def compute_blockdim(shape, axis, dtype):
# strategy: all the shape except reduce axis can be used for multicore
blockdim_limit = 2 if utils.product_is_mini() else 32
blockdim = 1
if isinstance(shape, int):
shape = [shape]
if not isinstance(axis, list):
axis = list(axis)
for a in axis:
if a < 0:
a += len(shape)
axis = sorted(axis)
red_sh = 1
if isinstance(shape, (list, tuple)):
for i, sh in enumerate(shape):
if not isinstance(sh, int):
raise TypeError(
"Shape to compute blockdim must be a list/tuple of integer"
)
if i in axis:
red_sh *= sh
else:
blockdim = blockdim * sh
else:
raise TypeError(
"Shape to compute blockdim must be a list/tuple of integer")
if red_sh < 32 / get_bytes(dtype):
# when reduce axis is too small, multicore may not always increase performace
blockdim = 1
return min(blockdim_limit, blockdim)
def shape_dtype_max_size_check(shape, dtype):
"""check validation of tensor's shape."""
if shape:
for x in shape:
if not isinstance(x, int):
return
mul = get_bytes(dtype) * int(reduce(lambda x, y: int(x) * int(y), shape))
if mul > MAX_DATA_SIZE:
error_msg = "*".join([str(sh) for sh in shape])
raise RuntimeError("Invalid shape, data is {} bytes ({}), which "
"exceed max data size {} bytes"
.format(mul, error_msg, MAX_DATA_SIZE))
def conv_shape_check(shape):
if (not isinstance(shape, (tuple, list))) or (len(shape) != 4):
raise RuntimeError("conv tensor shape should be 4d list or tuple")
conv_dtype = "float16"
size = get_bytes(conv_dtype)
for i in shape:
if (not isinstance(i, int)) or (i <= 0):
raise RuntimeError("conv tensor shape should be 4d list or "
"tuple of positive integer")
size *= i
if size > MAX_DATA_SIZE:
raise RuntimeError("runtime can not support tensor more than 2G size")
def get_input_pad_shape(shape, dtype):
"""Function for getting input pad shape."""
pad_unit = ft_util.get_bytes(dtype, allow_none=True)
if pad_unit is None:
logging.warning(
"%s is not support in TensorAddPad, the result is not undefined.",
dtype)
return shape
lastdim = int(math.ceil(shape[-1] / pad_unit) * pad_unit)
pad_shape = [*shape[:-1], '{},{}'.format(shape[-1], lastdim)
] if lastdim != shape[-1] else shape
return pad_shape
def compute_blockdim(shape, axis, dtype):
# strategy: all the shape before reduce axis can be used for multicore
blockdim_limit = 2 if utils.product_is_mini() else 32
blockdim = 1
if isinstance(shape, int):
shape = [shape]
if axis < 0:
axis += len(shape)
if isinstance(shape, (list, tuple)):
for i, sh in enumerate(shape):
if not isinstance(sh, int):
raise TypeError("Shape to compute blockdim must be a list/tuple of integer")
if i == axis:
if sh < 32 / get_bytes(dtype):
# when reduce axis is too small, multicore may not always increase performace
blockdim = 1
break
blockdim = blockdim * sh
else:
raise TypeError("Shape to compute blockdim must be a list/tuple of integer")
return min(blockdim_limit, blockdim)
def four2five(data, format_, dst_dtype='float16', need_custom_tiling=True):
"""
Convert 4-dims "data" to 5-dims,the format of "data" is defined in "format_"
Args:
data (tvm.tensor.Tensor): 4-dims tensor of type float16, float32
format_ (str): a str defined the format of "data"
dst_dtype (str): a str defined the type of output, could be float16 or float32
Returns:
5-dims tvm.tensor.Tensor,type is defined by dst_dtype,
which shape is [N, ceil(C / 16), H, W, 16] and attr about tiling args
Raises:
ValueError: If the type of format_ is invalid.
"""
# Check dtype
vc_util.ops_dtype_check(data.dtype, vc_util.DtypeForDavinci.ALL_FLOAT)
# Check shape
shape = get_shape(data)
vc_util.davinci_format_check(shape, format_, dim=4)
# Check format
if format_ not in ['NCHW', 'NHWC']:
raise ValueError(
"{} format is not support, four2five only support NCHW and NHWC format input"
.format(format_))
last_channel = 16
if format_ == "NCHW":
bs, c, h, w = get_shape(data)
else:
bs, h, w, c = get_shape(data)
pad_c = c
if c % last_channel != 0:
pad_c = (c + 15) // last_channel * last_channel
c1 = pad_c // last_channel
c0 = last_channel
is_dynamic = ds.shape_is_dynamic(data)
if not is_dynamic:
attrs = get_attrs()
else:
attrs = get_dynamic_attrs()
# Check size c when casting happens
if data.dtype != dst_dtype and c0 * c1 >= C_LIMIT_FOR_CAST:
raise ValueError(
"When input and output data type is not matched, shape of 'c' axis should not exceed {}, "
"while currently set is {}".format(C_LIMIT_FOR_CAST, c0 * c1))
@script(capture=locals())
def nchw_to_nc1hwc0_step(inputs, bs, c1, h, w, c0):
output = allocate((bs, c1, h, c0, w), inputs.dtype, "local")
for n_i in range(bs):
for c_i in range(c1):
for h_i in range(h):
for w_i in range(w):
for c_i0 in range(c0):
output[n_i, c_i, h_i, c_i0,
w_i] = inputs[n_i,
c_i * last_channel + c_i0,
h_i, w_i]
output1 = allocate((bs, c1, h, w, c0), inputs.dtype, "local")
for n_i in range(bs):
for c_i in range(c1):
for h_i in range(h):
for w_i in range(w):
for c_i0 in range(c0):
output1[n_i, c_i, h_i, w_i,
c_i0] = output[n_i, c_i, h_i, c_i0, w_i]
return output1
@script(capture=locals())
def nchw_to_nc1hwc0(inputs, bs, c1, h, w, c0):
output = allocate((bs, c1, h, w, c0), inputs.dtype, "local")
for n_i in range(bs):
for c_i in range(c1):
for h_i in range(h):
for w_i in range(w):
for c_i0 in range(c0):
output[n_i, c_i, h_i, w_i,
c_i0] = inputs[n_i,
c_i * last_channel + c_i0,
h_i, w_i]
return output
@script(capture=locals())
def nhwc_to_nc1hwc0(inputs, zero, bs, c1, h, w, c0):
output = allocate((bs, c1, h, w, c0), inputs.dtype, "local")
for n_i in range(bs):
for c_i in range(c1):
for h_i in range(h):
for w_i in range(w):
for c_i0 in range(c0):
if c_i * last_channel + c_i0 < c:
output[n_i, c_i, h_i, w_i,
c_i0] = inputs[n_i, h_i, w_i,
c_i * last_channel +
c_i0]
else:
output[n_i, c_i, h_i, w_i, c_i0] = zero
return output
cast_data = data
need_cast = data.dtype == 'float32' and dst_dtype == 'float16'
if c % last_channel != 0 or need_cast:
expansion = int(ct_util.BLOCK_SIZE / get_bytes(data.dtype))
else:
expansion = None
# float32 -> float16, need to cast before transform
if need_cast:
cast_data = akg.lang.cce.cast_to(data, dst_dtype)
zero_ = akg.tvm.const(0.0, cast_data.dtype)
if format_ == "NCHW":
if c % last_channel != 0:
pad_shape = [bs, pad_c, h, w]
if h == 1 and w == 1:
# if h and w both are 1, it is pad last dim case
output_shape = [bs, pad_c // last_channel, h, w, last_channel]
output = akg.tvm.compute(
output_shape,
lambda i, c1, k, l, c0: akg.tvm.expr.Select(
c0 < c - c1 * last_channel, cast_data[
i, c1 * last_channel + c0, k, l],
akg.tvm.const(0, cast_data.dtype)),
name="output")
else:
# if need to pad c dim, separate transpose to two steps
# first is nchw -> nc1hc0w, second is nc1hc0w -> nc1hwc0
pad_data = akg.tvm.compute(
pad_shape,
lambda i, j, k, l: akg.tvm.expr.Select(
j < c, cast_data[i, j, k, l], zero_),
name="pad_data")
output = nchw_to_nc1hwc0_step(pad_data, to_tvm_const(bs),
to_tvm_const(c1),
to_tvm_const(h), to_tvm_const(w),
to_tvm_const(c0))
else:
if not is_dynamic and data.dtype == "float16" and h * w % last_channel == 0 and h * w < 3600:
output_shape = [bs, c1, h, w, c0]
output = akg.tvm.compute(
output_shape,
lambda n, c1, h, w, c0: akg.lang.cce.four2five_nchw(
cast_data[n, c1 * last_channel + c0, h, w]),
name="output")
else:
output = nchw_to_nc1hwc0(cast_data, to_tvm_const(bs),
to_tvm_const(c1), to_tvm_const(h),
to_tvm_const(w), to_tvm_const(c0))
else:
if not is_dynamic and c < last_channel:
rank = 5 # (n, c1, h, w, c0)
pad_before = []
pad_after = []
for _ in range(rank):
pad_before.append(0)
pad_after.append(0)
pad_after[-1] = last_channel - c
# As c < last_channel, c1 is 1
output = akg.tvm.compute(
(bs, c1, h, w, c),
lambda bs_i, _, h_i, w_i, c_i: cast_data[bs_i, h_i, w_i, c_i],
name="output")
output = tvm_pad(output,
pad_before,
pad_after=pad_after,
name='pad_output')
else:
output = nhwc_to_nc1hwc0(cast_data, zero_, to_tvm_const(bs),
to_tvm_const(c1), to_tvm_const(h),
to_tvm_const(w), to_tvm_const(c0))
# float16 -> float32, need to cast after transform
if data.dtype == 'float16' and dst_dtype == 'float32':
output = akg.lang.cce.cast_to(output, dst_dtype)
vc_util.davinci_format_check(output.shape, "NC1HWC0", dim=5)
if not is_dynamic:
dim_info, _ = four2five_set_dim_func(data, format_, dst_dtype)
if dim_info != "":
attrs["dim"] = dim_info
if need_custom_tiling:
attrs["custom_tiling"] = four2five_tiling_strategy(
output, format_, expansion)
elif need_custom_tiling:
attrs["custom_tiling"] = four2five_tiling_strategy_dynamic(
output, format_)
if is_dynamic:
attrs["enable_feature_library_pre_poly"] = True
return output, attrs
def gen_random_shape(shape_dim, slope=0, min_value=None, max_value=None):
"""
Generate a list of random integer with length equals shape_dim within range [min_value, max_value];
Args:
shape_dim : length of output random shape
slope : only represents the tendency of random shape's value, not mathematical slope of random shape;
slope = -1 tend to generate random shape list with largest value at the beginning and smallest value at the end
slope = 0 tend to generate random shape list with nearly average value among list
slope = 1 tend to generate random shape list with smallest value at the beginning and largest value at the end
"""
if shape_dim <= 0:
raise ValueError("Shape dim should be positive.")
def _build_limit(limit, default):
if limit is None:
limit = default
res = list()
nonlocal shape_dim
if isinstance(limit, (tuple, list)):
if len(limit) != shape_dim:
raise ValueError(
"Min/Max value should have same length with shape_dim")
res = limit
elif isinstance(limit, int):
res = [limit] * shape_dim
else:
raise TypeError(
"Min/Max value should be int or list of int with same length of shape_dim"
)
return res
device_limit = MAX_DATA_SIZE // get_bytes("float32")
if max_value is None and shape_dim > 1:
limit_avg = int(math.pow(device_limit, 1 / shape_dim))
if slope == 0:
max_value = [limit_avg] * shape_dim
else:
ratio = np.arange(-1 / 2, 1 / 2 + 1 / shape_dim, 1 / shape_dim)
if len(ratio) > shape_dim:
new_ratio = list()
for i, r in enumerate(ratio):
if i == len(ratio) // 2 - 1:
new_ratio.append(0)
elif i == len(ratio) // 2:
continue
else:
new_ratio.append(r)
ratio = new_ratio
if slope == -1:
ratio.reverse()
max_value = list()
for i, r in enumerate(ratio):
max_value.append(int((1 + ratio[i]) * limit_avg))
shape_min = _build_limit(min_value, 1)
shape_extent = _build_limit(max_value, device_limit)
random_shape = list()
for mn, mx in zip(shape_min, shape_extent):
random_shape.append(random.randint(mn, mx))
return random_shape
def five2four(data, shape4d, dst_type, format_):
"""
Convert 5-dims "data" to 4-dims,the format of "data" is defined in "format_"
Args:
data (tvm.tensor.Tensor): 5-dims tensor of type float16, float32
shape4d (Union[list, tuple]): a list has 4 nums, shape of output Tensor
dst_type (str): data type of output Tensor
format_ (str): a str defined the format of returns, support NCHW and NHWC
Returns:
4-dims tvm.tensor.Tensor.
"""
vc_util.ops_dtype_check([data.dtype, dst_type],
vc_util.DtypeForDavinci.ALL_FLOAT)
shape5d = get_shape(data)
if not shape_is_dynamic(data):
if len(shape5d) != 5 or shape5d[-1] != 16:
raise ValueError(
"five2four_cce only support 5-dim data and last dim should be 16"
)
bs, c1, h, w, c0 = shape5d
if not shape_is_dynamic(data):
vc_util.davinci_format_check(shape5d, "NC1HWC0", dim=5)
# Check format
if format_ not in ['NCHW', 'NHWC']:
raise ValueError(
"{} format is not support, five2four only support NCHW and NHWC format input"
.format(format_))
if format_ == "NCHW":
if shape_is_dynamic(data):
shape4d = [bs, c1 * c0, h, w]
_, c, h_4d, w_4d = shape4d
else:
if shape_is_dynamic(data):
shape4d = [bs, h, w, c1 * c0]
_, h_4d, w_4d, c = shape4d
vc_util.davinci_format_check(shape4d, format_, dim=4)
# Check is shape4d and shape5d match
if False not in [
isinstance(s, (int, akg.tvm.expr.IntImm)) for s in shape5d
]:
if h_4d != h or w_4d != w:
raise ValueError(
"five2four_cce's shape4d h and w should equal to data shape's h and w"
)
if c > c1 * c0 or c <= (c1 - 1) * c0:
raise ValueError(
"five2four_cce's shape4d c should in set ((c1 - 1) * c0, c1 * c0]"
)
# Check size c when casting happens
if not shape_is_dynamic(data):
if data.dtype != dst_type and c >= C_LIMIT_FOR_CAST:
raise ValueError(
"When input and output data type is not matched, shape of 'c' axis should not exceed {}, "
"while currently set is {}".format(C_LIMIT_FOR_CAST, c))
@script(capture=locals())
def nc1hwc0_to_nhwc(inputs, bs, h, w, c, c1, c0):
output = allocate((bs, h, w, c), inputs.dtype, "local")
for n_i in range(bs):
for h_i in range(h):
for w_i in range(w):
for c_i in range(c1):
for c_i0 in range(c0):
output[n_i, h_i, w_i,
c_i * c0 + c_i0] = inputs[n_i, c_i, h_i,
w_i, c_i0]
return output
@script(capture=locals())
def nc1hwc0_to_nchw(inputs, bs, h, w, c, c1, c0):
output = allocate((bs, c, h, w), inputs.dtype, "local")
for n_i in range(bs):
for c_i in range(c1):
for h_i in range(h):
for w_i in range(w):
for c_i0 in range(c0):
output[n_i, c_i * c0 + c_i0, h_i,
w_i] = inputs[n_i, c_i, h_i, w_i, c_i0]
return output
# if c % 16 == 0, h and w == 1, five2four is a reshape operation
if shape_is_dynamic(data):
call_reshape = isinstance(h, int) and isinstance(
w, int) and h == 1 and w == 1
else:
call_reshape = h == 1 and w == 1 and c % 16 == 0
c_value = None
expansion = None
if format_ == "NHWC":
if call_reshape:
output = akg.topi.reshape(data, (bs, h, w, c))
if shape_is_dynamic(data):
output = akg.tvm.compute((bs, h, w, c),
lambda *indice: output(*indice),
name="reshape")
elif c < c0:
reshape_output = akg.topi.reshape(data, (bs, h, w, c0))
output = akg.tvm.compute((bs, h, w, c),
lambda *i: reshape_output(*i),
name='slice_output')
else:
output = nc1hwc0_to_nhwc(data, to_tvm_const(bs), to_tvm_const(h),
to_tvm_const(w), to_tvm_const(c),
to_tvm_const(c1), to_tvm_const(c0))
else:
if call_reshape:
output = akg.topi.reshape(data, (bs, c, h, w))
if shape_is_dynamic(data):
output = akg.tvm.compute((bs, c, h, w),
lambda *indice: output(*indice),
name="reshape")
else:
output = nc1hwc0_to_nchw(data, to_tvm_const(bs), to_tvm_const(h),
to_tvm_const(w), to_tvm_const(c),
to_tvm_const(c1), to_tvm_const(c0))
# two special cases for tiling strategy
if not shape_is_dynamic(data):
if c < c0 or output.dtype != dst_type:
c_value = c
if c % c0 != 0 and output.dtype != dst_type:
expansion = int(ct_util.BLOCK_SIZE / get_bytes(data.dtype))
attrs = get_attrs()
if not call_reshape:
attrs["custom_tiling"] = five2four_tiling_strategy(
data, c_value, expansion)
if output.dtype != dst_type:
output = akg.topi.cast(output, dst_type)
return output, attrs