57 lines
1.5 KiB
Common Lisp
57 lines
1.5 KiB
Common Lisp
|
#define TSIZE 32
|
||
|
|
#define WPT 16
|
||
|
|
#define RSIZE (TSIZE/WPT)
|
||
|
|
|
||
|
|
// super super slow sgemm kernel by heehoon
|
||
|
|
__kernel void sgemm(__global float *A, __global float *B, __global float *C, int M, int N, int K) {
|
||
|
|
|
||
|
|
const int row = get_local_id(0);
|
||
|
|
const int col = get_local_id(1);
|
||
|
|
const int grow = TSIZE * get_group_id(0) + row;
|
||
|
|
const int gcolumn = TSIZE * get_group_id(1) + col;
|
||
|
|
const int t_num = (K+TSIZE-1)/TSIZE;
|
||
|
|
|
||
|
|
float result[WPT];
|
||
|
|
__local float Sub_A[TSIZE][TSIZE];
|
||
|
|
__local float Sub_B[TSIZE][TSIZE];
|
||
|
|
|
||
|
|
for (int i=0; i<WPT; i++) {
|
||
|
|
result[i] = 0.0f;
|
||
|
|
}
|
||
|
|
|
||
|
|
for (int j=0; j<t_num; j++) {
|
||
|
|
for (int i=0; i<WPT; i++) {
|
||
|
|
const int trow = j * TSIZE + row;
|
||
|
|
const int tcolumn = j * TSIZE + col;
|
||
|
|
if(grow + i * RSIZE >= M || tcolumn >= K){
|
||
|
|
Sub_A[row + i * RSIZE][col] = 0.0f;
|
||
|
|
}
|
||
|
|
else{
|
||
|
|
Sub_A[row + i * RSIZE][col] = A[(grow + i * RSIZE) * K + tcolumn];
|
||
|
|
}
|
||
|
|
if(trow + i*RSIZE >=K || gcolumn >=N){
|
||
|
|
Sub_B[row + i*RSIZE][col] = 0.0f;
|
||
|
|
}
|
||
|
|
else{
|
||
|
|
Sub_B[row + i * RSIZE][col] = B[(trow + i * RSIZE) * N + gcolumn];
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
barrier(CLK_LOCAL_MEM_FENCE);
|
||
|
|
|
||
|
|
for (int k=0; k<TSIZE; k++) {
|
||
|
|
for (int i=0; i<WPT; i++) {
|
||
|
|
result[i] += Sub_A[row + i * RSIZE][k] * Sub_B[k][col];
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
barrier(CLK_LOCAL_MEM_FENCE);
|
||
|
|
}
|
||
|
|
|
||
|
|
for (int i=0; i<WPT; i++) {
|
||
|
|
if(grow + i * RSIZE >= M || gcolumn >=N) continue;
|
||
|
|
C[(grow + i * RSIZE) * N + gcolumn] = result[i];
|
||
|
|
}
|
||
|
|
|
||
|
|
}
|