Code Modifications - 2023.2 English

Vitis Tutorials: Hardware Acceleration (XD099)

Document ID

XD099

Release Date

2023-11-13

Version

2023.2 English

Navigate to $LAB_WORK_DIR/reference_files, and with a file editor, open run_sw_overlap_multiDDR.cpp.
From the host code, you will need to send the words to both DDR banks alternatively. The DDR bank assignment in the host code is supported by a AMD vendor extension to the OpenCL™ API. Two AMD extension pointer objects (cl_mem_ext_ptr_t) are created, buffer_words_ext[0] and buffer_words_ext[1]. Theflags will determine which DDR bank the buffer will be send to, so that kernel can access it.
```
 cl_mem_ext_ptr_t buffer_words_ext[2];

 buffer_words_ext[0].flags = 1 | XCL_MEM_TOPOLOGY; // DDR[1]
 buffer_words_ext[0].param = 0;
 buffer_words_ext[0].obj   = input_doc_words;
 buffer_words_ext[1].flags = 2 | XCL_MEM_TOPOLOGY; // DDR[2]
 buffer_words_ext[1].param = 0;
 buffer_words_ext[1].obj   = input_doc_words;
```

The next two buffers, buffer_doc_words[0] and buffer_doc_words[1], are created in DDR[1] and DDR[2] as follows.

buffer_doc_words[0] = cl::Buffer(context, CL_MEM_EXT_PTR_XILINX | CL_MEM_USE_HOST_PTR | CL_MEM_READ_ONLY, total_size*sizeof(uint), &buffer_words_ext[0]);
buffer_doc_words[1] = cl::Buffer(context, CL_MEM_EXT_PTR_XILINX | CL_MEM_USE_HOST_PTR | CL_MEM_READ_ONLY, total_size*sizeof(uint), &buffer_words_ext[1]);
buffer_inh_flags    = cl::Buffer(context, CL_MEM_USE_HOST_PTR | CL_MEM_WRITE_ONLY, total_size*sizeof(char),output_inh_flags);
buffer_bloom_filter = cl::Buffer(context, CL_MEM_USE_HOST_PTR | CL_MEM_READ_ONLY, bloom_filter_size*sizeof(uint),bloom_filter);

// Set buffer kernel arguments (needed to migrate the buffers in the correct memory)
kernel.setArg(0, buffer_inh_flags);
kernel.setArg(1, buffer_doc_words[0]);
kernel.setArg(2, buffer_bloom_filter);

// Make buffers resident in the device
q.enqueueMigrateMemObjects({buffer_bloom_filter, buffer_doc_words[0], buffer_doc_words[1], buffer_inh_flags}, CL_MIGRATE_MEM_OBJECT_CONTENT_UNDEFINED);

// Create sub-buffers, one for each transaction
unsigned subbuf_doc_sz = total_doc_size/num_iter;
unsigned subbuf_inh_sz = total_doc_size/num_iter;

cl_buffer_region subbuf_inh_info[num_iter];
cl_buffer_region subbuf_doc_info[num_iter];
cl::Buffer subbuf_inh_flags[num_iter];
cl::Buffer subbuf_doc_words[num_iter];

for (int i=0; i<num_iter; i++) {
    subbuf_inh_info[i]={i*subbuf_inh_sz*sizeof(char), subbuf_inh_sz*sizeof(char)};
    subbuf_doc_info[i]={i*subbuf_doc_sz*sizeof(uint), subbuf_doc_sz*sizeof(uint)};
    subbuf_inh_flags[i] = buffer_inh_flags.createSubBuffer(CL_MEM_WRITE_ONLY, CL_BUFFER_CREATE_TYPE_REGION, &subbuf_inh_info[i]);
    // The doc words sub-buffers will be alternating in DDR[1] and DDR[2]
    subbuf_doc_words[i] = buffer_doc_words[i%2].createSubBuffer (CL_MEM_READ_ONLY,  CL_BUFFER_CREATE_TYPE_REGION, &subbuf_doc_info[i]);
}

The kernel argument, input word is set to the array of sub-buffers created from buffer_doc_words[0] and buffer_doc_words[1] alternatively; hence, data is sent to DDR bank 1 and 2 alternatively in each kernel execution.

for (int i=0; i<num_iter; i++)
{
  cl::Event buffDone, krnlDone, flagDone;
  total_size = subbuf_doc_info[i].size / sizeof(uint);
  load_filter = false;
  kernel.setArg(0, subbuf_inh_flags[i]);
  kernel.setArg(1, subbuf_doc_words[i]);
  kernel.setArg(3, total_size);
  kernel.setArg(4, load_filter);
  q.enqueueMigrateMemObjects({subbuf_doc_words[i]}, 0, &wordWait, &buffDone);
  wordWait.push_back(buffDone);
  q.enqueueTask(kernel, &wordWait, &krnlDone);
  krnlWait.push_back(krnlDone);
  q.enqueueMigrateMemObjects({subbuf_inh_flags[i]}, CL_MIGRATE_MEM_OBJECT_HOST, &krnlWait, &flagDone);
  flagWait.push_back(flagDone);
}