64 lines
1.4 KiB
Common Lisp
64 lines
1.4 KiB
Common Lisp
// super super slow sgemm kernel by heehoon
|
|
|
|
#define TS 32
|
|
#define WPT 8
|
|
#define RTS TS/WPT
|
|
|
|
__kernel void sgemm(__global float *A, __global float *B, __global float *C, int M, int N, int K) {
|
|
|
|
int row = get_local_id(0);
|
|
int col = get_local_id(1);
|
|
int global_row = TS*get_group_id(0) + row ; // row index of C
|
|
int global_col = TS*get_group_id(1) + col ; // column index of C
|
|
|
|
__local float Asub[TS*TS];
|
|
__local float Bsub[TS*TS];
|
|
|
|
float acc[WPT];
|
|
for (int w=0; w<WPT; w++) {
|
|
acc[w] = 0.0f;
|
|
}
|
|
|
|
const int numTiles = (K+TS-1)/TS;
|
|
for (int t=0; t<numTiles; t++) {
|
|
for (int w=0; w < WPT; w++) {
|
|
|
|
const int tiled_row = TS*t + row;
|
|
const int tiled_col = TS*t + col;
|
|
|
|
if(global_row + w*RTS >=M || tiled_col >=K) {
|
|
Asub[(row + w*RTS)*TS + col] = 0.0f;
|
|
}
|
|
else {
|
|
Asub[(row + w*RTS)*TS + col] = A[(global_row + w*RTS)*K + tiled_col];
|
|
|
|
}
|
|
|
|
if(tiled_row + w*RTS >=K || global_col >= N){
|
|
Bsub[(row + w*RTS)*TS + col] = 0.0f;
|
|
}
|
|
else {
|
|
Bsub[(row + w*RTS)*TS + col] = B[(tiled_row + w*RTS)*N + global_col];
|
|
}
|
|
|
|
}
|
|
|
|
barrier(CLK_LOCAL_MEM_FENCE);
|
|
|
|
for (int k = 0; k < TS; k++) {
|
|
//acc += Asub[i * K + k] * B[k * N + j];
|
|
for (int w = 0; w< WPT; w++) {
|
|
acc[w] += Asub[(row + w*RTS)* TS + k] * Bsub[k * TS + col];
|
|
}
|
|
}
|
|
|
|
barrier(CLK_LOCAL_MEM_FENCE);
|
|
|
|
}
|
|
for (int w=0; w<WPT; w++){
|
|
if((global_row+w*RTS < M) && (global_col < N)) {
|
|
C[(global_row+w*RTS)*N + global_col] = acc[w];
|
|
}
|
|
}
|
|
}
|