devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merged in benoitsteiner/opencl (pull request PR-253)

Browse Source 
OpenCL improvements
This commit is contained in:
Benoit Steiner
2016-11-19 04:44:43 +00:00
parent 8649e16c2a a357fe1fb9
commit 1bdf1b9ce0
18 changed files with 780 additions and 355 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
unsupported/Eigen/CXX11/Tensor
View File

@@ -13,7 +13,7 @@
#include "../../../Eigen/Core"
#ifdef EIGEN_USE_SYCL
#if defined(EIGEN_USE_SYCL)
#undef min
#undef max
#undef isnan

296
unsupported/Eigen/CXX11/src/Tensor/TensorDeviceSycl.h
View File

@@ -16,27 +16,33 @@
#define EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H
namespace Eigen {
struct SyclDevice {
/// class members:
/// sycl queue
mutable cl::sycl::queue m_queue;
#define ConvertToActualTypeSycl(T, buf_acc) reinterpret_cast<typename cl::sycl::global_ptr<T>::pointer_t>((&(*buf_acc.get_pointer())))
struct QueueInterface {
/// class members:
bool exception_caught_ = false;
/// std::map is the container used to make sure that we create only one buffer
/// per pointer. The lifespan of the buffer now depends on the lifespan of SyclDevice.
/// If a non-read-only pointer is needed to be accessed on the host we should manually deallocate it.
mutable std::map<const void *, std::shared_ptr<void>> buffer_map;
mutable std::map<const uint8_t *, cl::sycl::buffer<uint8_t, 1>> buffer_map;
/// sycl queue
mutable cl::sycl::queue m_queue;
/// creating device by using selector
template<typename dev_Selector> explicit SyclDevice(dev_Selector s):
/// SyclStreamDevice is not owned. it is the caller's responsibility to destroy it.
template<typename dev_Selector> explicit QueueInterface(dev_Selector s):
#ifdef EIGEN_EXCEPTIONS
m_queue(cl::sycl::queue(s, [=](cl::sycl::exception_list l) {
m_queue(cl::sycl::queue(s, [&](cl::sycl::exception_list l) {
for (const auto& e : l) {
try {
std::rethrow_exception(e);
} catch (cl::sycl::exception e) {
std::cout << e.what() << std::endl;
if (e) {
exception_caught_ = true;
std::rethrow_exception(e);
}
} catch (cl::sycl::exception e) {
std::cerr << e.what() << std::endl;
}
}
}))
#else
@@ -44,63 +50,119 @@ struct SyclDevice {
#endif
{}
// destructor
~SyclDevice() { deallocate_all(); }
/// creating device by using selector
/// SyclStreamDevice is not owned. it is the caller's responsibility to destroy it.
explicit QueueInterface(cl::sycl::device d):
#ifdef EIGEN_EXCEPTIONS
m_queue(cl::sycl::queue(d, [&](cl::sycl::exception_list l) {
for (const auto& e : l) {
try {
if (e) {
exception_caught_ = true;
std::rethrow_exception(e);
}
} catch (cl::sycl::exception e) {
std::cerr << e.what() << std::endl;
}
}
}))
#else
m_queue(cl::sycl::queue(d))
#endif
{}
/// Allocating device pointer. This pointer is actually an 8 bytes host pointer used as key to access the sycl device buffer.
/// The reason is that we cannot use device buffer as a pointer as a m_data in Eigen leafNode expressions. So we create a key
/// pointer to be used in Eigen expression construction. When we convert the Eigen construction into the sycl construction we
/// use this pointer as a key in our buffer_map and we make sure that we dedicate only one buffer only for this pointer.
/// The device pointer would be deleted by calling deallocate function.
EIGEN_STRONG_INLINE void* allocate(size_t num_bytes) const {
auto buf = cl::sycl::buffer<uint8_t,1>(cl::sycl::range<1>(num_bytes));
auto ptr =buf.get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::host_buffer>().get_pointer();
buf.set_final_data(nullptr);
buffer_map.insert(std::pair<const uint8_t *, cl::sycl::buffer<uint8_t, 1>>(ptr,buf));
return static_cast<void*>(ptr);
}
/// This is used to deallocate the device pointer. p is used as a key inside
/// the map to find the device buffer and delete it.
template <typename T> EIGEN_STRONG_INLINE void deallocate(T *p) const {
auto it = buffer_map.find(p);
EIGEN_STRONG_INLINE void deallocate(const void *p) const {
auto it = buffer_map.find(static_cast<const uint8_t*>(p));
if (it != buffer_map.end()) {
buffer_map.erase(it);
internal::aligned_free(p);
}
}
/// This is called by the SyclDevice destructor to release all allocated memory if the user didn't already do so.
/// We also free the host pointer that we have dedicated as a key to accessing the device buffer.
EIGEN_STRONG_INLINE void deallocate_all() const {
std::map<const void *, std::shared_ptr<void>>::iterator it=buffer_map.begin();
while (it!=buffer_map.end()) {
auto p=it->first;
buffer_map.erase(it);
internal::aligned_free(const_cast<void*>(p));
it=buffer_map.begin();
EIGEN_STRONG_INLINE std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1>>::iterator find_buffer(const void* ptr) const {
auto it1 = buffer_map.find(static_cast<const uint8_t*>(ptr));
if (it1 != buffer_map.end()){
return it1;
}
buffer_map.clear();
else{
for(std::map<const uint8_t *, cl::sycl::buffer<uint8_t,1>>::iterator it=buffer_map.begin(); it!=buffer_map.end(); ++it){
auto size = it->second.get_size();
if((it->first < (static_cast<const uint8_t*>(ptr))) && ((static_cast<const uint8_t*>(ptr)) < (it->first + size)) ) return it;
}
}
//eigen_assert("No sycl buffer found. Make sure that you have allocated memory for your buffer by calling allocate function in SyclDevice");
std::cerr << "No sycl buffer found. Make sure that you have allocated memory for your buffer by calling allocate function in SyclDevice"<< std::endl;
abort();
//return buffer_map.end();
}
// This function checks if the runtime recorded an error for the
// underlying stream device.
EIGEN_STRONG_INLINE bool ok() const {
return !exception_caught_;
}
// destructor
~QueueInterface() { buffer_map.clear(); }
};
template <typename T> class MemCopyFunctor {
public:
typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer> read_accessor;
typedef cl::sycl::accessor<uint8_t, 1, cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer> write_accessor;
MemCopyFunctor(read_accessor src_acc, write_accessor dst_acc, size_t rng, size_t i, size_t offset): m_src_acc(src_acc), m_dst_acc(dst_acc), m_rng(rng), m_i(i), m_offset(offset) {}
void operator()(cl::sycl::nd_item<1> itemID) {
auto src_ptr = ConvertToActualTypeSycl(T, m_src_acc);
auto dst_ptr = ConvertToActualTypeSycl(T, m_dst_acc);
auto globalid = itemID.get_global_linear_id();
if (globalid < m_rng) {
dst_ptr[globalid + m_i] = src_ptr[globalid + m_offset];
}
}
private:
read_accessor m_src_acc;
write_accessor m_dst_acc;
size_t m_rng;
size_t m_i;
size_t m_offset;
};
struct SyclDevice {
// class member.
QueueInterface* m_queue_stream;
/// QueueInterface is not owned. it is the caller's responsibility to destroy it.
explicit SyclDevice(QueueInterface* queue_stream) : m_queue_stream(queue_stream){}
/// Creation of sycl accessor for a buffer. This function first tries to find
/// the buffer in the buffer_map. If found it gets the accessor from it, if not,
/// the function then adds an entry by creating a sycl buffer for that particular pointer.
template <cl::sycl::access::mode AcMd, typename T> EIGEN_STRONG_INLINE cl::sycl::accessor<T, 1, AcMd, cl::sycl::access::target::global_buffer>
get_sycl_accessor(size_t num_bytes, cl::sycl::handler &cgh, const T * ptr) const {
return (get_sycl_buffer<T>(num_bytes, ptr)->template get_access<AcMd, cl::sycl::access::target::global_buffer>(cgh));
}
/// Inserting a new sycl buffer. For every allocated device pointer only one buffer would be created. The buffer type is a device- only buffer.
/// The key pointer used to access the device buffer(the device pointer(ptr) ) must be initialised by the allocate function.
template<typename T> EIGEN_STRONG_INLINE std::pair<std::map<const void *, std::shared_ptr<void>>::iterator,bool> add_sycl_buffer(size_t num_bytes, const T *ptr) const {
using Type = cl::sycl::buffer<T, 1>;
std::pair<std::map<const void *, std::shared_ptr<void>>::iterator,bool> ret;
if(ptr!=nullptr){
ret= buffer_map.insert(std::pair<const void *, std::shared_ptr<void>>(ptr, std::shared_ptr<void>(new Type(cl::sycl::range<1>(num_bytes)),
[](void *dataMem) { delete static_cast<Type*>(dataMem); })));
(static_cast<Type*>(ret.first->second.get()))->set_final_data(nullptr);
} else {
eigen_assert("The device memory is not allocated. Please call allocate on the device!!");
}
return ret;
template <cl::sycl::access::mode AcMd> EIGEN_STRONG_INLINE cl::sycl::accessor<uint8_t, 1, AcMd, cl::sycl::access::target::global_buffer>
get_sycl_accessor(size_t num_bytes, cl::sycl::handler &cgh, const void* ptr) const {
return (get_sycl_buffer(num_bytes, ptr).template get_access<AcMd, cl::sycl::access::target::global_buffer>(cgh));
}
/// Accessing the created sycl device buffer for the device pointer
template <typename T> EIGEN_STRONG_INLINE cl::sycl::buffer<T, 1>* get_sycl_buffer(size_t num_bytes,const T * ptr) const {
return static_cast<cl::sycl::buffer<T, 1>*>(add_sycl_buffer(num_bytes, ptr).first->second.get());
EIGEN_STRONG_INLINE cl::sycl::buffer<uint8_t, 1>& get_sycl_buffer(size_t , const void * ptr) const {
return m_queue_stream->find_buffer(ptr)->second;
}
/// This is used to prepare the number of threads and also the number of threads per block for sycl kernels
EIGEN_STRONG_INLINE void parallel_for_setup(size_t n, size_t &tileSize, size_t &rng, size_t &GRange) const {
tileSize =m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>()/2;
tileSize =sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_work_group_size>()/2;
rng = n;
if (rng==0) rng=1;
GRange=rng;
@@ -110,58 +172,35 @@ struct SyclDevice {
if (xMode != 0) GRange += (tileSize - xMode);
}
}
/// Allocating device pointer. This pointer is actually an 8 bytes host pointer used as key to access the sycl device buffer.
/// The reason is that we cannot use device buffer as a pointer as a m_data in Eigen leafNode expressions. So we create a key
/// pointer to be used in Eigen expression construction. When we convert the Eigen construction into the sycl construction we
/// use this pointer as a key in our buffer_map and we make sure that we dedicate only one buffer only for this pointer.
/// The device pointer would be deleted by calling deallocate function.
EIGEN_STRONG_INLINE void *allocate(size_t) const {
return internal::aligned_malloc(8);
/// allocate device memory
EIGEN_STRONG_INLINE void *allocate(size_t num_bytes) const {
return m_queue_stream->allocate(num_bytes);
}
/// deallocate device memory
EIGEN_STRONG_INLINE void deallocate(const void *p) const {
m_queue_stream->deallocate(p);
}
// some runtime conditions that can be applied here
EIGEN_STRONG_INLINE bool isDeviceSuitable() const { return true; }
template <typename T> EIGEN_STRONG_INLINE std::map<const void *, std::shared_ptr<void>>::iterator find_nearest(const T* ptr) const {
auto it1 = buffer_map.find(ptr);
if (it1 != buffer_map.end()){
return it1;
}
else{
for(std::map<const void *, std::shared_ptr<void>>::iterator it=buffer_map.begin(); it!=buffer_map.end(); ++it){
auto size = ((cl::sycl::buffer<T, 1>*)it->second.get())->get_size();
if((static_cast<const T*>(it->first) < ptr) && (ptr < (static_cast<const T*>(it->first)) + size)) return it;
}
}
return buffer_map.end();
}
/// the memcpy function
template<typename T> EIGEN_STRONG_INLINE void memcpy(void *dst, const T *src, size_t n) const {
auto it1 = find_nearest(src);
auto it2 = find_nearest(static_cast<T*>(dst));
if ((it1 != buffer_map.end()) && (it2!=buffer_map.end())) {
auto offset= (src - (static_cast<const T*>(it1->first)));
auto i= ((static_cast<T*>(dst)) - const_cast<T*>((static_cast<const T*>(it2->first))));
size_t rng, GRange, tileSize;
parallel_for_setup(n/sizeof(T), tileSize, rng, GRange);
m_queue.submit([&](cl::sycl::handler &cgh) {
auto src_acc =((cl::sycl::buffer<T, 1>*)it1->second.get())-> template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
auto dst_acc =((cl::sycl::buffer<T, 1>*)it2->second.get())-> template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
typedef decltype(src_acc) DevToDev;
cgh.parallel_for<DevToDev>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), [=](cl::sycl::nd_item<1> itemID) {
auto globalid=itemID.get_global_linear_id();
if (globalid< rng) {
dst_acc[globalid+i ]=src_acc[globalid+offset];
}
});
});
m_queue.throw_asynchronous();
} else{
eigen_assert("no source or destination device memory found.");
}
//::memcpy(dst, src, n);
auto it1 = m_queue_stream->find_buffer((void*)src);
auto it2 = m_queue_stream->find_buffer(dst);
auto offset= (static_cast<const uint8_t*>(static_cast<const void*>(src))) - it1->first;
auto i= (static_cast<const uint8_t*>(dst)) - it2->first;
offset/=sizeof(T);
i/=sizeof(T);
size_t rng, GRange, tileSize;
parallel_for_setup(n/sizeof(T), tileSize, rng, GRange);
sycl_queue().submit([&](cl::sycl::handler &cgh) {
auto src_acc =it1->second.template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
auto dst_acc =it2->second.template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
cgh.parallel_for(cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<T>(src_acc, dst_acc, rng, 0, offset));
});
sycl_queue().throw_asynchronous();
}
/// The memcpyHostToDevice is used to copy the device only pointer to a host pointer. Using the device
@@ -170,8 +209,7 @@ struct SyclDevice {
/// buffer to host. Then we use the memcpy to copy the data to the host accessor. The first time that
/// this buffer is accessed, the data will be copied to the device.
template<typename T> EIGEN_STRONG_INLINE void memcpyHostToDevice(T *dst, const T *src, size_t n) const {
auto host_acc= get_sycl_buffer(n, dst)-> template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::host_buffer>();
auto host_acc= get_sycl_buffer(n, dst). template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::host_buffer>();
::memcpy(host_acc.get_pointer(), src, n);
}
/// The memcpyDeviceToHost is used to copy the data from host to device. Here, in order to avoid double copying the data. We create a sycl
@@ -180,57 +218,53 @@ struct SyclDevice {
/// buffer with map_allocator on the gpu in parallel. At the end of the function call the destination buffer would be destroyed and the data
/// would be available on the dst pointer using fast copy technique (map_allocator). In this case we can make sure that we copy the data back
/// to the cpu only once per function call.
template<typename T> EIGEN_STRONG_INLINE void memcpyDeviceToHost(T *dst, const T *src, size_t n) const {
auto it = find_nearest(src);
auto offset = src- (static_cast<const T*>(it->first));
if (it != buffer_map.end()) {
template<typename T> EIGEN_STRONG_INLINE void memcpyDeviceToHost(void *dst, const T *src, size_t n) const {
auto it = m_queue_stream->find_buffer(src);
auto offset =static_cast<const uint8_t*>(static_cast<const void*>(src))- it->first;
offset/=sizeof(T);
size_t rng, GRange, tileSize;
parallel_for_setup(n/sizeof(T), tileSize, rng, GRange);
// Assuming that the dst is the start of the destination pointer
auto dest_buf = cl::sycl::buffer<T, 1, cl::sycl::map_allocator<T>>(dst, cl::sycl::range<1>(rng));
typedef decltype(dest_buf) SYCLDTOH;
m_queue.submit([&](cl::sycl::handler &cgh) {
auto src_acc= (static_cast<cl::sycl::buffer<T, 1>*>(it->second.get()))-> template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
auto dest_buf = cl::sycl::buffer<uint8_t, 1, cl::sycl::map_allocator<uint8_t> >(static_cast<uint8_t*>(dst), cl::sycl::range<1>(rng*sizeof(T)));
sycl_queue().submit([&](cl::sycl::handler &cgh) {
auto src_acc= it->second.template get_access<cl::sycl::access::mode::read, cl::sycl::access::target::global_buffer>(cgh);
auto dst_acc =dest_buf.template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
cgh.parallel_for<SYCLDTOH>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), [=](cl::sycl::nd_item<1> itemID) {
cgh.parallel_for( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), MemCopyFunctor<T>(src_acc, dst_acc, rng, 0, offset));
});
sycl_queue().throw_asynchronous();
}
/// returning the sycl queue
EIGEN_STRONG_INLINE cl::sycl::queue& sycl_queue() const { return m_queue_stream->m_queue;}
/// Here is the implementation of memset function on sycl.
template<typename T> EIGEN_STRONG_INLINE void memset(T *buff, int c, size_t n) const {
size_t rng, GRange, tileSize;
parallel_for_setup(n/sizeof(T), tileSize, rng, GRange);
sycl_queue().submit([&](cl::sycl::handler &cgh) {
auto buf_acc =get_sycl_buffer(n, static_cast<uint8_t*>(static_cast<void*>(buff))). template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
cgh.parallel_for<SyclDevice>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), [=](cl::sycl::nd_item<1> itemID) {
auto globalid=itemID.get_global_linear_id();
if (globalid< dst_acc.get_size()) {
dst_acc[globalid] = src_acc[globalid + offset];
if (globalid< buf_acc.get_size()) {
for(size_t i=0; i<sizeof(T); i++)
buf_acc[globalid*sizeof(T) + i] = c;
}
});
});
m_queue.throw_asynchronous();
} else{
eigen_assert("no device memory found. The memory might be destroyed before creation");
}
}
/// Here is the implementation of memset function on sycl.
template<typename T> EIGEN_STRONG_INLINE void memset(T *buff, int c, size_t n) const {
size_t rng, GRange, tileSize;
parallel_for_setup(n/sizeof(T), tileSize, rng, GRange);
m_queue.submit([&](cl::sycl::handler &cgh) {
auto buf_acc =get_sycl_buffer(n, buff)-> template get_access<cl::sycl::access::mode::discard_write, cl::sycl::access::target::global_buffer>(cgh);
cgh.parallel_for<SyclDevice>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), [=](cl::sycl::nd_item<1> itemID) {
auto globalid=itemID.get_global_linear_id();
auto buf_ptr= reinterpret_cast<typename cl::sycl::global_ptr<unsigned char>::pointer_t>((&(*buf_acc.get_pointer())));
if (globalid< buf_acc.get_size()) {
for(size_t i=0; i<sizeof(T); i++)
buf_ptr[globalid*sizeof(T) + i] = c;
}
});
});
m_queue.throw_asynchronous();
sycl_queue().throw_asynchronous();
}
/// No need for sycl it should act the same as CPU version
EIGEN_STRONG_INLINE int majorDeviceVersion() const {
return 1;
EIGEN_STRONG_INLINE int majorDeviceVersion() const { return 1; }
/// There is no need to synchronise the buffer in sycl as it is automatically handled by sycl runtime scheduler.
EIGEN_STRONG_INLINE void synchronize() const {
sycl_queue().wait_and_throw();
}
// This function checks if the runtime recorded an error for the
// underlying stream device.
EIGEN_STRONG_INLINE bool ok() const {
return m_queue_stream->ok();
}
/// There is no need to synchronise the stream in sycl as it is automatically handled by sycl runtime scheduler.
EIGEN_STRONG_INLINE void synchronize() const {}
};
} // end namespace Eigen
#endif // EIGEN_CXX11_TENSOR_TENSOR_DEVICE_SYCL_H

40
unsupported/Eigen/CXX11/src/Tensor/TensorReductionSycl.h
View File

@@ -27,7 +27,7 @@ namespace internal {
template<typename CoeffReturnType, typename KernelName> struct syclGenericBufferReducer{
template<typename BufferTOut, typename BufferTIn>
static void run(BufferTOut* bufOut, BufferTIn& bufI, const Eigen::SyclDevice& dev, size_t length, size_t local){
static void run(BufferTOut& bufOut, BufferTIn& bufI, const Eigen::SyclDevice& dev, size_t length, size_t local){
do {
auto f = [length, local, bufOut, &bufI](cl::sycl::handler& h) mutable {
cl::sycl::nd_range<1> r{cl::sycl::range<1>{std::max(length, local)},
@@ -37,7 +37,7 @@ static void run(BufferTOut* bufOut, BufferTIn& bufI, const Eigen::SyclDevice& de
auto aI =
bufI.template get_access<cl::sycl::access::mode::read_write>(h);
auto aOut =
bufOut->template get_access<cl::sycl::access::mode::discard_write>(h);
bufOut.template get_access<cl::sycl::access::mode::discard_write>(h);
cl::sycl::accessor<CoeffReturnType, 1, cl::sycl::access::mode::read_write,
cl::sycl::access::target::local>
scratch(cl::sycl::range<1>(local), h);
@@ -61,7 +61,7 @@ static void run(BufferTOut* bufOut, BufferTIn& bufI, const Eigen::SyclDevice& de
/* Apply the reduction operation between the current local
* id and the one on the other half of the vector. */
if (globalid < length) {
int min = (length < local) ? length : local;
auto min = (length < local) ? length : local;
for (size_t offset = min / 2; offset > 0; offset /= 2) {
if (localid < offset) {
scratch[localid] += scratch[localid + offset];
@@ -72,14 +72,15 @@ static void run(BufferTOut* bufOut, BufferTIn& bufI, const Eigen::SyclDevice& de
if (localid == 0) {
aI[id.get_group(0)] = scratch[localid];
if((length<=local) && globalid ==0){
aOut[globalid]=scratch[localid];
auto aOutPtr = ConvertToActualTypeSycl(CoeffReturnType, aOut);
aOutPtr[0]=scratch[0];
}
}
}
});
};
dev.m_queue.submit(f);
dev.m_queue.throw_asynchronous();
dev.sycl_queue().submit(f);
dev.sycl_queue().throw_asynchronous();
/* At this point, you could queue::wait_and_throw() to ensure that
* errors are caught quickly. However, this would likely impact
@@ -116,7 +117,7 @@ struct FullReducer<Self, Op, const Eigen::SyclDevice, Vectorizable> {
if(rng ==0) {
red_factor=1;
};
size_t tileSize =dev.m_queue.get_device(). template get_info<cl::sycl::info::device::max_work_group_size>()/2;
size_t tileSize =dev.sycl_queue().get_device(). template get_info<cl::sycl::info::device::max_work_group_size>()/2;
size_t GRange=std::max((size_t )1, rng);
// convert global range to power of 2 for redecution
@@ -134,7 +135,9 @@ struct FullReducer<Self, Op, const Eigen::SyclDevice, Vectorizable> {
/// if the shared memory is less than the GRange, we set shared_mem size to the TotalSize and in this case one kernel would be created for recursion to reduce all to one.
if (GRange < outTileSize) outTileSize=GRange;
// getting final out buffer at the moment the created buffer is true because there is no need for assign
auto out_buffer =dev.template get_sycl_buffer<typename Eigen::internal::remove_all<CoeffReturnType>::type>(self.dimensions().TotalSize(), output);
// auto out_buffer =dev.template get_sycl_buffer<typename Eigen::internal::remove_all<CoeffReturnType>::type>(self.dimensions().TotalSize(), output);
auto out_buffer =dev.get_sycl_buffer(self.dimensions().TotalSize(), output);
/// creating the shared memory for calculating reduction.
/// This one is used to collect all the reduced value of shared memory as we dont have global barrier on GPU. Once it is saved we can
/// recursively apply reduction on it in order to reduce the whole.
@@ -142,7 +145,7 @@ struct FullReducer<Self, Op, const Eigen::SyclDevice, Vectorizable> {
typedef typename Eigen::internal::remove_all<decltype(self.xprDims())>::type Dims;
Dims dims= self.xprDims();
Op functor = reducer;
dev.m_queue.submit([&](cl::sycl::handler &cgh) {
dev.sycl_queue().submit([&](cl::sycl::handler &cgh) {
// create a tuple of accessors from Evaluator
auto tuple_of_accessors = TensorSycl::internal::createTupleOfAccessors(cgh, self.impl());
auto tmp_global_accessor = temp_global_buffer. template get_access<cl::sycl::access::mode::read_write, cl::sycl::access::target::global_buffer>(cgh);
@@ -161,16 +164,16 @@ struct FullReducer<Self, Op, const Eigen::SyclDevice, Vectorizable> {
auto globalid=itemID.get_global_linear_id();
if(globalid<rng)
tmp_global_accessor.get_pointer()[globalid]=InnerMostDimReducer<decltype(device_self_evaluator), Op, false>::reduce(device_self_evaluator, red_factor*globalid, red_factor, const_cast<Op&>(functor));
tmp_global_accessor.get_pointer()[globalid]=InnerMostDimReducer<decltype(device_self_evaluator), Op, false>::reduce(device_self_evaluator, static_cast<typename DevExpr::Index>(red_factor*globalid), red_factor, const_cast<Op&>(functor));
else
tmp_global_accessor.get_pointer()[globalid]=static_cast<CoeffReturnType>(0);
if(remaining!=0 && globalid==0 )
// this will add the rest of input buffer when the input size is not devidable to red_factor.
tmp_global_accessor.get_pointer()[globalid]+=InnerMostDimReducer<decltype(device_self_evaluator), Op, false>::reduce(device_self_evaluator, red_factor*(rng), remaining, const_cast<Op&>(functor));
tmp_global_accessor.get_pointer()[0]+=InnerMostDimReducer<decltype(device_self_evaluator), Op, false>::reduce(device_self_evaluator, static_cast<typename DevExpr::Index>(red_factor*(rng)), static_cast<typename DevExpr::Index>(remaining), const_cast<Op&>(functor));
});
});
dev.m_queue.throw_asynchronous();
dev.sycl_queue().throw_asynchronous();
/// This is used to recursively reduce the tmp value to an element of 1;
syclGenericBufferReducer<CoeffReturnType,HostExpr>::run(out_buffer, temp_global_buffer,dev, GRange, outTileSize);
@@ -198,7 +201,7 @@ struct InnerReducer<Self, Op, const Eigen::SyclDevice> {
Dims dims= self.xprDims();
Op functor = reducer;
dev.m_queue.submit([&](cl::sycl::handler &cgh) {
dev.sycl_queue().submit([&](cl::sycl::handler &cgh) {
// create a tuple of accessors from Evaluator
auto tuple_of_accessors = TensorSycl::internal::createTupleOfAccessors(cgh, self.impl());
auto output_accessor = dev.template get_sycl_accessor<cl::sycl::access::mode::discard_write>(num_coeffs_to_preserve,cgh, output);
@@ -212,19 +215,20 @@ struct InnerReducer<Self, Op, const Eigen::SyclDevice> {
const auto device_self_expr= TensorReductionOp<Op, Dims, decltype(device_expr.expr) ,MakeGlobalPointer>(device_expr.expr, dims, functor);
/// This is the evaluator for device_self_expr. This is exactly similar to the self which has been passed to run function. The difference is
/// the device_evaluator is detectable and recognisable on the device.
typedef Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice> DeiceSelf;
typedef Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice> DeviceSelf;
auto device_self_evaluator = Eigen::TensorEvaluator<decltype(device_self_expr), Eigen::DefaultDevice>(device_self_expr, Eigen::DefaultDevice());
auto output_accessor_ptr =ConvertToActualTypeSycl(typename DeviceSelf::CoeffReturnType, output_accessor);
/// const cast added as a naive solution to solve the qualifier drop error
auto globalid=itemID.get_global_linear_id();
if (globalid< range) {
typename DeiceSelf::CoeffReturnType accum = functor.initialize();
GenericDimReducer<DeiceSelf::NumReducedDims-1, DeiceSelf, Op>::reduce(device_self_evaluator, device_self_evaluator.firstInput(globalid),const_cast<Op&>(functor), &accum);
typename DeviceSelf::CoeffReturnType accum = functor.initialize();
GenericDimReducer<DeviceSelf::NumReducedDims-1, DeviceSelf, Op>::reduce(device_self_evaluator, device_self_evaluator.firstInput(static_cast<typename DevExpr::Index>(globalid)),const_cast<Op&>(functor), &accum);
functor.finalize(accum);
output_accessor.get_pointer()[globalid]= accum;
output_accessor_ptr[globalid]= accum;
}
});
});
dev.m_queue.throw_asynchronous();
dev.sycl_queue().throw_asynchronous();
return false;
}
};

19
unsupported/Eigen/CXX11/src/Tensor/TensorSyclExprConstructor.h
View File

@@ -30,7 +30,8 @@ namespace internal {
template <typename PtrType, size_t N, typename... Params>
struct EvalToLHSConstructor {
PtrType expr;
EvalToLHSConstructor(const utility::tuple::Tuple<Params...> &t): expr((&(*(utility::tuple::get<N>(t).get_pointer())))) {}
EvalToLHSConstructor(const utility::tuple::Tuple<Params...> &t) : expr(ConvertToActualTypeSycl(typename Eigen::internal::remove_all<PtrType>::type, utility::tuple::get<N>(t))) {}
//EvalToLHSConstructor(const utility::tuple::Tuple<Params...> &t): expr((&(*(utility::tuple::get<N>(t).get_pointer())))) {}
};
/// \struct ExprConstructor is used to reconstruct the expression on the device and
@@ -53,9 +54,11 @@ CVQual PlaceHolder<CVQual TensorMap<T, Options3_, MakePointer_>, N>, Params...>{
Type expr;\
template <typename FuncDetector>\
ExprConstructor(FuncDetector &fd, const utility::tuple::Tuple<Params...> &t)\
: expr(Type((&(*(utility::tuple::get<N>(t).get_pointer()))), fd.dimensions())) {}\
: expr(Type(ConvertToActualTypeSycl(typename Type::Scalar, utility::tuple::get<N>(t)), fd.dimensions())){}\
};
//: expr(Type((&(*(utility::tuple::get<N>(t).get_pointer()))), fd.dimensions())) {}
TENSORMAP(const)
TENSORMAP()
@@ -163,7 +166,7 @@ struct ExprConstructor<CVQual TensorAssignOp<OrigLHSExpr, OrigRHSExpr>, CVQual
ASSIGN()
#undef ASSIGN
/// specialisation of the \ref ExprConstructor struct when the node type is
/// TensorEvalToOp
/// TensorEvalToOp /// 0 here is the output number in the buffer
#define EVALTO(CVQual)\
template <typename OrigExpr, typename Expr, typename... Params>\
struct ExprConstructor<CVQual TensorEvalToOp<OrigExpr, MakeGlobalPointer>, CVQual TensorEvalToOp<Expr>, Params...> {\
@@ -189,12 +192,13 @@ template <typename OrigExpr, typename DevExpr, size_t N, typename... Params>\
struct ExprConstructor<CVQual TensorForcedEvalOp<OrigExpr, MakeGlobalPointer>,\
CVQual PlaceHolder<CVQual TensorForcedEvalOp<DevExpr>, N>, Params...> {\
typedef CVQual TensorMap<Tensor<typename TensorForcedEvalOp<DevExpr, MakeGlobalPointer>::Scalar,\
TensorForcedEvalOp<DevExpr, MakeGlobalPointer>::NumDimensions, 0, typename TensorForcedEvalOp<DevExpr>::Index>, 0, MakeGlobalPointer> Type;\
TensorForcedEvalOp<DevExpr, MakeGlobalPointer>::NumDimensions, Eigen::internal::traits<TensorForcedEvalOp<DevExpr, MakeGlobalPointer>>::Layout, typename TensorForcedEvalOp<DevExpr>::Index>, Eigen::internal::traits<TensorForcedEvalOp<DevExpr, MakeGlobalPointer>>::Layout, MakeGlobalPointer> Type;\
Type expr;\
template <typename FuncDetector>\
ExprConstructor(FuncDetector &fd, const utility::tuple::Tuple<Params...> &t)\
: expr(Type((&(*(utility::tuple::get<N>(t).get_pointer()))), fd.dimensions())) {}\
: expr(Type(ConvertToActualTypeSycl(typename Type::Scalar, utility::tuple::get<N>(t)), fd.dimensions())) {}\
};
//: expr(Type((&(*(utility::tuple::get<N>(t).get_pointer()))), fd.dimensions())) {}
FORCEDEVAL(const)
FORCEDEVAL()
@@ -214,12 +218,13 @@ struct ExprConstructor<CVQual TensorReductionOp<OP, Dim, OrigExpr, MakeGlobalPoi
CVQual PlaceHolder<CVQual TensorReductionOp<OP, Dim, DevExpr>, N>, Params...> {\
static const size_t NumIndices= ValueCondition< TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>::NumDimensions==0, 1, TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>::NumDimensions >::Res;\
typedef CVQual TensorMap<Tensor<typename TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>::Scalar,\
NumIndices, 0, typename TensorReductionOp<OP, Dim, DevExpr>::Index>, 0, MakeGlobalPointer> Type;\
NumIndices, Eigen::internal::traits<TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>>::Layout, typename TensorReductionOp<OP, Dim, DevExpr>::Index>, Eigen::internal::traits<TensorReductionOp<OP, Dim, DevExpr, MakeGlobalPointer>>::Layout, MakeGlobalPointer> Type;\
Type expr;\
template <typename FuncDetector>\
ExprConstructor(FuncDetector &fd, const utility::tuple::Tuple<Params...> &t)\
: expr(Type((&(*(utility::tuple::get<N>(t).get_pointer()))), fd.dimensions())) {}\
:expr(Type(ConvertToActualTypeSycl(typename Type::Scalar, utility::tuple::get<N>(t)), fd.dimensions())) {}\
};
//: expr(Type((&(*(utility::tuple::get<N>(t).get_pointer()))), fd.dimensions())) {}
SYCLREDUCTIONEXPR(const)
SYCLREDUCTIONEXPR()

5
unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractAccessor.h
View File

@@ -57,9 +57,8 @@ struct AccessorConstructor{
return utility::tuple::append(ExtractAccessor<Arg1>::getTuple(cgh, eval1),utility::tuple::append(ExtractAccessor<Arg2>::getTuple(cgh, eval2), ExtractAccessor<Arg3>::getTuple(cgh, eval3)));
}
template< cl::sycl::access::mode AcM, typename Arg> static inline auto getAccessor(cl::sycl::handler& cgh, Arg eval)
-> decltype(utility::tuple::make_tuple( eval.device().template get_sycl_accessor<AcM,
typename Eigen::internal::remove_all<typename Arg::CoeffReturnType>::type>(eval.dimensions().TotalSize(), cgh,eval.data()))){
return utility::tuple::make_tuple(eval.device().template get_sycl_accessor<AcM, typename Eigen::internal::remove_all<typename Arg::CoeffReturnType>::type>(eval.dimensions().TotalSize(), cgh,eval.data()));
-> decltype(utility::tuple::make_tuple( eval.device().template get_sycl_accessor<AcM>(eval.dimensions().TotalSize(), cgh,eval.data()))){
return utility::tuple::make_tuple(eval.device().template get_sycl_accessor<AcM>(eval.dimensions().TotalSize(), cgh,eval.data()));
}
};

2
unsupported/Eigen/CXX11/src/Tensor/TensorSyclExtractFunctors.h
View File

@@ -148,7 +148,7 @@ template<typename InDim>
template<typename Dim> struct DimConstr<Dim, 0> {
template<typename InDim>
static inline Dim getDim(InDim dims ) {return Dim(dims.TotalSize());}
static inline Dim getDim(InDim dims ) {return Dim(static_cast<Dim>(dims.TotalSize()));}
};
template<typename Op, typename Dims, typename ArgType, template <class> class MakePointer_, typename Device>

8
unsupported/Eigen/CXX11/src/Tensor/TensorSyclRun.h
View File

@@ -37,11 +37,11 @@ void run(Expr &expr, Dev &dev) {
typedef typename internal::createPlaceHolderExpression<Expr>::Type PlaceHolderExpr;
auto functors = internal::extractFunctors(evaluator);
dev.m_queue.submit([&](cl::sycl::handler &cgh) {
dev.sycl_queue().submit([&](cl::sycl::handler &cgh) {
// create a tuple of accessors from Evaluator
auto tuple_of_accessors = internal::createTupleOfAccessors<decltype(evaluator)>(cgh, evaluator);
size_t range, GRange, tileSize;
dev.parallel_for_setup(utility::tuple::get<0>(tuple_of_accessors).get_range()[0], tileSize, range, GRange);
dev.parallel_for_setup(utility::tuple::get<0>(tuple_of_accessors).get_range()[0]/sizeof(typename Expr::Scalar), tileSize, range, GRange);
// run the kernel
cgh.parallel_for<PlaceHolderExpr>( cl::sycl::nd_range<1>(cl::sycl::range<1>(GRange), cl::sycl::range<1>(tileSize)), [=](cl::sycl::nd_item<1> itemID) {
@@ -49,11 +49,11 @@ void run(Expr &expr, Dev &dev) {
auto device_expr =internal::createDeviceExpression<DevExpr, PlaceHolderExpr>(functors, tuple_of_accessors);
auto device_evaluator = Eigen::TensorEvaluator<decltype(device_expr.expr), Eigen::DefaultDevice>(device_expr.expr, Eigen::DefaultDevice());
if (itemID.get_global_linear_id() < range) {
device_evaluator.evalScalar(static_cast<int>(itemID.get_global_linear_id()));
device_evaluator.evalScalar(static_cast<typename DevExpr::Index>(itemID.get_global_linear_id()));
}
});
});
dev.m_queue.throw_asynchronous();
dev.sycl_queue().throw_asynchronous();
}
evaluator.cleanup();