Seeking your assistance in solving the second bullet point correctly. I am not clearly understanding if cudaMemcpyHostToDevice will copy over a variable value from CPU $($System Shared Memory?$)$ to GPU $($Global Shared Memory on the device?$).$ My understanding is that both the matrix A and the vector b needs to be copied over to GPU in this scenario, and the copy destination is the Global Shared Memory, and then, b needs to be copied over to the shared memory on each Streaming Multiprocessor. And the host code may have these actions within:

$//$ Allocate memory for matrix A and vector b using cudaMalloc

$//$ Copy A$,$ b$,$ and c into GPU via cudaMalloc

$//$ Launch kernel threads by invoking $\_\_$global$\_\_$ matrixVectorMultiply which I need to write. dimGrid and dimBlock need to be set before invoking matrixVectorMultiply $($should the argument to dimGrid be $1$K$,$ meaning $1024,$ and dimBlock $1$ here?$)$

Please advise

$---------$ Attached Problem $--------------$

$6[15$ points$]$

Matrix$-$vector multiplication using CUDA

Design a CUDA program to perform matrix$-$vector multiplication c$=$A xx b$.$ The size

of matrix A is $1$K xx$1$K$.$ The size of vectors b and c is $1$K xx$1.$ Your program should

use $1$K threads in total. Assume the shared memory is large enough to hold the entire

vector b$.$ The input matrix and the vector are initially stored in the host memory.

Write a pseudo code for the host function and kernel function. Note that your kernel

function must use shared memory to store vector b$.$

Assume each element of A$,$b$,$c is $4$ bytes; data transfer between CPU and GPU is

through PCIe whose bandwidth is $16$GB$//$s in each direction; the clock rate of GPU

is $1$ GHz ; the access latency to global memory and shared memory is $100$ clock cycles

and $10$ clock cycles, respectively; multiply$-$add operations are overlapped with memory

access operations. What is the execution time of your CUDA program in the best case? $6[15$ points$]$

Matrix$-$vector multiplication using CUDA

$-$ Design a CUDA program to perform matrix$-$vector multiplication $\backslash ($ c$=$A $\backslash$times b $\backslash ).$ The size of matrix A is $\backslash (1$ K $\backslash$times $1$ K $\backslash ).$ The size of vectors $\backslash ($ b $\backslash )$ and $\backslash ($ c $\backslash )$ is $\backslash (1$ K $\backslash$times $1\backslash ).$ Your program should use $\backslash (1$ K $\backslash )$ threads in total. Assume the shared memory is large enough to hold the entire vector $\backslash ($ b $\backslash ).$ The input matrix and the vector are initially stored in the host memory. Write a pseudo code for the host function and kernel function. Note that your kernel function must use shared memory to store vector $\backslash ($ b $\backslash ).$

$-$ Assume each element of $\backslash ($ A$,$ b$,$ c $\backslash )$ is $4$ bytes; data transfer between CPU and GPU is through PCIe whose bandwidth is $\backslash (16\backslash$mathrm$\{$~GB$\}/\backslash$mathrm$\{$s$\}\backslash )$ in each direction; the clock rate of GPU is $1$ GHz ; the access latency to global memory and shared memory is $100$ clock cycles and $10$ clock cycles, respectively; multiply$-$add operations are overlapped with memory access operations. What is the execution time of your CUDA program in the best case?