static void get_out_idx(Buffer *out_data,
Param &otmp, Param &rtmp,
Param &in, uint threads_x,
uint groups_x, uint groups_y)
{
std::string refName = std::string("get_out_idx_kernel_") + std::string(dtype_traits<T>::getName());
int device = getActiveDeviceId();
kc_entry_t entry = kernelCache(device, refName);
if (entry.prog==0 && entry.ker==0) {
ToNumStr<T> toNumStr;
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D zero=" << toNumStr(scalar<T>(0))
<< " -D 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[] = {where_cl};
const int ker_lens[] = {where_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "get_out_idx_kernel");
addKernelToCache(device, refName, entry);
}
NDRange local(threads_x, THREADS_PER_GROUP / threads_x);
NDRange global(local[0] * groups_x * in.info.dims[2], local[1] * groups_y * in.info.dims[3]);
uint lim = divup(otmp.info.dims[0], (threads_x * groups_x));
auto whereOp = KernelFunctor< Buffer, Buffer, KParam, Buffer, KParam,
Buffer, KParam, uint, uint, uint>(*entry.ker);
whereOp(EnqueueArgs(getQueue(), global, local),
*out_data, *otmp.data, otmp.info,
*rtmp.data, rtmp.info, *in.data, in.info,
groups_x, groups_y, lim);
CL_DEBUG_FINISH(getQueue());
}
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);
}
void iota(Param out)
{
try {
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static std::map<int, Program*> iotaProgs;
static std::map<int, Kernel*> iotaKernels;
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName();
options << " -D rep=" << rep;
if (std::is_same<T, double>::value ||
std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
Program prog;
buildProgram(prog, iota_cl, iota_cl_len, options.str());
iotaProgs[device] = new Program(prog);
iotaKernels[device] = new Kernel(*iotaProgs[device], "iota_kernel");
});
auto iotaOp = make_kernel<Buffer, const KParam, const dim_type, const dim_type>
(*iotaKernels[device]);
NDRange local(TX, TY, 1);
dim_type blocksPerMatX = divup(out.info.dims[0], TILEX);
dim_type blocksPerMatY = divup(out.info.dims[1], TILEY);
NDRange global(local[0] * blocksPerMatX * out.info.dims[2],
local[1] * blocksPerMatY * out.info.dims[3],
1);
iotaOp(EnqueueArgs(getQueue(), global, local),
*out.data, out.info, blocksPerMatX, blocksPerMatY);
CL_DEBUG_FINISH(getQueue());
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
void triangle(Param out, const Param in)
{
std::string refName = std::string("triangle_kernel_") + std::string(dtype_traits<T>::getName()) +
std::to_string(is_upper) + std::to_string(is_unit_diag);
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 is_upper=" << is_upper
<< " -D is_unit_diag=" << is_unit_diag
<< " -D ZERO=(T)(" << scalar_to_option(scalar<T>(0)) << ")"
<< " -D ONE=(T)(" << scalar_to_option(scalar<T>(1)) << ")";
if (std::is_same<T, double>::value || std::is_same<T, cdouble>::value)
options << " -D USE_DOUBLE";
const char* ker_strs[] = {triangle_cl};
const int ker_lens[] = {triangle_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "triangle_kernel");
addKernelToCache(device, refName, entry);
}
NDRange local(TX, TY);
int groups_x = divup(out.info.dims[0], TILEX);
int groups_y = divup(out.info.dims[1], TILEY);
NDRange global(groups_x * out.info.dims[2] * local[0], groups_y * out.info.dims[3] * local[1]);
auto triangleOp = KernelFunctor< Buffer, KParam, const Buffer, KParam,
const int, const int >(*entry.ker);
triangleOp(EnqueueArgs(getQueue(), global, local),
*out.data, out.info, *in.data, in.info, groups_x, groups_y);
CL_DEBUG_FINISH(getQueue());
}
void iota(Param out, const af::dim4 &sdims, const af::dim4 &tdims)
{
std::string refName = std::string("iota_kernel_") + std::string(dtype_traits<T>::getName());
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();
if (std::is_same<T, double>::value || std::is_same<T, cdouble>::value)
options << " -D USE_DOUBLE";
const char* ker_strs[] = {iota_cl};
const int ker_lens[] = {iota_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "iota_kernel");
addKernelToCache(device, refName, entry);
}
auto iotaOp = KernelFunctor<Buffer, const KParam,
const int, const int, const int, const int,
const int, const int, const int, const int,
const int, const int> (*entry.ker);
NDRange local(IOTA_TX, IOTA_TY, 1);
int blocksPerMatX = divup(out.info.dims[0], TILEX);
int blocksPerMatY = divup(out.info.dims[1], TILEY);
NDRange global(local[0] * blocksPerMatX * out.info.dims[2],
local[1] * blocksPerMatY * out.info.dims[3], 1);
iotaOp(EnqueueArgs(getQueue(), global, local),
*out.data, out.info, sdims[0], sdims[1], sdims[2], sdims[3],
tdims[0], tdims[1], tdims[2], tdims[3], blocksPerMatX, blocksPerMatY);
CL_DEBUG_FINISH(getQueue());
}
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);
}
void hsv2rgb_convert(Param out, const Param in)
{
std::string refName = std::string("hsvrgb_convert_") +
std::string(dtype_traits<T>::getName()) + std::to_string(isHSV2RGB);
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();
if(isHSV2RGB) options << " -D isHSV2RGB";
if (std::is_same<T, double>::value) options << " -D USE_DOUBLE";
const char* ker_strs[] = {hsv_rgb_cl};
const int ker_lens[] = {hsv_rgb_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "convert");
addKernelToCache(device, refName, entry);
}
NDRange local(THREADS_X, THREADS_Y);
int blk_x = divup(in.info.dims[0], THREADS_X);
int blk_y = divup(in.info.dims[1], THREADS_Y);
// all images are three channels, so batch
// parameter would be along 4th dimension
NDRange global(blk_x * in.info.dims[3] * THREADS_X, blk_y * THREADS_Y);
auto hsvrgbOp = KernelFunctor<Buffer, KParam, Buffer, KParam, int> (*entry.ker);
hsvrgbOp(EnqueueArgs(getQueue(), global, local), *out.data, out.info, *in.data, in.info, blk_x);
CL_DEBUG_FINISH(getQueue());
}
static void identity(Param out) {
std::string refName = std::string("identity_kernel") +
std::string(dtype_traits<T>::getName());
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 ONE=(T)("
<< scalar_to_option(scalar<T>(1)) << ")"
<< " -D ZERO=(T)(" << scalar_to_option(scalar<T>(0)) << ")";
if (std::is_same<T, double>::value || std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
const char* ker_strs[] = {identity_cl};
const int ker_lens[] = {identity_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "identity_kernel");
addKernelToCache(device, refName, entry);
}
NDRange local(32, 8);
int groups_x = divup(out.info.dims[0], local[0]);
int groups_y = divup(out.info.dims[1], local[1]);
NDRange global(groups_x * out.info.dims[2] * local[0],
groups_y * out.info.dims[3] * local[1]);
auto identityOp = KernelFunctor<Buffer, const KParam, int, int>(*entry.ker);
identityOp(EnqueueArgs(getQueue(), global, local), *(out.data), out.info,
groups_x, groups_y);
CL_DEBUG_FINISH(getQueue());
}
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());
}
void
set(Buffer &ptr, T val, const size_t &elements)
{
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static Program setProgs[DeviceManager::MAX_DEVICES];
static Kernel setKernels[DeviceManager::MAX_DEVICES];
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
Program::Sources setSrc;
setSrc.emplace_back(set_cl, set_cl_len);
setProgs[device] = Program(getContext(), setSrc);
string opt = string("-D T=") + dtype_traits<T>::getName();
setProgs[device].build(opt.c_str());
setKernels[device] = Kernel(setProgs[device], "set");
});
auto setKern = make_kernel<Buffer, T, const unsigned long>(setKernels[device]);
setKern(EnqueueArgs(getQueue(), NDRange(elements)), ptr, val, elements);
}
void copy(Param dst, const Param src, int ndims, outType default_value, double factor)
{
try {
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static std::map<int, Program*> cpyProgs;
static std::map<int, Kernel*> cpyKernels;
int device = getActiveDeviceId();
std::call_once(compileFlags[device], [&]() {
std::ostringstream options;
options << " -D inType=" << dtype_traits<inType>::getName()
<< " -D outType=" << dtype_traits<outType>::getName()
<< " -D inType_" << dtype_traits<inType>::getName()
<< " -D outType_" << dtype_traits<outType>::getName()
<< " -D SAME_DIMS=" << same_dims;
if (std::is_same<inType, double>::value ||
std::is_same<inType, cdouble>::value ||
std::is_same<outType, double>::value ||
std::is_same<outType, cdouble>::value) {
options << " -D USE_DOUBLE";
}
Program prog;
buildProgram(prog, copy_cl, copy_cl_len, options.str());
cpyProgs[device] = new Program(prog);
cpyKernels[device] = new Kernel(*cpyProgs[device], "copy");
});
NDRange local(DIM0, DIM1);
size_t local_size[] = {DIM0, DIM1};
local_size[0] *= local_size[1];
if (ndims == 1) {
local_size[1] = 1;
}
int blk_x = divup(dst.info.dims[0], local_size[0]);
int blk_y = divup(dst.info.dims[1], local_size[1]);
NDRange global(blk_x * dst.info.dims[2] * DIM0,
blk_y * dst.info.dims[3] * DIM1);
dims_t trgt_dims;
if (same_dims) {
trgt_dims= {{dst.info.dims[0], dst.info.dims[1], dst.info.dims[2], dst.info.dims[3]}};
} else {
dim_t trgt_l = std::min(dst.info.dims[3], src.info.dims[3]);
dim_t trgt_k = std::min(dst.info.dims[2], src.info.dims[2]);
dim_t trgt_j = std::min(dst.info.dims[1], src.info.dims[1]);
dim_t trgt_i = std::min(dst.info.dims[0], src.info.dims[0]);
trgt_dims= {{trgt_i, trgt_j, trgt_k, trgt_l}};
}
auto copyOp = KernelFunctor<Buffer, KParam, Buffer, KParam,
outType, float, dims_t,
int, int
>(*cpyKernels[device]);
copyOp(EnqueueArgs(getQueue(), global, local),
*dst.data, dst.info, *src.data, src.info,
default_value, (float)factor, trgt_dims, blk_x, blk_y);
CL_DEBUG_FINISH(getQueue());
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
void csrmm_nt(Param out, const Param &values, const Param &rowIdx,
const Param &colIdx, const Param &rhs, const T alpha,
const T beta) {
bool use_alpha = (alpha != scalar<T>(1.0));
bool use_beta = (beta != scalar<T>(0.0));
// Using greedy indexing is causing performance issues on many platforms
// FIXME: Figure out why
bool use_greedy = false;
std::string ref_name = std::string("csrmm_nt_") +
std::string(dtype_traits<T>::getName()) +
std::string("_") + std::to_string(use_alpha) +
std::string("_") + std::to_string(use_beta) +
std::string("_") + std::to_string(use_greedy);
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();
options << " -D USE_ALPHA=" << use_alpha;
options << " -D USE_BETA=" << use_beta;
options << " -D USE_GREEDY=" << use_greedy;
options << " -D THREADS_PER_GROUP=" << THREADS_PER_GROUP;
if (std::is_same<T, double>::value || std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
if (std::is_same<T, cfloat>::value || std::is_same<T, cdouble>::value) {
options << " -D IS_CPLX=1";
} else {
options << " -D IS_CPLX=0";
}
const char *ker_strs[] = {csrmm_cl};
const int ker_lens[] = {csrmm_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel[2];
entry.ker[0] = Kernel(*entry.prog, "csrmm_nt");
// FIXME: Change this after adding another kernel
entry.ker[1] = Kernel(*entry.prog, "csrmm_nt");
addKernelToCache(device, ref_name, entry);
}
auto csrmm_nt_kernel = entry.ker[0];
auto csrmm_nt_func =
KernelFunctor<Buffer, Buffer, Buffer, Buffer, int, int, Buffer, KParam,
T, T, Buffer>(csrmm_nt_kernel);
NDRange local(THREADS_PER_GROUP, 1);
int M = rowIdx.info.dims[0] - 1;
int N = rhs.info.dims[0];
int groups_x = divup(N, local[0]);
int groups_y = divup(M, REPEAT);
groups_y = std::min(groups_y, MAX_CSRMM_GROUPS);
NDRange global(local[0] * groups_x, local[1] * groups_y);
std::vector<int> count(groups_x);
cl::Buffer *counter = bufferAlloc(count.size() * sizeof(int));
getQueue().enqueueWriteBuffer(
*counter, CL_TRUE, 0, count.size() * sizeof(int), (void *)count.data());
csrmm_nt_func(EnqueueArgs(getQueue(), global, local), *out.data,
*values.data, *rowIdx.data, *colIdx.data, M, N, *rhs.data,
rhs.info, alpha, beta, *counter);
bufferFree(counter);
}
void approx1(Param out, const Param in, const Param pos, const float offGrid)
{
try {
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static std::map<int, Program*> approxProgs;
static std::map<int, Kernel*> approxKernels;
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
ToNum<Ty> toNum;
std::ostringstream options;
options << " -D Ty=" << dtype_traits<Ty>::getName()
<< " -D Tp=" << dtype_traits<Tp>::getName()
<< " -D ZERO=" << toNum(scalar<Ty>(0));
if((af_dtype) dtype_traits<Ty>::af_type == c32 ||
(af_dtype) dtype_traits<Ty>::af_type == c64) {
options << " -D CPLX=1";
} else {
options << " -D CPLX=0";
}
if (std::is_same<Ty, double>::value ||
std::is_same<Ty, cdouble>::value) {
options << " -D USE_DOUBLE";
}
switch(method) {
case AF_INTERP_NEAREST: options << " -D INTERP=NEAREST";
break;
case AF_INTERP_LINEAR: options << " -D INTERP=LINEAR";
break;
default:
break;
}
Program prog;
buildProgram(prog, approx1_cl, approx1_cl_len, options.str());
approxProgs[device] = new Program(prog);
approxKernels[device] = new Kernel(*approxProgs[device], "approx1_kernel");
});
auto approx1Op = make_kernel<Buffer, const KParam, const Buffer, const KParam,
const Buffer, const KParam, const float, const int>
(*approxKernels[device]);
NDRange local(THREADS, 1, 1);
int blocksPerMat = divup(out.info.dims[0], local[0]);
NDRange global(blocksPerMat * local[0] * out.info.dims[1],
out.info.dims[2] * out.info.dims[3] * local[0],
1);
approx1Op(EnqueueArgs(getQueue(), global, local),
*out.data, out.info, *in.data, in.info,
*pos.data, pos.info, offGrid, blocksPerMat);
CL_DEBUG_FINISH(getQueue());
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
void fast(unsigned* out_feat,
Param &x_out,
Param &y_out,
Param &score_out,
Param in,
const float thr,
const float feature_ratio,
const unsigned edge)
{
try {
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static std::map<int, Program*> fastProgs;
static std::map<int, Kernel*> lfKernel;
static std::map<int, Kernel*> nmKernel;
static std::map<int, Kernel*> gfKernel;
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D ARC_LENGTH=" << arc_length
<< " -D NONMAX=" << static_cast<unsigned>(nonmax);
if (std::is_same<T, double>::value ||
std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
cl::Program prog;
buildProgram(prog, fast_cl, fast_cl_len, options.str());
fastProgs[device] = new Program(prog);
lfKernel[device] = new Kernel(*fastProgs[device], "locate_features");
nmKernel[device] = new Kernel(*fastProgs[device], "non_max_counts");
gfKernel[device] = new Kernel(*fastProgs[device], "get_features");
});
const unsigned max_feat = ceil(in.info.dims[0] * in.info.dims[1] * feature_ratio);
// Matrix containing scores for detected features, scores are stored in the
// same coordinates as features, dimensions should be equal to in.
cl::Buffer *d_score = bufferAlloc(in.info.dims[0] * in.info.dims[1] * sizeof(float));
std::vector<float> score_init(in.info.dims[0] * in.info.dims[1], (float)0);
getQueue().enqueueWriteBuffer(*d_score, CL_TRUE, 0, in.info.dims[0] * in.info.dims[1] * sizeof(float), &score_init[0]);
cl::Buffer *d_flags = d_score;
if (nonmax) {
d_flags = bufferAlloc(in.info.dims[0] * in.info.dims[1] * sizeof(T));
}
const int blk_x = divup(in.info.dims[0]-edge*2, FAST_THREADS_X);
const int blk_y = divup(in.info.dims[1]-edge*2, FAST_THREADS_Y);
// Locate features kernel sizes
const NDRange local(FAST_THREADS_X, FAST_THREADS_Y);
const NDRange global(blk_x * FAST_THREADS_X, blk_y * FAST_THREADS_Y);
auto lfOp = make_kernel<Buffer, KParam,
Buffer, const float, const unsigned,
LocalSpaceArg> (*lfKernel[device]);
lfOp(EnqueueArgs(getQueue(), global, local),
*in.data, in.info, *d_score, thr, edge,
cl::Local((FAST_THREADS_X + 6) * (FAST_THREADS_Y + 6) * sizeof(T)));
CL_DEBUG_FINISH(getQueue());
const int blk_nonmax_x = divup(in.info.dims[0], 64);
const int blk_nonmax_y = divup(in.info.dims[1], 64);
// Nonmax kernel sizes
const NDRange local_nonmax(FAST_THREADS_NONMAX_X, FAST_THREADS_NONMAX_Y);
const NDRange global_nonmax(blk_nonmax_x * FAST_THREADS_NONMAX_X, blk_nonmax_y * FAST_THREADS_NONMAX_Y);
unsigned count_init = 0;
cl::Buffer *d_total = bufferAlloc(sizeof(unsigned));
getQueue().enqueueWriteBuffer(*d_total, CL_TRUE, 0, sizeof(unsigned), &count_init);
//size_t *global_nonmax_dims = global_nonmax();
size_t blocks_sz = blk_nonmax_x * FAST_THREADS_NONMAX_X * blk_nonmax_y * FAST_THREADS_NONMAX_Y * sizeof(unsigned);
cl::Buffer *d_counts = bufferAlloc(blocks_sz);
cl::Buffer *d_offsets = bufferAlloc(blocks_sz);
auto nmOp = make_kernel<Buffer, Buffer, Buffer,
Buffer, Buffer,
KParam, const unsigned> (*nmKernel[device]);
nmOp(EnqueueArgs(getQueue(), global_nonmax, local_nonmax),
*d_counts, *d_offsets, *d_total, *d_flags, *d_score, in.info, edge);
CL_DEBUG_FINISH(getQueue());
unsigned total;
getQueue().enqueueReadBuffer(*d_total, CL_TRUE, 0, sizeof(unsigned), &total);
total = total < max_feat ? total : max_feat;
if (total > 0) {
size_t out_sz = total * sizeof(float);
x_out.data = bufferAlloc(out_sz);
y_out.data = bufferAlloc(out_sz);
score_out.data = bufferAlloc(out_sz);
auto gfOp = make_kernel<Buffer, Buffer, Buffer,
Buffer, Buffer, Buffer,
KParam, const unsigned,
const unsigned> (*gfKernel[device]);
gfOp(EnqueueArgs(getQueue(), global_nonmax, local_nonmax),
*x_out.data, *y_out.data, *score_out.data,
*d_flags, *d_counts, *d_offsets,
in.info, total, edge);
CL_DEBUG_FINISH(getQueue());
}
*out_feat = total;
x_out.info.dims[0] = total;
x_out.info.strides[0] = 1;
y_out.info.dims[0] = total;
y_out.info.strides[0] = 1;
score_out.info.dims[0] = total;
score_out.info.strides[0] = 1;
for (int k = 1; k < 4; k++) {
x_out.info.dims[k] = 1;
x_out.info.strides[k] = total;
y_out.info.dims[k] = 1;
y_out.info.strides[k] = total;
score_out.info.dims[k] = 1;
score_out.info.strides[k] = total;
}
bufferFree(d_score);
if (nonmax) bufferFree(d_flags);
bufferFree(d_total);
bufferFree(d_counts);
bufferFree(d_offsets);
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
void reorderOutputHelper(Param out,
Param packed,
Param sig,
Param filter,
const int baseDim,
ConvolveBatchKind kind)
{
try {
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static std::map<int, Program*> fftconvolveProgs;
static std::map<int, Kernel*> roKernel;
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D ROUND_OUT=" << (int)roundOut
<< " -D EXPAND=" << (int)expand;
if ((af_dtype) dtype_traits<convT>::af_type == c32) {
options << " -D CONVT=float";
}
else if ((af_dtype) dtype_traits<convT>::af_type == c64 && isDouble) {
options << " -D CONVT=double"
<< " -D USE_DOUBLE";
}
cl::Program prog;
buildProgram(prog,
fftconvolve_reorder_cl,
fftconvolve_reorder_cl_len,
options.str());
fftconvolveProgs[device] = new Program(prog);
roKernel[device] = new Kernel(*fftconvolveProgs[device], "reorder_output");
});
Param sig_tmp, filter_tmp;
calcParamSizes(sig_tmp, filter_tmp, packed, sig, filter, baseDim, kind);
// Number of packed complex elements in dimension 0
int sig_half_d0 = divup(sig.info.dims[0], 2);
int blocks = divup(out.info.strides[3] * out.info.dims[3], THREADS);
NDRange local(THREADS);
NDRange global(blocks * THREADS);
auto roOp = make_kernel<Buffer, KParam,
Buffer, KParam,
KParam, const int,
const int> (*roKernel[device]);
if (kind == ONE2MANY) {
roOp(EnqueueArgs(getQueue(), global, local),
*out.data, out.info,
*filter_tmp.data, filter_tmp.info,
filter.info, sig_half_d0, baseDim);
}
else {
roOp(EnqueueArgs(getQueue(), global, local),
*out.data, out.info,
*sig_tmp.data, sig_tmp.info,
filter.info, sig_half_d0, baseDim);
}
CL_DEBUG_FINISH(getQueue());
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
void unwrap(Param out, const Param in, const dim_t wx, const dim_t wy,
const dim_t sx, const dim_t sy, const dim_t px, const dim_t py,
const dim_t nx, const bool is_column) {
std::string ref_name = std::string("unwrap_") +
std::string(dtype_traits<T>::getName()) +
std::string("_") + std::to_string(is_column);
int device = getActiveDeviceId();
kc_entry_t entry = kernelCache(device, ref_name);
if (entry.prog == 0 && entry.ker == 0) {
ToNumStr<T> toNumStr;
std::ostringstream options;
options << " -D is_column=" << is_column
<< " -D ZERO=" << toNumStr(scalar<T>(0))
<< " -D T=" << dtype_traits<T>::getName();
if (std::is_same<T, double>::value || std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
Program prog;
buildProgram(prog, unwrap_cl, unwrap_cl_len, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "unwrap_kernel");
addKernelToCache(device, ref_name, entry);
}
dim_t TX = 1, TY = 1;
dim_t BX = 1;
const dim_t BY = out.info.dims[2] * out.info.dims[3];
dim_t reps = 1;
if (is_column) {
TX = std::min(THREADS_PER_GROUP, nextpow2(out.info.dims[0]));
TY = THREADS_PER_GROUP / TX;
BX = divup(out.info.dims[1], TY);
reps = divup((wx * wy), TX);
} else {
TX = THREADS_X;
TY = THREADS_Y;
BX = divup(out.info.dims[0], TX);
reps = divup((wx * wy), TY);
}
NDRange local(TX, TY);
NDRange global(local[0] * BX, local[1] * BY);
auto unwrapOp =
KernelFunctor<Buffer, const KParam, const Buffer, const KParam,
const int, const int, const int, const int, const int,
const int, const int, const int>(*entry.ker);
unwrapOp(EnqueueArgs(getQueue(), global, local), *out.data, out.info,
*in.data, in.info, wx, wy, sx, sy, px, py, nx, reps);
CL_DEBUG_FINISH(getQueue());
}
void exampleFunc(Param c, const Param a, const Param b, const af_someenum_t p)
{
try {
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static std::map<int, Program*> egProgs;
static std::map<int, Kernel*> egKernels;
int device = getActiveDeviceId();
// std::call_once is used to ensure OpenCL kernels
// are compiled only once for any given device and combination
// of template parameters to this kernel wrapper function 'exampleFunc<T>'
std::call_once( compileFlags[device], [device] () {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName();
// You can pass any template parameters as compile options
// to kernel the compilation step. This is equivalent of
// having templated kernels in CUDA
// The following option is passed to kernel compilation
// if template parameter T is double or complex double
// to enable FP64 extension
if (std::is_same<T, double>::value ||
std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
Program prog;
// below helper function 'buildProgram' uses the option string
// we just created and compiles the kernel string
// 'example_cl' which was created by our opencl kernel code obfuscation
// stage
buildProgram(prog, example_cl, example_cl_len, options.str());
// create a cl::Program object on heap
egProgs[device] = new Program(prog);
// create a cl::Kernel object on heap
egKernels[device] = new Kernel(*egProgs[device], "example");
});
// configure work group parameters
NDRange local(THREADS_X, THREADS_Y);
int blk_x = divup(c.info.dims[0], THREADS_X);
int blk_y = divup(c.info.dims[1], THREADS_Y);
// configure global launch parameters
NDRange global(blk_x * THREADS_X, blk_y * THREADS_Y);
// create a kernel functor from the cl::Kernel object
// corresponding to the device on which current execution
// is happending.
auto exampleFuncOp = KernelFunctor<Buffer, KParam, Buffer, KParam,
Buffer, KParam, int>(*egKernels[device]);
// launch the kernel
exampleFuncOp(EnqueueArgs(getQueue(), global, local),
*c.data, c.info, *a.data, a.info, *b.data, b.info, (int)p);
// Below Macro activates validations ONLY in DEBUG
// mode as its name indicates
CL_DEBUG_FINISH(getQueue());
} catch (cl::Error err) { // Catch all cl::Errors and convert them
// to appropriate ArrayFire error codes
CL_TO_AF_ERROR(err);
}
}
void swapdblk(int n, int nb,
cl_mem dA, size_t dA_offset, int ldda, int inca,
cl_mem dB, size_t dB_offset, int lddb, int incb)
{
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static std::map<int, Program*> swpProgs;
static std::map<int, Kernel*> swpKernels;
int device = getActiveDeviceId();
std::call_once(compileFlags[device], [device] () {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName();
if (std::is_same<T, double>::value ||
std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
cl::Program prog;
buildProgram(prog, swapdblk_cl, swapdblk_cl_len, options.str());
swpProgs[device] = new Program(prog);
swpKernels[device] = new Kernel(*swpProgs[device], "swapdblk");
});
int nblocks = n / nb;
if(nblocks == 0)
return;
int info = 0;
if (n < 0) {
info = -1;
} else if (nb < 1 || nb > 1024) {
info = -2;
} else if (ldda < (nblocks-1)*nb*inca + nb) {
info = -4;
} else if (inca < 0) {
info = -5;
} else if (lddb < (nblocks-1)*nb*incb + nb) {
info = -7;
} else if (incb < 0) {
info = -8;
}
if (info != 0) {
AF_ERROR("Invalid configuration", AF_ERR_INTERNAL);
return;
}
NDRange local(nb);
NDRange global(nblocks * nb);
auto swapdOp = make_kernel<int,
cl_mem, unsigned long long, int, int,
cl_mem, unsigned long long, int, int>(*swpKernels[device]);
swapdOp(EnqueueArgs(getQueue(), global, local),
nb,
dA, dA_offset, ldda, inca,
dB, dB_offset, lddb, incb);
}
void mean_first_launcher(Param out, Param owt,
Param in, Param inWeight,
const int threads_x,
const uint groups_x,
const uint groups_y)
{
bool input_weight = ((inWeight.info.dims[0] *
inWeight.info.dims[1] *
inWeight.info.dims[2] *
inWeight.info.dims[3]) != 0);
bool output_weight = (( owt.info.dims[0] *
owt.info.dims[1] *
owt.info.dims[2] *
owt.info.dims[3]) != 0);
std::string ref_name =
std::string("mean_0_") +
std::string(dtype_traits<Ti>::getName()) +
std::string("_") +
std::string(dtype_traits<Tw>::getName()) +
std::string("_") +
std::string(dtype_traits<To>::getName()) +
std::string("_") +
std::to_string(threads_x) +
std::string("_") +
std::to_string(input_weight) +
std::string("_") +
std::to_string(output_weight);
int device = getActiveDeviceId();
kc_entry_t entry = kernelCache(device, ref_name);
if (entry.prog==0 && entry.ker==0) {
Binary<To, af_add_t> mean;
ToNumStr<To> toNumStr;
ToNumStr<Tw> twNumStr;
Transform<uint, Tw, af_add_t> transform_weight;
std::ostringstream options;
options << " -D Ti=" << dtype_traits<Ti>::getName()
<< " -D Tw=" << dtype_traits<Tw>::getName()
<< " -D To=" << dtype_traits<To>::getName()
<< " -D DIMX=" << threads_x
<< " -D THREADS_PER_GROUP=" << THREADS_PER_GROUP
<< " -D init_To=" << toNumStr(mean.init())
<< " -D init_Tw=" << twNumStr(transform_weight(0))
<< " -D one_Tw=" << twNumStr(transform_weight(1));
if (input_weight) { options << " -D INPUT_WEIGHT"; }
if (output_weight) { options << " -D OUTPUT_WEIGHT"; }
if (std::is_same<Ti, double>::value ||
std::is_same<Ti, cdouble>::value ||
std::is_same<To, double>::value) {
options << " -D USE_DOUBLE";
}
const char *ker_strs[] = {mean_ops_cl, mean_first_cl};
const int ker_lens[] = {mean_ops_cl_len, mean_first_cl_len};
Program prog;
buildProgram(prog, 2, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "mean_first_kernel");
addKernelToCache(device, ref_name, entry);
}
NDRange local(threads_x, THREADS_PER_GROUP / threads_x);
NDRange global(groups_x * in.info.dims[2] * local[0],
groups_y * in.info.dims[3] * local[1]);
uint repeat = divup(in.info.dims[0], (local[0] * groups_x));
if (input_weight && output_weight) {
auto meanOp = KernelFunctor<
Buffer, KParam,
Buffer, KParam,
Buffer, KParam,
Buffer, KParam,
uint, uint, uint>(*entry.ker);
meanOp(EnqueueArgs(getQueue(), global, local),
*out.data, out.info,
*owt.data, owt.info,
*in.data, in.info,
*inWeight.data, inWeight.info,
groups_x, groups_y, repeat);
} else if (!input_weight && !output_weight) {
auto meanOp = KernelFunctor<
Buffer, KParam,
Buffer, KParam,
uint, uint, uint>(*entry.ker);
meanOp(EnqueueArgs(getQueue(), global, local),
*out.data, out.info,
*in.data, in.info,
groups_x, groups_y, repeat);
} else if ( input_weight && !output_weight) {
auto meanOp = KernelFunctor<
Buffer, KParam,
Buffer, KParam,
Buffer, KParam,
uint, uint, uint>(*entry.ker);
meanOp(EnqueueArgs(getQueue(), global, local),
*out.data, out.info,
*in.data, in.info,
*inWeight.data, inWeight.info,
groups_x, groups_y, repeat);
} else if (!input_weight && output_weight) {
auto meanOp = KernelFunctor<
Buffer, KParam,
Buffer, KParam,
Buffer, KParam,
uint, uint, uint>(*entry.ker);
meanOp(EnqueueArgs(getQueue(), global, local),
*out.data, out.info,
*owt.data, owt.info,
*in.data, in.info,
groups_x, groups_y, repeat);
}
CL_DEBUG_FINISH(getQueue());
}
void csrmv(Param out,
const Param &values, const Param &rowIdx, const Param &colIdx,
const Param &rhs, const T alpha, const T beta)
{
bool use_alpha = (alpha != scalar<T>(1.0));
bool use_beta = (beta != scalar<T>(0.0));
// Using greedy indexing is causing performance issues on many platforms
// FIXME: Figure out why
bool use_greedy = false;
// FIXME: Find a better number based on average non zeros per row
int threads = 64;
std::string ref_name =
std::string("csrmv_") +
std::string(dtype_traits<T>::getName()) +
std::string("_") +
std::to_string(use_alpha) +
std::string("_") +
std::to_string(use_beta) +
std::string("_") +
std::to_string(use_greedy) +
std::string("_") +
std::to_string(threads);
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();
options << " -D USE_ALPHA=" << use_alpha;
options << " -D USE_BETA=" << use_beta;
options << " -D USE_GREEDY=" << use_greedy;
options << " -D THREADS=" << threads;
if (std::is_same<T, double>::value ||
std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
if (std::is_same<T, cfloat>::value ||
std::is_same<T, cdouble>::value) {
options << " -D IS_CPLX=1";
} else {
options << " -D IS_CPLX=0";
}
const char *ker_strs[] = {csrmv_cl};
const int ker_lens[] = {csrmv_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel[2];
entry.ker[0] = Kernel(*entry.prog, "csrmv_thread");
entry.ker[1] = Kernel(*entry.prog, "csrmv_block");
addKernelToCache(device, ref_name, entry);
}
int count = 0;
cl::Buffer *counter = bufferAlloc(sizeof(int));
getQueue().enqueueWriteBuffer(*counter, CL_TRUE,
0,
sizeof(int),
(void *)&count);
// TODO: Figure out the proper way to choose either csrmv_thread or csrmv_block
bool is_csrmv_block = true;
auto csrmv_kernel = is_csrmv_block ? entry.ker[1] : entry.ker[0];
auto csrmv_func = KernelFunctor<Buffer,
Buffer, Buffer, Buffer,
int,
Buffer, KParam, T, T, Buffer>(csrmv_kernel);
NDRange local(is_csrmv_block ? threads : THREADS_PER_GROUP, 1);
int M = rowIdx.info.dims[0] - 1;
int groups_x = is_csrmv_block ? divup(M, REPEAT) : divup(M, REPEAT * local[0]);
groups_x = std::min(groups_x, MAX_CSRMV_GROUPS);
NDRange global(local[0] * groups_x, 1);
csrmv_func(EnqueueArgs(getQueue(), global, local),
*out.data, *values.data, *rowIdx.data, *colIdx.data,
M, *rhs.data, rhs.info, alpha, beta, *counter);
CL_DEBUG_FINISH(getQueue());
bufferFree(counter);
}
void swapdblk(int n, int nb, cl_mem dA, size_t dA_offset, int ldda, int inca,
cl_mem dB, size_t dB_offset, int lddb, int incb,
cl_command_queue queue) {
std::string refName =
std::string("swapdblk_") + std::string(dtype_traits<T>::getName());
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();
if (std::is_same<T, double>::value || std::is_same<T, cdouble>::value)
options << " -D USE_DOUBLE";
const char* ker_strs[] = {swapdblk_cl};
const int ker_lens[] = {swapdblk_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "swapdblk");
addKernelToCache(device, refName, entry);
}
int nblocks = n / nb;
if (nblocks == 0) return;
int info = 0;
if (n < 0) {
info = -1;
} else if (nb < 1 || nb > 1024) {
info = -2;
} else if (ldda < (nblocks - 1) * nb * inca + nb) {
info = -4;
} else if (inca < 0) {
info = -5;
} else if (lddb < (nblocks - 1) * nb * incb + nb) {
info = -7;
} else if (incb < 0) {
info = -8;
}
if (info != 0) {
AF_ERROR("Invalid configuration", AF_ERR_INTERNAL);
return;
}
NDRange local(nb);
NDRange global(nblocks * nb);
cl::Buffer dAObj(dA, true);
cl::Buffer dBObj(dB, true);
auto swapdOp =
KernelFunctor<int, Buffer, unsigned long long, int, int, Buffer,
unsigned long long, int, int>(*entry.ker);
cl::CommandQueue q(queue);
swapdOp(EnqueueArgs(q, global, local), nb, dAObj, dA_offset, ldda, inca,
dBObj, dB_offset, lddb, incb);
}
void complexMultiplyHelper(Param packed,
Param sig,
Param filter,
const int baseDim,
ConvolveBatchKind kind)
{
try {
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static std::map<int, Program*> fftconvolveProgs;
static std::map<int, Kernel*> cmKernel;
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D ONE2ONE=" << (int)ONE2ONE
<< " -D MANY2ONE=" << (int)MANY2ONE
<< " -D ONE2MANY=" << (int)ONE2MANY
<< " -D MANY2MANY=" << (int)MANY2MANY;
if ((af_dtype) dtype_traits<convT>::af_type == c32) {
options << " -D CONVT=float";
}
else if ((af_dtype) dtype_traits<convT>::af_type == c64 && isDouble) {
options << " -D CONVT=double"
<< " -D USE_DOUBLE";
}
cl::Program prog;
buildProgram(prog,
fftconvolve_multiply_cl,
fftconvolve_multiply_cl_len,
options.str());
fftconvolveProgs[device] = new Program(prog);
cmKernel[device] = new Kernel(*fftconvolveProgs[device], "complex_multiply");
});
Param sig_tmp, filter_tmp;
calcParamSizes(sig_tmp, filter_tmp, packed, sig, filter, baseDim, kind);
int sig_packed_elem = sig_tmp.info.strides[3] * sig_tmp.info.dims[3];
int filter_packed_elem = filter_tmp.info.strides[3] * filter_tmp.info.dims[3];
int mul_elem = (sig_packed_elem < filter_packed_elem) ?
filter_packed_elem : sig_packed_elem;
int blocks = divup(mul_elem, THREADS);
NDRange local(THREADS);
NDRange global(blocks * THREADS);
// Multiply filter and signal FFT arrays
auto cmOp = make_kernel<Buffer, KParam,
Buffer, KParam,
Buffer, KParam,
const int, const int> (*cmKernel[device]);
cmOp(EnqueueArgs(getQueue(), global, local),
*packed.data, packed.info,
*sig_tmp.data, sig_tmp.info,
*filter_tmp.data, filter_tmp.info,
mul_elem, (int)kind);
CL_DEBUG_FINISH(getQueue());
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
void nearest_neighbour(Param idx,
Param dist,
Param query,
Param train,
const dim_t dist_dim,
const unsigned n_dist)
{
try {
const unsigned feat_len = query.info.dims[dist_dim];
const To max_dist = maxval<To>();
// Determine maximum feat_len capable of using shared memory (faster)
cl_ulong avail_lmem = getDevice().getInfo<CL_DEVICE_LOCAL_MEM_SIZE>();
size_t lmem_predef = 2 * THREADS * sizeof(unsigned) + feat_len * sizeof(T);
size_t ltrain_sz = THREADS * feat_len * sizeof(T);
bool use_lmem = (avail_lmem >= (lmem_predef + ltrain_sz)) ? true : false;
size_t lmem_sz = (use_lmem) ? lmem_predef + ltrain_sz : lmem_predef;
unsigned unroll_len = nextpow2(feat_len);
if (unroll_len != feat_len) unroll_len = 0;
std::string ref_name =
std::string("knn_") +
std::to_string(dist_type) +
std::string("_") +
std::to_string(use_lmem) +
std::string("_") +
std::string(dtype_traits<T>::getName()) +
std::string("_") +
std::to_string(unroll_len);
int device = getActiveDeviceId();
kc_t::iterator cache_idx = kernelCaches[device].find(ref_name);
kc_entry_t entry;
if (cache_idx == kernelCaches[device].end()) {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D To=" << dtype_traits<To>::getName()
<< " -D THREADS=" << THREADS
<< " -D FEAT_LEN=" << unroll_len;
switch(dist_type) {
case AF_SAD: options <<" -D DISTOP=_sad_"; break;
case AF_SSD: options <<" -D DISTOP=_ssd_"; break;
case AF_SHD: options <<" -D DISTOP=_shd_ -D __SHD__";
break;
default: break;
}
if (std::is_same<T, double>::value ||
std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
if (use_lmem)
options << " -D USE_LOCAL_MEM";
cl::Program prog;
buildProgram(prog,
nearest_neighbour_cl,
nearest_neighbour_cl_len,
options.str());
entry.prog = new Program(prog);
entry.ker = new Kernel[3];
entry.ker[0] = Kernel(*entry.prog, "nearest_neighbour_unroll");
entry.ker[1] = Kernel(*entry.prog, "nearest_neighbour");
entry.ker[2] = Kernel(*entry.prog, "select_matches");
kernelCaches[device][ref_name] = entry;
} else {
entry = cache_idx->second;
}
const dim_t sample_dim = (dist_dim == 0) ? 1 : 0;
const unsigned nquery = query.info.dims[sample_dim];
const unsigned ntrain = train.info.dims[sample_dim];
unsigned nblk = divup(ntrain, THREADS);
const NDRange local(THREADS, 1);
const NDRange global(nblk * THREADS, 1);
cl::Buffer *d_blk_idx = bufferAlloc(nblk * nquery * sizeof(unsigned));
cl::Buffer *d_blk_dist = bufferAlloc(nblk * nquery * sizeof(To));
// For each query vector, find training vector with smallest Hamming
// distance per CUDA block
if (unroll_len > 0) {
auto huOp = KernelFunctor<Buffer, Buffer,
Buffer, KParam,
Buffer, KParam,
const To,
LocalSpaceArg> (entry.ker[0]);
huOp(EnqueueArgs(getQueue(), global, local),
*d_blk_idx, *d_blk_dist,
*query.data, query.info, *train.data, train.info,
max_dist, cl::Local(lmem_sz));
}
else {
auto hmOp = KernelFunctor<Buffer, Buffer,
Buffer, KParam,
Buffer, KParam,
const To, const unsigned,
LocalSpaceArg> (entry.ker[1]);
hmOp(EnqueueArgs(getQueue(), global, local),
*d_blk_idx, *d_blk_dist,
*query.data, query.info, *train.data, train.info,
max_dist, feat_len, cl::Local(lmem_sz));
}
CL_DEBUG_FINISH(getQueue());
const NDRange local_sm(32, 8);
const NDRange global_sm(divup(nquery, 32) * 32, 8);
// Reduce all smallest Hamming distances from each block and store final
// best match
auto smOp = KernelFunctor<Buffer, Buffer, Buffer, Buffer,
const unsigned, const unsigned,
const To> (entry.ker[2]);
smOp(EnqueueArgs(getQueue(), global_sm, local_sm),
*idx.data, *dist.data,
*d_blk_idx, *d_blk_dist,
nquery, nblk, max_dist);
CL_DEBUG_FINISH(getQueue());
bufferFree(d_blk_idx);
bufferFree(d_blk_dist);
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
void packDataHelper(Param packed,
Param sig,
Param filter,
const int baseDim,
ConvolveBatchKind kind)
{
try {
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static std::map<int, Program*> fftconvolveProgs;
static std::map<int, Kernel*> pdKernel;
static std::map<int, Kernel*> paKernel;
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName();
if ((af_dtype) dtype_traits<convT>::af_type == c32) {
options << " -D CONVT=float";
}
else if ((af_dtype) dtype_traits<convT>::af_type == c64 && isDouble) {
options << " -D CONVT=double"
<< " -D USE_DOUBLE";
}
cl::Program prog;
buildProgram(prog, fftconvolve_pack_cl, fftconvolve_pack_cl_len, options.str());
fftconvolveProgs[device] = new Program(prog);
pdKernel[device] = new Kernel(*fftconvolveProgs[device], "pack_data");
paKernel[device] = new Kernel(*fftconvolveProgs[device], "pad_array");
});
Param sig_tmp, filter_tmp;
calcParamSizes(sig_tmp, filter_tmp, packed, sig, filter, baseDim, kind);
int sig_packed_elem = sig_tmp.info.strides[3] * sig_tmp.info.dims[3];
int filter_packed_elem = filter_tmp.info.strides[3] * filter_tmp.info.dims[3];
// Number of packed complex elements in dimension 0
int sig_half_d0 = divup(sig.info.dims[0], 2);
int sig_half_d0_odd = sig.info.dims[0] % 2;
int blocks = divup(sig_packed_elem, THREADS);
// Locate features kernel sizes
NDRange local(THREADS);
NDRange global(blocks * THREADS);
// Pack signal in a complex matrix where first dimension is half the input
// (allows faster FFT computation) and pad array to a power of 2 with 0s
auto pdOp = make_kernel<Buffer, KParam,
Buffer, KParam,
const int, const int> (*pdKernel[device]);
pdOp(EnqueueArgs(getQueue(), global, local),
*sig_tmp.data, sig_tmp.info, *sig.data, sig.info,
sig_half_d0, sig_half_d0_odd);
CL_DEBUG_FINISH(getQueue());
blocks = divup(filter_packed_elem, THREADS);
global = NDRange(blocks * THREADS);
// Pad filter array with 0s
auto paOp = make_kernel<Buffer, KParam,
Buffer, KParam> (*paKernel[device]);
paOp(EnqueueArgs(getQueue(), global, local),
*filter_tmp.data, filter_tmp.info,
*filter.data, filter.info);
CL_DEBUG_FINISH(getQueue());
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
void morph(Param out,
const Param in,
const Param mask)
{
try {
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static std::map<int, Program*> morProgs;
static std::map<int, Kernel*> morKernels;
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
ToNumStr<T> toNumStr;
T init = isDilation ? Binary<T, af_max_t>().init() : Binary<T, af_min_t>().init();
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D isDilation="<< isDilation
<< " -D init=" << toNumStr(init)
<< " -D windLen=" << windLen;
if (std::is_same<T, double>::value ||
std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
Program prog;
buildProgram(prog, morph_cl, morph_cl_len, options.str());
morProgs[device] = new Program(prog);
morKernels[device] = new Kernel(*morProgs[device], "morph");
});
auto morphOp = KernelFunctor<Buffer, KParam,
Buffer, KParam,
Buffer, cl::LocalSpaceArg,
int, int
>(*morKernels[device]);
NDRange local(THREADS_X, THREADS_Y);
int blk_x = divup(in.info.dims[0], THREADS_X);
int blk_y = divup(in.info.dims[1], THREADS_Y);
// launch batch * blk_x blocks along x dimension
NDRange global(blk_x * THREADS_X * in.info.dims[2],
blk_y * THREADS_Y * in.info.dims[3]);
// copy mask/filter to constant memory
cl_int se_size = sizeof(T)*windLen*windLen;
cl::Buffer *mBuff = bufferAlloc(se_size);
getQueue().enqueueCopyBuffer(*mask.data, *mBuff, 0, 0, se_size);
// calculate shared memory size
const int halo = windLen/2;
const int padding = 2*halo;
const int locLen = THREADS_X + padding + 1;
const int locSize = locLen * (THREADS_Y+padding);
morphOp(EnqueueArgs(getQueue(), global, local),
*out.data, out.info, *in.data, in.info, *mBuff,
cl::Local(locSize*sizeof(T)), blk_x, blk_y);
bufferFree(mBuff);
CL_DEBUG_FINISH(getQueue());
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
void orb(unsigned* out_feat,
Param& x_out,
Param& y_out,
Param& score_out,
Param& ori_out,
Param& size_out,
Param& desc_out,
Param image,
const float fast_thr,
const unsigned max_feat,
const float scl_fctr,
const unsigned levels,
const bool blur_img)
{
try {
static std::once_flag compileFlags[DeviceManager::MAX_DEVICES];
static Program orbProgs[DeviceManager::MAX_DEVICES];
static Kernel hrKernel[DeviceManager::MAX_DEVICES];
static Kernel kfKernel[DeviceManager::MAX_DEVICES];
static Kernel caKernel[DeviceManager::MAX_DEVICES];
static Kernel eoKernel[DeviceManager::MAX_DEVICES];
int device = getActiveDeviceId();
std::call_once( compileFlags[device], [device] () {
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D BLOCK_SIZE=" << ORB_THREADS_X;
if (std::is_same<T, double>::value ||
std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
buildProgram(orbProgs[device],
orb_cl,
orb_cl_len,
options.str());
hrKernel[device] = Kernel(orbProgs[device], "harris_response");
kfKernel[device] = Kernel(orbProgs[device], "keep_features");
caKernel[device] = Kernel(orbProgs[device], "centroid_angle");
eoKernel[device] = Kernel(orbProgs[device], "extract_orb");
});
unsigned patch_size = REF_PAT_SIZE;
unsigned min_side = std::min(image.info.dims[0], image.info.dims[1]);
unsigned max_levels = 0;
float scl_sum = 0.f;
for (unsigned i = 0; i < levels; i++) {
min_side /= scl_fctr;
// Minimum image side for a descriptor to be computed
if (min_side < patch_size || max_levels == levels) break;
max_levels++;
scl_sum += 1.f / (float)pow(scl_fctr,(float)i);
}
std::vector<cl::Buffer*> d_x_pyr(max_levels);
std::vector<cl::Buffer*> d_y_pyr(max_levels);
std::vector<cl::Buffer*> d_score_pyr(max_levels);
std::vector<cl::Buffer*> d_ori_pyr(max_levels);
std::vector<cl::Buffer*> d_size_pyr(max_levels);
std::vector<cl::Buffer*> d_desc_pyr(max_levels);
std::vector<unsigned> feat_pyr(max_levels);
unsigned total_feat = 0;
// Compute number of features to keep for each level
std::vector<unsigned> lvl_best(max_levels);
unsigned feat_sum = 0;
for (unsigned i = 0; i < max_levels-1; i++) {
float lvl_scl = (float)pow(scl_fctr,(float)i);
lvl_best[i] = ceil((max_feat / scl_sum) / lvl_scl);
feat_sum += lvl_best[i];
}
lvl_best[max_levels-1] = max_feat - feat_sum;
// Maintain a reference to previous level image
Param prev_img;
Param lvl_img;
const unsigned gauss_len = 9;
T* h_gauss = nullptr;
Param gauss_filter;
gauss_filter.data = nullptr;
for (unsigned i = 0; i < max_levels; i++) {
const float lvl_scl = (float)pow(scl_fctr,(float)i);
if (i == 0) {
// First level is used in its original size
lvl_img = image;
prev_img = image;
}
else if (i > 0) {
// Resize previous level image to current level dimensions
lvl_img.info.dims[0] = round(image.info.dims[0] / lvl_scl);
lvl_img.info.dims[1] = round(image.info.dims[1] / lvl_scl);
lvl_img.info.strides[0] = 1;
lvl_img.info.strides[1] = lvl_img.info.dims[0];
for (int k = 2; k < 4; k++) {
lvl_img.info.dims[k] = 1;
lvl_img.info.strides[k] = lvl_img.info.dims[k - 1] * lvl_img.info.strides[k - 1];
}
lvl_img.info.offset = 0;
lvl_img.data = bufferAlloc(lvl_img.info.dims[3] * lvl_img.info.strides[3] * sizeof(T));
resize<T, AF_INTERP_BILINEAR>(lvl_img, prev_img);
if (i > 1)
bufferFree(prev_img.data);
prev_img = lvl_img;
}
unsigned lvl_feat = 0;
Param d_x_feat, d_y_feat, d_score_feat;
// Round feature size to nearest odd integer
float size = 2.f * floor(patch_size / 2.f) + 1.f;
// Avoid keeping features that might be too wide and might not fit on
// the image, sqrt(2.f) is the radius when angle is 45 degrees and
// represents widest case possible
unsigned edge = ceil(size * sqrt(2.f) / 2.f);
// Detect FAST features
fast<T, 9, true>(&lvl_feat, d_x_feat, d_y_feat, d_score_feat,
lvl_img, fast_thr, 0.15f, edge);
if (lvl_feat == 0) {
feat_pyr[i] = 0;
if (i > 0 && i == max_levels-1)
bufferFree(lvl_img.data);
continue;
}
bufferFree(d_score_feat.data);
unsigned usable_feat = 0;
cl::Buffer* d_usable_feat = bufferAlloc(sizeof(unsigned));
getQueue().enqueueWriteBuffer(*d_usable_feat, CL_TRUE, 0, sizeof(unsigned), &usable_feat);
cl::Buffer* d_x_harris = bufferAlloc(lvl_feat * sizeof(float));
cl::Buffer* d_y_harris = bufferAlloc(lvl_feat * sizeof(float));
cl::Buffer* d_score_harris = bufferAlloc(lvl_feat * sizeof(float));
// Calculate Harris responses
// Good block_size >= 7 (must be an odd number)
const dim_type blk_x = divup(lvl_feat, ORB_THREADS_X);
const NDRange local(ORB_THREADS_X, ORB_THREADS_Y);
const NDRange global(blk_x * ORB_THREADS_X, ORB_THREADS_Y);
unsigned block_size = 7;
float k_thr = 0.04f;
auto hrOp = make_kernel<Buffer, Buffer, Buffer,
Buffer, Buffer, const unsigned,
Buffer, Buffer, KParam,
const unsigned, const float, const unsigned> (hrKernel[device]);
hrOp(EnqueueArgs(getQueue(), global, local),
*d_x_harris, *d_y_harris, *d_score_harris,
*d_x_feat.data, *d_y_feat.data, lvl_feat,
*d_usable_feat, *lvl_img.data, lvl_img.info,
block_size, k_thr, patch_size);
CL_DEBUG_FINISH(getQueue());
getQueue().enqueueReadBuffer(*d_usable_feat, CL_TRUE, 0, sizeof(unsigned), &usable_feat);
bufferFree(d_x_feat.data);
bufferFree(d_y_feat.data);
bufferFree(d_usable_feat);
if (usable_feat == 0) {
feat_pyr[i] = 0;
bufferFree(d_x_harris);
bufferFree(d_y_harris);
bufferFree(d_score_harris);
if (i > 0 && i == max_levels-1)
bufferFree(lvl_img.data);
continue;
}
// Sort features according to Harris responses
Param d_harris_sorted;
Param d_harris_idx;
d_harris_sorted.info.dims[0] = usable_feat;
d_harris_idx.info.dims[0] = usable_feat;
d_harris_sorted.info.strides[0] = 1;
d_harris_idx.info.strides[0] = 1;
for (int k = 1; k < 4; k++) {
d_harris_sorted.info.dims[k] = 1;
d_harris_idx.info.dims[k] = 1;
d_harris_sorted.info.strides[k] = d_harris_sorted.info.dims[k - 1] * d_harris_sorted.info.strides[k - 1];
d_harris_idx.info.strides[k] = d_harris_idx.info.dims[k - 1] * d_harris_idx.info.strides[k - 1];
}
d_harris_sorted.info.offset = 0;
d_harris_idx.info.offset = 0;
d_harris_sorted.data = d_score_harris;
d_harris_idx.data = bufferAlloc((d_harris_idx.info.dims[0]) * sizeof(unsigned));
sort0_index<float, false>(d_harris_sorted, d_harris_idx);
cl::Buffer* d_x_lvl = bufferAlloc(usable_feat * sizeof(float));
cl::Buffer* d_y_lvl = bufferAlloc(usable_feat * sizeof(float));
cl::Buffer* d_score_lvl = bufferAlloc(usable_feat * sizeof(float));
usable_feat = min(usable_feat, lvl_best[i]);
// Keep only features with higher Harris responses
const dim_type keep_blk = divup(usable_feat, ORB_THREADS);
const NDRange local_keep(ORB_THREADS, 1);
const NDRange global_keep(keep_blk * ORB_THREADS, 1);
auto kfOp = make_kernel<Buffer, Buffer, Buffer,
Buffer, Buffer, Buffer, Buffer,
const unsigned> (kfKernel[device]);
kfOp(EnqueueArgs(getQueue(), global_keep, local_keep),
*d_x_lvl, *d_y_lvl, *d_score_lvl,
*d_x_harris, *d_y_harris, *d_harris_sorted.data, *d_harris_idx.data,
usable_feat);
CL_DEBUG_FINISH(getQueue());
bufferFree(d_x_harris);
bufferFree(d_y_harris);
bufferFree(d_harris_sorted.data);
bufferFree(d_harris_idx.data);
cl::Buffer* d_ori_lvl = bufferAlloc(usable_feat * sizeof(float));
cl::Buffer* d_size_lvl = bufferAlloc(usable_feat * sizeof(float));
// Compute orientation of features
const dim_type centroid_blk_x = divup(usable_feat, ORB_THREADS_X);
const NDRange local_centroid(ORB_THREADS_X, ORB_THREADS_Y);
const NDRange global_centroid(centroid_blk_x * ORB_THREADS_X, ORB_THREADS_Y);
auto caOp = make_kernel<Buffer, Buffer, Buffer,
const unsigned, Buffer, KParam,
const unsigned> (caKernel[device]);
caOp(EnqueueArgs(getQueue(), global_centroid, local_centroid),
*d_x_lvl, *d_y_lvl, *d_ori_lvl,
usable_feat, *lvl_img.data, lvl_img.info,
patch_size);
CL_DEBUG_FINISH(getQueue());
Param lvl_filt;
Param lvl_tmp;
if (blur_img) {
lvl_filt = lvl_img;
lvl_tmp = lvl_img;
lvl_filt.data = bufferAlloc(lvl_filt.info.dims[0] * lvl_filt.info.dims[1] * sizeof(T));
lvl_tmp.data = bufferAlloc(lvl_tmp.info.dims[0] * lvl_tmp.info.dims[1] * sizeof(T));
// Calculate a separable Gaussian kernel
if (h_gauss == nullptr) {
h_gauss = new T[gauss_len];
gaussian1D(h_gauss, gauss_len, 2.f);
gauss_filter.info.dims[0] = gauss_len;
gauss_filter.info.strides[0] = 1;
for (int k = 1; k < 4; k++) {
gauss_filter.info.dims[k] = 1;
gauss_filter.info.strides[k] = gauss_filter.info.dims[k - 1] * gauss_filter.info.strides[k - 1];
}
dim_type gauss_elem = gauss_filter.info.strides[3] * gauss_filter.info.dims[3];
gauss_filter.data = bufferAlloc(gauss_elem * sizeof(T));
getQueue().enqueueWriteBuffer(*gauss_filter.data, CL_TRUE, 0, gauss_elem * sizeof(T), h_gauss);
}
// Filter level image with Gaussian kernel to reduce noise sensitivity
convolve2<T, convAccT, 0, false, gauss_len>(lvl_tmp, lvl_img, gauss_filter);
convolve2<T, convAccT, 1, false, gauss_len>(lvl_filt, lvl_tmp, gauss_filter);
bufferFree(lvl_tmp.data);
}
// Compute ORB descriptors
cl::Buffer* d_desc_lvl = bufferAlloc(usable_feat * 8 * sizeof(unsigned));
unsigned* h_desc_lvl = new unsigned[usable_feat * 8];
for (int j = 0; j < (int)usable_feat * 8; j++)
h_desc_lvl[j] = 0;
getQueue().enqueueWriteBuffer(*d_desc_lvl, CL_TRUE, 0, usable_feat * 8 * sizeof(unsigned), h_desc_lvl);
delete[] h_desc_lvl;
auto eoOp = make_kernel<Buffer, const unsigned,
Buffer, Buffer, Buffer, Buffer,
Buffer, KParam,
const float, const unsigned> (eoKernel[device]);
if (blur_img) {
eoOp(EnqueueArgs(getQueue(), global_centroid, local_centroid),
*d_desc_lvl, usable_feat,
*d_x_lvl, *d_y_lvl, *d_ori_lvl, *d_size_lvl,
*lvl_filt.data, lvl_filt.info,
lvl_scl, patch_size);
CL_DEBUG_FINISH(getQueue());
bufferFree(lvl_filt.data);
}
else {
eoOp(EnqueueArgs(getQueue(), global_centroid, local_centroid),
*d_desc_lvl, usable_feat,
*d_x_lvl, *d_y_lvl, *d_ori_lvl, *d_size_lvl,
*lvl_img.data, lvl_img.info,
lvl_scl, patch_size);
CL_DEBUG_FINISH(getQueue());
}
// Store results to pyramids
total_feat += usable_feat;
feat_pyr[i] = usable_feat;
d_x_pyr[i] = d_x_lvl;
d_y_pyr[i] = d_y_lvl;
d_score_pyr[i] = d_score_lvl;
d_ori_pyr[i] = d_ori_lvl;
d_size_pyr[i] = d_size_lvl;
d_desc_pyr[i] = d_desc_lvl;
if (i > 0 && i == max_levels-1)
bufferFree(lvl_img.data);
}
if (gauss_filter.data != nullptr)
bufferFree(gauss_filter.data);
if (h_gauss != nullptr)
delete[] h_gauss;
// If no features are found, set found features to 0 and return
if (total_feat == 0) {
*out_feat = 0;
return;
}
// Allocate output memory
x_out.info.dims[0] = total_feat;
x_out.info.strides[0] = 1;
y_out.info.dims[0] = total_feat;
y_out.info.strides[0] = 1;
score_out.info.dims[0] = total_feat;
score_out.info.strides[0] = 1;
ori_out.info.dims[0] = total_feat;
ori_out.info.strides[0] = 1;
size_out.info.dims[0] = total_feat;
size_out.info.strides[0] = 1;
desc_out.info.dims[0] = 8;
desc_out.info.strides[0] = 1;
desc_out.info.dims[1] = total_feat;
desc_out.info.strides[1] = desc_out.info.dims[0];
for (int k = 1; k < 4; k++) {
x_out.info.dims[k] = 1;
x_out.info.strides[k] = x_out.info.dims[k - 1] * x_out.info.strides[k - 1];
y_out.info.dims[k] = 1;
y_out.info.strides[k] = y_out.info.dims[k - 1] * y_out.info.strides[k - 1];
score_out.info.dims[k] = 1;
score_out.info.strides[k] = score_out.info.dims[k - 1] * score_out.info.strides[k - 1];
ori_out.info.dims[k] = 1;
ori_out.info.strides[k] = ori_out.info.dims[k - 1] * ori_out.info.strides[k - 1];
size_out.info.dims[k] = 1;
size_out.info.strides[k] = size_out.info.dims[k - 1] * size_out.info.strides[k - 1];
if (k > 1) {
desc_out.info.dims[k] = 1;
desc_out.info.strides[k] = desc_out.info.dims[k - 1] * desc_out.info.strides[k - 1];
}
}
if (total_feat > 0) {
size_t out_sz = total_feat * sizeof(float);
x_out.data = bufferAlloc(out_sz);
y_out.data = bufferAlloc(out_sz);
score_out.data = bufferAlloc(out_sz);
ori_out.data = bufferAlloc(out_sz);
size_out.data = bufferAlloc(out_sz);
size_t desc_sz = total_feat * 8 * sizeof(unsigned);
desc_out.data = bufferAlloc(desc_sz);
}
unsigned offset = 0;
for (unsigned i = 0; i < max_levels; i++) {
if (feat_pyr[i] == 0)
continue;
if (i > 0)
offset += feat_pyr[i-1];
getQueue().enqueueCopyBuffer(*d_x_pyr[i], *x_out.data, 0, offset*sizeof(float), feat_pyr[i] * sizeof(float));
getQueue().enqueueCopyBuffer(*d_y_pyr[i], *y_out.data, 0, offset*sizeof(float), feat_pyr[i] * sizeof(float));
getQueue().enqueueCopyBuffer(*d_score_pyr[i], *score_out.data, 0, offset*sizeof(float), feat_pyr[i] * sizeof(float));
getQueue().enqueueCopyBuffer(*d_ori_pyr[i], *ori_out.data, 0, offset*sizeof(float), feat_pyr[i] * sizeof(float));
getQueue().enqueueCopyBuffer(*d_size_pyr[i], *size_out.data, 0, offset*sizeof(float), feat_pyr[i] * sizeof(float));
getQueue().enqueueCopyBuffer(*d_desc_pyr[i], *desc_out.data, 0, offset*8*sizeof(unsigned), feat_pyr[i] * 8 * sizeof(unsigned));
bufferFree(d_x_pyr[i]);
bufferFree(d_y_pyr[i]);
bufferFree(d_score_pyr[i]);
bufferFree(d_ori_pyr[i]);
bufferFree(d_size_pyr[i]);
bufferFree(d_desc_pyr[i]);
}
// Sets number of output features
*out_feat = total_feat;
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
void conv2Helper(const conv_kparam_t& param, Param out, const Param signal, const Param filter)
{
try {
int f0 = filter.info.dims[0];
int f1 = filter.info.dims[1];
std::string ref_name =
std::string("conv2_") +
std::string(dtype_traits<T>::getName()) +
std::string("_") +
std::string(dtype_traits<aT>::getName()) +
std::string("_") +
std::to_string(expand) +
std::string("_") +
std::to_string(f0) +
std::string("_") +
std::to_string(f1);
int device = getActiveDeviceId();
kc_t::iterator idx = kernelCaches[device].find(ref_name);
kc_entry_t entry;
if (idx == kernelCaches[device].end()) {
size_t LOC_SIZE = (THREADS_X+2*(f0-1))*(THREADS_Y+2*(f1-1));
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D accType="<< dtype_traits<aT>::getName()
<< " -D BASE_DIM="<< 2 /* hard constant specific to this convolution type */
<< " -D FLEN0=" << f0
<< " -D FLEN1=" << f1
<< " -D EXPAND="<< expand
<< " -D C_SIZE="<< LOC_SIZE;
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());
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "convolve");
kernelCaches[device][ref_name] = entry;
} else {
entry = idx->second;
}
auto convOp = cl::KernelFunctor<Buffer, KParam, Buffer, KParam,
Buffer, KParam, int, int,
int, int,
int, int
>(*entry.ker);
convOp(EnqueueArgs(getQueue(), param.global, param.local),
*out.data, out.info, *signal.data, signal.info,
*param.impulse, filter.info, param.nBBS0, param.nBBS1,
param.o[1], param.o[2], param.s[1], param.s[2]);
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
