clsparseStatus
atomic_reduce(clsparse::array_base<T>& pR,
const clsparse::array_base<T>& pX,
const cl_ulong wg_size,
const clsparseControl control)
{
assert(wg_size == pX.size());
std::string params = std::string()
+ " -DSIZE_TYPE=" + OclTypeTraits<cl_ulong>::type
+ " -DVALUE_TYPE=" + OclTypeTraits<T>::type
+ " -DWG_SIZE=" + std::to_string(wg_size)
+ " -D" + ReduceOperatorTrait<OP>::operation;
if (typeid(cl_float) == typeid(T))
{
std::string options = std::string() + " -DATOMIC_FLOAT";
params.append(options);
}
else if (typeid(cl_double) == typeid(T))
{
std::string options = std::string() + " -DATOMIC_DOUBLE";
params.append(options);
}
else if (typeid(cl_int) == typeid(T))
{
std::string options = std::string() + " -DATOMIC_INT";
params.append(options);
}
else
{
return clsparseInvalidType;
}
cl::Kernel kernel = KernelCache::get(control->queue,
"atomic_reduce", "reduce_block",
params);
KernelWrap kWrapper(kernel);
kWrapper << pR.data();
kWrapper << pX.data();
int blocksNum = (pX.size() + wg_size - 1) / wg_size;
int globalSize = blocksNum * wg_size;
cl::NDRange local(wg_size);
cl::NDRange global(globalSize);
cl_int status = kWrapper.run(control, global, local);
if (status != CL_SUCCESS)
{
return clsparseInvalidKernelExecution;
}
return clsparseSuccess;
}
clsparseStatus dot(clsparse::array_base<T>& pR,
const clsparse::array_base<T>& pX,
const clsparse::array_base<T>& pY,
const clsparseControl control)
{
cl_int status;
//not necessary to have it, but remember to init the pR with the proper value
init_scalar(pR, (T)0, control);
// with REDUCE_BLOCKS_NUMBER = 256 final reduction can be performed
// within one block;
const cl_ulong REDUCE_BLOCKS_NUMBER = 256;
/* For future optimisation
//workgroups per compute units;
const cl_uint WG_PER_CU = 64;
const cl_ulong REDUCE_BLOCKS_NUMBER = control->max_compute_units * WG_PER_CU;
*/
const cl_ulong REDUCE_BLOCK_SIZE = 256;
cl_ulong xSize = pX.size();
cl_ulong ySize = pY.size();
assert (xSize == ySize);
cl_ulong size = xSize;
if (size > 0)
{
cl::Context context = control->getContext();
//partial result
clsparse::vector<T> partial(control, REDUCE_BLOCKS_NUMBER, 0,
CL_MEM_READ_WRITE, false);
status = inner_product<T>(partial, pX, pY, size, REDUCE_BLOCKS_NUMBER,
REDUCE_BLOCK_SIZE, control);
if (status != clsparseSuccess)
{
return clsparseInvalidKernelExecution;
}
status = atomic_reduce<T>(pR, partial, REDUCE_BLOCK_SIZE,
control);
if (status != CL_SUCCESS)
{
return clsparseInvalidKernelExecution;
}
}
return clsparseSuccess;
}
clsparseStatus
csrmv_adaptive( const clsparse::array_base<T>& pAlpha,
const clsparseCsrMatrixPrivate* pCsrMatx,
const clsparse::array_base<T>& pX,
const clsparse::array_base<T>& pBeta,
clsparse::array_base<T>& pY,
clsparseControl control )
{
const cl_uint group_size = 256;
std::string params = std::string( )
+ " -DROWBITS=" + std::to_string( ROW_BITS )
+ " -DWGBITS=" + std::to_string( WG_BITS )
+ " -DBLOCKSIZE=" + std::to_string( BLKSIZE );
if(typeid(T) == typeid(cl_double))
{
std::string options = std::string() + " -DDOUBLE";
params.append(options);
}
cl::Kernel kernel = KernelCache::get( control->queue,
"csrmv_adaptive",
"csrmv_adaptive",
params );
KernelWrap kWrapper( kernel );
kWrapper << pCsrMatx->values
<< pCsrMatx->colIndices << pCsrMatx->rowOffsets
<< pX.data() << pY.data()
<< pCsrMatx->rowBlocks
<< pAlpha.data() << pBeta.data();
//<< h_alpha << h_beta;
// if NVIDIA is used it does not allow to run the group size
// which is not a multiplication of group_size. Don't know if that
// have an impact on performance
cl_uint global_work_size = ( pCsrMatx->rowBlockSize - 1 ) * group_size;
cl::NDRange local( group_size );
cl::NDRange global( global_work_size > local[ 0 ] ? global_work_size : local[ 0 ] );
cl_int status = kWrapper.run( control, global, local );
if( status != CL_SUCCESS )
{
return clsparseInvalidKernelExecution;
}
return clsparseSuccess;
}
clsparseStatus
axpy(clsparse::array_base<T>& pY,
const clsparse::array_base<T>& pAlpha,
const clsparse::array_base<T>& pX,
const clsparse::array_base<T>& pZ,
const clsparseControl control)
{
const int group_size = 256; // this or higher? control->max_wg_size?
const std::string params = std::string()
+ " -DSIZE_TYPE=" + OclTypeTraits<cl_ulong>::type
+ " -DVALUE_TYPE=" + OclTypeTraits<T>::type
+ " -DWG_SIZE=" + std::to_string( group_size )
+ " -D" + ElementWiseOperatorTrait<OP>::operation;
cl::Kernel kernel = KernelCache::get(control->queue, "blas1", "axpy",
params);
KernelWrap kWrapper(kernel);
cl_ulong size = pY.size();
cl_ulong offset = 0;
kWrapper << size
<< pY.data()
<< offset
<< pAlpha.data()
<< offset
<< pX.data()
<< offset
<< pZ.data()
<< offset;
int blocksNum = (size + group_size - 1) / group_size;
int globalSize = blocksNum * group_size;
cl::NDRange local(group_size);
cl::NDRange global (globalSize);
cl_int status = kWrapper.run(control, global, local);
if (status != CL_SUCCESS)
{
return clsparseInvalidKernelExecution;
}
return clsparseSuccess;
}
clsparseStatus
inner_product (clsparse::array_base<T>& partial,
const clsparse::array_base<T>& pX,
const clsparse::array_base<T>& pY,
const cl_ulong size,
const cl_ulong REDUCE_BLOCKS_NUMBER,
const cl_ulong REDUCE_BLOCK_SIZE,
const clsparseControl control)
{
cl_ulong nthreads = REDUCE_BLOCK_SIZE * REDUCE_BLOCKS_NUMBER;
std::string params = std::string()
+ " -DSIZE_TYPE=" + OclTypeTraits<cl_ulong>::type
+ " -DVALUE_TYPE=" + OclTypeTraits<T>::type
+ " -DWG_SIZE=" + std::to_string(REDUCE_BLOCK_SIZE)
+ " -DREDUCE_BLOCK_SIZE=" + std::to_string(REDUCE_BLOCK_SIZE)
+ " -DN_THREADS=" + std::to_string(nthreads);
cl::Kernel kernel = KernelCache::get(control->queue,
"dot", "inner_product", params);
KernelWrap kWrapper(kernel);
kWrapper << size
<< partial.data()
<< pX.data()
<< pY.data();
cl::NDRange local(REDUCE_BLOCK_SIZE);
cl::NDRange global(REDUCE_BLOCKS_NUMBER * REDUCE_BLOCK_SIZE);
cl_int status = kWrapper.run(control, global, local);
if (status != CL_SUCCESS)
{
return clsparseInvalidKernelExecution;
}
return clsparseSuccess;
}
clsparseStatus
scale(clsparse::array_base<T>& pVector,
const clsparse::array_base<T>& pAlpha,
clsparseControl control)
{
const int group_size = 256;
//const int group_size = control->max_wg_size;
const std::string params = std::string()
+ " -DSIZE_TYPE=" + OclTypeTraits<cl_ulong>::type
+ " -DVALUE_TYPE="+ OclTypeTraits<T>::type
+ " -DWG_SIZE=" + std::to_string(group_size);
cl::Kernel kernel = KernelCache::get(control->queue,
"blas1", "scale",
params);
KernelWrap kWrapper(kernel);
cl_ulong size = pVector.size();
cl_ulong offset = 0;
kWrapper << size
<< pVector.data()
<< offset
<< pAlpha.data()
<< offset;
int blocksNum = (size + group_size - 1) / group_size;
int globalSize = blocksNum * group_size;
cl::NDRange local(group_size);
cl::NDRange global (globalSize);
cl_int status = kWrapper.run(control, global, local);
if (status != CL_SUCCESS)
{
return clsparseInvalidKernelExecution;
}
return clsparseSuccess;
}
clsparseStatus
scale( clsparse::array_base<T>& pResult,
const clsparse::array_base<T>& pAlpha,
const clsparse::array_base<T>& pVector,
clsparseControl control)
{
const int group_size = 256;
//const int group_size = control->max_wg_size;
std::string params = std::string()
+ " -DVALUE_TYPE="+ OclTypeTraits<T>::type
+ " -DWG_SIZE=" + std::to_string(group_size);
if (sizeof(clsparseIdx_t) == 8)
{
std::string options = std::string()
+ " -DSIZE_TYPE=" + OclTypeTraits<cl_ulong>::type;
params.append(options);
}
else
{
std::string options = std::string()
+ " -DSIZE_TYPE=" + OclTypeTraits<cl_uint>::type;
params.append(options);
}
if(typeid(T) == typeid(cl_double))
{
params.append(" -DDOUBLE");
if (!control->dpfp_support)
{
#ifndef NDEBUG
std::cerr << "Failure attempting to run double precision kernel on device without DPFP support." << std::endl;
#endif
return clsparseInvalidDevice;
}
}
cl::Kernel kernel = KernelCache::get(control->queue,
"blas1", "scale",
params);
KernelWrap kWrapper(kernel);
clsparseIdx_t size = pResult.size();
clsparseIdx_t offset = 0;
kWrapper << size
<< pResult.data()
<< offset
<< pVector.data()
<< offset
<< pAlpha.data()
<< offset;
clsparseIdx_t blocksNum = (size + group_size - 1) / group_size;
clsparseIdx_t globalSize = blocksNum * group_size;
cl::NDRange local(group_size);
cl::NDRange global (globalSize);
cl_int status = kWrapper.run(control, global, local);
if (status != CL_SUCCESS)
{
return clsparseInvalidKernelExecution;
}
return clsparseSuccess;
}
clsparseStatus
csrmv_vector(const clsparse::array_base<T>& pAlpha,
const clsparseCsrMatrixPrivate* pMatx,
const clsparse::array_base<T>& pX,
const clsparse::array_base<T>& pBeta,
clsparse::array_base<T>& pY,
clsparseControl control)
{
cl_uint nnz_per_row = pMatx->nnz_per_row(); //average nnz per row
cl_uint wave_size = control->wavefront_size;
cl_uint group_size = 256; // 256 gives best performance!
cl_uint subwave_size = wave_size;
// adjust subwave_size according to nnz_per_row;
// each wavefron will be assigned to the row of the csr matrix
if(wave_size > 32)
{
//this apply only for devices with wavefront > 32 like AMD(64)
if (nnz_per_row < 64) { subwave_size = 32; }
}
if (nnz_per_row < 32) { subwave_size = 16; }
if (nnz_per_row < 16) { subwave_size = 8; }
if (nnz_per_row < 8) { subwave_size = 4; }
if (nnz_per_row < 4) { subwave_size = 2; }
const std::string params = std::string() +
"-DINDEX_TYPE=" + OclTypeTraits<cl_int>::type
+ " -DVALUE_TYPE=" + OclTypeTraits<T>::type
+ " -DSIZE_TYPE=" + OclTypeTraits<cl_ulong>::type
+ " -DWG_SIZE=" + std::to_string(group_size)
+ " -DWAVE_SIZE=" + std::to_string(wave_size)
+ " -DSUBWAVE_SIZE=" + std::to_string(subwave_size);
cl::Kernel kernel = KernelCache::get(control->queue,
"csrmv_general",
"csrmv_general",
params);
KernelWrap kWrapper(kernel);
cl_ulong offset = 0;
kWrapper << pMatx->num_rows
<< pAlpha.data() << offset
<< pMatx->rowOffsets
<< pMatx->colIndices
<< pMatx->values
<< pX.data() << offset
<< pBeta.data() << offset
<< pY.data() << offset;
// subwave takes care of each row in matrix;
// predicted number of subwaves to be executed;
cl_uint predicted = subwave_size * pMatx->num_rows;
// if NVIDIA is used it does not allow to run the group size
// which is not a multiplication of group_size. Don't know if that
// have an impact on performance
cl_uint global_work_size =
group_size* ((predicted + group_size - 1 ) / group_size);
cl::NDRange local(group_size);
//cl::NDRange global(predicted > local[0] ? predicted : local[0]);
cl::NDRange global(global_work_size > local[0] ? global_work_size : local[0]);
cl_int status = kWrapper.run(control, global, local);
if (status != CL_SUCCESS)
{
return clsparseInvalidKernelExecution;
}
return clsparseSuccess;
}
clsparseStatus
csrmv_adaptive( const clsparse::array_base<T>& pAlpha,
const clsparseCsrMatrixPrivate* pCsrMatx,
const clsparse::array_base<T>& pX,
const clsparse::array_base<T>& pBeta,
clsparse::array_base<T>& pY,
clsparseControl control )
{
const cl_uint group_size = 256;
std::string params = std::string( )
+ " -DROWBITS=" + std::to_string( ROW_BITS )
+ " -DWGBITS=" + std::to_string( WG_BITS )
+ " -DVALUE_TYPE=" + OclTypeTraits<T>::type
+ " -DWG_SIZE=" + std::to_string( group_size )
+ " -DBLOCKSIZE=" + std::to_string( BLKSIZE )
+ " -DBLOCK_MULTIPLIER=" + std::to_string( BLOCK_MULTIPLIER )
+ " -DROWS_FOR_VECTOR=" + std::to_string( ROWS_FOR_VECTOR );
if( sizeof( clsparseIdx_t ) == 8 )
{
std::string options = std::string()
+ " -DINDEX_TYPE=" + OclTypeTraits<cl_ulong>::type;
params.append(options);
}
else
{
std::string options = std::string()
+ " -DINDEX_TYPE=" + OclTypeTraits<cl_uint>::type;
params.append(options);
}
std::string options;
if(typeid(T) == typeid(cl_double))
{
options = std::string() + " -DDOUBLE";
if (!control->dpfp_support)
{
#ifndef NDEBUG
std::cerr << "Failure attempting to run double precision kernel on device without DPFP support." << std::endl;
#endif
return clsparseInvalidDevice;
}
}
else if(typeid(T) == typeid(cl_ulong))
options = std::string() + " -DLONG";
else if(typeid(T) == typeid(cl_long))
options = std::string() + " -DLONG";
if(control->extended_precision)
options += " -DEXTENDED_PRECISION";
params.append(options);
cl::Kernel kernel = KernelCache::get( control->queue,
"csrmv_adaptive",
"csrmv_adaptive",
params );
KernelWrap kWrapper( kernel );
const matrix_meta* meta_ptr = static_cast< const matrix_meta* >( pCsrMatx->meta );
kWrapper << pCsrMatx->values
<< pCsrMatx->col_indices << pCsrMatx->row_pointer
<< pX.data() << pY.data()
<< meta_ptr->rowBlocks
<< pAlpha.data() << pBeta.data();
//<< h_alpha << h_beta;
// if NVIDIA is used it does not allow to run the group size
// which is not a multiplication of group_size. Don't know if that
// have an impact on performance
// Setting global work size to half the row block size because we are only
// using half the row blocks buffer for actual work.
// The other half is used for the extended precision reduction.
clsparseIdx_t global_work_size = ( ( meta_ptr->rowBlockSize/2) - 1 ) * group_size;
cl::NDRange local( group_size );
cl::NDRange global( global_work_size > local[ 0 ] ? global_work_size : local[ 0 ] );
cl_int status = kWrapper.run( control, global, local );
if( status != CL_SUCCESS )
{
return clsparseInvalidKernelExecution;
}
return clsparseSuccess;
}
