Add ARM architecture support for adsim (#522)

Summary:
Pull Request resolved: #522

- Extend build options and instruction to support ARM beyond x86
- Correct package installation errors in Ubuntu environment
- Fix failed cp operations that blocked build process
- Upgrade FBGEMM version to 1.4 for consistency with fbgemm benchmark

Pull Request resolved: #258

Test Plan: Tested on x86 Bergamo and aarch64 t11_GRC_ARM and the installation works for both.

Reviewed By: excelle08

Differential Revision: D96225562

Pulled By: ahmadelyoussef

fbshipit-source-id: a0e3135763d6b2959bd38d0a1dcb49ef54966040

Author: zaibie01

packages/adsim/buildfiles/fbgemm/cmake.txt

@@ -39,6 +39,9 @@ cp libfbgemm.so* "${ADSIM_STAGING_DIR}/lib/" 2>/dev/null || true
 # Copy headers from source
 cp -r ../include/fbgemm "${ADSIM_STAGING_DIR}/include/" 2>/dev/null || true
 
+# Copy asmjit headers (new dependency in FBGEMM v1.4.0)
+cp -r ../external/asmjit/src/asmjit "${ADSIM_STAGING_DIR}/include/"
+
 echo "[BUILD] FBGEMM files copied to staging directory"
 
 # Try make install but don't fail if it errors

packages/adsim/install_fbgemm.sh

@@ -16,10 +16,10 @@
 MINICONDA_PREFIX=${FBGEMM_STAGING_DIR}/miniconda
 
 # Version of FBGEMM to install
-FBGEMM_VERSION=v1.2.0
+FBGEMM_VERSION=v1.4.0
 
 # Version of PyTorch to install
-PYTORCH_VERSION=2.7.0
+PYTORCH_VERSION=2.8.0
 
 MINICONDA_VERSION="5.1-0"
 

packages/adsim/patches/treadmill.patch

@@ -1,7 +1,7 @@
 diff -wbBdu -ruN '--exclude=.git' '--exclude=*.rej' '--exclude=*.orig' '--exclude=gen-cpp2' '--exclude=build' '--exclude=third_party' treadmill-src/build.sh treadmill/build.sh
 --- treadmill-src/build.sh	1969-12-31 16:00:00.000000000 -0800
 +++ treadmill/build.sh	2025-08-04 15:44:21.731476550 -0700
-@@ -0,0 +1,42 @@
+@@ -0,0 +1,47 @@
 +#!/bin/bash
 +
 +# Exit on error
@@ -15,7 +15,12 @@ diff -wbBdu -ruN '--exclude=.git' '--exclude=*.rej' '--exclude=*.orig' '--exclud
 +INSTALL=false
 +
 +
-+sudo dnf install -y numactl-devel
++if command -v dnf >/dev/null 2>&1; then
++    sudo dnf install -y numactl-devel
++elif command -v apt-get >/dev/null 2>&1; then
++    sudo apt-get update
++    sudo apt-get install -y libnuma-dev numactl
++fi
 +# Create build directory
 +mkdir -p build
 +cd build
@@ -6977,7 +6982,7 @@ diff -wbBdu -ruN '--exclude=.git' '--exclude=*.rej' '--exclude=*.orig' '--exclud
 diff -wbBdu -ruN '--exclude=.git' '--exclude=*.rej' '--exclude=*.orig' '--exclude=gen-cpp2' '--exclude=build' '--exclude=third_party' treadmill-src/src/Scheduler.cpp treadmill/src/Scheduler.cpp
 --- treadmill-src/src/Scheduler.cpp	1969-12-31 16:00:00.000000000 -0800
 +++ treadmill/src/Scheduler.cpp	2025-08-04 15:44:21.785297704 -0700
-@@ -0,0 +1,107 @@
+@@ -0,0 +1,113 @@
 +/*
 + *  Copyright (c) 2014, Facebook, Inc.
 + *  All rights reserved.
@@ -7039,8 +7044,14 @@ diff -wbBdu -ruN '--exclude=.git' '--exclude=*.rej' '--exclude=*.orig' '--exclud
 +     to avoid memory order violation, which greatly improves its performance.
 +     http://siyobik.info.gf/main/reference/instruction/PAUSE */
 +  for (auto start = nowNs(); nowNs() - start < ns;) {
-+    asm volatile("pause");
-+  }
++#if defined(__x86_64__) || defined(__i386__)
++          asm volatile("pause");
++#elif defined(__aarch64__) || defined(__arm64__)
++          asm volatile("yield" ::: "memory");
++#else
++          asm volatile("" ::: "memory");
++#endif
++}
 +}
 +
 +/**

packages/adsim/src/cpp2/server/CMakeLists.txt

@@ -9,8 +9,14 @@ elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64")
 endif()
 
 target_compile_definitions(folly_memcpy_obj PRIVATE FOLLY_MEMCPY_IS_MEMCPY)
-target_compile_options(folly_memcpy_obj PRIVATE
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64")
+    target_compile_options(folly_memcpy_obj PRIVATE
     -mavx2 -march=haswell)
+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64")
+    target_compile_options(folly_memcpy_obj PRIVATE
+    -march=armv8.5-a+sve2
+    -mcpu=native)
+endif()
 
 # Data objects library
 add_library(data_objects DataObjects.cpp)
@@ -61,13 +67,18 @@ target_link_libraries(adsim_server
         Atomic::Atomic
         ${FMT_LIBRARIES}
         ${JEMALLOC_LIB}
-        $<LINK_LIBRARY:WHOLE_ARCHIVE,folly_memcpy_obj>
 )
 
+# Link folly_memcpy_obj only on x86 — on ARM the custom memmove causes
+# R_AARCH64_CALL26 relocation overflows and is a no-op anyway
+# (FOLLY_MEMCPY_IS_MEMCPY is defined).
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64")
+    target_link_libraries(adsim_server
+        PUBLIC $<LINK_LIBRARY:WHOLE_ARCHIVE,folly_memcpy_obj>)
+    target_compile_options(adsim_server PRIVATE
+        -fno-builtin-mempcpy
+        -fno-builtin-memmove)
+endif()
 
 # Linker and compiler options
 target_link_options(adsim_server PRIVATE "-Wl,--export-dynamic")
-target_compile_options(adsim_server PRIVATE
-    -fno-builtin-mempcpy
-    -fno-builtin-memmove
-)

packages/adsim/src/cpp2/server/dwarfs/CMakeLists.txt

@@ -20,12 +20,19 @@ add_library(gemm GEMM.cc GEMM.h
     ${FBGEMM_SRC_DIR}/test/QuantizationHelpers.cc
 )
 
-target_compile_options(gemm PRIVATE
-    ${COROUTINES_FLAG}
-    -m64
-    -mavx2
-    -mfma
-    -masm=intel)
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64")
+    target_compile_options(gemm PRIVATE
+        ${COROUTINES_FLAG}
+        -m64
+        -mavx2
+        -mfma
+        -masm=intel)
+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64")
+    target_compile_options(gemm PRIVATE
+        ${COROUTINES_FLAG}
+        -march=armv8.5-a+sve2
+        -mcpu=native)
+endif()
 target_link_directories(gemm
     PUBLIC
         ${ADSIM_STAGING_DIR}/include
@@ -40,12 +47,19 @@ add_dependencies(gemm fbgemm)
 
 add_library(embedding Embedding.cc Embedding.h)
 
-target_compile_options(embedding PRIVATE
-    ${COROUTINES_FLAG}
-    -m64
-    -mavx2
-    -mfma
-    -masm=intel)
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64")
+    target_compile_options(gemm PRIVATE
+        ${COROUTINES_FLAG}
+        -m64
+        -mavx2
+        -mfma
+        -masm=intel)
+elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64")
+    target_compile_options(gemm PRIVATE
+        ${COROUTINES_FLAG}
+        -march=armv8.5-a+sve2
+        -mcpu=native)
+endif()
 target_link_directories(embedding
     PUBLIC
         ${ADSIM_STAGING_DIR}/include