Actions
Forum Thread Details
  • Replies 1 reply
  • Subscribers 532 subscribers
  • Views 233 views
  • Users 0 members are here
Related

in L3 Corner Track example. Why the flow buffer needs to be created everytime in loop?

bugtech

How can i reduce the execution time of "L3 Corner Track example" using my own video. I have observed in Vitis Analyzer that optical flow buffers are created in a loop which are adding more time to the execution process. I already managed to move other buffers of (pyr_down) to the beginning of the code so that they must not be created in a loop. But i am unable to move "Optical Flow" buffers out of the loop, if i do so, it does not give the correct optical flow. How this buffer can be managed?

I am executing a hardware build on "ZCU102"

Any leads?

 

Vitis Analyzer

Summary (API CALLS)

for (int l = NUM_LEVELS - 1; l >= 0; l--)

{

// compute current level height

int curr_height = pyr_h[l];

int curr_width = pyr_w[l];

 

////////////////MOVED THESE BUFFERS OUT OF THE LOOP/////////////////WORKS FINE//////////////

 

in_img_py1_buf = cl::Buffer(context, CL_MEM_USE_HOST_PTR | CL_MEM_READ_WRITE,

(pyr_h[l] * pyr_w[l] * CH_TYPE), (ap_uint<INPUT_PTR_WIDTH> *)imagepyr1[l].data);

in_img_py2_buf = cl::Buffer(context, CL_MEM_USE_HOST_PTR | CL_MEM_READ_WRITE,

(pyr_h[l] * pyr_w[l] * CH_TYPE), (ap_uint<INPUT_PTR_WIDTH> *)imagepyr2[l].data);

 

/////////////////////////////////////////////////////////////////////////////////////////

 

in_img_py1_Vec.push_back(in_img_py1_buf);

in_img_py2_Vec.push_back(in_img_py2_buf);

 

q.enqueueMigrateMemObjects(in_img_py1_Vec, 0 /* 0 means from host*/);

q.enqueueMigrateMemObjects(in_img_py2_Vec, 0 /* 0 means from host*/);

 

/////////////////////////////UNABLE TO MOVE THIS////////////////////////////////////////

flow_buf = cl::Buffer(context, CL_MEM_USE_HOST_PTR | CL_MEM_READ_WRITE, (pyr_h[l] * pyr_w[l] * 4),

(ap_uint<OUTPUT_PTR_WIDTH> *)flow.data);

////////////////////////////////////////////////////////////////////////////////////////////////////////

std::cout << "\n Buffers created" << std::endl;

// compute the flow vectors for the current pyramid level iteratively

std::cout << "\n *********OF Computation Level =" << l << "*********" << std::endl;

q.finish();

 

std::cout << "\n OF done" << std::endl;flow.init(pyr_h[l], pyr_w[l], 0);

flow_in.init(pyr_h[l], pyr_w[l], 0);

  •   we, all have to start before you can run..  Also you have a buffer in_img_py1_buffer what is that? Do you know in a year (or less) you will forget what it is. that goes for your other cryptic things lick pyr_h[1], pyr_w[1]  ok, you declared them as int and are set in lines 7 & 9  why?? but where are you getting the values from space?? and why don't you just use a static or what I like in #define curr_width  [to some value] (no equal sign) now some soup why do you need an array of 1 when a int variable would do??   ~~ that my two cents.. Cris H.