Tensor changes in master graph and ring buffer #17
Conversation
Changed all instances from Image to Tensor Remove the cycle idx loop in output routine Removed video output routine Removed copy out API Remove API to get output width, height from info Remove output tensor allocation
Remove master buffer and intrduce sub buffer for HOST
| #endif | ||
| _convert_time.end(); | ||
| return Status::OK; | ||
| std::vector<void*> output_ptr = _ring_buffer.get_read_buffers(); |
There was a problem hiding this comment.
Is this call doing a copy or returning reference to a vector? Can you please use auto call like the image pipeline here?
There was a problem hiding this comment.
Changed this to auto. This call would return a copy of the vector of pointers
| for(unsigned i = 0; i < _internal_tensor_list.size(); i++) | ||
| _output_tensor_list[i]->set_mem_handle(output_ptr[i]); | ||
|
|
||
| return &_output_tensor_list; |
There was a problem hiding this comment.
In case of framework integration: how does it copy to framework tensor?
There was a problem hiding this comment.
We would return the tensor to user and in the iterator we call a tensor.copy_data(user_buffer) where we pass a user buffer and the contents of tensor is copied to user buffer. This API is added as a member function of rocalTensor class in tensor.cpp
There was a problem hiding this comment.
Hi Fionna, In that case we need to make copy_data() uses optimized code to get the best performance
| { | ||
| ERR("Exception thrown in the process routine: " + STR(e.what()) + STR("\n")); | ||
| _processing = false; | ||
| _ring_buffer.release_all_blocked_calls(); |
There was a problem hiding this comment.
why is this exception handling code removed?
There was a problem hiding this comment.
We have removed the output routine of the video, so it is showing this as removed but the code is present for the same
| @@ -1493,163 +783,6 @@ bool MasterGraph::no_more_processed_data() | |||
| return (_output_routine_finished_processing && _ring_buffer.empty()); | |||
| } | |||
|
|
|||
| MasterGraph::Status | |||
| MasterGraph::copy_out_tensor_planar(void *out_ptr, RocalTensorFormat format, float multiplier0, float multiplier1, | |||
There was a problem hiding this comment.
this was required for supporting planar output from the pipeline (like for CFar10 data set). Is this handled in tensor pipeline?
There was a problem hiding this comment.
Yes this is handled in tensor pipeline, some of our trainings require planar output i.e NCHW we are able to return a planar output from rocAL and normalization would happen in RPP.
| @@ -159,7 +151,7 @@ void RingBuffer::init(RocalMemType mem_type, void *devres, unsigned sub_buffer_s | |||
|
|
|||
| } | |||
| } | |||
| else | |||
| else | |||
There was a problem hiding this comment.
make sure the code works for all 3 backends (ENABLE_HIP, ENABLE_OPENCL, CPU)
| @@ -316,7 +316,7 @@ unsigned rocalTensor::copy_data(hipStream_t stream, void *host_memory, bool sync | |||
| #endif | |||
|
|
|||
| unsigned rocalTensor::copy_data(void *user_buffer) { | |||
| if (_info._type != rocalTensorInfo::Type::HANDLE) return 0; | |||
There was a problem hiding this comment.
why is this changed? Are they both have the same meaning
There was a problem hiding this comment.
In our previous case we would associate all the image related rocalTensors with a vx_tensor and we set the type as handle for the output tensors where we create vx_tensors with handle.
But with our current changes we are not creating the output tensor (which is to be returned to the user) as a vx_tensor hence handle is not set, so we could not copy the data. Hence I have changed the condition to check if it is a null pointer.
|
Hi @rrawther and @LakshmiKumar23 I have addressed all the PR comments. |
@rrawther can I merge this PR? |
In Master Graph
In Ring Buffer