56 lines
1.6 KiB
Common Lisp
56 lines
1.6 KiB
Common Lisp
|
#define TS 32
|
||
|
|
#define WPT 16
|
||
|
|
#define RTS (int)(TS/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) {
|
||
|
|
// Thread identifiers
|
||
|
|
// Thread identifiers
|
||
|
|
const int row = get_local_id(0); // Local row ID (max: TS/WPT == RTS)
|
||
|
|
const int col = get_local_id(1); // Local col ID (max: TS)
|
||
|
|
const int globalRow = TS*get_group_id(0) + row; // Row ID of C (0..M)
|
||
|
|
const int globalCol = 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 numTiles = (K + TS -1)/TS;
|
||
|
|
for (int t=0; t<numTiles; t++) {
|
||
|
|
|
||
|
|
for (int w=0; w<WPT; w++) {
|
||
|
|
const int tiledRow = TS*t + row;
|
||
|
|
const int tiledCol = TS*t + col;
|
||
|
|
|
||
|
|
if(tiledCol < K && globalRow+w*RTS < M)
|
||
|
|
Asub[row + w*RTS][col] = A[(globalRow + w*RTS)*K + tiledCol];
|
||
|
|
else
|
||
|
|
Asub[row + w*RTS][col] = 0.0f;
|
||
|
|
|
||
|
|
if(globalCol < N && tiledRow+w*RTS < K)
|
||
|
|
Bsub[row + w*RTS][col] = B[(tiledRow + w*RTS)*N + globalCol];
|
||
|
|
else
|
||
|
|
Bsub[row + w*RTS][col] = 0.0f;
|
||
|
|
}
|
||
|
|
|
||
|
|
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(globalRow + w*RTS < M && globalCol < N)
|
||
|
|
C[(globalRow + w*RTS)*N + globalCol] = acc[w];
|
||
|
|
}
|
||
|
|
}
|