Heterogeneous computing

1. The SIMD computational model refers to: (5 points)

   • Model in which a thread processes individual data
   • Model in which multiple threads process the same individual data
   • Model in which mutiple threads process multiple data at the same time
   • Model in which multiple threads process individual data

2. According to Flynn’s Taxonomy. What model does the processing found in multi-core CPUs correspond to? (5 points)

   • SISD
   • SIMD
   • MISD
   • MIMD

3. In the CUDA programming language for Nvidia GPUs, what is the purpose of the cudaMalloc(…) call? (5 points)

   • Reserve memory in host
   • Launch a running kernel
   • Reserve memory in GPU device
   • Copy data to GPU device

4. In SYCL and DPC++ there are different classes for managing kernels and data. Indicate for each of the following classes its purpose: (12 points)

   Class	SRAM	DRAM
   sycl::device	☐	☐
   sycl::queue	☐	☐
   sycl::buffer	☐	☐
   sycl::handler	☐	☐
   sycl::accessor	☐	☐
   sycl::range	☐	☐

5. Given the following code in DPC++, What type of accessor should be defines for A and B? (10 points)

   int main(){
       queue q;
       std::vector<int> v1(N);
       std::vector<int> v2(N);
       ...
       buffer buff1(v1);
       buffer buff2(v2);
       q.submit([&](handler& h){
           accessor A(buff1, h, XXXXX);
           accessor B(buff2, h, XXXXX);
           h.parallel_for(N, [=](auto i){
               B[i] = i + A[i];
           });
       }).wait();
   }
   

   • A: read_write and B: read_write
   • A: read_only and B: read_only
   • A: read_only and B: write_only
   • A: write_only and B: write_only

6. Given the following kernel definition, how many threads will be created for parallel_for execution? (10 points)

   h.parallel_for(nd_range<3>(range<3>({4,5,2}), range<3>({4,2,1})), [=](nd_item<3> item){
       auto global_ = item.get_global_id();
       auto local_id = item.get_local_id();
       ...
   });
   

   • 48
   • 4
   • 40
   • 3
   • 20

7. Given the following kernel definition. How many workgroups are defined for the parallel_for? (10 points)

   h.parallel_for(nd_range<3>(range<3>({4,5,2}), range<3>({4,2,1})), [=](nd_item<3> item){
       auto global_ = item.get_global_id();
       auto local_id = item.get_local_id();
       ...
   });
   

   • 40
   • 3
   • 8
   • 1
   • 5

8. GPU Computing Talk: What do the ASICs in the new Nvidia GPUs correspond to? (5 points)

   • Specialized hardware for certain specific tasks
   • Hardware available to run general-purpose computation
   • New programmable L1 cache available in Raytracing and Tensor cores
   • Streaming multiprocesors available on the chip

9. GPU Computing Talk: Explain what the term Multi-chip-module refers to in modern GPUs (8 points)

   ---

    Type your answer here
   

   ---