void vkeGameRendererDynamic::setNodeData(VkeNodeData::List *inData){
m_node_data = inData;
if (m_node_data != NULL){
uint32_t cnt = m_node_data->count();
uint32_t transformsSize = 64 * 64;
uint32_t sz = sizeof(VkeNodeUniform) * 100;
sz += (transformsSize);
VkBufferUsageFlags usageFlags = VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT;
bufferCreate(&m_uniforms_buffer_staging, sz, (VkBufferUsageFlagBits)usageFlags);
bufferAlloc(&m_uniforms_buffer_staging, &m_uniforms_staging, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT);
bufferCreate(&m_uniforms_buffer, sz, VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT);
bufferAlloc(&m_uniforms_buffer, &m_uniforms_memory, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
m_uniforms_descriptor.buffer = m_uniforms_buffer;
m_uniforms_descriptor.offset = 0;
m_uniforms_descriptor.range = sizeof(VkeNodeUniform) * 100;
m_transforms_descriptor.buffer = m_uniforms_buffer;
m_transforms_descriptor.offset = sizeof(VkeNodeUniform) * 100;
m_transforms_descriptor.range = transformsSize; //(4 * 64);
}
}
void mean_first(Param out, Param in, Param inWeight)
{
uint threads_x = nextpow2(std::max(32u, (uint)in.info.dims[0]));
threads_x = std::min(threads_x, THREADS_PER_GROUP);
uint threads_y = THREADS_PER_GROUP / threads_x;
uint groups_x = divup(in.info.dims[0], threads_x * REPEAT);
uint groups_y = divup(in.info.dims[1], threads_y);
Param tmpOut = out;
Param noWeight;
noWeight.info.offset = 0;
for (int k = 0; k < 4; ++k) {
noWeight.info.dims[k] = 0;
noWeight.info.strides[k] = 0;
}
// Does not matter what the value is it will not be used. Just needs to be valid.
noWeight.data = inWeight.data;
Param tmpWeight = noWeight;
if (groups_x > 1) {
tmpOut.data = bufferAlloc(groups_x *
in.info.dims[1] *
in.info.dims[2] *
in.info.dims[3] *
sizeof(To));
tmpWeight.data = bufferAlloc(groups_x *
in.info.dims[1] *
in.info.dims[2] *
in.info.dims[3] *
sizeof(Tw));
tmpOut.info.dims[0] = groups_x;
for (int k = 1; k < 4; k++) tmpOut.info.strides[k] *= groups_x;
tmpWeight.info = tmpOut.info;
}
mean_first_launcher<Ti, Tw, To>(tmpOut, tmpWeight, in, inWeight, threads_x, groups_x, groups_y);
if (groups_x > 1) {
// No Weight is needed when writing out the output.
mean_first_launcher<Ti, Tw, To>(out, noWeight, tmpOut, tmpWeight, threads_x, 1, groups_y);
bufferFree(tmpOut.data);
bufferFree(tmpWeight.data);
}
}
void vkeGameRendererDynamic::setNodeData(VkeNodeData::List *inData){
m_node_data = inData;
if (m_node_data != NULL){
uint32_t cnt = m_node_data->count();
uint32_t transformsSize = 64 * m_instance_count;
uint32_t sz = sizeof(VkeNodeUniform) * cnt;
sz += (transformsSize);
m_uniforms_local = (float*)malloc( sz);
bufferCreate(&m_uniforms_buffer, sz, VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT);
bufferAlloc(&m_uniforms_buffer, &m_uniforms_memory, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
m_uniforms_descriptor.buffer = m_uniforms_buffer;
m_uniforms_descriptor.offset = 0;
m_uniforms_descriptor.range = sizeof(VkeNodeUniform) * cnt;
m_transforms_descriptor.buffer = m_uniforms_buffer;
m_transforms_descriptor.offset = sizeof(VkeNodeUniform) * cnt;
m_transforms_descriptor.range = transformsSize;
}
}
void vkeGameRendererDynamic::setMaterialData(VkeMaterial::List *inData){
m_materials = inData;
if (m_materials != NULL){
uint32_t cnt = m_materials->count();
uint32_t sz = sizeof(VkeMaterialUniform) * cnt;
VkBufferUsageFlags usageFlags = VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT;
bufferCreate(&m_material_buffer_staging, sz, (VkBufferUsageFlagBits)usageFlags);
bufferAlloc(&m_material_buffer_staging, &m_material_staging, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_CACHED_BIT);
VkeMaterialUniform *uniforms = NULL;
VKA_CHECK_ERROR(vkMapMemory(getDefaultDevice(), m_material_staging, 0, sz, 0, (void **)&uniforms), "Could not map buffer memory.\n");
for (uint32_t i = 0; i < cnt; ++i){
VkeMaterial *mat = m_materials->getMaterial(i);
mat->initVKBufferData(m_material_buffer_staging);
mat->updateVKBufferData(uniforms);
}
vkUnmapMemory(getDefaultDevice(), m_material_staging);
}
}
static void scan_dim(Param &out, const Param &in, int dim)
{
uint threads_y = std::min(THREADS_Y, nextpow2(out.info.dims[dim]));
uint threads_x = THREADS_X;
uint groups_all[] = {divup((uint)out.info.dims[0], threads_x),
(uint)out.info.dims[1],
(uint)out.info.dims[2],
(uint)out.info.dims[3]};
groups_all[dim] = divup(out.info.dims[dim], threads_y * REPEAT);
if (groups_all[dim] == 1) {
scan_dim_launcher<Ti, To, op, inclusive_scan>(out, out, in,
dim, true,
threads_y,
groups_all);
} else {
Param tmp = out;
tmp.info.dims[dim] = groups_all[dim];
tmp.info.strides[0] = 1;
for (int k = 1; k < 4; k++) {
tmp.info.strides[k] = tmp.info.strides[k - 1] * tmp.info.dims[k - 1];
}
int tmp_elements = tmp.info.strides[3] * tmp.info.dims[3];
// FIXME: Do I need to free this ?
tmp.data = bufferAlloc(tmp_elements * sizeof(To));
scan_dim_launcher<Ti, To, op, inclusive_scan>(out, tmp, in,
dim, false,
threads_y,
groups_all);
int gdim = groups_all[dim];
groups_all[dim] = 1;
if (op == af_notzero_t) {
scan_dim_launcher<To, To, af_add_t, true>(tmp, tmp, tmp,
dim, true,
threads_y,
groups_all);
} else {
scan_dim_launcher<To, To, op, true>(tmp, tmp, tmp,
dim, true,
threads_y,
groups_all);
}
groups_all[dim] = gdim;
bcast_dim_launcher<To, To, op, inclusive_scan>(out, tmp,
dim, true,
threads_y,
groups_all);
bufferFree(tmp.data);
}
}
void conv2(conv_kparam_t& p, Param& out, const Param& sig, const Param& filt)
{
size_t se_size = filt.info.dims[0] * filt.info.dims[1] * sizeof(aT);
p.impulse = bufferAlloc(se_size);
int f0Off = filt.info.offset;
for (int b3=0; b3<filt.info.dims[3]; ++b3) {
int f3Off = b3 * filt.info.strides[3];
for (int b2=0; b2<filt.info.dims[2]; ++b2) {
int f2Off = b2 * filt.info.strides[2];
// FIXME: if the filter array is strided, direct copy of symbols
// might cause issues
getQueue().enqueueCopyBuffer(*filt.data, *p.impulse,
(f2Off+f3Off+f0Off)*sizeof(aT),
0, se_size);
p.o[1] = (p.outHasNoOffset ? 0 : b2);
p.o[2] = (p.outHasNoOffset ? 0 : b3);
p.s[1] = (p.inHasNoOffset ? 0 : b2);
p.s[2] = (p.inHasNoOffset ? 0 : b3);
conv2Helper<T, aT, expand>(p, out, sig, filt);
}
}
}
void convolve2(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] () {
const size_t C0_SIZE = (THREADS_X+2*(fLen-1))* THREADS_Y;
const size_t C1_SIZE = (THREADS_Y+2*(fLen-1))* THREADS_X;
size_t locSize = (conv_dim==0 ? C0_SIZE : C1_SIZE);
std::ostringstream options;
options << " -D T=" << dtype_traits<T>::getName()
<< " -D accType="<< dtype_traits<accType>::getName()
<< " -D CONV_DIM="<< conv_dim
<< " -D EXPAND="<< expand
<< " -D FLEN="<< fLen
<< " -D LOCAL_MEM_SIZE="<<locSize;
if (std::is_same<T, double>::value ||
std::is_same<T, cdouble>::value) {
options << " -D USE_DOUBLE";
}
Program prog;
buildProgram(prog, convolve_separable_cl, convolve_separable_cl_len, options.str());
convProgs[device] = new Program(prog);
convKernels[device] = new Kernel(*convProgs[device], "convolve");
});
auto convOp = make_kernel<Buffer, KParam, Buffer, KParam, Buffer,
int, int>(*convKernels[device]);
NDRange local(THREADS_X, THREADS_Y);
int blk_x = divup(out.info.dims[0], THREADS_X);
int blk_y = divup(out.info.dims[1], THREADS_Y);
NDRange global(blk_x*signal.info.dims[2]*THREADS_X,
blk_y*signal.info.dims[3]*THREADS_Y);
cl::Buffer *mBuff = bufferAlloc(fLen*sizeof(accType));
// FIX ME: if the filter array is strided, direct might cause issues
getQueue().enqueueCopyBuffer(*filter.data, *mBuff, 0, 0, fLen*sizeof(accType));
convOp(EnqueueArgs(getQueue(), global, local),
*out.data, out.info, *signal.data, signal.info, *mBuff, blk_x, blk_y);
bufferFree(mBuff);
} catch (cl::Error err) {
CL_TO_AF_ERROR(err);
throw;
}
}
Array<T> index(const Array<T>& in, const af_index_t idxrs[])
{
kernel::IndexKernelParam_t p;
std::vector<af_seq> seqs(4, af_span);
// create seq vector to retrieve output
// dimensions, offsets & offsets
for (dim_t x=0; x<4; ++x) {
if (idxrs[x].isSeq) {
seqs[x] = idxrs[x].idx.seq;
}
}
// retrieve dimensions, strides and offsets
dim4 iDims = in.dims();
dim4 dDims = in.getDataDims();
dim4 oDims = toDims (seqs, iDims);
dim4 iOffs = toOffset(seqs, dDims);
dim4 iStrds= toStride(seqs, dDims);
for (dim_t i=0; i<4; ++i) {
p.isSeq[i] = idxrs[i].isSeq;
p.offs[i] = iOffs[i];
p.strds[i] = iStrds[i];
}
Buffer* bPtrs[4];
std::vector< Array<uint> > idxArrs(4, createEmptyArray<uint>(dim4()));
// look through indexs to read af_array indexs
for (dim_t x=0; x<4; ++x) {
// set index pointers were applicable
if (!p.isSeq[x]) {
idxArrs[x] = castArray<uint>(idxrs[x].idx.arr);
bPtrs[x] = idxArrs[x].get();
// set output array ith dimension value
oDims[x] = idxArrs[x].elements();
}
else {
// alloc an 1-element buffer to avoid OpenCL from failing
bPtrs[x] = bufferAlloc(sizeof(uint));
}
}
Array<T> out = createEmptyArray<T>(oDims);
if(oDims.elements() == 0) { return out; }
kernel::index<T>(out, in, p, bPtrs);
for (dim_t x=0; x<4; ++x) {
if (p.isSeq[x]) bufferFree(bPtrs[x]);
}
return out;
}
void morph3d(Param out,
const Param in,
const Param mask)
{
std::string refName = std::string("morph3d_") +
std::string(dtype_traits<T>::getName()) +
std::to_string(isDilation) + std::to_string(SeLength);
int device = getActiveDeviceId();
kc_entry_t entry = kernelCache(device, refName);
if (entry.prog==0 && entry.ker==0) {
std::string options = generateOptionsString<T, isDilation, SeLength>();
const char* ker_strs[] = {morph_cl};
const int ker_lens[] = {morph_cl_len};
Program prog;
buildProgram(prog, 1, ker_strs, ker_lens, options);
entry.prog = new Program(prog);
entry.ker = new Kernel(*entry.prog, "morph3d");
addKernelToCache(device, refName, entry);
}
auto morphOp = KernelFunctor< Buffer, KParam, Buffer, KParam, Buffer,
cl::LocalSpaceArg, int >(*entry.ker);
NDRange local(CUBE_X, CUBE_Y, CUBE_Z);
int blk_x = divup(in.info.dims[0], CUBE_X);
int blk_y = divup(in.info.dims[1], CUBE_Y);
int blk_z = divup(in.info.dims[2], CUBE_Z);
// launch batch * blk_x blocks along x dimension
NDRange global(blk_x * CUBE_X * in.info.dims[3], blk_y * CUBE_Y, blk_z * CUBE_Z);
// copy mask/filter to constant memory
cl_int se_size = sizeof(T)*SeLength*SeLength*SeLength;
cl::Buffer *mBuff = bufferAlloc(se_size);
getQueue().enqueueCopyBuffer(*mask.data, *mBuff, 0, 0, se_size);
// calculate shared memory size
const int padding = (SeLength%2==0 ? (SeLength-1) : (2*(SeLength/2)));
const int locLen = CUBE_X+padding+1;
const int locArea = locLen *(CUBE_Y+padding);
const int locSize = locArea*(CUBE_Z+padding);
morphOp(EnqueueArgs(getQueue(), global, local),
*out.data, out.info, *in.data, in.info,
*mBuff, cl::Local(locSize*sizeof(T)), blk_x);
bufferFree(mBuff);
CL_DEBUG_FINISH(getQueue());
}
unsigned nonMaximal(cl::Buffer* x_out, cl::Buffer* y_out, cl::Buffer* resp_out,
const unsigned idim0, const unsigned idim1,
const cl::Buffer* resp_in, const unsigned edge,
const unsigned max_corners) {
unsigned corners_found = 0;
std::string refName =
std::string("non_maximal_") + 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 NONMAX";
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, "non_maximal");
addKernelToCache(device, refName, entry);
}
cl::Buffer* d_corners_found = bufferAlloc(sizeof(unsigned));
getQueue().enqueueWriteBuffer(*d_corners_found, CL_TRUE, 0,
sizeof(unsigned), &corners_found);
auto nonMaximalOp =
KernelFunctor<Buffer, Buffer, Buffer, Buffer, unsigned, unsigned,
Buffer, unsigned, 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]);
nonMaximalOp(EnqueueArgs(getQueue(), global, local), *x_out, *y_out,
*resp_out, *d_corners_found, idim0, idim1, *resp_in, edge,
max_corners);
getQueue().enqueueReadBuffer(*d_corners_found, CL_TRUE, 0, sizeof(unsigned),
&corners_found);
bufferFree(d_corners_found);
return corners_found;
}
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 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 VkeCubeTexture::loadTextureFiles(const char **inPath){
bool imagesOK = true;
VKA_INFO_MSG("Loading Cube Texture.\n");
for (uint32_t i = 0; i < 6; ++i){
if (!loadTexture(inPath[i], NULL, NULL, &m_width, &m_height)){
VKA_ERROR_MSG("Error loading texture image.\n");
printf("Texture : %d not available (%s).\n", i, inPath[i]);
return;
}
}
VulkanDC::Device::Queue::Name queueName = "DEFAULT_GRAPHICS_QUEUE";
VulkanDC::Device::Queue::CommandBufferID cmdID = INIT_COMMAND_ID;
VulkanDC *dc = VulkanDC::Get();
VulkanDC::Device *device = dc->getDefaultDevice();
VulkanDC::Device::Queue *queue = device->getQueue(queueName);
VkCommandBuffer cmd = VK_NULL_HANDLE;
queue->beginCommandBuffer(cmdID, &cmd, VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT);
imageCreateAndBind(
&m_data.image,
&m_data.memory,
m_format, VK_IMAGE_TYPE_2D,
m_width, m_height, 1, 6,
VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
(VkImageUsageFlagBits)( VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT ),
VK_IMAGE_TILING_OPTIMAL);
VkBuffer cubeMapBuffer;
VkDeviceMemory cubeMapMem;
bufferCreate(&cubeMapBuffer, m_width*m_height * 4 * 6, VK_BUFFER_USAGE_TRANSFER_SRC_BIT);
bufferAlloc(&cubeMapBuffer, &cubeMapMem, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT);
VkDeviceSize dSize = m_width * m_height * 4;
uint32_t rowPitch = m_width * 4;
if (m_memory_flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT){
imageSetLayoutBarrier(cmdID, queueName, m_data.image, VK_IMAGE_ASPECT_COLOR_BIT, VK_IMAGE_LAYOUT_PREINITIALIZED, VK_IMAGE_LAYOUT_GENERAL);
for (uint32_t i = 0; i < 6; ++i){
void *data = NULL;
VkDeviceSize ofst = dSize*i;
VKA_CHECK_ERROR(vkMapMemory(getDefaultDevice(),cubeMapMem, ofst, dSize, 0, &data), "Could not map memory for image.\n");
if (!loadTexture(inPath[i], (uint8_t**)&data, rowPitch, &m_width, &m_height)){
VKA_ERROR_MSG("Could not load final image.\n");
}
vkUnmapMemory(getDefaultDevice(), cubeMapMem);
}
VkBufferImageCopy biCpyRgn[6];
for (uint32_t k = 0; k < 6; ++k){
VkDeviceSize ofst = dSize*k;
biCpyRgn[k].bufferOffset = ofst;
biCpyRgn[k].bufferImageHeight = 0;
biCpyRgn[k].bufferRowLength = 0;
biCpyRgn[k].imageExtent.width = m_width;
biCpyRgn[k].imageExtent.height = m_height;
biCpyRgn[k].imageExtent.depth = 1;
biCpyRgn[k].imageOffset.x = 0;
biCpyRgn[k].imageOffset.y = 0;
biCpyRgn[k].imageOffset.z = 0;
biCpyRgn[k].imageSubresource.baseArrayLayer = k;
biCpyRgn[k].imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
biCpyRgn[k].imageSubresource.layerCount = 1;
biCpyRgn[k].imageSubresource.mipLevel = 0;
}
VkFence copyFence;
VkFenceCreateInfo fenceInfo;
memset(&fenceInfo, 0, sizeof(fenceInfo));
fenceInfo.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO;
vkCreateFence(device->getVKDevice(), &fenceInfo,NULL , ©Fence);
vkCmdCopyBufferToImage(cmd, cubeMapBuffer, m_data.image, m_data.imageLayout, 6, biCpyRgn);
queue->flushCommandBuffer(cmdID , ©Fence);
vkWaitForFences(device->getVKDevice(), 1, ©Fence, VK_TRUE, 100000000000);
vkDestroyBuffer(device->getVKDevice(), cubeMapBuffer, NULL);
vkFreeMemory(device->getVKDevice(), cubeMapMem, NULL);
}
VkSamplerCreateInfo sampler;
sampler.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO;
sampler.pNext = NULL;
sampler.magFilter = VK_FILTER_NEAREST;
sampler.minFilter = VK_FILTER_NEAREST;
sampler.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST;
sampler.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
sampler.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
sampler.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
sampler.mipLodBias = 0.0f;
sampler.maxAnisotropy = 1;
sampler.compareOp = VK_COMPARE_OP_NEVER;
sampler.minLod = 0.0f;
sampler.maxLod = 0.0f;
sampler.borderColor = VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE;
VkImageViewCreateInfo view;
view.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO;
view.pNext = NULL;
view.viewType = VK_IMAGE_VIEW_TYPE_CUBE;
view.format = m_format;
view.components = {
VK_COMPONENT_SWIZZLE_R, VK_COMPONENT_SWIZZLE_G, VK_COMPONENT_SWIZZLE_B, VK_COMPONENT_SWIZZLE_A
};
view.subresourceRange = { VK_IMAGE_ASPECT_COLOR_BIT, 0, 1, 0, 0 };
view.subresourceRange.baseArrayLayer = 0;
view.subresourceRange.levelCount = 1;
view.subresourceRange.baseMipLevel = 0;
view.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
view.subresourceRange.layerCount = 1;
VKA_CHECK_ERROR(vkCreateSampler(getDefaultDevice(), &sampler,NULL, &m_data.sampler), "Could not create sampler for image texture.\n");
view.image = m_data.image;
VKA_CHECK_ERROR(vkCreateImageView(getDefaultDevice(), &view,NULL, &m_data.view), "Could not create image view for texture.\n");
VKA_INFO_MSG("Created CUBE Image Texture.\n");
}
void VkeCubeTexture::loadCubeDDS(const char *inFile){
std::string searchPaths[] = {
std::string(PROJECT_NAME),
NVPWindow::sysExePath() + std::string(PROJECT_RELDIRECTORY),
std::string(PROJECT_ABSDIRECTORY)
};
nv_dds::CDDSImage ddsImage;
for (uint32_t i = 0; i < 3; ++i){
std::string separator = "";
uint32_t strSize = searchPaths[i].size();
if(searchPaths[i].substr(strSize-1,strSize) != "/") separator = "/";
std::string filePath = searchPaths[i] + separator + std::string("images/") + std::string(inFile);
ddsImage.load(filePath, true);
if (ddsImage.is_valid()) break;
}
if (!ddsImage.is_valid()){
perror("Could not cube load texture image.\n");
exit(1);
}
uint32_t imgW = ddsImage.get_width();
uint32_t imgH = ddsImage.get_height();
uint32_t comCount = ddsImage.get_components();
uint32_t fmt = ddsImage.get_format();
bool isCube = ddsImage.is_cubemap();
bool isComp = ddsImage.is_compressed();
VkFormat vkFmt = VK_FORMAT_R8G8B8A8_UNORM;
switch (fmt){
case GL_COMPRESSED_RGBA_S3TC_DXT1_EXT:
vkFmt = VK_FORMAT_BC1_RGB_SRGB_BLOCK;
break;
case GL_COMPRESSED_RGBA_S3TC_DXT3_EXT:
vkFmt = VK_FORMAT_BC2_UNORM_BLOCK;
break;
case GL_COMPRESSED_RGBA_S3TC_DXT5_EXT:
vkFmt = VK_FORMAT_BC3_UNORM_BLOCK;
break;
default:
break;
}
m_width = imgW;
m_height = imgH;
m_format = vkFmt;
VulkanDC::Device::Queue::Name queueName = "DEFAULT_GRAPHICS_QUEUE";
VulkanDC::Device::Queue::CommandBufferID cmdID = INIT_COMMAND_ID;
VulkanDC *dc = VulkanDC::Get();
VulkanDC::Device *device = dc->getDefaultDevice();
VulkanDC::Device::Queue *queue = device->getQueue(queueName);
VkCommandBuffer cmd = VK_NULL_HANDLE;
queue->beginCommandBuffer(cmdID, &cmd, VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT);
imageCreateAndBind(
&m_data.image,
&m_data.memory,
m_format, VK_IMAGE_TYPE_2D,
m_width, m_height, 1, 6,
VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
(VkImageUsageFlagBits)(VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT),
VK_IMAGE_TILING_OPTIMAL);
VkBuffer cubeMapBuffer;
VkDeviceMemory cubeMapMem;
bufferCreate(&cubeMapBuffer, m_width*m_height * 3 * 6, VK_BUFFER_USAGE_TRANSFER_SRC_BIT);
bufferAlloc(&cubeMapBuffer, &cubeMapMem, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT);
if (m_memory_flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT){
imageSetLayoutBarrier(cmdID, queueName, m_data.image, VK_IMAGE_ASPECT_COLOR_BIT, VK_IMAGE_LAYOUT_PREINITIALIZED, VK_IMAGE_LAYOUT_GENERAL);
for (uint32_t i = 0; i < 6; ++i){
void *data = NULL;
VkSubresourceLayout layout;
VkImageSubresource subres;
subres.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
subres.mipLevel = m_mip_level;
subres.arrayLayer = i;
vkGetImageSubresourceLayout(getDefaultDevice(), m_data.image, &subres, &layout);
VKA_CHECK_ERROR(vkMapMemory(getDefaultDevice(), cubeMapMem, layout.offset, layout.size, 0, &data), "Could not map memory for image.\n");
const nv_dds::CTexture &mipmap = ddsImage.get_cubemap_face(i);
memcpy(data, (void *)mipmap, layout.size);
vkUnmapMemory(getDefaultDevice(), cubeMapMem);
}
VkBufferImageCopy biCpyRgn[6];
for (uint32_t k = 0; k < 6; ++k){
VkSubresourceLayout layout;
VkImageSubresource subres;
subres.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
subres.mipLevel = m_mip_level;
subres.arrayLayer = k;
vkGetImageSubresourceLayout(getDefaultDevice(), m_data.image, &subres, &layout);
biCpyRgn[k].bufferOffset = layout.offset;
biCpyRgn[k].bufferImageHeight = 0;
biCpyRgn[k].bufferRowLength = 0;
biCpyRgn[k].imageExtent.width = m_width;
biCpyRgn[k].imageExtent.height = m_height;
biCpyRgn[k].imageExtent.depth = 1;
biCpyRgn[k].imageOffset.x = 0;
biCpyRgn[k].imageOffset.y = 0;
biCpyRgn[k].imageOffset.z = 0;
biCpyRgn[k].imageSubresource.baseArrayLayer = k;
biCpyRgn[k].imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
biCpyRgn[k].imageSubresource.layerCount = 1;
biCpyRgn[k].imageSubresource.mipLevel = 0;
}
VkFence copyFence;
VkFenceCreateInfo fenceInfo;
memset(&fenceInfo, 0, sizeof(fenceInfo));
fenceInfo.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO;
vkCreateFence(device->getVKDevice(), &fenceInfo, NULL, ©Fence);
vkCmdCopyBufferToImage(cmd, cubeMapBuffer, m_data.image, m_data.imageLayout, 6, biCpyRgn);
queue->flushCommandBuffer(cmdID, ©Fence);
vkWaitForFences(device->getVKDevice(), 1, ©Fence, VK_TRUE, 100000000000);
vkDestroyBuffer(device->getVKDevice(), cubeMapBuffer, NULL);
vkFreeMemory(device->getVKDevice(), cubeMapMem, NULL);
}
VkSamplerCreateInfo sampler;
sampler.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO;
sampler.pNext = NULL;
sampler.magFilter = VK_FILTER_NEAREST;
sampler.minFilter = VK_FILTER_NEAREST;
sampler.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST;
sampler.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
sampler.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
sampler.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
sampler.mipLodBias = 0.0f;
sampler.maxAnisotropy = 1;
sampler.compareOp = VK_COMPARE_OP_NEVER;
sampler.minLod = 0.0f;
sampler.maxLod = 0.0f;
sampler.borderColor = VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE;
VkImageViewCreateInfo view;
view.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO;
view.pNext = NULL;
view.viewType = VK_IMAGE_VIEW_TYPE_CUBE;
view.format = m_format;
view.components.r = VK_COMPONENT_SWIZZLE_R;
view.components.g = VK_COMPONENT_SWIZZLE_G;
view.components.b = VK_COMPONENT_SWIZZLE_B;
view.components.a = VK_COMPONENT_SWIZZLE_A;
view.subresourceRange.baseArrayLayer = 0;
view.subresourceRange.levelCount = 1;
view.subresourceRange.baseMipLevel = 0;
view.subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
view.subresourceRange.layerCount = 1;
VKA_CHECK_ERROR(vkCreateSampler(getDefaultDevice(), &sampler, NULL, &m_data.sampler), "Could not create sampler for image texture.\n");
view.image = m_data.image;
VKA_CHECK_ERROR(vkCreateImageView(getDefaultDevice(), &view, NULL, &m_data.view), "Could not create image view for texture.\n");
}
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);
}
static void where(Param &out, Param &in)
{
uint threads_x = nextpow2(std::max(32u, (uint)in.info.dims[0]));
threads_x = std::min(threads_x, THREADS_PER_GROUP);
uint threads_y = THREADS_PER_GROUP / threads_x;
uint groups_x = divup(in.info.dims[0], threads_x * REPEAT);
uint groups_y = divup(in.info.dims[1], threads_y);
Param rtmp;
Param otmp;
rtmp.info.dims[0] = groups_x;
otmp.info.dims[0] = in.info.dims[0];
rtmp.info.strides[0] = 1;
otmp.info.strides[0] = 1;
rtmp.info.offset = 0;
otmp.info.offset = 0;
for (int k = 1; k < 4; k++) {
rtmp.info.dims[k] = in.info.dims[k];
rtmp.info.strides[k] = rtmp.info.strides[k - 1] * rtmp.info.dims[k - 1];
otmp.info.dims[k] = in.info.dims[k];
otmp.info.strides[k] = otmp.info.strides[k - 1] * otmp.info.dims[k - 1];
}
int rtmp_elements = rtmp.info.strides[3] * rtmp.info.dims[3];
rtmp.data = bufferAlloc(rtmp_elements * sizeof(uint));
int otmp_elements = otmp.info.strides[3] * otmp.info.dims[3];
otmp.data = bufferAlloc(otmp_elements * sizeof(uint));
scan_first_launcher<T, uint, af_notzero_t>(otmp, rtmp, in, false, groups_x, groups_y, threads_x);
// Linearize the dimensions and perform scan
Param ltmp = rtmp;
ltmp.info.offset = 0;
ltmp.info.dims[0] = rtmp_elements;
for (int k = 1; k < 4; k++) {
ltmp.info.dims[k] = 1;
ltmp.info.strides[k] = rtmp_elements;
}
scan_first<uint, uint, af_add_t>(ltmp, ltmp);
// Get output size and allocate output
uint total;
getQueue().enqueueReadBuffer(*rtmp.data, CL_TRUE,
sizeof(uint) * (rtmp_elements - 1),
sizeof(uint),
&total);
out.data = bufferAlloc(total * sizeof(uint));
out.info.dims[0] = total;
out.info.strides[0] = 1;
for (int k = 1; k < 4; k++) {
out.info.dims[k] = 1;
out.info.strides[k] = total;
}
if (total > 0)
get_out_idx<T>(out.data, otmp, rtmp, in, threads_x, groups_x, groups_y);
bufferFree(rtmp.data);
bufferFree(otmp.data);
}
void vkeGameRendererDynamic::initIndirectCommands(){
if (!m_node_data) return;
VulkanDC *dc = VulkanDC::Get();
VulkanDC::Device *device = dc->getDefaultDevice();
VulkanDC::Device::Queue *queue = dc->getDefaultQueue();
uint32_t cnt = m_node_data->count();
uint32_t sz = sizeof(VkDrawIndexedIndirectCommand)*cnt;
VkBuffer sceneIndirectStaging;
VkDeviceMemory sceneIndirectMemStaging;
VkBufferUsageFlags usageFlags = VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT;
bufferCreate(&m_scene_indirect_buffer, sz, (VkBufferUsageFlagBits)usageFlags);
bufferAlloc(&m_scene_indirect_buffer, &m_scene_indirect_memory, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);
usageFlags = VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
bufferCreate(&sceneIndirectStaging, sz, (VkBufferUsageFlagBits)usageFlags);
bufferAlloc(&sceneIndirectStaging, &sceneIndirectMemStaging, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT);
VkDrawIndexedIndirectCommand *commands = NULL;
VKA_CHECK_ERROR(vkMapMemory(device->getVKDevice(), sceneIndirectMemStaging, 0, sz, 0, (void **)&commands), "Could not map indirect buffer memory.\n");
for (uint32_t i = 0; i < cnt; ++i){
VkeMesh *mesh = m_node_data->getData(i)->getMesh();
commands[i].firstIndex = mesh->getFirstIndex();
commands[i].firstInstance = i*m_instance_count;
commands[i].vertexOffset = mesh->getFirstVertex();
commands[i].indexCount = mesh->getIndexCount();
commands[i].instanceCount = m_instance_count;
}
vkUnmapMemory(device->getVKDevice(), sceneIndirectMemStaging);
VkBufferCopy bufCpy;
bufCpy.dstOffset = 0;
bufCpy.srcOffset = 0;
bufCpy.size = sz;
VkCommandBuffer copyCmd;
VkCommandBufferAllocateInfo cmdBufInfo = { VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO };
cmdBufInfo.commandBufferCount = 1;
cmdBufInfo.commandPool = queue->getCommandPool();
cmdBufInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
VKA_CHECK_ERROR(vkAllocateCommandBuffers(device->getVKDevice(), &cmdBufInfo, ©Cmd), "Could not allocate command buffers.\n");
VkCommandBufferBeginInfo cmdBeginInfo = { VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO };
cmdBeginInfo.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
VKA_CHECK_ERROR(vkBeginCommandBuffer(copyCmd, &cmdBeginInfo), "Could not begin commmand buffer.\n");
vkCmdCopyBuffer(copyCmd, sceneIndirectStaging, m_scene_indirect_buffer, 1, &bufCpy);
VKA_CHECK_ERROR(vkEndCommandBuffer(copyCmd), "Could not end command buffer.\n");
VkSubmitInfo subInfo = { VK_STRUCTURE_TYPE_SUBMIT_INFO };
subInfo.commandBufferCount = 1;
subInfo.pCommandBuffers = ©Cmd;
VkFence theFence;
VkFenceCreateInfo fenceInfo = { VK_STRUCTURE_TYPE_FENCE_CREATE_INFO };
VKA_CHECK_ERROR(vkCreateFence(device->getVKDevice(), &fenceInfo, NULL, &theFence), "Could not create fence.\n");
VKA_CHECK_ERROR(vkQueueSubmit(queue->getVKQueue(), 1, &subInfo, theFence), "Could not submit queue for indirect buffer copy.\n");
VKA_CHECK_ERROR(vkWaitForFences(device->getVKDevice(), 1, &theFence, VK_TRUE, UINT_MAX), "Could not wait for fence.\n");
vkFreeCommandBuffers(device->getVKDevice(), queue->getCommandPool(), 1, ©Cmd);
vkDestroyFence(device->getVKDevice(), theFence, NULL);
}
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 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 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;
}
}
T *memAlloc(const size_t &elements)
{
managerInit();
return (T *)bufferAlloc(elements * sizeof(T));
}
