migrate();
}
}
void InteropDataBuffer::registerPrimaryUse(const std::vector<const InteropDataBuffer*>& writeList, const std::vector<const InteropDataBuffer*>& readList) {
for (const auto &v:writeList) {
if (v == nullptr)
continue;
}
}
void InteropDataBuffer::preparePrimaryUse(const std::vector<const InteropDataBuffer*>& writeList, const std::vector<const InteropDataBuffer*>& readList, bool synchronizeWritables) {
for (const auto &v:readList) {
if (v == nullptr)
continue;
v->getDataBuffer()->syncToPrimary(LaunchContext::defaultContext());
}
}
void InteropDataBuffer::expand(size_t newlength) {
_dataBuffer->expand(newlength * DataTypeUtils::sizeOf(_dataBuffer->getDataType()));
}
void InteropDataBuffer::setDeviceId(int deviceId) {
_dataBuffer->setDeviceId(deviceId);
}
}