def linear(op: Linear, memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
w = op.inputs["w"]
y = op.outputs["y"]
assert x.order == OrderNC or x.order == OrderNHWC
assert w.order == OrderCN or w.order == OrderHWCN
assert y.order == OrderNC or y.order == OrderNHWC
assert w.ndim == x.ndim
buffer_injector = BufferInjector()
buffer_injector.register({
"linear_X": memory_layout[x],
"linear_Y": memory_layout[y],
"linear_W": memory_layout[w],
"linear_M": y.shape_dict[Axis.N],
"linear_N": y.size // y.shape_dict[Axis.N],
"linear_K": x.size // x.shape_dict[Axis.N],
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(
{name_injector.name: source},
name_injector.name,
buffer_injector.buffer,
buffer_injector.unresolved_value_list
)
return [kernel]
def sum_handler(op: Sum, memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
axis = op.parameters["axis"]
buffer_injector = BufferInjector()
buffer_injector.register({
"sum_X": memory_layout[x],
"sum_Y": memory_layout[y],
"sum_y_stride": y.stride,
"sum_y_shape": y.shape,
"sum_x_stride": [x.stride_dict[a] for a in y.order.axes],
"sum_D": y.ndim,
"sum_N": x.shape_dict[axis],
"sum_MAX_GID": y.size,
"sum_x_target_axis_stride": x.stride_dict[axis]
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(
{name_injector.name: source},
name_injector.name,
GPUSize(8, 1, 1),
GPUSize(MAX_THREADS_PER_THREADGROUP, 1, 1),
buffer_injector.buffer,
buffer_injector.unresolved_value_list
)
return [kernel]
def arg_min_handler(op: ArgMin, memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
axis = op.parameters["axis"]
buffer_injector = BufferInjector()
buffer_injector.register({
"argmin_X": memory_layout[x],
"argmin_Y": memory_layout[y],
"argmin_y_stride": y.stride,
"argmin_y_shape": y.shape,
"argmin_x_stride": [x.stride_dict[a] for a in y.order.axes],
"argmin_D": y.ndim,
"argmin_N": x.shape_dict[axis],
"argmin_MAX_GID": y.size,
"argmin_x_target_axis_stride": x.stride_dict[axis]
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(
{name_injector.name: source},
name_injector.name,
buffer_injector.buffer,
buffer_injector.unresolved_value_list
)
return [kernel]
def softmax_same_order(op: Softmax,
memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
target_axis = op.parameters["axis"]
target_axis_index = x.order.axes_dict[target_axis]
D1 = mul(x.shape[:target_axis_index])
D2 = x.shape[target_axis_index]
D3 = mul(x.shape[target_axis_index + 1:])
buffer_injector = BufferInjector()
buffer_injector.register({
"softmax_X": memory_layout[x],
"softmax_Y": memory_layout[y],
"softmax_D1": D1,
"softmax_D2": D2,
"softmax_D3": D3
})
name_injector = KernelNameInjector(op)
source = template_same_order
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def reshape(op: Reshape, memory_layout: MemoryLayout) -> List[Kernel]:
# Operation without need for transposition is currently supported
x = op.inputs["x"]
y = op.outputs["y"]
if memory_layout[x] == memory_layout[y]:
# This is inplace operation
return []
assert x.order == op.parameters["in_order"]
assert y.order == op.parameters["out_order"]
assert y.size == mul(op.parameters["out_shape"])
buffer_injector = BufferInjector()
buffer_injector.register({
"reshape_x": memory_layout[x],
"reshape_y": memory_layout[y],
"reshape_N": y.size,
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def depth2space(op: Depth2Space, memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
r = op.parameters['r']
assert x.order == OrderNHWC
assert y.order == OrderNHWC
buffer_injector = BufferInjector()
buffer_injector.register({
"depth2space_x": memory_layout[x],
"depth2space_y": memory_layout[y],
'depth2space_r': r,
"depth2space_N": x.shape_dict[Axis.N],
"depth2space_C1": x.shape_dict[Axis.C],
"depth2space_C2": y.shape_dict[Axis.C],
"depth2space_H1": x.shape_dict[Axis.H],
"depth2space_H2": y.shape_dict[Axis.H],
"depth2space_W1": x.shape_dict[Axis.W],
"depth2space_W2": y.shape_dict[Axis.W],
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def space2depth(op: Space2Depth, memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
r = op.parameters['r']
assert x.order == OrderNHWC
assert y.order == OrderNHWC
buffer_injector = BufferInjector()
buffer_injector.register({
"space2depth_x": memory_layout[x],
"space2depth_y": memory_layout[y],
'space2depth_r': r,
"space2depth_N": x.shape_dict[Axis.N],
"space2depth_C1": x.shape_dict[Axis.C],
"space2depth_C2": y.shape_dict[Axis.C],
"space2depth_H1": x.shape_dict[Axis.H],
"space2depth_H2": y.shape_dict[Axis.H],
"space2depth_W1": x.shape_dict[Axis.W],
"space2depth_W2": y.shape_dict[Axis.W],
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def reinterpret_axis(op: ReinterpretAxis,
memory_layout: MemoryLayout) -> List[Kernel]:
# Operation without need for transposition is currently supported
x = memory_layout[op.inputs["x"]]
y = memory_layout[op.outputs["y"]]
assert x.variable.order == op.parameters["in_order"]
assert y.variable.order == op.parameters["out_order"]
buffer_injector = BufferInjector()
buffer_injector.register({
"reinterpret_axis_x": x,
"reinterpret_axis_y": y,
"reinterpret_axis_N": y.variable.size,
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def axiswise_bias_same_order(op: AxiswiseBias,
memory_layout: MemoryLayout) -> List[Kernel]:
x = memory_layout[op.inputs["x"]]
b = memory_layout[op.inputs["b"]]
y = memory_layout[op.outputs["y"]]
target_axis_index = x.variable.order.axes_dict[op.axis]
D1 = mul(x.variable.shape[:target_axis_index])
D2 = x.variable.shape[target_axis_index]
D3 = mul(x.variable.shape[target_axis_index + 1:])
buffer_injector = BufferInjector()
buffer_injector.register({
"axiswise_bias_X": x,
"axiswise_bias_B": b,
"axiswise_bias_Y": y,
"axiswise_bias_D1": D1,
"axiswise_bias_D2": D2,
"axiswise_bias_D3": D3
})
name_injector = KernelNameInjector(op)
source = generate_template_same_order(D1, D3)
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def sgemm(op: Sgemm, memory_layout: MemoryLayout) -> List[Kernel]:
A = memory_layout[op.inputs["A"]]
B = memory_layout[op.inputs["B"]]
C = memory_layout[op.outputs["C"]]
buffer_injector = BufferInjector()
buffer_injector.register({
"sgemm_A": A,
"sgemm_B": B,
"sgemm_C": C,
"sgemm_M": op.M,
"sgemm_N": op.N,
"sgemm_K": op.K
})
name_injector = KernelNameInjector(op)
# transpose_X assumes fortran-order data. True means X is C-order, False means Fortran-order.
# In default convolution, transpose_A == transpose_B == True.
# The order of output matrix C is C-order.
source = generate_template_64(op.transpose_A, op.transpose_B, op.M, op.N,
op.K)
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize((op.M + 64 - 1) // 64, (op.N + 64 - 1) // 64, 1),
GPUSize(64, 1, 1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def tile(op: Tile, memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
buffer_injector = BufferInjector()
buffer_injector.register({
"tile_x":
memory_layout[x],
"tile_y":
memory_layout[y],
"tile_y_stride":
y.stride,
"tile_x_stride": [x.stride_dict[a] for a in y.order.axes],
"tile_x_shape": [x.shape_dict[a] for a in y.order.axes],
"tile_D":
x.ndim,
"tile_MAX_GID":
y.size,
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def space2depth(op: Space2Depth, memory_layout: MemoryLayout) -> List[Kernel]:
x = memory_layout[op.inputs["x"]]
y = memory_layout[op.outputs["y"]]
r = op.parameters['r']
assert x.variable.order == OrderNHWC
assert y.variable.order == OrderNHWC
buffer_injector = BufferInjector()
buffer_injector.register({
"space2depth_x": x,
"space2depth_y": y,
'space2depth_r': r,
"space2depth_N": x.variable.shape_dict[Axis.N],
"space2depth_C1": x.variable.shape_dict[Axis.C],
"space2depth_C2": y.variable.shape_dict[Axis.C],
"space2depth_H1": x.variable.shape_dict[Axis.H],
"space2depth_H2": y.variable.shape_dict[Axis.H],
"space2depth_W1": x.variable.shape_dict[Axis.W],
"space2depth_W2": y.variable.shape_dict[Axis.W],
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def embedding(op: Embedding, memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
w = op.inputs["w"]
y = op.outputs["y"]
assert x.order == OrderNT
assert w.order == OrderCN
assert y.order == OrderNTC
buffer_injector = BufferInjector()
buffer_injector.register({
"embedding_X": memory_layout[x],
"embedding_Y": memory_layout[y],
"embedding_W": memory_layout[w],
"embedding_vocabulary": w.shape_dict[Axis.C],
"embedding_sequence_len": x.shape_dict[Axis.T],
"embedding_batch_size": x.shape_dict[Axis.N],
"embedding_dim": w.shape_dict[Axis.N]
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def flatten(op: Flatten,
memory_layout: MemoryLayout) -> List[Kernel]:
x = memory_layout[op.inputs["x"]]
y = memory_layout[op.outputs["y"]]
# assert x.variable.order == y.variable.order
buffer_injector = BufferInjector()
buffer_injector.register({
"flatten_x": x,
"flatten_y": y,
"flatten_N": y.variable.size,
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(
{name_injector.name: source},
name_injector.name,
GPUSize(8, 1, 1),
GPUSize(MAX_THREADS_PER_THREADGROUP, 1, 1),
buffer_injector.buffer,
buffer_injector.unresolved_value_list
)
return [kernel]
def reshape(op: Reshape, memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
if memory_layout[x].offset == memory_layout[y].offset:
# Inplace
return []
assert x.order == op.parameters["in_order"]
assert y.order == op.parameters["out_order"]
assert y.size == mul(op.parameters["out_shape"])
buffer_injector = BufferInjector()
buffer_injector.register({
"reshape_x": memory_layout[x],
"reshape_y": memory_layout[y],
"reshape_N": y.size,
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def reshape(op: Reshape, memory_layout: MemoryLayout) -> List[Kernel]:
x = memory_layout[op.inputs["x"]]
y = memory_layout[op.outputs["y"]]
assert x.variable.order == op.parameters["in_order"]
assert y.variable.order == op.parameters["out_order"]
assert y.variable.size == mul(op.parameters["out_shape"])
buffer_injector = BufferInjector()
buffer_injector.register({
"reshape_x": x,
"reshape_y": y,
"reshape_N": y.variable.size,
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(1024, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def softmax(op: Softmax, memory_layout: MemoryLayout) -> List[Kernel]:
x = memory_layout[op.inputs["x"]]
y = memory_layout[op.outputs["y"]]
assert y.variable.order == x.variable.order
assert y.variable.shape == x.variable.shape
axis = op.parameters["axis"]
assert axis == x.variable.order.axes[
-1], "[Webassembly] Softmax supports only for aggregating last axis."
buffer_injector = BufferInjector()
buffer_injector.register({
"softmax_X":
x,
"softmax_Y":
y,
"softmax_N":
y.variable.size // y.variable.shape_dict[axis],
"softmax_C":
y.variable.shape_dict[axis],
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def embedding(op: Embedding, memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
w = op.inputs["w"]
y = op.outputs["y"]
assert x.order == OrderNT
assert w.order == OrderCN
assert y.order == OrderNTC
buffer_injector = BufferInjector()
buffer_injector.register({
"embedding_X": memory_layout[x],
"embedding_Y": memory_layout[y],
"embedding_W": memory_layout[w],
"embedding_T": x.shape_dict[Axis.T],
"embedding_N": x.shape_dict[Axis.N],
"embedding_C": w.shape_dict[Axis.N]
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def axiswise_scale_same_order(op: AxiswiseScale,
memory_layout: MemoryLayout) -> List[Kernel]:
x = memory_layout[op.inputs["x"]]
s = memory_layout[op.inputs["s"]]
y = memory_layout[op.outputs["y"]]
target_axis_index = x.variable.order.axes_dict[op.axis]
D1 = mul(x.variable.shape[:target_axis_index])
D2 = x.variable.shape[target_axis_index]
D3 = mul(x.variable.shape[target_axis_index + 1:])
buffer_injector = BufferInjector()
buffer_injector.register({
"axiswise_scale_X": x,
"axiswise_scale_S": s,
"axiswise_scale_Y": y,
"axiswise_scale_D1": D1,
"axiswise_scale_D2": D2,
"axiswise_scale_D3": D3
})
name_injector = KernelNameInjector(op)
source = template_same_order
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def reinterpret_axis(op: ReinterpretAxis,
memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
if memory_layout[x] == memory_layout[y]:
# This is inplace operation
return []
assert x.order == op.parameters["in_order"]
assert y.order == op.parameters["out_order"]
buffer_injector = BufferInjector()
buffer_injector.register({
"reinterpret_axis_x": memory_layout[x],
"reinterpret_axis_y": memory_layout[y],
"reinterpret_axis_N": y.size,
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def axiswise_scale_general(op: AxiswiseScale,
memory_layout: MemoryLayout) -> List[Kernel]:
x = memory_layout[op.inputs["x"]]
s = memory_layout[op.inputs["s"]]
y = memory_layout[op.outputs["y"]]
x_shape = x.variable.shape
target_axis_index = x.variable.order.axes_dict[op.axis]
D1 = mul(x_shape[:target_axis_index])
D2 = x_shape[target_axis_index]
D3 = mul(x_shape[target_axis_index + 1:])
y_strides = []
stride = 1
for sh in reversed(y.variable.shape):
y_strides.insert(0, stride)
stride *= sh
x_stride_in_y = [
y_strides[y.variable.order.axes_dict[axis]]
for axis in x.variable.order.axes
]
buffer_injector = BufferInjector()
buffer_injector.register({
"axiswise_scale_X":
x,
"axiswise_scale_S":
s,
"axiswise_scale_Y":
y,
"axiswise_scale_D1":
D1,
"axiswise_scale_D2":
D2,
"axiswise_scale_D3":
D3,
"axiswise_scale_D":
x.variable.ndim,
"axiswise_scale_d_target":
x.variable.order.axes_dict[op.axis],
"axiswise_scale_x_shape":
x_shape,
"axiswise_scale_x_stride_in_y":
x_stride_in_y,
})
name_injector = KernelNameInjector(op)
source = template_general
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def local_response_normalization_general(
op: LocalResponseNormalization,
memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
target_axis = Axis.C
x_shape = x.shape
y_strides = []
stride = 1
for s in reversed(y.shape):
y_strides.insert(0, stride)
stride *= s
x_stride_in_y = [
y_strides[y.order.axes_dict[axis]] for axis in x.order.axes
]
buffer_injector = BufferInjector()
buffer_injector.register({
"local_response_normalization_X":
memory_layout[x],
"local_response_normalization_Y":
memory_layout[y],
"local_response_normalization_D":
x.ndim,
"local_response_normalization_d_target":
x.order.axes_dict[target_axis],
"local_response_normalization_x_shape":
x_shape,
"local_response_normalization_x_stride_in_y":
x_stride_in_y,
"local_response_normalization_param_half_n":
int(op.parameters["n"] // 2),
"local_response_normalization_param_k":
float(op.parameters["k"]),
"local_response_normalization_param_alpha":
float(op.parameters["alpha"]),
"local_response_normalization_param_minus_beta":
float(-op.parameters["beta"])
})
name_injector = KernelNameInjector(op)
source = template_general
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def average_pooling_2d(op: AveragePooling2D,
memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
assert x.order == OrderNHWC
assert y.order == OrderNHWC
buffer_injector = BufferInjector()
buffer_injector.register({
"average_pooling_2d_X":
memory_layout[x],
"average_pooling_2d_Y":
memory_layout[y],
"average_pooling_2d_N":
x.shape_dict[Axis.N],
"average_pooling_2d_H1":
x.shape_dict[Axis.H],
"average_pooling_2d_W1":
x.shape_dict[Axis.W],
"average_pooling_2d_C":
x.shape_dict[Axis.C],
"average_pooling_2d_H2":
y.shape_dict[Axis.H],
"average_pooling_2d_W2":
y.shape_dict[Axis.W],
"average_pooling_2d_KH":
op.parameters["ksize"][0],
"average_pooling_2d_KW":
op.parameters["ksize"][1],
"average_pooling_2d_SH":
op.parameters["stride"][0],
"average_pooling_2d_SW":
op.parameters["stride"][1],
"average_pooling_2d_PH":
op.parameters["padding"][0],
"average_pooling_2d_PW":
op.parameters["padding"][1],
})
name_injector = KernelNameInjector(op)
source = template
for key, statement in statement_divide_without_padding[
op.parameters["divide_without_padding"]].items():
source = source.replace(key, statement)
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def tensordot(op: Tensordot, memory_layout: MemoryLayout) -> List[Kernel]:
A = op.inputs["A"]
B = op.inputs["B"]
C = op.outputs["C"]
axes = op.axes
# Reduced axes must be located on inside of input variables.
assert A.order.axes[-len(axes[0]):] == axes[0]
assert B.order.axes[-len(axes[1]):] == axes[1]
# output variable's axes order must be as [*a_remained_axes, *b_remained_axes]
assert C.order.axes[:A.ndim - len(axes[0])] == A.order.axes[:-len(axes[0])]
assert C.order.axes[-(B.ndim -
len(axes[1])):] == B.order.axes[:-len(axes[1])]
assert C.ndim == A.ndim - len(axes[0]) + B.ndim - len(axes[1])
K = mul(A.shape_dict[a] for a in axes[0])
M = A.size // K
N = B.size // K
buffer_injector = BufferInjector()
buffer_injector.register({
"sgemm_A": memory_layout[A],
"sgemm_B": memory_layout[B],
"sgemm_C": memory_layout[C],
"sgemm_M": M,
"sgemm_N": N,
"sgemm_K": K
})
if op.has_attribute(UseEigenAttribute):
source = generate_template_eigen(True, False)
buffer_injector.register({
"sgemm_A": memory_layout[A],
"sgemm_B": memory_layout[B],
"sgemm_C": memory_layout[C]
})
else:
source = generate_template(True, False)
buffer_injector.register({
"sgemm_A": memory_layout[A],
"sgemm_B": memory_layout[B],
"sgemm_C": memory_layout[C],
"sgemm_M": op.M,
"sgemm_N": op.N,
"sgemm_K": op.K
})
name_injector = KernelNameInjector(op)
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def concat(op: Concat, memory_layout: MemoryLayout) -> List[Kernel]:
xs = [
memory_layout[op.inputs[f"x{str(i)}"]] for i in range(len(op.inputs))
]
y = memory_layout[op.outputs["y"]]
target_axis = op.axis
x_offsets = [x.offset for x in xs]
x_shapes = [x.variable.shape for x in xs]
y_strides = []
stride = 1
for s in reversed(y.variable.shape):
y_strides.insert(0, stride)
stride *= s
# x_strides[i][j] is stride size of xs[i].order.axes[j] in y
x_strides_in_y = [[] for _ in xs]
for x, strides in zip(xs, x_strides_in_y):
for axis in x.variable.order.axes:
strides.append(y_strides[y.variable.order.axes_dict[axis]])
# x_offsets[i] is memory offset of xs[i]'s data in y.
y_offsets = []
target_axis_offset = 0
for x in xs:
y_offsets.append(target_axis_offset *
y_strides[y.variable.order.axes_dict[target_axis]])
target_axis_offset += x.variable.shape_dict[target_axis]
buffer_injector = BufferInjector()
buffer_injector.register({
"concat_y": y,
"concat_D": len(y.variable.shape),
"concat_N": len(xs),
"concat_xs": xs,
"concat_x_strides_in_y": x_strides_in_y,
"concat_x_shapes": x_shapes,
"concat_y_offsets": y_offsets
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def average_pooling_2d(op: AveragePooling2D,
memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
assert x.order == OrderNHWC
assert y.order == OrderNHWC
buffer_injector = BufferInjector()
buffer_injector.register({
"average_pooling_2d_X":
memory_layout[x],
"average_pooling_2d_Y":
memory_layout[y],
"average_pooling_2d_N":
x.shape_dict[Axis.N],
"average_pooling_2d_H1":
x.shape_dict[Axis.H],
"average_pooling_2d_W1":
x.shape_dict[Axis.W],
"average_pooling_2d_C":
x.shape_dict[Axis.C],
"average_pooling_2d_H2":
y.shape_dict[Axis.H],
"average_pooling_2d_W2":
y.shape_dict[Axis.W],
"average_pooling_2d_KH":
op.parameters["ksize"][0],
"average_pooling_2d_KW":
op.parameters["ksize"][1],
"average_pooling_2d_SH":
op.parameters["stride"][0],
"average_pooling_2d_SW":
op.parameters["stride"][1],
"average_pooling_2d_PH":
op.parameters["padding"][0],
"average_pooling_2d_PW":
op.parameters["padding"][1],
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def im2col(op: Im2Col, memory_layout: MemoryLayout) -> List[Kernel]:
im = op.inputs["im"]
col = op.outputs["col"]
assert im.order == OrderNHWC
assert col.order == OrderNHWC or col.order == OrderCNHW
N = im.shape_dict[Axis.N]
C1 = im.shape_dict[Axis.C]
H1 = im.shape_dict[Axis.H]
W1 = im.shape_dict[Axis.W]
H1P = H1 + 2 * op.PH
W1P = W1 + 2 * op.PW
buffer_injector = BufferInjector()
buffer_injector.register({
"im2col_im": memory_layout[im],
"im2col_col": memory_layout[col],
"im2col_N": N,
"im2col_C1": C1,
"im2col_H1": im.shape_dict[Axis.H],
"im2col_W1": im.shape_dict[Axis.W],
"im2col_H2": col.shape_dict[Axis.H],
"im2col_W2": col.shape_dict[Axis.W],
"im2col_KH": op.KH,
"im2col_KW": op.KW,
"im2col_DH": op.DH,
"im2col_DW": op.DW,
"im2col_SH": op.SH,
"im2col_SW": op.SW,
"im2col_PH": op.PH,
"im2col_PW": op.PW,
})
name_injector = KernelNameInjector(op)
source = template_CNHW if col.order == OrderCNHW else generate_template_NHWC(op.SH, op.SW, op.DH, op.DW, C1)
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel(
{name_injector.name: source},
name_injector.name,
GPUSize(N * H1P * W1P, 1, 1),
GPUSize(64, 1, 1),
buffer_injector.buffer,
buffer_injector.unresolved_value_list
)
return [kernel]
def slice_handler(op: Slice, memory_layout: MemoryLayout) -> List[Kernel]:
x = op.inputs["x"]
y = op.outputs["y"]
remained_axes_in_y_order = [a for a in y.order.axes if a in x.order.axes]
removed_axes = [a for a in x.order.axes if a not in y.order.axes]
x_index_offset = 0
x_strides = []
for axis in remained_axes_in_y_order:
assert isinstance(op.indices[axis], slice)
index = normalize_slice(op.indices[axis], x.shape_dict[axis])
x_index_offset += x.stride_dict[axis] * index.start
x_strides.append(x.stride_dict[axis] * index.step)
for axis in removed_axes:
assert isinstance(op.indices[axis], int)
x_index_offset += x.stride_dict[axis] * op.indices[axis]
buffer_injector = BufferInjector()
buffer_injector.register({
"slice_ndim":
len(remained_axes_in_y_order),
"slice_X":
memory_layout[x],
"slice_x_stride_in_y_order":
x_strides,
"slice_x_index_offset":
x_index_offset,
"slice_Y":
memory_layout[y],
"slice_y_size":
y.size,
"slice_y_shape": [y.shape_dict[a] for a in remained_axes_in_y_order],
"slice_y_stride": [y.stride_dict[a] for a in remained_axes_in_y_order]
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def axiswise_bias_general(op: AxiswiseBias,
memory_layout: MemoryLayout) -> List[Kernel]:
x = memory_layout[op.inputs["x"]]
b = memory_layout[op.inputs["b"]]
y = memory_layout[op.outputs["y"]]
x_shape = x.variable.shape
y_strides = []
stride = 1
for s in reversed(y.variable.shape):
y_strides.insert(0, stride)
stride *= s
x_stride_in_y = [
y_strides[y.variable.order.axes_dict[axis]]
for axis in x.variable.order.axes
]
buffer_injector = BufferInjector()
buffer_injector.register({
"axiswise_bias_X":
x,
"axiswise_bias_B":
b,
"axiswise_bias_Y":
y,
"axiswise_bias_D":
x.variable.ndim,
"axiswise_bias_d_target":
x.variable.order.axes_dict[op.axis],
"axiswise_bias_x_shape":
x_shape,
"axiswise_bias_x_stride_in_y":
x_stride_in_y,
})
name_injector = KernelNameInjector(op)
source = template_general
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
def split_axis(op: SplitAxis, memory_layout: MemoryLayout) -> List[Kernel]:
x = memory_layout[op.inputs["x"]]
ys = [
memory_layout[op.outputs[f"y{str(i)}"]] for i in range(len(op.outputs))
]
target_axis = op.parameters["axis"]
y_shapes = [y.variable.shape for y in ys]
# y_strides[i][j] is stride size of ys[i].order.axes[j] in x
y_strides_in_x = [[
x.variable.stride_dict[axis] for axis in y.variable.order.axes
] for y in ys]
# x_offsets[i] is memory offset of ys[i]'s data in x.
x_offsets = []
target_axis_offset = 0
for y in ys:
x_offsets.append(
target_axis_offset *
x.variable.stride[x.variable.order.axes_dict[target_axis]])
target_axis_offset += y.variable.shape_dict[target_axis]
buffer_injector = BufferInjector()
buffer_injector.register({
"split_axis_x": x,
"split_axis_D": len(x.variable.shape),
"split_axis_N": len(ys),
"split_axis_ys": ys,
"split_axis_y_strides_in_x": y_strides_in_x,
"split_axis_y_shapes": y_shapes,
"split_axis_x_offsets": x_offsets
})
name_injector = KernelNameInjector(op)
source = template
source = buffer_injector.inject(source)
source = name_injector.inject(source)
kernel = Kernel({name_injector.name: source}, name_injector.name,
GPUSize(8, 1, 1), GPUSize(MAX_THREADS_PER_THREADGROUP, 1,
1), buffer_injector.buffer,
buffer_injector.unresolved_value_list)
return [kernel]
