cavis/libnd4j/include/execution/cuda/ContextBuffers.cu
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

238 lines
7.3 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 <execution/ContextBuffers.h>
#include <exceptions/cuda_exception.h>
#include <logger.h>
#include <AffinityManager.h>
#include <cuda.h>
#include <cuda_runtime_api.h>
#include <cuda_runtime.h>
#include <cuda_device_runtime_api.h>
namespace nd4j {
ContextBuffers::ContextBuffers() {
//nd4j_printf("Creating ContextBuffers for device [%i]\n", AffinityManager::currentDeviceId());
_deviceId = AffinityManager::currentDeviceId();
}
ContextBuffers::ContextBuffers(const ContextBuffers &other) {
release();
this->_initialized = other._initialized;
this->_allocated = other._allocated;
this->_deviceId = other._deviceId;
this->_specialStream = other._specialStream;
this->_execStream = other._execStream;
this->_allocationPointer = other._allocationPointer;
this->_reductionPointer = other._reductionPointer;
this->_scalarPointer = other._scalarPointer;
}
ContextBuffers& ContextBuffers::operator=(const ContextBuffers& other) {
release();
this->_initialized = other._initialized;
this->_allocated = other._allocated;
this->_deviceId = other._deviceId;
this->_specialStream = other._specialStream;
this->_execStream = other._execStream;
this->_allocationPointer = other._allocationPointer;
this->_reductionPointer = other._reductionPointer;
this->_scalarPointer = other._scalarPointer;
return *this;
}
ContextBuffers& ContextBuffers::operator=(ContextBuffers&& other) {
release();
this->_initialized = other._initialized;
this->_allocated = other._allocated;
this->_deviceId = other._deviceId;
this->_specialStream = other._specialStream;
this->_execStream = other._execStream;
this->_allocationPointer = other._allocationPointer;
this->_reductionPointer = other._reductionPointer;
this->_scalarPointer = other._scalarPointer;
return *this;
}
void ContextBuffers::release() {
if (_allocated) {
//nd4j_printf("Releasing ContextBuffers on device [%i]\n", _deviceId);
if (_allocationPointer != nullptr)
cudaFree(_allocationPointer);
if (_scalarPointer != nullptr)
cudaFree(_scalarPointer);
if (_allocationPointer != nullptr)
cudaFree(_reductionPointer);
auto _cudaStream = reinterpret_cast<cudaStream_t*>(_execStream);
auto _cudaSpecialStream = reinterpret_cast<cudaStream_t*>(_specialStream);
cudaStreamSynchronize(*_cudaStream);
cudaStreamSynchronize(*_cudaSpecialStream);
cudaStreamDestroy(*_cudaStream);
cudaStreamDestroy(*_cudaSpecialStream);
delete _cudaStream;
delete _cudaSpecialStream;
//////
_allocated = false;
_deviceId = -1;
this->_specialStream = nullptr;
this->_execStream = nullptr;
this->_allocationPointer = nullptr;
this->_reductionPointer = nullptr;
this->_scalarPointer = nullptr;
}
_initialized = false;
}
ContextBuffers::~ContextBuffers() {
release();
}
ContextBuffers::ContextBuffers(void* rPointer, void* sPointer, void* aPointer, bool isOwner) {
_reductionPointer = rPointer;
_scalarPointer = sPointer;
_allocationPointer = aPointer;
_allocated = isOwner;
}
void ContextBuffers::initialize() {
_deviceId = AffinityManager::currentNativeDeviceId();
//nd4j_printf("Initializing buffers on deviceId [%i]\n", _deviceId);
auto res = cudaMalloc(reinterpret_cast<void**>(&_reductionPointer), 1024 * 1024 * 8);
if (res != 0)
throw cuda_exception::build("_reductionPointer allocation failed", res);
res = cudaHostAlloc(reinterpret_cast<void**>(&_scalarPointer), 16, cudaHostAllocDefault);
if (res != 0)
throw cuda_exception::build("_scalarPointer allocation failed", res);
res = cudaMalloc(reinterpret_cast<void**>(&_allocationPointer), 1024 * 1024 * 8);
if (res != 0)
throw cuda_exception::build("_allocationPointer allocation failed", res);
_execStream = new cudaStream_t();
_specialStream = new cudaStream_t();
if (nullptr == _execStream || nullptr == _specialStream)
throw std::runtime_error("Failed to allocate memory for new CUDA stream");
res = cudaStreamCreate(reinterpret_cast<cudaStream_t*>(_execStream));
if (res != 0)
throw cuda_exception::build("Failed to create default CUDA stream with launch context", res);
res = cudaStreamCreate(reinterpret_cast<cudaStream_t*>(_specialStream));
if (res != 0)
throw cuda_exception::build("Failed to create special CUDA stream with launch context", res);
_allocated = true;
_initialized = true;
}
void* ContextBuffers::reductionBuffer() {
if (!_initialized)
initialize();
return _reductionPointer;
}
void* ContextBuffers::scalarBuffer() {
if (!_initialized)
initialize();
return _scalarPointer;
}
void* ContextBuffers::allocationBuffer() {
if (!_initialized)
initialize();
return _allocationPointer;
}
void ContextBuffers::setReductionBuffer(void* pointer) {
_reductionPointer = pointer;
}
void ContextBuffers::setScalarBuffer(void* pointer) {
_scalarPointer = pointer;
}
void ContextBuffers::setAllocationBuffer(void* pointer) {
_allocationPointer = pointer;
}
void ContextBuffers::triggerOwnership(bool isOwner) {
_allocated = isOwner;
}
int ContextBuffers::deviceId() {
return _deviceId;
}
void* ContextBuffers::execStream() {
if (!_initialized) {
//nd4j_printf("execStream not initialized\n", "");
initialize();
} else {
//nd4j_printf("execStream is initialized\n", "");
}
return _execStream;
}
void* ContextBuffers::specialStream() {
if (!_initialized) {
//nd4j_printf("specialStream not initialized\n", "");
initialize();
} else {
//nd4j_printf("specialStream is initialized\n", "");
}
return _specialStream;
}
bool ContextBuffers::isInitialized() {
return _initialized;
}
sd::ErrorReference* ContextBuffers::errorReference() {
return &_errorReference;
}
}