areusch commented on a change in pull request #5417: URL: https://github.com/apache/incubator-tvm/pull/5417#discussion_r416139282
########## File path: topi/python/topi/arm_cpu/cortex_m7/micro_kernel/gemm.py ########## @@ -0,0 +1,221 @@ +# 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. +# pylint: disable=invalid-name, no-value-for-parameter +"""Defines gemm intrinsics for SIMD matrix multiplication.""" + +import random +import string + +import tvm +from tvm import te + +########################## +# MxKxN MatMul Intrinsic # +########################## + +# NOTE this is transposed matmul (A * B^T) +def intrin_gemm_MxKxN(M, K, N, in_dtype, out_dtype): + """Defines a SIMD-accelerated transposed matmul.""" + # we generate a unique ID for every intrinsic definition, to prevent name + # collisions in the generated source (e.g., if there are multiple operators + # in the same module that use the same intrinsic) + # + # TODO(weberlo, areusch): to cut down on memory usage, we should cache each intrinsic + # instantiation and include it only once, eliminating the need for unique + # IDs + UNIQ_ID_LEN = 8 + uniq_id = ''.join(random.choices(string.ascii_uppercase, k=UNIQ_ID_LEN)) + + if isinstance(M, tvm.tir.IntImm): + M = M.value + if isinstance(K, tvm.tir.IntImm): + K = K.value + if isinstance(N, tvm.tir.IntImm): + N = N.value + assert K % 4 == 0 + # TODO(weberlo, areusch): support more dtypes? + assert in_dtype == 'int8' + assert out_dtype == 'int32' + A = te.placeholder((M, K), name='a', dtype=in_dtype) + B = te.placeholder((N, K), name='b', dtype=in_dtype) + k = te.reduce_axis((0, K), name='k') + C = te.compute( + (M, N), + lambda i, j: te.sum(A[i, k].astype(out_dtype) * B[j, k].astype(out_dtype), axis=k), + name='c') + A_buf = tvm.tir.decl_buffer( + A.shape, A.dtype, + name="A", + offset_factor=1, + strides=[te.var("A_s"), 1]) + B_buf = tvm.tir.decl_buffer( + B.shape, B.dtype, + name="B", + offset_factor=1, + strides=[te.var("B_s"), 1]) + C_buf = tvm.tir.decl_buffer( + C.shape, C.dtype, + name="C", + offset_factor=1, + strides=[te.var("C_s"), 1]) + def intrin_func(ins, outs): + aa, bb = ins + cc = outs[0] + def _reduce_update(): + ib = tvm.tir.ir_builder.create() + ib.emit(tvm.tir.call_extern("int32", f"gemm_{M}x{K}x{N}_update_{uniq_id}", + aa.access_ptr("r"), + bb.access_ptr("r"), + cc.access_ptr("w"), + aa.strides[0], + bb.strides[0], + cc.strides[0])) + return ib.get() + def _reduce_reset(): + ib = tvm.tir.ir_builder.create() + ib.emit(tvm.tir.call_extern("int32", f"gemm_{M}x{K}x{N}_reset_{uniq_id}", + cc.access_ptr("w"), + cc.strides[0])) + return ib.get() + def _body(): + ib = tvm.tir.ir_builder.create() + # # NOTE we need the reset in the body for cases where the buffer Review comment: done ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: us...@infra.apache.org
