wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

tensor: generate 8x8 in correctness script

Browse Source
This commit is contained in:
Hansung Kim
2024-06-07 18:13:57 -07:00
parent 483f975439
commit 2cac995db9
2 changed files with 12 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
kernel/src/vx_spawn.c
View File

@@ -254,6 +254,7 @@ void vx_spawn_tasks_cluster(int num_tasks, vx_spawn_tasks_cb callback, void *arg
vx_wspawn_wait(); vx_wspawn_wait();
} }
// TODO: this is incomplete
// TODO: Instead of launching an additional wave just to work on remaining // TODO: Instead of launching an additional wave just to work on remaining
// threads, handle this in the last wave amongst other full warps. // threads, handle this in the last wave amongst other full warps.
if (rem_threads_in_last_warp != 0 && core_id_in_cluster == 0) { if (rem_threads_in_last_warp != 0 && core_id_in_cluster == 0) {

15
tests/kernel/tensor/check_correctness.py
View File

@@ -82,16 +82,23 @@ with open(file) as f:
expected = np.load("abc.npz") expected = np.load("abc.npz")
expected_A = expected['A_array'] # expected_A = expected['A_array']
expected_B = expected['B_array'] # expected_B = expected['B_array']
expected_C = expected['C_array'] # expected_C = expected['C_array']
expected_A = expected['A_array'][0:8, 0:8]
expected_B = expected['B_array'][0:8, 0:8]
expected_C = expected['C_array'][0:8, 0:8]
expected_C = expected_C + expected_A @ expected_B expected_C = expected_C + expected_A @ expected_B
print('expected A:')
print(expected_A)
print('expected B:')
print(expected_B)
print('expected C:') print('expected C:')
print(expected_C[0:8, 0:8]) print(expected_C[0:8, 0:8])
print('got C:') print('got C:')
print(C_array[0:8, 0:8]) print(C_array[0:8, 0:8])
print('diff C:') print('diff C:')
print((expected_C - C_array)[0:8, 0:8]) print(expected_C[0:8, 0:8] - C_array[0:8, 0:8])
expected_C.astype('float32').tofile("c_expected.bin") expected_C.astype('float32').tofile("c_expected.bin")