119 lines
3.6 KiB
Common Lisp
119 lines
3.6 KiB
Common Lisp
|
// super super slow sgemm kernel by heehoon
|
||
|
|
||
|
|
||
|
#define TS 32 // The square-root of the 2D tile-size (== work-group dims)
|
||
|
#define WPT 8 // The amount of work-per-thread, i.e. the thread-coarsening factor
|
||
|
#define RTS (TS/WPT) // The reduced tile-size in one dimension
|
||
|
|
||
|
|
||
|
__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); // Local row ID (max: TS)
|
||
|
const int col = get_local_id(1); // Local col ID (max: TS/WPT == RTS)
|
||
|
|
||
|
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 memory to fit a tile of TS*TS elements of A and B
|
||
|
__local float Asub[TS][TS];
|
||
|
__local float Bsub[TS][TS];
|
||
|
|
||
|
float acc = 0.0f;
|
||
|
const int numTiles = (K-1) / TS + 1;
|
||
|
|
||
|
for(int t=0; t<numTiles; t++)
|
||
|
{
|
||
|
const int t_row = 32*t+row;
|
||
|
const int t_col = 32*t+col;
|
||
|
|
||
|
if(t_col < K && globalRow < M) Asub[row][col] = A[globalRow * K + t_col];
|
||
|
else
|
||
|
Asub[row][col] = 0;
|
||
|
|
||
|
if(t_row < K && globalCol < N ) Bsub[row][col] = B[t_row * N + globalCol];
|
||
|
else
|
||
|
Bsub[row][col] = 0;
|
||
|
|
||
|
// Synchronise to make sure the tile is loaded
|
||
|
barrier(CLK_LOCAL_MEM_FENCE);
|
||
|
|
||
|
for(int k=0; k<32; k++)
|
||
|
{
|
||
|
acc += Asub[row][k] * Bsub[k][col];
|
||
|
}
|
||
|
|
||
|
// Synchronise to make sure the tile is loaded
|
||
|
barrier(CLK_LOCAL_MEM_FENCE);
|
||
|
}
|
||
|
if (globalRow < M && globalCol < N) {
|
||
|
C[globalRow*N + globalCol] = acc;
|
||
|
}
|
||
|
}
|
||
|
*/
|
||
|
|
||
|
// 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 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 memory to fit a tile of TS*TS elements of A and B
|
||
|
__local float Asub[TS][TS];
|
||
|
__local float Bsub[TS][TS];
|
||
|
|
||
|
// Initialise the accumulation registers
|
||
|
float acc[WPT];
|
||
|
for (int w=0; w<WPT; w++) {
|
||
|
acc[w] = 0.0f;
|
||
|
}
|
||
|
|
||
|
// Loop over all tiles
|
||
|
const int numTiles = (K+ TS -1) / TS;
|
||
|
for (int t=0; t<numTiles; t++) {
|
||
|
|
||
|
// Load one tile of A and B into local memory
|
||
|
for (int w=0; w<WPT; w++) {
|
||
|
const int tiledRow = TS*t + row;
|
||
|
const int tiledCol = TS*t + col;
|
||
|
|
||
|
if(globalRow + w*RTS >= M || tiledCol >=K)
|
||
|
{
|
||
|
Asub[row + w*RTS][col] = 0.0f;
|
||
|
}
|
||
|
else
|
||
|
{
|
||
|
Asub[row + w*RTS][col] = A[(globalRow + w*RTS)*K + tiledCol];
|
||
|
}
|
||
|
|
||
|
if(tiledRow + w*RTS >= K || globalCol >= N)
|
||
|
{
|
||
|
Bsub[row + w*RTS][col] = 0.0f;
|
||
|
}
|
||
|
else
|
||
|
{
|
||
|
Bsub[row + w*RTS][col] = B[(tiledRow + w*RTS)*N + globalCol];
|
||
|
}
|
||
|
}
|
||
|
|
||
|
// Synchronise to make sure the tile is loaded
|
||
|
barrier(CLK_LOCAL_MEM_FENCE);
|
||
|
|
||
|
// Perform the computation for a single tile
|
||
|
for (int k=0; k<TS; k++) {
|
||
|
for (int w=0; w<WPT; w++) {
|
||
|
acc[w] += Asub[row + w*RTS][k] * Bsub[k][col];
|
||
|
}
|
||
|
}
|
||
|
|
||
|
// Synchronise before loading the next tile
|
||
|
barrier(CLK_LOCAL_MEM_FENCE);
|
||
|
}
|
||
|
|
||
|
// Store the final results in C
|
||
|
for (int w=0; w<WPT; w++) {
|
||
|
if (globalRow + w*RTS >= M || globalCol >= N) continue;
|
||
|
C[(globalRow + w*RTS)*N + globalCol] = acc[w];
|
||
|
}
|
||
|
|
||
|
}
|