guberti commented on code in PR #12969: URL: https://github.com/apache/tvm/pull/12969#discussion_r988317552
########## python/tvm/topi/arm_cpu/mprofile/dsp/micro_kernel/tensordot.py: ########## @@ -0,0 +1,136 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""Computes a "jumpy tensordot" operator, which can be used to tensorize many common operators +including regular conv2d, depthwise conv2d, and grouped conv2d provided the data and kernel layouts +are the optimal ones. When groups=1, the optimal data layout is NHWC and kernel layout is OHWI. When +this is a depthwise convolution, the optimal data layout is NCHW and kernel layout is OIHW.""" + +import textwrap + +from tvm import te, tir + +from .common import num_simd_lanes_per_word + + +def _get_func_name(in_dtype, tensor_h, jump, tensor_w, suffix): + """Gets the C function name of the tensordot function.""" + return f"tensordot_{in_dtype}_h{tensor_h}_j{jump}_w{tensor_w}_{suffix}" + + +def make_intrin_tensordot(slices, strides, tensordot_params): + """Helper function for constructing tensordot intrinsic. We can't construct the whole thing here + (as multiple schedules use tensordot and each must build the intrinstic differently) but we can + build part here to simplify the code.""" + + # in_dtype, tensor_h, jump, tensor_w, suffix = tensordot_params + data, kernel, output = slices + data_strides, kernel_strides = strides + + data_buf = tir.decl_buffer( + data.shape, data.dtype, name="data", offset_factor=1, strides=data_strides + ) + kernel_buf = tir.decl_buffer( + kernel.shape, + kernel.dtype, + name="kernel", + offset_factor=1, + strides=kernel_strides, + ) + output_buf = tir.decl_buffer( + output.shape, output.dtype, name="output", offset_factor=1, strides=[1] + ) + + def intrin_func(ins, outs): + builder = tir.ir_builder.create() + builder.emit( + tir.call_extern( + "int32", + _get_func_name(*tensordot_params), + outs[0].access_ptr("w"), + ins[0].access_ptr("r"), + ins[1].access_ptr("r"), + ) + ) + return builder.get() + + return te.decl_tensor_intrin( + output.op, + intrin_func, + binds={data: data_buf, kernel: kernel_buf, output: output_buf}, + ) + + +def tensordot_impl(in_dtype, tensor_h, jump, tensor_w, suffix): + """Generates C code for tensordot. The int8 and int16 versions have Arm v7e-m DSP assembly.""" Review Comment: Wrote a detailed docstring and some pseudocode. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: commits-unsubscr...@tvm.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org
