Fix rocAL Tensor build issues PR 4 - Pipeline and Pybind #28
Conversation
fiona-gladwin
commented
Jul 7, 2023
fiona-gladwin
commented
- Re introduced the copy tensor API in master graph as it gives performance in tensor trainings
- Re-introduce some external API to get width and height from master graph as suggested
- Introduce API and changes required to get original ROI from tensor (Similar to TOT changes)
- Fix build issues with pybind
fiona-gladwin
force-pushed
the
PR_tensor_w4
branch
from
2651ffa to
0a25c27
Compare
| @@ -50,6 +50,25 @@ THE SOFTWARE. | |||
| #define MAX_NUM_ANCHORS 8732 // Num of bbox achors used in SSD training | |||
| #define MAX_MASK_BUFFER 10000 | |||
|
|
|||
| #if ENABLE_SIMD | |||
There was a problem hiding this comment.
this seems to be missing in tot MIVIsionX. Why is this needed here
There was a problem hiding this comment.
This portion was a part of master_graph.cpp.
We have moved the masks as constants into the master_graph.h hence we have moved the include part also from the cpp file into master_graph.h file
|
|
||
| _convert_time.start(); | ||
| // Copies to the output context given by the user | ||
| auto dims = _output_tensor_list[0]->info().dims(); |
There was a problem hiding this comment.
please use another pointer to point to output_tensor_list[0]->info() and use that everywhere.
| const size_t c = dims[3]; | ||
| const size_t h = dims[1]; | ||
| const size_t w = dims[2]; | ||
| const size_t single_output_image_size = _output_tensor_list[0]->info().data_size(); |
There was a problem hiding this comment.
rename to tensor specific name
| @@ -112,6 +112,8 @@ void TensorInfo::reset_tensor_roi_buffers() { | |||
| size_t roi_size = (_layout == RocalTensorlayout::NFCHW || _layout == RocalTensorlayout::NFHWC) ? _dims[0] * _dims[1] : _batch_size; // For Sequences pre allocating the ROI to N * F to replicate in OpenVX extensions | |||
| allocate_host_or_pinned_mem((void **)&_roi_buf, roi_size * 4 * sizeof(unsigned), _mem_type); | |||
| } | |||
| if(!_original_roi_height->size()) _original_roi_height = std::make_shared<std::vector<uint32_t>>(_batch_size); | |||
There was a problem hiding this comment.
Why is changed from struct roi to width and height
There was a problem hiding this comment.
Removed these _original_roi_width and _original_roi_height as they are not used anymore
rocAL/rocAL_pybind/rocal_pybind.cpp
Outdated
| numpy_array = py::array(py::buffer_info( | ||
| ((unsigned char *)(output_tensor.buffer())) + idx * (info.strides()[0]/sizeof(float)), | ||
| ((unsigned char *)(output_tensor.buffer())) + idx * (output_tensor.strides()[0]/sizeof(float)), |
There was a problem hiding this comment.
why is this dividing by sizeof(float) for UINT8
fiona-gladwin
force-pushed
the
PR_tensor_w4
branch
from
0a25c27 to
2651ffa
Compare
Avoid the conflicting function name between tensor and image files
@rrawther and @LakshmiKumar23 I have addressed all review comments. Please check. |
