121 lines
3.0 KiB
Common Lisp
121 lines
3.0 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) {
|
|
|
|
|
|
const int row = get_local_id(0);
|
|
const int col = get_local_id(1);
|
|
const int global_row = TS * get_group_id(0) + row;
|
|
const int global_col = TS * get_group_id(1) + col;
|
|
|
|
__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;
|
|
for (int t=0; t<numTiles; t++) {
|
|
|
|
for (int w=0; w<WPT; w++) {
|
|
const int t_row = TS * t + row;
|
|
const int t_col = TS * t + col;
|
|
Asub[row + w*RTS][col] = A[(global_row + w*RTS)*K + t_col];
|
|
Bsub[row + w*RTS][col] = B[(t_row + w*RTS)*N + global_col];
|
|
}
|
|
|
|
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++) {
|
|
C[(global_row + w*RTS)*N + global_col] = acc[w];
|
|
}
|
|
|
|
//if(global_row == 0 && global_col == 0){
|
|
// printf("A 0 0 : %f\n", A[0]);
|
|
// printf("A 0 1 : %f\n", A[1]);
|
|
// printf("B 0 0 : %f\n", B[0]);
|
|
// printf("B 1 0 : %f\n", B[N]);
|
|
//}
|
|
|
|
|
|
}
|
|
|
|
|
|
__kernel void paddingAddZeroes(const int M, const int N, const __global float* input, const int M_XL, const int N_XL,__global float* output) {
|
|
|
|
// Thread identifiers
|
|
const int row = get_group_id(0)*16 + get_local_id(0); // 0..P_XL
|
|
const int col = get_group_id(1)*16 + get_local_id(1); // 0..Q_XL
|
|
|
|
// Check whether we are within bounds of the XL matrix
|
|
if (row < M_XL && col < N_XL) {
|
|
|
|
// Copy the input or pad a zero
|
|
float value;
|
|
if (row < M && col < N) {
|
|
value = input[row*N + col];
|
|
}
|
|
else {
|
|
value = 0.0f;
|
|
}
|
|
|
|
// Store the result
|
|
output[row*N_XL + col] = value;
|
|
}
|
|
}
|
|
|
|
__kernel void mergeResult(__global float * C1, __global float * C2, __global float * C3, __global float * C4, __global float * result, int M_xl, int N_xl, int M_single, int M, int N){
|
|
|
|
const int row = get_group_id(0) * 32 + get_local_id(0); // M
|
|
const int col = get_group_id(1) * 32 + get_local_id(1); // N
|
|
|
|
//if(row == 0 && col == 0){
|
|
// printf("M_xl : %d\n", M_xl);
|
|
// printf("N_xl : %d\n", N_xl);
|
|
// printf("M : %d\n", M);
|
|
// printf("N : %d\n", N);
|
|
//}
|
|
|
|
if(row >= M || col >= N) return;
|
|
|
|
float value;
|
|
int aligned_row;
|
|
|
|
if(row < M_single){
|
|
aligned_row = row;
|
|
value = C1[aligned_row * N_xl + col];
|
|
}
|
|
else if(row < 2*M_single){
|
|
aligned_row = row - M_single;
|
|
value = C2[aligned_row * N_xl + col];
|
|
}
|
|
else if(row < 3*M_single){
|
|
aligned_row = row - 2*M_single;
|
|
value = C3[aligned_row * N_xl + col];
|
|
}
|
|
else{
|
|
aligned_row = row - 3*M_single;
|
|
value = C4[aligned_row * N_xl + col];
|
|
}
|
|
|
|
result[row * N + col] = value;
|
|
|
|
|
|
}
|