chundoong-lab-ta/SamsungDS22/submissions/HW5/seong81.kim/kernel.cl

69 lines
2.3 KiB
Common Lisp

#define TS 64
#define WPT 32
#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 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 tempVal[WPT];
for (int w=0; w<WPT; w++)
tempVal[w] = 0.0f;
// Loop over all tiles
const int numTiles = (K + TS - 1) / TS ; // Ceiling to Tile Size
//const int numTiles = (K / TS) + 1 - (K % TS == 0);
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;
const int orgM = globalRow + w*RTS ;
const int orgK = tiledRow + w*RTS ;
// Copy Global A to local
if ( (orgM < M) && (tiledCol < K) )
Asub[row + w*RTS][col] = A[orgM*K + tiledCol] ;
else
Asub[row + w*RTS][col] = 0 ;
// Copy Global B to local
if ( (orgK < K) && (globalCol < N) )
Bsub[row + w*RTS][col] = B[orgK*N + globalCol] ;
else
Bsub[row + w*RTS][col] = 0 ;
}
// 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++) {
tempVal[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++) {
const int orgM = globalRow + w*RTS ;
if ( (orgM < M) && (globalCol < N) )
C[orgM * N + globalCol] = tempVal[w];
}
}