|
|
@@ -1898,10 +1898,11 @@ kernel void kernel_mul_mm(device const uchar * src0,
|
|
|
threadgroup float *temp_str = ((threadgroup float *)shared_memory) \
|
|
|
+ 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
|
|
|
for (int i = 0; i < 8; i++) {
|
|
|
+ threadgroup_barrier(mem_flags::mem_device);
|
|
|
simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
|
|
|
}
|
|
|
|
|
|
- threadgroup_barrier(mem_flags::mem_threadgroup);
|
|
|
+ threadgroup_barrier(mem_flags::mem_device);
|
|
|
device float *C = dst + BLOCK_SIZE_M * r0 + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
|
|
|
if (sgitg==0) {
|
|
|
for (int i = 0; i < n_rows; i++) {
|
Shouzheng Liu