quic-sanirudh commented on code in PR #12204: URL: https://github.com/apache/tvm/pull/12204#discussion_r951795171
########## src/runtime/hexagon/ops/conv_utils.cc: ########## @@ -0,0 +1,243 @@ +/* + * 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. + */ + +#include "tvm/runtime/hexagon/ops/conv2d.h" + +namespace tvm { +namespace runtime { +namespace hexagon { + +/** + * @brief Function to "blockize" the flat input data + * The term "blockize" is used to mention that the data is stored in non-contiguous blocks + * + * The input is mapped into the below mentioned layout (notation similar to index map used for + * transform layout): + * + * lambda n, h, w, c: n, h//8, w//4, c//32, AXIS_SEPARATOR, h%8, (w%4)//2, c%32, w%2 + * + * where AXIS_SEPARATOR represents split up in the physical layout + * + * @param out Pre-allocated output memory pointer + * @param inp_flat Flat input data pointer + * @param height + * @param width + * @param depth + */ +void blockize_hwc_16b(void* out, void* inp_flat, int height, int width, int depth) { + auto inp_data = static_cast<uint16_t*>(inp_flat); + auto out_data = static_cast<uintptr_t*>(out); + const int stride_x = depth; + const int stride_y = stride_x * width; + + for (int cy = 0; cy < height; cy += 8) { + for (int cx = 0; cx < width; cx += 4) { + for (int cc = 0; cc < depth; cc += 32) { + auto block = reinterpret_cast<uint16_t*>(*out_data++); + int max_y = std::min(8, height - cy); + int max_x = std::min(4, width - cx); + int max_c = std::min(32, depth - cc); + for (int y = 0; y < max_y; ++y) { + for (int x = 0; x < max_x; ++x) { + for (int c = 0; c < max_c; ++c) { + block[xyc_to_sm_16b(y, x, c)] = + inp_data[(cy + y) * stride_y + (cx + x) * stride_x + (cc + c)]; + } + for (int c = max_c; c < 32; ++c) block[xyc_to_sm_16b(y, x, c)] = 0; + } + for (int x = max_x; x < 4; ++x) { + for (int c = 0; c < 32; ++c) block[xyc_to_sm_16b(y, x, c)] = 0; + } + } + + for (int y = max_y; y < 8; ++y) + for (int x = 0; x < 4; ++x) + for (int c = 0; c < 32; ++c) block[xyc_to_sm_16b(y, x, c)] = 0; + } // cc + } // cx + } // cy +} + +/** + * @brief Convert back from non-contguous layout to a flat layout + * + * @param out_float Pre-allocated output memory pointer + * @param inp Blockized input data pointer + * @param height + * @param width + * @param depth + */ +void deblockize_hwc_16b(void* out_flat, void* inp, int height, int width, int depth) { + uintptr_t* inp_data = static_cast<uintptr_t*>(inp); + uint16_t* out_data = static_cast<uint16_t*>(out_flat); + const int stride_x = depth; + const int stride_y = stride_x * width; + + for (int cy = 0; cy < height; cy += 8) { + for (int cx = 0; cx < width; cx += 4) { + for (int cc = 0; cc < depth; cc += 32) { + auto block = reinterpret_cast<uint16_t*>(*inp_data); + int max_y = std::min(8, height - cy); + int max_x = std::min(4, width - cx); + int max_c = std::min(32, depth - cc); + for (int y = 0; y < max_y; ++y) { + for (int x = 0; x < max_x; ++x) { + for (int c = 0; c < max_c; ++c) { + out_data[(cy + y) * stride_y + (cx + x) * stride_x + (cc + c)] = + block[xyc_to_sm_16b(y, x, c)]; + } + } + } + + inp_data++; + } + } + } +} + +/** + * @brief Convert the layout of weights from flat to "chunked". The term chunked is explained below: + * + * Weights are packed into the below mentioned layout (notation similar to index map): + * Since weights cannot be exactly represented into a index map notation, the + * base split up is mentioned below with a few gotchas + * + * lambda h, w, i, o: h//8, w//4, o//32, i//32, h%8, w%4, (i%32)//2, o%32, i%2 + * + * The gotchas are: + * - (w%4) is actually stored in the right to left order, as in 3,2,1,0 instead of 0,1,2,3 + * - The h%8 and (w%4) dimensions are not padded up, leading to chunks of different sizes + * (thereby the name "chunked" instead of packed) + * - The thinnest chunk of width is stored first. For example, if a kernel is 5x5, the first + * chunk along the width has size 1 (representing index 0) and then next one has size 4 + * representing indices (1,2,3,4) Review Comment: I've moved it to the conv2d.h file as well, and I'll create a separate PR with the documentation, thanks for the suggestions. -- 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: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
