def scalar_affine(op: ScalarAffine, constants_layout: MemoryLayout, variables_layout: MemoryLayout) -> List[Kernel]: x = variables_layout[op.inputs["x"]] y = variables_layout[op.outputs["y"]] assert x.variable.shape == y.variable.shape meta_injector = MetaInjector() meta_injector.register({ "affine_transform_X_offset": x.offset, "affine_transform_Y_offset": y.offset, "affine_transform_N": y.variable.size, "affine_transform_scale": float(op.scale), "affine_transform_bias": float(op.bias) }) name_injector = KernelNameInjector(op) source = template source = meta_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), meta_injector.buffer) 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 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 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 tanh(op: Tanh, constants_layout: MemoryLayout, variables_layout: MemoryLayout) -> List[Kernel]: x = variables_layout[op.inputs["x"]] y = variables_layout[op.outputs["y"]] assert x.variable.shape == y.variable.shape assert x.variable.order == y.variable.order meta_injector = MetaInjector() meta_injector.register({ "tanh_X_offset": x.offset, "tanh_Y_offset": y.offset, "tanh_N": y.variable.size }) name_injector = KernelNameInjector(op) source = template source = meta_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), meta_injector.buffer) return [kernel]
def max_pooling_2d(op: MaxPooling2D, constants_layout: MemoryLayout, variables_layout: MemoryLayout) -> List[Kernel]: x = variables_layout[op.inputs["x"]] y = variables_layout[op.outputs["y"]] assert x.variable.order == OrderNHWC assert y.variable.order == OrderNHWC meta_injector = MetaInjector() meta_injector.register({ "max_pooling_2d_X_offset": x.offset, "max_pooling_2d_Y_offset": y.offset, "max_pooling_2d_N": x.variable.shape_dict[Axis.N], "max_pooling_2d_H1": x.variable.shape_dict[Axis.H], "max_pooling_2d_W1": x.variable.shape_dict[Axis.W], "max_pooling_2d_C": x.variable.shape_dict[Axis.C], "max_pooling_2d_H2": y.variable.shape_dict[Axis.H], "max_pooling_2d_W2": y.variable.shape_dict[Axis.W], "max_pooling_2d_K": op.parameters["ksize"][0], "max_pooling_2d_S": op.parameters["stride"][0], "max_pooling_2d_P": op.parameters["padding"][0], }) name_injector = KernelNameInjector(op) source = template source = meta_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), meta_injector.buffer) 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 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 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 axiswise_bias_same_order(op: AxiswiseBias, constants_layout: MemoryLayout, variables_layout: MemoryLayout) -> List[Kernel]: x = variables_layout[op.inputs["x"]] b = constants_layout[op.inputs["b"]] y = variables_layout[op.outputs["y"]] target_axis_index = x.variable.order.axes_dict[op.axis] D1 = int(np.prod(x.variable.shape[:target_axis_index])) D2 = x.variable.shape[target_axis_index] D3 = int(np.prod(x.variable.shape[target_axis_index + 1:])) meta_injector = MetaInjector() meta_injector.register({ "axiswise_bias_X_offset": x.offset, "axiswise_bias_B_offset": b.offset, "axiswise_bias_Y_offset": y.offset, "axiswise_bias_D1": D1, "axiswise_bias_D2": D2, "axiswise_bias_D3": D3 }) name_injector = KernelNameInjector(op) source = generate_template_same_order(D1, D3) source = meta_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), meta_injector.buffer) 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 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 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 axiswise_scale(op: AxiswiseScale, constants_layout: MemoryLayout, variables_layout: MemoryLayout) -> List[Kernel]: x = variables_layout[op.inputs["x"]] s = constants_layout[op.inputs["s"]] y = variables_layout[op.outputs["y"]] assert x.variable.order == OrderNC or x.variable.order == OrderNHWC or x.variable.order == OrderHWNC assert y.variable.order == OrderNC or y.variable.order == OrderNHWC or y.variable.order == OrderHWNC assert op.parameters[ "axis"] == Axis.C, "[WebGPU] AxiswiseScale supports only channelwise bias." meta_injector = MetaInjector() meta_injector.register({ "axiswise_scale_X_offset": x.offset, "axiswise_scale_Y_offset": y.offset, "axiswise_scale_S_offset": s.offset, "axiswise_scale_N": y.variable.size, "axiswise_scale_C": y.variable.shape_dict[Axis.C], }) name_injector = KernelNameInjector(op) source = template source = meta_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), meta_injector.buffer) 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 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 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 elementwise_sum(op: AxiswiseScale, constants_layout: MemoryLayout, variables_layout: MemoryLayout) -> List[Kernel]: x0 = variables_layout[op.inputs["x0"]] x1 = variables_layout[op.inputs["x1"]] y = variables_layout[op.outputs["y"]] assert len(op.inputs) == 2, "[WebGPU] ElementwiseSum operator currently supported only 2 inputs." assert x0.variable.shape == x1.variable.shape == y.variable.shape meta_injector = MetaInjector() meta_injector.register({ "elementwise_sum_X0_offset": x0.offset, "elementwise_sum_X1_offset": x1.offset, "elementwise_sum_Y_offset": y.offset, "elementwise_sum_N": y.variable.size }) inline_injector = InlineInjector(op) name_injector = KernelNameInjector(op) source = generate_template(y.variable.size, inline_injector.has_inline) source = meta_injector.inject(source) source = inline_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), meta_injector.buffer ) 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 local_response_normalization_general( op: LocalResponseNormalization, constants_layout: MemoryLayout, variables_layout: MemoryLayout) -> List[Kernel]: x = variables_layout[op.inputs["x"]] y = variables_layout[op.outputs["y"]] target_axis = Axis.C 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 ] meta_injector = MetaInjector() meta_injector.register({ "local_response_normalization_X_offset": x.offset, "local_response_normalization_Y_offset": y.offset, "local_response_normalization_D": x.variable.ndim, "local_response_normalization_d_target": x.variable.order.axes_dict[target_axis], "local_response_normalization_x_shape": np.array(x_shape, dtype=np.int32).tobytes(), "local_response_normalization_x_stride_in_y": np.array(x_stride_in_y, dtype=np.int32).tobytes(), "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 = meta_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), meta_injector.buffer) 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 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 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 elementwise_kernel(op: Elementwise, 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"]] item = _registered_items[op.__class__] parameters = {key: fn(op) for key, fn in item.parameters.items()} 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]]) buffer_injector = BufferInjector() buffer_injector.register({ "elementwise_Y": y, "elementwise_D": len(y.variable.shape), "elementwise_N": xs[0].variable.size, "elementwise_Xs": xs, "elementwise_X_strides_in_Y": x_strides_in_y, "elementwise_X_shapes": x_shapes }) buffer_injector.register({ f"elementwise_parameters_{key}": val for key, val in parameters.items() }) name_injector = KernelNameInjector(op) source = _generate_source(xs, y, item.code, parameters) 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 elementwise_kernel_base(op: Elementwise, command_buffer: CommandBuffer, buffer_injector: BufferInjector): name_injector = KernelNameInjector(op) source = encode_command(command_buffer) 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 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 axiswise_bias_general(op: AxiswiseBias, constants_layout: MemoryLayout, variables_layout: MemoryLayout) -> List[Kernel]: x = variables_layout[op.inputs["x"]] b = constants_layout[op.inputs["b"]] y = variables_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 ] meta_injector = MetaInjector() meta_injector.register({ "axiswise_bias_X_offset": x.offset, "axiswise_bias_B_offset": b.offset, "axiswise_bias_Y_offset": y.offset, "axiswise_bias_D": x.variable.ndim, "axiswise_bias_d_target": x.variable.order.axes_dict[op.axis], "axiswise_bias_x_shape": np.array(x_shape, dtype=np.int32).tobytes(), "axiswise_bias_x_stride_in_y": np.array(x_stride_in_y, dtype=np.int32).tobytes(), }) name_injector = KernelNameInjector(op) source = template_general source = meta_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), meta_injector.buffer) 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]
def im2col(op: Im2Col, constants_layout: MemoryLayout, variables_layout: MemoryLayout) -> List[Kernel]: im = variables_layout[op.inputs["im"]] col = variables_layout[op.outputs["col"]] assert im.variable.order == OrderNHWC assert col.variable.order == OrderNHWC or col.variable.order == OrderCNHW N = im.variable.shape_dict[Axis.N] C1 = im.variable.shape_dict[Axis.C] H1 = im.variable.shape_dict[Axis.H] W1 = im.variable.shape_dict[Axis.W] H1P = H1 + 2 * op.PH W1P = W1 + 2 * op.PW meta_injector = MetaInjector() meta_injector.register({ "im2col_im_offset": im.offset, "im2col_col_offset": col.offset, "im2col_N": col.variable.shape_dict[Axis.N], "im2col_C1": C1, "im2col_H1": im.variable.shape_dict[Axis.H], "im2col_W1": im.variable.shape_dict[Axis.W], "im2col_H2": col.variable.shape_dict[Axis.H], "im2col_W2": col.variable.shape_dict[Axis.W], "im2col_KH": op.KH, "im2col_KW": op.KW, "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.variable.order == OrderCNHW else generate_template_NHWC( op.SH, op.SW, C1) source = meta_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), meta_injector.buffer) return [kernel]