diff --git a/CMakeLists.txt b/CMakeLists.txt index 7759cfa885299a0f53405cb3c8fe7077aee9bf48..97ce7eb15e64860701486ae00f85cc39f7d97cb0 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -123,11 +123,21 @@ set(KLEIDIAI_FILES_NEON kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxps1s0_qsu4cxs1s0_neon.c +) +set(KLEIDIAI_FILES_NEON_DOTPROD_ASM + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod_asm.S + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod_asm.S + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod_asm.S + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod_asm.S + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c ) set(KLEIDIAI_FILES_NEON_DOTPROD - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c + ${KLEIDIAI_FILES_NEON_DOTPROD_ASM} kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c @@ -136,32 +146,29 @@ set(KLEIDIAI_FILES_NEON_DOTPROD kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod_asm.S - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod_asm.S - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod_asm.S - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod_asm.S +) + +set(KLEIDIAI_FILES_NEON_I8MM_ASM + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm_asm.S + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm_asm.S + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm_asm.S + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c ) set(KLEIDIAI_FILES_NEON_I8MM + ${KLEIDIAI_FILES_NEON_I8MM_ASM} kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.c - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.c - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.c kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm_asm.S - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm_asm.S - kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm_asm.S ) set(KLEIDIAI_FILES_SME @@ -196,13 +203,7 @@ add_library(kleidiai) target_sources(kleidiai PRIVATE ${KLEIDIAI_FILES_SCALAR}) # Selectively enable architecture features. -# -# Note: MSVC compiler does not support inline assembly for ArmĀ® processors hence it cannot be used to compile the -# optimized KleidiAI micro-kernels. It can however still be used to compile scalar routines. -# https://learn.microsoft.com/en-us/cpp/assembler/inline/inline-assembler?view=msvc-170 -# -if((CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "ARM64" OR CMAKE_SYSTEM_NAME STREQUAL "iOS") AND NOT MSVC) - +if(NOT MSVC) target_sources(kleidiai PRIVATE ${KLEIDIAI_FILES_NEON}) target_sources(kleidiai PRIVATE ${KLEIDIAI_FILES_NEON_FP16}) target_sources(kleidiai PRIVATE ${KLEIDIAI_FILES_NEON_BF16}) @@ -223,6 +224,20 @@ if((CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL # Use -fno-tree-vectorize option to disable compiler based vectorization set_source_files_properties(${KLEIDIAI_FILES_SME} PROPERTIES COMPILE_OPTIONS "-fno-tree-vectorize;-march=armv8.2-a+sve+sve2${KLEIDIAI_INTERNAL_EXTRA_ARCH}") set_source_files_properties(${KLEIDIAI_FILES_SME2} PROPERTIES COMPILE_OPTIONS "-fno-tree-vectorize;-march=armv8.2-a+sve+sve2${KLEIDIAI_INTERNAL_EXTRA_ARCH}") +else() + target_sources(kleidiai PRIVATE ${KLEIDIAI_FILES_NEON_DOTPROD_ASM}) + target_sources(kleidiai PRIVATE ${KLEIDIAI_FILES_NEON_I8MM_ASM}) + + set_source_files_properties(${KLEIDIAI_FILES_SCALAR} PROPERTIES COMPILE_OPTIONS /arch:armv8.0${KLEIDIAI_INTERNAL_EXTRA_ARCH}) + set_source_files_properties(${KLEIDIAI_FILES_NEON_I8MM_ASM} PROPERTIES COMPILE_OPTIONS /arch:armv8.2${KLEIDIAI_INTERNAL_EXTRA_ARCH}) + set_source_files_properties(${KLEIDIAI_FILES_NEON_DOTPROD_ASM} PROPERTIES COMPILE_OPTIONS /arch:armv8.2${KLEIDIAI_INTERNAL_EXTRA_ARCH}) + + set(KLEIDIAI_FILES_ASM + ${KLEIDIAI_FILES_NEON_DOTPROD_ASM} + ${KLEIDIAI_FILES_NEON_I8MM_ASM}) + list(FILTER KLEIDIAI_FILES_ASM INCLUDE REGEX "^.*\.S$") + + set_source_files_properties(${KLEIDIAI_FILES_ASM} PROPERTIES LANGUAGE ASM_MARMASM) endif() target_include_directories(kleidiai