You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
58 lines
1.6 KiB
58 lines
1.6 KiB
// TODO: Simple matrix-vector multiplication, every thread computes a complete dot product
|
|
//
|
|
// i := get_global_id(0)
|
|
//
|
|
// IF ID < n THEN:
|
|
// yi := b[i]
|
|
// LOOP j := 0 .. m DO:
|
|
// yi += A[j + i * m] * x[j]
|
|
// END LOOP
|
|
// y[i] := yi
|
|
// END IF
|
|
__kernel
|
|
void simpleMV(const int n, const int m, __global float* y, __global float* A, __global float* x, __global float* b){
|
|
|
|
}
|
|
|
|
// TODO: Matrix-vector multiplication with parallelization of the dot product
|
|
// Assumptions: M = 2^k, M <= maximum workgroup size
|
|
//
|
|
// i = get_group_id(0)
|
|
// j = get_local_id(0)
|
|
//
|
|
// Q[j] := A[i * M + j] * x[j]
|
|
// BARRIER
|
|
//
|
|
// Sum scan on Q (reduction)
|
|
//
|
|
// IF j = 0 THEN:
|
|
// y[i] = Q[0] + b[i]
|
|
//
|
|
__kernel
|
|
void reduceMV(const int n, const int M, __global float* y, __global float* A, __global float* x, __global float* b, __local float* Q){
|
|
|
|
}
|
|
|
|
// TODO: General solution for matrix-vector multiplication, every thread processes a chunk of the dot product and visits multiple rows of the result
|
|
//
|
|
// t := get_local_id(0) / Z
|
|
// z := get_local_id(0) % Z
|
|
//
|
|
// FOR i := t ; i < n ; i := i + T :
|
|
// Compute Q[t * Z + z] as shown in the lecture
|
|
// Sum scan on Q (reduction)
|
|
// IF z = 0 THEN:
|
|
// y[i] = Q[t * Z + 0] + b[i]
|
|
//
|
|
// END FOR
|
|
__kernel
|
|
void largeMV(const int n, const int m, __global float* y, __global float* A, __global float* x, __global float* b, const int T, const int Z, __local float* Q){
|
|
|
|
}
|
|
|
|
// TODO: Gaussian elimination as shown in the lecture
|
|
// (execute the 2nd loop of the sequential implemential in parallel)
|
|
__kernel void gaussian(const int n, const int m, __global float* A){
|
|
|
|
}
|
|
|
|
|