| WRITE of size 8 at 0x0065720656c0 thread T0 | |
| #0 /usr/local/google/home/benoitjacob/iree/iree/hal/dylib/dylib_executable.cc:93:21 iree::hal::dylib::DyLibExecutable::Initialize(iree::hal::ExecutableSpec) | |
| #1 /usr/local/google/home/benoitjacob/iree/iree/hal/dylib/dylib_executable.cc:28:3 iree::hal::dylib::DyLibExecutable::Load(iree::hal::ExecutableSpec) | |
| #2 /usr/local/google/home/benoitjacob/iree/iree/hal/dylib/dylib_executable_cache.cc:46:10 iree::hal::dylib::DyLibExecutableCache::PrepareExecutable(iree::hal::ExecutableLayout*, iree::hal::ExecutableCachingMode, iree::hal::ExecutableSpec const&) | |
| #3 /usr/local/google/home/benoitjacob/iree/iree/hal/api.cc:1970:3 iree_hal_executable_cache_prepare_executable | |
| #4 /usr/local/google/home/benoitjacob/iree/iree/modules/hal/hal_module.cc:667:5 iree::hal::(anonymous namespace)::HALModuleState::ExecutableCachePrepare(iree::vm::ref<iree_hal_executable_cache> const&, iree::vm::ref<iree_hal_executable_layout> const&, unsigned int, iree::vm::ref<iree_vm_ro_byte_buffer_t> const&) | |
| #5 |
| #include <sys/mman.h> | |
| #include <sys/time.h> | |
| #include <sys/resource.h> | |
| #include <sys/types.h> | |
| #include <sys/stat.h> | |
| #include <fcntl.h> | |
| #include <unistd.h> | |
| #include <cstdio> | |
| #include <cstdint> | |
| #include <cstdlib> |
This is just a 8bit PRNG. Randomness is as bad as it could possibly get. The period is 256 in the scalar implementation. The implementation is widened 64x for SIMD purposes, and that has the unintentional side benefit of multiplying the period by that amount too.
Sample output on Pixel4 (pinned on the biggest CPU core by taskset 80):
To scribble LDR qN, [xN, offset] into MOVI vN.16b, 0, do:
hack_arm_binary some_aarch64_binary_file neon-load-to-movi-zero
| ���� JFIF �� C | |
| $.' ",# (7),01444 '9=82<.342�� C 2! !22222222222222222222222222222222222222222222222222�� �'" �� | |
| �� � } !1AQa"q2���#B��R��$3br� | |
| %&'()*456789:CDEFGHIJSTUVWXYZcdefghijstuvwxyz��������������������������������������������������������������������������� | |
| �� � w !1AQaq"2�B���� #3R�br� | |
| $4�%�&'()*56789:CDEFGHIJSTUVWXYZcdefghijstuvwxyz�������������������������������������������������������������������������� ? ��k�5��r�~ꎭ\�|I����G��?¨�Hv�}�y�|���ɮ���|�_�����6M����_5�� `u w<n���vF0���K�U���� ���(� ��k� @����� | |
| ������}͇��V���� �i$���o�� ~Q��皭�h�l> �WՖ���w/migd��@���+`e� �sҧ�������eZ� �:o��?�YV�������K�H/��m���8�E���B��#pLeW2�@!A$��4ۏ�?��Ƨk���#�p���e;xϖF0U��9�{(�7�eZ� �:o��?�YV�������GJ�������^Z���H�<�B7�b��Hޮ?��t [��l�����-�&P'��� 89-�s�����c�� ��k� @����� | |
| ?�eZ� |
| Some images for IREE Tracy docs |
| 0000000000003980 T ruy::Kernel8bitNeonDotprod(ruy::KernelParams8bit<8, 8> const&) | |
| 0000000000003756 T ruy::Pack8bitRowMajorForNeon(unsigned char const*, int, int, int, int, int, int, signed char*, int, int, int*, int, int) | |
| 0000000000003544 T ruy::Kernel8bitNeonDotprodA55ish(ruy::KernelParams8bit<8, 8> const&) | |
| 0000000000003068 T ruy::Kernel8bitNeonA55ish(ruy::KernelParams8bit<4, 4> const&) | |
| 0000000000002708 T ruy::Kernel8bitNeon(ruy::KernelParams8bit<4, 4> const&) | |
| 0000000000001704 T ruy::Pack8bitColMajorForNeonDotprod(void const*, void const*, void const*, void const*, int, int, int, int, int, int, signed char*, int*, int) | |
| 0000000000001316 T ruy::Kernel8bitNeon1Col(ruy::KernelParams8bit<4, 4> const&) | |
| 0000000000001260 T ruy::Kernel8bitNeonDotprod1Col(ruy::KernelParams8bit<8, 8> const&) | |
| 0000000000001208 T ruy::MakeBlockMap(int, int, int, int, int, int, int, int, ruy::CpuCacheParams const&, ruy::BlockMap*) | |
| 0000000000001108 t ruy::(anonymous namespace)::TrMulTask::Run() |
| #!/bin/bash | |
| # Copyright 2020 Google LLC | |
| # | |
| # Licensed 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 | |
| # | |
| # https://www.apache.org/licenses/LICENSE-2.0 | |
| # |
| * 86.55% FullyConnected | |
| * 86.53% cpu_backend_gemm::Gemm | |
| * 86.53% Mul | |
| * 41.12% matmul shape: 128x512x384 | |
| * 41.12% TrMul (Path=0x20, max_num_threads=1, is_prepacked=(0,0)) | |
| * 41.12% TrMulImpl, general case | |
| * 36.59% Kernel (kNeon) | |
| * 4.47% Pack (kNeon) | |
| * 0.04% [other] | |
| * 0.02% MakeBlockMap |
Visualization of the MobileNet-v3-large shapes (ordered similarly by decreasing time percentage, so the most important shapes come first).
