natas_hw
/
openvino
0
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[GPU] fix tensor allocation (#22551) 

### Details:
 - *fix tensor allocation*

### Tickets:
 - *CVS-130946*
This commit is contained in:
Andrei Gorbachev 2024-02-19 13:48:08 +00:00 committed by GitHub
parent bd2fce78f9
commit 069bd84f60
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 32 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
src/plugins/intel_gpu/include/intel_gpu/plugin/common_utils.hpp
View File

@ -121,6 +121,7 @@ inline void ForceExit() {
void convert_and_copy(const ov::ITensor* src, cldnn::memory::ptr dst, cldnn::stream& stream);
void convert_and_copy(const cldnn::memory::ptr src, ov::ITensor const* dst, const cldnn::stream& stream);
void convert_and_copy(const ov::ITensor* src, ov::ITensor const* dst, const cldnn::stream& stream);
void convert_and_copy(const cldnn::memory::ptr src, cldnn::memory::ptr dst, cldnn::stream& stream);
} // namespace intel_gpu

14
src/plugins/intel_gpu/src/plugin/common_utils.cpp
View File

@ -138,6 +138,20 @@ void convert_and_copy(const cldnn::memory::ptr src, ov::ITensor const* dst, cons
return ::convert_and_copy(src_ptr, src_et, dst_ptr, dst_et, size, src->get_layout());
}
void convert_and_copy(const cldnn::memory::ptr src, cldnn::memory::ptr dst, cldnn::stream& stream) {
const bool blocking = true;
auto src_et = src->get_layout().data_type;
auto dst_et = dst->get_layout().data_type;
cldnn::mem_lock<uint8_t, cldnn::mem_lock_type::read> src_lock(src, stream);
const void* src_ptr = src_lock.data();
size_t size = ov::shape_size(src->get_layout().get_shape());
ov::Tensor tmp_tensor(dst_et, src->get_layout().get_shape());
::convert_and_copy(src_ptr, src_et, tmp_tensor.data(), dst_et, size, src->get_layout());
dst->copy_from(stream, tmp_tensor.data(), blocking);
}
void convert_and_copy(const ov::ITensor* src, ov::ITensor const* dst, const cldnn::stream& stream) {
auto src_et = src->get_element_type();
auto dst_et = dst->get_element_type();

25
src/plugins/intel_gpu/src/plugin/sync_infer_request.cpp
View File

@ -675,7 +675,13 @@ std::vector<cldnn::event::ptr> SyncInferRequest::prepare_input(const std::string
bool convert_needed = is_convert_required(element_type, device_tensor_et);
if (is_remote) {
m_plugin_inputs[name] = user_tensor_wrapper;
if (convert_needed) {
m_plugin_inputs[name] = { create_device_tensor(pshape,
cldnn::element_type_to_data_type(element_type),
false), TensorOwner::PLUGIN };
} else {
m_plugin_inputs[name] = user_tensor_wrapper;
}
} else if (is_usm_host_tensor && !convert_needed && can_use_usm_host(engine)) {
if (element_type != cldnn::element_type_to_data_type(element_type)) {
m_plugin_inputs[name] = {std::make_shared<RemoteTensorImpl>(m_context,
@ -753,14 +759,17 @@ std::vector<cldnn::event::ptr> SyncInferRequest::prepare_input(const std::string
}
cldnn::event::ptr ret_event = nullptr;
if (!is_remote) {
if (convert_needed) {
convert_and_copy(user_tensor.get(), device_tensor.get(), stream);
if (!is_remote && !convert_needed) {
auto src_ptr = static_cast<uint8_t*>(user_tensor->data());
if (!same_host_mem(memory, src_ptr)) {
ret_event = memory->copy_from(stream, src_ptr, false);
}
}
if (convert_needed) {
if (is_remote) {
convert_and_copy(remote_ptr->get_memory(), device_tensor->get_memory(), stream);
} else {
auto src_ptr = static_cast<uint8_t*>(user_tensor->data());
if (!same_host_mem(memory, src_ptr)) {
ret_event = memory->copy_from(stream, src_ptr, false);
}
convert_and_copy(user_tensor.get(), device_tensor.get(), stream);
}
}