cyb70289

diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 9a4ee49..137e3ea 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -337,6 +337,11 @@ set(GGML_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of
 set(GGML_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
 set(GGML_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
 
+foreach(lib "ggml" "ggml-base")
+    target_link_libraries(${lib} PUBLIC numa)

cyb70289

[root@lnd-jm02 iscsi-hardlock]# ./test-lse
Counters:    17171    48463    25182    21011    24232    24657    20559    15806    22230    42319    32158    16342    33570    29741    26467    19637    24702    26923
Counters:    32542    38140    21447    21856    25708    21327    13062    23763    24566    35316    36268    24727    28855    23722    34471    28493    17219    13948
Counters:    39521    29413    29365    26218    24139    10212    34962    16706     6899    30642    54875    30239    29795    22390    13964    29797    14388    21804
Counters:    39066    24223    10731    19933    41360    19116    17735    38524    24549    34395    22350    21955    28910    27429    33503    20678    24180    16701
Counters:    45116    30269    18649    32161    14482    25578    34041    29304    38479    20491    30742    29352    26589    17246    22371    23332     9264    18094
Counters:    35308    41495    16516    28869    27017    19990    22366    28377    18060    13942    29316    25233    

cyb70289

From 2eee837216e575abd9e48a30c65161a42ad59117 Mon Sep 17 00:00:00 2001
From: Yibo Cai <[email protected]>
Date: Wed, 21 Aug 2024 06:40:48 -0400
Subject: [PATCH 2/2] optimize varint with lookup table

---
 src/google/protobuf/io/coded_stream.h | 20 ++++++++++++--------
 1 file changed, 12 insertions(+), 8 deletions(-)

diff --git a/src/google/protobuf/io/coded_stream.h b/src/google/protobuf/io/coded_stream.h

cyb70289

From a5239aaf894334f4b4c331f1d40245f5f64a2cd4 Mon Sep 17 00:00:00 2001
From: Yibo Cai <[email protected]>
Date: Wed, 21 Aug 2024 06:40:29 -0400
Subject: [PATCH 1/2] add string list and map benchmarks

---
 benchmarks/benchmark.cc     | 65 +++++++++++++++++++++++++++++++++++++
 benchmarks/descriptor.proto |  8 +++++
 2 files changed, 73 insertions(+)

cyb70289

diff --git a/cmake/set_arch_flags.cmake b/cmake/set_arch_flags.cmake
index 538ddfe..81c40e4 100644
--- a/cmake/set_arch_flags.cmake
+++ b/cmake/set_arch_flags.cmake
@@ -2,8 +2,8 @@ function(set_arch_flags target arch)
   message(STATUS "Setting architecture flags for ${arch}")
   if(arch MATCHES "x86_64")
     target_compile_options(${target} PRIVATE -mavx2 -mpclmul -mbmi -mlzcnt)
-  elseif(arch MATCHES "arm")
-    target_compile_options(${target} PRIVATE -march=armv8-a)

cyb70289

diff --git a/cmake/set_arch_flags.cmake b/cmake/set_arch_flags.cmake
index 538ddfe..6dc7754 100644
--- a/cmake/set_arch_flags.cmake
+++ b/cmake/set_arch_flags.cmake
@@ -2,8 +2,8 @@ function(set_arch_flags target arch)
   message(STATUS "Setting architecture flags for ${arch}")
   if(arch MATCHES "x86_64")
     target_compile_options(${target} PRIVATE -mavx2 -mpclmul -mbmi -mlzcnt)
-  elseif(arch MATCHES "arm")
-    target_compile_options(${target} PRIVATE -march=armv8-a)

cyb70289

diff --git a/src/cpu/aarch64/jit_uni_reorder.cpp b/src/cpu/aarch64/jit_uni_reorder.cpp
index 5c7c6d5..dc55c69 100644
--- a/src/cpu/aarch64/jit_uni_reorder.cpp
+++ b/src/cpu/aarch64/jit_uni_reorder.cpp
@@ -2680,6 +2680,55 @@ status_t jit_uni_reorder_t::pd_t::create(reorder_pd_t **reorder_pd,
     return safe_ptr_assign(*reorder_pd, _pd.release());
 }
 
+#define MY_REORDER
+

cyb70289

##################################################################
# profile.py
##################################################################

import tensorflow as tf
import timeit
import os

n_threads = int(os.getenv('OMP_NUM_THREADS'))
if n_threads < 1 or n_threads > 999:

cyb70289

Bluewhale memory latency vs. bandwidth
======================================

max bw = 574*8*32 MB/s = 147GB/s

bw = 0%
-------
$ numactl -m0 -N0 /usr/lib/lmbench/bin/lat_mem_rd -P 1 512 4096
"stride=4096
0.00391 1.429

cyb70289

// tested with g++-10.5, probably okay for other versions as
// the code is quite simple, check assembly to make sure

// g++ -std=c++11 -O3 -pthread -static bw-test.cc -o bw-test
// XXX: it costs about half minute to compile this file

#include <cstdlib>
#include <iostream>
#include <thread>
#include <vector>