SparseArray<T> sparseConvertDenseToStorage(const Array<T> &in_)
{
in_.eval();
// MKL only has dns->csr.
// CSR <-> CSC is only supported if input is square
uint nNZ = reduce_all<af_notzero_t, T, uint>(in_);
SparseArray<T> sparse_ = createEmptySparseArray<T>(in_.dims(), nNZ, AF_STORAGE_CSR);
sparse_.eval();
auto func = [=] (SparseArray<T> sparse, const Array<T> in) {
// Read: https://software.intel.com/en-us/node/520848
// But job description is incorrect with regards to job[1]
// 0 implies row major and 1 implies column major
int j1 = 1, j2 = 0;
const int job[] = {0, j1, j2, 2, (int)sparse.elements(), 1};
const int M = in.dims()[0];
const int N = in.dims()[1];
int ldd = in.strides()[1];
int info = 0;
// Have to mess up all const correctness because MKL dnscsr function
// is bidirectional and has input/output on all pointers
Array<T > &values = sparse.getValues();
Array<int> &rowIdx = sparse.getRowIdx();
Array<int> &colIdx = sparse.getColIdx();
dnscsr_func<T>()(
job, &M, &N,
reinterpret_cast<ptr_type<T>>(const_cast<T*>(in.get())), &ldd,
reinterpret_cast<ptr_type<T>>(values.get()),
colIdx.get(),
rowIdx.get(),
&info);
};
getQueue().enqueue(func, sparse_, in_);
if(stype == AF_STORAGE_CSR)
return sparse_;
else
AF_ERROR("CPU Backend only supports Dense to CSR or COO", AF_ERR_NOT_SUPPORTED);
return sparse_;
}
Array<T> sparseConvertCOOToDense(const SparseArray<T> &in)
{
in.eval();
Array<T> dense = createValueArray<T>(in.dims(), scalar<T>(0));
dense.eval();
const Array<T> values = in.getValues();
const Array<int> rowIdx = in.getRowIdx();
const Array<int> colIdx = in.getColIdx();
getQueue().enqueue(kernel::coo2dense<T>, dense, values, rowIdx, colIdx);
return dense;
}
static void bcast_dim_launcher(Param &out,
Param &tmp,
const uint groups_all[4])
{
Kernel ker = get_scan_dim_kernels<Ti, To, op, dim, isFinalPass, threads_y>(1);
NDRange local(THREADS_X, threads_y);
NDRange global(groups_all[0] * groups_all[2] * local[0],
groups_all[1] * groups_all[3] * local[1]);
uint lim = divup(out.info.dims[dim], (threads_y * groups_all[dim]));
auto bcastOp = make_kernel<Buffer, KParam,
Buffer, KParam,
uint, uint,
uint, uint>(ker);
bcastOp(EnqueueArgs(getQueue(), global, local),
out.data, out.info, tmp.data, tmp.info,
groups_all[0], groups_all[1], groups_all[dim], lim);
CL_DEBUG_FINISH(getQueue());
}
Array<T>::Array(af::dim4 dims, af::dim4 strides, dim_t offset_,
const T * const in_data, bool is_device) :
info(getActiveDeviceId(), dims, offset_, strides, (af_dtype)dtype_traits<T>::af_type),
data(is_device ? (T*)in_data : memAlloc<T>(info.total()).release(), memFree<T>),
data_dims(dims),
node(bufferNodePtr<T>()),
ready(true),
owner(true)
{
if (!is_device) {
// Ensure the memory being written to isnt used anywhere else.
getQueue().sync();
copy(in_data, in_data + info.total(), data.get());
}
}
Array<T> rotate(const Array<T> &in, const float theta, const af::dim4 &odims,
const af_interp_type method)
{
in.eval();
Array<T> out = createEmptyArray<T>(odims);
switch(method) {
case AF_INTERP_NEAREST:
getQueue().enqueue(kernel::rotate<T, AF_INTERP_NEAREST>, out, in, theta);
break;
case AF_INTERP_BILINEAR:
getQueue().enqueue(kernel::rotate<T, AF_INTERP_BILINEAR>, out, in, theta);
break;
case AF_INTERP_LOWER:
getQueue().enqueue(kernel::rotate<T, AF_INTERP_LOWER>, out, in, theta);
break;
default:
AF_ERROR("Unsupported interpolation type", AF_ERR_ARG);
break;
}
return out;
}
void sort0(Param val)
{
try {
compute::command_queue c_queue(getQueue()());
compute::buffer val_buf((*val.data)());
for(int w = 0; w < val.info.dims[3]; w++) {
int valW = w * val.info.strides[3];
for(int z = 0; z < val.info.dims[2]; z++) {
int valWZ = valW + z * val.info.strides[2];
for(int y = 0; y < val.info.dims[1]; y++) {
int valOffset = valWZ + y * val.info.strides[1];
if(isAscending) {
compute::stable_sort(
compute::make_buffer_iterator<T>(val_buf, valOffset),
compute::make_buffer_iterator<T>(val_buf, valOffset + val.info.dims[0]),
compute::less<T>(), c_queue);
} else {
compute::stable_sort(
compute::make_buffer_iterator<T>(val_buf, valOffset),
compute::make_buffer_iterator<T>(val_buf, valOffset + val.info.dims[0]),
compute::greater<T>(), c_queue);
}
}
}
}
CL_DEBUG_FINISH(getQueue());
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
void convNHelper(const conv_kparam_t& param, Param& out, const Param& signal, const Param& filter)
{
std::string ref_name = std::string("convolveND_") +
std::string(dtype_traits<T>::getName()) + std::string(dtype_traits<aT>::getName()) +
std::to_string(bDim) + std::to_string(expand);
int device = getActiveDeviceId();
kc_entry_t entry = kernelCache(device, ref_name);
if (entry.prog==0 && entry.ker==0) {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D Ti=" << dtype_traits<T>::getName()
<< " -D To=" << dtype_traits<aT>::getName()
<< " -D accType=" << dtype_traits<aT>::getName()
<< " -D BASE_DIM=" << bDim
<< " -D EXPAND=" << expand
<< " -D " << binOpName<af_mul_t>();
if((af_dtype) dtype_traits<T>::af_type == c32 ||
(af_dtype) dtype_traits<T>::af_type == c64) {
options << " -D CPLX=1";
} else {
options << " -D CPLX=0";
}
if (std::is_same<T, double>::value || std::is_same<T, cdouble>::value)
options << " -D USE_DOUBLE";
const char *ker_strs[] = {ops_cl, convolve_cl};
const int ker_lens[] = {ops_cl_len, convolve_cl_len};
Program prog;
buildProgram(prog, 2, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "convolve");
addKernelToCache(device, ref_name, entry);
}
auto convOp = cl::KernelFunctor<Buffer, KParam, Buffer, KParam, cl::LocalSpaceArg, Buffer, KParam,
int, int, int, int, int, int, int, int >(*entry.ker);
convOp(EnqueueArgs(getQueue(), param.global, param.local),
*out.data, out.info, *signal.data, signal.info, cl::Local(param.loc_size),
*param.impulse, filter.info, param.nBBS0, param.nBBS1,
param.o[0], param.o[1], param.o[2], param.s[0], param.s[1], param.s[2]);
}
/*
* Print Entire queue: Debugging
*/
void printQueue() {
BoardNode boardToPrint = getQueue(NULL)->start;
int row, col, queueN = 0;
while(boardToPrint != NULL) {
queueN++;
printf("Queue Item %d\n",queueN);
for(row = 0; row < MAXROW; row++) {
for(col = 0; col < MAXCOL; col++) {
printf("%d ",boardToPrint->board[row][col]);
}
pNL();
}
boardToPrint = boardToPrint->next;
}
}
void random(cl::Buffer out, dim_type elements)
{
try {
static unsigned counter;
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static Program ranProgs[DeviceManager::MAX_DEVICES];
static Kernel ranKernels[DeviceManager::MAX_DEVICES];
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
Program::Sources setSrc;
setSrc.emplace_back(random_cl, random_cl_len);
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D repeat="<< REPEAT
<< " -D " << random_name<T, isRandu>().name();
if (std::is_same<T, double>::value) {
options << " -D USE_DOUBLE";
options << " -D IS_64";
}
if (std::is_same<T, char>::value) {
options << " -D IS_BOOL";
}
buildProgram(ranProgs[device], random_cl, random_cl_len, options.str());
ranKernels[device] = Kernel(ranProgs[device], "random");
});
auto randomOp = make_kernel<cl::Buffer, uint, uint, uint, uint>(ranKernels[device]);
uint groups = divup(elements, THREADS * REPEAT);
counter += divup(elements, THREADS * groups);
NDRange local(THREADS, 1);
NDRange global(THREADS * groups, 1);
randomOp(EnqueueArgs(getQueue(), global, local),
out, elements, counter, random_seed[0], random_seed[1]);
} catch(cl::Error ex) {
CL_TO_AF_ERROR(ex);
}
}
/*
*Checks start of action Queue for command, and actions it if all criteria are met
*/
int popToTower() {
ActionQueueStructure queue = getQueue(NULL);
GameProperties Game = getGame(NULL);
int needed;
if(queue->start != NULL) {
needed = calculateCosts(queue->start->command,queue->start->option,queue->start->target);
switch(queue->start->command) {
case cmd_upgrade:
if (checkQueue(queue, Game,needed)) {
upgradeTowerStat(queue->start->option,queue->start->target);
useMemory(Game, needed);
removeQueueItem();
}
break;
case cmd_mktwr:
if (checkQueue(queue,Game,needed)) {
switch(queue->start->option) {
case mktwr_int:
createTowerTypeFromPositions(queue->start->target,INT_TYPE);
break;
case mktwr_char:
createTowerTypeFromPositions(queue->start->target,CHAR_TYPE);
break;
default:
fprintf(stderr,"Unrecognised tower type\n");
break;
}
//createTowerFromPositions(queue->start->target);
useMemory(Game, needed);
removeQueueItem();
}
break;
case cmd_aptget:
if(checkQueue(queue,Game,needed)) {
unlock_ability(KILL);
useMemory(Game, needed);
removeQueueItem();
}
default:
break;
}
} else {
return 0;
}
return 1;
}
//=============================================================================
// METHOD: SPELLipcMessageMailbox::place
//=============================================================================
bool SPELLipcMessageMailbox::place( std::string id, const SPELLipcMessage& msg )
{
DEBUG(NAME + "Place message on queue with id " + id + " (" + msg.getSequenceStr() + ")");
SPELLipcMessageQueue* queue = getQueue(id);
if(queue)
{
DEBUG(NAME + "Place message IN");
queue->push(msg);
DEBUG(NAME + "Place message OUT");
return true;
}
else
{
LOG_ERROR("###### No queue to place response " + msg.dataStr());
return false;
}
}
int cholesky_inplace(Array<T> &in, const bool is_upper)
{
if(OpenCLCPUOffload()) {
return cpu::cholesky_inplace(in, is_upper);
}
dim4 iDims = in.dims();
int N = iDims[0];
magma_uplo_t uplo = is_upper ? MagmaUpper : MagmaLower;
int info = 0;
cl::Buffer *in_buf = in.get();
magma_potrf_gpu<T>(uplo, N,
(*in_buf)(), in.getOffset(), in.strides()[1],
getQueue()(), &info);
return info;
}
void laset(int m, int n,
T offdiag, T diag,
cl_mem dA, size_t dA_offset, magma_int_t ldda)
{
std::string refName = laset_name<uplo>() + std::string("_") +
std::string(dtype_traits<T>::getName()) +
std::to_string(uplo);
int device = getActiveDeviceId();
kc_entry_t entry = kernelCache(device, refName);
if (entry.prog==0 && entry.ker==0) {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D BLK_X=" << BLK_X
<< " -D BLK_Y=" << BLK_Y
<< " -D IS_CPLX=" << af::iscplx<T>();
if (std::is_same<T, double>::value || std::is_same<T, cdouble>::value)
options << " -D USE_DOUBLE";
const char* ker_strs[] = {laset_cl};
const int ker_lens[] = {laset_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, laset_name<uplo>());
addKernelToCache(device, refName, entry);
}
int groups_x = (m - 1) / BLK_X + 1;
int groups_y = (n - 1) / BLK_Y + 1;
NDRange local(BLK_X, 1);
NDRange global(groups_x * local[0], groups_y * local[1]);
// retain the cl_mem object during cl::Buffer creation
cl::Buffer dAObj(dA, true);
auto lasetOp = KernelFunctor<int, int, T, T, Buffer, unsigned long long, int>(*entry.ker);
lasetOp(EnqueueArgs(getQueue(), global, local), m, n, offdiag, diag, dAObj, dA_offset, ldda);
}
Array<T> triangleSolve(const Array<T> &A, const Array<T> &b, const af_mat_prop options)
{
trsm_func<T> gpu_trsm;
Array<T> B = copyArray<T>(b);
int N = B.dims()[0];
int NRHS = B.dims()[1];
const cl::Buffer* A_buf = A.get();
cl::Buffer* B_buf = B.get();
cl_event event = 0;
cl_command_queue queue = getQueue()();
std::string pName = getPlatformName(getDevice());
if(pName.find("NVIDIA") != std::string::npos && (options & AF_MAT_UPPER))
{
Array<T> AT = transpose<T>(A, true);
cl::Buffer* AT_buf = AT.get();
gpu_trsm(clblasColumnMajor,
clblasLeft,
clblasLower,
clblasConjTrans,
options & AF_MAT_DIAG_UNIT ? clblasUnit : clblasNonUnit,
N, NRHS, scalar<T>(1),
(*AT_buf)(), AT.getOffset(), AT.strides()[1],
(*B_buf)(), B.getOffset(), B.strides()[1],
1, &queue, 0, nullptr, &event);
} else {
gpu_trsm(clblasColumnMajor,
clblasLeft,
options & AF_MAT_LOWER ? clblasLower : clblasUpper,
clblasNoTrans,
options & AF_MAT_DIAG_UNIT ? clblasUnit : clblasNonUnit,
N, NRHS, scalar<T>(1),
(*A_buf)(), A.getOffset(), A.strides()[1],
(*B_buf)(), B.getOffset(), B.strides()[1],
1, &queue, 0, nullptr, &event);
}
return B;
}
int32_t InnerUdtServer::sendMessage(idgs::actor::ActorMessagePtr& msg) {
int32_t memberId = msg->getDestMemberId();
if(memberId < 0) {
LOG(ERROR) << "Invalid member ID: " << memberId;
return RC_ERROR;
}
auto q = getQueue(memberId);
msg->freePbMemory();
q->push(msg);
std::shared_ptr<InnerUdtConnection> conn = getConnection(memberId);
if(conn) {
conn->sendMessage(msg);
}
return 0;
}
Array<T> triangleSolve(const Array<T> &A, const Array<T> &b, const af_mat_prop options)
{
gpu_blas_trsm_func<T> gpu_blas_trsm;
Array<T> B = copyArray<T>(b);
int N = B.dims()[0];
int NRHS = B.dims()[1];
const cl::Buffer* A_buf = A.get();
cl::Buffer* B_buf = B.get();
cl_event event = 0;
cl_command_queue queue = getQueue()();
if(getActivePlatform() == AFCL_PLATFORM_NVIDIA && (options & AF_MAT_UPPER))
{
Array<T> AT = transpose<T>(A, true);
cl::Buffer* AT_buf = AT.get();
CLBLAS_CHECK(gpu_blas_trsm(
clblasLeft,
clblasLower,
clblasConjTrans,
options & AF_MAT_DIAG_UNIT ? clblasUnit : clblasNonUnit,
N, NRHS, scalar<T>(1),
(*AT_buf)(), AT.getOffset(), AT.strides()[1],
(*B_buf)(), B.getOffset(), B.strides()[1],
1, &queue, 0, nullptr, &event));
} else {
CLBLAS_CHECK(gpu_blas_trsm(
clblasLeft,
options & AF_MAT_LOWER ? clblasLower : clblasUpper,
clblasNoTrans,
options & AF_MAT_DIAG_UNIT ? clblasUnit : clblasNonUnit,
N, NRHS, scalar<T>(1),
(*A_buf)(), A.getOffset(), A.strides()[1],
(*B_buf)(), B.getOffset(), B.strides()[1],
1, &queue, 0, nullptr, &event));
}
return B;
}
void convNHelper(const conv_kparam_t& param, Param& out, const Param& signal, const Param& filter)
{
try {
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static std::map<int, Program*> convProgs;
static std::map<int, Kernel*> convKernels;
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D accType="<< dtype_traits<aT>::getName()
<< " -D BASE_DIM="<< bDim
<< " -D EXPAND=" << expand;
if (std::is_same<T, double>::value ||
std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
Program prog;
buildProgram(prog, convolve_cl, convolve_cl_len, options.str());
convProgs[device] = new Program(prog);
convKernels[device] = new Kernel(*convProgs[device], "convolve");
});
auto convOp = cl::KernelFunctor<Buffer, KParam, Buffer, KParam,
cl::LocalSpaceArg, Buffer, KParam,
int, int,
int, int, int,
int, int, int
>(*convKernels[device]);
convOp(EnqueueArgs(getQueue(), param.global, param.local),
*out.data, out.info, *signal.data, signal.info, cl::Local(param.loc_size),
*param.impulse, filter.info, param.nBBS0, param.nBBS1,
param.o[0], param.o[1], param.o[2], param.s[0], param.s[1], param.s[2]);
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
void susan(cl::Buffer* out, const cl::Buffer* in, const unsigned in_off,
const unsigned idim0, const unsigned idim1, const float t,
const float g, const unsigned edge) {
std::string refName = std::string("susan_responses_") +
std::string(dtype_traits<T>::getName()) +
std::to_string(radius);
int device = getActiveDeviceId();
kc_entry_t entry = kernelCache(device, refName);
if (entry.prog == 0 && entry.ker == 0) {
const size_t LOCAL_MEM_SIZE =
(SUSAN_THREADS_X + 2 * radius) * (SUSAN_THREADS_Y + 2 * radius);
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D LOCAL_MEM_SIZE=" << LOCAL_MEM_SIZE
<< " -D BLOCK_X=" << SUSAN_THREADS_X
<< " -D BLOCK_Y=" << SUSAN_THREADS_Y << " -D RADIUS=" << radius
<< " -D RESPONSE";
if (std::is_same<T, double>::value || std::is_same<T, cdouble>::value)
options << " -D USE_DOUBLE";
const char* ker_strs[] = {susan_cl};
const int ker_lens[] = {susan_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "susan_responses");
addKernelToCache(device, refName, entry);
}
auto susanOp = KernelFunctor<Buffer, Buffer, unsigned, unsigned, unsigned,
float, float, unsigned>(*entry.ker);
NDRange local(SUSAN_THREADS_X, SUSAN_THREADS_Y);
NDRange global(divup(idim0 - 2 * edge, local[0]) * local[0],
divup(idim1 - 2 * edge, local[1]) * local[1]);
susanOp(EnqueueArgs(getQueue(), global, local), *out, *in, in_off, idim0,
idim1, t, g, edge);
}
int cholesky_inplace(Array<T> &in, const bool is_upper)
{
try {
initBlas();
dim4 iDims = in.dims();
int N = iDims[0];
magma_uplo_t uplo = is_upper ? MagmaUpper : MagmaLower;
int info = 0;
cl::Buffer *in_buf = in.get();
magma_potrf_gpu<T>(uplo, N,
(*in_buf)(), in.getOffset(), in.strides()[1],
getQueue()(), &info);
return info;
} catch (cl::Error &err) {
CL_TO_AF_ERROR(err);
}
}
Array<T> solveLU(const Array<T> &A, const Array<int> &pivot,
const Array<T> &b, const af_mat_prop options)
{
A.eval();
pivot.eval();
b.eval();
int N = A.dims()[0];
int NRHS = b.dims()[1];
Array< T > B = copyArray<T>(b);
auto func = [=] (Array<T> A, Array<T> B, Array<int> pivot, int N, int NRHS) {
getrs_func<T>()(AF_LAPACK_COL_MAJOR, 'N',
N, NRHS, A.get(), A.strides()[1],
pivot.get(), B.get(), B.strides()[1]);
};
getQueue().enqueue(func, A, B, pivot, N, NRHS);
return B;
}
/* Input: An empty queue, and an initialized sudoku s, an empty queue
* for the solutions, and a pointer to a variable to store the
* number of guesses
* Returns: NULL if no possible solutions exist or the queue is not
* empty, and otherwise the solution to the input puzzle.
* -----------------------------------------------------------
* Solves the puzzle using backtracking. The solver is initialized by
* putting the puzzle into the empty queue. Then, each iteration will
* pull a board out of the queue, perform a simple reduction on that
* board, and then make a guess on the cell which has the least number
* of possibilities. If verbose is set, it prints each board before
* making a guess, giving a sense of the whole solution process.
*
* If there is nothing to pull out of the queue, there are no possible
* solutions to the puzzle, and the function returns an error value of
* NULL. If the queue is not empty at initialization, the function
* prints an error message regardless of the state of the flags, and
* returns an error.
*/
sudoku solve(queue q, sudoku s, int * guesses)
{
if (!isEmptyQueue(q)) {
printf("Error: call to solve with a non-empty queue");
return NULL;
}
putQueue(q, (void *) s);
*guesses = 0;
int slvd = 0;
while (1)
{
if (getQueue(q, (void **) &s)){
return NULL;
}
reduce(s);
if(verbose) {
system("clear");
printSudoku(s, pretty);
printf("\n");
}
slvd = checkSudoku(s);
if (slvd == -1) deleteSudoku(s);
else if (slvd == 1) return s;
else
{
if(guess(q, s)) {
printf("Error: Full queue");
return NULL;
}
deleteSudoku(s);
(*guesses)++;
}
}
}
void EmitterSystem::update(Camera & camera, float frametime)
{
Vector3 near = camera.unProject(Mouse::getPosition(), 0.f);
Vector3 far = camera.unProject(Mouse::getPosition(), 1.f);
Vector3 dir = (far - near);
dir.normalize();
dir.x *= lerp(1.f, 4.f, std::abs(Mouse::getPosition().x - 0.5f) / 0.5f);
dir.y *= lerp(1.f, 4.f, std::abs(Mouse::getPosition().y - 0.5f) / 0.5f);
cl::Event event;
cl_int err = 0;
glFinish();
if (Keyboard::isKeyPressed(GLFW_KEY_F))
init();
if (Keyboard::isKeyPressed(GLFW_KEY_H))
m_disableVelocity = !m_disableVelocity;
if (m_disableVelocity)
frametime = 0.f;
std::vector<cl::Memory> buffers;
buffers.push_back(m_glBuffer[Index::Particles]);
err = getQueue().enqueueAcquireGLObjects(&buffers, NULL, NULL);
if (err != CL_SUCCESS)
std::cout << "Failed acquiring GL object : " << getError(err) << std::endl;
err = getQueue().finish();
if (err != CL_SUCCESS)
std::cout << "ERROR" << std::endl;
err = getKernel().setArg(3, frametime);
if (err != CL_SUCCESS)
std::cout << "ERROR kernel args" << std::endl;
err = getQueue().enqueueNDRangeKernel(getKernel(), cl::NullRange, cl::NDRange(m_particleCount), cl::NullRange, NULL, &event);
if (err != CL_SUCCESS)
std::cout << "Failed enqueueing kernel : " << getError(err) << std::endl;
err = event.wait();
if (err != CL_SUCCESS)
std::cout << "ERROR" << std::endl;
getQueue().enqueueCopyBuffer(m_clBuffer[Index::Particles], m_glBuffer[Index::Particles], 0, 0, m_particleCount * sizeof(Particle), NULL, NULL);
err = getQueue().enqueueReleaseGLObjects(&buffers, NULL, NULL);
if (err != CL_SUCCESS)
std::cout << "Failed releasing GL object : " << getError(err) << std::endl;
getQueue().finish();
}
Array<T> solveLU(const Array<T> &A, const Array<int> &pivot,
const Array<T> &b, const af_mat_prop options)
{
int N = A.dims()[0];
int NRHS = b.dims()[1];
std::vector<int> ipiv(N);
copyData(&ipiv[0], pivot);
Array< T > B = copyArray<T>(b);
const cl::Buffer *A_buf = A.get();
cl::Buffer *B_buf = B.get();
int info = 0;
magma_getrs_gpu<T>(MagmaNoTrans, N, NRHS,
(*A_buf)(), A.getOffset(), A.strides()[1],
&ipiv[0],
(*B_buf)(), B.getOffset(), B.strides()[1],
getQueue()(), &info);
return B;
}
Array<T> triangleSolve(const Array<T> &A, const Array<T> &b, const af_mat_prop options)
{
A.eval();
b.eval();
Array<T> B = copyArray<T>(b);
int N = B.dims()[0];
int NRHS = B.dims()[1];
auto func = [=] (Array<T> A, Array<T> B, int N, int NRHS, const af_mat_prop options) {
trtrs_func<T>()(AF_LAPACK_COL_MAJOR,
options & AF_MAT_UPPER ? 'U' : 'L',
'N', // transpose flag
options & AF_MAT_DIAG_UNIT ? 'U' : 'N',
N, NRHS,
A.get(), A.strides()[1],
B.get(), B.strides()[1]);
};
getQueue().enqueue(func, A, B, N, NRHS, options);
return B;
}
Array<T>::Array(const dim4 &dims, T *const in_data, bool is_device,
bool copy_device)
: info(getActiveDeviceId(), dims, 0, calcStrides(dims),
(af_dtype)dtype_traits<T>::af_type)
, data((is_device & !copy_device) ? (T *)in_data
: memAlloc<T>(dims.elements()).release(),
memFree<T>)
, data_dims(dims)
, node(bufferNodePtr<T>())
, ready(true)
, owner(true) {
static_assert(is_standard_layout<Array<T>>::value,
"Array<T> must be a standard layout type");
static_assert(
offsetof(Array<T>, info) == 0,
"Array<T>::info must be the first member variable of Array<T>");
if (!is_device || copy_device) {
// Ensure the memory being written to isnt used anywhere else.
getQueue().sync();
copy(in_data, in_data + dims.elements(), data.get());
}
}
void lu(Array<T> &lower, Array<T> &upper, Array<int> &pivot,
const Array<T> &in) {
lower.eval();
upper.eval();
pivot.eval();
in.eval();
dim4 iDims = in.dims();
int M = iDims[0];
int N = iDims[1];
Array<T> in_copy = copyArray<T>(in);
pivot = lu_inplace(in_copy);
// SPLIT into lower and upper
dim4 ldims(M, min(M, N));
dim4 udims(min(M, N), N);
lower = createEmptyArray<T>(ldims);
upper = createEmptyArray<T>(udims);
getQueue().enqueue(kernel::lu_split<T>, lower, upper, in_copy);
}
Array<T>* setIntersect(const Array<T> &first,
const Array<T> &second,
const bool is_unique)
{
if ((std::is_same<T, double>::value || std::is_same<T, cdouble>::value) &&
!isDoubleSupported(getActiveDeviceId())) {
OPENCL_NOT_SUPPORTED();
}
Array<T> unique_first = first;
Array<T> unique_second = second;
if (!is_unique) {
unique_first = *setUnique(first, false);
unique_second = *setUnique(second, false);
}
size_t out_size = std::max(unique_first.dims()[0], unique_second.dims()[0]);
Array<T> *out = createEmptyArray<T>(dim4(out_size, 1, 1, 1));
compute::command_queue queue(getQueue()());
compute::buffer first_data((*unique_first.get())());
compute::buffer second_data((*unique_second.get())());
compute::buffer out_data((*out->get())());
compute::buffer_iterator<T> first_begin(first_data, 0);
compute::buffer_iterator<T> first_end(first_data, unique_first.dims()[0]);
compute::buffer_iterator<T> second_begin(second_data, 0);
compute::buffer_iterator<T> second_end(second_data, unique_second.dims()[0]);
compute::buffer_iterator<T> out_begin(out_data, 0);
compute::buffer_iterator<T> out_end = compute::set_intersection(
first_begin, first_end, second_begin, second_end, out_begin, queue
);
out->resetDims(dim4(std::distance(out_begin, out_end), 1, 1, 1));
return out;
}
void testPushToQueue() {
cmdType nCommand_1=cmd_upgrade;
cmdOption nStat_1=upgrade_power;
int tar_1 = 1;
cmdType nCommand_2=cmd_upgrade;
cmdOption nStat_2=upgrade_range;
int tar_2 = 2;
ActionQueueStructure newQueue = getQueue(NULL);
sput_fail_unless(pushToQueue(newQueue,nCommand_1,nStat_1,tar_1) == 1,"Valid: 1 Queue Item");
sput_fail_unless(pushToQueue(newQueue,nCommand_2,nStat_2,tar_2) == 2,"Valid: 2 Queue Items");
sput_fail_unless(getFirstCommand(newQueue) == cmd_upgrade,"Valid: Top of Queue Upgrade Command");
sput_fail_unless(getFirstOption(newQueue) == upgrade_power,"Valid: Top of Queue Power Option");
sput_fail_unless(getLastCommand(newQueue) == cmd_upgrade,"Valid: Last in Queue upgrade Command");
sput_fail_unless(getLastOption(newQueue) == upgrade_range,"Valid: Last of Queue range Option");
pushToQueue(newQueue,cmd_mktwr,mktwr_int,2);
sput_fail_unless(getLastCommand(newQueue) == cmd_mktwr,"Valid: Last in Queue make tower command");
sput_fail_unless(getLastOption(newQueue) == mktwr_int,"Valid: Last option in Queue is int tower");
clearQueue();
}
Array<T> dot(const Array<T> &lhs, const Array<T> &rhs,
af_blas_transpose optLhs, af_blas_transpose optRhs)
{
initBlas();
int N = lhs.dims()[0];
dot_func<T> dot;
cl::Event event;
auto out = createEmptyArray<T>(af::dim4(1));
cl::Buffer scratch(getContext(), CL_MEM_READ_WRITE, sizeof(T) * N);
clblasStatus err;
err = dot(N,
(*out.get())(), out.getOffset(),
(*lhs.get())(), lhs.getOffset(), lhs.strides()[0],
(*rhs.get())(), rhs.getOffset(), rhs.strides()[0],
scratch(),
1, &getQueue()(), 0, nullptr, &event());
if(err) {
throw runtime_error(std::string("CLBLAS error: ") + std::to_string(err));
}
return out;
}
Array<T> sparseConvertStorageToDense(const SparseArray<T> &in_)
{
in_.eval();
Array<T> dense_ = createValueArray<T>(in_.dims(), scalar<T>(0));
dense_.eval();
auto func = [=] (Array<T> dense, const SparseArray<T> in) {
Array<T > values = in.getValues();
Array<int> rowIdx = in.getRowIdx();
Array<int> colIdx = in.getColIdx();
kernel::csr_dense<T>()(dense, values, rowIdx, colIdx);
};
getQueue().enqueue(func, dense_, in_);
if(stype == AF_STORAGE_CSR)
return dense_;
else
AF_ERROR("CPU Backend only supports Dense to CSR or COO", AF_ERR_NOT_SUPPORTED);
return dense_;
}
