def test_fused_mul_div_rsqrt_mul_isfinite_red(shape,
dtype='float32',
poly_sch=False):
input = gen_data(shape, dtype)
expect = compute_expect(input)
input_shape = [shape, shape]
input_dtype = [dtype, dtype]
if poly_sch:
mod = utils.op_build_test(
fused_mul_div_rsqrt_mul_isfinite_red,
input_shape,
input_dtype,
kernel_name="fused_mul_div_rsqrt_mul_isfinite_red",
attrs={"target": "cuda"})
outputs = [np.full(
(1, ), False, 'bool')] + [np.full(shape, np.nan, dtype)] * 3
output = utils.mod_launch(mod, [*input, *outputs],
outputs=list(range(-len(outputs), 0)),
expect=expect)
ret = compare_tensor(output[0], expect[0], rtol=5e-03, atol=1.e-08)
ret &= compare_tensor(output[1], expect[1], rtol=5e-03, atol=1.e-08)
ret &= compare_tensor(output[2], expect[2], rtol=5e-03, atol=1.e-08)
ret &= compare_tensor(output[3], expect[3], rtol=5e-03, atol=1.e-08)
print("Test {}".format("Pass" if ret else "Failed"))
if not ret:
print("Error cuda:========================")
print(mod.imported_modules[0].get_source())
raise AssertionError("Test fail")
data = to_tvm_nd_array(input)
expect = to_tvm_nd_array(expect)
gpu_profiling(mod, *data, *expect, 400)
def globalavgpool_run(n, c, h, w, pool_type, dtype, attrs):
dshape = (n, c, h, w)
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(globalavgpool.globalavgpool, [dshape],
[pool_type],
kernel_name=kernel_name,
attrs=attrs,
tuning=t)
if t:
args, exp_output, input = gen_data(c, dshape, dtype, h, n,
pool_type, w)
return mod, exp_output, args
else:
return mod
else:
# Result_globalavgpool
mod = globalavgpool.globalavgpool(n, c, h, w, pool_type, attrs)
args, exp_output, input = gen_data(c, dshape, dtype, h, n, pool_type,
w)
acu_output = utils.mod_launch(mod, args, expect=exp_output)
# compare result
TestCase_Result = compare_tensor(acu_output,
exp_output,
rtol=5e-03,
equal_nan=True)
return input, acu_output, exp_output, TestCase_Result
"""
def test_ms_resize_grad(shape, size, dtype, align_corners, poly_sch=False):
op_attr = [size, align_corners]
if poly_sch:
mod = utils.op_build_test(
resize_nearest_neighbor_grad_auto, [shape], [dtype],
op_attr,
kernel_name="resize_nearest_neighbor_grad_auto",
attrs={"target": "cuda"})
else:
mod = utils.op_build_test(
resize_nearest_neighbor_grad_manual, [shape], [dtype],
op_attr,
kernel_name="resize_nearest_neighbor_grad_manual")
data, output, expect = gen_data(shape, size, align_corners, dtype)
output = utils.mod_launch(mod, (data, output), expect=expect)
compare_res = compare_tensor(output, expect, rtol=5e-03, atol=1e-08)
def apply_ftrl_run(shape, dtype, attrs=None):
"""run function for dsl function apply_ftrl."""
scalar_shape = (1, )
var_shape, accum_shape, linear_shape, grad_shape = [shape] * 4
lr_shape, l1_shape, l2_shape, lr_power_shape = [scalar_shape] * 4
shapes = [
var_shape, accum_shape, linear_shape, grad_shape, lr_shape, l1_shape,
l2_shape, lr_power_shape
]
dtypes = [dtype] * 9
mod = utils.op_build_test(apply_ftrl,
shapes,
dtypes,
kernel_name='apply_ftrl',
attrs=attrs)
expects, (var, accum, linear, grad), (lr, l1, l2,
lr_power) = gen_data(dtype, shape)
outputs = utils.mod_launch(
mod, (var, accum, linear, grad, lr, l1, l2, lr_power),
outputs=(0, 1, 2))
rtol, atol = get_rtol_atol("apply_ftrl", dtype)
compare_result = list(
map(lambda x, y: compare_tensor(x, y, rtol=rtol, atol=atol), outputs,
expects))
inputs = (var, accum, linear, grad, lr, l1, l2, lr_power)
return inputs, outputs, expects, all(compare_result)
def test_ms_equal(shapes, dtype, poly_sch=False):
if poly_sch:
mod = utils.op_build_test(equal, shapes, [dtype, dtype], kernel_name="equal", attrs={"target": "cuda"})
inputs1, output1, expect1 = gen_data(shapes, dtype)
output1 = utils.mod_launch(mod, (*inputs1, output1), expect=expect1)
if shapes[0] == shapes[1]:
inputs2 = []
inputs2.append(inputs1[0])
inputs2.append(inputs1[0])
expect2 = np.equal(inputs2[0], inputs2[1])
output2 = np.full(expect2.shape, 0, bool)
output2 = utils.mod_launch(mod, (*inputs2, output2), expect=expect1)
res = np.allclose(output1, expect1, rtol=5e-03, atol=1.e-8) and np.allclose(output2, expect2, rtol=5e-03, atol=1.e-8)
print("Test {}".format("Pass" if res else "Fail"))
if not res:
print("Error cuda:========================")
print(mod.imported_modules[0].get_source())
raise AssertionError("Test fail")
return True
else:
res = np.allclose(output1, expect1, rtol=5e-03, atol=1.e-8)
print("Test {}".format("Pass" if res else "Fail"))
if not res:
print("Error cuda:========================")
print(mod.imported_modules[0].get_source())
raise AssertionError("Test fail")
return True
def logprob_ad_run(shape, dtype, kernel_name="", attrs=None):
expects, head, x, mean, scale, outputs = gen_data(dtype, shape)
mod = utils.op_build_test(
distr_normal_diag_logprob_ad.normal_diag_logprob_ad,
[head.shape, x.shape, mean.shape, scale.shape],
[dtype, dtype, dtype, dtype],
kernel_name=kernel_name,
op_attrs=None,
attrs=None,
log_code=True,
dump_code=True,
polyhedral=True,
)
outputs = utils.mod_launch(mod, [head, x, mean, scale, *outputs],
outputs=tuple(range(-len(outputs), 0)),
expect=expects)
outputs = list(outputs)
result = True
for i in range(len(outputs)):
result &= compare_tensor(outputs[i],
expects[i],
rtol=5e-03,
equal_nan=True)
return (head, x, mean, scale), outputs, expects, 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 softmax_ad_run(shape, dtype, axis, kernel_name, optimized, attrs):
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
if optimized:
mod = utils.op_build_test(softmax_ad.softmax_ad_optimized,
[shape, shape], [dtype, dtype],
kernel_name=kernel_name,
op_attrs=[axis],
attrs=attrs,
tuning=t)
else:
mod = utils.op_build_test(softmax_ad.softmax_ad, [shape, shape],
[dtype, dtype],
kernel_name=kernel_name,
op_attrs=[axis],
attrs=attrs,
tuning=t)
if t:
expect, input_head, inputs, output = gen_data(dtype, shape)
return mod, expect, (input_head, inputs, output)
else:
return mod
else:
expect, input_head, inputs, output = gen_data(dtype, shape)
if optimized:
mod = utils.op_build_test(softmax_ad.softmax_ad_optimized,
[shape, shape], [dtype, dtype],
kernel_name="softmax_ad_optimized",
op_attrs=[axis],
attrs=attrs)
else:
mod = utils.op_build_test(softmax_ad.softmax_ad, [shape, shape],
[dtype, dtype],
kernel_name="softmax_ad",
op_attrs=[axis],
attrs=attrs)
print(mod.imported_modules[0].get_source())
output = utils.mod_launch(mod, [input_head, inputs, output],
expect=expect)
return [input_head,
inputs], output, expect, np.allclose(output,
expect,
rtol=5e-03,
atol=0.1,
equal_nan=True)
def test_fused_relu_grad_bn_update_grad(shape,
out_shape,
dtype="float16",
layout="NHWC",
out_dtype="float32",
poly_sch=False):
shape_list = [out_shape, shape, shape, shape]
dtype_list = [out_dtype, dtype, dtype, dtype]
op_attrs = [layout]
if poly_sch:
mod = utils.op_build_test(
fused_relu_grad_bn_update_grad_auto,
shape_list,
dtype_list,
op_attrs=op_attrs,
kernel_name="fused_relu_grad_bn_update_grad_auto",
attrs={
"target": "cuda",
"register_memory_depth": 3
})
else:
mod = utils.op_build_test(
fused_relu_grad_bn_update_grad_manual,
shape_list,
dtype_list,
kernel_name="fused_relu_grad_bn_update_grad_manual",
op_attrs=op_attrs)
head, data_sum, in_bn, in_active, output, expect = gen_data(
shape, out_shape, dtype, out_dtype, layout)
outputs = [output, output]
inputs = [data_sum, in_bn, head, in_active]
arg_list = inputs + outputs
outputs = utils.mod_launch(mod,
arg_list,
outputs=tuple(range(-len(outputs), 0)),
expect=expect)
res = np.allclose(outputs, expect, rtol=5e-03, atol=1.e-8)
print("Test {}".format("Pass" if res else "Fail"))
if not res:
print("Error cuda:========================")
print(mod.imported_modules[0].get_source())
raise AssertionError("Test fail")
inputs = to_tvm_nd_array(inputs)
expect = to_tvm_nd_array(expect)
gpu_profiling(mod, *inputs, *expect, 400)
def batch_matmul_run(shape1,
shape2,
dtype,
out_dtype="float32",
layout1="NHDT",
layout2="NHDT",
layout_out="NHDT",
shape_bias=None,
add_bias=False,
tensor_core=True,
poly_sch=True,
attrs=None):
op_attrs = [out_dtype, layout1, layout2, layout_out, tensor_core, add_bias]
default_attrs = attrs
if not attrs:
default_attrs = {"target": "cuda"}
if default_attrs["target"] == "cuda" and tensor_core:
default_attrs.update({
"pragma_enable_matmul": True,
"enable_auto_inline": False
})
elif default_attrs["target"] == "llvm":
if "pragma_enable_matmul" not in default_attrs.keys():
default_attrs["pragma_enable_matmul"] = True
if "feature" not in default_attrs.keys():
default_attrs["feature"] = "avx"
mod = utils.op_build_test(BatchMatMul, (shape1, shape2, shape_bias),
(dtype, dtype, out_dtype),
op_attrs=op_attrs,
attrs=default_attrs,
polyhedral=poly_sch,
kernel_name="batch_matmul")
lhs, rhs, bias, output, expect = gen_data(shape1, shape2, dtype, out_dtype,
layout1, layout2, layout_out,
shape_bias, add_bias)
args = (lhs, rhs, bias, output)
output = utils.mod_launch(mod, args, expect=expect)
res = np.allclose(output, expect, rtol=5e-03, atol=1.e-8)
print("Test {}".format("Pass" if res else "Fail"))
target_name = default_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"]:
args = to_tvm_nd_array(args, akg.tvm.context(target_name, 0))
target_profiling(mod,
*args,
target=target_name,
repeat_time=attrs["repeat_times"])
return (lhs, rhs, bias), output, expect, res
def fused_gather_gather_add_mul_max_exp_scatter_add_run(
input1_shape,
input2_shape,
input3_shape,
input4_shape,
data_dtype,
indices_type,
axis,
poly_sch=True,
attrs=None):
op_attrs = [axis]
default_attrs = {"target": "cuda"}
if attrs:
default_attrs.update(attrs)
mod = utils.op_build_test(
fused_gather_gather_add_mul_max_exp_scatter_add,
[input1_shape, input2_shape, input3_shape, input4_shape],
[data_dtype, indices_type, data_dtype, indices_type],
op_attrs=op_attrs,
attrs=default_attrs,
polyhedral=poly_sch,
kernel_name="fused_gather_gather_add_mul_max_exp_scatter_add",
)
# gen data
input1, input2, input3, input4, expect1, expect2 = gen_data(
input1_shape, input2_shape, input3_shape, input4_shape, data_dtype,
indices_type, axis)
output1 = np.zeros(expect1.shape, expect1.dtype)
output2 = deepcopy(input1)
output1, output2 = utils.mod_launch(
mod, (input1, input2, input3, input4, output1, output2),
outputs=(-2, -1))
atol, rtol = get_rtol_atol(
"fused_gather_gather_add_mul_max_exp_scatter_add", data_dtype)
res = compare_tensor(output1, expect1, rtol=rtol, atol=atol)
res &= compare_tensor(output2, expect2, 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"]:
inputs = to_tvm_nd_array(
[input1, input2, input3, input4, output1, output2],
akg.tvm.context(target_name, 0))
target_profiling(mod,
*inputs,
target=target_name,
repeat_time=attrs["repeat_times"])
return (input1, input2, input3, input4), (output1, output2), (expect1,
expect2), res
def assign_run(ref_shape,
val_shape,
dtype,
kernel_name="assign",
attrs_op={},
cce_path="./",
attrs={}):
attrs.update(attrs_op)
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(Assign, [ref_shape, val_shape],
[dtype, dtype],
kernel_name=kernel_name,
attrs=attrs,
tuning=t)
if t:
ref, val, expect = gen_data(dtype, ref_shape, val_shape)
return mod, expect, (ref, val)
else:
return mod
else:
ref, val, expect = gen_data(dtype, ref_shape, val_shape)
mod = utils.op_build_test(Assign, [ref_shape, val_shape],
[dtype, dtype],
kernel_name=kernel_name,
attrs=attrs)
fake_output = np.full(val_shape, np.nan, dtype)
result, _ = utils.mod_launch(mod, (ref, val, fake_output),
outputs=(0, -1),
expect=expect)
if attrs.get("profiling", False):
target_name = attrs["target"].split()[0]
ref, val, output = to_tvm_nd_array([ref, val, fake_output],
akg.tvm.context(target_name, 0))
target_profiling(mod,
ref,
val,
output,
target=target_name,
repeat_time=attrs["repeat_times"])
return (ref, val), result, expect, compare_tensor(result,
expect,
atol=5e-01,
rtol=5e-03,
equal_nan=True)
def test_fused_bn_double_follow_relu(in_shape,
in_dtype='float16',
layout='NHWC',
out_dtype='float16',
poly_sch=False):
if layout != "NHWC" and layout != "NCHW":
raise NotImplementedError('Layout not supported {} '.format(layout))
inter_dtype = 'float32'
inputs, output, expect = gen_data(in_shape, in_dtype, inter_dtype, layout,
out_dtype)
input_shape_list = [i.shape for i in inputs]
input_dtype_list = [inter_dtype] * 4 + [in_dtype
] + [inter_dtype] * 4 + [in_dtype]
op_attrs = [layout, out_dtype]
if poly_sch:
mod = utils.op_build_test(
fused_bn_double_follow_relu_auto,
input_shape_list,
input_dtype_list,
kernel_name="fused_bn_double_follow_relu_auto",
op_attrs=op_attrs,
attrs={"target": "cuda"})
else:
mod = utils.op_build_test(
fused_bn_double_follow_relu_manual,
input_shape_list,
input_dtype_list,
kernel_name="fused_bn_double_follow_relu_manual",
op_attrs=op_attrs)
outputs = [output]
arglist = inputs + outputs
output = utils.mod_launch(mod, arglist, expect=expect)
res = np.allclose(output, expect, rtol=5e-03, atol=1.e-8)
print("Test {}".format("Pass" if res else "Fail"))
if not res:
print("Error cuda:========================")
print(mod.imported_modules[0].get_source())
raise AssertionError("Test fail")
inputs = to_tvm_nd_array(inputs)
expect = to_tvm_nd_array(expect)
gpu_profiling(mod, *inputs, expect, 400)
def ones_like_run(shape, dtype, attrs):
mod = utils.op_build_test(ones_like.ones_like, [shape], [dtype], kernel_name='ones_like', attrs=attrs)
input, expect, output = gen_data(dtype, shape)
output = utils.mod_launch(mod, (input, output), expect=expect)
rtol, atol = get_rtol_atol("ones_like", dtype)
# compare result
compare_res = compare_tensor(output, expect, rtol=rtol, atol=atol)
return input, output, expect, compare_res
def softplus_ad_run(shape, dtype, kernel_name, attrs):
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(softplus_ad, [shape, shape], [dtype, dtype], kernel_name=kernel_name, attrs=attrs, tuning=t)
if t:
expect, head_np, input_np, output = gen_data(dtype, shape)
return mod, expect, (head_np, input_np, output)
else:
return mod
else:
expect, head_np, input_np, output = gen_data(dtype, shape)
mod = utils.op_build_test(softplus_ad, [shape, shape], [dtype, dtype], kernel_name="softplus", attrs=attrs)
output = utils.mod_launch(mod, [head_np, input_np, output], expect=expect)
rtol, atol = get_rtol_atol("softplus", dtype)
return (head_np, input_np), output, expect, compare_tensor(output, expect, rtol=rtol, atol=atol, equal_nan=True)
def test_ms_maximum(shape1, shape2, dtype, poly_sch=False):
if poly_sch:
mod = utils.op_build_test(maximum_auto, (shape1, shape2), (dtype, dtype), kernel_name="maximum_auto", attrs={"target": "cuda"})
else:
mod = utils.op_build_test(maximum_manual, (shape1, shape2), (dtype, dtype), kernel_name="maximum_manual")
lhs, rhs, output, expect = gen_data(shape1, shape2, dtype)
args = (lhs, rhs, output)
output = utils.mod_launch(mod, args, expect=expect)
res = compare_tensor(output, expect, rtol=5e-03, atol=1.e-8)
print("Test {}".format("Pass" if res else "Fail"))
if not res:
print("Error cuda:========================")
print(mod.imported_modules[0].get_source())
raise AssertionError("Test fail")
lhs, rhs, expect = to_tvm_nd_array([lhs, rhs, expect])
gpu_profiling(mod, lhs, rhs, expect, 400)
def test_expand_dims(shape1, axis, dtype, poly_sch=False):
if poly_sch:
mod = utils.op_build_test(expand_dims_auto, [shape1], [dtype], op_attrs=[axis], attrs={"target": "cuda"}, kernel_name="expand_dims_auto")
else:
mod = utils.op_build_test(expand_dims_manual, [shape1], [dtype], op_attrs=[axis], kernel_name="expand_dims_manual")
expect, input1, output = gen_data(axis, dtype, shape1)
args = (input1, output)
output = utils.mod_launch(mod, args, expect=expect)
res = np.allclose(output, expect, rtol=5e-03, atol=1.e-8)
print("Test {}".format("Pass" if res else "Fail"))
if not res:
print("Error cuda:========================")
print(mod.imported_modules[0].get_source())
raise AssertionError("Test fail")
input1, expect = to_tvm_nd_array([input1, expect])
gpu_profiling(mod, input1, expect, 400)
def test_ms_cast(shape, srcType, dstType, poly_sch=False):
if poly_sch:
mod = utils.op_build_test(cast_auto, [shape], [
srcType], [dstType], attrs={"target": "cuda"}, kernel_name="cast_auto")
else:
mod = utils.op_build_test(cast_manual, [shape], [srcType], [dstType], kernel_name="cast_manual")
output, expect, inputs = gen_data(shape, srcType, dstType)
output = utils.mod_launch(mod, (inputs, output), expect=expect)
res = np.allclose(output, expect, rtol=5e-03, atol=1.e-8)
print("Test {}".format("Pass" if res else "Fail"))
if not res:
print("Error cuda:========================")
print(mod.imported_modules[0].get_source())
raise AssertionError("Test fail")
inputs, expect = to_tvm_nd_array([inputs, expect])
gpu_profiling(mod, inputs, expect, 400)
def addn_compile(shape, dtype, n, attrs, kernel_name="addn", tuning=False):
shapes = []
for _ in range(n):
shapes.append(shape)
return utils.op_build_test(Addn, [shapes], [dtype],
kernel_name=kernel_name,
attrs=attrs,
tuning=tuning), shapes
def slice_run(shape, begin, size, dtype, attrs):
op_attr = [begin, size]
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(op_slice, [shape], [dtype], op_attr, kernel_name=kernel_name, attrs=attrs, tuning=t)
if t:
expect, input, output = gen_data(begin, dtype, shape, size)
return mod, expect, (input, output)
else:
return mod
else:
mod = utils.op_build_test(op_slice, [shape], [dtype], op_attr, kernel_name='slice', attrs=attrs)
expect, input, output = gen_data(begin, dtype, shape, size)
output = utils.mod_launch(mod, (input, output), expect=expect) # unified launch
return input, output, expect, compare_tensor(output, expect, rtol=5e-03, equal_nan=True)
def log1p_run(shape, dtype, kernel_name, attrs):
"""run function for dsl function log1p."""
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(log1p.log1p, [shape], [dtype], kernel_name=kernel_name, attrs=attrs, tuning=t)
if t:
expect, inputs, output = gen_data(dtype, shape)
return mod, expect, (inputs, output)
return mod
mod = utils.op_build_test(log1p.log1p, [shape], [dtype], kernel_name=kernel_name, attrs=attrs)
expect, inputs, output = gen_data(dtype, shape)
output = utils.mod_launch(mod, (inputs, output), expect=expect)
return inputs, output, expect, compare_tensor(output, expect, rtol=5e-03, atol=5e-03, equal_nan=True)
def mean_square_run(shape, reduce_axis, keepdims, dtype, attrs):
op_attrs = [reduce_axis, keepdims]
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(mean_square.mean_square, [shape], [dtype], op_attrs, kernel_name=kernel_name, attrs=attrs, tuning=t)
if t:
expect, input1, output = gen_data(dtype, keepdims, reduce_axis, shape)
return mod, expect, (input1, output)
else:
return mod
else:
mod = utils.op_build_test(mean_square.mean_square, [shape], [dtype], op_attrs, attrs=attrs)
expect, input1, output = gen_data(dtype, keepdims, reduce_axis, shape)
output = utils.mod_launch(mod, (input1, output), expect=expect)
return input1, output, expect, compare_tensor(output, expect, rtol=5e-3, atol=5e-3, equal_nan=True)
def avgpool_run(shape, kernel, stride, strategy, dtype, attrs):
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(avgpool.avgpool, [shape], [dtype], op_attrs=[kernel, stride, strategy],
kernel_name=kernel_name, attrs=attrs, tuning=t)
if t:
expect, input, output = gen_data(dtype, kernel, shape, strategy, stride)
return mod, expect, (input, output)
else:
return mod
else:
mod = utils.op_build_test(avgpool.avgpool, [shape], [dtype], op_attrs=[kernel, stride, strategy],
kernel_name='avgpool', attrs=attrs)
expect, input, output = gen_data(dtype, kernel, shape, strategy, stride)
output = utils.mod_launch(mod, [input, output], expect=expect)
return input, output, expect, compare_tensor(output, expect, rtol=5e-03, atol=5e-03, equal_nan=True)
def add_ad_run(ashape,
bshape,
dtype,
kernel_name="add",
scale=1.0,
attrs={},
polyhedral=True):
if type(scale) is not float or not int:
if type(attrs) is not bool:
scale, attrs = 1.0, scale
else:
scale, attrs, polyhedral = 1.0, scale, attrs
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
a = random_gaussian(ashape, miu=1, sigma=0.1).astype(dtype)
b = random_gaussian(bshape, miu=1, sigma=0.1).astype(dtype)
out = np.add(a, b * scale)
mod = utils.op_build_test(add_ad, [out.shape, ashape, bshape],
[dtype, dtype, dtype],
kernel_name=kernel_name,
op_attrs=[scale],
attrs=attrs,
polyhedral=polyhedral,
tuning=t)
if t:
expect, head_np, output = gen_data(dtype, out)
return mod, expect, (head_np, a, b, output)
else:
return mod
else:
a = random_gaussian(ashape, miu=1, sigma=0.1).astype(dtype)
b = random_gaussian(bshape, miu=1, sigma=0.1).astype(dtype)
out = np.add(a, b * scale)
mod = utils.op_build_test(add_ad, [out.shape, ashape, bshape],
[dtype, dtype, dtype],
kernel_name='add_ad',
op_attrs=[scale],
attrs=attrs,
polyhedral=polyhedral)
expect, head_np, output = gen_data(dtype, out)
output = utils.mod_launch(mod, (head_np, a, b, output), expect=expect)
return (head_np, a, b), output, expect, compare_tensor(output,
expect,
atol=0.1)
def logprob_run(shape, dtype, kernelname="", attrs = None):
expect, x, mean, scale, output = gen_data(dtype, shape)
mod = utils.op_build_test(logprob_op, [x.shape, mean.shape, scale.shape],
[dtype, dtype, dtype], kernel_name=kernelname,
op_attrs=[], attrs=None, log_cce=True, dump_cce=True, polyhedral=True)
output = utils.mod_launch(mod, [x, mean, scale, output], expect = expect)
return (x, mean, scale), output, expect, compare_tensor(output, expect, rtol=5e-03, equal_nan=True)
def conv_fusion_run(shape_data,
shape_filter1,
shape_filter2,
stride1,
stride2,
padding1,
padding2,
dilation1,
dilation2,
dtype,
out_dtype="float32",
poly_sch=True,
attrs=None):
if not attrs:
attrs = {"target": "cuda"}
op_attrs = [stride1, stride2, padding1, padding2, dilation1, dilation2]
attrs.update({
"enable_auto_fuse": False,
"shared_memory_tensors": "out input_1 input_2 input_3",
"pragma_disable_loop_fusion": True,
"dim": "3 0 1 1 3 1 1 1 3 2 4 4 3 3 52 52 3 4 64 64"
})
mod = utils.op_build_test(ConvFusion,
(shape_data, shape_filter1, shape_filter2),
(dtype, dtype, dtype),
op_attrs=op_attrs,
attrs=attrs,
polyhedral=poly_sch,
kernel_name="conv_fusion_auto")
data, weight1, weight2, output, expect = fusion_gen_data(
shape_data, shape_filter1, shape_filter2, stride1, stride2, padding1,
padding2, dilation1, dilation2, dtype, out_dtype)
args = (data, weight1, weight2, output)
output = utils.mod_launch(mod, args, expect=expect)
res = np.allclose(output, expect, rtol=5e-3, atol=1.e-8)
print("Test {}".format("Pass"))
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"]:
data, weight1, weight2, output = to_tvm_nd_array(
[data, weight1, weight2, output], akg.tvm.context(target_name, 0))
target_profiling(mod,
data,
weight1,
weight2,
output,
target=target_name,
repeat_time=attrs["repeat_times"])
return (data, weight1, weight2), output, expect, res
def conv_bn1_run(fmap_shape,
filter_shape,
pad,
stride,
dilation,
use_bias=False,
attrs=None):
vc_util.convolution_format_check(fmap_shape, filter_shape, pad, stride,
dilation)
if use_bias:
raise ValueError("do not support bias yet !!!")
conv_dtype = 'float16'
conv_param = {'stride': stride, 'pad': pad, 'dilation': dilation}
stride, pad, dilation = conv_param_prepare(conv_param)
fm_shape, w_shape, out_shape = conv_shape_4d(fmap_shape, filter_shape, pad,
stride, dilation)
IN, IC, IH, IW = fm_shape
WN, WC, WH, WW = w_shape
C0 = 16
input_shape = [(IN, IC // C0, IH, IW, C0),
(WC // C0 * WH * WW, WN // 16, 16, C0)]
mod = utils.op_build_test(conv_bn1.conv_bn1, [input_shape], [conv_dtype],
op_attrs=[
fmap_shape, filter_shape, pad, stride,
dilation, use_bias, attrs
],
kernel_name='conv_bn1',
attrs=attrs)
fmap_data, filter_data, bias_data, conv_expect = \
gen_data(fmap_shape, filter_shape, pad, stride, dilation, use_bias)
axes = (0, 2, 3)
conv_mean = np.mean(conv_expect, axis=axes, keepdims=True)
conv_square = np.power(conv_expect, 2)
conv_var_part = np.mean(conv_square, axis=axes, keepdims=True)
expects = (conv_expect, conv_var_part, conv_mean)
out_datas = [np.full(e.shape, 0, 'float16') for e in expects]
out_datas[1] = out_datas[1].astype(np.float32)
out_datas[2] = out_datas[2].astype(np.float32)
in_data = [fmap_data, filter_data]
args = in_data
for out in out_datas:
args.append(out)
args = tuple(args)
outputs = utils.mod_launch(mod, args, outputs=(-3, -2, -1), expect=expects)
rtol, atol = get_rtol_atol("conv_bn1", conv_dtype)
cmp_res = list(
map(lambda x, y: compare_tensor(x, y, rtol=rtol, atol=atol), outputs,
expects))
return (fmap_data, filter_data, bias_data), outputs, expects, all(cmp_res)
def common_run(shape, dtype, axis, attrs, method):
if attrs is None:
attrs = {}
attrs["enable_algebra_simplify"] = True
if attrs.get("dynamic"):
build_shape = []
for i in range(len(shape)):
build_shape.append(tvm.var("I" + str(i)))
else:
build_shape = shape
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
if method is "min":
mod = utils.op_build_test(argmin.argmin, [build_shape], [dtype], op_attrs=[axis], kernel_name=kernel_name,
attrs=attrs, tuning=t)
elif method is "max":
mod = utils.op_build_test(argmax.argmax, [build_shape], [dtype], op_attrs=[axis], kernel_name=kernel_name,
attrs=attrs, tuning=t)
else:
raise RuntimeError("not support " + method)
if t:
args, exp_output, input = gen_data(axis, dtype, method, shape)
return mod, exp_output, args
else:
return mod
else:
if method is "min":
mod = utils.op_build_test(argmin.argmin, [build_shape], [dtype], op_attrs=[axis], kernel_name="argmin",
attrs=attrs)
elif method is "max":
mod = utils.op_build_test(argmax.argmax, [build_shape], [dtype], op_attrs=[axis], kernel_name="argmax",
attrs=attrs)
else:
raise RuntimeError("not support " + method)
args, exp_output, input = gen_data(axis, dtype, method, shape)
if attrs.get("dynamic"):
for i in range(len(shape)):
args.append(shape[i])
block_dim = compute_blockdim(shape)
args.append(block_dim)
res = utils.mod_launch(mod, args, outputs=(1,), expect=exp_output)
acu_output = res.astype("int32")
rtol, atol = get_rtol_atol("argmax_min_common", dtype)
return input, acu_output, exp_output, compare_tensor(acu_output, exp_output, rtol=rtol, atol=atol, equal_nan=True)
def bounding_box_encode_run(anchor_box_shape, groundtruth_box_shape, anchor_samples_shape, dtype, scale_factors,
epsilon, kernel_name, attrs={}):
# check_shape:
# bachsize
assert (groundtruth_box_shape[0] == anchor_samples_shape[0])
# num_archors
assert (anchor_box_shape[0] == anchor_samples_shape[1])
assert (not scale_factors or len(scale_factors) == COORDINATES_LEN) and (
anchor_box_shape[-1] == COORDINATES_PAD_LEN) and (
groundtruth_box_shape[-1] == COORDINATES_PAD_LEN)
op_attrs = [scale_factors, epsilon]
if 'tuning' in attrs.keys():
t = attrs.get("tuning", False)
kernel_name = attrs.get("kernel_name", False)
mod = utils.op_build_test(bounding_box_encode.bouding_box_encode,
[anchor_box_shape, groundtruth_box_shape, anchor_samples_shape],
[dtype, dtype, "int32"],
op_attrs, kernel_name=kernel_name, attrs=attrs, dump_code=True, tuning=t)
if t:
anchor_box_data, anchor_samples_data, expect, groundtruth_box_data, output_data = gen_data(anchor_box_shape,
anchor_samples_shape,
dtype, epsilon,
groundtruth_box_shape,
scale_factors)
return mod, expect, (anchor_box_data, groundtruth_box_data, anchor_samples_data, output_data)
else:
return mod
else:
mod = utils.op_build_test(bounding_box_encode.bouding_box_encode,
[anchor_box_shape, groundtruth_box_shape, anchor_samples_shape],
[dtype, dtype, "int32"],
op_attrs, kernel_name=kernel_name, attrs=attrs, dump_code=True)
anchor_box_data, anchor_samples_data, expect, groundtruth_box_data, output_data = gen_data(anchor_box_shape,
anchor_samples_shape,
dtype, epsilon,
groundtruth_box_shape,
scale_factors)
output = utils.mod_launch(mod, (anchor_box_data, groundtruth_box_data, anchor_samples_data, output_data),
expect=expect)
# compare result
compare_result = compare_tensor(output, expect, rtol=5e-3, equal_nan=True)
return (anchor_box_data, groundtruth_box_data, anchor_samples_data), output, expect, compare_result
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)