53 lines
1.5 KiB
Common Lisp
53 lines
1.5 KiB
Common Lisp
#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) {
|
|
// Thread identifiers
|
|
const int row = get_local_id(0); // Local row ID (max: TS)
|
|
const int col = get_local_id(1); // Local col ID (max: TS/WPT == RTS)
|
|
const int global_row = TS*get_group_id(0) + row; // Row ID of C (0..M)
|
|
const int global_col = TS*get_group_id(1) + col; // Col ID of C (0..N)
|
|
|
|
__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 num_tiles = (K + TS-1) / TS;
|
|
for (int t = 0; t < num_tiles; t++) {
|
|
for (int w=0; w<WPT; w++) {
|
|
const int t_row = TS * t + row;
|
|
const int t_col = TS * t + col;
|
|
|
|
if( t_col < K && global_row < M )
|
|
Asub[row + w*RTS][col] = A[(global_row + w*RTS)*K + t_col];
|
|
else
|
|
Asub[row + w*RTS][col] = 0;
|
|
|
|
if( t_row < K && global_col < N )
|
|
Bsub[row + w*RTS][col] = B[(t_row + w*RTS)*N + global_col];
|
|
else
|
|
Bsub[row + w*RTS][col] = 0;
|
|
}
|
|
|
|
barrier(CLK_LOCAL_MEM_FENCE);
|
|
|
|
for (int k = 0; k < TS; k++) {
|
|
for (int w=0; w<WPT; w++) {
|
|
acc[w] += Asub[row + w*RTS][k] * Bsub[k][col];
|
|
}
|
|
}
|
|
|
|
barrier(CLK_LOCAL_MEM_FENCE);
|
|
}
|
|
|
|
for (int w=0; w<WPT; w++) {
|
|
if (global_row < M && global_col < N)
|
|
C[(global_row + w*RTS)*N + global_col] = acc[w];
|
|
}
|
|
|
|
} |