def test_accel(A, B, w_width, a_width):
assert A.shape[1] == B.shape[1], "sliced shape not match"
ctx = tvm.cpu(0)
f = tsim.load_module()
a_arr = []
b_arr = []
packedA = pack_8(A)
packedB = pack_8(B)
dtype = 'uint64'
for i in range(packedA.shape[0]):
a_arr.append(tvm.nd.array(np.array([packedA[i]]).astype(dtype), ctx))
for i in range(packedB.shape[0]):
b_arr.append(tvm.nd.array(np.array([packedB[i]]).astype(dtype), ctx))
cycles = 0
accum = tvm.nd.array(np.array([0]).astype("uint64"), ctx)
for i in range(len(a_arr)):
for j in range(len(b_arr)):
shift = np.uint8(i * w_width + j * a_width)
if i == 0 and j == 0:
cycles += f(a_arr[i], b_arr[j], shift, accum,
np.uint32(1)) # reset accumulator
else:
cycles += f(a_arr[i], b_arr[j], shift, accum,
np.uint32(0)) # no reset
return (accum.asnumpy()[0], cycles)
def test_accel(A, B, i_width, w_width):
assert A.shape[1] == B.shape[2], "sliced shape not match"
dtype = A.dtype
ctx = tvm.cpu(0)
f = tsim.load_module()
a_arr = []
b_arr = []
for i in range(A.shape[0]):
list_a = np.zeros(A.shape[1]).astype(dtype)
for j in range(A.shape[1]):
list_a[j] = A[i][j]
a_arr.append(tvm.nd.array(list_a.astype(dtype), ctx))
for i in range(B.shape[0]):
# transpose
list_b = np.zeros((B.shape[2], B.shape[1])).astype(dtype)
for j in range(B.shape[2]):
for k in range(B.shape[1]):
list_b[j][k] = B[i][j][k]
b_arr.append(tvm.nd.array(list_b.astype(dtype), ctx))
cycles = 0
accum = tvm.nd.array(np.zeros(A.shape[1]).astype("uint32"), ctx)
for i in range(len(a_arr)):
for j in range(len(b_arr)):
shift = np.uint8(i * i_width + j * w_width)
if i == 0 and j == 0:
cycles += f(b_arr[j], a_arr[i], shift, accum,
np.uint32(1)) # reset accumulator
else:
cycles += f(b_arr[j], a_arr[i], shift, accum,
np.uint32(0)) # no reset
return (accum.asnumpy(), cycles)
def test_accel():
rmax = 64
dtype = "uint64"
n = 1 << np.random.randint(0, 5)
ctx = tvm.cpu(0)
a = tvm.nd.array(np.random.randint(rmax, size=n).astype(dtype), ctx)
b = tvm.nd.array(np.random.randint(rmax, size=n).astype(dtype), ctx)
c = tvm.nd.array(np.zeros(n).astype(dtype), ctx)
f = tsim.load_module()
f(a, b, c)
msg = "n:{}".format(n)
np.testing.assert_equal(c.asnumpy(), a.asnumpy() + b.asnumpy(), err_msg = "[FAIL] " + msg)
print("[PASS] " + msg)
def test_accel():
rmax = 64
dtype = "uint64"
n = np.random.randint(1, rmax)
c = np.random.randint(0, rmax)
ctx = tvm.cpu(0)
a = tvm.nd.array(np.random.randint(rmax, size=n).astype(dtype), ctx)
b = tvm.nd.array(np.zeros(n).astype(dtype), ctx)
f = tsim.load_module()
cycles = f(a, b, c)
msg = "cycles:{0:4} n:{1:2} c:{2:2}".format(cycles, n, c)
np.testing.assert_equal(b.asnumpy(), a.asnumpy() + c, err_msg = "[FAIL] " + msg)
print("[PASS] " + msg)
def test_accel(A, B, slice_width):
assert A.shape == B.shape, "sliced shape not match"
dtype = A.dtype
ctx = tvm.cpu(0)
f = tsim.load_module()
print(dtype)
a_arr = []
b_arr = []
# print("this is driver array")
for i in range(A.shape[0]):
list_a = np.zeros(A.shape[1]).astype(dtype)
list_b = np.zeros(A.shape[1]).astype(dtype)
for j in range(A.shape[1]):
list_a[j] = A[i][j]
list_b[j] = B[i][j]
# print("a[" + str(i) + "]")
# print(list_a)
# print("b[" + str(i) + "]")
# print(list_b)
a_arr.append(tvm.nd.array(list_a.astype(dtype), ctx))
b_arr.append(tvm.nd.array(list_b.astype(dtype), ctx))
cycles = 0
print("here")
accum = tvm.nd.array(np.array([0]).astype("uint64"), ctx)
for i in range(len(a_arr)):
for j in range(len(a_arr)):
shift = np.uint8((i + j) * slice_width)
if i == 0 and j == 0:
cycles += f(a_arr[i], b_arr[j], shift, accum,
np.uint32(1)) # reset accumulator
else:
cycles += f(a_arr[i], b_arr[j], shift, accum,
np.uint32(0)) # no reset
print(a_arr[i])
print(b_arr[j])
print("\n")
# cycles = f(a, b, c)
# msg = "cycles:{0:4} n:{1:2} c:{2:2}".format(cycles, n, c)
# np.testing.assert_equal(b.asnumpy(), a.asnumpy() + c, err_msg = "[FAIL] " + msg)
# print("[PASS] " + msg)
return (accum.asnumpy()[0], cycles)