def fill_v2_d(y, value, shape, kernel_name="fill_v2_d"):
"""
interface of fill_v2_d
:param y: output
:param value: value to fill the shape, float32
:param shape: list int, output shape
:param kernel_name: fill_v2_d
:return:
"""
# check kernel name
util.check_kernel_name(kernel_name)
# shape to list
shape = te.lang.cce.util.shape_to_list(shape)
util.check_shape_rule(shape)
# pseudo input, won't be used.
data_x = tvm.placeholder(shape, dtype="float32", name="data_x")
# do compute
res = fill_v2_compute(data_x, value, shape, y, kernel_name)
# new schedule
schedule = [tvm.create_schedule(res.op)]
elewise_sch = te.lang.cce.te_schedule.cce_schedule.ElewiseSchedule()
elewise_sch._get_emit_insn_map = types.MethodType(_get_emit_insn_map, elewise_sch)
elewise_sch._do_buffer_tile = types.MethodType(_do_buffer_tile, elewise_sch)
elewise_sch.do_schedule([res], schedule, [])
schedule = schedule[0]
schedule.cce_special = {"tensor_list": (), "orign_out_tensor": [res], "real_out_tensor": [res]}
# build operater
config = {"name": kernel_name,
"tensor_list": (data_x, res)}
te.lang.cce.cce_build_code(schedule, config)
def _unpack_schedule(self,
block_tiling_axis,
right_dim_in,
ub_tiling_axis,
split_factor):
"""
unpack schedule function
Parameters
----------
block_tiling_axis: int
identify spilt axis for multicore
right_dim_in: tvm.var
the var identify right_dim of output_shape
ub_tiling_axis: int
identify spilt axis for ub_tiling
split_factor: tvm.var
the var identify spilt_factor
Returns
---------
sch: tvm.schedule
the compute schedule
build_list: list
include tvm.tensor of input and tvm.tensor of res
"""
build_list = [self._input_placeholder]
for res_tensor in self.res_tensor_list:
build_list.append(res_tensor)
sch = tvm.create_schedule(self.virtual_node.op)
sch.disable_allocate(cce_params.scope_ubuf)
for tensor in self.ub_tensor_list:
sch[tensor].set_scope(cce_params.scope_ubuf)
right_dim_outer, right_dim_inner = sch[self.virtual_node].split(
self.virtual_node.op.axis[block_tiling_axis], factor=right_dim_in)
sch[self.virtual_node].bind(right_dim_outer, self.block_idx)
if ub_tiling_axis == 0:
axis_outer, axis_inner = sch[self.virtual_node].split(
self.virtual_node.op.axis[0], factor=1)
else:
axis_outer, axis_inner = sch[self.virtual_node].split(
right_dim_inner, factor=split_factor)
for i in range(self.num):
sch[self.ub_tensor_list[i]].compute_at(
sch[self.virtual_node], axis_outer)
sch[self.res_tensor_list[i]].compute_at(
sch[self.virtual_node], axis_outer)
sch[self.ub_tensor_list[i]].emit_insn(
self.ub_tensor_list[i].op.axis[ub_tiling_axis],
insn_cmd.DMA_COPY)
sch[self.res_tensor_list[i]].emit_insn(
self.res_tensor_list[i].op.axis[ub_tiling_axis],
insn_cmd.DMA_COPY)
sch[self.virtual_node].emit_insn(axis_inner, insn_cmd.PHONY_INSN)
return sch, build_list
def load_to_l1(input_x, output_x, kernel_name="load_to_l1"):
"""
copy data from ddr to l1
Parameters
----------
input_x : TVM tensor
the input tensor
output_x : dict
dict of output_x, include keys(shape and dtype)
kernel_name : str
kernel name, default value is "load_to_l1"
Returns
-------
None
"""
input_shape = input_x.get("shape")
input_dtype = input_x.get("dtype")
input_tensor = tvm.placeholder(input_shape,
name="input_tensor",
dtype=input_dtype)
res = load_to_l1_compute(input_tensor, output_x, kernel_name=kernel_name)
sch = tvm.create_schedule([res.op])
sch[res].set_scope(cce.scope_cbuf_fusion)
sch[res].emit_insn(res.op.axis[0], 'dma_copy')
tensor_list = [input_tensor, res]
with build_config:
tvm.build(sch, tensor_list, "cce", name=kernel_name)
def custom_equal(shape_x, shape_y, dtype, kernel_name="cce_tf_equal", need_build=False,
need_print=False):
"""
do element-wise equal operation between two input tensors
Parameters:
----------
shape_x : shape of input x
shape_y : shape of input y
dtype : source data type, support float16,float32,int32,int8,uint8
kernel_name : cce kernel name, default value is "cce_tf_equal"
need_buid : if need to build CCEC kernel, default value is False
need_print : if need to print the ir, default value is False
Returns
-------
None
"""
util.check_kernel_name(kernel_name)
util.check_shape_rule(shape_x)
util.check_shape_rule(shape_y)
check_list = ["float16", "float32", "int32", "int8", "uint8", "bool"]
dtype = dtype.lower()
if not (dtype in check_list):
raise RuntimeError(
"tf_equal_cce only support %s while dtype is %s" % (",".join(check_list), dtype))
util.check_shape_size(shape_x, SHAPE_SIZE_LIMIT)
util.check_shape_size(shape_y, SHAPE_SIZE_LIMIT)
shape_x, shape_y, shape_max = util.produce_shapes(shape_x, shape_y)
util.check_shape_size(shape_max, SHAPE_SIZE_LIMIT)
x = tvm.placeholder(shape_x, dtype=dtype, name="x")
y = tvm.placeholder(shape_y, dtype=dtype, name="y")
x_tmp = te.lang.cce.broadcast(x, shape_max)
y_tmp = te.lang.cce.broadcast(y, shape_max)
res = tvm.compute(shape_max, lambda *i: x_tmp(*i) == y_tmp(*i), name='res')
sch = tvm.create_schedule(res.op)
if need_print:
with build_config:
print(tvm.lower(sch, [x, y, res], simple_mode=True))
if need_build:
with build_config:
tvm.build(sch, [x, y, res], "cce", name=kernel_name)
def custom_logical_not(shape,
dtype,
kernel_name="cce_tf_logical_not",
need_build=False,
need_print=False):
"""
logical not for the input tensor
Parameters
----------
shape : input shape of data
dtype : the data type, support bool
kernel_name : cce kernel name, default value is "cce_logical_not"
need_buid : if need to build CCEC kernel, default value is False
need_print : if need to print the ir, default value is False
Returns
-------
None
"""
util.check_kernel_name(kernel_name)
util.check_shape_rule(shape)
check_list = ["bool"]
if not dtype.lower() in check_list:
raise RuntimeError(
"logical_not_cce ony supports %s while dtype is %s" %
(",".join(check_list), dtype))
util.check_shape_size(shape, SHAPE_SIZE_LIMIT)
inp_dtype = dtype.lower()
data = tvm.placeholder(shape, name="data", dtype=inp_dtype)
with tvm.target.cce():
result = tvm.compute(
shape,
lambda *i: tvm.select(data[i] is True, False, True),
name="result")
schedule = tvm.create_schedule(result.op)
if need_print:
with build_config:
print(tvm.lower(schedule, [data, result], simple_mode=True))
if need_build:
with build_config:
tvm.build(schedule, [data, result], "cce", name=kernel_name)
def depthwise_weight_6d_2_4d(x,
y,
src_format,
dst_format,
kernel_name="depthwise_weight_6d_2_4d"):
"""Operation and Schedule for depthwise_weight_6d_2_4d.
Parameters
----------
x: shape and dtype of input, the dtype support float16, float32,
int32, uint16.
y: the shape and dtype of outputs, the dtype same as input.
src_format: the source data_format
dst_format: the target data_format
kernel_name : cce kernel name, default value is "depthwise_weight_6d_2_4d"
Returns
-------
convert C1HWNCoC0 tp HWCN
"""
_check_parameters(x, y, src_format, dst_format)
output_shape = y.get("shape")
channel_size = output_shape[2]
input_shape = x.get("shape")
dtype = x.get("dtype")
channel_4d = channel_size
op_utils.check_shape(input_shape, param_name="x")
check_list = ("float16", "float32", "int32", "uint16")
dtype = dtype.lower()
op_utils.check_dtype(dtype, check_list, param_name="x")
input_data = tvm.placeholder(input_shape, name="input_data", dtype=dtype)
six2four = _Six2FourParam(input_shape, channel_4d)
res = tvm.extern(
[six2four.get_out_shape()], [input_data],
lambda ins, outs: _intrin_factor(six2four, dtype, ins, outs),
name="res",
dtype=dtype)
sch = tvm.create_schedule(res.op)
build_list = [input_data, res]
with build_config:
tvm.build(sch, build_list, "cce", name=kernel_name)
def depthwise_weight_4d_2_6d(x,
y,
src_format,
dst_format,
kernel_name="depthwise_weight_4d_2_6d"):
"""Operation and Schedule for depthwise_weight_4d_2_6d.
Parameters
----------
x: shape and dtype of input, the dtype support float16,
float32, int32, uint16.
y: the shape and dtype of outputs, the dtype same as input.
src_format: the source data_format
dst_format: the target data_format
kernel_name : cce kernel name, default value is "depthwise_weight_4d_2_6d"
Returns
-------
convert HWCN to C1HWNCoC0
"""
if src_format.lower() != "hwcn":
raise RuntimeError("dst_format must be HWCN!")
if dst_format.lower() != "c1hwncoc0":
raise RuntimeError("src_format must be C1HWNCoC0 !")
input_shape = x.get("shape")
dtype = x.get("dtype")
op_utils.check_shape(input_shape, param_name="x")
check_list = ("float16", "float32", "int32", "uint16")
dtype = dtype.lower()
op_utils.check_dtype(dtype, check_list, param_name="x")
input_data = tvm.placeholder(input_shape, name="input_data", dtype=dtype)
four2six = _Four2SixParam(input_shape)
res = tvm.extern(
[four2six.get_out_shape()], [input_data],
lambda ins, outs: _intrin_factor(four2six, dtype, ins, outs),
name="res",
dtype=dtype)
sch = tvm.create_schedule(res.op)
build_list = [input_data, res]
with build_config:
tvm.build(sch, build_list, "cce", name=kernel_name)
def zn_2_hwcn(src, dst, src_format, dst_format, kernel_name='zn_2_hwcn'):
"""
algorithm: zn_2_hwcn
calculating: change data format from Zn to HWCN
Parameters
----------
src: dict
contains shape and dtype information of input tensor
dst: dict
contains shape and dtype information of output tensor
src_format: str
represents the format of input tensor, only support "Zn"
dst_format: str
represents the format of output tensor, only support "HWCN"
kernel_name: str
cce kernel name, default value is "zn_2_hwcn"
Returns
-------
None
"""
_check_parameters(src, dst, src_format, dst_format, kernel_name)
dst_shape = dst.get("shape")
dtype = src.get("dtype")
h_i, w_i, c_i, n_i = dst_shape
c_0 = 16
if dtype == "int8":
c_0 = 32
c_1 = _ceil_div(c_i, c_0)
n_ni = 16
n_no = _ceil_div(n_i, n_ni)
shape_zn = [c_1*h_i*w_i, n_no, n_ni, c_0]
branch = _get_ir_branch(shape_zn, dtype)
data = tvm.placeholder(shape_zn, dtype=dtype, name="data")
if branch == "more_row":
res = tvm.extern(dst_shape, [data],
lambda ins, outs: _more_row_ir(outs[0], ins[0], c_0),
name="res", dtype=dtype)
else:
res = tvm.extern(dst_shape, [data],
lambda ins, outs: _split_row_ir(outs[0], ins[0]),
name="res", dtype=dtype)
tensor_list = [data, res]
sch = tvm.create_schedule(res.op)
with build_config:
tvm.build(sch, tensor_list, "cce", name=kernel_name)
def schedule_tile_cce(out):
"""Schedule for cce tile arithmetic operator.
Parameters
----------
out: TVM Tensor
The computation graph description of cce tile.
Returns
-------
s: Schedule
The computation schedule for tile.
"""
sch = tvm.create_schedule(out.op)
return sch
def upsample(x, y, scale=1.0, stride_h=2, stride_w=2, kernel_name="upsample"):
"""
calculating data
Parameters
---------
x : dict
include shape dtype and format
stride_h : int
the shape change axis h
stride_w : int
the shape change axis w
scale : float
the value of tensor change axis, default value is 1.0
y :output
kernel_name : str
kernel name, default value is "upsample"
Returns
-------
None
"""
upsample_check(x, stride_h, stride_w, kernel_name)
dtype = x.get("dtype")
op_list, ins_list, tensor_dic, feature, y \
= gen_upsample(x, dtype, scale, stride_h, stride_w)
schedule = tvm.create_schedule(y.op)
# skip the res buffer
buffer_mapping(schedule, op_list[:-1])
tilling_spilt_axis_dic \
= tilling_spilt_axis(schedule, tensor_dic, stride_h, stride_w)
cal_axis_dic, axis \
= cal_axis_spilt(x, stride_h, stride_w,
tilling_spilt_axis_dic, tensor_dic, schedule)
axis_list = upsample_compute(schedule, cal_axis_dic, tensor_dic)
res_op = tensor_dic.get("res")
ins_emit(schedule, op_list, axis_list, ins_list)
if axis == 0:
schedule[y].bind(cal_axis_dic.get("axis_xo"),
tvm.thread_axis("blockIdx.x"))
else:
res_out, _ = bind_multcore(axis, x, schedule, res_op)
schedule[y].bind(res_out, tvm.thread_axis("blockIdx.x"))
with build_config:
tvm.build(schedule, [feature, y], "cce", name=kernel_name)
def store_to_gm(input_x, output_x, kernel_name="store_to_gm"):
"""
copy data from l1 to ddr (l1 --> ub --> ddr)
Parameters
----------
input_x : TVM tensor
the input tensor
output_x : dict
dict of output_x, include keys(shape and dtype)
kernel_name : str
kernel name, default value is "store_to_gm"
Returns
-------
None
"""
input_shape = input_x.get("shape")
input_dtype = input_x.get("dtype")
input_tensor = tvm.placeholder(input_shape,
name="input_tensor",
dtype=input_dtype)
res, res_ub = store_to_gm_compute(input_tensor,
output_x,
kernel_name=kernel_name)
sch = tvm.create_schedule([res.op])
split_axis, split_factor = _tilling_axis(input_shape, input_dtype)
axis_outer, axis_inner = sch[res].split(res.op.axis[split_axis],
factor=split_factor)
sch[res_ub].compute_at(sch[res], axis_outer)
sch[input_tensor].set_scope(cce.scope_cbuf_fusion)
sch[res_ub].set_scope(cce.scope_ubuf)
sch[res_ub].emit_insn(res_ub.op.axis[split_axis], 'dma_copy')
sch[res].emit_insn(axis_inner, 'dma_copy')
tensor_list = [input_tensor, res]
with build_config:
tvm.build(sch, tensor_list, "cce", name=kernel_name)
def flatten(x, y, kernel_name="flatten"):
"""return a copy of the tensor collapsed into one dimension.
Parameters
----------
x : dict
shape and dtype of input.
y : dict
shape and dtype of output.
kernel_name : str
kernel name, default value is "flatten"
Returns
-------
None
"""
shape = x.get("shape")
dtype = x.get("dtype")
dtype_lower = dtype.lower()
check_list = ("int8", "int16", "int32", "int64", "uint8", "uint16",
"uint32", "uint64", "float16", "float32")
check_shape(shape, param_name="x")
check_dtype(dtype_lower, check_list, param_name="x")
size = 1
for i, _ in enumerate(shape):
size = size * shape[i]
shape_new = [size]
data = tvm.placeholder(shape_new, name="data", dtype=dtype_lower)
data_ub = tvm.compute(shape_new, lambda *i: data(*i), name='data_ub')
res = tvm.compute(shape_new, lambda *i: data_ub(*i), name='res')
sch = tvm.create_schedule(res.op)
sch[data_ub].set_scope(tbe_platform.scope_ubuf)
sch_new = _tile_axis([sch, data_ub, res], shape_new, dtype_lower)
with build_config:
tvm.build(sch_new, [data, res], "cce", name=kernel_name)
def space_to_batch_d(x,
y,
block_size,
paddings,
kernel_name="space_to_batch_d"):
"""
the main function of space_to_batch_d
Parameters
----------
x: dict,shape and datatype,datatype supports float16,float32
y: dict,shape and datatype,datatype supports float16,float32
block_size: must be greater than one. It indicates the block size
paddings: (tuple, list),the padding of the input with zeros across the
spatial dimensions as follows:
paddings = [[pad_top, pad_bottom], [pad_left, pad_right]]
kernel_name: cce kernel name, default value is "space_to_batch_d"
Returns
-------
None
"""
if len(paddings) == 4:
paddings = [[paddings[0], paddings[1]], [paddings[2], paddings[3]]]
_check_param(x, y, paddings, block_size, kernel_name)
input_shape = x.get("shape")
input_dtype = x.get("dtype").lower()
block_shape = [block_size, block_size]
data = tvm.placeholder(input_shape, name="data", dtype=input_dtype)
res = space_to_batch_nd_d_compute(data, y, block_shape, paddings,
kernel_name)
sch = tvm.create_schedule(res.op)
with build_config:
tvm.build(sch, [data, res], "cce", name=kernel_name)
def avg_pool_grad_schedule(res):
"""
the tiling avg pool grad schedule
"""
s = tvm.create_schedule(res.op)
mad_cast = res.op.input_tensors[0]
mad_res = mad_cast.op.input_tensors[0]
dout_col_pad = mad_res.op.input_tensors[0]
weight_rotated = mad_res.op.input_tensors[1]
weight = weight_rotated.op.input_tensors[0]
dout_col = dout_col_pad.op.input_tensors[0]
dout_dilated = dout_col.op.input_tensors[0]
dout_mul = dout_dilated.op.input_tensors[0]
dout = dout_mul.op.input_tensors[0]
dvealuemean = dout_mul.op.input_tensors[1]
dout_ubuf = s.cache_read(dout, tbe_platform.scope_ubuf, [dout_mul])
dvealuemean_ubuf = s.cache_read(dvealuemean, tbe_platform.scope_ubuf,
[dout_mul])
dout_mul_ubuf = s.cache_write(dout_mul, tbe_platform.scope_ubuf)
dout_cbuf_nc1hwc0 = s.cache_write(dout_dilated, tbe_platform.scope_cbuf)
dout_dilated_ubuf = s.cache_write(dout_cbuf_nc1hwc0,
tbe_platform.scope_ubuf)
dout_cbuf_row_major = s.cache_write(dout_col, tbe_platform.scope_cbuf)
dout_ca = s.cache_write(dout_col_pad, tbe_platform.scope_ca)
s[dout_mul].compute_inline()
s[dout_dilated].compute_inline()
s[dout_col].compute_inline()
s[dout_col_pad].compute_inline()
weight_cbuf = s.cache_read(weight, tbe_platform.scope_cbuf,
[weight_rotated])
weight_cb = s.cache_write(weight_rotated, tbe_platform.scope_cb)
s[weight_rotated].compute_inline()
mad_cc = s.cache_write(mad_res, tbe_platform.scope_cc)
mad_ubuf = s.cache_write(mad_cast, tbe_platform.scope_ubuf)
s[mad_res].compute_inline()
s[mad_cast].compute_inline()
# get shape value
dilated_pad_top = res.op.attrs['dilated_pad'][0].value
dilated_pad_bottom = res.op.attrs['dilated_pad'][1].value
dilated_pad_left = res.op.attrs['dilated_pad'][2].value
dilated_pad_right = res.op.attrs['dilated_pad'][3].value
k_height = res.op.attrs['weight_height'].value
k_width = res.op.attrs['weight_width'].value
block_size = dout.op.shape[len(dout.op.shape) - 1].value
_, _, _, dout_dilated_h, dout_dilated_w, _ = dout_dilated.shape
input_w = dout_dilated_w.value + dilated_pad_left \
+ dilated_pad_right - k_width + 1
input_h = dout_dilated_h.value + dilated_pad_top \
+ dilated_pad_bottom - k_height + 1
stride = dout_dilated.op.attrs["strides"][0].value
weight_shape = [int(i.value) for i in weight.shape]
dout_shape = [int(i.value) for i in dout.shape]
dout_dilated_shape = [int(i.value) for i in dout_dilated.shape]
mad_cc_axis_n, mad_cc_axis_cg, mad_cc_axis_co1, mad_cc_axis_howomad, \
mad_cc_axis_co0 = mad_cc.op.axis
mad_ubuf_axis_n, mad_ubuf_axis_cg, mad_ubuf_axis_co1, \
mad_ubuf_axis_howomad, mad_ubuf_axis_co0 = mad_ubuf.op.axis
mad_res_shape = [int(i.value) for i in mad_res.shape]
res_block_n, res_block_cgroup, _, _, _ = mad_res_shape
#tiling
res_l1, tile_input_h, tile_dile_h_ub, tile_m, \
tile_k, tile_n = avg_pool_grad_tiling(
input_w, input_h, weight_shape, dout_shape, res, stride)
mad_cc_Ncut_o, mad_cc_Ncut_i = s[mad_cc].split(mad_cc_axis_n, factor=1)
mad_cc_mcut_o, mad_cc_mcut_i = s[mad_cc].split(mad_cc_axis_howomad,
factor=tile_m)
mad_cc_kcut_o, mad_cc_kcut_i = s[mad_cc].split(mad_cc.op.reduce_axis[0],
factor=tile_k)
mad_cc_ncut_o, mad_cc_ncut_i = s[mad_cc].split(mad_cc_axis_co1,
factor=tile_n)
s[mad_cc].reorder(mad_cc_Ncut_o, mad_cc_axis_cg, mad_cc_ncut_o,
mad_cc_mcut_o, mad_cc_kcut_o, mad_cc_Ncut_i,
mad_cc_ncut_i, mad_cc_mcut_i, mad_cc_axis_co0,
mad_cc_kcut_i, mad_cc.op.reduce_axis[1])
s[dout_ca].compute_at(s[mad_cc], mad_cc_kcut_o)
s[weight_cb].compute_at(s[mad_cc], mad_cc_kcut_o)
mad_ubuf_Ncut_o, mad_ubuf_Ncut_i = s[mad_ubuf].split(mad_ubuf_axis_n,
factor=1)
mad_ubuf_mcut_o, mad_ubuf_mcut_i = s[mad_ubuf].split(mad_ubuf_axis_howomad,
factor=tile_m)
mad_ubuf_ncut_o, mad_ubuf_ncut_i = s[mad_ubuf].split(mad_ubuf_axis_co1,
factor=tile_n)
s[mad_ubuf].reorder(mad_ubuf_Ncut_o, mad_ubuf_axis_cg, mad_ubuf_ncut_o,
mad_ubuf_mcut_o, mad_ubuf_Ncut_i, mad_ubuf_ncut_i,
mad_ubuf_mcut_i, mad_ubuf_axis_co0)
s[mad_cc].compute_at(s[mad_ubuf], mad_ubuf_mcut_o)
conv_Ncut_o, conv_Ncut_i = s[res].split(res.op.axis[0], factor=1)
conv_hcut_o, conv_hcut_i = s[res].split(res.op.axis[3], factor=(res_l1))
conv_mcut_o, conv_mcut_i = s[res].split(conv_hcut_i, factor=tile_m)
s[res].reorder(conv_Ncut_o, res.op.axis[1], conv_hcut_o, conv_mcut_o,
conv_Ncut_i, res.op.axis[2], conv_mcut_i, res.op.axis[4])
s[mad_ubuf].buffer_align((1, 1), (1, 1), (1, 1), (1, block_size),
(1, block_size))
s[mad_ubuf].compute_at(s[res], conv_mcut_o)
s[dout_cbuf_row_major].buffer_align((1, 1), (1, 1), (input_w, input_w),
(1, 1), (1, 1), (1, 1),
(1, block_size))
s[dout_cbuf_row_major].compute_at(s[res], conv_hcut_o)
s[dout_cbuf_nc1hwc0].compute_at(s[res], conv_hcut_o)
s[weight_cbuf].compute_at(s[res], conv_hcut_o)
dout_dilated_w = dout_dilated_shape[4]
ub_l1hcut_o, ub_l1hcut_i = s[dout_cbuf_nc1hwc0].split(
dout_cbuf_nc1hwc0.op.axis[3], factor=tile_dile_h_ub)
if stride > 1:
dila_o_h, dila_i_h = s[dout_dilated_ubuf].split(
dout_dilated_ubuf.op.axis[3], factor=stride)
dila_o_w, dila_i_w = s[dout_dilated_ubuf].split(
dout_dilated_ubuf.op.axis[4], factor=stride)
s[dout_dilated_ubuf].reorder(dila_i_h, dila_i_w, dila_o_h, dila_o_w)
s[dout_dilated_ubuf].unroll(dila_i_h)
s[dout_dilated_ubuf].unroll(dila_i_w)
s[dout_dilated_ubuf].compute_at(s[dout_cbuf_nc1hwc0], ub_l1hcut_o)
s[dout_dilated_ubuf].emit_insn(dout_dilated_ubuf.op.axis[0],
insn_cmd.DMA_PADDING)
else:
s[dout_dilated_ubuf].compute_inline()
s[dout_mul_ubuf].compute_at(s[dout_cbuf_nc1hwc0], ub_l1hcut_o)
s[dout_ubuf].compute_at(s[dout_cbuf_nc1hwc0], ub_l1hcut_o)
s[dvealuemean_ubuf].compute_at(s[dout_cbuf_nc1hwc0], ub_l1hcut_o)
s[dout_ubuf].emit_insn(dout_ubuf.op.axis[0], insn_cmd.DMA_COPY)
s[dvealuemean_ubuf].emit_insn(dvealuemean_ubuf.op.axis[0],
insn_cmd.DMA_COPY)
s[dout_mul_ubuf].emit_insn(dout_mul_ubuf.op.axis[0], insn_cmd.MUL)
s[dout_cbuf_nc1hwc0].emit_insn(ub_l1hcut_i, insn_cmd.DMA_COPY)
# emit convolution params.
setfmatrix_dict = {
"conv_kernel_h": res.op.attrs['weight_height'],
"conv_kernel_w": res.op.attrs['weight_width'],
"conv_padding_top": res.op.attrs['dilated_pad'][0],
"conv_padding_bottom": res.op.attrs['dilated_pad'][1],
"conv_padding_left": res.op.attrs['dilated_pad'][2],
"conv_padding_right": res.op.attrs['dilated_pad'][3],
"conv_stride_h": res.op.attrs['dilated_strides'][0],
"conv_stride_w": res.op.attrs['dilated_strides'][1],
"conv_fm_c": dout_dilated.shape[2] * dout_dilated.shape[5],
"conv_fm_h": dout_dilated.shape[3],
"conv_fm_w": dout_dilated.shape[4]
}
s[dout_cbuf_row_major].emit_insn(dout_cbuf_row_major.op.axis[1],
insn_cmd.SET_FMATRIX, setfmatrix_dict)
s[dout_ca].emit_insn(dout_ca.op.axis[1], insn_cmd.IM2COL)
s[weight_cbuf].emit_insn(weight_cbuf.op.axis[0], insn_cmd.DMA_COPY)
s[weight_cb].emit_insn(weight_cb.op.axis[3], insn_cmd.DMA_COPY)
s[mad_ubuf].emit_insn(mad_ubuf_Ncut_i, insn_cmd.DMA_COPY)
mad_dict = {
"mad_pattern": tbe_platform.cce_params.CONV_MODE,
"k_outer": mad_cc_kcut_o
}
s[mad_cc].emit_insn(mad_cc_Ncut_i, insn_cmd.MAD, mad_dict)
s[res].emit_insn(conv_Ncut_i, insn_cmd.DMA_COPY)
s[dout_ca].double_buffer()
s[weight_cb].double_buffer()
s[mad_cc].double_buffer()
# for multi cores
if res_block_n < 16:
res_NNCut_o, res_NNCut_i = s[res].split(conv_Ncut_o,
nparts=res_block_n)
res_ccCut_o, res_ccCut_i = s[res].split(res.op.axis[1],
nparts=res_block_cgroup)
s[res].reorder(res_NNCut_o, res_ccCut_o, res_NNCut_i, res_ccCut_i)
out_fused = s[res].fuse(res_NNCut_o, res_ccCut_o)
out_fused_out, _ = s[res].split(out_fused,
nparts=res_block_n * res_block_cgroup)
bind_out, _ = s[res].split(out_fused_out, 1)
blockidx = tvm.thread_axis("blockIdx.x")
s[res].bind(bind_out, blockidx)
else:
block = tvm.thread_axis("blockIdx.x")
s[res].bind(conv_Ncut_o, block)
return s
def custom_truncatemod(shape1, shape2, dtype, kernel_name="cce_tf_truncatemod",
need_build=False, need_print=False):
"""
do element-wise truncatemod operation between two input tensors
Parameters:
----------
shape1 : shape of input data1
shape2 : shape of input data2
dtype : source data type, support float16,float32,int32
kernel_name : cce kernel name, default value is "cce_tf_truncatemod"
need_buid : if need to build CCEC kernel, default value is False
need_print : if need to print the ir, default value is False
Returns
-------
None
"""
max_dim = 8
shape1_len = len(shape1)
shape2_len = len(shape2)
if shape1_len > max_dim or shape2_len > max_dim:
raise RuntimeError(
"mod_cce only support up to %d dimensions while the shape's \
dimensions is %d, %d" % (max_dim, shape1_len, shape2_len))
util.check_kernel_name(kernel_name)
util.check_shape_rule(shape1)
util.check_shape_rule(shape2)
util.check_shape_size(shape1, SHAPE_SIZE_LIMIT)
util.check_shape_size(shape2, SHAPE_SIZE_LIMIT)
check_list = ["float16", "float32", "int32"]
device_api_map = {"float16": "cc_device_truncatemod_float16",
"float32": "cc_device_truncatemod_float",
"int32": "cc_device_truncatemod_int32"}
dtype = dtype.lower()
if dtype not in check_list:
raise RuntimeError(
"tf_truncatemod_cce only support %s while dtype is %s" % (
",".join(check_list), dtype))
shape1, shape2, shape_out = util.produce_shapes(shape1, shape2)
util.check_shape_size(shape_out, SHAPE_SIZE_LIMIT)
inp_dtype = dtype.lower()
device_api = device_api_map[inp_dtype]
# block
block_num = "block_num"
block_idx = "block_idx"
# x param
v_xndim_cnt = tvm.const(len(shape1), "int32")
p_xshape = util.create_param_ptr(shape1, "int32", "p_xshape")
xpad_c0 = tvm.const(0, "int32")
data_input_x = tvm.placeholder(shape1, name="data_input_x",
dtype=inp_dtype)
# y param
v_yndim_cnt = tvm.const(len(shape2), "int32")
p_yshape = util.create_param_ptr(shape2, "int32", "p_yshape")
ypad_c0 = tvm.const(0, "int32")
data_input_y = tvm.placeholder(shape2, name="data_input_y",
dtype=inp_dtype)
# output
v_out_ndim_cnt = tvm.const(len(shape_out), "int32")
p_out_shape = util.create_param_ptr(shape_out, "int32", "p_yshape")
out_padc0 = tvm.const(0, "int32")
output = tvm.extern(shape_out,
[p_xshape, data_input_x, p_yshape, data_input_y,
p_out_shape], lambda ins, outs:
tvm.call_extern("int32_t", device_api,
block_num,
block_idx,
v_xndim_cnt,
ins[0].access_ptr("r"), # shape x
xpad_c0,
ins[1].access_ptr("r"), # input x
v_yndim_cnt,
ins[2].access_ptr("r"), # shape y
ypad_c0,
ins[3].access_ptr("r"), # input y
v_out_ndim_cnt,
ins[4].access_ptr("r"), # shape out
out_padc0,
outs[0].access_ptr("w")),
name="output", dtype=inp_dtype)
schedule = tvm.create_schedule(output.op)
# print IR
if need_print:
with build_config:
print(tvm.lower(schedule, [data_input_x, data_input_y, output],
simple_mode=True))
# Compile to generate the cce file
if need_build:
with build_config:
tvm.build(schedule, [data_input_x, data_input_y, output], "cce",
name=kernel_name)
def _unpack_schedule(input_place, output_shape, y, num, axis, dtype):
"""Create unpack schedule.
Parameters
----------
input_place: TVM tensor
the tensor of input.
output_shape: tuple or list
the shape of output tensor.
y: tuple or list
the list of output tensor.
num : int.
the length of the dim axis.
axis: int.
the axis to unpack along.
dtype: str.
the dtype of input.
Returns
-------
sch: schedule
the created schedule.
build_list: list
the list of input and output tensors, tensor type is TVM tensor.
"""
_, ele_each_block, device_core_num = _get_public_param(dtype)
befordim, afterdim = output_shape[0], output_shape[-1]
block_idx = tvm.thread_axis('blockIdx.x')
# can open multi-core scene
if befordim >= ele_each_block and afterdim < ele_each_block:
befordim_in = ele_each_block // afterdim + 1
befordim_out = (befordim + befordim_in - 1) // befordim_in
while (befordim + befordim_out -
1) // befordim_out * afterdim < ele_each_block:
befordim_out -= 1
if befordim_out >= device_core_num:
befordim_out = device_core_num
afterdim_in = afterdim
gm2ub_tensor, ub2ub_tensor_list, ub2gm_tensor_list, virtual_node = _unpack_compute_scalar(
input_place, y, num, axis)
res_op = []
build_list = [input_place]
for ub2gm_tensor in ub2gm_tensor_list:
res_op.append(ub2gm_tensor.op)
build_list.append(ub2gm_tensor)
sch = tvm.create_schedule(virtual_node.op)
sch[gm2ub_tensor].set_scope(tbe_platform.scope_ubuf)
for tensor in ub2ub_tensor_list:
sch[tensor].set_scope(tbe_platform.scope_ubuf)
befordim_outer, befordim_inner = sch[virtual_node].split(
virtual_node.op.axis[0], nparts=befordim_out)
afterdim_outer, afterdim_inner = sch[virtual_node].split(
virtual_node.op.axis[2], factor=afterdim_in)
sch[virtual_node].reorder(befordim_outer, afterdim_outer,
befordim_inner, afterdim_inner)
fused_axis = sch[virtual_node].fuse(befordim_outer, afterdim_outer)
sch[virtual_node].bind(fused_axis, block_idx)
new_shape = ((befordim + befordim_out - 1) // befordim_out, num,
afterdim_in)
split_axis, split_factor = _tiling_axis(new_shape, dtype)
if split_axis == 0:
axis_outer, axis_inner = sch[virtual_node].split(
befordim_inner, factor=split_factor)
else:
axis_outer, axis_inner = sch[virtual_node].split(
afterdim_inner, factor=split_factor)
sch[gm2ub_tensor].compute_at(sch[virtual_node], axis_outer)
sch[gm2ub_tensor].emit_insn(gm2ub_tensor.op.axis[split_axis],
insn_cmd.DMA_COPY)
for i in range(num):
sch[ub2gm_tensor_list[i]].compute_at(sch[virtual_node], axis_outer)
sch[ub2ub_tensor_list[i]].compute_at(sch[virtual_node], axis_outer)
sch[ub2ub_tensor_list[i]].emit_insn(
ub2ub_tensor_list[i].op.axis[split_axis], insn_cmd.DATA_MOV)
sch[ub2gm_tensor_list[i]].emit_insn(
ub2gm_tensor_list[i].op.axis[split_axis], insn_cmd.DMA_COPY)
sch[virtual_node].emit_insn(axis_inner, insn_cmd.PHONY_INSN)
else:
gm2ub_tensor_list, ub2gm_tensor_list, virtual_node = _unpack_compute_copy(
input_place, y, num, axis)
res_op = []
build_list = [input_place]
for ub2gm_tensor in ub2gm_tensor_list:
res_op.append(ub2gm_tensor.op)
build_list.append(ub2gm_tensor)
sch = tvm.create_schedule(virtual_node.op)
for tensor in gm2ub_tensor_list:
sch[tensor].set_scope(tbe_platform.scope_ubuf)
# can open multi-core scene
if afterdim >= ele_each_block:
if befordim >= device_core_num:
befordim_out = device_core_num
afterdim_in = afterdim
elif befordim == 1:
befordim_out = befordim
afterdim_in = (afterdim + device_core_num -
1) // device_core_num
else:
afterdim_outer = device_core_num // befordim
afterdim_in = (afterdim + afterdim_outer - 1) // afterdim_outer
while afterdim % afterdim_in < ele_each_block:
afterdim_in += 1
befordim_out = befordim
befordim_outer, befordim_inner = sch[virtual_node].split(
virtual_node.op.axis[0], nparts=befordim_out)
afterdim_outer, afterdim_inner = sch[virtual_node].split(
virtual_node.op.axis[2], factor=afterdim_in)
sch[virtual_node].reorder(befordim_outer, afterdim_outer,
befordim_inner, afterdim_inner)
fused_axis = sch[virtual_node].fuse(befordim_outer, afterdim_outer)
sch[virtual_node].bind(fused_axis, block_idx)
new_shape = ((befordim + befordim_out - 1) // befordim_out, 1,
afterdim_in)
split_axis, split_factor = _tiling_axis(new_shape, dtype)
if split_axis == 0:
axis_outer, axis_inner = sch[virtual_node].split(
befordim_inner, factor=split_factor)
else:
axis_outer, axis_inner = sch[virtual_node].split(
afterdim_inner, factor=split_factor)
else:
split_axis, split_factor = _tiling_axis(output_shape, dtype)
axis_outer, axis_inner = sch[virtual_node].split(
virtual_node.op.axis[split_axis], factor=split_factor)
for i in range(num):
storage_axis = split_axis - 1 if split_axis != 0 else 0
sch[gm2ub_tensor_list[i]].storage_align(
gm2ub_tensor_list[i].op.axis[storage_axis], ele_each_block, 0)
sch[gm2ub_tensor_list[i]].double_buffer()
sch[gm2ub_tensor_list[i]].compute_at(sch[virtual_node], axis_outer)
sch[ub2gm_tensor_list[i]].compute_at(sch[virtual_node], axis_outer)
sch[gm2ub_tensor_list[i]].emit_insn(
gm2ub_tensor_list[i].op.axis[split_axis], insn_cmd.DMA_COPY)
sch[ub2gm_tensor_list[i]].emit_insn(
ub2gm_tensor_list[i].op.axis[split_axis], insn_cmd.DMA_COPY)
sch[virtual_node].emit_insn(axis_inner, insn_cmd.PHONY_INSN)
return sch, build_list
def cast(input_x, output_y, dst_type, kernel_name="cast"):
"""
cast a tensor/scaler with input shape form src data type to dst data
type. restrictions of input algorithms are as follow
only types' groups blow are support tensor process:
float16->float32
float16->int32
float32->float16
float32->int32
int8->float32
uint8->float32
int8->float16
uint8->float16
int8->int32
uint8->int32
int32->uint8 // number out of [0,255] can get unexpected result
int32->int8 // number out of [-128,127] can get unexpected result
int32->float32 // For tans with fp16, only guarantees
number in [-1023,1023] get correct result
int32->float16 // only guarantees
number in [-1023,1023] get correct result
scale convert support:(means only support shape [1,])
int64->int32
int64->float32
Parameters
----------
input_x : dict
shape and dtype of input, only support float16, float32
output_y: dict
shape and dtype of output, should be same shape as input,
and the dtype is the dst dtype need to cast
kernel_name : str
cce kernel name, default value is cast
Returns
-------
None
"""
shape = util.scalar2tensor_one(input_x.get("shape"))
src_type = input_x.get("dtype").lower()
check_shape(shape, param_name="input_x")
if src_type == "bool":
src_type = "int8"
dst_type = _cast_dsttype_conversion(dst_type)
fuseshape = [1]
fuseshape[0] = reduceIns(lambda x, y: x * y, shape)
data = tvm.placeholder(fuseshape, name="data", dtype=src_type)
if src_type == "int64":
check_dtype(dst_type, ("float32", "int32"), param_name="dst_type")
res = tvm.extern(
[fuseshape], [data],
lambda ins, outs: _kernel_ir(outs, ins, dst_type, "int64"),
name="res",
dtype=dst_type)
tensor_list = [data, res]
schedule = tvm.create_schedule(res.op)
with build_config:
tvm.build(schedule, tensor_list, "cce", name=kernel_name)
else:
with tvm.target.cce():
res = cast_compute(data, output_y, dst_type, kernel_name)
sch = generic.auto_schedule(res)
config = {
"print_ir": False,
"name": kernel_name,
"tensor_list": [data, res]
}
te.lang.cce.cce_build_code(sch, config)
def drop_out_do_mask(input_tensor,
input_mask,
input_keep_prob,
output,
kernel_name="dropout_do_mask"):
"""
algorithm: tf_dropout_do_mask
scale_x = x*(1 / keep_prob)
res = select(mask == 1, scale_x, 0)
Parameters
----------
input_tensor : dict,shape and dtype of input_tensor,only support float16 and float32
input_mask : dict,shape and dtype of input_mask
shape of mask,1D, dtype == uint8
length=(size(shape_tensor)+ELEMS_BATCH_PROCESS_FP16
-1)/ELEMS_BATCH_PROCESS_FP16*ELEMS_BATCH_PROCESS_FP16/8
eg. shape_tensor=[2,5,8] shape_mask=[16] shape_res=[2,5,8]
shape_tensor=[15,17,19] shape_mask=[608] shape_res=[15,17,19]
input_keep_prob : dict,shape and dtype of input_keep_prob
shape of keep_prob, only 1 parament and equals to (1)
prob scale (0.0,1.0] NOTICE: type same as dytpe
output : dict,shape and dtype of output
kernel_name : str
cce kernel name, default value is "dropout_do_mask"
Returns
-------
None
"""
shape_tensor = input_tensor.get("shape")
shape_mask = input_mask.get("shape")
shape_keep_prob = input_keep_prob.get("shape")
dtype = input_tensor.get("dtype")
if shape_keep_prob == 1:
shape_keep_prob = (shape_keep_prob, )
check_shape(shape_tensor, param_name="input_tensor")
check_dtype(dtype.lower(), ["float16", "float32"],
param_name="input_tensor")
if len(shape_mask) != 1:
raise RuntimeError("The length of mask shape must be 1")
if shape_keep_prob not in [(1, ), [
1,
]]:
raise RuntimeError("Only support shape (1, ) or [1, ]")
# functools_reduce: product of all dimension
# Align to ELEMS_BATCH_PROCESS_FP16
product_mask = (functools_reduce(lambda x, y: x*y, shape_tensor[:]) +
ELEMS_BATCH_PROCESS_FP16 - 1) // \
ELEMS_BATCH_PROCESS_FP16 * ELEMS_BATCH_PROCESS_FP16 // 8
if product_mask != shape_mask[0]:
raise RuntimeError("The mask[0] should=%d, but now=%d" %
(product_mask, shape_mask[0]))
data_tensor = tvm.placeholder(
(functools_reduce(lambda x, y: x * y, shape_tensor), ),
dtype=dtype,
name="data_tensor")
data_mask = tvm.placeholder(
(functools_reduce(lambda x, y: x * y, shape_mask), ),
dtype='uint8',
name="data_mask")
keep_prob_tensor = tvm.placeholder(shape_keep_prob,
dtype=dtype,
name="keep_prob_tensor")
const_1 = tvm.const(1.0, dtype=dtype)
res = tvm.extern([shape_tensor, shape_mask, shape_keep_prob],
[data_tensor, data_mask, keep_prob_tensor],
lambda ins, outs: _kernel_ir(outs, ins, const_1),
name="res",
dtype=dtype)
tensor_list = [data_tensor, data_mask, keep_prob_tensor, res]
schedule = tvm.create_schedule(res.op)
with build_config:
tvm.build(schedule, tensor_list, "cce", name=kernel_name)
def custom_round(shape,
dtype,
kernel_name="cce_round",
need_build=False,
need_print=False):
"""
doing round operations, calculating data type is float16 or float32 or int32
Parameters
----------
shape : shape of data
dtype : the data type, assume src_dtype equals dst_dtype
kernel_name : cce kernel name, default value is "cce_round"
need_buid : if need to build CCEC kernel, default value is False
need_print : if need to print the ir, default value is False
Returns
-------
None
"""
check_list = ["float16", "float32", "int32"]
device_api_map = {
"float16": "cc_device_round_float16",
"float32": "cc_device_round_float",
"int32": "cc_device_round_int32"
}
max_dim = 8
shape_len = len(shape)
if shape_len > max_dim:
raise RuntimeError(
"round_cce only support up to %d dimensions while the shape's dimension is %d"
% (max_dim, shape_len))
util.check_kernel_name(kernel_name)
util.check_shape_rule(shape)
util.check_shape_size(shape, SHAPE_SIZE_LIMIT)
if not (dtype.lower() in check_list):
raise RuntimeError("round_cce only support %s while dtype is %s" %
(",".join(check_list), dtype))
inp_dtype = dtype.lower()
shape = util.shape_refine(shape)
data_input = tvm.placeholder(shape, name="data_input", dtype=inp_dtype)
device_api = device_api_map[inp_dtype]
block_num = "block_num"
block_idx = "block_idx"
v_ndim = tvm.const(len(shape), "int32")
padC0 = tvm.const(0, "int32")
p_shape = util.create_param_ptr(shape, "int32", "p_shape")
output = tvm.extern(
shape,
[data_input, p_shape],
lambda ins, outs: tvm.call_extern(
"int32_t",
device_api,
block_num,
block_idx,
v_ndim,
ins[1].access_ptr("r"), # shape
padC0,
ins[0].access_ptr("r"), # input x
outs[0].access_ptr("w")),
name="output",
dtype=inp_dtype)
s = tvm.create_schedule(output.op)
if need_print:
with build_config:
print(tvm.lower(s, [data_input, output], simple_mode=True))
if need_build:
with build_config:
tvm.build(s, [data_input, output], "cce", name=kernel_name)
def strided_slice_d(input_x,
output_x,
begin,
end,
strides=None,
begin_mask=0,
end_mask=0,
ellipsis_mask=0,
new_axis_mask=0,
shrink_axis_mask=0,
kernel_name="strided_slice_d"):
"""
Extracts a strided slice of a tensor (generalized python array indexing).
Roughly speaking, this op extracts a slice of size (end-begin)/stride
from the given input_ tensor.
Starting at the location specified by begin the slice continues
by adding stride to the index
until all dimensions are not less than end. Note that a stride
can be negative, which causes a reverse slice.
Parameters
----------
input_x : dict
shape and dtype of input
output_x : dict
shape and dtype of out
begin: list.
represents the index of the first value to select.
end: list.
represents the index of the last value to select.
strides: list or tuple.
step length to select.
begin_mask: int
a bitmask where a bit i being 1 means to ignore the begin
value and instead use the
largest interval possible.
end_mask: int
analogous to `begin_mask`.
ellipsis_mask: int
a bitmask where bit `i` being 1 means the `i`th position
is actually an ellipsis.
new_axis_mask: int
a bitmask where bit `i` being 1 means the `i`th specification creates a
new shape 1 dimension.
shrink_axis_mask: int
a bitmask where bit `i` implies that the `i`th specification
should shrink the dimensionality.
kernel_name : str
cce kernel name, default value is "strided_slice_d"
Returns
-------
None
"""
input_shape = input_x.get("shape")
input_dtype = input_x.get("dtype").lower()
check_list = ("float16", "float32", "int32", "uint8", "bool", "int8")
check_dtype(input_dtype, check_list, param_name="input_x")
check_shape(input_shape, param_name="input_x")
begin = list(begin)
end = list(end)
if not _check_parameter(input_shape, begin, end, strides, ellipsis_mask,
new_axis_mask, shrink_axis_mask):
raise RuntimeError("Parameter Invalid!")
if strides is None:
strides = _fill_list_with_ones(len(input_shape))
else:
strides = list(strides)
input_tensor = tvm.placeholder(input_shape,
dtype=input_dtype,
name='input_tensor')
[output, out_shape] = strided_slice_d_compute(input_tensor,
output_x,
begin,
end,
strides,
begin_mask,
end_mask,
ellipsis_mask,
new_axis_mask,
shrink_axis_mask,
kernel_name=kernel_name)
# pylint: disable=locally-disabled,unnecessary-lambda
out_tensor = tvm.compute(out_shape,
lambda *i: output(*i),
name='out_tensor',
tag='strided_slice_d|3')
input_size = functools_reduce(lambda x, y: x * y, input_shape[0:])
out_size = functools_reduce(lambda x, y: x * y, out_shape[0:])
output_dtype = output_x.get("dtype").lower()
output_shape = output_x.get("shape")
if input_size == out_size:
if output_dtype == "bool":
input_x["dtype"] = "int8"
output_x["dtype"] = "int8"
if len(output_shape) == 0:
output_x["shape"] = (1, )
copy_only(input_x, output_x, kernel_name)
return
output_shape_one = list(output_shape)
if ellipsis_mask == 0 and shrink_axis_mask != 0:
for i, _ in enumerate(list(input_shape)):
if (shrink_axis_mask & 2**i) == 2**i:
output_shape_one.insert(i, 1)
output_shape = tuple(output_shape_one)
# for RL tune getting res
fusion_manager.set_op_res(out_tensor)
ret, sch = rl_bank.query_rl_bank([out_tensor])
if ret and sch:
with build_config:
tvm.build(sch, [input_tensor, out_tensor], "cce", name=kernel_name)
return
sch = tvm.create_schedule(out_tensor.op)
sch[output].set_scope(tbe_platform.scope_ubuf)
sch_input_shape = []
for dim in output.shape:
sch_input_shape.append(dim.value)
check_result = _check_last_axis_situation(sch_input_shape, begin, end,
strides)
if check_result:
_schedule_last_axis(sch, sch_input_shape, output, out_tensor,
input_dtype)
with build_config:
tvm.build(sch, [input_tensor, out_tensor], "cce", name=kernel_name)
return
if _check_tik_branch(input_shape, output_shape, begin, end, strides):
begin_shape = copy.deepcopy(begin)
end_shape = copy.deepcopy(end)
stride_shape = list(strides)
stride_shape = copy.deepcopy(stride_shape)
input_list = list(input_shape)
# update begin_shape, end_shape
begin_shape, end_shape, stride_shape = _init_parameter(
input_list, begin_shape, end_shape, stride_shape, begin_mask,
end_mask, ellipsis_mask, new_axis_mask, shrink_axis_mask)
head_size = 1
for i in range(0, (len(input_shape) - 1)):
head_size = head_size * input_shape[i]
if input_dtype == "float32" and input_shape[-1] == 2 and \
begin_shape[len(begin_shape) - 1] == 0 and end_shape[len(begin_shape) - 1] == 1 \
and head_size > 128:
strided_slice_two_turn_one(input_x, output_x, kernel_name)
return
if input_list[-1] > 80 and output_shape[-1] == 80:
res1 = strided_slice_last_dim_only(input_shape, input_dtype,
output_shape, begin_shape,
kernel_name)
if res1:
return
if input_list[-1] >= 32 and input_list[-1] < 7500 and len(output_shape) > 1 and \
output_shape[-1] >= 32:
res = strided_slice_last_dim_mte(input_shape, input_dtype,
output_shape, begin_shape,
kernel_name)
if res:
return
res = strided_slice_last_dim(input_shape, input_dtype, output_shape,
begin_shape, end_shape, stride_shape,
kernel_name)
if res:
return
else:
res1 = strided_slice_last_dim_one(input_shape, input_dtype,
output_shape, begin_shape,
kernel_name)
if res1:
return
split_axis, split_factor = _tilling_axis(out_shape, dtype=input_dtype)
core_state = _get_multicore(out_shape, input_dtype, split_axis,
split_factor)
axis_outer, axis_inner = sch[out_tensor].split(
out_tensor.op.axis[split_axis], factor=split_factor)
if split_axis == 0:
core_num = _get_target_core_num(out_shape[split_axis] // split_factor)
axis_outer_outer, axis_outer_inter = sch[out_tensor].split(
axis_outer, nparts=core_num)
else:
core_num = _get_target_core_num(out_shape[0])
axis_outer_outer, axis_outer_inter = sch[out_tensor].split(
out_tensor.op.axis[0], nparts=core_num)
for i in range(1, split_axis):
axis_outer_inter = sch[out_tensor].fuse(axis_outer_inter,
out_tensor.op.axis[i])
axis_outer_inter = sch[out_tensor].fuse(axis_outer_inter, axis_outer)
sch[output].compute_at(sch[out_tensor], axis_outer_inter)
sch[output].emit_insn(output.op.axis[0], insn_cmd.DMA_COPY) # gm-ub
if len(out_shape) >= 2:
# Convert bytes to Bytes
dtype_bytes_size = tbe_platform.cce_intrin.get_bit_len(
input_dtype) // 8
# 32 means one block size(32 Bytes), divide by 32 to
# get the numbers of data that
# can be stored in one block.
element = 32 // dtype_bytes_size
align_axis = _get_align_axis(out_shape)
sch[output].storage_align(output.op.axis[align_axis], element, 0)
if core_state:
thread_block = tvm.thread_axis("blockIdx.x")
sch[out_tensor].bind(axis_outer_outer, thread_block)
sch[out_tensor].emit_insn(axis_inner, insn_cmd.DMA_COPY) # ub-gm
with build_config:
tvm.build(sch, [input_tensor, out_tensor], "cce", name=kernel_name)
def custom_pow(shape,
shape_y,
dtype,
kernel_name="cce_tf_pow",
need_build=False,
need_print=False):
"""
calculate x^y, calculating data type is float16 or float32 or int32
when x < 0 , the output is a meaningless value.
Parameters
----------
shape : shape of data
dtype : the data type, assume src_dtype equals dst_dtype, only support
float16, float32, int32
kernel_name : cce kernel name, default value is "tf_pow_cce"
need_buid : if need to build CCEC kernel, default value is False
need_print : if need to print the ir, default value is False
Returns
-------
None
"""
supported_dtypes = ["float16", "float32", "int32"]
device_api = "cc_device_pow"
util.check_kernel_name(kernel_name)
util.check_shape_rule(shape)
util.check_shape_size(shape, SHAPE_SIZE_LIMIT)
if not dtype.lower() in supported_dtypes:
raise RuntimeError("tf_pow_cce only support %s while dtype is %s" %
(",".join(supported_dtypes), dtype))
inp_dtype = dtype.lower()
shape = util.shape_refine(shape)
data_lhs = tvm.placeholder(shape, name="data_lhs", dtype=inp_dtype)
data_rhs = tvm.placeholder(shape, name="data_rhs", dtype=inp_dtype)
v_datatype = util.get_device_api_dtype(inp_dtype)
v_ndim = len(shape)
block_num = "block_num"
block_idx = "block_idx"
pad_c0 = 0
p_scale = util.create_param_ptr([0], inp_dtype, "p_scale")
p_shift = util.create_param_ptr([0], inp_dtype, "p_shift")
p_power = util.create_param_ptr([0], inp_dtype, "p_power")
p_shape = util.create_param_ptr(shape, "int32", "p_shape")
output = tvm.extern(
shape,
[data_lhs, data_rhs, p_scale, p_shift, p_power, p_shape],
lambda ins, outs: tvm.call_extern(
"int32_t",
device_api,
block_num,
block_idx,
v_datatype,
ins[2].access_ptr("r"), # scale
ins[3].access_ptr("r"), # shift
ins[4].access_ptr("r"), # power
v_ndim,
ins[5].access_ptr("r"), # shape
pad_c0,
ins[0].access_ptr("r"), # input x
v_ndim,
v_ndim,
ins[5].access_ptr("r"), # shape
pad_c0,
ins[1].access_ptr("r"), # input y
outs[0].access_ptr("w")),
name="output",
dtype=inp_dtype)
schedule = tvm.create_schedule(output.op)
if need_print:
with build_config:
print(
tvm.lower(schedule, [data_lhs, data_rhs, output],
simple_mode=True))
if need_build:
with build_config:
tvm.build(schedule, [data_lhs, data_rhs, output],
"cce",
name=kernel_name)
def custom_Upsample(shape,
dtype,
scale,
data_format="channels_last",
kernel_name="cce_darknet_upsample",
need_build=False,
need_print=False):
"""
Parameters
----------
shape: input tensor's shape
dtype: input tensor's dtype, support:`float16,float32
scale: the upsampling factors
data_format: "channels_last" or "channels_first"
kernel_name : kernel name, default value is "MyUpsample"
need_buid : if need to build CCEC kernel, default value is False
need_print : if need to print the ir, default value is False
Returns
-------
None
"""
"""
TODO:
Please refer to the TE DSL Manual, And code here with TE DSL.
"""
inp_dtype = dtype.lower()
check_list = ["float16", "float32", "int32", "int8", "uint8"]
if inp_dtype not in check_list:
raise RuntimeError("upsample only support %s while dtype is %s" %
(",".join(check_list), dtype))
util.check_kernel_name(kernel_name)
util.check_shape_rule(shape)
util.check_shape_size(shape, SHAPE_SIZE_LIMIT)
size = (scale, scale)
shape_size = len(shape)
if not (shape_size == 4 or shape_size == 5):
raise RuntimeError(
"upsample only support 4D or 5D while len(shape):%d" % len(shape))
input_tensor = tvm.placeholder(shape, name="input_tensor", dtype=inp_dtype)
res = None
if shape_size == 5:
# shape_size == 5 D-sepecial (N, C1, H, W, C0)
output_shape = (shape[0], shape[1], shape[2] * size[0],
shape[3] * size[1], shape[4])
res = tvm.compute(
output_shape, lambda n, c0, h, w, c: input_tensor[n, c0, h // size[
0], w // size[1], c])
else:
if data_format == "channels_last":
output_shape = (shape[0], shape[1] * size[0], shape[2] * size[1],
shape[3])
res = tvm.compute(
output_shape, lambda n, h, w, c: input_tensor[n, h // size[0],
w // size[1], c])
elif data_format == "channels_first":
output_shape = (shape[0], shape[1], shape[2] * size[0],
shape[3] * size[1])
res = tvm.compute(
output_shape, lambda n, c, h, w: input_tensor[n, c, h // size[
0], w // size[1]])
else:
raise RuntimeError(
"upsample only support channels_last|channels_first "
"while input type %s" % data_format)
schedule = tvm.create_schedule(res.op)
if need_print:
with build_config:
print(tvm.lower(schedule, [input_tensor, res], simple_mode=True))
if need_build:
with build_config:
tvm.build(schedule, [input_tensor, res], "cce", name=kernel_name)
def custom_logical_and(shape_x,
shape_y,
dtype,
kernel_name="cce_tf_logical_and",
need_build=False,
need_print=False):
"""
do element-wise logical-and operation between two input tensors
Parameters:
----------
shape_x : shape of input data1
shape_y : shape of input data2
dtype : source data type, support "bool"
kernel_name : cce kernel name, default value is "cce_tf_logical_and"
need_buid : if need to build CCEC kernel, default value is False
need_print : if need to print the ir, default value is False
Returns
-------
None
"""
util.check_kernel_name(kernel_name)
util.check_shape_rule(shape_x)
util.check_shape_rule(shape_y)
check_list = ["bool"]
if not (dtype.lower() in check_list):
raise RuntimeError(
"logical_and_cce only support %s while dtype is %s" %
(",".join(check_list), dtype))
util.check_shape_size(shape_x, SHAPE_SIZE_LIMIT)
util.check_shape_size(shape_y, SHAPE_SIZE_LIMIT)
inp_dtype = dtype.lower()
shape_x, shape_y, shape_max = util.produce_shapes(shape_x, shape_y)
data1 = tvm.placeholder(shape_x, dtype=inp_dtype, name="data1")
data2 = tvm.placeholder(shape_y, dtype=inp_dtype, name="data2")
with tvm.target.cce():
data1_tmp1 = te.lang.cce.broadcast(data1, shape_max)
data1_tmp2 = te.lang.cce.broadcast(data2, shape_max)
min_value = tvm.const(0, dtype=inp_dtype)
res = tvm.compute(
shape_max,
lambda *i: tvm.select(
tvm.all(
tvm.any(
data1_tmp1(*i) > min_value,
data1_tmp1(*i) < -min_value),
tvm.any(
data1_tmp2(*i) > min_value,
data1_tmp2(*i) < -min_value)), True, False),
name="res")
sch = tvm.create_schedule(res.op)
if need_print:
with build_config:
print(tvm.lower(sch, [data1, data2, res], simple_mode=True))
if need_build:
with build_config:
tvm.build(sch, [data1, data2, res], "cce", name=kernel_name)
def custom_expm1(shape,
dtype,
kernel_name="cce_tf_expm1",
need_build=False,
need_print=False):
"""
algorithm: expm1
calculating data's expm1, y= (e ** x) - 1,dtype is float16 or float32.
Parameters
----------
shape : shape of data.
dtype : the data type, assume src_dtype equals dst_dtype, only support float16, float32.
kernel_name : cce kernel name, default value is "cce_tf_expm1".
need_buid : if need to build CCEC kernel, default value is False.
need_print : if need to print the ir, default value is False.
Returns
-------
None
"""
# [aicpu] int32_t cc_device_exp(uint32_t blockNum, uint32_t blockIdx, int32_t dataType, const void *scale, const void *shift,
# const void *base, int32_t dimCnt, int32_t *shape, uint32_t padC0, const void *x, void *y);
supported_dtypes = ["float16", "float32"]
util.check_kernel_name(kernel_name)
util.check_shape_rule(shape)
util.check_shape_size(shape, SHAPE_SIZE_LIMIT)
if not (dtype.lower() in supported_dtypes):
raise RuntimeError("tf_expm1_cce only support %s while dtype is %s" %
(",".join(supported_dtypes), dtype))
inp_dtype = dtype.lower()
shape = util.shape_refine(shape)
data_input = tvm.placeholder(shape, name="data_input", dtype=inp_dtype)
# step 1. calculate y = exp ** x by aicpu api
device_api = "DeviceExp"
v_datatype = util.get_device_api_dtype(inp_dtype)
v_ndim = len(shape)
block_num = "block_num"
block_idx = "block_idx"
padC0 = 0
p_scale = util.create_param_ptr([1], inp_dtype, "p_scale")
p_shift = util.create_param_ptr([0], inp_dtype, "p_shift")
p_base = util.create_param_ptr([-1], inp_dtype, "p_base")
p_shape = util.create_param_ptr(shape, "int32", "p_shape")
output_exp = tvm.extern(
shape,
[data_input, p_scale, p_shift, p_base, p_shape],
lambda ins, outs: tvm.call_extern(
"int32_t",
device_api,
block_num,
block_idx,
v_datatype,
ins[1].access_ptr("r"), # scale
ins[2].access_ptr("r"), # shift
ins[3].access_ptr("r"), # base
v_ndim,
ins[4].access_ptr("r"), # shape
padC0,
ins[0].access_ptr("r"), # input x
outs[0].access_ptr("w")),
name="output_exp",
dtype=inp_dtype)
offset = tvm.const((-1), dtype=inp_dtype)
# step 2. cauculate y = exp ** x - 1 by tvm
output = tvm.compute(
shape,
lambda *indice: output_exp(*indice) + offset.astype(inp_dtype),
name="output")
# step 3. schedule the computation by tvm
s = tvm.create_schedule(output.op)
# step 4. build by tvm
if need_print:
with build_config:
print(tvm.lower(s, [data_input, output], simple_mode=True))
if need_build:
with build_config:
tvm.build(s, [data_input, output], "cce", name=kernel_name)
def max_pool_grad_grad_with_argmax(x,
grad,
argmax,
y,
ksize,
strides,
padding="VALID",
kernel_name="cce_max_pool_grad_grad"
"_with_argmax"):
"""
Computes second-order gradients of the maxpooling function.
Parameters
----------
x: dict
Include info about ori_input,
format, ori_format, shape, ori_shape, dtype.
grad: dict
Include info about grad of ori_input,
format, ori_format, shape, ori_shape, dtype.
argmax: dict
Include info about ori_input,
format, ori_format, shape, ori_shape, dtype.
y: dict
Include info about result of function,
format, ori_format, shape, ori_shape, dtype.
ksize: list or tuple
The size of the window for each dimension of the input tensor.
strides: list or tuple
The stride of the sliding window of the input tensor.
padding: str
The type of padding algorithm to use.
Only support "VALID" or "SAME"
kernel_name: str
Cce kernel name,
default value is "cce_max_pool_grad_grad_with_argmax"
Returns
-------
None
"""
check_shape_and_format_vailded(x, grad, argmax, y, ksize, strides, padding,
kernel_name)
shape_x = x.get("shape")
shape_grad = grad.get("shape")
shape_argmax = argmax.get("shape")
shape_argmax = (shape_argmax[0], shape_argmax[1], shape_argmax[2],
shape_argmax[3] * shape_argmax[4], 1)
dtype_x = x.get("dtype").lower()
dtype_grad = grad.get("dtype").lower()
ori_format_x = x.get("ori_format")
x_tensor = tvm.placeholder(shape_x, dtype=dtype_x, name="input_x")
# argmax is continuous bool, real type is uint16
_, _, _, howo, _ = shape_argmax
shape_argmax_boolean = (shape_argmax[0], shape_argmax[1] * shape_argmax[2],
howo // 16, 16, shape_argmax[4])
shape_argmax_boolean = list(shape_argmax_boolean[:-1]) + list(
[shape_argmax_boolean[-1] * 16])
argmax_tensor = tvm.placeholder(shape_argmax_boolean,
dtype="bool",
name="argmax")
grad_tensor = tvm.placeholder(shape_grad,
dtype=dtype_grad,
name="input_grad")
compute_list = _max_pool_grad_grad_with_argmax_compute(
[x_tensor, argmax_tensor, grad_tensor], x, argmax, grad, y, ksize,
strides, padding, ori_format_x, kernel_name)
res = compute_list[-1]
sch = tvm.create_schedule(res.op)
_max_pool_grad_grad_with_argmax_schedule(compute_list, [sch])
tensor_list = [x_tensor, grad_tensor, argmax_tensor, res]
new_config = build_config_update(build_config, "dummy_placeholder", True)
with new_config:
tvm.build(sch, tensor_list, "cce", name=kernel_name)
def custom_Reduction(shape,
dtype,
axis,
op,
coeff,
kernel_name="cce_reductionLayer",
need_build=False,
need_print=False):
"""
Reduce a tensor on a certain axis, and scale output with coeff
Parameters
----------
shape : shape of data
dtype : source data type, only support float16, float32, int8, uint8
axis : the first axis to reduce, may be negative to index from the end
(e.g., -1 for the last axis).
If axis == 0, the output Blob always has the empty shape (count 1),
performing reduction across the entire input.
op : can only be one of "SUM, ASUM (sum of abs), SUMSQ (sum of sqr), MEAN"
coeff : scale for output
kernel_name : cce kernel name, default value is "cce_reductionLayer"
need_buid : if need to build CCEC kernel, default value is False
need_print : if need to print the ir, default value is False
Returns
-------
None
"""
util.check_kernel_name(kernel_name)
util.check_shape_rule(shape)
check_list = ["float16", "float32", "int8", "uint8"]
if not dtype.lower() in check_list:
raise RuntimeError(
"reductionLayer_cce only support %s while dtype is %s" %
(",".join(check_list), dtype))
reduction_op = ("SUM", "ASUM", "SUMSQ", "MEAN")
if not isinstance(axis, int):
raise RuntimeError("type of axis value should be int")
if op not in reduction_op:
raise RuntimeError("op can only be one of SUM, ASUM, SUMSQ , MEAN")
if not isinstance(coeff, int) and not isinstance(coeff, float):
raise RuntimeError("coeff must be a value")
axis_origin = axis
shape_origin = shape
axis = util.axis_check(len(shape), axis)
util.check_reduce_shape_rule(shape)
shape = list(shape)
shape1 = shape[:axis] + [
functools_reduce(lambda x, y: x * y, shape[axis:])
]
shape1, axis = util.shape_refine(shape1, axis)
if not axis:
axis = [0]
shape1 = [1] + shape1
inp_dtype = dtype.lower()
data = tvm.placeholder(shape1, name="data_input", dtype=inp_dtype)
with tvm.target.cce():
res = caffe_reduction_layer_compute([data], shape_origin, dtype,
axis_origin, op, coeff,
kernel_name, need_build,
need_print)
if op == "MEAN" and (inp_dtype == "int8" or inp_dtype == "uint8"):
util.check_shape_size(shape, SHAPE_SIZE_LIMIT)
res = te.lang.cce.cast_to(res, inp_dtype)
schedule = tvm.create_schedule(res.op)
if need_print:
with build_config:
print(tvm.lower(schedule, [data, res], simple_mode=True))
if need_build:
with build_config:
tvm.build(schedule, [data, res], "cce", name=kernel_name)
else:
with tvm.target.cce():
sch = generic.auto_schedule(res)
config = {
"print_ir": need_print,
"need_build": need_build,
"name": kernel_name,
"tensor_list": [data, res]
}
te.lang.cce.cce_build_code(sch, config)
def histogram_fixed_width_d(x,
range,
y,
nbins,
dtype="int32",
kernel_name='histogram_fixed_width_d'):
"""this operation returns a rank 1 histogram counting
the number of entries in `values` that fell into every bin.
The bins are equal width and determined by the arguments
`value_range` and `nbins`.
Parameters
----------
x: dict
dict info of input value, must include the keys(shape and dtype).
range: dict
dict info of input value_range, must include the keys(shape and dtype).
the shape must be (2,) or [2]
y: dict
dict info of output
nbins: int
number of histogram bins.
dtype: str
data type for returned histogram.
kernel_name: str
cce kernel name, default value is "histogram_fixed_width"
returns
-------
None
"""
input_shape_list = [x.get("shape"), range.get("shape")]
input_dtype = x.get("dtype")
dtype_input = input_dtype.lower()
check_shape(input_shape_list[0], param_name="x")
check_shape(input_shape_list[1], param_name="range")
util.compare_tensor_dict_key(x, range, "dtype")
data_shape_size = util.check_tensor_shape_size(list(input_shape_list[0]))
data_range_shape_size = util.check_tensor_shape_size(
list(input_shape_list[1]))
check_dtype(dtype_input, ("float16", "float32", "int32"), param_name="x")
if data_range_shape_size != 2:
raise RuntimeError("the shape of range must be (2,) or [2]")
if nbins <= 0:
raise RuntimeError("the nbins must be > 0")
data = tvm.placeholder([data_shape_size],
dtype=dtype_input,
name="input_data")
range_data = tvm.placeholder([data_range_shape_size],
dtype=dtype_input,
name="input_range_data")
res = histogram_fixed_width_d_compute(data, range_data, y, nbins,
kernel_name)
sch = tvm.create_schedule(res.op)
with build_config:
tvm.build(sch, [data, range_data, res], "cce", name=kernel_name)
def custom_batch_matmul(shape_x,
shape_y,
dtype,
trans_a=False,
trans_b=False,
kernel_name="cce_tf_batch_matmul",
need_build=False,
need_print=False):
"""
Multiplies slices of two tensors in batches(each slice can be viewed
as an element of a batch), the output is of the same batch size.
Each of the individual slices can optionally be transposed before
multiplication by setting the trans_a or trans_b flag to True, which
are by default False. The input tensors are 2-D or higher with the
shape [..., r_x, c_x] and [..., r_y, c_y]. The output tensor is 2-D
or higher with the shape [..., r_o, c_o], where
r_o = c_x if trans_a else r_x
c_o = r_y if trans_b else c_y
Parameters
----------
shape_x : shape of the first tensor x with rank > 1
shape_y : shape of the second tensor y with the same type and shape with x
dtype : the data type, support int8, uint8,float16,float32,int32
kernel_name : cce kernel name, default value is "cce_batch_matmul"
trans_a : if True, shape_x is transposed before multiplication
trans_b : if True, shape_y is transposed before multiplication
need_buid : if need to build CCEC kernel, default value is False
need_print : if need to print the ir, default value is False
Returns
-------
None
"""
util.check_kernel_name(kernel_name)
util.check_shape_rule(shape_x)
util.check_shape_rule(shape_y)
util.check_shape_size(shape_x, SHAPE_SIZE_LIMIT)
util.check_shape_size(shape_y, SHAPE_SIZE_LIMIT)
data_dtype = dtype.lower()
check_list = ["int8", "uint8", "float16", "float32", "int32"]
if data_dtype not in check_list:
raise RuntimeError(
"batch_matmul_cce ony supports %s while dtype is %s" %
(",".join(check_list), dtype))
def transpose_tensor(shape, size):
"""Transpose the shape, e.g., the shape [..., r_x, c_x] is transposed
to [..., c_x, r_x].
Parameters
----------
shape : shape of a tensor
size : length of the shape
Returns
-------
shape_ori : the transposed shape
"""
shape_ori = ()
if size == 1:
shape_ori = shape_ori + shape
elif size == 2:
shape_ori = shape_ori + (shape[1], ) + (shape[0], )
else:
shape_ori = shape_ori + (shape[:(size - 2)]) + (
shape[size - 1], ) + (shape[size - 2], )
return shape_ori
def check_matmul(shape_x, shape_y):
"""Check whether batch_matmul is supported or not.
Parameters
----------
shape_x : shape of the first tensor x
shape_y : shape of the second tensor y with the same type and shape
with x
Returns
-------
None
"""
len_x = len(shape_x)
len_y = len(shape_y)
if (len_x < 2) or (len_y < 2):
raise RuntimeError("Only tensors of rank>=2 are supported!")
if shape_x[len_x - 1] != shape_y[len_y - 2]:
raise RuntimeError(
"Invalid matrix multiplication for the inner 2 dimensions!")
if (len_x == len_y) and (len_x > 2):
for i in range(len_x - 2):
if shape_x[i] != shape_y[i]:
raise RuntimeError("Outer dimensions do not match!")
return
elif (len_x == len_y) and (len_x == 2):
return
else:
raise RuntimeError("The input tensors are not with the same rank!")
def _compute(output_shape, x, y, K, trans_a, trans_b, *indices):
"""matmul compuation in terms of the output shape and the transposes
Parameters
----------
output_shape : the final output shape, e.g., shape_x = (2, 6),
shape_y = (8, 2), trans_a = True, True_b = True, then,
output_shape = (6, 8).
x : the first input tensor according to shape_x.
y : the second input tensor according to shape_y.
K : the number of the axis for sum, in the above example, K = 2.
trans_a : if True, x needs to be transposed.
trans_b : if True, y needs to be transposed.
*indices : the output shape space for tvm.compute.
Returns
-------
tvm.Tensor
"""
n_len = len(output_shape)
k = tvm.reduce_axis((0, K), 'k')
if trans_a is True and trans_b is False:
# For example, A: (6, 7, 8), B: (6, 7, 9), so the length is n = 3
# C = A' * B : (6, 8, 9), A' means the transpose of A
# indices means the space of (6, 8, 9), k = 7
# x_indices = indices[:1]+(7, )+indices[1:2] = (6, 7, 8)
# y_indices = indices[:1]+(7, )+indices[2:] = (6, 7, 9)
x_indices = indices[:(n_len - 2)] + (k, ) + indices[(n_len - 2):
(n_len - 1)]
y_indices = indices[:(n_len - 2)] + (k, ) + indices[(n_len - 1):]
return tvm.sum(x(*x_indices) * y(*y_indices), axis=k)
elif not trans_a and trans_b:
# For example, A: (6, 7, 8), B: (6, 9, 8), C = A * B' : (6, 7, 9)
# indices means the space of (6, 7, 9), n=3, k = 8
# x_indices = indices[:2]+(8, ) = (6, 7, 8)
# y_indices = indices[:1]+indices[2:]+(8, ) = (6, 9, 8)
x_indices = indices[:(n_len - 1)] + (k, )
y_indices = indices[:(n_len - 2)] + indices[(n_len - 1):] + (k, )
return tvm.sum(x(*x_indices) * y(*y_indices), axis=k)
elif trans_a and trans_b:
# For example, A: (6, 8, 10), B: (6, 12, 8), C = A' * B' : \
# (6, 10, 12)
# indices means the space of (6, 10, 12), n=3, k = 8
# x_indices = indices[:1]+(8, )+indices[1:2] = (6, 8, 10)
# y_indices = indices[:1]+indices[2:]+(8, ) = (6, 12, 8)
x_indices = indices[:(n_len - 2)] + (k, ) + indices[(n_len - 2):
(n_len - 1)]
y_indices = indices[:(n_len - 2)] + indices[(n_len - 1):] + (k, )
return tvm.sum(x(*x_indices) * y(*y_indices), axis=k)
else:
# For example, A: (6, 15, 16), B: (6, 16, 18), C = A * B : \
# (6, 15, 18)
# indices means the space of (6, 15, 18), n=3, k = 16
# x_indices = indices[:2]+(16, ) = (6, 15, 16)
# y_indices = indices[:1]+(16, )+indices[2:] = (6, 16, 18)
x_indices = indices[:(n_len - 1)] + (k, )
y_indices = indices[:(n_len - 2)] + (k, ) + indices[(n_len - 1):]
return tvm.sum(x(*x_indices) * y(*y_indices), axis=k)
def check_supportted_shape_size(shape_x, shape_y, limit, trans_a, trans_b):
"""
check shape size for operator
----------
shape: shape of data
limit: limit of the product
Returns
-------
None
"""
# This function is used to check whether the shape is too large to \
# cause a timeout.
# shape_x = (a,b,c,d,e,k) shape_y = (a,b,c,d,k,f)
# t_1 : time consumed by each addition operation
# t_2 : time consumed by each multiplication operation
# t_all : time consumed by a complete calculation
# t_all is approximately equal to (a*b*c*d)*(e*k*f)*(t_1+t_2)
# As (t_1 + t_2) is a constant, so t_all is proportional to \
# (a * b * c * d * e * k * f)
len_x = len(shape_x)
len_y = len(shape_y)
if (len_x < 2) or (len_y < 2):
raise RuntimeError("Only tensors of rank>=2 are supported!")
shape_x = list(shape_x)
shape_y = list(shape_y)
tmp_shape_x = shape_x[:]
if trans_a:
tmp_shape_x = shape_x[:-2] + [shape_x[-1], shape_x[-2]]
tmp_shape_y = shape_y[:]
if trans_b:
tmp_shape_y = shape_y[:-2] + [shape_y[-1], shape_y[-2]]
union_shape = tmp_shape_x + [tmp_shape_y[-1]]
union_size = reduce(lambda i, j: i * j, union_shape)
if union_size > limit:
raise RuntimeError("the shape is too large to calculate")
if data_dtype in ["float16", "float32", "int32"]:
type_shape_map = {
'float16': SHAPE_SIZE_FP16_LIMIT,
'float32': SHAPE_SIZE_FP32_LIMIT,
'int32': SHAPE_SIZE_INT32_LIMIT
}
check_supportted_shape_size(shape_x, shape_y,
type_shape_map[data_dtype], trans_a,
trans_b)
x_size = len(shape_x)
y_size = len(shape_y)
shape_a = shape_x
shape_b = shape_y
if trans_a is True:
shape_x = transpose_tensor(shape_x, x_size)
if trans_b is True:
shape_y = transpose_tensor(shape_y, y_size)
check_matmul(shape_x, shape_y)
last_axis = shape_x[x_size - 1]
x_temp = tvm.placeholder(shape_a, name="input_1", dtype=data_dtype)
y_temp = tvm.placeholder(shape_b, name="input_2", dtype=data_dtype)
# output shape
output_shape = ()
for i in range(x_size - 1):
output_shape = output_shape + (shape_x[i], )
output_shape = output_shape + (shape_y[x_size - 1], )
result = tvm.compute(
output_shape,
lambda *indices: _compute(output_shape, x_temp, y_temp, last_axis,
trans_a, trans_b, *indices),
name="result")
schedule = tvm.create_schedule(result.op)
if need_print:
with build_config:
print(
tvm.lower(schedule, [x_temp, y_temp, result],
simple_mode=True))
if need_build:
with build_config:
tvm.build(schedule, [x_temp, y_temp, result],
"cce",
name=kernel_name)
def custom_Exp(shape,
dtype,
gamma,
alpha,
beta,
kernel_name="cce_exp",
need_build=False,
need_print=False):
"""
calculate gamma **(alpha * data + beta),
calculate exp(log(gamma) * alpha * data) * (gamma ** beta)
Parameters
----------
shape : shape of data
dtype : the data type, assume src_dtype equals dst_dtype, only support \
float16, float32
gamma : the data type must be same with dtype parameter
args in (alpha * data + beta) ** gamma, base
alpha : the data type must be same with dtype parameter
args in (alpha * data + beta) ** gamma, scale
beta : the data type must be same with dtype parameter
args in (alpha * data + beta) ** gamma, shift
kernel_name : cce kernel name, default value is "cce_exp"
need_buid : if need to build CCEC kernel, default value is False
need_print : if need to print the ir, default value is False
Returns
-------
None
"""
supported_dtypes = ["float16", "float32"]
device_api = "DeviceExp"
util.check_kernel_name(kernel_name)
util.check_shape_rule(shape)
util.check_shape_size(shape, SHAPE_SIZE_LIMIT)
if not dtype.lower() in supported_dtypes:
raise RuntimeError(
"caffe_exp_layer_cce only support %s while dtype is %s" %
(",".join(supported_dtypes), dtype))
if gamma != -1 and gamma <= 0:
# api cc_device_exp_c handle gamma == -1 as e
raise ValueError(
"please ensure gamma is greater than 0, where gamma = %s" %
str(gamma))
inp_dtype = dtype.lower()
shape = util.shape_refine(shape)
data_input = tvm.placeholder(shape, name="data_input", dtype=inp_dtype)
v_datatype = util.get_device_api_dtype(inp_dtype)
v_ndim = len(shape)
block_num = "block_num"
block_idx = "block_idx"
pad_c0 = 0
p_scale = util.create_param_ptr([alpha], inp_dtype, "p_scale")
p_shift = util.create_param_ptr([beta], inp_dtype, "p_shift")
p_base = util.create_param_ptr([gamma], inp_dtype, "p_base")
p_shape = util.create_param_ptr(shape, "int32", "p_shape")
# scale --> alpha, shitf --> beta, base --> gamma
output = tvm.extern(
shape,
[data_input, p_scale, p_shift, p_base, p_shape],
lambda ins, outs: tvm.call_extern(
"int32_t",
device_api,
block_num,
block_idx,
v_datatype,
ins[1].access_ptr("r"), # scale
ins[2].access_ptr("r"), # shift
ins[3].access_ptr("r"), # base
v_ndim,
ins[4].access_ptr("r"), # shape
pad_c0,
ins[0].access_ptr("r"), # input x
outs[0].access_ptr("w")),
name="output",
dtype=inp_dtype)
schedule = tvm.create_schedule(output.op)
if need_print:
with build_config:
print(tvm.lower(schedule, [data_input, output], simple_mode=True))
if need_build:
with build_config:
tvm.build(schedule, [data_input, output], "cce", name=kernel_name)
def SpatialTransformer(input_shape, out_shape, dtype="float32", kernel_name="SpatialTransformer", need_build = True, need_print = False):
"""Spatial Transformer Layer
Implements a spatial transformer layer as described in [1]_.
Based on [2]_.
Parameters
----------
input_shape :
the shape of input tensor
[num_batch, height, width, num_channels]
out_shape: float
the height and width of output tensor
[out_height, out_width].
out_size: tuple of two ints
The size of the output of the network (height, width)
dtype: data type
kernel_name : kernel name, default value is "SpatialTransformer"
need_buid : if need to build CCEC kernel, default value is True
need_print : if need to print the ir, default value is False
Returns
-------
tvm.Tensor
References
----------
.. [1] Spatial Transformer Networks
Max Jaderberg, Karen Simonyan, Andrew Zisserman, Koray Kavukcuoglu
.. [2] https://github.com/tensorflow/models/tree/master/research/transformer
"""
def _meshgrid(height, width):
y0 = tvm.compute((height,), lambda i: -1 + i * 2.0 / (height - 1), name = 'y0')
x0 = tvm.compute((width,), lambda i: -1 + i * 2.0 / (width - 1), name = 'x0')
y = tvm.compute((height * width,), lambda i: y0[i // width], name = 'y')
x = tvm.compute((height * width,), lambda i: x0[i % width], name = 'x')
y = topi.reshape(y, (1, height * width))
x = topi.reshape(x, (1, height * width))
ones = tvm.compute((1, height * width), lambda i,j:1, name = 'ones')
grid = tvm.compute((3, height * width),lambda i,j: 0.5 * (i - 1) * (i - 2) * x[0,j] + i * (2 - i) * y[0,j] + 0.5 * i * (i-1) * ones[0,j], name = 'grid')
#grid = topi.concatenate((x,y,ones),0) #can not use topi.concatenate
return grid
def _interpolate(im, im_shape, x, y, out_size, dtype):
num_batch = im_shape[0]
height = im_shape[1]
width = im_shape[2]
channels = im_shape[3]
out_height = out_size[0]
out_width = out_size[1]
max_y = int(im_shape[1] - 1)
max_x = int(im_shape[2] - 1)
#[-1,1] -> [0, width-1]
x = topi.multiply(topi.add(x, tvm.const(1, dtype=dtype)), width / tvm.const(2, dtype=dtype))
y = topi.multiply(topi.add(y, tvm.const(1, dtype=dtype)), height / tvm.const(2, dtype=dtype))
# do sampling
dim3 = out_height * out_width * num_batch
x0 = topi.cast(topi.floor(x), 'int32')
y0 = topi.cast(topi.floor(y), 'int32')
x1 = topi.add(x0,tvm.const(1, dtype="int32"))
y1 = topi.add(y0,tvm.const(1, dtype="int32"))
x0 = topi.clip(x0, 0, max_x)
x1 = topi.clip(x1, 0, max_x)
y0 = topi.clip(y0, 0, max_y)
y1 = topi.clip(y1, 0, max_y)
dim2 = width
dim1 = width * height
base = tvm.compute((dim3,),lambda i:(i // (out_height * out_width)) * width * height, name = 'base')
base_y0 = topi.add(base, topi.multiply(y0, dim2))
base_y1 = topi.add(base, topi.multiply(y1, dim2))
idx_a = topi.add(base_y0, x0)
idx_b = topi.add(base_y1, x0)
idx_c = topi.add(base_y0, x1)
idx_d = topi.add(base_y1, x1)
im_flat = topi.reshape(im, (num_batch * height * width, channels))
im_flat = topi.cast(im_flat, dtype)
Ia = tvm.compute((dim3, channels),lambda i,j: im_flat[idx_a[i], j], name = 'Ia')
Ib = tvm.compute((dim3, channels),lambda i,j: im_flat[idx_b[i], j], name = 'Ib')
Ic = tvm.compute((dim3, channels),lambda i,j: im_flat[idx_c[i], j], name = 'Ic')
Id = tvm.compute((dim3, channels),lambda i,j: im_flat[idx_d[i], j], name = 'Id')
x0_f = topi.cast(x0, dtype)
x1_f = topi.cast(x1, dtype)
y0_f = topi.cast(y0, dtype)
y1_f = topi.cast(y1, dtype)
wa = topi.expand_dims(topi.multiply(topi.subtract(x1_f, x), topi.subtract(y1_f, y)), 1)
wb = topi.expand_dims(topi.multiply(topi.subtract(x1_f, x), topi.subtract(y, y0_f)), 1)
wc = topi.expand_dims(topi.multiply(topi.subtract(x, x0_f), topi.subtract(y1_f, y)), 1)
wd = topi.expand_dims(topi.multiply(topi.subtract(x, x0_f), topi.subtract(y, y0_f)), 1)
output = topi.add(topi.add(topi.add(topi.multiply(wa, Ia), topi.multiply(wb, Ib)),topi.multiply(wc, Ic)), topi.multiply(wd, Id))
return output
def _transform(theta, input_dim, out_size, input_shape, dtype):
num_batch = input_shape[0]
height = input_shape[1]
width = input_shape[2]
num_channels = input_shape[3]
theta = topi.reshape(theta, (num_batch, 2, 3))
theta = topi.cast(theta, dtype)
out_height = out_size[0]
out_width = out_size[1]
grid = _meshgrid(out_height, out_width)
grid = topi.reshape(grid, (num_batch, 3, out_height*out_width))
grid = topi.cast(grid, dtype=dtype)
k = tvm.reduce_axis((0, 3), 'k')
T_g = tvm.compute((num_batch, 2, out_height*out_width),lambda b, y, x: tvm.sum(theta[b, y, k] * grid[b, k, x], axis = k), name = 'T_g')
x_s = tvm.compute((num_batch, 1, out_height*out_width), lambda i,j,k:T_g[i,0,k], name = 'x_s')
y_s = tvm.compute((num_batch, 1, out_height*out_width), lambda i,j,k:T_g[i,1,k], name = 'y_s')
x_s_flat = topi.reshape(x_s, (num_batch*out_height*out_width,))
y_s_flat = topi.reshape(y_s, (num_batch*out_height*out_width,))
input_transformed = _interpolate(input_dim, input_shape, x_s_flat, y_s_flat, out_size, dtype)
output = topi.reshape(input_transformed, [num_batch, out_height, out_width, num_channels])
return output
num_batch = input_shape[0]
input_height = input_shape[1]
input_width = input_shape[2]
channel = input_shape[3]
U = tvm.placeholder((num_batch, input_height, input_width, channel), name="U", dtype=dtype)
theta = tvm.placeholder((num_batch, 6, 1, 1), dtype=dtype)
output = _transform(theta, U, out_shape, input_shape, dtype)
s = tvm.create_schedule(output.op)
if need_print:
with build_config:
print(tvm.lower(s, [U, theta, output], simple_mode=True))
if need_build:
with build_config:
tvm.build(s, [U, theta, output], "cce", name=kernel_name)