def conv_backprop_filter_run(fmap_shape, filter_shape, pad_, stride_, dilation_, attrs=None):
block_size = 16
conv_dtype = 'float16'
in_n, in_c, in_h, in_w = fmap_shape
cout, cin, w_h, w_w = filter_shape
in_c = (in_c + block_size - 1) // block_size * block_size
cout = (cout + block_size - 1) // block_size * block_size
pad_top, pad_bottom, pad_left, pad_right = pad_
stride_h, stride_w = stride_
out_n = in_n
out_c = cout
out_h = (in_h + pad_top + pad_bottom - w_h) // stride_h + 1
out_w = (in_w + pad_left + pad_right - w_w) // stride_w + 1
x_shape = (in_n, in_c, in_h, in_w)
w_shape = (cout, in_c, w_h, w_w)
y_shape = (out_n, out_c, out_h, out_w)
inN, inC, inH, inW = x_shape
input_shape_nc1hwc0 = (inN, inC // block_size, inH, inW, block_size)
o_n, o_c, o_h, o_w = y_shape
y_shape_nc1hwc0 = (o_n, o_c // block_size, o_h, o_w, block_size)
input_shape = [y_shape_nc1hwc0, input_shape_nc1hwc0]
input_file = os.environ.get("RANDOM_DATA_DISK_PATH", "")
expect_file = input_file + "/" + gen_kernel_name([input_shape], [conv_dtype],
op_attrs=[fmap_shape, filter_shape, pad_, stride_, dilation_],
kernel_name='conv_backprop_filter', attrs=attrs) + ".bin"
print("gen_data begin.")
dy_data, dx_data, expect = gen_data(x_shape, w_shape, pad_, stride_, dilation_, expect_file, attrs=attrs)
assert (dy_data.shape == y_shape_nc1hwc0)
print("gen_data finished.")
out_data = np.full(expect.shape, 0, 'float32')
input = (dy_data, dx_data)
flag_w = os.environ.get("WRITE_TO_DISK", "No")
if flag_w == "Yes":
return input, out_data, expect, True
mod = utils.op_build_test(conv_backprop_filter, [input_shape], [conv_dtype],
op_attrs=[fmap_shape, filter_shape, pad_, stride_, dilation_],
kernel_name='conv_backprop_filter', attrs=attrs)
args = (dy_data, dx_data, out_data)
out_data = utils.mod_launch(mod, args, expect=expect)
rtol, atol = get_rtol_atol("conv_backprop_filter", conv_dtype)
return input, out_data, expect, compare_tensor(out_data, expect, rtol=rtol, atol=atol, equal_nan=True)
def rint_run(shape, dtype, attrs=None):
"""rint_run"""
if attrs is None:
attrs = {}
mod = utils.op_build_test(rint.rint, [shape], [dtype], kernel_name='rint', attrs=attrs)
args, exp_output, input_x = gen_data(shape, dtype)
acu_output = utils.mod_launch(mod, args, expect=exp_output)
# compare result
rtol, atol = get_rtol_atol("rint", dtype)
testcase_result = compare_tensor(acu_output, exp_output, rtol=rtol, atol=atol, equal_nan=True)
return input_x, acu_output, exp_output, testcase_result
def laplacian_of_gaussian_ad_run(shape, dtype, attrs):
expect, head_np, input_np = gen_data(dtype, shape)
mod = utils.op_build_test(laplacian_of_gaussian_ad, [head_np.shape, shape],
[dtype, dtype],
kernel_name='mulexp',
attrs=attrs)
output = np.full(expect.shape, np.nan, dtype)
output = utils.mod_launch(mod, (head_np, input_np, output), expect=expect)
rtol, atol = get_rtol_atol("laplacian_of_gaussian", dtype)
return (head_np, input_np), output, expect, compare_tensor(output,
expect,
rtol=rtol,
atol=atol)
def bessel_i1e_run(x_shape, x_dtype, attrs):
shapes = [x_shape]
dtypes = [x_dtype]
mod = utils.op_build_test(bessel_i1e,
shapes,
dtypes,
kernel_name="bessel_i1e",
attrs=attrs)
bench_mark, inputs, output = gen_data(dtypes, shapes)
output = utils.mod_launch(mod, inputs + [output], expect=bench_mark)
rtol, atol = get_rtol_atol("bessel_i1e", dtypes[0])
compare_res = compare_tensor(output, bench_mark, rtol=rtol, atol=atol)
return inputs, output, bench_mark, compare_res
def conv_filter_ad_run(fmap_shape, filter_shape, pad_, stride_, dilation_, attrs=None):
block_size = 16
conv_dtype = 'float16'
in_n, in_c, in_h, in_w = fmap_shape
cout, cin, w_h, w_w = filter_shape
assert(in_c == cin)
in_c = (in_c + block_size - 1) // block_size * block_size
cout = (cout + block_size - 1) // block_size * block_size
pad_top, pad_bottom, pad_left, pad_right = pad_
stride_h, stride_w = stride_
out_n = in_n
out_c = cout
out_h = (in_h + pad_top + pad_bottom - w_h) // stride_h + 1
out_w = (in_w + pad_left + pad_right - w_w) // stride_w + 1
x_shape = (in_n, in_c, in_h, in_w)
w_shape = (cout, in_c, w_h, w_w)
x_5D_shape = (in_n, in_c // block_size, in_h, in_w, block_size)
y_5D_shape = (out_n, out_c // block_size, out_h, out_w, block_size)
forward_input_output_shapes = [y_5D_shape, x_5D_shape]
dw_input_shapes = [y_5D_shape, x_5D_shape]
input_file = os.environ.get("RANDOM_DATA_DISK_PATH", "")
expect_file = input_file + "/" + gen_kernel_name([dw_input_shapes], [conv_dtype],
op_attrs=[fmap_shape, filter_shape, pad_, stride_, dilation_],
kernel_name='conv_filter_ad', attrs=attrs) + ".bin"
print("gen_data begin.")
dy_data, dx_data, expect = gen_data_dw(x_shape, w_shape, pad_, stride_, dilation_, expect_file, attrs=attrs)
print("gen_data finished.")
out_data = np.full(expect.shape, 0, 'float32')
np_input = (dy_data, dx_data)
flag_w = os.environ.get("WRITE_TO_DISK", "No")
if flag_w == "Yes":
return np_input, out_data, expect, True
mod = utils.op_build_test(ConvFilterAd, [dw_input_shapes], [conv_dtype],
op_attrs=[fmap_shape, filter_shape, pad_, stride_, dilation_], kernel_name='conv_filter_ad', attrs=attrs, dump_code = True)
args = (dy_data, dx_data, out_data)
out_data = utils.mod_launch(mod, args, expect=expect)
rtol, atol = get_rtol_atol("conv_filter_ad", conv_dtype)
assert_res = compare_tensor(out_data, expect, rtol=rtol, atol=atol, equal_nan=True)
return np_input, out_data, expect, assert_res
def batchmatmul_execute(bs, m, n, k, bias_shape, dtype, trans_a, trans_b,
kernel_name, attrs):
# Generate data
_, _, out_shape = get_shape(bs, m, n, k, trans_a, trans_b)
# this part is for auto-tuning
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = batchmatmul_compile(bs, m, n, k, bias_shape, dtype, trans_a,
trans_b, kernel_name, attrs, t)
if t:
m_a, m_b, matrix_bias, expect = gen_data(bs, m, n, k, bias_shape,
trans_a, trans_b, dtype)
output = np.full(out_shape, np.nan, dtype=dtype)
return mod, expect, (m_a, m_b, matrix_bias,
output) if len(bias_shape) > 0 else (m_a, m_b,
output)
else:
return mod
mod, args = batchmatmul_compile(bs, m, n, k, bias_shape, dtype, trans_a,
trans_b, kernel_name, attrs)
m_a, m_b, matrix_bias, expect = gen_data(bs, m, n, k, bias_shape, trans_a,
trans_b, dtype)
output = np.full(out_shape, np.nan, dtype=dtype)
launch_args = []
outputs = []
if len(bias_shape) > 0:
launch_args = [m_a, m_b, matrix_bias, output]
outputs = [
3,
]
else:
launch_args = [m_a, m_b, output]
outputs = [
2,
]
if attrs.get("dynamic"):
launch_args = launch_args + args
block_dim = compute_blockdim(bs, (m, n))
launch_args.append(block_dim)
output = utils.mod_launch(mod, launch_args, outputs=outputs, expect=expect)
rtol, atol = get_rtol_atol("batchmatmul", dtype)
print("-----------------")
print(output.dtype)
print(expect.dtype)
res = compare_tensor(output, expect, rtol=rtol, atol=atol, equal_nan=True)
# if not res:
# res = utils.double_fivethou_compare(output, expect, r_tol=5e-3, equal_nan=True)
return (m_a, m_b), output, expect, res
def matmul_execute(shape_x, shape_y, bias, left_format, right_format,
out_format, adj_x, adj_y, dtype, out_dtype, kernel_name,
attrs):
'''
There are four types of fractal format in Davinci core: zZ, zN, nZ, nN
general matmul format
left_trans: False right_trans False: zZ * nZ = zN
left_trans: True right_trans False: nN * nZ = zN
left_trans: False right_trans True : zZ * zN = zN
left_trans: True right_trans True : nN * zN = zN
Now we need to support: zN * nZ = zN
use left_format to specify, left matrix data format
use right_format to specify, right matrix data format
'''
batch_tuple, m, k, n = extract_dim(shape_x, shape_y, adj_x, adj_y)
m = (m + 15) // 16 * 16
n = (n + 15) // 16 * 16
k = (k + 15) // 16 * 16
shape_xx, shape_yy, bias_shape, out_shape, k = get_converted_shapes(
m, n, k, batch_tuple, adj_x, adj_y, bias, left_format, right_format,
out_format)
mod = dynamic_matmul_compile(shape_x, shape_y, bias, left_format,
right_format, out_format, adj_x, adj_y, dtype,
out_dtype, kernel_name, attrs)
# Generate data
m_x, m_y, bench_mark, bias_data = matmul_data(batch_tuple, m, k, n, dtype,
out_dtype, bias, adj_x,
adj_y, left_format,
right_format, out_format)
# mod launch
output = np.full(out_shape, np.nan, out_dtype)
if bias == 0:
output = utils.mod_launch(
mod, (m_x, m_y, output, 1, 1, 1, 1, 1, 1, 1, 1, 1),
outputs=(2, ),
expect=bench_mark)
elif bias == 1:
output = utils.mod_launch(mod, (m_x, m_y, bias_data, output),
expect=bench_mark)
# compare result
rtol, atol = get_rtol_atol("matmul", dtype)
compare_result = compare_tensor(output,
bench_mark,
rtol=rtol,
atol=atol,
equal_nan=True)
# compare_result = utils.result_compare(output, bench_mark, r_tol=5e-3)
return (m_x, m_y), output, bench_mark, compare_result
def acosh_grad_run(shape, dtype, attrs):
"""run function for dsl function acosh_grad."""
shapes = [shape, shape]
dtypes = [dtype, dtype]
mod = utils.op_build_test(acosh_grad,
shapes,
dtypes,
kernel_name="acosh_grad",
attrs=attrs)
bench_mark, inputs, output = gen_data(dtype, shape)
output = utils.mod_launch(mod, inputs + [output], expect=bench_mark)
rtol, atol = get_rtol_atol("acosh_grad", dtype)
compare_res = compare_tensor(output, bench_mark, rtol=rtol, atol=atol)
return inputs, output, bench_mark, compare_res
def sinh_run(shape, dtype, attrs=None):
mod = utils.op_build_test(Sinh, [shape], [dtype],
kernel_name="sinh",
attrs=attrs)
expect, inputs, output = gen_data(dtype, shape)
output = utils.mod_launch(mod, (inputs, output), expect=expect)
rtol, atol = get_rtol_atol("sinh", dtype)
TestCase_Result = compare_tensor(output,
expect,
rtol=rtol,
atol=atol,
equal_nan=False)
return inputs, output, expect, TestCase_Result
def div_no_nan_execute(shapes, dtype, attrs):
exp_output, inputs, args = gen_data(dtype, shapes)
mod = div_no_nan_compile(shapes, dtype, attrs)
# result_tvm
acu_output = utils.mod_launch(mod, args, expect=exp_output)
# compare result
rtol, atol = get_rtol_atol("div_no_nan", dtype)
TestCase_Result = compare_tensor(acu_output,
exp_output,
rtol=rtol,
atol=atol,
equal_nan=True)
return inputs, acu_output, exp_output, TestCase_Result
def truncate_div_run(shape1, dtype1, shape2, dtype2, attrs):
"""run function for truncate_div"""
expect, inputs, output = gen_data(dtype1, dtype2, shape1, shape2)
mod = utils.op_build_test(truncate_div.truncate_div, [shape1, shape2],
[dtype1, dtype2],
kernel_name="truncate_div",
attrs=attrs)
output = utils.mod_launch(mod, (*inputs, output), expect=expect)
rtol, atol = get_rtol_atol("truncate_div", dtype1)
return inputs, output, expect, compare_tensor(output,
expect,
rtol=rtol,
atol=atol,
equal_nan=True)
def clear_zero_run(shape, dtype, attrs, kernel_name="clear_zero"):
expect = np.full(shape, 0, dtype)
data = np.full(shape, np.nan, dtype)
inout = data
mod = utils.op_build_test(ClearZero, [shape], [dtype],
kernel_name=kernel_name)
inout = utils.mod_launch(mod, (inout, ), outputs=(-1, ), expect=expect)
rtol, atol = get_rtol_atol("clear_zero", dtype)
return (data), inout, expect, compare_tensor(inout,
expect,
rtol=rtol,
atol=atol,
equal_nan=True)
def asinh_run(x_shape, x_dtype, attrs):
"""run function for dsl function asinh."""
shapes = [x_shape]
dtypes = [x_dtype]
mod = utils.op_build_test(asinh,
shapes,
dtypes,
kernel_name="asinh",
attrs=attrs)
bench_mark, input_datas, output = gen_data(x_dtype, x_shape)
output = utils.mod_launch(mod, input_datas + [output], expect=bench_mark)
rtol, atol = get_rtol_atol("asinh", x_dtype)
compare_res = compare_tensor(output, bench_mark, rtol=rtol, atol=atol)
return input_datas, output, bench_mark, compare_res
def fused_bn_grad_5D_run_1(shape, dtype, kernel_name, attrs):
""" test bnGrad_1 """
def get_expect(dy, data, mean):
if dy.dtype == "float16":
dy = dy.astype("float32")
data = data.astype("float32")
data_minus_mean = data - mean
dgamma_red_hw = np.sum(dy * data_minus_mean, axis=(2,3), keepdims=True)
dbeta_red_hw = np.sum(dy, axis=(2,3), keepdims=True)
return [dgamma_red_hw, dbeta_red_hw, data_minus_mean]
shape_nc1c0 = (shape[0], shape[1], 1, 1, shape[4])
shape_c1c0 = (1, shape[1], 1, 1, shape[4])
bng1_shapes = [shape, shape, shape_c1c0]
bng1_dtypes = [dtype, dtype, "float32"]
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(fused_bn_grad1,
bng1_shapes, bng1_dtypes,
kernel_name=kernel_name + "_step1", attrs=attrs, tuning=t)
if t:
inputs = [np.random.rand(*s).astype(t) for (s, t) in zip(bng1_shapes, bng1_dtypes)]
inputs[2] = np.mean(inputs[1], axis=(0, 2, 3), keepdims=True).astype(bng1_dtypes[2])
out_shapes = [shape_nc1c0, shape_nc1c0, shape]
outputs = [np.full(s, np.nan, "float32") for s in out_shapes]
expects = get_expect(*inputs)
return mod, expects, {"args": (*inputs, *outputs), 'outputs': tuple(range(-len(outputs), 0)),
'tuning': False}
else:
return mod
mod = utils.op_build_test(fused_bn_grad1,
bng1_shapes, bng1_dtypes,
kernel_name=kernel_name + "_step1", attrs=attrs)
# np.random.seed(0)
inputs = [np.random.rand(*s).astype(t) for (s, t) in zip(bng1_shapes, bng1_dtypes)]
inputs[2] = np.mean(inputs[1], axis=(0, 2, 3), keepdims=True).astype(bng1_dtypes[2])
out_shapes = [shape_nc1c0, shape_nc1c0, shape]
outputs = [np.full(s, np.nan, "float32") for s in out_shapes]
outputs = list(utils.mod_launch(mod, (*inputs, *outputs), outputs=tuple(range(-len(outputs), 0)),
expect=get_expect(*inputs)))
expects = get_expect(*inputs)
rtol, atol = get_rtol_atol("fused_batch_norm_grad", dtype)
results = list(map(lambda x, y: np.allclose(x, y, rtol=rtol, atol=atol), outputs, expects))
print("results", results)
return inputs, outputs, expects, all(results)
def reduction_layer_execute(shape, dtype, axis, op, coeff, attrs):
exp_output, inputs, args = gen_data(shape, dtype, axis, op, coeff)
mod = reduction_layer_compile(shape, dtype, axis, op, coeff, attrs)
# result_tvm
acu_output = utils.mod_launch(mod, args, expect=exp_output)
# compare result
rtol, atol = get_rtol_atol("reduction_layer", dtype)
TestCase_Result = compare_tensor(acu_output,
exp_output,
rtol=rtol,
atol=atol,
equal_nan=True)
return inputs, acu_output, exp_output, TestCase_Result
def bitwise_and_run(shape1, dtype1, shape2, dtype2, kernel_name, attrs):
mod = utils.op_build_test(bitwise_and.bitwise_and, [shape1, shape2],
[dtype1, dtype2],
kernel_name=kernel_name,
attrs=attrs)
expect, inputs, output = gen_data(shape1, shape2, dtype1, dtype2)
actual = utils.mod_launch(mod, (*inputs, output), expect=expect)
rtol, atol = get_rtol_atol("bitwise_and", dtype1)
testcase_result = compare_tensor(actual,
expect,
rtol=rtol,
atol=atol,
equal_nan=True)
return input, actual, expect, testcase_result
def leaky_relu_execute(shape, dtype, negative_slop, attrs):
exp_output, inputs, args = gen_data(dtype, shape, negative_slop)
mod = leaky_relu_compile(shape, dtype, negative_slop, attrs)
# result_tvm
acu_output = utils.mod_launch(mod, args, expect=exp_output)
# compare result
rtol, atol = get_rtol_atol("leaky_relu", dtype)
TestCase_Result = compare_tensor(acu_output,
exp_output,
rtol=rtol,
atol=atol,
equal_nan=True)
return inputs, acu_output, exp_output, TestCase_Result
def abs_run(shape, dtype, attrs={}):
# Result_Numpy
input_shape = [shape]
input_dtype = [dtype]
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(Abs,
input_shape,
input_dtype,
kernel_name=kernel_name,
attrs=attrs,
tuning=t)
if t:
exp_output, inputs, output = gen_date(dtype, shape)
return mod, exp_output, (inputs, output)
else:
return mod
else:
mod = utils.op_build_test(Abs,
input_shape,
input_dtype,
kernel_name='abs',
attrs=attrs)
exp_output, inputs, output = gen_date(dtype, shape)
acu_output = utils.mod_launch(mod, (inputs, output), expect=exp_output)
# compare result
rtol, atol = get_rtol_atol("abs", dtype)
TestCase_Result = compare_tensor(acu_output,
exp_output,
rtol=rtol,
atol=atol,
equal_nan=True)
target_name = attrs["target"].split()[0]
if attrs.get("profiling", False):
target_name = attrs["target"].split()[0]
data, output = to_tvm_nd_array([inputs, output],
akg.tvm.context(target_name, 0))
target_profiling(mod,
data,
output,
target=target_name,
repeat_time=attrs["repeat_times"])
return inputs, acu_output, exp_output, TestCase_Result
def fused_bn_grad_5D_run_2(shape, dtype, eps, kernel_name, attrs):
""" test bnGrad_2 """
def get_expect(dgamma_red_hw, dbeta_red_hw, var, gamma, eps, data_shape):
m = data_shape[0] * data_shape[2] * data_shape[3]
neg_m_rec = -1.0 / m
eps = np.array([eps], dtype=var.dtype).reshape([1] * 5)
neg_m_rec = np.array([neg_m_rec], dtype=var.dtype).reshape([1] * 5)
s = (1.0 / np.sqrt(var + eps)).astype(var.dtype)
dgamma = s * np.sum(dgamma_red_hw, axis=0, keepdims=True)
dbeta = np.sum(dbeta_red_hw, axis=0, keepdims=True)
rs = gamma * s
dgamma_dx = neg_m_rec * rs * s * dgamma
dbeta_dx = neg_m_rec * rs * dbeta
return [dgamma, dbeta, rs, dgamma_dx, dbeta_dx]
shape_nc1c0 = (shape[0], shape[1], 1, 1, shape[4])
shape_c1c0 = (1, shape[1], 1, 1, shape[4])
bng2_shapes = [shape_nc1c0, shape_nc1c0, shape_c1c0, shape_c1c0]
bng2_dtypes = ["float32"] * len(bng2_shapes)
bng2_opattrs = [eps, shape]
# np.random.seed(0)
inputs = [np.random.rand(*s).astype(t) for (s, t) in zip(bng2_shapes, bng2_dtypes)]
out_shapes = [shape_c1c0, shape_c1c0, shape_c1c0, shape_c1c0, shape_c1c0]
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(fused_bn_grad2,
bng2_shapes, bng2_dtypes, bng2_opattrs,
kernel_name=kernel_name + "_step2", attrs=attrs, tuning=t)
if t:
outputs = [np.full(s, np.nan, "float32") for s in out_shapes]
expects = get_expect(*inputs, *bng2_opattrs)
return mod, expects, {"args": (*inputs, *outputs), 'outputs': tuple(range(-len(outputs), 0)),
'tuning': False}
else:
return mod
mod = utils.op_build_test(fused_bn_grad2,
bng2_shapes, bng2_dtypes, bng2_opattrs,
kernel_name=kernel_name + "_step2", attrs=attrs)
outputs = [np.full(s, np.nan, "float32") for s in out_shapes]
outputs = list(utils.mod_launch(mod, (*inputs, *outputs), outputs=tuple(range(-len(outputs), 0)),
expect=get_expect(*inputs, *bng2_opattrs)))
expects = get_expect(*inputs, *bng2_opattrs)
rtol, atol = get_rtol_atol("fused_batch_norm_grad", dtype)
results = list(map(lambda x, y: np.allclose(x, y, rtol=rtol, atol=atol), outputs, expects))
print("results", results)
return inputs, outputs, expects, all(results)
def broadcast_to_run(x_shape, x_dtype, shape, attrs):
shapes = [x_shape]
dtypes = [x_dtype]
op_attrs = [shape]
op_name = "broadcast_to"
mod = utils.op_build_test(broadcast_to,
shapes,
dtypes,
op_attrs=op_attrs,
kernel_name=op_name,
attrs=attrs)
bench_mark, inputs, output = gen_data(dtypes, shapes, shape)
output = utils.mod_launch(mod, inputs + [output], expect=bench_mark)
rtol, atol = get_rtol_atol(op_name, x_dtype)
compare_res = compare_tensor(output, bench_mark, rtol=rtol, atol=atol)
return inputs, output, bench_mark, compare_res
def xlogy_grad_run(shape1, shape2, dtype, attrs):
_, _, grad_shape = produce_shapes(shape1, shape2)
mod = utils.op_build_test(xlogy_grad.xlogy_grad,
[shape1, shape2, grad_shape],
[dtype, dtype, dtype],
kernel_name="xlogy_grad", attrs=attrs)
expects, inputs, outputs = gen_data(shape1, shape2, dtype)
reses = utils.mod_launch(
mod, (*inputs, *outputs), expect=expects,
outputs=(-2, -1))
rtol, atol = get_rtol_atol("xlogy_grad", dtype)
TestCase_Results = list(map(lambda x, y: compare_tensor(
x, y, rtol=rtol, atol=atol, equal_nan=True), reses, expects))
return inputs, reses, expects, all(TestCase_Results)
def unsorted_segment_sum_run_others(data_shape,
data_type,
indices_shape,
indices_type,
num,
attrs=None):
mod = unsortedsegmentsum_compile(data_shape,
indices_shape,
num,
data_type,
attrs,
kernel_name='unsortedsegmentsum_run',
tuning=False)
# gen data
input1, input2, expect = gen_data(data_shape, data_type, indices_shape,
indices_type, num)
output_shape = expect.shape
if len(expect.shape) == 0:
output_shape = (1, )
#output = np.full(output_shape, np.nan, expect.dtype)
output = np.zeros(output_shape, expect.dtype)
output = utils.mod_launch(mod, (input1, input2, output), expect=expect)
atol, rtol = get_rtol_atol("unsorted_segment_sum", data_type)
res = compare_tensor(output, expect, rtol=rtol, atol=atol)
print("Test {}".format("Pass" if res else "Failed"))
target_name = attrs["target"].split()[0]
if not res:
mod_source = mod
if target_name != "llvm":
mod_source = mod.imported_modules[0]
print("Error {}:========================".format(target_name))
print(mod_source.get_source())
raise AssertionError("Test fail")
if attrs["profiling"]:
input1, input2, output = to_tvm_nd_array([input1, input2, output],
akg.tvm.context(
target_name, 0))
target_profiling(mod,
input1,
input2,
output,
target=target_name,
repeat_time=attrs["repeat_times"])
return (input1, input2), output, expect, res
def five2four_execute(shape4d, out_dtype, format, dtype, attrs):
# Generate data
op_attrs = [shape4d, out_dtype, format]
if attrs is None:
attrs = {}
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
input, bench_mark = gen_data(shape4d, dtype, out_dtype, format)
shape_5d = input.shape
mod = five2four_compile(shape_5d,
dtype,
op_attrs,
attrs,
kernel_name=kernel_name,
tuning=t)
if t:
output = np.full(shape4d, np.nan, out_dtype)
return mod, bench_mark, (input, output)
else:
return mod
else:
input, bench_mark = gen_data(shape4d, dtype, out_dtype, format)
# mod launch
shape_5d = input.shape
mod = five2four_compile(shape_5d, dtype, op_attrs, attrs)
output = np.full(shape4d, np.nan, out_dtype)
args = [input, output]
# if attrs.get("dynamic"):
# for i in range(len(shape4d) - 1, -1, -1):
# args.append(shape4d[i])
if attrs.get("dynamic"):
args.append(shape_5d[0])
args.append(shape_5d[1])
args.append(shape_5d[4])
block_dim = compute_blockdim(shape4d)
args.append(block_dim)
output = utils.mod_launch(mod, args, outputs=(1, ), expect=bench_mark)
# compare result
rtol, atol = get_rtol_atol("five2four", dtype)
compare_result = compare_tensor(output,
bench_mark,
rtol=rtol,
atol=atol,
equal_nan=True)
return input, output, bench_mark, compare_result
def reduce_max_run(shape,
dtype,
axis,
keepdims,
kernel_name="reduce_max",
attrs=None):
"""run function for dsl function reduce_max"""
if attrs is None:
attrs = {}
op_attrs = [axis, keepdims]
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(reduce_max, [shape], [dtype],
op_attrs=op_attrs,
kernel_name=kernel_name,
attrs=attrs,
tuning=t)
if t:
expect, inputs, output = gen_data(axis, dtype, keepdims, shape)
return mod, expect, (inputs, output)
return mod
mod = utils.op_build_test(reduce_max, [shape], [dtype],
op_attrs=op_attrs,
kernel_name=kernel_name,
attrs=attrs)
expect, inputs, output = gen_data(axis, dtype, keepdims, shape)
output = utils.mod_launch(mod, (inputs, output), expect=expect)
rtol, atol = get_rtol_atol("reduce_max", dtype)
if attrs.get("profiling", False):
import akg
target_name = attrs["target"].split()[0]
args_list = to_tvm_nd_array([inputs, output],
akg.tvm.context(target_name, 0))
target_profiling(mod,
*args_list,
target=target_name,
repeat_time=attrs["repeat_times"])
return inputs, output, expect, compare_tensor(output,
expect,
rtol=rtol,
atol=atol,
equal_nan=True)
def gather_run(shape1,
dtype1,
shape2,
dtype2,
axis,
poly_sch=True,
attrs=None):
if not attrs:
attrs = {"target": "cuda"}
op_attrs = [axis]
mod = utils.op_build_test(gather, [shape1, shape2], [dtype1, dtype2],
op_attrs=op_attrs,
polyhedral=poly_sch,
attrs=attrs,
kernel_name="gather")
# gen data
params, indices, expect = gen_data(shape1, dtype1, shape2, dtype2, axis)
output_shape = expect.shape
if len(expect.shape) == 0:
output_shape = (1, )
output = np.zeros(output_shape, expect.dtype)
output = utils.mod_launch(mod, (params, indices, output), expect=expect)
atol, rtol = get_rtol_atol("gather", dtype1)
res = compare_tensor(output, expect, rtol=rtol, atol=atol)
print("Test {}".format("Pass" if res else "Failed"))
target_name = attrs["target"].split()[0]
if not res:
mod_source = mod
if target_name != "llvm":
mod_source = mod.imported_modules[0]
print("Error {}:========================".format(target_name))
print(mod_source.get_source())
raise AssertionError("Test fail")
if attrs["profiling"]:
params, indices, output = to_tvm_nd_array([params, indices, output],
akg.tvm.context(
target_name, 0))
target_profiling(mod,
params,
indices,
output,
target=target_name,
repeat_time=attrs["repeat_times"])
return (params, indices), output, expect, res
def fake_quant_with_min_max_vars_per_channel_gradient_run(
shape_gradient,
shape_input,
shape_min,
shape_max,
dtype,
num_bits=8,
narror_range=False,
attrs=None):
"""fake_quant_with_min_max_vars_per_channel_gradient_run"""
mod = utils.op_build_test(
fake_quant_with_min_max_vars_per_channel_gradient.
fake_quant_with_min_max_vars_per_channel_gradient,
[shape_gradient, shape_input, shape_min, shape_max],
[dtype, dtype, dtype, dtype], [num_bits, narror_range],
kernel_name='fake_quant_with_min_max_vars_per_channel_gradient',
attrs=attrs)
args, exp_output, input_gradient, input_data, input_min, input_max = gen_data(
shape_gradient, shape_input, shape_min, shape_max, dtype, num_bits,
narror_range)
acu_output = utils.mod_launch(mod,
args,
expect=exp_output,
outputs=(-3, -2, -1))
# compare result
rtol, atol = get_rtol_atol(
"fake_quant_with_min_max_vars_per_channel_gradient", dtype)
testcase_result_0 = compare_tensor(acu_output[0],
exp_output[0],
rtol=rtol,
atol=atol,
equal_nan=True)
testcase_result_1 = compare_tensor(acu_output[1],
exp_output[1],
rtol=rtol,
atol=atol,
equal_nan=True)
testcase_result_2 = compare_tensor(acu_output[2],
exp_output[2],
rtol=rtol,
atol=atol,
equal_nan=True)
testcase_result = list(
map(lambda x, y: compare_tensor(x, y, rtol=rtol, atol=atol),
acu_output, exp_output))
return [input_gradient, input_data, input_min,
input_max], acu_output, exp_output, all(testcase_result)
def square_execute(shape, dtype, kernel_name, attrs):
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = square_compile(shape, dtype, kernel_name, attrs, tuning=t)
if t:
args, exp_output, input = method_name(dtype, shape)
return mod, exp_output, args
else:
return mod
else:
mod = square_compile(shape, dtype, kernel_name, attrs)
args, exp_output, input = method_name(dtype, shape)
acu_output = utils.mod_launch(mod, args, expect=exp_output)
rtol, atol = get_rtol_atol("square", dtype)
testcase_result = compare_tensor(acu_output, exp_output, rtol=rtol, atol=atol, equal_nan=True)
return input, acu_output, exp_output, testcase_result
def selu_run(shape, dtype, attrs):
"""selu_run implementation"""
mod = utils.op_build_test(selu.selu, [shape], [dtype],
kernel_name='selu',
op_attrs=[],
attrs=attrs)
args, exp_output, input_data = gen_data(dtype, shape)
acu_output = utils.mod_launch(mod, args, expect=exp_output)
# compare result
rtol, atol = get_rtol_atol("selu", dtype)
testcase_result = compare_tensor(acu_output,
exp_output,
rtol=rtol,
atol=atol,
equal_nan=True)
return input_data, acu_output, exp_output, testcase_result
def unpack_run(shape, dtype, tensor_format, num, axis, attrs):
"""run function for unpack"""
mod = utils.op_build_test(unpack.unpack, [shape], [dtype],
op_attrs=[tensor_format, num, axis],
kernel_name="unpack",
attrs=attrs)
data, expects, out_bufs = gen_data(shape, dtype, axis)
outputs = utils.mod_launch(mod, (data, *out_bufs),
expect=expects,
outputs=list(range(-len(out_bufs), 0)))
rtol, atol = get_rtol_atol("unpack", dtype)
cmp_res = list(
map(lambda x, y: compare_tensor(x, y, rtol=rtol, atol=atol), outputs,
expects))
return data, outputs, expects, all(cmp_res)
def bn_2_run(shape, dtype, momentum, eps, kernel_name, attrs):
"""Test run function for second part of splited bn"""
in_shapes, in_dtypes = get_compile_param(shape, dtype, 2)
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(fused_bn2,
in_shapes, in_dtypes,
op_attrs=[momentum],
kernel_name=kernel_name,
attrs=attrs, tuning=t)
if t:
inputs, output_buffers, expects = gen_data(shape, dtype, momentum, eps, 2)
inplace_binds = ((2, 1), (3, 2))
output_places2 = list(range(-len(output_buffers), 0))
if inplace_binds is not None:
for bind in inplace_binds:
output_places2[bind[1]] = bind[0]
return mod, expects, {
"args": (*inputs, *output_buffers),
'outputs': output_places2,
'tuning': False}
return mod
mod_2 = utils.op_build_test(fused_bn2,
in_shapes, in_dtypes,
op_attrs=[momentum],
kernel_name="fusedbn2_"+kernel_name,
attrs=attrs)
inputs, output_buffers, expects = gen_data(shape, dtype, momentum, eps, 2)
inplace_binds = ((2, 1), (3, 2))
output_places2 = list(range(-len(output_buffers), 0))
if inplace_binds is not None:
for bind in inplace_binds:
output_places2[bind[1]] = bind[0]
res_2 = utils.mod_launch(mod_2, [*inputs, *output_buffers],
outputs=output_places2, expect=expects)
rtol, atol = get_rtol_atol("bn_split", dtype)
cmp_res = list(map(lambda x, y:
compare_tensor(x, y, rtol=rtol, atol=atol),
res_2, expects))
return inputs, res_2, expects, all(cmp_res)