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

tensor: Fix dimensions for fp16 in script

Browse Source
This commit is contained in:
Hansung Kim
2024-08-12 15:20:27 -07:00
parent c1906ebb4f
commit 07dd9e35a0
2 changed files with 25 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
tests/kernel/tensor/check_correctness.py
View File

@@ -41,9 +41,13 @@ def C_index(threadgroup, thread, register):
def check_sim_output(): def check_sim_output():
file = input("simulator output filename: ") file = input("simulator output filename: ")
A_array = np.zeros((16, 8)) M = 8
B_array = np.zeros((8, 16)) N = 8
C_array = np.zeros((16, 16)) K = 16
A_array = np.zeros((M, K))
B_array = np.zeros((K, N))
C_array = np.zeros((M, N))
with open(file) as f: with open(file) as f:
for line in f.readlines(): for line in f.readlines():
@@ -87,26 +91,23 @@ def check_sim_output():
if __name__ == "__main__": if __name__ == "__main__":
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_A) print(expected_A)
print('expected B:') print('expected B:')
print(expected_B) print(expected_B)
print('expected C:') print('expected C:')
print(expected_C[0:8, 0:8]) print(expected_C)
expected_C.astype('float32').tofile("c_expected.bin") expected_C.astype('float32').tofile("c_expected.bin")
[got_A, got_B, got_C] = check_sim_output() [got_A, got_B, got_C] = check_sim_output()
print('got C:') print('got C:')
print(C_array[0:8, 0:8]) print(C_array)
print('diff C:') print('diff C:')
print(expected_C[0:8, 0:8] - C_array[0:8, 0:8]) print(expected_C - C_array)
assert np.allclose(expected_A, got_A) assert np.allclose(expected_A, got_A)
assert np.allclose(expected_B, got_B) assert np.allclose(expected_B, got_B)
assert np.allclose(expected_C, got_C) assert np.allclose(expected_C, got_C)

20
tests/kernel/tensor/generate_matrix.py
View File

@@ -1,10 +1,14 @@
import numpy as np import numpy as np
M = 8
N = 8
K = 16
# A_array = np.random.rand(8, 16) # A_array = np.random.rand(8, 16)
A_array = np.arange(8 * 8).reshape([8, 8]) A_array = np.arange(M * K).reshape([M, K])
B_array = np.arange(8 * 8).reshape([8, 8]) B_array = np.arange(K * N).reshape([K, N])
# C_array = np.random.rand(16, 16) # C_array = np.random.rand(16, 16)
C_array = np.zeros([8, 8]) C_array = np.zeros([M, N])
# A_array = np.zeros((16, 8)) # A_array = np.zeros((16, 8))
# B_array = np.zeros((8, 16)) # B_array = np.zeros((8, 16))
# A_array[0,:] = 1.0 # A_array[0,:] = 1.0
@@ -56,10 +60,10 @@ if __name__ == "__main__":
np.savez("abc", A_array=A_array, B_array=B_array, C_array=C_array) np.savez("abc", A_array=A_array, B_array=B_array, C_array=C_array)
A_array.astype('float32').tofile("input.a.bin") # A_array.astype('float32').tofile("input.a.bin")
B_array.astype('float32').tofile("input.b.bin") # B_array.astype('float32').tofile("input.b.bin")
# A_array.astype('float16').tofile("input.a.bin") A_array.astype('float16').tofile("input.a.bin")
# B_array = pack_fp16_by_column(B_array) B_array = pack_fp16_by_column(B_array)
# B_array.astype('float16').tofile("input.b.bin") B_array.astype('float16').tofile("input.b.bin")
print(B_array) print(B_array)