perf: Improve CPU performance on x64 and arm64

.github/scripts/build-cpu.sh

@@ -4,7 +4,11 @@ declare build_os
 
 set -xeuo pipefail
 
-pip install cmake==3.28.3
+if [[ "${build_os}" == windows* ]]; then
+    pip install cmake==3.30.9
+else
+    pip install cmake==3.28.3
+fi
 
 if [ "${build_os:0:5}" == macos ] && [ "${build_arch}" == aarch64 ]; then
 	cmake -DCMAKE_OSX_ARCHITECTURES=arm64 -DCOMPUTE_BACKEND=cpu .

CMakeLists.txt

@@ -118,6 +118,11 @@ if (BUILD_CPU)
     set(CMAKE_CXX_STANDARD 17)
     set(CMAKE_CXX_STANDARD_REQUIRED ON)
     string(TOLOWER "${CMAKE_SYSTEM_PROCESSOR}" HOST_ARCH)
+    if(MSVC)
+        # Use the experimental OpenMP runtime for persistent thread pool support.
+        # Requires CMake 3.30+; silently ignored on older CMake versions.
+        set(OpenMP_RUNTIME_MSVC "experimental")
+    endif()
     find_package(OpenMP)
 endif()
 
@@ -350,35 +355,52 @@ set_source_files_properties(${CPP_FILES} PROPERTIES LANGUAGE CXX)
 add_library(bitsandbytes SHARED ${SRC_FILES})
 target_compile_features(bitsandbytes PUBLIC cxx_std_17)
 target_include_directories(bitsandbytes PUBLIC csrc)
+set_target_properties(bitsandbytes PROPERTIES VISIBILITY_INLINES_HIDDEN ON)
 
 if (BUILD_CPU)
+    include(CheckIPOSupported)
+    check_ipo_supported(RESULT ipo_supported OUTPUT ipo_output)
+    if (ipo_supported AND NOT MSVC)
+        set_property(TARGET bitsandbytes PROPERTY INTERPROCEDURAL_OPTIMIZATION TRUE)
+    endif()
+
     if (OpenMP_CXX_FOUND)
         target_link_libraries(bitsandbytes PRIVATE OpenMP::OpenMP_CXX)
         add_definitions(-DHAS_OPENMP)
     endif()
 
-    if ((HOST_ARCH MATCHES "x86_64|amd64") AND (NOT MSVC))
-        include(CheckCXXCompilerFlag)
-        check_cxx_compiler_flag(-mavx512f HAS_AVX512F_FLAG)
-        check_cxx_compiler_flag(-mavx512bf16 HAS_AVX512BF16_FLAG)
-        if (HAS_AVX512F_FLAG)
-            target_compile_options(bitsandbytes PRIVATE -mavx512f)
-            target_compile_options(bitsandbytes PRIVATE -mavx512dq)
-            target_compile_options(bitsandbytes PRIVATE -mavx512bw)
-            target_compile_options(bitsandbytes PRIVATE -mavx512vl)
+    if (NOT MSVC)
+        if (CMAKE_SYSTEM_NAME STREQUAL "Linux")
+            target_compile_options(bitsandbytes PRIVATE -fno-semantic-interposition)
         endif()
-        if (HAS_AVX512BF16_FLAG)
-            target_compile_options(bitsandbytes PRIVATE -mavx512bf16)
+
+        if (HOST_ARCH MATCHES "x86_64|amd64")
+            include(CheckCXXCompilerFlag)
+            check_cxx_compiler_flag(-mavx512f HAS_AVX512F_FLAG)
+            check_cxx_compiler_flag(-mavx512bf16 HAS_AVX512BF16_FLAG)
+            if (HAS_AVX512F_FLAG)
+                target_compile_options(
+                    bitsandbytes PRIVATE
+                    -mavx512f
+                    -mavx512bw
+                    -mavx512dq
+                    -mavx512vl
+                )
+            endif()
+            if (HAS_AVX512BF16_FLAG)
+                target_compile_options(bitsandbytes PRIVATE -mavx512bf16)
+            endif()
+            target_compile_options(
+                bitsandbytes PRIVATE
+                -mprefer-vector-width=256
+                -mfma
+                -mavx2
+                -mf16c
+                -mlzcnt
+                -mbmi
+                -mbmi2
+            )
         endif()
-        target_compile_options(
-            bitsandbytes PRIVATE
-            -mprefer-vector-width=256
-            -mfma
-            -mavx2
-            -mlzcnt
-            -mbmi
-            -mbmi2
-        )
     endif()
 endif()
 

bitsandbytes/backends/cpu/ops.py

@@ -20,8 +20,10 @@
 # However, we can overflow if we use this without AVX512_VNNI support.
 # This is fixed in torch 2.6+, so we set this as the minimum to be safe.
 # For more information: https://github.com/pytorch/pytorch/pull/136942
-# TODO(matthewdouglas): aarch64?
-if torch.__version__ >= (2, 6):
+#
+# Without AVX-512 (including aarch64), torch._int_mm uses a scalar fallback
+# that is much slower than fp32 matmul. Only use it when AVX-512 is available.
+if torch.__version__ >= (2, 6) and _has_avx512:
 
     @register_kernel("bitsandbytes::int8_linear_matmul", "cpu")
     def _(A: torch.Tensor, B: torch.Tensor):